repo
stringlengths
2
99
file
stringlengths
13
225
code
stringlengths
0
18.3M
file_length
int64
0
18.3M
avg_line_length
float64
0
1.36M
max_line_length
int64
0
4.26M
extension_type
stringclasses
1 value
inversegraphics
inversegraphics-master/test.py
#!/usr/bin/env python3.4m import matplotlib matplotlib.use('Agg') import scene_io_utils from blender_utils import * from generative_models import * from tabulate import tabulate import matplotlib.pyplot as plt from matplotlib.font_manager import FontProperties import cv2 plt.ioff() numpy.random.seed(1) inchToMeter = 0.0254 rendersDir = '../data/output/' baseTestDir = '../data/aztest/' groundTruth, imageFiles, segmentFiles,segmentSingleFiles, unoccludedFiles, prefixes = loadGroundTruth(rendersDir) width = 110 height = 110 useCycles = False useGPU = False distance = 0.45 numSamples = 16 originalLoc = mathutils.Vector((0,-distance , 0)) numpy.random.seed(1) minThresTemplate = 10 maxThresTemplate = 100 minThresImage = 50 maxThresImage = 150 baseDir = '../databaseFull/models/' experimentTeapots = [2] outputExperiments = [] # distanceTypes = ['chamferDataToModel', 'robustChamferDataToModel', 'sqDistImages', 'robustSqDistImages'] distanceTypes = ['negLogLikelihood'] # distanceTypes = ['negLogLikelihood','negLogLikelihoodRobust'] masks = numpy.array([]) segmentImages = [] for segment in segmentSingleFiles: print(segment) segmentImg = cv2.imread(segment, cv2.IMREAD_ANYDEPTH)/255.0 segmentImg = segmentImg[..., numpy.newaxis] segmentImages = segmentImages + [segmentImg] masks = numpy.concatenate([aux for aux in segmentImages], axis=-1) layerPrior = globalLayerPrior(masks) backgroundModels = ['UNOCCLUDED'] # backgroundModels = ['SINGLE', 'UNOCCLUDED'] completeScene = True if 'FULL' not in backgroundModels: completeScene = False # sortedScenesIndices = numpy.argsort(groundTruth[:,6]) # sortedGroundTruth = groundTruth[sortedScenesIndices, :] # sortedMasks = masks[:,:,sortedScenesIndices] # sortedImageFiles = imageFiles[sortedScenesIndices] filterScenes = [19] indicesOccluded = (groundTruth[:,5] < 0.90) & (groundTruth[:,5] > 0.05) indicesScenes = (groundTruth[:,6] != 19) partsOccluded = (groundTruth[:,14] == 1) | (groundTruth[:,13] == 1) indicesPrefix = [ x==y for (x,y) in zip(prefixes, ['_occluded']*len(prefixes))] [targetScenes, targetModels, transformations] = scene_io_utils.loadTargetModels(experimentTeapots) numExperiments = 100 spout = mathutils.Vector((-6.2, -0.16, 6.25)) handle = mathutils.Vector((6.2, 0.2, 5.7)) tip = mathutils.Vector((0, 0, 8)) for teapotIdx, teapotTest in enumerate(experimentTeapots): teapot = targetModels[teapotIdx] teapot.layers[0] = True teapot.layers[1] = True transformation = transformations[teapotIdx] spout = transformation * spout handle = transformation * handle tip = transformation * tip print("Experiment on teapot " + teapot.name) indicesTeapot = (groundTruth[:, 3] == teapotTest) indices = numpy.where(indicesTeapot & indicesOccluded & indicesScenes & partsOccluded ) selTest = indices[0] numTests = len(selTest) print ("Total of " + str(numTests) + " tests") # expSelTest = [0,1] numExperiments = min(numTests, numExperiments) expSelTest = numpy.array(numpy.arange(0,numTests-1,int(numTests-1)/numExperiments), dtype=numpy.int) # numpy.arange(0,numTests,int(numTests)) print ("Running only " + str(len(expSelTest)) + " tests") # selTest = numpy.random.permutation(selTest) currentScene = -1 performance = {} elevations = {} groundTruthRelAzimuths = {} groundTruthAzimuths = {} bestRelAzimuths= {} bestAzimuths= {} occlusions = {} for backgroundModelOut in backgroundModels: for distanceTypeOut in distanceTypes: performance[(backgroundModelOut, distanceTypeOut)] = numpy.array([]) elevations[(backgroundModelOut, distanceTypeOut)] = numpy.array([]) groundTruthRelAzimuths[(backgroundModelOut, distanceTypeOut)] = numpy.array([]) groundTruthAzimuths[(backgroundModelOut, distanceTypeOut)] = numpy.array([]) bestRelAzimuths[(backgroundModelOut, distanceTypeOut)]= numpy.array([]) bestAzimuths[(backgroundModelOut, distanceTypeOut)] = numpy.array([]) occlusions[(backgroundModelOut, distanceTypeOut)] = numpy.array([]) for selTestNum in expSelTest: test = selTest[selTestNum] testResults = {} testAzimuths = {} groundTruthAz = groundTruth[test,0] groundTruthObjAz = groundTruth[test,1] groundTruthRelAz = numpy.arctan2(numpy.sin((groundTruthAz-groundTruthObjAz)*numpy.pi/180), numpy.cos((groundTruthAz-groundTruthObjAz)*numpy.pi/180))*180/numpy.pi groundTruthEl = groundTruth[test,2] occlusion = groundTruth[test,5] sceneNum = int(groundTruth[test,6]) targetIndex= int(groundTruth[test,7]) sampleDir = baseTestDir + 'teapot' + str(teapotTest) + '/' 'test_samples/num' + str(test) + '_azim' + str(int(groundTruthAz)) + '_elev' + str(int(groundTruthEl)) + '_occlusion' + str(occlusion) + '/' if not os.path.exists(sampleDir): os.makedirs(sampleDir) if currentScene != sceneNum: if currentScene != -1: # Cleanup for objnum, obji in enumerate(scene.objects): if obji.name != teapot.name and obji.type == 'EMPTY' and obji.dupli_type == 'GROUP': deleteInstance(obji) elif obji.type == 'LAMP': obji.data.user_clear() bpy.data.lamps.remove(obji.data) obji.user_clear() bpy.data.objects.remove(obji) elif obji.name != teapot.name: obji.user_clear() bpy.data.objects.remove(obji) camera.user_clear() bpy.data.objects.remove(camera) cam.user_clear() bpy.data.cameras.remove(cam) world.user_clear() bpy.data.worlds.remove(world) bpy.context.screen.scene = bpy.data.scenes['Scene'] scene.user_clear() bpy.data.scenes.remove(scene) currentScene = sceneNum sceneFileName = '{0:05d}'.format(sceneNum) print("Experiment on scene " + sceneFileName) instances = scene_io_utils.getSceneInstancesInfo('../databaseFull/scenes/scene' + sceneFileName + '.txt') targetParentPosition = instances[targetIndex][2] targetParentIndex = instances[targetIndex][1] [blenderScenes, modelInstances] = scene_io_utils.importBlenderScenes(instances, completeScene, targetIndex) # targetParentInstance = modelInstances[targetParentIndex] roomName = '' for model in modelInstances: reg = re.compile('(room[0-9]+)') res = reg.match(model.name) if res: roomName = res.groups()[0] scene = scene_io_utils.composeScene(modelInstances, targetIndex) # bpy.context.scene.use_nodes = True # tree = bpy.context.scene.node_tree # rl = tree.nodes[0] # links = tree.links # v = tree.nodes.new('CompositorNodeViewer') # v.location = 750,210 # v.use_alpha = False # links.new(rl.outputs[0], v.inputs[0]) # link Image output to Viewer input scene.update() scene.render.threads = 8 scene.render.threads_mode = 'AUTO' bpy.context.screen.scene = scene cycles = bpy.context.scene.cycles scene.render.tile_x = 55 scene.render.tile_y = 55 cam = bpy.data.cameras.new("MainCamera") camera = bpy.data.objects.new("MainCamera", cam) world = bpy.data.worlds.new("MainWorld") setupScene(scene, targetIndex, roomName, world, distance, camera, width, height, numSamples, useCycles, useGPU) scene.objects.link(teapot) teapot.layers[1] = True scene.update() teapot.matrix_world = mathutils.Matrix.Translation(targetParentPosition) center = centerOfGeometry(teapot.dupli_group.objects, teapot.matrix_world) original_matrix_world = teapot.matrix_world.copy() teapot.matrix_world = original_matrix_world # bpy.ops.mesh.primitive_uv_sphere_add(size=0.005, location=spout) # sphereSpout = scene.objects[0] # bpy.ops.mesh.primitive_uv_sphere_add(size=0.005, location=handle) # sphereHandle = scene.objects[0] # bpy.ops.mesh.primitive_uv_sphere_add(size=0.005, location=tip) # sphereTip = scene.objects[0] # sphereTip.layers[1] = True # sphereSpout.layers[1] = True # sphereHandle.layers[1] = True azimuthRot = mathutils.Matrix.Rotation(radians(-groundTruthObjAz), 4, 'Z') groundTruthTransf = mathutils.Matrix.Translation(original_matrix_world.to_translation()) * azimuthRot * (mathutils.Matrix.Translation(-original_matrix_world.to_translation())) teapot.matrix_world = groundTruthTransf * original_matrix_world sphereSpout_matrix_world = teapot.matrix_world * mathutils.Matrix.Translation(spout) sphereHandle_matrix_world = teapot.matrix_world * mathutils.Matrix.Translation(handle) sphereTip_matrix_world = teapot.matrix_world * mathutils.Matrix.Translation(tip) for backgroundModel in backgroundModels: variances = numpy.ones([height, width, 3])*2/255.0 print("Experiment on background model " + backgroundModel) scene.layers[1] = True scene.layers[0] = False scene.render.layers[0].use = False scene.render.layers[1].use = True teapot.layers[1] = True if backgroundModel == 'FULL': scene.layers[1] = False scene.layers[0] = True scene.render.layers[0].use = True scene.render.layers[1].use = False teapot.layers[0] = True scene.update() sqDistsSeq = [] for distIdx, distanceType in enumerate(distanceTypes): print("Experiment on model " + distanceType) computingSqDists = False if distIdx == 0: computingSqDists = True if not computingSqDists and not backgroundModel == 'FULL': sqRes = numpy.concatenate([aux[..., numpy.newaxis] for aux in sqDistsSeq], axis=-1) variances = computeVariances(sqRes) variances[numpy.where(variances <= 1)] = 2.0/255.0 # robust = not robust # if robust is False: # robustScale = 0 scores = [] relAzimuths = [] azimuths = [] directory = baseTestDir + 'teapot' + str(teapotTest) + '/' + backgroundModel + '_' + distanceType if not os.path.exists('../data/aztest/' + 'teapot' + str(teapotTest) + '/'): os.makedirs('../data/aztest/' + 'teapot' + str(teapotTest) + '/') if not os.path.exists(directory + 'test_samples'): os.makedirs(directory + 'test_samples') numDir = directory + 'test_samples/num' + str(test) + '_azim' + str(int(groundTruthAz)) + '_elev' + str(int(groundTruthEl)) + '_occlusion' + str(occlusion) + '/' if not os.path.exists(numDir): os.makedirs(numDir) testImage = cv2.imread(imageFiles[test])/255.0 if backgroundModel == 'UNOCCLUDED': testImage = cv2.imread(unoccludedFiles[test])/255.0 # testImage = cv2.cvtColor(numpy.float32(rgbTestImage*255), cv2.COLOR_RGB2BGR)/255.0 testMask = masks[:,:,test] testImageEdges = cv2.Canny(numpy.uint8(testImage*255), minThresImage,maxThresImage) cv2.imwrite(numDir + "image_canny_" + backgroundModel + ".png" , testImageEdges) cv2.imwrite(numDir + "image_" + backgroundModel + ".png" , numpy.uint8(testImage*255)) cv2.imwrite(sampleDir + "image_canny_" + backgroundModel + ".png" , testImageEdges) cv2.imwrite(sampleDir + "image_" + backgroundModel + ".png" , numpy.uint8(testImage*255)) score = numpy.finfo(numpy.float64).max elevation = groundTruthEl azimuth = 0 elevationRot = mathutils.Matrix.Rotation(radians(-elevation), 4, 'X') for azimuth in numpy.arange(0,360,5): azimuthRot = mathutils.Matrix.Rotation(radians(-azimuth), 4, 'Z') elevationRot = mathutils.Matrix.Rotation(radians(-elevation), 4, 'X') location = center + azimuthRot * elevationRot * originalLoc camera.location = location scene.update() look_at(camera, center) scene.update() # projMat = projection_matrix(scene.camera.data, scene) # ipdb.set_trace() spoutlocation = image_project(scene, camera, sphereSpout_matrix_world.to_translation()) handlelocation = image_project(scene, camera,sphereHandle_matrix_world.to_translation()) spherelocation = image_project(scene, camera, sphereTip_matrix_world.to_translation()) # touched = closestCameraIntersection(scene, sphereSpout.matrix_world.to_translation()) # scene.objects.unlink(sphereSpout) # scene.objects.unlink(sphereHandle) # scene.objects.unlink(sphereTip) result, object, matrix, location, normal = scene.ray_cast(scene.camera.location, sphereSpout_matrix_world.to_translation()) scene.render.filepath = numDir + '_blender.png' bpy.ops.render.render( write_still=True ) # image2 = numpy.flipud(numpy.array(bpy.data.images['Viewer Node'].extract_render(scene=scene)).reshape([256,256,4]))[:,:,0:3] image = cv2.imread(scene.render.filepath) # blendImage = bpy.data.images['Render Result'] # # # image2 = numpy.flipud(numpy.array(blendImage.extract_render(scene=scene)).reshape([height,width,4]))[:,:,0:3] # # ipdb.set_trace() # Truncate intensities larger than 1. # image = image * 1.5 # image2[numpy.where(image2 > 1)] = 1 # ipdb.set_trace() # image[0:20, 75:100, :] = 0 # image2 = cv2.cvtColor(numpy.float32(image2), cv2.COLOR_RGB2BGR) image = image/255.0 # image2 = image2/255.0 methodParams = {'backgroundModel': backgroundModel, 'testMask': testMask, 'variances':variances, 'layerPrior': layerPrior, 'minThresImage': minThresImage, 'maxThresImage': maxThresImage, 'minThresTemplate': minThresTemplate, 'maxThresTemplate': maxThresTemplate} distance = scoreImage(testImage, image, distanceType, methodParams) # cv2.imwrite(numDir + 'image' + "_az" + '%.1f' % azimuth + '_dist' + '%.1f' % distance + '.png', numpy.uint8(image*255.0)) scores.append(distance) relAzimuth = numpy.mod(numpy.arctan2(numpy.sin((azimuth-groundTruthObjAz)*numpy.pi/180), numpy.cos((azimuth-groundTruthObjAz)*numpy.pi/180))*180/numpy.pi , 360) azimuths.append(azimuth) if distance <= score: imageEdges = cv2.Canny(numpy.uint8(image*255.0), minThresTemplate,maxThresTemplate) bestImageEdges = imageEdges bestImage = image score = distance bestRelAzimuth = relAzimuth bestAzimuth = azimuth # if robust is False: # robustScale = 1.4826 * numpy.sqrt(numpy.median(scores)) # error = numpy.arctan2(numpy.sin((groundTruthRelAz-bestRelAzimuth)*numpy.pi/180), numpy.cos((groundTruthRelAz-bestRelAzimuth)*numpy.pi/180))*180/numpy.pi error = numpy.arctan2(numpy.sin((groundTruthAz-bestAzimuth)*numpy.pi/180), numpy.cos((groundTruthAz-bestAzimuth)*numpy.pi/180))*180/numpy.pi performance[(backgroundModel, distanceType)] = numpy.append(performance[(backgroundModel, distanceType)], error) elevations[(backgroundModel, distanceType)] = numpy.append(elevations[(backgroundModel, distanceType)], elevation) bestAzimuths[(backgroundModel, distanceType)] = numpy.append(bestAzimuths[(backgroundModel, distanceType)], bestAzimuth) groundTruthAzimuths[(backgroundModel, distanceType)] = numpy.append(groundTruthAzimuths[(backgroundModel, distanceType)], groundTruthAz) occlusions[(backgroundModel, distanceType)] = numpy.append(occlusions[(backgroundModel, distanceType)], occlusion) testResults[(backgroundModel, distanceType)] = scores testAzimuths[(backgroundModel, distanceType)] = azimuths sqDist = sqDistImages(bestImage, testImage) disp = cv2.normalize(sqDist, sqDist, alpha=0, beta=255, norm_type=cv2.NORM_MINMAX, dtype=cv2.CV_8U) cv2.imwrite(numDir + 'sqDists' + "_az" + '%.1f' % bestAzimuth + '_dist' + '%.1f' % score + '.png', numpy.uint8(disp)) if computingSqDists: sqDistsSeq = sqDistsSeq + [sqDist] if distanceType == 'negLogLikelihoodRobust': pixLik = pixelLikelihoodRobust(testImage, bestImage, testMask, backgroundModel, layerPrior, variances) fig = plt.figure() plt.imshow(pixLik) plt.colorbar() fig.savefig(numDir + 'pixelLikelihoodRobustPlot' + "_az" + '%.1f' % bestAzimuth + '_dist' + '%.1f' % score + '.png') plt.close(fig) disp = cv2.normalize(pixLik, pixLik, alpha=0, beta=255, norm_type=cv2.NORM_MINMAX, dtype=cv2.CV_8U) cv2.imwrite(numDir + 'pixelLikelihoodRobust' + "_az" + '%.1f' % bestAzimuth + '_dist' + '%.1f' % score + '.png', numpy.uint8(disp)) fgpost, bgpost = layerPosteriorsRobust(testImage, bestImage, testMask, backgroundModel, layerPrior, variances) # plt.imshow(testImage) # plt.show() # plt.imshow(bestImage) # plt.show() # z = fgpost + bgpost # plt.imshow(z) # plt.show() # plt.imshow(fgpost) # plt.show() # plt.imshow(bgpost) # plt.show() # ipdb.set_trace assert(numpy.abs(numpy.sum(fgpost + bgpost - testMask)) < 0.01) fig = plt.figure() plt.imshow(fgpost) plt.colorbar() fig.savefig(numDir + 'fgPosteriorPlot' + "_az" + '%.1f' % bestAzimuth + '_dist' + '%.1f' % score + '.png') plt.close(fig) fig = plt.figure() plt.imshow(bgpost) plt.colorbar() fig.savefig(numDir + 'bgPosteriorPlot' + "_az" + '%.1f' % bestAzimuth + '_dist' + '%.1f' % score + '.png') plt.close(fig) disp = cv2.normalize(fgpost, fgpost, alpha=0, beta=255, norm_type=cv2.NORM_MINMAX, dtype=cv2.CV_8U) cv2.imwrite(numDir + 'fgPosterior' + "_az" + '%.1f' % bestAzimuth + '_dist' + '%.1f' % score + '.png', numpy.uint8(disp)) disp = cv2.normalize(bgpost, bgpost, alpha=0, beta=255, norm_type=cv2.NORM_MINMAX, dtype=cv2.CV_8U) cv2.imwrite(numDir + 'bgPosterior' + "_az" + '%.1f' % bestAzimuth + '_dist' + '%.1f' % score + '.png', numpy.uint8(disp)) if distanceType == 'negLogLikelihood': pixLik = pixelLikelihood(testImage, bestImage, testMask, backgroundModel, variances) fig = plt.figure() plt.imshow(pixLik) plt.colorbar() fig.savefig(numDir + 'pixelLikelihoodPlot' + "_az" + '%.1f' % bestAzimuth + '_dist' + '%.1f' % score + '.png') plt.close(fig) disp = cv2.normalize(pixLik, pixLik, alpha=0, beta=255, norm_type=cv2.NORM_MINMAX, dtype=cv2.CV_8U) cv2.imwrite(numDir + 'pixelLikelihood' + "_az" + '%.1f' % bestAzimuth + '_dist' + '%.1f' % score + '.png', numpy.uint8(disp)) cv2.imwrite(numDir + 'bestImage' + "_canny_az" + '%.1f' % bestAzimuth + '_dist' + '%.1f' % score + '.png' , bestImageEdges) cv2.imwrite(numDir + 'bestImage' + "_az" + '%.1f' % bestAzimuth + '_dist' + '%.1f' % score + '.png', numpy.uint8(bestImage*255.0)) imgEdges = cv2.Canny(numpy.uint8(testImage*255), minThresImage,maxThresImage) bwEdges1 = cv2.distanceTransform(~imgEdges, cv2.DIST_L2, 5) disp = cv2.normalize(bwEdges1, bwEdges1, alpha=0, beta=255, norm_type=cv2.NORM_MINMAX, dtype=cv2.CV_8U) cv2.imwrite(numDir + 'dist_transform' + '.png', disp) # ipdb.set_trace() fig = plt.figure() plt.plot(azimuths, numpy.array(scores)) plt.xlabel('Azimuth (degrees)') plt.ylabel('Negative Log Likelihood') plt.title(distanceType) plt.axvline(x=bestAzimuth, linewidth=2, color='b', label='Minimum score azimuth') plt.axvline(x=groundTruthAz, linewidth=2, color='g', label='Ground truth azimuth') plt.axvline(x=(bestAzimuth + 180) % 360, linewidth=1, color='b', ls='--', label='Minimum distance azimuth + 180') fontP = FontProperties() fontP.set_size('small') x1,x2,y1,y2 = plt.axis() plt.axis((0,360,y1,y2)) # plt.legend() fig.savefig(numDir + 'performance.png') plt.close(fig) fig = plt.figure() backgroundModelOut = 'UNOCCLUDED' distanceTypeOut = 'negLogLikelihood' maxY = numpy.max(numpy.array(testResults[(backgroundModelOut, distanceTypeOut)])) minY = numpy.min(numpy.array(testResults[(backgroundModelOut, distanceTypeOut)])) plt.plot(testAzimuths[(backgroundModelOut, distanceTypeOut)], (numpy.array(testResults[(backgroundModelOut, distanceTypeOut)]) - minY)/ (maxY-minY), color='g', label='Clean model') plt.axvline(x=testAzimuths[(backgroundModelOut, distanceTypeOut)][numpy.argmin(numpy.array(testResults[(backgroundModelOut, distanceTypeOut)]))], linewidth=1, color='g', ls='--') backgroundModelOut = 'SINGLE' distanceTypeOut = 'negLogLikelihood' maxY = numpy.max(numpy.array(testResults[(backgroundModelOut, distanceTypeOut)])) minY = numpy.min(numpy.array(testResults[(backgroundModelOut, distanceTypeOut)])) plt.plot(testAzimuths[(backgroundModelOut, distanceTypeOut)], (numpy.array(testResults[(backgroundModelOut, distanceTypeOut)]) - minY)/ (maxY-minY), color='r', label='Normal model') plt.axvline(x=testAzimuths[(backgroundModelOut, distanceTypeOut)][numpy.argmin(numpy.array(testResults[(backgroundModelOut, distanceTypeOut)]))], linewidth=1, color='r', ls='--') backgroundModelOut = 'SINGLE' distanceTypeOut = 'negLogLikelihoodRobust' maxY = numpy.max(numpy.array(testResults[(backgroundModelOut, distanceTypeOut)])) minY = numpy.min(numpy.array(testResults[(backgroundModelOut, distanceTypeOut)])) plt.plot(testAzimuths[(backgroundModelOut, distanceTypeOut)], (numpy.array(testResults[(backgroundModelOut, distanceTypeOut)])- minY) / (maxY-minY), color='c', label='Robust model') plt.axvline(x=testAzimuths[(backgroundModelOut, distanceTypeOut)][numpy.argmin(numpy.array(testResults[(backgroundModelOut, distanceTypeOut)]))], linewidth=1, color='c', ls='--') plt.xlabel('Azimuth (degrees)') plt.ylabel('Negative Log Likelihood') fontP = FontProperties() fontP.set_size('small') # plt.axvline(x=bestAzimuth, linewidth=2, color='r', label='Minimum score azimuth') plt.axvline(x=groundTruthAz, linewidth=1, color='b', label='GT azimuth') # plt.axvline(x=(bestAzimuth + 180) % 360, linewidth=1, color='r', ls='--', label='Minimum distance azimuth + 180') x1,x2,y1,y2 = plt.axis() plt.axis((0,360,y1,y2)) lgd = plt.legend(bbox_to_anchor=(0., 1.02, 1., .102), loc=3, ncol=2, mode="expand", borderaxespad=0.) fig.savefig(sampleDir + 'performance.png', bbox_extra_artists=(lgd,), bbox_inches='tight') plt.close(fig) scene.objects.unlink(teapot) for backgroundModelOut in backgroundModels: for distanceTypeOut in distanceTypes: directory = baseTestDir + 'teapot' + str(teapotTest) + '/' + backgroundModelOut + distanceTypeOut experiment = {'distanceType':distanceTypeOut, 'backgroundModel':backgroundModelOut, 'rendersDir':rendersDir, 'scores':scores, 'azimuths':azimuths,'teapot':teapotTest, 'bestAzimuths':bestAzimuths[(backgroundModelOut, distanceTypeOut)], 'performance': performance[(backgroundModelOut, distanceTypeOut)], 'elevations':elevations[(backgroundModelOut, distanceTypeOut)], 'groundTruthAzimuths': groundTruthRelAzimuths[(backgroundModelOut, distanceTypeOut)], 'occlusions': occlusions[(backgroundModelOut, distanceTypeOut)],'selTest':selTest, 'test':test} with open(directory + 'experiment.pickle', 'wb') as pfile: pickle.dump(experiment, pfile) fig = plt.figure() plt.scatter(elevations[(backgroundModelOut, distanceTypeOut)], performance[(backgroundModelOut, distanceTypeOut)]) plt.xlabel('Elevation (degrees)') plt.ylabel('Angular error') x1,x2,y1,y2 = plt.axis() plt.axis((0,90,-180,180)) plt.title('Performance scatter plot') fig.savefig(directory + '_elev-performance-scatter.png') plt.close(fig) fig = plt.figure() plt.scatter(occlusions[(backgroundModelOut, distanceTypeOut)]*100.0, performance[(backgroundModelOut, distanceTypeOut)]) plt.xlabel('Occlusion (%)') plt.ylabel('Angular error') x1,x2,y1,y2 = plt.axis() plt.axis((0,100,-180,180)) plt.title('Performance scatter plot') fig.savefig(directory + '_occlusion-performance-scatter.png') plt.close(fig) fig = plt.figure() plt.scatter(groundTruthAzimuths[(backgroundModelOut, distanceTypeOut)], performance[(backgroundModelOut, distanceTypeOut)]) plt.xlabel('Azimuth (degrees)') plt.ylabel('Angular error') x1,x2,y1,y2 = plt.axis() plt.axis((0,360,-180,180)) plt.title('Performance scatter plot') fig.savefig(directory + '_azimuth-performance-scatter.png') plt.close(fig) fig = plt.figure() plt.hist(performance[(backgroundModelOut, distanceTypeOut)], bins=36) plt.xlabel('Angular error') plt.ylabel('Counts') x1,x2,y1,y2 = plt.axis() plt.axis((-180,180,y1, y2)) plt.title('Performance histogram') fig.savefig(directory + '_performance-histogram.png') plt.close(fig) # experimentFile = 'aztest/teapotsc7549b28656181c91bff71a472da9153Teapot N311012_cleaned.pickle' # with open(experimentFile, 'rb') as pfile: # experiment = pickle.load( pfile) headers=["Best global fit", ""] table = [["Mean angular error", numpy.mean(numpy.abs(performance[(backgroundModelOut, distanceTypeOut)]))],["Median angualar error",numpy.median(numpy.abs(performance[(backgroundModelOut, distanceTypeOut)]))]] performanceTable = tabulate(table, tablefmt="latex", floatfmt=".1f") with open(directory + 'performance.tex', 'w') as expfile: expfile.write(performanceTable) experiment = {'distanceTypes':distanceTypes, 'backgroundModels':backgroundModels, 'rendersDir':rendersDir,'teapot':teapotTest,'testResults':testResults,'testAzimuths':testAzimuths, 'bestAzimuths':bestAzimuths, 'performance': performance, 'elevations':elevations, 'groundTruthAzimuths': groundTruthRelAzimuths, 'occlusions': occlusions} with open(baseTestDir + 'teapot' + str(teapotTest) + '/' + 'experiment.pickle', 'wb') as pfile: pickle.dump(experiment, pfile) print("Finished the experiment")
29,478
47.645215
559
py
inversegraphics
inversegraphics-master/differentiable_renderer.py
import chumpy as ch from chumpy import depends_on, Ch import cv2 import numpy as np import scipy.sparse as sp from chumpy.utils import row, col from opendr.geometry import Rodrigues import warnings #Make simple experiment. def nanmean(a, axis): # don't call nan_to_num in here, unless you check that # occlusion_test.py still works after you do it! result = np.nanmean(a, axis=axis) return result def nangradients(arr): with warnings.catch_warnings(): warnings.simplefilter("ignore", category=RuntimeWarning) dy = np.expand_dims(arr[:-1,:,:] - arr[1:,:,:], axis=3) dx = np.expand_dims(arr[:,:-1,:] - arr[:, 1:, :], axis=3) dy = np.concatenate((dy[1:,:,:], dy[:-1,:,:]), axis=3) dy = np.nanmean(dy, axis=3) dx = np.concatenate((dx[:,1:,:], dx[:,:-1,:]), axis=3) dx = np.nanmean(dx, axis=3) if arr.shape[2] > 1: gy, gx, _ = np.gradient(arr) else: gy, gx = np.gradient(arr.squeeze()) gy = np.atleast_3d(gy) gx = np.atleast_3d(gx) gy[1:-1,:,:] = -dy gx[:,1:-1,:] = -dx return gy, gx class SQErrorRenderer(Ch): terms = ['renderer', 'params_list'] dterms = ['params'] def compute_r(self): return self.renderer.errors def compute_dr_wrt(self, wrt): import ipdb # ipdb.set_trace() inParamsList = False if wrt is self.params: return None if not inParamsList: for param in self.params_list: if wrt is param: inParamsList = True if inParamsList: return self.gradient_pred(wrt) else: errJac = sp.csc_matrix((2*self.renderer.r - 2*self.imageGT).ravel()[:,None]) return errJac.multiply(self.renderer.dr_wrt(wrt)) def gradient_pred(self, paramWrt): observed = self.renderer.color_image boundaryid_image = self.renderer.boundaryid_image barycentric = self.renderer.barycentric_image visibility = self.renderer.visibility_image visible = np.nonzero(visibility.ravel() != 4294967295)[0] jacIm = self.dImage_wrt_2dVerts_predict(observed, paramWrt, visible, visibility, barycentric, observed.shape[0], observed.shape[1], self.renderer.v.shape[0], self.renderer.f, boundaryid_image != 4294967295) return jacIm.dot(self.renderer.camera.dr_wrt(paramWrt)) def dImage_wrt_2dVerts_predict(self, observed, paramWrt, visible, visibility, barycentric, image_width, image_height, num_verts, f, bnd_bool): """Construct a sparse jacobian that relates 2D projected vertex positions (in the columns) to pixel values (in the rows). This can be done in two steps.""" bnd_bool = np.logical_and(bnd_bool, self.renderer.visibility_image != 4294967295) camJac = self.renderer.camera.dr_wrt(paramWrt) xdiff = self.renderer.dEdx ydiff = self.renderer.dEdy # visible.ravel()[bidxs_out[bidx]] = True # import ipdb # ipdb.set_trace() visible = np.nonzero(visibility.ravel() != 4294967295)[0] n_channels = np.atleast_3d(observed).shape[2] shape = visibility.shape #2: Take the data and copy the corresponding dxs and dys to these new pixels. # Step 1: get the structure ready, ie the IS and the JS IS = np.tile(col(visible), (1, 2*f.shape[1])).ravel() JS = col(f[visibility.ravel()[visible]].ravel()) JS = np.hstack((JS*2, JS*2+1)).ravel() pxs = np.asarray(visible % shape[1], np.int32) pys = np.asarray(np.floor(np.floor(visible) / shape[1]), np.int32) if n_channels > 1: IS = np.concatenate([IS*n_channels+i for i in range(n_channels)]) JS = np.concatenate([JS for i in range(n_channels)]) datas = [] # The data is weighted according to barycentric coordinates bc0 = col(barycentric[pys, pxs, 0]) bc1 = col(barycentric[pys, pxs, 1]) bc2 = col(barycentric[pys, pxs, 2]) for k in range(n_channels): dxs = xdiff[pys, pxs, k] dys = ydiff[pys, pxs, k] if f.shape[1] == 3: datas.append(np.hstack((col(dxs)*bc0,col(dys)*bc0,col(dxs)*bc1,col(dys)*bc1,col(dxs)*bc2,col(dys)*bc2)).ravel()) else: datas.append(np.hstack((col(dxs)*bc0,col(dys)*bc0,col(dxs)*bc1,col(dys)*bc1)).ravel()) data = np.concatenate(datas) ij = np.vstack((IS.ravel(), JS.ravel())) result = sp.csc_matrix((data, ij), shape=(image_width*image_height*n_channels, num_verts*2)) return result class DifferentiableRenderer(Ch): terms = ['renderer', 'params_list'] dterms = ['params'] def compute_r(self): return self.renderer.r def compute_dr_wrt(self, wrt): import ipdb for param in self.params_list: if wrt is param: return self.gradient_pred(wrt) return self.renderer.dr_wrt(wrt) def gradients(self): # self._call_on_changed() observed = self.renderer.color_image boundaryid_image = self.renderer.boundaryid_image barycentric = self.renderer.barycentric_image visibility = self.renderer.visibility_image visible = np.nonzero(visibility.ravel() != 4294967295)[0] return self.dImage_wrt_2dVerts_bnd_gradient(observed, barycentric, observed.shape[0], observed.shape[1], boundaryid_image != 4294967295) def dImage_wrt_2dVerts_bnd_gradient(self, observed, barycentric, image_width, image_height, bnd_bool): bnd_bool = np.logical_and(self.renderer.visibility_image != 4294967295, bnd_bool) n_channels = np.atleast_3d(observed).shape[2] shape = [image_height, image_width] bndf = bnd_bool.astype(np.float64) bnd_nan = bndf.reshape((observed.shape[0], observed.shape[1], -1)).copy() bnd_nan.ravel()[bnd_nan.ravel()>0] = np.nan bnd_nan += 1 obs_nonbnd = np.atleast_3d(observed) * bnd_nan ydiffnb, xdiffnb = nangradients(obs_nonbnd) observed = np.atleast_3d(observed) if observed.shape[2] > 1: ydiffbnd, xdiffbnd, _ = np.gradient(observed) else: ydiffbnd, xdiffbnd = np.gradient(observed.squeeze()) ydiffbnd = np.atleast_3d(ydiffbnd) xdiffbnd = np.atleast_3d(xdiffbnd) # This corrects for a bias imposed boundary differences begin spread over two pixels # (by np.gradients or similar) but only counted once (since OpenGL's line # drawing spans 1 pixel) xdiffbnd *= 2.0 ydiffbnd *= 2.0 xdiffnb = -xdiffnb ydiffnb = -ydiffnb xdiffbnd = -xdiffbnd ydiffbnd = -ydiffbnd idxs = np.isnan(xdiffnb.ravel()) xdiffnb.ravel()[idxs] = xdiffbnd.ravel()[idxs] idxs = np.isnan(ydiffnb.ravel()) ydiffnb.ravel()[idxs] = ydiffbnd.ravel()[idxs] xdiff = xdiffnb ydiff = ydiffnb dybt = -np.vstack([np.diff(observed, n=1, axis=0), np.zeros([1,observed.shape[1],3])]) dytb = np.vstack([np.zeros([1,observed.shape[1],3]), np.flipud(np.diff(np.flipud(observed), n=1, axis=0))]) dxrl = -np.hstack([np.diff(observed, n=1, axis=1), np.zeros([observed.shape[0],1,3])]) dxlr = np.hstack([np.zeros([observed.shape[0],1,3]),np.fliplr(np.diff(np.fliplr(observed), n=1, axis=1))]) bary_sl = np.roll(barycentric , shift=-1, axis=1) bary_sr = np.roll(barycentric , shift=1, axis=1) bary_st = np.roll(barycentric , shift=-1, axis=0) bary_sb = np.roll(barycentric , shift=1, axis=0) return xdiff, ydiff, dybt, dxrl, dytb, dxlr, bary_sl, bary_sr, bary_st, bary_sb def gradient_pred(self, paramWrt): observed = self.renderer.color_image boundaryid_image = self.renderer.boundaryid_image barycentric = self.renderer.barycentric_image visibility = self.renderer.visibility_image visible = np.nonzero(visibility.ravel() != 4294967295)[0] jacIm = self.dImage_wrt_2dVerts_predict(observed, paramWrt, visible, visibility, barycentric, observed.shape[0], observed.shape[1], self.renderer.v.shape[0], self.renderer.f, boundaryid_image != 4294967295) return jacIm.dot(self.renderer.camera.dr_wrt(paramWrt)) def boundary_neighborhood(self): boundary = self.renderer.boundaryid_image != 4294967295 visibility = self.renderer.visibility_image != 4294967295 boundary = np.logical_and(visibility, boundary) shape = boundary.shape notboundary = np.logical_not(boundary) horizontal = np.hstack((np.diff(boundary.astype(np.int8),axis=1), np.zeros((shape[0],1), dtype=np.int8))) # horizontal = np.hstack((np.diff(boundary.astype(np.int8),axis=1), np.zeros((shape[0],1), dtype=np.int8))) vertical = np.vstack((np.diff(boundary.astype(np.int8), axis=0), np.zeros((1,shape[1]), dtype=np.int8))) # vertical = np.vstack((np.diff(boundary.astype(np.int8), axis=0), np.zeros((1,shape[1]), dtype=np.int8))) pixl = (horizontal == 1) pixr = (horizontal == -1) pixt = (vertical == 1) pixb = (vertical == -1) # plt.imshow((pixrl | pixlr | pixtb | pixbt)) #Quicker, convolve (FFT) and take mask * etc. lidxs_out = np.where(pixl.ravel())[0] ridxs_out = np.where(pixr.ravel())[0] + 1 tidxs_out = np.where(pixt.ravel())[0] bidxs_out = np.where(pixb.ravel())[0] + shape[1] lidxs_int = np.where(pixl.ravel())[0] + 1 ridxs_int = np.where(pixr.ravel())[0] tidxs_int = np.where(pixt.ravel())[0] + shape[1] bidxs_int = np.where(pixb.ravel())[0] return pixr, pixl, pixt, pixb, lidxs_out, ridxs_out, tidxs_out, bidxs_out, lidxs_int, ridxs_int, tidxs_int, bidxs_int def dImage_wrt_2dVerts_predict(self, observed, paramWrt, visible, visibility, barycentric, image_width, image_height, num_verts, f, bnd_bool): """Construct a sparse jacobian that relates 2D projected vertex positions (in the columns) to pixel values (in the rows). This can be done in two steps.""" bnd_bool = np.logical_and(bnd_bool, self.renderer.visibility_image != 4294967295) camJac = self.renderer.camera.dr_wrt(paramWrt) xdiff, ydiff, dybt, dxrl, dytb, dxlr, bary_sl, bary_sr, bary_st, bary_sb = self.gradients() pixr, pixl, pixt, pixb, lidxs_out, ridxs_out, tidxs_out, bidxs_out, lidxs_int, ridxs_int, tidxs_int, bidxs_int = self.boundary_neighborhood() lidxs_out = np.where(bnd_bool.ravel())[0]-1 ridxs_out = np.where(bnd_bool.ravel())[0]+1 tidxs_out = np.where(bnd_bool.ravel())[0]-bnd_bool.shape[1] bidxs_out = np.where(bnd_bool.ravel())[0]+bnd_bool.shape[1] lidxs_int = np.where(bnd_bool.ravel())[0] ridxs_int= np.where(bnd_bool.ravel())[0] tidxs_int= np.where(bnd_bool.ravel())[0] bidxs_int = np.where(bnd_bool.ravel())[0] #Where are triangles moving wrt to the image coordinates at the boundaries? lintGrad = camJac[f[visibility.ravel()[lidxs_int]]*2] rintGrad = camJac[f[visibility.ravel()[ridxs_int]]*2] tintGrad = camJac[f[visibility.ravel()[tidxs_int]]*2+1] bintGrad = camJac[f[visibility.ravel()[bidxs_int]]*2+1] lidx = lintGrad[:,0,0] < -0.0001 xdiff.reshape([-1,3])[lidxs_out[lidx]] = xdiff.reshape([-1,3])[lidxs_int[lidx]] barycentric.reshape([-1,3])[lidxs_out[lidx]] = barycentric.reshape([-1,3])[lidxs_int[lidx]] visibility.ravel()[lidxs_out[lidx]] = visibility.ravel()[lidxs_int[lidx]] xdiff.reshape([-1,3])[lidxs_int[lidx]] = dxrl.reshape([-1,3])[lidxs_int[lidx]] # visible.ravel()[lidxs_out[lidx]] = True ridx = rintGrad[:,0,0] > 0.0001 xdiff.reshape([-1,3])[ridxs_out[ridx]] = xdiff.reshape([-1,3])[ridxs_int[ridx]] barycentric.reshape([-1,3])[ridxs_out[ridx]] = barycentric.reshape([-1,3])[ridxs_int[ridx]] visibility.ravel()[ridxs_out[ridx]] = visibility.ravel()[ridxs_int[ridx]] xdiff.reshape([-1,3])[ridxs_int[ridx]] = dxlr.reshape([-1,3])[ridxs_int[ridx]] # visible.ravel()[ridxs_out[ridx]] = True tidx = tintGrad[:,0,0] > 0.0001 ydiff.reshape([-1,3])[tidxs_out[tidx]] = ydiff.reshape([-1,3])[tidxs_int[tidx]] barycentric.reshape([-1,3])[tidxs_out[tidx]] = barycentric.reshape([-1,3])[tidxs_int[tidx]] visibility.ravel()[tidxs_out[tidx]] = visibility.ravel()[tidxs_int[tidx]] ydiff.reshape([-1,3])[tidxs_int[tidx]] = dybt.reshape([-1,3])[tidxs_int[tidx]] # visible.ravel()[tidxs_out[tidx]] = True bidx = bintGrad[:,0,0] < -0.0001 ydiff.reshape([-1,3])[bidxs_out[bidx]] = ydiff.reshape([-1,3])[bidxs_int[bidx]] barycentric.reshape([-1,3])[bidxs_out[bidx]] = barycentric.reshape([-1,3])[bidxs_int[bidx]] visibility.ravel()[bidxs_out[bidx]] = visibility.ravel()[bidxs_int[bidx]] ydiff.reshape([-1,3])[bidxs_int[bidx]] = dytb.reshape([-1,3])[bidxs_int[bidx]] # visible.ravel()[bidxs_out[bidx]] = True # import ipdb # ipdb.set_trace() visible = np.nonzero(visibility.ravel() != 4294967295)[0] n_channels = np.atleast_3d(observed).shape[2] shape = visibility.shape #2: Take the data and copy the corresponding dxs and dys to these new pixels. # Step 1: get the structure ready, ie the IS and the JS IS = np.tile(col(visible), (1, 2*f.shape[1])).ravel() JS = col(f[visibility.ravel()[visible]].ravel()) JS = np.hstack((JS*2, JS*2+1)).ravel() pxs = np.asarray(visible % shape[1], np.int32) pys = np.asarray(np.floor(np.floor(visible) / shape[1]), np.int32) if n_channels > 1: IS = np.concatenate([IS*n_channels+i for i in range(n_channels)]) JS = np.concatenate([JS for i in range(n_channels)]) datas = [] # The data is weighted according to barycentric coordinates bc0 = col(barycentric[pys, pxs, 0]) bc1 = col(barycentric[pys, pxs, 1]) bc2 = col(barycentric[pys, pxs, 2]) for k in range(n_channels): dxs = xdiff[pys, pxs, k] dys = ydiff[pys, pxs, k] if f.shape[1] == 3: datas.append(np.hstack((col(dxs)*bc0,col(dys)*bc0,col(dxs)*bc1,col(dys)*bc1,col(dxs)*bc2,col(dys)*bc2)).ravel()) else: datas.append(np.hstack((col(dxs)*bc0,col(dys)*bc0,col(dxs)*bc1,col(dys)*bc1)).ravel()) data = np.concatenate(datas) ij = np.vstack((IS.ravel(), JS.ravel())) result = sp.csc_matrix((data, ij), shape=(image_width*image_height*n_channels, num_verts*2)) return result
14,887
40.127072
214
py
inversegraphics
inversegraphics-master/lasagne_nn.py
#!/usr/bin/env python """ Usage example employing Lasagne for digit recognition using the MNIST dataset. This example is deliberately structured as a long flat file, focusing on how to use Lasagne, instead of focusing on writing maximally modular and reusable code. It is used as the foundation for the introductory Lasagne tutorial: http://lasagne.readthedocs.org/en/latest/user/tutorial.html More in-depth examples and reproductions of paper results are maintained in a separate repository: https://github.com/Lasagne/Recipes """ from __future__ import print_function import sys import os import time import numpy as np import theano import theano.tensor as T import pickle # import lasagne.layers.dnn import lasagne from lasagne.layers import InputLayer, DenseLayer, DropoutLayer # from lasagne.layers.dnn import Conv2DDNNLayer as ConvLayer from lasagne.layers import Conv2DLayer as ConvLayer from lasagne.layers import MaxPool2DLayer as PoolLayer from lasagne.layers import LocalResponseNormalization2DLayer as NormLayer from lasagne.utils import floatX from utils import getTriplets import ipdb # ##################### Build the neural network model ####################### # This script supports three types of models. For each one, we define a # function that takes a Theano variable representing the input and returns # the output layer of a neural network model built in Lasagne. def build_mlp(input_var=None): # This creates an MLP of two hidden layers of 800 units each, followed by # a softmax output layer of 10 units. It applies 20% dropout to the input # data and 50% dropout to the hidden layers. # Input layer, specifying the expected input shape of the network # (unspecified batchsize, 1 channel, 28 rows and 28 columns) and # linking it to the given Theano variable `input_var`, if any: l_in = lasagne.layers.InputLayer(shape=(None, 1, 28, 28), input_var=input_var) # Apply 20% dropout to the input data: l_in_drop = lasagne.layers.DropoutLayer(l_in, p=0.2) # Add a fully-connected layer of 800 units, using the linear rectifier, and # initializing weights with Glorot's scheme (which is the default anyway): l_hid1 = lasagne.layers.DenseLayer( l_in_drop, num_units=800, nonlinearity=lasagne.nonlinearities.rectify, W=lasagne.init.GlorotUniform()) # We'll now add dropout of 50%: l_hid1_drop = lasagne.layers.DropoutLayer(l_hid1, p=0.5) # Another 800-unit layer: l_hid2 = lasagne.layers.DenseLayer( l_hid1_drop, num_units=800, nonlinearity=lasagne.nonlinearities.rectify) # 50% dropout again: l_hid2_drop = lasagne.layers.DropoutLayer(l_hid2, p=0.5) # Finally, we'll add the fully-connected output layer, of 10 softmax units: l_out = lasagne.layers.DenseLayer( l_hid2_drop, num_units=10, nonlinearity=lasagne.nonlinearities.softmax) # Each layer is linked to its incoming layer(s), so we only need to pass # the output layer to give access to a network in Lasagne: return l_out def build_custom_mlp(input_var=None, depth=2, width=800, drop_input=.2, drop_hidden=.5): # By default, this creates the same network as `build_mlp`, but it can be # customized with respect to the number and size of hidden layers. This # mostly showcases how creating a network in Python code can be a lot more # flexible than a configuration file. Note that to make the code easier, # all the layers are just called `network` -- there is no need to give them # different names if all we return is the last one we created anyway; we # just used different names above for clarity. # Input layer and dropout (with shortcut `dropout` for `DropoutLayer`): network = lasagne.layers.InputLayer(shape=(None, 1, 28, 28), input_var=input_var) if drop_input: network = lasagne.layers.dropout(network, p=drop_input) # Hidden layers and dropout: nonlin = lasagne.nonlinearities.rectify for _ in range(depth): network = lasagne.layers.DenseLayer( network, width, nonlinearity=nonlin) if drop_hidden: network = lasagne.layers.dropout(network, p=drop_hidden) # Output layer: softmax = lasagne.nonlinearities.softmax network = lasagne.layers.DenseLayer(network, 10, nonlinearity=softmax) return network def build_cnn(input_var=None): # As a third model, we'll create a CNN of two convolution + pooling stages # and a fully-connected hidden layer in front of the output layer. # Input layer, as usual: network = lasagne.layers.InputLayer(shape=(None, 1, 150, 150), input_var=input_var) # This time we do not apply input dropout, as it tends to work less well # for convolutional layers. # Convolutional layer with 32 kernels of size 5x5. Strided and padded # convolutions are supported as well; see the docstring. network = ConvLayer( network, num_filters=32, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify, W=lasagne.init.GlorotUniform()) # Expert note: Lasagne provides alternative convolutional layers that # override Theano's choice of which implementation to use; for details # please see http://lasagne.readthedocs.org/en/latest/user/tutorial.html. # Max-pooling layer of factor 2 in both dimensions: network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=32, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=256, nonlinearity=lasagne.nonlinearities.rectify) # And, finally, the 10-unit output layer with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=9, nonlinearity=lasagne.nonlinearities.linear) return network def build_cnn_small(input_var=None): # As a third model, we'll create a CNN of two convolution + pooling stages # and a fully-connected hidden layer in front of the output layer. # Input layer, as usual: network = lasagne.layers.InputLayer(shape=(None, 1, 150, 150), input_var=input_var) # This time we do not apply input dropout, as it tends to work less well # for convolutional layers. # Convolutional layer with 32 kernels of size 5x5. Strided and padded # convolutions are supported as well; see the docstring. network = ConvLayer( network, num_filters=16, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify, W=lasagne.init.GlorotUniform()) # Expert note: Lasagne provides alternative convolutional layers that # override Theano's choice of which implementation to use; for details # please see http://lasagne.readthedocs.org/en/latest/user/tutorial.html. # Max-pooling layer of factor 2 in both dimensions: network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=16, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=64, nonlinearity=lasagne.nonlinearities.rectify) # And, finally, the 10-unit output layer with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=9, nonlinearity=lasagne.nonlinearities.linear) return network def build_cnn_pose_embedding(input_var=None, sizeIm=150): # As a third model, we'll create a CNN of two convolution + pooling stages # and a fully-connected hidden layer in front of the output layer. # Input layer, as usual: network = lasagne.layers.InputLayer(shape=(None, 1, sizeIm, sizeIm), input_var=input_var) # This time we do not apply input dropout, as it tends to work less well # for convolutional # Convolutional layer with 32 kernels of size 5x5. Strided and padded # convolutions are supported as well; see the docstring. network = ConvLayer( network, num_filters=128, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=128, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) # # network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=128, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) # network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=128, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) # network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # # # A fully-connected layer of 256 units with 50% dropout on its inputs: # network = lasagne.layers.DenseLayer( # lasagne.layers.dropout(network, p=.5), # num_units=128, # nonlinearity=lasagne.nonlinearities.rectify) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( network, num_units=32, nonlinearity=lasagne.nonlinearities.linear) return network def build_cnn2(input_var=None): # As a third model, we'll create a CNN of two convolution + pooling stages # and a fully-connected hidden layer in front of the output layer. # Input layer, as usual: network = lasagne.layers.InputLayer(shape=(None, 1, 150, 150), input_var=input_var) # This time we do not apply input dropout, as it tends to work less well # for convolutional layers. # Convolutional layer with 32 kernels of size 5x5. Strided and padded # convolutions are supported as well; see the docstring. network = ConvLayer( network, num_filters=32, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify, W=lasagne.init.GlorotUniform()) # Expert note: Lasagne provides alternative convolutional layers that # override Theano's choice of which implementation to use; for details # please see http://lasagne.readthedocs.org/en/latest/user/tutorial.html. # Max-pooling layer of factor 2 in both dimensions: network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=32, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=32, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=256, nonlinearity=lasagne.nonlinearities.rectify) # And, finally, the 10-unit output layer with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=9, nonlinearity=lasagne.nonlinearities.linear) return network def build_cnn_pose(input_var=None): # As a third model, we'll create a CNN of two convolution + pooling stages # and a fully-connected hidden layer in front of the output layer. # Input layer, as usual: network = lasagne.layers.InputLayer(shape=(None, 1, 150, 150), input_var=input_var) # This time we do not apply input dropout, as it tends to work less well # for convolutional # Convolutional layer with 32 kernels of size 5x5. Strided and padded # convolutions are supported as well; see the docstring. network = ConvLayer( network, num_filters=64, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=64, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=128, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=256, nonlinearity=lasagne.nonlinearities.rectify) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=32, nonlinearity=lasagne.nonlinearities.rectify) # And, finally, the 10-unit output layer with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( network, num_units=4, nonlinearity=lasagne.nonlinearities.linear) return network def build_cnn_pose_color(input_var=None): # As a third model, we'll create a CNN of two convolution + pooling stages # and a fully-connected hidden layer in front of the output layer. # Input layer, as usual: network = lasagne.layers.InputLayer(shape=(None, 3, 150, 150), input_var=input_var) # This time we do not apply input dropout, as it tends to work less well # for convolutional # Convolutional layer with 32 kernels of size 5x5. Strided and padded # convolutions are supported as well; see the docstring. network = ConvLayer( network, num_filters=64, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=64, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=128, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=256, nonlinearity=lasagne.nonlinearities.rectify) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=32, nonlinearity=lasagne.nonlinearities.rectify) # And, finally, the 10-unit output layer with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( network, num_units=4, nonlinearity=lasagne.nonlinearities.linear) return network def build_cnn_appLight(input_var=None): # As a third model, we'll create a CNN of two convolution + pooling stages # and a fully-connected hidden layer in front of the output layer. # Input layer, as usual: input = lasagne.layers.InputLayer(shape=(None, 3, 150, 150), input_var=input_var) # This time we do not apply input dropout, as it tends to work less well # for convolutional # Convolutional layer with 32 kernels of size 5x5. Strided and padded # convolutions are supported as well; see the docstring. network = ConvLayer( input, num_filters=64, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=64, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=128, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=256, nonlinearity=lasagne.nonlinearities.rectify) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=32, nonlinearity=lasagne.nonlinearities.rectify) # And, finally, the 10-unit output layer with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( network, num_units=12, nonlinearity=lasagne.nonlinearities.linear) return network def build_cnn_shape(input_var=None): # As a third model, we'll create a CNN of two convolution + pooling stages # and a fully-connected hidden layer in front of the output layer. # Input layer, as usual: network = lasagne.layers.InputLayer(shape=(None, 1, 150, 150), input_var=input_var) # This time we do not apply input dropout, as it tends to work less well # for convolutional # Convolutional layer with 32 kernels of size 5x5. Strided and padded # convolutions are supported as well; see the docstring. network = ConvLayer( network, num_filters=64, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=64, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=128, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=256, nonlinearity=lasagne.nonlinearities.rectify) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=32, nonlinearity=lasagne.nonlinearities.rectify) # And, finally, the 10-unit output layer with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( network, num_units=10, nonlinearity=lasagne.nonlinearities.linear) return network def build_cnn_shape_k(input_var=None): # As a third model, we'll create a CNN of two convolution + pooling stages # and a fully-connected hidden layer in front of the output layer. # Input layer, as usual: input = lasagne.layers.InputLayer(shape=(None, 3, 150, 150), input_var=input_var) # This time we do not apply input dropout, as it tends to work less well # for convolutional # Convolutional layer with 32 kernels of size 5x5. Strided and padded # convolutions are supported as well; see the docstring. network = ConvLayer( input, num_filters=32, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=64, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = ConvLayer( network, num_filters=64, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=128, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = ConvLayer( network, num_filters=128, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.0), num_units=128, nonlinearity=lasagne.nonlinearities.rectify) ## A fully-connected layer of 256 units with 50% dropout on its inputs: #network = lasagne.layers.DenseLayer( # lasagne.layers.dropout(network, p=.0), # num_units=128, # nonlinearity=lasagne.nonlinearities.rectify) # And, finally, the 10-unit output layer with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( network, num_units=10, nonlinearity=lasagne.nonlinearities.linear) return network def build_cnn_shape_medium(input_var=None): # As a third model, we'll create a CNN of two convolution + pooling stages # and a fully-connected hidden layer in front of the output layer. # Input layer, as usual: input = lasagne.layers.InputLayer(shape=(None, 3, 150, 150), input_var=input_var) # This time we do not apply input dropout, as it tends to work less well # for convolutional # Convolutional layer with 32 kernels of size 5x5. Strided and padded # convolutions are supported as well; see the docstring. network = ConvLayer( input, num_filters=128, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=128, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=128, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=256, nonlinearity=lasagne.nonlinearities.rectify) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=128, nonlinearity=lasagne.nonlinearities.rectify) # And, finally, the 10-unit output layer with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( network, num_units=10, nonlinearity=lasagne.nonlinearities.linear) return network def build_cnn_shape_large(input_var=None): # As a third model, we'll create a CNN of two convolution + pooling stages # and a fully-connected hidden layer in front of the output layer. # Input layer, as usual: network = lasagne.layers.InputLayer(shape=(None, 3, 150, 150), input_var=input_var) # This time we do not apply input dropout, as it tends to work less well # for convolutional layers. # Convolutional layer with 32 kernels of size 5x5. Strided and padded # convolutions are supported as well; see the docstring. network = ConvLayer( network, num_filters=256, filter_size=(7, 7), nonlinearity=lasagne.nonlinearities.rectify, W=lasagne.init.GlorotUniform()) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(3, 3)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=256, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=256, filter_size=(3, 3), nonlinearity=lasagne.nonlinearities.rectify) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=256, filter_size=(3, 3), nonlinearity=lasagne.nonlinearities.rectify) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=256, nonlinearity=lasagne.nonlinearities.rectify) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=256, nonlinearity=lasagne.nonlinearities.rectify) # And, finally, the 10-unit output layer with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( network, num_units=10, nonlinearity=lasagne.nonlinearities.linear) return network def build_cnn_light(input_var=None): # As a third model, we'll create a CNN of two convolution + pooling stages # and a fully-connected hidden layer in front of the output layer. # Input layer, as usual: input = lasagne.layers.InputLayer(shape=(None, 3, 150, 150), input_var=input_var) # This time we do not apply input dropout, as it tends to work less well # for convolutional # Convolutional layer with 32 kernels of size 5x5. Strided and padded # convolutions are supported as well; see the docstring. network = ConvLayer( input, num_filters=64, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=64, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=128, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=256, nonlinearity=lasagne.nonlinearities.rectify) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=32, nonlinearity=lasagne.nonlinearities.rectify) # And, finally, the 10-unit output layer with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( network, num_units=9, nonlinearity=lasagne.nonlinearities.linear) return network def build_cnn_app(input_var=None): # As a third model, we'll create a CNN of two convolution + pooling stages # and a fully-connected hidden layer in front of the output layer. # Input layer, as usual: input = lasagne.layers.InputLayer(shape=(None, 3, 150, 150), input_var=input_var) # This time we do not apply input dropout, as it tends to work less well # for convolutional # Convolutional layer with 32 kernels of size 5x5. Strided and padded # convolutions are supported as well; see the docstring. network = ConvLayer( input, num_filters=64, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=64, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=128, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=256, nonlinearity=lasagne.nonlinearities.rectify) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=32, nonlinearity=lasagne.nonlinearities.rectify) # And, finally, the 10-unit output layer with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( network, num_units=3, nonlinearity=lasagne.nonlinearities.linear) return network class MeanLayer(lasagne.layers.Layer): def get_output_for(self, input, **kwargs): return input.mean(axis=1).mean(axis=1) def get_output_shape_for(self, input_shape): return [input_shape[0],1] def build_cnn_mask(input_var=None): # As a third model, we'll create a CNN of two convolution + pooling stages # and a fully-connected hidden layer in front of the output layer. # Input layer, as usual: input = lasagne.layers.InputLayer(shape=(None, 3, 150, 150), input_var=input_var) # This time we do not apply input dropout, as it tends to work less well # for convolutional # Convolutional layer with 32 kernels of size 5x5. Strided and padded # convolutions are supported as well; see the docstring. network = ConvLayer( input, num_filters=64, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=64, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=128, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=256, nonlinearity=lasagne.nonlinearities.rectify) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=225, nonlinearity=lasagne.nonlinearities.rectify) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=2250, nonlinearity=lasagne.nonlinearities.linear) numSmallMask = 2250 scaleFactor = 10 network = lasagne.layers.ReshapeLayer(network, ([0],1, numSmallMask)) # A fully-connected layer of 256 units with 50% dropout on its inputs: mask = lasagne.layers.Upscale1DLayer( incoming=network, scale_factor=scaleFactor) mask = lasagne.layers.ReshapeLayer(mask, ([0], numSmallMask*scaleFactor)) mask = lasagne.layers.NonlinearityLayer( mask, nonlinearity=lasagne.nonlinearities.sigmoid) return mask def build_cnn_mask_large(input_var=None): # As a third model, we'll create a CNN of two convolution + pooling stages # and a fully-connected hidden layer in front of the output layer. # Input layer, as usual: input = lasagne.layers.InputLayer(shape=(None, 3, 150, 150), input_var=input_var) # This time we do not apply input dropout, as it tends to work less well # for convolutional # Convolutional layer with 32 kernels of size 5x5. Strided and padded # convolutions are supported as well; see the docstring. network = ConvLayer( input, num_filters=256, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) lasagne.layers.get_output_shape(network) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=256, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=256, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) network = ConvLayer( network, num_filters=256, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) # network = lasagne.layers.MaxPool2DLayer(network, pool_size=(1, 1)) network = ConvLayer( network, num_filters=256, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) # network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=1125, nonlinearity=lasagne.nonlinearities.rectify) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=1125, nonlinearity=lasagne.nonlinearities.rectify) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network), num_units=2500, nonlinearity=lasagne.nonlinearities.linear) # numSmallMask = 2250 # scaleFactor = 10 # network = lasagne.layers.ReshapeLayer(network, ([0],1, numSmallMask)) # # # A fully-connected layer of 256 units with 50% dropout on its inputs: # mask = lasagne.layers.Upscale1DLayer( # incoming=network, # scale_factor=scaleFactor) # # mask = lasagne.layers.ReshapeLayer(mask, ([0], numSmallMask*scaleFactor)) mask = lasagne.layers.NonlinearityLayer( network, nonlinearity=lasagne.nonlinearities.sigmoid) return mask def build_cnn_appmask(input_var=None): # As a third model, we'll create a CNN of two convolution + pooling stages # and a fully-connected hidden layer in front of the output layer. # Input layer, as usual: input = lasagne.layers.InputLayer(shape=(None, 3, 150, 150), input_var=input_var) # This time we do not apply input dropout, as it tends to work less well # for convolutional # Convolutional layer with 32 kernels of size 5x5. Strided and padded # convolutions are supported as well; see the docstring. network = ConvLayer( input, num_filters=64, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=64, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=128, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=256, nonlinearity=lasagne.nonlinearities.rectify) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=2250, nonlinearity=lasagne.nonlinearities.rectify) # A fully-connected layer of 256 units with 50% dropout on its inputs: mask = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=22500, nonlinearity=lasagne.nonlinearities.rectify) inputR = lasagne.layers.SliceLayer(input, 0, axis=1) inputG = lasagne.layers.SliceLayer(input, 1, axis=1) inputB = lasagne.layers.SliceLayer(input, 2, axis=1) reshapedMask = lasagne.layers.ReshapeLayer(mask, ([0],150, 150)) outR = MeanLayer(lasagne.layers.ElemwiseMergeLayer(incomings=[inputR, reshapedMask], merge_function=theano.tensor.mul)) outG = MeanLayer(lasagne.layers.ElemwiseMergeLayer(incomings=[inputG, reshapedMask], merge_function=theano.tensor.mul)) outB = MeanLayer(lasagne.layers.ElemwiseMergeLayer(incomings=[inputB, reshapedMask], merge_function=theano.tensor.mul)) # And, finally, the 10-unit output layer with 50% dropout on its inputs: network = lasagne.layers.ConcatLayer(incomings=[outR, outG, outB], axis=0) network = lasagne.layers.ReshapeLayer(network, ([0],3)) return network def build_cnn_pose_norm(input_var=None): # As a third model, we'll create a CNN of two convolution + pooling stages # and a fully-connected hidden layer in front of the output layer. # Input layer, as usual: network = lasagne.layers.InputLayer(shape=(None, 1, 150, 150), input_var=input_var) # This time we do not apply input dropout, as it tends to work less well # for convolutional # Convolutional layer with 32 kernels of size 5x5. Strided and padded # convolutions are supported as well; see the docstring. network = lasagne.layers.normalization.batch_norm(ConvLayer( network, num_filters=64, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify)) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = lasagne.layers.normalization.batch_norm(ConvLayer( network, num_filters=64, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify)) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = lasagne.layers.normalization.batch_norm(ConvLayer( network, num_filters=128, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify)) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.normalization.batch_norm(lasagne.layers.DenseLayer( network, num_units=256, nonlinearity=lasagne.nonlinearities.rectify)) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.normalization.batch_norm(lasagne.layers.DenseLayer( network, num_units=32, nonlinearity=lasagne.nonlinearities.rectify)) # And, finally, the 10-unit output layer with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( network, num_units=4, nonlinearity=lasagne.nonlinearities.linear) return network def build_cnn_pose_large(input_var=None): # As a third model, we'll create a CNN of two convolution + pooling stages # and a fully-connected hidden layer in front of the output layer. # Input layer, as usual: network = lasagne.layers.InputLayer(shape=(None, 1, 150, 150), input_var=input_var) # This time we do not apply input dropout, as it tends to work less well # for convolutional layers. # Convolutional layer with 32 kernels of size 5x5. Strided and padded # convolutions are supported as well; see the docstring. network = ConvLayer( network, num_filters=96, filter_size=(7, 7), nonlinearity=lasagne.nonlinearities.rectify, W=lasagne.init.GlorotUniform()) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(3, 3)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=256, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=512, filter_size=(3, 3), nonlinearity=lasagne.nonlinearities.rectify) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=512, filter_size=(3, 3), nonlinearity=lasagne.nonlinearities.rectify) # Another convolution with 32 5x5 kernels, and another 2x2 pooling: network = ConvLayer( network, num_filters=512, filter_size=(3, 3), nonlinearity=lasagne.nonlinearities.rectify) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(3, 3)) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=4096, nonlinearity=lasagne.nonlinearities.rectify) # A fully-connected layer of 256 units with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( lasagne.layers.dropout(network, p=.5), num_units=1024, nonlinearity=lasagne.nonlinearities.rectify) # And, finally, the 10-unit output layer with 50% dropout on its inputs: network = lasagne.layers.DenseLayer( network, num_units=4, nonlinearity=lasagne.nonlinearities.linear) return network def build_cnn_pose_large_norm(input_var=None): network = lasagne.layers.InputLayer(shape=(None, 1, 150, 150), input_var=input_var) network = lasagne.layers.normalization.batch_norm(ConvLayer( network, num_filters=96, filter_size=(7, 7), nonlinearity=lasagne.nonlinearities.rectify, W=lasagne.init.GlorotUniform())) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(3, 3)) network = lasagne.layers.normalization.batch_norm(ConvLayer( network, num_filters=256, filter_size=(5, 5), nonlinearity=lasagne.nonlinearities.rectify)) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2)) network = lasagne.layers.normalization.batch_norm(ConvLayer( network, num_filters=512, filter_size=(3, 3), nonlinearity=lasagne.nonlinearities.rectify)) network = lasagne.layers.normalization.batch_norm(ConvLayer( network, num_filters=512, filter_size=(3, 3), nonlinearity=lasagne.nonlinearities.rectify)) network = lasagne.layers.normalization.batch_norm(ConvLayer( network, num_filters=512, filter_size=(3, 3), nonlinearity=lasagne.nonlinearities.rectify)) network = lasagne.layers.MaxPool2DLayer(network, pool_size=(3, 3)) network = lasagne.layers.normalization.batch_norm(lasagne.layers.DenseLayer( network, num_units=4096, nonlinearity=lasagne.nonlinearities.rectify)) network = lasagne.layers.normalization.batch_norm(lasagne.layers.DenseLayer( network, num_units=1024, nonlinearity=lasagne.nonlinearities.rectify)) network = lasagne.layers.DenseLayer( network, num_units=4, nonlinearity=lasagne.nonlinearities.linear) return network def iterate_minibatches(inputs, targets, batchsize, shuffle=False): assert len(inputs) == len(targets) if shuffle: indices = np.arange(len(inputs)) np.random.shuffle(indices) for start_idx in range(0, len(inputs) - batchsize + 1, batchsize): if shuffle: excerpt = indices[start_idx:start_idx + batchsize] else: excerpt = slice(start_idx, start_idx + batchsize) yield inputs[excerpt], targets[excerpt] def iterate_minibatches_triplets(inputs, inputs_t1, inputs_t2, batchsize, shuffle=False): assert len(inputs) == len(inputs_t1) == len(inputs_t2) if shuffle: indices = np.arange(len(inputs)) np.random.shuffle(indices) for start_idx in range(0, len(inputs) - batchsize + 1, batchsize): if shuffle: excerpt = indices[start_idx:start_idx + batchsize] else: excerpt = slice(start_idx, start_idx + batchsize) yield inputs[excerpt], inputs_t1[excerpt], inputs_t2[excerpt] def iterate_minibatches_h5(inputs_h5, trainSet, trainValSet, targets, batchsize, shuffle=False): assert len(trainValSet) == len(targets) print("Loading minibatch set") if shuffle: indices = np.arange(len(trainValSet)) np.random.shuffle(indices) for start_idx in range(0, len(trainValSet), batchsize): if shuffle: excerpt = indices[start_idx:start_idx + batchsize] else: excerpt = slice(start_idx, start_idx + batchsize) boolSet = np.zeros(len(inputs_h5)).astype(np.bool) boolSet[trainSet[trainValSet[excerpt]]] = True yield inputs_h5[boolSet,:,:], targets[excerpt] print("Ended loading minibatch set") def load_network(modelType='cnn', param_values=[], imgSize=150): # Load the dataset # Prepare Theano variables for inputs and targets input_var = T.tensor4('inputs') # Create neural network model (depending on first command line parameter) print("Building model and compiling functions...") if modelType == 'mlp': network = build_mlp(input_var) elif modelType.startswith('custom_mlp:'): depth, width, drop_in, drop_hid = modelType.split(':', 1)[1].split(',') network = build_custom_mlp(input_var, int(depth), int(width), float(drop_in), float(drop_hid)) elif modelType == 'cnn': network = build_cnn(input_var) elif modelType == 'cnn2': network = build_cnn2(input_var) elif modelType == 'cnn_pose': network = build_cnn_pose(input_var) elif modelType == 'cnn_pose_embedding': network = build_cnn_pose_embedding(input_var, imgSize) elif modelType == 'cnn_pose_large': network = build_cnn_pose_large(input_var) elif modelType == 'cnn_pose_color': network = build_cnn_pose_color(input_var) elif modelType == 'cnn_pose_norm': network = build_cnn_pose_norm(input_var) elif modelType == 'cnn_pose_large_norm': network = build_cnn_pose_large_norm(input_var) elif modelType == 'cnn_light': network = build_cnn_light(input_var) elif modelType == 'cnn_shape': network = build_cnn_shape(input_var) elif modelType == 'cnn_shape_large': network = build_cnn_shape_large(input_var) elif modelType == 'cnn_shape_medium': network = build_cnn_shape_medium(input_var) elif modelType == 'cnn_appLight': network = build_cnn_appLight(input_var) elif modelType == 'cnn_app': network = build_cnn_app(input_var) elif modelType == 'cnn_appmask': network = build_cnn_appmask(input_var) elif modelType == 'cnn_mask': network = build_cnn_mask(input_var) elif modelType == 'cnn_mask_large': network = build_cnn_mask_large(input_var) else: print("Unrecognized model type %r." % modelType) if param_values: lasagne.layers.set_all_param_values(network, param_values) return network def get_prediction_fun(network): # Load the dataset # Prepare Theano variables for inputs and targets input_var = lasagne.layers.get_all_layers(network)[0].input_var prediction = lasagne.layers.get_output(network, deterministic=True) prediction_fn = theano.function([input_var], prediction) return prediction_fn def get_prediction_fun_nondeterministic(network): # Load the dataset # Prepare Theano variables for inputs and targets input_var = lasagne.layers.get_all_layers(network)[0].input_var prediction = lasagne.layers.get_output(network, deterministic=False) prediction_fn = theano.function([input_var], prediction) return prediction_fn def train_nn_h5(X_h5, trainSetVal, y_train, y_val, meanImage, network, modelType = 'cnn', num_epochs=150, saveModelAtEpoch=True, modelPath='tmp/nnmodel.pickle', param_values=[]): # Load the dataset print("Loading validation set") if meanImage.ndim == 2: meanImage = meanImage[:,:,None] X_val = X_h5[trainSetVal::].astype(np.float32) - meanImage.reshape([1,meanImage.shape[2], meanImage.shape[0],meanImage.shape[1]]).astype(np.float32) print("Ended loading validation set") model = {} model['mean'] = meanImage model['type'] = modelType if param_values: lasagne.layers.set_all_param_values(network, param_values) input_var = lasagne.layers.get_all_layers(network)[0].input_var target_var = T.fmatrix('targets') prediction = lasagne.layers.get_output(network) params = lasagne.layers.get_all_params(network, trainable=True) test_prediction = lasagne.layers.get_output(network, deterministic=True) if modelType == 'cnn_mask': loss = lasagne.objectives.binary_crossentropy(prediction, target_var) loss = loss.mean() test_loss = lasagne.objectives.binary_crossentropy(test_prediction, target_var) test_loss = test_loss.mean() else: loss = lasagne.objectives.squared_error(prediction, target_var) loss = loss.mean() test_loss = lasagne.objectives.squared_error(test_prediction, target_var) test_loss = test_loss.mean() updates = lasagne.updates.nesterov_momentum( loss, params, learning_rate=0.01, momentum=0.9) train_fn = theano.function([input_var, target_var], loss, updates=updates) # Compile a second function computing the validation loss and accura# cy: val_fn = theano.function([input_var, target_var], test_loss) # Finally, launch the training loop. print("Starting training...") # We iterate over epochs: patience = 20 best_valid = np.inf best_valid_epoch = 0 best_weights = None batchSize = 128 for epoch in range(num_epochs): # In each epoch, we do a full pass over the training data: train_err = 0 train_batches = 0 start_time = time.time() slicesize = 20000 sliceidx = 0 for start_idx in range(0, trainSetVal, slicesize): sliceidx += 1 print("Working on slice " + str(sliceidx) + " of " + str(int(trainSetVal/slicesize))) X_train = X_h5[start_idx:min(start_idx + slicesize,trainSetVal)].astype(np.float32) - meanImage.reshape([1,meanImage.shape[2], meanImage.shape[0],meanImage.shape[1]]).astype(np.float32) for batch in iterate_minibatches(X_train, y_train[start_idx:min(start_idx + slicesize,trainSetVal)], batchSize, shuffle=True): # print("Batch " + str(train_batches)) inputs, targets = batch train_err += train_fn(inputs, targets) train_batches += 1 # And a full pass over the validation data: val_err = 0 val_batches = 0 for batch in iterate_minibatches(X_val, y_val, batchSize, shuffle=False): inputs, targets = batch err = val_fn(inputs, targets) val_err += err val_batches += 1 if val_err < best_valid: best_weights = lasagne.layers.get_all_param_values(network) if saveModelAtEpoch: model['params'] = best_weights with open(modelPath, 'wb') as pfile: pickle.dump(model, pfile) best_valid = val_err best_valid_epoch = epoch elif best_valid_epoch + patience < epoch: print("Early stopping.") # print("Best valid loss was {:.6f} at epoch {}.".format( # best_valid, best_valid_epoch)) break # Then we print the results for this epoch: print("Epoch {} of {} took {:.3f}s".format( epoch + 1, num_epochs, time.time() - start_time)) print(" training loss:\t\t{:.6f}".format(train_err / train_batches)) print(" validation loss:\t\t{:.6f}".format(val_err / val_batches)) model['params'] = best_weights return model from lasagne.regularization import regularize_layer_params_weighted, l2, l1 def train_triplets_h5(X_h5, Xt1_h5, Xt2_h5, trainSetVal, meanImage, network, modelType = 'cnn', num_epochs=150, saveModelAtEpoch=True, modelPath='tmp/nnmodel.pickle', param_values=[]): # Load the dataset print("Loading validation set") if meanImage.ndim == 2: meanImage = meanImage[:,:,None] X_val = X_h5[trainSetVal::].astype(np.float32) X_t1_val = Xt1_h5[trainSetVal::].astype(np.float32) X_t2_val = Xt2_h5[trainSetVal::].astype(np.float32) # height = X_h5[0].shape[2] # width = X_h5[0].shape[3] print("Ended loading validation set") model = {} model['mean'] = meanImage model['type'] = modelType if param_values: lasagne.layers.set_all_param_values(network, param_values) input_var = lasagne.layers.get_all_layers(network)[0].input_var # target_var = T.fmatrix('targets') predictions = lasagne.layers.get_output(network) params = lasagne.layers.get_all_params(network, trainable=True) batchSize = 36 batchStepSize = int(batchSize / 3) test_predictions = lasagne.layers.get_output(network, deterministic=True) test_predictions_t0 = test_predictions[0:batchStepSize] test_predictions_t1 = test_predictions[batchStepSize:2*batchStepSize] test_predictions_t2 = test_predictions[2*batchStepSize::] # loss = lasagne.objectives.binary_crossentropy(prediction, prediction_t1, prediction_t2) m = 0.001 predictions_t0 = predictions[0:batchStepSize] predictions_t1 = predictions[batchStepSize:2*batchStepSize] predictions_t2 = predictions[2*batchStepSize::] loss = T.sum((predictions_t1 - predictions_t0)**2, axis=1) / (T.sum((predictions_t2 - predictions_t0)**2, axis=1) + m) loss = loss.mean() # ipdb.set_trace() layers = lasagne.layers.get_all_layers(network) regLayers = {layer:0.01 for layer in layers} l2_penalty = regularize_layer_params_weighted(regLayers,l2) # loss = loss test_loss = T.sum((test_predictions_t1 - test_predictions_t0)**2, axis=1) / (T.sum((test_predictions_t2 - test_predictions_t0)**2, axis=1) + m) test_loss = test_loss.mean() # test_loss = test_loss updates = lasagne.updates.nesterov_momentum( loss, params, learning_rate=0.01, momentum=0.9) train_fn = theano.function([input_var], loss, updates=updates) pred_t0_fn = theano.function([input_var], test_predictions_t0) pred_t1_fn = theano.function([input_var], test_predictions_t1) pred_t2_fn = theano.function([input_var], test_predictions_t2) # Compile a second function computing the validation loss and accura# cy: val_fn = theano.function([input_var], test_loss) # Finally, launch the training loop. print("Starting training...") # We iterate over epochs: patience = 20 best_valid = np.inf best_valid_epoch = 0 best_weights = None for epoch in range(num_epochs): # In each epoch, we do a full pass over the training data: train_err = 0 train_batches = 0 start_time = time.time() slicesize = 10000 sliceidx = 0 for start_idx in range(0, trainSetVal, slicesize): sliceidx += 1 print("Working on slice " + str(sliceidx) + " of " + str(int(trainSetVal/slicesize))) X_train = X_h5[start_idx:min(start_idx + slicesize,trainSetVal)].astype(np.float32) X_t1_train = Xt1_h5[start_idx:min(start_idx + slicesize,trainSetVal)].astype(np.float32) X_t2_train = Xt2_h5[start_idx:min(start_idx + slicesize, trainSetVal)].astype(np.float32) # parameterVals = y_train[start_idx:min(start_idx + slicesize,trainSetVal)] # anchorIdx, closeIdx, farIdx = getTriplets(parameterVals, closeDist=5 * np.pi / 180, farDist=15 * np.pi / 180, normConst=2 * np.pi, chunkSize=10) for batch in iterate_minibatches_triplets(X_train, X_t1_train, X_t2_train, batchStepSize, shuffle=True): # print("Batch " + str(train_batches)) inputs, inputs_t1, inputs_t2 = batch # # p0 = pred_t0_fn(np.concatenate([inputs, inputs_t1, inputs_t2], axis=0)) # p1 = pred_t1_fn(np.concatenate([inputs, inputs_t1, inputs_t2], axis=0)) # p2 = pred_t2_fn(np.concatenate([inputs, inputs_t1, inputs_t2], axis=0)) # nperr = np.sum((p1 - p0) ** 2, axis=1) / (np.sum((p2 - p0)**2, axis=1) + m) train_err += train_fn(np.concatenate([inputs, inputs_t1, inputs_t2],axis=0)) # # # p0 = pred_t0_fn(np.concatenate([inputs, inputs_t1, inputs_t2], axis=0)) # p1 = pred_t1_fn(np.concatenate([inputs, inputs_t1, inputs_t2], axis=0)) # p2 = pred_t2_fn(np.concatenate([inputs, inputs_t1, inputs_t2], axis=0)) # # nperr = np.sum((p1 - p0) ** 2, axis=1) / (np.sum((p2 - p0)**2, axis=1) + m) train_batches += 1 # And a full pass over the validation data: val_err = 0 val_batches = 0 for batch in iterate_minibatches_triplets(X_val, X_t1_val, X_t2_val, batchStepSize, shuffle=True): inputs, inputs_t1, inputs_t2 = batch err = val_fn(np.concatenate([inputs, inputs_t1, inputs_t2], axis=0)) val_err += err val_batches += 1 if val_err < best_valid: best_weights = lasagne.layers.get_all_param_values(network) if saveModelAtEpoch: model['params'] = best_weights with open(modelPath, 'wb') as pfile: pickle.dump(model, pfile) best_valid = val_err best_valid_epoch = epoch elif best_valid_epoch + patience < epoch: print("Early stopping.") # print("Best valid loss was {:.6f} at epoch {}.".format( # best_valid, best_valid_epoch)) break # Then we print the results for this epoch: print("Epoch {} of {} took {:.3f}s".format( epoch + 1, num_epochs, time.time() - start_time)) print(" training loss:\t\t{:.6f}".format(train_err / train_batches)) print(" validation loss:\t\t{:.6f}".format(val_err / val_batches)) model['params'] = best_weights return model
63,328
40.122727
197
py
inversegraphics
inversegraphics-master/densecrf_model.py
""" Usage: python util_inference_example.py image annotations Adapted from the dense_inference.py to demonstate the usage of the util functions. """ import sys import numpy as np import cv2 import pydensecrf.densecrf as dcrf import matplotlib.pylab as plt from skimage.segmentation import relabel_sequential import skimage from skimage.segmentation import slic import skimage.segmentation from skimage.segmentation import mark_boundaries from skimage.util import img_as_float import ipdb from pydensecrf.utils import compute_unary, create_pairwise_bilateral, \ create_pairwise_gaussian def unaryOcclusionModel(img, fgMask, probs): """ Simple classifier that is 50% certain that the annotation is correct (same as in the inference example). """ fg_energy = -np.log(probs[0]) occ_energy = -np.log(probs[1]) probBackground = 0.6 U = np.zeros((3, fgMask.size), dtype='float32') U[0, :] = -np.log(0.0001) U[1, :] = -np.log((1-probBackground)) U[2, :] = -np.log(probBackground) U[0, fgMask] = fg_energy U[1, fgMask] = occ_energy U[2, fgMask] = -np.log(0.0001) return U def superpixelUnary(img, U, fgMask, iouProb): segments = skimage.segmentation.quickshift(img, ratio=1, max_dist=10, convert2lab=True) # plt.imshow(segments) fgMaskReshaped = fgMask.reshape([img.shape[0], img.shape[1]]) for seg_i in np.arange(np.max(segments)): currentSegment = segments == seg_i masksCat = np.concatenate([fgMaskReshaped[:,:,None], currentSegment[:,:,None]], axis=2) segmentationIOU = np.sum(np.all(masksCat, axis=2)) / np.sum(currentSegment) if segmentationIOU > 0.05 and segmentationIOU < 0.95: U[0, currentSegment.ravel()*fgMask] = -np.log(1 - iouProb) U[1, currentSegment.ravel()*fgMask] = -np.log(iouProb) U[2, currentSegment.ravel()*(~fgMask)] = -np.log(1 - iouProb) U[1, currentSegment.ravel()*(~fgMask)] = -np.log(iouProb) # # Edge pixels inside the mask (and far from boundary) are occluder pixels. How to choose the right one? (interior pixels) # if segmentationIOU >= 0.95 : # # U[0, currentSegment.ravel()*fgMask] = -np.log(iouProb) # U[1, currentSegment.ravel()*fgMask] = -np.log(1-iouProb) # # Segments with edges that match with boundary of the mask are fg. # if segmentationIOU >= 0.99: # U[0, currentSegment*fgMask] = 1 - iouProb # U[1, currentSegment*fgMask] = iouProb # U[2, currentSegment*(~fgMask)] = 1 - iouProb # U[1, currentSegment*(~fgMask)] = iouProb return U def boundaryUnary(img, U, fgMask, boundProb): from damascene import damascene #img must be uint8 0-255 borders, textons, orientations = damascene(np.uint8(img[:,:,0:3]*255), device_num=0) return U def crfInference(imageGT, fgMask, probs): ################################## ### Read images and annotation ### ################################## # img = np.uint8(img*255) img = skimage.color.rgb2lab(imageGT) M = 3 # forground, background, occluding object. ########################### ### Setup the CRF model ### ########################### # Example using the DenseCRF class and the util functions crfmodel = dcrf.DenseCRF(img.shape[0] * img.shape[1], M) # get unary potentials (neg log probability) # U = compute_unary(labels, M) U = unaryOcclusionModel(img, fgMask, probs) U = superpixelUnary(imageGT, U, fgMask, 0.8) crfmodel.setUnaryEnergy(U) # This creates the color-independent features and then add them to the CRF feats = create_pairwise_gaussian(sdims=(3, 3), shape=img.shape[:2]) crfmodel.addPairwiseEnergy(feats, compat=3, kernel=dcrf.DIAG_KERNEL, normalization=dcrf.NORMALIZE_SYMMETRIC) # This creates the color-dependent features and then add them to the CRF feats = create_pairwise_bilateral(sdims=(30, 30), schan=(13, 13, 13), img=img, chdim=2) crfmodel.addPairwiseEnergy(feats, compat=10, kernel=dcrf.DIAG_KERNEL, normalization=dcrf.NORMALIZE_SYMMETRIC) #################################### ### Do inference and compute map ### #################################### Q = crfmodel.inference(5) mapseg = np.argmax(Q, axis=0).reshape(img.shape[:2]) # res = map.astype('float32') * 255 / map.max() # plt.imshow(res) # plt.show() # # Manually inference # Q, tmp1, tmp2 = crfmodel.startInference() # for i in range(5): # print("KL-divergence at {}: {}".format(i, crfmodel.klDivergence(Q))) # crfmodel.stepInference(Q, tmp1, tmp2) return mapseg, np.array(Q)
4,856
31.165563
131
py
inversegraphics
inversegraphics-master/probLineSearch.py
from scipy.special import erf import numpy as np from scipy.stats import mvn import ipdb def probLineSearch(func, x0, f0, df0, search_direction, alpha0, verbosity, outs, paras, var_f0, var_df0): # probLineSearch.m -- A probabilistic line search algorithm for nonlinear # optimization problems with noisy gradients. # # == INPUTS =============================================================== # [f,f', var_f, var_df] = func(x) -- function handle # input: # x -- column vectors (positions) (Dx1) # output: # f -- scalar function values # df -- vector gradients (Dx1) # var_f -- estimated noise for function values (scalar) # var_df -- estimated noise for gradients (Dx1) # x0 -- current position of optimizer (Dx1) # f0 -- function value at x0 (scalar, previous output y_tt) # df0 -- gradient at x0 ((Dx1), previous output dy_tt) # search_direction -- (- df(x0) does not need to be normalized) # alpha0: initial step size (scalar, previous output alpha0_out) # var_f0 -- variance of function values at x0. (scalar, previous output var_f_tt) # var_df0 -- variance of gradients at x0. ((Dx1), previous output var_df_tt) # verbosity -- level of stdout output. # 0 -- no output # 1 -- steps, function values, state printed to stdout # 2 -- plots with only function values # 3 -- plots including Gaussian process and Wolfe condition beliefs. # paras -- possible parameters that func needs. # outs -- struct with collected statistics # # == OUTPUTS ============================================================== # outs -- struct including counters and statistics # alpha0_out -- accepted stepsize * 1.3 (initial step size for next step) # x_tt -- accepted position # y_tt -- functin value at x_tt # dy_tt -- gradient at x_tt # var_f_tt -- variance of function values at x_tt # var_df_tt -- variance of gradients values at x_tt # # outs = {} # -- setup fixed parameters ----------------------------------------------- # if ~isfield(outs, 'counter') # outs['counter'] = 1 # end outs['alpha_stats'] = alpha0 # running average over accepted step sizes limit = 7 # maximum #function evaluations in one line search (+1) # constants for Wolfe conditions (must be chosen 0 < c1 < c2 < 1) c1 = 0.05 # <---- DECIDED FIXED 0.05 c2 = 0.8 # <---- DECIDED FIXED 0.8 # c2 = 0 extends until ascend location reached: lots of extrapolation # c2 = 1 accepts any point of increased gradient: almost no extrapolation WolfeThreshold = 0.3 # <---- DECIDED FIXED (0.3) # the new free parameter of this method relative to sgd: # search is terminated when probability(Wolfe conditions hold) > WolfeThreshold # not sensitive between 0.1 and 0.8 (0 = accept everyting, 1= accept nothing) offset = 10 # off-set, for numerical stability. EXT = 1 # extrapolation factor tt = 1 # initial step size in scaled space # -- set up GP ------------------------------------------------------------ # create variables with shared scope. Ugly, but necessary because # matlab does not discover shared variables if they are first created in a # nested function. d2m = np.array([]); d3m = np.array([]); V = np.array([]); Vd = np.array([]); dVd = np.array([]) m0 = np.array([]); dm0 = np.array([]); V0= np.array([]);Vd0 = np.array([]); dVd0= np.array([]) V0f = np.array([]); Vd0f= np.array([]); V0df=np.array([]); Vd0df = np.array([]) # kernel: k = lambda a,b: ((np.minimum(a+offset,b+offset)**3)/3 + 0.5 * np.abs(a-b) * np.minimum(a+offset,b+offset)**2) kd = lambda a,b: np.int32(a<b) * ((a+offset)**2)/2 + np.int32(a>=b) * (np.dot(a+offset,b+offset) - 0.5 * (b+offset)**2) dk = lambda a,b: np.int32(a>b) * ((b+offset)**2)/2 + np.int32(a<=b) * (np.dot(a+offset,b+offset) - 0.5 * (a+offset)**2) dkd= lambda a,b: np.minimum(a+offset,b+offset) # further derivatives ddk = lambda a,b: np.int32(a<=b) * (b-a) ddkd= lambda a,b: np.int32(a<=b) dddk= lambda a,b: -np.int32(a<=b) # -- helper functions ----------------------------------------------------- GaussCDF = lambda z: 0.5 * (1 + erf(z/np.sqrt(2))) GaussPDF = lambda z: np.exp( - 0.5 * z**2 ) / np.sqrt(2*np.pi) EI = lambda m,s,eta: (eta - m) * GaussCDF((eta-m)/s) + s * GaussPDF((eta-m)/s) # -- scale ---------------------------------------------------------------- beta = np.abs(np.dot(search_direction.T,df0)) # scale f and df according to 1/(beta*alpha0) # -- scaled noise --------------------------------------------------------- sigmaf = np.sqrt(var_f0)/(alpha0*beta) sigmadf = np.sqrt(np.dot((search_direction**2).T,var_df0))/beta # -- initiate data storage ------------------------------------------------ T = np.array([0]) Y = np.array([0]) dY = np.array(df0)[:,None] dY_projected = np.array([np.dot(df0.T,search_direction)/beta]) Sigmaf = np.array([var_f0]) Sigmadf = np.array(var_df0)[:,None] N = 1 m = [] d1m = [] d2m = [] d3m = [] V = [] Vd = [] dVd = [] m0 = [] dm0 = [] V0 = [] Vd0 = [] dVd0 = [] V0f = [] Vd0f = [] V0df = [] d0df = [] # -- helper functions ----------------------------------------------------- def updateGP(): # using multiscope variables to construct GP nonlocal m nonlocal d1m nonlocal d2m nonlocal d3m nonlocal V nonlocal Vd nonlocal dVd nonlocal m0 nonlocal dm0 nonlocal V0 nonlocal Vd0 nonlocal dVd0 nonlocal V0f nonlocal Vd0f nonlocal V0df nonlocal Vd0df # build Gram matrix kTT = np.zeros([N,N]); kdTT = np.zeros([N,N]); dkdTT = np.zeros([N,N]); for i in range(N): for j in range(N): kTT[i,j] = k(T[i], T[j]) kdTT[i,j] = kd(T[i], T[j]) dkdTT[i,j] = dkd(T[i],T[j]) # build noise matrix Sig = sigmaf**2 * np.ones([2*N, 1]); Sig[N::] = sigmadf**2 # build Gram matrix G = np.diag(Sig.ravel()) + np.r_[np.c_[kTT, kdTT], np.c_[kdTT.T, dkdTT]] A = np.linalg.solve(G, np.append(Y, dY_projected)) # posterior mean function and all its derivatives m = lambda t: np.dot(np.concatenate([k(t, T.T) , kd(t, T.T)]), A) d1m = lambda t: np.dot(np.concatenate([dk(t, T.T) , dkd(t, T.T)]), A) d2m = lambda t: np.dot(np.concatenate([ddk(t, T.T) ,ddkd(t, T.T)]), A) d3m = lambda t: np.dot(np.concatenate([dddk(t, T.T), np.zeros([N])]), A) # posterior marginal covariance between function and first derivative V = lambda t: k(t,t) - np.dot(np.concatenate([k(t, T.T) , kd(t, T.T)]), np.linalg.solve(G , np.concatenate([k(t, T.T) , kd(t, T.T)]).T)) Vd = lambda t: kd(t,t) - np.dot(np.concatenate([k(t, T.T) , kd(t, T.T)]), np.linalg.solve(G , np.concatenate([dk(t, T.T),dkd(t, T.T)]).T)) dVd = lambda t: dkd(t,t) - np.dot(np.concatenate([dk(t, T.T), dkd(t, T.T)]), np.linalg.solve(G , np.concatenate([dk(t, T.T),dkd(t, T.T)]).T)) # belief at starting point, used for Wolfe conditions m0 = m(0) dm0 = d1m(0) V0 = V(0) Vd0 = Vd(0) dVd0 = dVd(0) # covariance terms with function (derivative) values at origin V0f = lambda t: k(0,t) - np.dot(np.concatenate([k(0, T.T) , kd(0, T.T)]) , np.linalg.solve(G, np.concatenate([k(t, T.T) , kd(t, T.T)]).T)) Vd0f = lambda t: dk(0,t) - np.dot(np.concatenate([dk(0, T.T), dkd(0, T.T)]) , np.linalg.solve(G , np.concatenate([k(t, T.T) , kd(t, T.T)]).T)) V0df = lambda t: kd(0,t) - np.dot(np.concatenate([k(0, T.T), kd(0, T.T)]) , np.linalg.solve(G , np.concatenate([dk(t, T.T),dkd(t, T.T)]).T)) Vd0df = lambda t: dkd(0,t)- np.dot(np.concatenate([dk(0, T.T), dkd(0, T.T)]) , np.linalg.solve(G , np.concatenate([dk(t, T.T),dkd(t, T.T)]).T)) # -- update GP with new datapoint ----------------------------------------- updateGP() x_tt = [] y_tt = [] dy_tt = [] var_f_tt = [] var_df_tt = [] alpha0_out = [] def make_outs(y, dy, var_f, var_df): nonlocal x_tt nonlocal y_tt nonlocal dy_tt nonlocal var_f_tt nonlocal var_df_tt nonlocal alpha0_out nonlocal outs x_tt = x0 + tt*alpha0*search_direction # accepted position y_tt = y*(alpha0*beta) + f0 # function value at accepted position dy_tt = dy # gradient at accepted position var_f_tt = var_f # variance of function value at accepted position var_df_tt = var_df # variance of gradients at accepted position # set new set size # next initial step size is 1.3 times larger than last accepted step size alpha0_out = tt*alpha0 * 1.3 # running average for reset in case the step size becomes very small # this is a saveguard gamma = 0.9 outs['alpha_stats'] = gamma*outs['alpha_stats'] + (1-gamma)*tt*alpha0; # reset NEXT initial step size to average step size if accepted step # size is 100 times smaller or larger than average step size if (alpha0_out > 1e2*outs['alpha_stats']) or (alpha0_out < 1e-2*outs['alpha_stats']): if verbosity > 0: print('making a very small step, resetting alpha0') alpha0_out = outs['alpha_stats'] # reset step size def probWolfe(t): # probability for Wolfe conditions to be fulfilled # marginal for Armijo condition ma = m0 - m(t) + c1 * t * dm0 Vaa = V0 + (c1 * t)**2 * dVd0 + V(t) + 2 * (c1 * t * (Vd0 - Vd0f(t)) - V0f(t)) # marginal for curvature condition mb = d1m(t) - c2 * dm0 Vbb = c2**2 * dVd0 - 2 * c2 * Vd0df(t) + dVd(t) # covariance between conditions Vab = -c2 * (Vd0 + c1 * t * dVd0) + V0df(t) + c2 * Vd0f(t) + c1 * t * Vd0df(t) - Vd(t) if (Vaa < 1e-9) and (Vbb < 1e-9): # deterministic evaluations p = np.int32(ma >= 0) * np.int32(mb >= 0) return p, None # joint probability if Vaa <= 0 or Vbb <= 0: p = 0 p12 = np.array([0,0,0]) return p,p12 rho = Vab / np.sqrt(Vaa * Vbb) upper = 2 * c2 * ((np.abs(dm0)+2*np.sqrt(dVd0)-mb)/np.sqrt(Vbb) ) # p = bvn(-ma / np.sqrt(Vaa), np.inf, -mb / np.sqrt(Vbb), upper, rho) _, p, _ = mvn.mvndst(np.array([-ma / np.sqrt(Vaa),-mb / np.sqrt(Vbb)]), np.array([np.inf,upper]), np.array([1, 2]), rho) # if nargout > 1: # individual marginal probabilities for each condition # (for debugging) p12 = np.array([1 - GaussCDF(-ma/np.sqrt(Vaa)), GaussCDF(upper)- GaussCDF(-mb/np.sqrt(Vbb)), Vab / np.sqrt(Vaa * Vbb)]) return p, p12 def cubicMinimum(ts): nonlocal d1m nonlocal d2m nonlocal d3m # mean belief at ts is a cubic function. It is defined up to a constant by d1mt = d1m(ts) d2mt = d2m(ts) d3mt = d3m(ts) a = 0.5 * d3mt b = d2mt - ts * d3mt c = d1mt - d2mt * ts + 0.5 * d3mt * ts**2 if abs(d3mt) < 1e-9: # essentially a quadratic. Single extremum tm = - (d1mt - ts * d2mt) / d2mt return tm # compute the two possible roots: detmnt = b**2 - 4*a*c if detmnt < 0: # no roots tm = np.inf return tm LR = (-b - np.sign(a) * np.sqrt(detmnt)) / (2*a) # left root RR = (-b + np.sign(a) * np.sqrt(detmnt)) / (2*a) # right root # and the two values of the cubic at those points (up to constant) Ldt = LR - ts # delta t for left root Rdt = RR - ts # delta t for right root LCV = d1mt * Ldt + 0.5 * d2mt * Ldt**2 + (d3mt * Ldt**3)/6 # left cubic value RCV = d1mt * Rdt + 0.5 * d2mt * Rdt**2 + (d3mt * Rdt**3)/6 # right cubic value if LCV < RCV: tm = LR else: tm = RR return tm # -- search (until budget used or converged) ------------------------------ for N in range(2,limit+1): # -- evaluate function (change minibatch!) ---------------------------- outs['counter'] = outs['counter'] + 1 [y, dy, var_f, var_df] = func(x0 + tt*alpha0*search_direction) # y: function value at tt # # Test with matlab function TESTFUNCTION=3 # if N==2: # y = 140.9659 # dy = np.array([16.5853, 22.3174]) # var_f = 1 # var_df = np.array([1, 1]) # # if N == 3: # y = 133.2719 # dy = np.array([18.4566,22.1113]) # var_f = 1 # var_df = np.array([1,1]) # # if N == 4: # y = 118.0413 # dy = np.array([18.0786,21.8977]) # var_f = 1 # var_df = np.array([1,1]) # # if N == 5: # y = 85.3557 # dy = np.array([17.3779,18.2448]) # var_f = 1 # var_df = np.array([1,1]) # # if N == 6: # y = 43.5753 # dy = np.array([3.1085,11.2455]) # var_f = 1 # var_df = np.array([1,1]) if np.isinf(y) or np.isnan(y): # this does not happen often, but still needs a fix # e.g. if function return inf or nan (probably too large step), # evaluate again at 0.1*tt print('function values is inf or nan.') print("N " + str(N)) # -- scale function output -------------------------------------------- y = (y - f0)/(alpha0*beta) # substract offset dy_projected = np.dot(dy.T,search_direction)/beta # projected gradient # -- store output ----------------------------------------------------- T = np.append(T,tt) Y = np.append(Y,y) dY = np.hstack([dY,dy.reshape([-1,1])]) dY_projected = np.append(dY_projected, dy_projected) Sigmaf = np.append(Sigmaf, var_f) Sigmadf = np.hstack([Sigmadf,var_df.reshape([-1,1])]) # if N == 7: # ipdb.set_trace() # -- update GP with new datapoint ------------------------------------- updateGP() # store, update belief # -- check last evaluated point for acceptance ------------------------ (probWolfeVal, _) = probWolfe(tt) if probWolfeVal > WolfeThreshold: # are we done? if verbosity > 0: print('found acceptable point.') make_outs(y, dy, var_f, var_df) return outs, alpha0_out, y_tt, dy_tt, x_tt, var_f_tt, var_df_tt # -- find candidates (set up for EI) ---------------------------------- # decide where to probe next: evaluate expected improvement and prob # Wolfe conditions at a number of candidate points, take the most promising one. # lowest mean evaluation, and its gradient (needed for EI): M = np.zeros([N,1]) dM = np.zeros([N,1]) for l in range(N): M[l] = m(T[l]) dM[l] = d1m(T[l]) minj = np.argmin(M) # minm: lowest GP mean, minj: index in candidate list minm = M[minj] tmin = T[minj] # tt of point with lowest GP mean of function values dmin = dM[minj] # GP mean of gradient at lowest point # -- check this point as well for acceptance -------------------------- if np.abs(dmin) < 1e-5 and Vd(tmin) < 1e-4: # nearly deterministic tt = tmin; dy = dY[:, minj]; y = Y[minj]; var_f = Sigmaf[minj]; var_df = Sigmadf[:, minj]; print('found a point with almost zero gradient. Stopping, although Wolfe conditions not guaranteed.') make_outs(y, dy, var_f, var_df) return outs, alpha0_out, y_tt, dy_tt, x_tt, var_f_tt, var_df_tt # -- find candidates -------------------------------------------------- # CANDIDATES 1: minimal means between all evaluations: # iterate through all `cells' (O[N]), check minimum mean locations. Tcand = np.array([]) # positions of candidate points Mcand = np.array([]) # means of candidate points Scand = np.array([]) # standard deviation of candidate points Tsort = np.sort(T) Wolfes = np.array([]) # list of acceptable points. reeval = False for cel in range(N-1): # loop over cells Trep = Tsort[cel] + 1e-6 * (Tsort[cel+1] - Tsort[cel]) cc = cubicMinimum(Trep) # add point to candidate list if minimum lies in between T(cel) and T(cel+1) if cc > Tsort[cel] and cc < Tsort[cel+1]: Tcand = np.append(Tcand,cc) Mcand = np.append(Mcand,m(cc)) Scand = np.append(Scand, np.sqrt(V(cc))) else: # no minimum, just take half-way if cel==1 and d1m(0) > 0: # only in first cell if verbosity > 0: print('function seems very steep, reevaluating close to start.') reeval = True Tcand = np.array([0.1 * (Tsort[cel] + Tsort[cel+1])]) Mcand = np.append(Mcand, m(0.1 * (Tsort[cel] + Tsort[cel+1]))) Scand = np.append(Scand, np.sqrt(V(0.1 * (Tsort[cel] + Tsort[cel+1])))) break # check whether there is an acceptable point among already # evaluated points (since this changes each time the GP gets updated) probWolfeVal, _ = probWolfe(Tsort[cel]) if cel > 1 and (probWolfeVal > WolfeThreshold): Wolfes = np.append(Wolfes,Tsort[cel]) # list of acceptable points. # -- check if at least on point is acceptable ------------------------- if len(Wolfes) > 0: if verbosity > 0: # makePlot() print('found acceptable point.') # -- chose best point among Wolfes, return. ----------------------- MWolfes = 0 * Wolfes; for o in range(len(Wolfes)): MWolfes[o] = m(Wolfes[o]) # compute GP means of acceptable points tt = Wolfes[MWolfes == np.min(MWolfes)] # find corresponding gradient and variances dy = dY[:, T == tt].ravel(); y = Y[T == tt].ravel(); var_f = Sigmaf[T == tt].ravel(); var_df = Sigmadf[:, T==tt].ravel(); make_outs(y, dy, var_f, var_df) # if dy.shape == (2,1): # ipdb.set_trace() return outs, alpha0_out, y_tt, dy_tt, x_tt, var_f_tt, var_df_tt # Candidate 2: one extrapolation step if not reeval: Tcand = np.append(Tcand, np.max(T) + EXT) Mcand = np.append(Mcand, m(np.max(T) + EXT)) Scand = np.append(Scand, np.sqrt(V(np.max(T)+EXT))) # -- discriminate candidates through EI and prob Wolfe ---------------- EIcand = EI(Mcand, Scand, minm) # minm: lowest GP mean of collected evaluations (not candidates) PPcand = np.zeros(EIcand.shape) for ti in range(len(EIcand)): PPcand[ti], _ = probWolfe(Tcand[ti]) idx_best = np.argmax(EIcand * PPcand) # find best candidate if Tcand[idx_best] == tt + EXT: # extrapolating. Extend extrapolation step EXT = 2 * EXT tt = Tcand[idx_best] # makePlot() # -- algorithm reached limit without finding acceptable point. ------------ # Evaluate a final time, return "best" point (one with lowest function value) outs['counter'] = outs['counter'] + 1 [y, dy, var_f, var_df] = func(x0 + tt*alpha0*search_direction) # y: function value at tt if np.isinf(y) or np.isnan(y): # this does not happen often, but still needs a fix # e.g. if function return inf or nan (probably too large step), # evaluate again at 0.1*tt print('function values is inf or nan.') # -- scale function output ------------------------------------------------ y = (y - f0)/(alpha0*beta) # substract offset dy_projected = np.dot(dy.T,search_direction)/beta # projected gradient at tt; # -- store output ----------------------------------------------------- T = np.append(T,tt) Y = np.append(Y,y) dY = np.hstack([dY,dy.reshape([-1,1])]) dY_projected = np.append(dY_projected, dy_projected) N = limit + 1 Sigmaf = np.append(Sigmaf, var_f) Sigmadf = np.hstack([Sigmadf,var_df.reshape([-1,1])]) # -- update GP with new datapoint ----------------------------------------- updateGP() # -- check last evaluated point for acceptance ---------------------------- probWolfeVal, _ = probWolfe(tt) if probWolfeVal > WolfeThreshold: # are we done? if verbosity > 0: print('found acceptable point right at end of budget. Phew!') make_outs(y, dy, var_f, var_df) return outs, alpha0_out, y_tt, dy_tt, x_tt, var_f_tt, var_df_tt # -- return point with lowest mean ---------------------------------------- M = np.ones([N,1])*np.inf for ii in range(N): M[ii] = m(T[ii]) # compute all GP means of all evaluated locations minj = np.argmin(M) # index of minimal GP mean if verbosity > 0: print('reached evaluation limit. Returning ''best'' known point.'); # find corresponding tt, gradient and noise tt = T[minj]; dy = dY[:, minj]; y = Y[minj]; var_f = Sigmaf[minj]; var_df = Sigmadf[:, minj] make_outs(y, dy, var_f, var_df) return outs, alpha0_out, y_tt, dy_tt, x_tt, var_f_tt, var_df_tt
22,197
38.92446
151
py
inversegraphics
inversegraphics-master/render.py
#!/usr/bin/env python3.4m import matplotlib # matplotlib.use('Agg') import scene_io_utils import re from blender_utils import * from collision import * import matplotlib.pyplot as plt numpy.random.seed(1) inchToMeter = 0.0254 outputDir = '../data/output/' if not os.path.exists(outputDir): os.makedirs(outputDir) prefix = '_tmp' width = 110 height = 110 numSamples = 1024 useCycles = False useGPU = False distance = 0.45 numFrames = 200 batchSize = 10 completeScene = True cam = bpy.data.cameras.new("MainCamera") camera = bpy.data.objects.new("MainCamera", cam) world = bpy.data.worlds.new("MainWorld") renderTeapotsList = [2] [targetScenes, targetModels, transformations] = scene_io_utils.loadTargetModels(renderTeapotsList) replaceableScenesFile = '../databaseFull/fields/scene_replaceables.txt' sceneLines = [line.strip() for line in open(replaceableScenesFile)] sceneLineNums = numpy.arange(len(sceneLines)) spout = mathutils.Vector((-6.2, -0.16, 6.25)) handle = mathutils.Vector((6.2, 0.2, 5.7)) tip = mathutils.Vector((0, 0, 8)) occludeHandle = True for sceneNum in sceneLineNums[0:1]: sceneLine = sceneLines[sceneNum] sceneParts = sceneLine.split(' ') sceneFile = sceneParts[0] sceneNumber = int(re.search('.+?scene([0-9]+)\.txt', sceneFile, re.IGNORECASE).groups()[0]) sceneFileName = re.search('.+?(scene[0-9]+\.txt)', sceneFile, re.IGNORECASE).groups()[0] targetIndex = int(sceneParts[1]) instances = scene_io_utils.getSceneInstancesInfo('../databaseFull/scenes/' + sceneFileName) targetParentPosition = instances[targetIndex][2] targetParentIndex = instances[targetIndex][1] [blenderScenes, modelInstances] = scene_io_utils.importBlenderScenes(instances, completeScene, targetIndex) targetParentInstance = modelInstances[targetParentIndex] targetParentInstance.layers[2] = True roomName = '' for model in modelInstances: reg = re.compile('(room[0-9]+)') res = reg.match(model.name) if res: roomName = res.groups()[0] occludingObjName = 'c9fe86bef85fd1d0caeedf6b101df8f6' for model in modelInstances: reg = re.compile('(.*?' + occludingObjName + '.*?)') res = reg.match(model.name) if res: occludingObjName = res.groups()[0] scene = scene_io_utils.composeScene(modelInstances, targetIndex) roomInstance = scene.objects[roomName] ipdb.set_trace() roomInstance.layers[2] = True targetParentInstance.layers[2] = True occludingObjInstance = scene.objects[occludingObjName] newOccludingObjInstance = bpy.data.objects.new(occludingObjName + '_2', None) newOccludingObjInstance.dupli_type = 'GROUP' newOccludingObjInstance.dupli_group = occludingObjInstance.dupli_group newOccludingObjInstance.matrix_world = occludingObjInstance.matrix_world newOccludingObjInstance.pass_index = 0 scene.update() scene.render.threads = 20 scene.render.threads_mode = 'AUTO' bpy.context.screen.scene = scene cycles = bpy.context.scene.cycles scene.render.tile_x = 25 scene.render.tile_y = 25 originalLoc = mathutils.Vector((0,-distance , 0)) setupScene(scene, targetIndex,roomName, world, distance, camera, width, height, numSamples, useCycles, useGPU) bpy.context.user_preferences.system.prefetch_frames = batchSize bpy.context.user_preferences.system.memory_cache_limit = 1000 totalAzimuths = [] totalObjAzimuths = [] totalElevations = [] totalObjectIds = [] frameStart = 0 frameEnd = frameStart + numFrames for teapotIdx, teapotNum in enumerate(renderTeapotsList): director = outputDir teapot = targetModels[teapotIdx] teapot.layers[1] = True teapot.layers[2] = True transformation = transformations[teapotIdx] spout = transformation * spout handle = transformation * handle tip = transformation * tip azimuths = numpy.mod(numpy.random.uniform(270,450, numFrames), 360) # Avoid looking outside the room # azimuths = numpy.array([]) # while len(azimuths) < numFrames: # num = numpy.random.uniform(0,360, 1) # numpy.arccos(mathutils.Vector((-0.6548619270324707, 0.6106656193733215, -0.4452454447746277)) * mathutils.Vector((0.0, -1.0, 0.0))) # objAzimuths = numpy.append(azimuths, num) # objAzimuths = numpy.arange(0,360, 5) # Map it to non colliding rotations. objAzimuths = numpy.array([]) while len(objAzimuths) < numFrames: num = numpy.random.uniform(0,360, 1) # if not(num>= 250 and num<290) and not(num>= 80 and num<110): objAzimuths = numpy.append(objAzimuths, num) elevations = numpy.random.uniform(0,90, numFrames) spoutProjectionsX = numpy.zeros(numFrames) spoutProjectionsY = numpy.zeros(numFrames) handleProjectionsX = numpy.zeros(numFrames) handleProjectionsY = numpy.zeros(numFrames) tipProjectionsX = numpy.zeros(numFrames) tipProjectionsY = numpy.zeros(numFrames) spoutsOccluded = numpy.zeros(numFrames) handlesOccluded = numpy.zeros(numFrames) tipsOccluded = numpy.zeros(numFrames) scene.objects.link(teapot) teapot.layers[1] = True teapot.layers[2] = True teapot.matrix_world = mathutils.Matrix.Translation(targetParentPosition) center = centerOfGeometry(teapot.dupli_group.objects, teapot.matrix_world) original_matrix_world = teapot.matrix_world.copy() # ipdb.set_trace() for frame in range(frameStart, frameEnd): azimuth = azimuths[frame - frameStart] objAzimuth = objAzimuths[frame - frameStart] elevation = elevations[frame - frameStart] bpy.context.scene.frame_set(frame) azimuthRot = mathutils.Matrix.Rotation(radians(-azimuth), 4, 'Z') elevationRot = mathutils.Matrix.Rotation(radians(-elevation), 4, 'X') location = center + azimuthRot * elevationRot * originalLoc camera.location = location azimuthRot = mathutils.Matrix.Rotation(radians(-objAzimuth), 4, 'Z') teapot.matrix_world = mathutils.Matrix.Translation(original_matrix_world.to_translation()) * azimuthRot * (mathutils.Matrix.Translation(-original_matrix_world.to_translation())) * original_matrix_world scene.update() look_at(camera, center) scene.update() teapot.keyframe_insert(data_path="rotation_euler", frame=frame, index=-1) camera.keyframe_insert(data_path="location", frame=frame, index=-1) camera.keyframe_insert(data_path="rotation_euler", frame=frame, index=-1) scene.frame_set(frame) scene.update() sphereSpout_matrix_world = teapot.matrix_world * mathutils.Matrix.Translation(spout) sphereHandle_matrix_world = teapot.matrix_world * mathutils.Matrix.Translation(handle) sphereTip_matrix_world = teapot.matrix_world * mathutils.Matrix.Translation(tip) spoutlocation = image_project(scene, camera, sphereSpout_matrix_world.to_translation()) spoutProjectionsX[frame - frameStart] = spoutlocation[0] spoutProjectionsY[frame - frameStart] = spoutlocation[1] handlelocation = image_project(scene, camera,sphereHandle_matrix_world.to_translation()) handleProjectionsX[frame - frameStart] = handlelocation[0] handleProjectionsY[frame - frameStart] = handlelocation[1] tiplocation = image_project(scene, camera, sphereTip_matrix_world.to_translation()) tipProjectionsX[frame - frameStart] = tiplocation[0] tipProjectionsY[frame - frameStart] = tiplocation[1] #Occlude Handle result, object, matrix, location, normal = scene.ray_cast(scene.camera.location, sphereSpout_matrix_world.to_translation()) spoutOccluded = False if result and object.users_group[0].name != teapot.name: spoutOccluded = True spoutsOccluded[frame - frameStart] = int(spoutOccluded) result, object, matrix, location, normal = scene.ray_cast(scene.camera.location, sphereHandle_matrix_world.to_translation()) handleOccluded = False if result and object.users_group[0].name != teapot.name: handleOccluded = True handlesOccluded[frame - frameStart] = int(handleOccluded) #Occlude Handle: if scene.objects.find(newOccludingObjInstance.name) != -1: scene.objects.unlink(newOccludingObjInstance) if (not result) and occludeHandle and elevation < 35: scene.objects.link(newOccludingObjInstance) initLoc = newOccludingObjInstance.location handleLoc = sphereHandle_matrix_world.to_translation() initLocA = numpy.array(initLoc) handleLocA = numpy.array(handleLoc) newX = ((handleLocA - camera.location)/4 + camera.location)[0] newY = ((handleLocA - camera.location)/4 + camera.location)[1] newOccludingObjInstance.location = mathutils.Vector((newX, newY, initLoc.z)) newOccludingObjInstance.keyframe_insert(data_path="location", frame=frame, index=-1) scene.frame_set(frame) scene.update() result, object, matrix, location, normal = scene.ray_cast(scene.camera.location, sphereHandle_matrix_world.to_translation()) handleOccluded = False if result and object.users_group[0].name != teapot.name: handleOccluded = True handlesOccluded[frame - frameStart] = int(handleOccluded) result, object, matrix, location, normal = scene.ray_cast(scene.camera.location, sphereTip_matrix_world.to_translation()) tipOccluded = False if result and object.users_group[0].name != teapot.name: tipOccluded = True tipsOccluded[frame - frameStart] = int(tipOccluded) numBatches = int(numFrames / batchSize) for batch in range(numBatches): with open(director + 'groundtruth.txt', "a") as groundtruth: for batch_i in range(batchSize): print(str(azimuths[batch * batchSize + batch_i]) + ' ' + str(objAzimuths[batch * batchSize + batch_i]) + ' ' + str(elevations[batch * batchSize + batch_i]) + ' ' + str(teapotNum) + ' ' + str(batch * batchSize + batch_i + frameStart) + ' ' + str(sceneNumber) + ' ' + str(targetIndex) \ + ' ' + str(spoutProjectionsX[batch * batchSize + batch_i]) + ' ' + str(spoutProjectionsY[batch * batchSize + batch_i]) \ + ' ' + str(handleProjectionsX[batch * batchSize + batch_i]) + ' ' + str(handleProjectionsY[batch * batchSize + batch_i]) \ + ' ' + str(tipProjectionsX[batch * batchSize + batch_i]) + ' ' + str(tipProjectionsY[batch * batchSize + batch_i]) \ + ' ' + str(int(spoutsOccluded[batch * batchSize + batch_i])) \ + ' ' + str(int(handlesOccluded[batch * batchSize + batch_i])) \ + ' ' + str(int(tipsOccluded[batch * batchSize + batch_i])) \ + ' ' + prefix, file = groundtruth) scene.frame_start = frameStart + batch * batchSize scene.frame_end = min(frameStart + batch * batchSize + batchSize - 1, frameEnd) scene.layers[1] = True scene.layers[0] = False scene.layers[2] = False scene.render.layers[0].use = False scene.render.layers[2].use = False scene.render.layers[1].use = True cycles.samples = 1 scene.render.engine = 'CYCLES' scene.render.image_settings.file_format = 'OPEN_EXR_MULTILAYER' # scene.render.image_settings.file_format = 'PNG' scene.render.filepath = director + 'render' + prefix + '_obj' + str(teapotNum) + '_' + 'scene' + str(sceneNumber) + '_target' + str(targetIndex) + '_' + 'single_' scene.update() bpy.ops.render.render( animation=True ) scene.layers[1] = False scene.layers[2] = False scene.layers[0] = True scene.render.layers[0].use = True scene.render.layers[1].use = False scene.render.layers[2].use = False # scene.render.image_settings.file_format = 'OPEN_EXR_MULTILAYER' #scene.render.image_settings.file_format = 'PNG' scene.render.filepath = director + 'render' + prefix + '_obj' + str(teapotNum) + '_' + 'scene' + str(sceneNumber) + '_target' + str(targetIndex) + '_' scene.update() bpy.ops.render.render( animation=True ) if useCycles: cycles.samples = numSamples if not useCycles: scene.render.engine = 'BLENDER_RENDER' scene.render.image_settings.file_format = 'PNG' scene.render.filepath = director + 'images/' + 'render' + prefix + '_obj' + str(teapotNum) + '_' + 'scene' + str(sceneNumber) + '_target' + str(targetIndex) + '_' scene.update() bpy.ops.render.render( animation=True ) scene.layers[1] = False scene.layers[2] = True scene.layers[0] = False scene.render.layers[0].use = False scene.render.layers[1].use = False scene.render.layers[2].use = True scene.render.image_settings.file_format = 'PNG' scene.render.filepath = director + 'images/' + 'render' + prefix + '_obj' + str(teapotNum) + '_' + 'scene' + str(sceneNumber) + '_target' + str(targetIndex) + '_unoccluded' scene.update() bpy.ops.render.render( animation=True ) scene.objects.unlink(teapot) objectIds = [teapotNum]*numFrames # with open(director + 'groundtruth' + str(teapotNum) + '.txt', mode='wt', encoding='utf-8') as groundtruth: # print(str(azimuths.tolist())[1:-1], file = groundtruth) # print(str(objAzimuths.tolist())[1:-1], file = groundtruth) # print(str(elevations.tolist())[1:-1], file = groundtruth) # print(str(objectIds)[1:-1], file = groundtruth) totalAzimuths = totalAzimuths + azimuths.tolist() totalObjAzimuths = totalObjAzimuths + objAzimuths.tolist() totalElevations = totalElevations + elevations.tolist() totalObjectIds = totalObjectIds + objectIds # Cleanup for objnum, obji in enumerate(scene.objects): if obji.name != teapot.name: obji.user_clear() bpy.data.objects.remove(obji) scene.user_clear() bpy.data.scenes.remove(scene) # bpy.ops.scene.delete() print("Renders ended.") # with open(director + 'groundtruth_total.txt', mode='wt', encoding='utf-8') as groundtruth: # print(str(totalAzimuths)[1:-1], file = groundtruth) # print(str(totalObjAzimuths)[1:-1], file = groundtruth) # print(str(totalElevations)[1:-1], file = groundtruth) # print(str(totalObjectIds)[1:-1], file = groundtruth) # # with open(director + 'scene.pickle', 'wb') as pfile: # pickle.dump(scene, pfile)
15,462
41.833795
304
py
inversegraphics
inversegraphics-master/diffrender_test.py
test__author__ = 'pol' # from damascene import damascene import matplotlib # matplotlib.use('QT4Agg') import matplotlib.pyplot as plt plt.rcParams['animation.ffmpeg_path'] = '/usr/bin/ffmpeg' import scene_io_utils import mathutils from math import radians import timeit import time import opendr import chumpy as ch import geometry import image_processing import numpy as np import cv2 import glfw import generative_models import recognition_models from opendr_utils import * from utils import * import OpenGL.GL as GL import light_probes from OpenGL import contextdata from light_probes import SHProjection import theano theano.config.optimizer='fast_compile' theano.config.cycle_detection = 'fast' # theano.sandbox.cuda.use('cpu') import lasagne import lasagne_nn plt.ion() ######################################### # Test configuration ######################################### seed = 1 np.random.seed(seed) parameterRecognitionModels = set(['randForestAzs', 'randForestElevs', 'randForestVColors', 'linearRegressionVColors', 'neuralNetModelSHLight', ]) parameterRecognitionModels = set(['randForestAzs', 'randForestElevs', 'randForestVColors', 'linearRegressionVColors', 'linRegModelSHZernike' ]) parameterRecognitionModels = set(['randForestAzs', 'randForestElevs','linearRegressionVColors','neuralNetModelSHLight' ]) parameterRecognitionModels = set(['neuralNetPose', 'linearRegressionVColors','constantSHLight' ]) parameterRecognitionModels = set(['neuralNetPose', 'neuralNetApperanceAndLight', 'neuralNetVColors' ]) parameterRecognitionModels = set(['neuralNetPose', 'neuralNetModelSHLight', 'neuralNetVColors', 'neuralNetModelShape' ]) # parameterRecognitionModels = set(['neuralNetPose', 'neuralNetApperanceAndLight']) # parameterRecognitionModels = set(['randForestAzs', 'randForestElevs','randForestVColors','randomForestSHZernike' ]) # gtPrefix = 'train4_occlusion_shapemodel' # gtPrefix = 'train4_occlusion_shapemodel_synthetic_10K_test100-1100' gtPrefix = 'train4_occlusion_shapemodel_photorealistic_10K_test100-1100' # gtPrefix = 'objectnet3d_teapots' experimentPrefix = 'train4_occlusion_shapemodel_10k' # gtPrefix = 'train4_occlusion_multi' # experimentPrefix = 'train4_occlusion_multi' trainPrefixPose = 'train4_occlusion_shapemodel_10k' trainPrefixVColor = 'train4_occlusion_shapemodel_10k' trainPrefixLightCoeffs = 'train4_occlusion_shapemodel_10k' trainPrefixShapeParams = 'train4_occlusion_shapemodel_10k' trainModelsDirAppLight = 'train4_occlusion_shapemodel_10k' ######################################### # OpenDR Initialization starts here ######################################### #Main script options: glModes = ['glfw','mesa'] glMode = glModes[0] width, height = (150, 150) win = -1 multiObjects = False chThError = None global renderer if glMode == 'glfw': #Initialize base GLFW context for the Demo and to share context among all renderers. glfw.init() glfw.window_hint(glfw.CONTEXT_VERSION_MAJOR, 3) glfw.window_hint(glfw.CONTEXT_VERSION_MINOR, 3) # glfw.window_hint(glfw.OPENGL_FORWARD_COMPAT, GL.GL_TRUE) glfw.window_hint(glfw.OPENGL_PROFILE, glfw.OPENGL_CORE_PROFILE) glfw.window_hint(glfw.DEPTH_BITS,32) glfw.window_hint(glfw.VISIBLE, GL.GL_FALSE) win = glfw.create_window(width, height, "Demo", None, None) glfw.make_context_current(win) angle = 60 * 180 / np.pi clip_start = 0.05 clip_end = 10 frustum = {'near': clip_start, 'far': clip_end, 'width': width, 'height': height} camDistance = 0.4 teapots = [line.strip() for line in open('teapots.txt')] # renderTeapotsList = np.arange(len(teapots)) renderTeapotsList = np.arange(len(teapots))[0:1] hdridx = 0 targetModels = [] v_teapots, f_list_teapots, vc_teapots, vn_teapots, uv_teapots, haveTextures_list_teapots, textures_list_teapots, vflat, varray, center_teapots = scene_io_utils.loadTeapotsOpenDRData(renderTeapotsList, False, False, targetModels) azimuth = np.pi chCosAz = ch.Ch([np.cos(azimuth)]) chSinAz = ch.Ch([np.sin(azimuth)]) chAz = 2*ch.arctan(chSinAz/(ch.sqrt(chCosAz**2 + chSinAz**2) + chCosAz)) chAz = ch.Ch([0]) chObjAz = ch.Ch([0]) chAzRel = chAz - chObjAz elevation = 0 chLogCosEl = ch.Ch(np.log(np.cos(elevation))) chLogSinEl = ch.Ch(np.log(np.sin(elevation))) chEl = 2*ch.arctan(ch.exp(chLogSinEl)/(ch.sqrt(ch.exp(chLogCosEl)**2 + ch.exp(chLogSinEl)**2) + ch.exp(chLogCosEl))) chEl = ch.Ch([0.95993109]) chDist = ch.Ch([camDistance]) chLightSHCoeffs = ch.Ch(np.array([2, 0.25, 0.25, 0.12,-0.17,0.36,0.1,0.,0.])) # if multiObjects: chObjDist = ch.Ch([0]) chObjRotation = ch.Ch([0]) chObjAzMug = ch.Ch([0]) chObjDistMug = ch.Ch([0]) chObjRotationMug = ch.Ch([0]) chVColorsMug = ch.Ch([1,0,0]) clampedCosCoeffs = clampedCosineCoefficients() chComponent = chLightSHCoeffs * clampedCosCoeffs light_color = ch.ones(3) chVColors = ch.Ch([0.4,0.4,0.4]) chDisplacement = ch.Ch([0.0, 0.0,0.0]) chScale = ch.Ch([1.0,1.0,1.0]) # vcch[0] = np.ones_like(vcflat[0])*chVColorsGT.reshape([1,3]) renderer_teapots = [] for teapot_i in range(len(renderTeapotsList)): vmod = v_teapots[teapot_i] fmod_list = f_list_teapots[teapot_i] vcmod = vc_teapots[teapot_i] vnmod = vn_teapots[teapot_i] uvmod = uv_teapots[teapot_i] haveTexturesmod_list = haveTextures_list_teapots[teapot_i] texturesmod_list = textures_list_teapots[teapot_i] centermod = center_teapots[teapot_i] vmod, vnmod, _ = transformObject(vmod, vnmod, chScale, chObjAz, ch.Ch([0]), ch.Ch([0]), np.array([0,0,0])) renderer = createRendererTarget(glMode, chAz, chEl, chDist, centermod, vmod, vcmod, fmod_list, vnmod, light_color, chComponent, chVColors, 0, chDisplacement, width,height, uvmod, haveTexturesmod_list, texturesmod_list, frustum, win ) renderer.overdraw = True renderer.nsamples = 8 renderer.msaa = True renderer.initGL() renderer.initGLTexture() # renderer.initGL_AnalyticRenderer() renderer.imageGT = None renderer.r renderer_teapots = renderer_teapots + [renderer] currentTeapotModel = 0 center = center_teapots[currentTeapotModel] if multiObjects: mugs = [line.strip() for line in open('mugs.txt')] renderMugsList = np.arange(len(teapots))[0:1] v_mugs, f_list_mugs, vc_mugs, vn_mugs, uv_mugs, haveTextures_list_mugs, textures_list_mugs, vflat_mugs, varray_mugs, center_mugs = scene_io_utils.loadMugsOpenDRData( renderMugsList, False, False, None) v_mug = v_mugs[0][0] f_list_mug = f_list_mugs[0][0] chVColorsMug = ch.Ch([1, 0, 0]) vc_mug = [chVColorsMug * np.ones(v_mug[0].shape)] vn_mug = vn_mugs[0][0] uv_mug = uv_mugs[0][0] haveTextures_list_mug = haveTextures_list_mugs[0][0] textures_list_mug = textures_list_mugs[0][0] ######################################### # Initialization ends here ######################################### ######################################### # Generative model set up ######################################### global rendererGT rendererGT = ch.Ch(renderer.r.copy()) numPixels = width*height E_raw = renderer - rendererGT SE_raw = ch.sum(E_raw*E_raw, axis=2) SSqE_raw = ch.SumOfSquares(E_raw)/numPixels initialPixelStdev = 0.1 # finalPixelStdev = 0.05 stds = ch.Ch([initialPixelStdev]) variances = stds ** 2 globalPrior = ch.Ch([0.9]) negLikModel = -ch.sum(generative_models.LogGaussianModel(renderer=renderer, groundtruth=rendererGT, variances=variances)) negLikModelRobust = -ch.sum(generative_models.LogRobustModel(renderer=renderer, groundtruth=rendererGT, foregroundPrior=globalPrior, variances=variances)) pixelLikelihoodCh = generative_models.LogGaussianModel(renderer=renderer, groundtruth=rendererGT, variances=variances) pixelLikelihoodRobustCh = generative_models.LogRobustModel(renderer=renderer, groundtruth=rendererGT, foregroundPrior=globalPrior, variances=variances) post = generative_models.layerPosteriorsRobustCh(rendererGT, renderer, np.array([]), 'MASK', globalPrior, variances)[0] models = [negLikModel, negLikModelRobust] pixelModels = [pixelLikelihoodCh, pixelLikelihoodRobustCh] modelsDescr = ["Gaussian Model", "Outlier model" ] model = 1 pixelErrorFun = pixelModels[model] errorFun = models[model] global iterat iterat = 0 t = time.time() ######################################### # Generative model setup ends here. ######################################### ######################################### # Test code starts here: ######################################### gtDir = 'groundtruth/' + gtPrefix + '/' featuresDir = gtDir experimentDir = 'experiments/' + experimentPrefix + '/' trainModelsDirPose = 'experiments/' + trainPrefixPose + '/' + gtPrefix trainModelsDirVColor = 'experiments/' + trainPrefixVColor + '/' + gtPrefix trainModelsDirLightCoeffs = 'experiments/' + trainPrefixLightCoeffs + '/' + gtPrefix trainModelsDirShapeParams = 'experiments/' + trainPrefixShapeParams + '/' + gtPrefix trainModelsDirAppLight = 'experiments/' + trainModelsDirAppLight + '/' + gtPrefix useCRFOcclusionPred = False useShapeModel = True makeVideo = False reduceVariance = False getColorFromCRF = False syntheticGroundtruth = False evaluateWithGT = True ignoreGT = True ignore = [] if os.path.isfile(gtDir + 'ignore.npy'): ignore = np.load(gtDir + 'ignore.npy') testSet = np.load(experimentDir + 'test.npy') rangeTests = np.arange(len(testSet)) rangeTests = np.arange(100,1100) # rangeTests = np.arange(100,1100) # testSet = testSet[rangeTests] numTests = len(testSet) if evaluateWithGT: groundTruthFilename = gtDir + 'groundTruth.h5' gtDataFile = h5py.File(groundTruthFilename, 'r') shapeGT = gtDataFile[gtPrefix].shape # boolTestSet = np.zeros(shapeGT).astype(np.bool) # boolTestSet[testSet] = True boolTestSet = np.array([np.any(num == testSet) for num in gtDataFile[gtPrefix]['trainIds']]) dataIds = gtDataFile[gtPrefix][boolTestSet]['trainIds'] dataIdsTestIndices = np.array([np.where(dataIds==num)[0][0] for num in testSet]) # testGroundTruth = gtDataFile[gtPrefix][boolTestSet] # groundTruthTest = np.zeros(shapeGT, dtype=testGroundTruth.dtype) # groundTruthTest[boolTestSet] = testGroundTruth groundTruth = gtDataFile[gtPrefix][boolTestSet][dataIdsTestIndices] dataTeapotIdsTest = groundTruth['trainTeapotIds'] test = np.arange(len(testSet)) testSet = testSet[test] print("Reading experiment.") dataAzsGT = groundTruth['trainAzsGT'] dataObjAzsGT = groundTruth['trainObjAzsGT'] dataElevsGT = groundTruth['trainElevsGT'] dataLightAzsGT = groundTruth['trainLightAzsGT'] dataLightElevsGT = groundTruth['trainLightElevsGT'] dataLightIntensitiesGT = groundTruth['trainLightIntensitiesGT'] dataVColorGT = groundTruth['trainVColorGT'] dataScenes = groundTruth['trainScenes'] dataTeapotIds = groundTruth['trainTeapotIds'] dataEnvMaps = groundTruth['trainEnvMaps'] dataOcclusions = groundTruth['trainOcclusions'] dataTargetIndices = groundTruth['trainTargetIndices'] dataLightCoefficientsGT = groundTruth['trainLightCoefficientsGT'] dataLightCoefficientsGTRel = groundTruth['trainLightCoefficientsGTRel'] dataAmbientIntensityGT = groundTruth['trainAmbientIntensityGT'] dataIds = groundTruth['trainIds'] if multiObjects: dataObjDistGT = groundTruth['trainObjDistGT'] dataObjRotationGT = groundTruth['trainObjRotationGT'] dataObjDistMug = groundTruth['trainObjDistMug'] dataObjRotationMug = groundTruth['trainObjRotationMug'] dataObjAzMug = groundTruth['trainObjAzMug'] dataVColorsMug = groundTruth['trainVColorsMug'] dataObjAzMugRel = groundTruth['trainObjAzMugRel'] dataObjAzGTRel = groundTruth['trainObjAzGTRel'] dataMugPosOffset = groundTruth['trainMugPosOffset'] dataTeapotPosOffset = groundTruth['trainTeapotPosOffset'] if useShapeModel: dataShapeModelCoeffsGT = groundTruth['trainShapeModelCoeffsGT'] gtDtype = groundTruth.dtype testSetFixed = testSet whereBad = [] for test_it, test_id in enumerate(testSet): if test_id in ignore: bad = np.where(testSetFixed==test_id) testSetFixed = np.delete(testSetFixed, bad) whereBad = whereBad + [bad] # testSet = testSetFixed loadFromHdf5 = False synthPrefix = '_cycles' if syntheticGroundtruth: synthPrefix = '' if syntheticGroundtruth: imagesDir = gtDir + 'images_opendr/' else: imagesDir = gtDir + 'images/' if evaluateWithGT: images = readImages(imagesDir, dataIds, loadFromHdf5) else: import glob imageFiles = glob.glob1(imagesDir, "*.png") images = np.zeros([len(imageFiles), width, height, 3]) for imageFile_i, imageFile in enumerate(imageFiles): image = skimage.io.imread(imagesDir + imageFile) image = skimage.transform.resize(image , [height, width]) images[imageFile_i] = image testSet = np.arange(len(images)) print("Backprojecting and fitting estimates.") # testSet = np.arange(len(images))[0:10] if evaluateWithGT: testAzsGT = dataAzsGT testObjAzsGT = dataObjAzsGT testElevsGT = dataElevsGT testLightAzsGT = dataLightAzsGT testLightElevsGT = dataLightElevsGT testLightIntensitiesGT = dataLightIntensitiesGT testVColorGT = dataVColorGT if useShapeModel: testShapeParamsGT = dataShapeModelCoeffsGT testLightCoefficientsGTRel = dataLightCoefficientsGTRel * dataAmbientIntensityGT[:,None] testAzsRel = np.mod(testAzsGT - testObjAzsGT, 2*np.pi) testOcclusions = dataOcclusions testIds = dataIds if multiObjects: testObjAzMugRel = dataObjAzMugRel testObjAzGTRel = dataObjAzGTRel testMugPosOffset = dataMugPosOffset testTeapotPosOffset = dataTeapotPosOffset testVColorMug = dataVColorsMug testObjDistGT = dataObjDistGT testObjRotationGT = dataObjRotationGT testObjDistMug = dataObjDistMug testObjRotationMug = dataObjRotationMug testObjAzMug = dataObjAzMug testVColorsMug = dataVColorsMug else: #There is no GT. testAzsGT = np.zeros(len(testSet)) testObjAzsGT = np.zeros(len(testSet)) testElevsGT = np.zeros(len(testSet)) testLightAzsGT = np.zeros(len(testSet)) testLightElevsGT = np.zeros(len(testSet)) testLightIntensitiesGT = np.zeros(len(testSet)) testOcclusions = np.zeros(len(testSet)) testIds = np.zeros(len(testSet)) testVColorGT = np.zeros(len(testSet)) if useShapeModel: testShapeParamsGT = np.zeros(len(testSet)) testLightCoefficientsGTRel = np.zeros(len(testSet)) testAzsRel = np.zeros(len(testSet)) if multiObjects: testObjAzMugRel = np.zeros(len(testSet)) testObjAzGTRel = np.zeros(len(testSet)) testMugPosOffset = np.zeros(len(testSet)) testTeapotPosOffset = np.zeros(len(testSet)) testVColorMug = np.zeros(len(testSet)) testObjDistGT = np.zeros(len(testSet)) testObjRotationGT = np.zeros(len(testSet)) testObjDistMug = np.zeros(len(testSet)) testObjRotationMug = np.zeros(len(testSet)) testObjAzMug = np.zeros(len(testSet)) testVColorsMug = np.zeros(len(testSet)) ##Read Training set labels # trainGTPrefix = 'train4_occlusion_shapemodel' # # trainGTDir = 'groundtruth/' + trainGTPrefix + '/' # trainGroundTruthFilename = trainGTDir + 'groundTruth.h5' # trainGTDataFile = h5py.File(trainGroundTruthFilename, 'r') # # trainSet = np.load(experimentDir + 'train.npy') # # shapeGT = trainGTDataFile[trainGTPrefix].shape # boolTrainSet = np.zeros(shapeGT).astype(np.bool) # boolTrainSet[trainSet] = True # trainGroundTruth = trainGTDataFile[trainGTPrefix][boolTrainSet] # groundTruthTrain = np.zeros(shapeGT, dtype=trainGroundTruth.dtype) # groundTruthTrain[boolTrainSet] = trainGroundTruth # groundTruthTrain = groundTruthTrain[trainSet] # dataTeapotIdsTrain = groundTruthTrain['trainTeapotIds'] # train = np.arange(len(trainSet)) # # testSet = testSet[test] # # print("Reading experiment.") # trainAzsGT = groundTruthTrain['trainAzsGT'] # trainObjAzsGT = groundTruthTrain['trainObjAzsGT'] # trainElevsGT = groundTruthTrain['trainElevsGT'] # trainLightAzsGT = groundTruthTrain['trainLightAzsGT'] # trainLightElevsGT = groundTruthTrain['trainLightElevsGT'] # trainLightIntensitiesGT = groundTruthTrain['trainLightIntensitiesGT'] # trainVColorGT = groundTruthTrain['trainVColorGT'] # trainScenes = groundTruthTrain['trainScenes'] # trainTeapotIds = groundTruthTrain['trainTeapotIds'] # trainEnvMaps = groundTruthTrain['trainEnvMaps'] # trainOcclusions = groundTruthTrain['trainOcclusions'] # trainTargetIndices = groundTruthTrain['trainTargetIndices'] # trainLightCoefficientsGT = groundTruthTrain['trainLightCoefficientsGT'] # trainLightCoefficientsGTRel = groundTruthTrain['trainLightCoefficientsGTRel'] # trainAmbientIntensityGT = groundTruthTrain['trainAmbientIntensityGT'] # trainIds = groundTruthTrain['trainIds'] # # if useShapeModel: # trainShapeModelCoeffsGT = groundTruthTrain['trainShapeModelCoeffsGT'] # # trainLightCoefficientsGTRel = trainLightCoefficientsGTRel * trainAmbientIntensityGT[:,None] # # trainAzsRel = np.mod(trainAzsGT - trainObjAzsGT, 2*np.pi) # latexify(columns=2) # # directory = 'tmp/occlusions' # fig = plt.figure() # ax = fig.add_subplot(111) # ax.hist(trainOcclusions*100, bins=40) # ax.set_xlabel('Occlusion level (\%)') # ax.set_ylabel('Counts') # ax.set_title('Occlusion histogram') # ax.set_xlim(0,100) # fig.savefig(directory + '-histogram.pdf', bbox_inches='tight') # plt.close(fig) recognitionTypeDescr = ["near", "mean", "sampling"] recognitionType = 1 optimizationTypeDescr = ["predict", "optimize", "joint"] optimizationType = 1 computePredErrorFuns = True method = 5 model = 1 maxiter = 100 numSamples = 1 mintime = time.time() free_variables = [ chAz, chEl] boundEl = (-np.pi, np.pi) boundAz = (-3*np.pi, 3*np.pi) boundscomponents = (0,None) bounds = [boundAz,boundEl] bounds = [(None , None ) for sublist in free_variables for item in sublist] methods = ['dogleg', 'minimize', 'BFGS', 'L-BFGS-B', 'Nelder-Mead', 'TNC', 'SGDMom', 'probLineSearch'] # options = {'disp':False, 'maxiter':maxiter, 'lr':0.0001, 'momentum':0.1, 'decay':0.99} options = {'disp':False, 'maxiter':maxiter} azVar = 0.1 elVar = 0.1 vColorVar = 0.01 shCoeffsVar = 0.01 df_vars = np.concatenate([azVar*np.ones(chAz.shape), elVar*np.ones(chEl.shape), vColorVar*np.ones(chVColors.r.shape), shCoeffsVar*np.ones(chLightSHCoeffs.r.shape)]) options = {'disp':False, 'maxiter':maxiter, 'df_vars':df_vars} testRenderer = 0 # #Read texture UV mapping # sceneFilename = 'teapot.blend' # with bpy.data.libraries.load(filepath=sceneFilename) as (data_from, data_to): # for attr in dir(data_to): # setattr(data_to, attr, getattr(data_from, attr)) # teapot = bpy.data.scenes['Teapot'].objects[1] # vmod, fmod_list, vcmod, vnmod, uvmod, haveTexturesmod_list, texturesmod_list = scene_io_utils.unpackBlenderObject(teapot, '', False) # color = np.array([0,0,0]) # texturesmod_list[0][0][0] = ch.Ch(np.ones(texturesmod_list[0][0][0].shape)*color[None,None,:]) # testRenderer = 2 if useShapeModel: import shape_model #%% Load data filePath = 'data/teapotModel.pkl' teapotModel = shape_model.loadObject(filePath) faces = teapotModel['faces'] #%% Sample random shape Params latentDim = np.shape(teapotModel['ppcaW'])[1] shapeParams = np.zeros(latentDim) chShapeParams = ch.Ch(shapeParams.copy()) meshLinearTransform=teapotModel['meshLinearTransform'] W=teapotModel['ppcaW'] b=teapotModel['ppcaB'] chVertices = shape_model.VerticesModel(chShapeParams=chShapeParams,meshLinearTransform=meshLinearTransform,W = W,b=b) chVertices.init() chVertices = ch.dot(geometry.RotateZ(-np.pi/2)[0:3,0:3],chVertices.T).T smFaces = [[faces]] smVColors = [chVColors*np.ones(chVertices.shape)] # smUVs = ch.Ch(np.zeros([chVertices.shape[0],2])) smUVs = uvmod[0] # smHaveTextures = [[False]] smHaveTextures = [[False]] smTexturesList = [[None]] # smTexturesList = [[texturesmod_list[0][0][0]]] chVertices = chVertices - ch.mean(chVertices, axis=0) minZ = ch.min(chVertices[:,2]) chMinZ = ch.min(chVertices[:,2]) zeroZVerts = chVertices[:,2]- chMinZ chVertices = ch.hstack([chVertices[:,0:2] , zeroZVerts.reshape([-1,1])]) chVertices = chVertices*0.09 smCenter = ch.array([0,0,0.1]) smVertices = [chVertices] chNormals = shape_model.chGetNormals(chVertices, faces) smNormals = [chNormals] if useShapeModel: center = smCenter UVs = smUVs v = smVertices vn = smNormals Faces = smFaces VColors = smVColors HaveTextures = smHaveTextures TexturesList = smTexturesList else: v, vn = v_teapots[currentTeapotModel][0], vn_teapots[currentTeapotModel][0] Faces = f_list_teapots[currentTeapotModel][0] VColors = vc_teapots[currentTeapotModel][0] UVs = uv_teapots[currentTeapotModel][0] HaveTextures = haveTextures_list_teapots[currentTeapotModel][0] TexturesList = textures_list_teapots[currentTeapotModel][0] v, vn, teapotPosOffset = transformObject(v, vn, chScale, chObjAz, chObjDist, chObjRotation, np.array([0,0,0])) verticesCube, facesCube, normalsCube, vColorsCube, texturesListCube, haveTexturesCube = getCubeData() uvCube = np.zeros([verticesCube.shape[0],2]) chCubePosition = ch.Ch([0, 0, 0]) chCubeScale = ch.Ch([10.0]) chCubeAzimuth = ch.Ch([0]) chCubeVCColors = ch.Ch(np.ones_like(vColorsCube) * 1.0) v_transf, vn_transf = transformObject2([verticesCube], [normalsCube], chCubeScale, chCubeAzimuth, chCubePosition) v_scene = [v] f_list_scene = [smFaces] vc_scene = [smVColors] vn_scene = [vn] uv_scene = [smUVs] haveTextures_list_scene = [smHaveTextures] textures_list_scene = [smTexturesList] addObjectData(v_scene, f_list_scene, vc_scene, vn_scene, uv_scene, haveTextures_list_scene, textures_list_scene, v_transf, [[facesCube]],[chCubeVCColors], vn_transf, [uvCube], haveTexturesCube, texturesListCube) if multiObjects: verticesMug, normalsMug, mugPosOffset = transformObject(v_mug, vn_mug, chScale, chObjAzMug + np.pi / 2, chObjDistMug, chObjRotationMug, np.array([0,0,0])) VerticesB = [v] + [verticesMug] NormalsB = [vn] + [normalsMug] FacesB = [Faces] + [f_list_mug] VColorsB = [VColors] + [vc_mug] UVsB = [UVs] + [uv_mug] HaveTexturesB = [HaveTextures] + [haveTextures_list_mug] TexturesListB = [TexturesList] + [textures_list_mug] renderer = createRendererTarget(glMode, chAz, chEl, chDist, center, VerticesB, VColorsB, FacesB, NormalsB, light_color,chComponent, chVColors, np.array([0,0,0]), chDisplacement, width, height, UVsB, HaveTexturesB, TexturesListB, frustum, None) renderer.initGL() renderer.initGLTexture() else: # renderer = createRendererTarget(glMode, chAz, chEl, chDist, smCenter, [v], [smVColors], [smFaces], [vn], light_color, chComponent, chVColors, 0, chDisplacement, width,height, [smUVs], [smHaveTextures], [smTexturesList], frustum, win ) renderer = createRendererTarget(glMode, chAz, chEl, chDist, smCenter, v_scene, vc_scene, f_list_scene, vn_scene, light_color, chComponent, chVColors, 0, chDisplacement, width,height, uv_scene, haveTextures_list_scene, textures_list_scene, frustum, win ) renderer.overdraw = True renderer.nsamples = 8 renderer.msaa = True renderer.initGL() renderer.initGLTexture() # renderer.initGL_AnalyticRenderer() renderer.imageGT = None renderer.r # chShapeParams[:] = np.zeros([latentDim]) chVerticesMean = chVertices.r.copy() else: renderer = renderer_teapots[testRenderer] # plt.imsave('testrender.png', sqeRenderer.render_image) loadMask = True if loadMask: masksGT = loadMasks(gtDir + '/masks_occlusion/', testSet) # ### Groundtruth triplets generation # # groundTruthFilename = 'groundtruth/' + gtPrefix + '/' '/groundTruth.h5' # gtDataFileToRender = h5py.File(groundTruthFilename, 'r') # groundTruthToRender = gtDataFileToRender[gtPrefix] # # rangeGT = np.arange(len(groundTruthToRender)) # # for gtIdx in rangeGT[:]: # groundTruthToRender['trainEnvMapPhiOffsets'][gtIdx] # # color = groundTruthToRender['trainVColorGT'][gtIdx] # az = groundTruthToRender['trainObjAzsGT'][gtIdx] - groundTruthToRender['trainAzsGT'][gtIdx] # el = groundTruthToRender['trainElevsGT'][gtIdx] # lightCoefficientsRel = groundTruthToRender['trainLightCoefficientsGTRel'][gtIdx] * groundTruthToRender['trainAmbientIntensityGT'][gtIdx] # # if useShapeModel: # shapeParams = groundTruthToRender['trainShapeModelCoeffsGT'][gtIdx] # # chAz[:] = 0 # chEl[:] = el # chObjAz[:] = az # chVColors[:] = color # chLightSHCoeffs[:] = lightCoefficientsRel # if useShapeModel: # chShapeParams[:] = shapeParams # # image = renderer.r[:].copy() # lin2srgb(image) # cv2.imwrite(gtDir + '/backprojections/im' + str(gtIdx) + '.jpeg', 255 * image[:, :, [2, 1, 0]], [int(cv2.IMWRITE_JPEG_QUALITY), 100]) # # ipdb.set_trace() # vis_im = np.array(renderer.indices_image==1).copy().astype(np.bool) # im = skimage.io.imread('renderergt539.jpeg').astype(np.float32)/255. # rendererGT[:] = srgb2lin(im.copy()) # post = generative_models.layerPosteriorsRobustCh(rendererGT, renderer, vis_im, 'MASK', globalPrior, variances)[0].r>0.05 # render = renderer.r.copy() # render[~mask*vis_im] = np.concatenate([np.ones([1000,1000])[:,:,None], np.zeros([1000,1000])[:,:,None],np.zeros([1000,1000])[:,:,None]], axis=2)[~mask*vis_im] # # render[renderer.boundarybool_image.astype(np.bool)] = renderer.r[renderer.boundarybool_image.astype(np.bool)] # # cv2.imwrite('renderer' + '.jpeg' , 255*lin2srgb(render[:,:,[2,1,0]]), [int(cv2.IMWRITE_JPEG_QUALITY), 100]) # cv2.imwrite('rendererGT' + '.jpeg' , 255*lin2srgb(rendererGT.r[:,:,[2,1,0]]), [int(cv2.IMWRITE_JPEG_QUALITY), 100]) # plt.imsave('renderered.png', lin2srgb(render)) ##### Tests with optimizing multiple objects in scene and camera settings. # test_i = 0 # for test_i in range(len(testSet)): # # color = testVColorGT[test_i] # az = testAzsGT[test_i] # el = testElevsGT[test_i] # lightCoefficientsRel = testLightCoefficientsGTRel[test_i] # # if useShapeModel: # shapeParams = testShapeParamsGT[test_i] # # chAz[:] = 0 # chEl[:] = el # chVColors[:] = color # chLightSHCoeffs[:] = lightCoefficientsRel # if useShapeModel: # chShapeParams[:] = shapeParams # # chObjAz[:] = testObjAzsGT[test_i] - testAzsGT[test_i] # chObjDist[:] = testObjDistGT[test_i] # chObjRotation[:] = testObjRotationGT[test_i] - testAzsGT[test_i] # chObjAzMug[:] = testObjAzMug[test_i] - testAzsGT[test_i] # chObjDistMug[:] = testObjDistMug[test_i] # chObjRotationMug[:] = testObjRotationMug[test_i] - testAzsGT[test_i] # # chVColorsMug[:] = testVColorMug[test_i] # # image = skimage.transform.resize(images[test_i], [height, width]) # imageSrgb = image.copy() # rendererGT[:] = srgb2lin(image) # # masksGT = loadMasks(gtDir + '/masks_occlusion/', testSet) # masksMug = loadMasksMug(gtDir + '/masks_occlusion/', testSet) # # maskTeapot = masksGT[test_i] # maskMug = masksMug[test_i] # # coords = np.meshgrid(np.arange(width)-width/2, np.arange(height)-height/2) # # coordsMugX = coords[1][maskMug] # coordsMugY = coords[0][maskMug] # # bbRendererGT = rendererGT.r.copy() # # bbRendererGT[coordsMugX.min()+width/2, coordsMugY.min()+height/2 : coordsMugY.max() + height/2] = np.array([1,0,0]) # bbRendererGT[coordsMugX.max()+width/2, coordsMugY.min()+height/2 : coordsMugY.max() + height/2] = np.array([1,0,0]) # bbRendererGT[coordsMugX.min()+width/2 : coordsMugX.max() + width/2, coordsMugY.min()+height/2] = np.array([1,0,0]) # bbRendererGT[coordsMugX.min()+width/2:coordsMugX.max()+width/2, coordsMugY.max() + height/2] = np.array([1,0,0]) # # coordsTeapotX = coords[1][maskTeapot] # coordsTeapotY = coords[0][maskTeapot] # # bbRendererGT[coordsTeapotX.min()+width/2, coordsTeapotY.min()+height/2 : coordsTeapotY.max() + height/2] = np.array([1,0,0]) # bbRendererGT[coordsTeapotX.max()+width/2, coordsTeapotY.min()+height/2 : coordsTeapotY.max() + height/2] = np.array([1,0,0]) # bbRendererGT[coordsTeapotX.min()+width/2 : coordsTeapotX.max() + width/2, coordsTeapotY.min()+height/2] = np.array([1,0,0]) # bbRendererGT[coordsTeapotX.min()+width/2:coordsTeapotX.max()+width/2, coordsTeapotY.max() + height/2] = np.array([1,0,0]) # # # posMug = np.array([(coordsMugY.min() + coordsMugY.max())/2, (coordsMugX.min() + coordsMugX.max())/2]) # posTeapot = np.array([(coordsTeapotY.min() + coordsTeapotY.max())/2, (coordsTeapotX.min() + coordsTeapotX.max())/2]) # # # createRendererTarget(glMode, False, chAz, chEl, chDist, center, VerticesB, VColorsB, FacesB, NormalsB, light_color,chComponent, chVColors, np.array([0,0,0]), chDisplacement, width, height, UVsB, HaveTexturesB, TexturesListB, frustum, None) # chAzCam = ch.Ch([0]) # chElCam = ch.Ch(chEl.r.copy()) # # relPosTeapotGT = v[0].r.sum(axis=0)/v[0].r.shape[0] # relPosMugGT = verticesMug[0].r.sum(axis=0)/verticesMug[0].r.shape[0] # # bbRendererGT[posMug[1] - 2+width/2:posMug[1] + 2+width/2, posMug[0] - 2 + height/2: posMug[0] + 2 + height/2] = np.array([1,0,0]) # bbRendererGT[posTeapot[1] - 2+width/2:posTeapot[1] + 2+width/2, posTeapot[0] - 2 + height/2: posTeapot[0] + 2 + height/2] = np.array([1,0,0]) # # _, _, camTransfomMatGT = setupCamera(np.array([0,0,0]), chAz.r, chEl.r, chDist, np.array([0,0,0.1]), width, height) # camEyeGT = camTransfomMatGT[0:4,0:4].dot(np.array([0,0,0,1]))[0:3] # # vecMugToCamGT = camEyeGT - (mugPosOffset + np.array([0,0,0.1])) # mugCamElGT = 2*ch.arctan(ch.norm(ch.array([0,-1,0])*ch.norm(vecMugToCamGT) - vecMugToCamGT*ch.norm(ch.array([0,-1,0])))/ch.norm(ch.array([0,-1,0])*ch.norm(vecMugToCamGT) + ch.norm(ch.array([0,-1,0]))*vecMugToCamGT)) # # vecTeapotToCamGT = camEyeGT - (teapotPosOffset + np.array([0,0,0.1])) # teapotCamElGT = 2*ch.arctan(ch.norm(ch.array([0,-1,0])*ch.norm(vecTeapotToCamGT) - vecTeapotToCamGT*ch.norm(ch.array([0,-1,0])))/ch.norm(ch.array([0,-1,0])*ch.norm(vecTeapotToCamGT) + ch.norm(ch.array([0,-1,0]))*vecTeapotToCamGT)) # # objDisplacementMat = computeHemisphereTransformation(chObjRotationMug, 0, chObjDistMug, np.array([0, 0, 0.05])) # pointMug = objDisplacementMat[0:3, 3] # # # pointMug = ch.Ch([0,0,0.1]) # imPosMugCam, modelRotation, camTransfomMat = setupCamera(pointMug, chAzCam, chElCam, chDist, np.array([0,0,0.1]), width, height) # imPosMug = imPosMugCam - np.array([height/2, width/2]) # # objDisplacementMat = computeHemisphereTransformation(chObjRotation, 0, chObjDist, np.array([0, 0, 0.1])) # pointTeapot = objDisplacementMat[0:3, 3] # # pointTeapot = ch.Ch([0,0,0.1]) # # imPosTeapotCam, modelRotation, camTransfomMat = setupCamera(pointTeapot, chAzCam, chElCam, chDist, np.array([0,0,0.1]), width, height) # imPosTeapot = imPosTeapotCam - np.array([height/2, width/2]) # # errMug = ch.sum((imPosMug - posMug)**2) # errTeapot = ch.sum((imPosTeapot - posTeapot)**2) # # camEye = camTransfomMat[0:4,0:4].dot(np.array([0,0,0, 1]))[0:3] # # vecMugToCam = camEye - pointMug # mugCamEl = 2*ch.arctan(ch.norm(ch.array([0,-1,0])*ch.norm(vecMugToCam) - vecMugToCam*ch.norm(ch.array([0,-1,0])))/ch.norm(ch.array([0,-1,0])*ch.norm(vecMugToCam) + ch.norm(ch.array([0,-1,0]))*vecMugToCam)) # # vecTeapotToCam = camEye - pointTeapot # # teapotCamEl = 2*ch.arctan(ch.norm(ch.array([0,-1,0])*ch.norm(vecTeapotToCam) - vecTeapotToCam*ch.norm(ch.array([0,-1,0])))/ch.norm(ch.array([0,-1,0])*ch.norm(vecTeapotToCam) + ch.norm(ch.array([0,-1,0]))*vecTeapotToCam)) # # chElPredMug = mugCamElGT # chElPredTeapot = teapotCamElGT # # errElMug = ch.sum((chElPredMug*180/np.pi - mugCamEl*180/np.pi)**2) # errElTeapot = ch.sum((chElPredTeapot*180/np.pi - teapotCamEl*180/np.pi)**2) # # # spatialFreeVars = [pointTeapot[0:2], pointMug[0:2], chElCam] # spatialFreeVars = [chObjRotationMug, chObjRotation, chObjDistMug, chObjDist, chElCam] # # spatialErrorFun = errMug + errTeapot + errElMug+ errElTeapot # # def cbS(_): # pass # # bbRendererGT[imPosTeapot[1] - 2 +width/2:imPosTeapot[1] + 2+width/2, imPosTeapot[0] - 2 + height/2: imPosTeapot[0] + 2 + height/2] = np.array([0,0,1]) # bbRendererGT[imPosMug[1] - 2+width/2:imPosMug[1] + 2+width/2, imPosMug[0] - 2 + height/2: imPosMug[0] + 2 + height/2] = np.array([0,0,1]) # # plt.imsave('tmp/rendererPred' + str(test_i) + '.png', renderer.r) # # #Dumb initial state (center of table). # chObjDist[:] = 0 # chObjRotation[:] = 0 # chObjDistMug[:] = 0 # chObjRotationMug[:] = 0 # # plt.imsave('tmp/rendererInit' + str(test_i) + '.png', renderer.r) # # # ch.minimize({'raw': spatialErrorFun }, bounds=[(-0.5,0.5), (-0.5,0.5), (0,np.pi/2)], method=methods[1], x0=spatialFreeVars, callback=cbS, options={'disp':False, 'maxiter':10}) # ch.minimize({'raw': spatialErrorFun }, bounds=[(-2*np.pi,2*np.pi), (-2*np.pi,2*np.pi), (0,0.5), (0,0.5), (0,np.pi/2)], method=methods[2], x0=spatialFreeVars, callback=cbS, options={'disp':True, 'maxiter':50}) # # bbRendererGT[imPosTeapot[1] - 2+width/2:imPosTeapot[1] + 2+width/2, imPosTeapot[0] - 2 + height/2: imPosTeapot[0] + 2 + height/2] = np.array([0,1,0]) # bbRendererGT[imPosMug[1] - 2+width/2:imPosMug[1] + 2+width/2, imPosMug[0] - 2 + height/2: imPosMug[0] + 2 + height/2] = np.array([0,1,0]) # # plt.imsave('tmp/bbRendererGT' + str(test_i) + '.png', bbRendererGT) # # chEl[:] = chElCam.r.copy() # # plt.imsave('tmp/rendererOpt' + str(test_i) + '.png', renderer.r) import skimage.color nearGTOffsetRelAz = 0.01 nearGTOffsetEl = 0.01 nearGTOffsetLighCoeffs = np.zeros(9) nearGTOffsetVColor = np.zeros(3) ############ RECOGNITION MDOEL PREDICTIONS (COMPUTE AND SAVE, DON'T USE THIS EVERY TIME) #Load trained recognition models nnBatchSize = 100 azsPredictions = np.array([]) recomputeMeans = False includeMeanBaseline = False recomputePredictions = False recomputePredictionsPose = False if includeMeanBaseline: meanTrainLightCoefficientsGTRel = np.repeat(np.mean(trainLightCoefficientsGTRel, axis=0)[None,:], numTests, axis=0) meanTrainElevation = np.repeat(np.mean(trainElevsGT, axis=0), numTests, axis=0) meanTrainAzimuthRel = np.repeat(0, numTests, axis=0) meanTrainShapeParams = np.repeat(np.zeros([latentDim])[None,:], numTests, axis=0) meanTrainVColors = np.repeat(np.mean(trainVColorGT, axis=0)[None,:], numTests, axis=0) chShapeParams[:] = np.zeros([10]) meanTrainShapeVertices = np.repeat(chVertices.r.copy()[None,:], numTests, axis=0) # if recomputeMeans or not os.path.isfile(experimentDir + "meanTrainEnvMapProjections.npy"): # envMapTexture = np.zeros([180,360,3]) # approxProjectionsPredList = [] # for train_i in range(len(trainSet)): # pEnvMap = SHProjection(envMapTexture, np.concatenate([trainLightCoefficientsGTRel[train_i][:,None], trainLightCoefficientsGTRel[train_i][:,None], trainLightCoefficientsGTRel[train_i][:,None]], axis=1)) # approxProjectionPred = np.sum(pEnvMap, axis=(2,3)) # # approxProjectionsPredList = approxProjectionsPredList + [approxProjectionPred[None,:]] # approxProjections = np.vstack(approxProjectionsPredList) # meanTrainEnvMapProjections = np.mean(approxProjections, axis=0) # np.save(experimentDir + 'meanTrainEnvMapProjections.npy', meanTrainEnvMapProjections) # else: # meanTrainEnvMapProjections = np.load(experimentDir + 'meanTrainEnvMapProjections.npy') # meanTrainEnvMapProjections = np.repeat(meanTrainEnvMapProjections[None,:], numTests, axis=0) meanTrainEnvMapProjections = None rangeTests = np.arange(len(testSet)) if recomputePredictionsPose or not os.path.isfile(trainModelsDirPose + "elevsPred.npy"): if 'neuralNetPose' in parameterRecognitionModels: poseModel = "" with open(experimentDir + 'neuralNetModelPose.pickle', 'rb') as pfile: neuralNetModelPose = pickle.load(pfile) meanImage = neuralNetModelPose['mean'] modelType = neuralNetModelPose['type'] param_values = neuralNetModelPose['params'] grayTestImages = 0.3*images[:,:,:,0] + 0.59*images[:,:,:,1] + 0.11*images[:,:,:,2] grayTestImages = grayTestImages[:,None, :,:] grayTestImages = grayTestImages - meanImage.reshape([1,1, grayTestImages.shape[2],grayTestImages.shape[3]]) network = lasagne_nn.load_network(modelType=modelType, param_values=param_values) posePredictionFun = lasagne_nn.get_prediction_fun(network) posePredictions = np.zeros([len(grayTestImages), 4]) for start_idx in range(0, len(grayTestImages), nnBatchSize): posePredictions[start_idx:start_idx + nnBatchSize] = posePredictionFun(grayTestImages.astype(np.float32)[start_idx:start_idx + nnBatchSize]) # posePredictions = posePredictionFun(grayTestImages.astype(np.float32)) cosAzsPred = posePredictions[:,0] sinAzsPred = posePredictions[:,1] cosElevsPred = posePredictions[:,2] sinElevsPred = posePredictions[:,3] elevsPred = np.arctan2(sinElevsPred, cosElevsPred) azsPred = np.arctan2(sinAzsPred, cosAzsPred) np.save(trainModelsDirPose + 'elevsPred.npy', elevsPred) np.save(trainModelsDirPose + 'azsPred.npy', azsPred) # ##Get predictions with dropout on to get samples. # with open(trainModelsDirPose + 'neuralNetModelPose.pickle', 'rb') as pfile: # neuralNetModelPose = pickle.load(pfile) # # meanImage = neuralNetModelPose['mean'] # modelType = neuralNetModelPose['type'] # param_values = neuralNetModelPose['params'] # # # network = lasagne_nn.load_network(modelType=modelType, param_values=param_values) # nonDetPosePredictionFun = lasagne_nn.get_prediction_fun_nondeterministic(network) # posePredictionsSamples = [] # cosAzsPredSamples = [] # sinAzsPredSamples = [] # cosElevsPredSamples = [] # sinElevsPredSamples = [] # for i in range(100): # posePredictionsSample = np.zeros([len(grayTestImages), 4]) # for start_idx in range(0, len(grayTestImages), nnBatchSize): # posePredictionsSample[start_idx:start_idx + nnBatchSize] = nonDetPosePredictionFun(grayTestImages.astype(np.float32)[start_idx:start_idx + nnBatchSize]) # # cosAzsPredSample = posePredictionsSample[:,0] # sinAzsPredSample = posePredictionsSample[:,1] # cosAzsPredSamples = cosAzsPredSamples + [cosAzsPredSample[:,None]] # sinAzsPredSamples = sinAzsPredSamples + [sinAzsPredSample[:,None]] # # cosElevsPredSample = posePredictionsSample[:,2] # sinElevsPredSample = posePredictionsSample[:,3] # cosElevsPredSamples = cosElevsPredSamples + [cosElevsPredSample[:,None]] # sinElevsPredSamples = sinElevsPredSamples + [sinElevsPredSample[:,None]] # # cosAzsPredSamples = np.hstack(cosAzsPredSamples) # sinAzsPredSamples = np.hstack(sinAzsPredSamples) # # cosElevsPredSamples = np.hstack(cosElevsPredSamples) # sinElevsPredSamples = np.hstack(sinElevsPredSamples) # # azsPredictions = np.arctan2(sinAzsPredSamples, cosAzsPredSamples) # elevsPredictions = np.arctan2(sinElevsPredSamples, cosElevsPredSamples) # # np.save(trainModelsDirPose + 'azsPredictions.npy', azsPredictions) # np.save(trainModelsDirPose + 'elevsPredictions.npy', elevsPredictions) # ##Get predictions with dropout on to get samples. # with open(trainModelsDirPose + 'neuralNetModelPose.pickle', 'rb') as pfile: # neuralNetModelPose = pickle.load(pfile) # else: elevsPred = np.load(trainModelsDirPose + 'elevsPred.npy')[rangeTests] azsPred = np.load(trainModelsDirPose + 'azsPred.npy')[rangeTests] # azsPredictions = np.load(trainModelsDirPose + 'azsPredictions.npy')[rangeTests] azsPredictions = None # elevsPredictions = np.load(trainModelsDirPose + 'elevsPredictions.npy')[rangeTests] elevsPredictions = None if recomputePredictions or not os.path.isfile(trainModelsDirVColor + "vColorsPred.npy"): if 'neuralNetVColors' in parameterRecognitionModels: import theano # theano.sandbox.cuda.use('cpu') import lasagne import lasagne_nn nnModel = "" with open(experimentDir + 'neuralNetModelAppearance.pickle', 'rb') as pfile: neuralNetModelAppearance = pickle.load(pfile) meanImage = neuralNetModelAppearance['mean'] modelType = neuralNetModelAppearance['type'] param_values = neuralNetModelAppearance['params'] testImages = images.reshape([images.shape[0],3,images.shape[1],images.shape[2]]) - meanImage.reshape([1,meanImage.shape[2], meanImage.shape[0],meanImage.shape[1]]).astype(np.float32) network = lasagne_nn.load_network(modelType=modelType, param_values=param_values) appPredictionFun = lasagne_nn.get_prediction_fun(network) appPredictions = np.zeros([len(testSet), 3]) for start_idx in range(0, len(testSet), nnBatchSize): appPredictions[start_idx:start_idx + nnBatchSize] = appPredictionFun(testImages.astype(np.float32)[start_idx:start_idx + nnBatchSize]) vColorsPred = appPredictions np.save(trainModelsDirPose + 'vColorsPred.npy', vColorsPred) if 'randForestVColors' in parameterRecognitionModels: with open(trainModelsDirVColor + 'randForestModelVColor.pickle', 'rb') as pfile: randForestModelVColor = pickle.load(pfile) colorWindow = 30 image = images[0] croppedImages = images[:,image.shape[0]/2-colorWindow:image.shape[0]/2+colorWindow,image.shape[1]/2-colorWindow:image.shape[1]/2+colorWindow,:] vColorsPred = recognition_models.testRandomForest(randForestModelVColor, croppedImages.reshape([len(testSet),-1])) if 'linearRegressionVColors' in parameterRecognitionModels: with open(trainModelsDirVColor + 'linearRegressionModelVColor.pickle', 'rb') as pfile: linearRegressionModelVColor = pickle.load(pfile) colorWindow = 30 image = images[0] croppedImages = images[:,image.shape[0]/2-colorWindow:image.shape[0]/2+colorWindow,image.shape[1]/2-colorWindow:image.shape[1]/2+colorWindow,:] vColorsPred = recognition_models.testLinearRegression(linearRegressionModelVColor, croppedImages.reshape([len(testSet),-1])) if 'medianVColors' in parameterRecognitionModels: # recognition_models.medianColor(image, win) colorWindow = 30 imagesWin = images[:,images.shape[1]/2-colorWindow:images.shape[1]/2+colorWindow,images.shape[2]/2-colorWindow:images.shape[2]/2+colorWindow,:] vColorsPred = np.median(imagesWin.reshape([images.shape[0],-1,3]), axis=1)/1.4 # return color else: vColorsPred = np.load(trainModelsDirVColor + 'vColorsPred.npy')[rangeTests] SHModel = "" import theano import theano.tensor as T ## Theano NN error function 1. # import lasagne_nn # import lasagne # # with open(trainModelsDirPose + 'neuralNetModelAppearance.pickle', 'rb') as pfile: # neuralNetModelPose = pickle.load(pfile) # # meanImage = neuralNetModelPose['mean'].reshape([150,150,3]) # # modelType = neuralNetModelPose['type'] # param_values = neuralNetModelPose['params'] # network = lasagne_nn.load_network(modelType=modelType, param_values=param_values) # layer = lasagne.layers.get_all_layers(network)[6] # inputLayer = lasagne.layers.get_all_layers(network)[0] # layer_output = lasagne.layers.get_output(layer, deterministic=True) # dim_output= layer.output_shape[1] # # networkGT = lasagne_nn.load_network(modelType=modelType, param_values=param_values) # layerGT = lasagne.layers.get_all_layers(networkGT)[6] # inputLayerGT = lasagne.layers.get_all_layers(networkGT)[0] # layer_outputGT = lasagne.layers.get_output(layerGT, deterministic=True) # # rendererGray = 0.3*renderer[:,:,0] + 0.59*renderer[:,:,1] + 0.11*renderer[:,:,2] # rendererGrayGT = 0.3*rendererGT[:,:,0] + 0.59*rendererGT[:,:,1] + 0.11*rendererGT[:,:,2] # # chThError = TheanoFunOnOpenDR(theano_input=inputLayer.input_var, theano_output=layer_output, opendr_input=renderer - meanImage, dim_output = dim_output, # theano_input_gt=inputLayerGT.input_var, theano_output_gt=layer_outputGT, opendr_input_gt=rendererGT - meanImage) # # chThError.compileFunctions(layer_output, theano_input=inputLayer.input_var, dim_output=dim_output, theano_input_gt=inputLayerGT.input_var, theano_output_gt=layer_outputGT) # # chThError.r # ## Theano NN error function finite differences. # with open(trainModelsDirPose + 'neuralNetModelAzimuthTriplet.pickle', 'rb') as pfile: # neuralNetModelPose = pickle.load(pfile) #meanImage = neuralNetModelPose['mean'].reshape([150,150]) # modelType = neuralNetModelPose['type'] # param_values = neuralNetModelPose['params'] # # network = lasagne_nn.load_network(modelType=modelType, param_values=param_values, imgSize=75) # # # layer = lasagne.layers.get_all_layers(network)[-2] # inputLayer = lasagne.layers.get_all_layers(network)[0] # layer_output = lasagne.layers.get_output(network, deterministic=True) # dim_output= network.output_shape[1] # # networkGT = lasagne_nn.load_network(modelType=modelType, param_values=param_values, imgSize=75) # # layerGT = lasagne.layers.get_all_layers(networkGT)[-2] # inputLayerGT = lasagne.layers.get_all_layers(networkGT)[0] # # layer_outputGT = lasagne.layers.get_output(layerGT, deterministic=True) # layer_outputGT = lasagne.layers.get_output(networkGT, deterministic=True) # # rendererGray = 0.3*renderer[:,:,0] + 0.59*renderer[:,:,1] + 0.11*renderer[:,:,2] # rendererGrayGT = 0.3*rendererGT[:,:,0] + 0.59*rendererGT[:,:,1] + 0.11*rendererGT[:,:,2] # chThError = TheanoFunFiniteDiff(theano_input=inputLayer.input_var, theano_output=layer_output, opendr_input=rendererGray, dim_output = dim_output, # theano_input_gt=inputLayerGT.input_var, theano_output_gt=layer_outputGT, opendr_input_gt=rendererGrayGT, imSize=75) # # chThError.compileFunctions(layer_output, theano_input=inputLayer.input_var, dim_output=dim_output, theano_input_gt=inputLayerGT.input_var, theano_output_gt=layer_outputGT) # # chThError.r if recomputePredictions or not os.path.isfile(trainModelsDirLightCoeffs + "relLightCoefficientsPred.npy"): if 'neuralNetModelSHLight' in parameterRecognitionModels: import theano # theano.sandbox.cuda.use('cpu') import lasagne import lasagne_nn nnModel = "" with open(experimentDir + 'neuralNetModelLight.pickle', 'rb') as pfile: neuralNetModelSHLight = pickle.load(pfile) meanImage = neuralNetModelSHLight['mean'] modelType = neuralNetModelSHLight['type'] param_values = neuralNetModelSHLight['params'] testImages = images.reshape([images.shape[0],3,images.shape[1],images.shape[2]]) - meanImage.reshape([1,meanImage.shape[2], meanImage.shape[0],meanImage.shape[1]]).astype(np.float32) network = lasagne_nn.load_network(modelType=modelType, param_values=param_values) lightPredictionFun = lasagne_nn.get_prediction_fun(network) lightPredictions = np.zeros([len(testImages), 9]) for start_idx in range(0, len(testImages), nnBatchSize): lightPredictions[start_idx:start_idx + nnBatchSize] = lightPredictionFun(testImages.astype(np.float32)[start_idx:start_idx + nnBatchSize]) relLightCoefficientsPred = lightPredictions np.save(trainModelsDirLightCoeffs + 'relLightCoefficientsPred.npy', relLightCoefficientsPred) else: relLightCoefficientsPred = np.load(trainModelsDirLightCoeffs + 'relLightCoefficientsPred.npy')[rangeTests] if recomputePredictions or not os.path.isfile(trainModelsDirShapeParams + "shapeParamsPred.npy"): if 'neuralNetModelShape' in parameterRecognitionModels and useShapeModel: import theano # theano.sandbox.cuda.use('cpu') import lasagne import lasagne_nn nnModel = "" with open(experimentDir + 'neuralNetModelShape.pickle', 'rb') as pfile: neuralNetModelSHLight = pickle.load(pfile) meanImage = neuralNetModelSHLight['mean'] modelType = neuralNetModelSHLight['type'] param_values = neuralNetModelSHLight['params'] grayTestImages = 0.3*images[:,:,:,0] + 0.59*images[:,:,:,1] + 0.11*images[:,:,:,2] grayTestImages = grayTestImages[:,None, :,:] grayTestImages = grayTestImages - meanImage.reshape([1,1, grayTestImages.shape[2],grayTestImages.shape[3]]) network = lasagne_nn.load_network(modelType=modelType, param_values=param_values) shapePredictionFun = lasagne_nn.get_prediction_fun(network) shapePredictions = np.zeros([len(grayTestImages), latentDim]) for start_idx in range(0, len(grayTestImages), nnBatchSize): shapePredictions[start_idx:start_idx + nnBatchSize] = shapePredictionFun(grayTestImages.astype(np.float32)[start_idx:start_idx + nnBatchSize]) shapeParamsPred = shapePredictions np.save(trainModelsDirShapeParams + 'shapeParamsPred.npy', shapeParamsPred) # #Samples: # shapeParamsPredSamples = [] # shapeParamsNonDetFun = lasagne_nn.get_prediction_fun_nondeterministic(network) # # for i in range(100): # shapeParamsPredictionsSample = np.zeros([len(grayTestImages), 10]) # for start_idx in range(0, len(grayTestImages), nnBatchSize): # shapeParamsPredictionsSample[start_idx:start_idx + nnBatchSize] = shapeParamsNonDetFun(grayTestImages.astype(np.float32)[start_idx:start_idx + nnBatchSize]) # # shapeParamsPredSamples = shapeParamsPredSamples + [shapeParamsPredictionsSample[:,:][:,:,None]] # # shapeParamsPredSamples = np.concatenate(shapeParamsPredSamples, axis=2) # # np.save(trainModelsDirShapeParams + 'shapeParamsPredSamples.npy', shapeParamsPredSamples) else: shapeParamsPred = np.load(trainModelsDirShapeParams + 'shapeParamsPred.npy')[rangeTests] # shapeParamsPredSamples = np.load(trainModelsDirShapeParams + 'shapeParamsPredSamples.npy')[rangeTests] # if recomputePredictions or not os.path.isfile(trainModelsDirShapeParams + "neuralNetModelMaskLarge.npy"): # if 'neuralNetModelMask' in parameterRecognitionModels: # # import theano # # theano.sandbox.cuda.use('cpu') # import lasagne # import lasagne_nn # # # nnModel = "" # with open(trainModelsDirLightCoeffs + 'neuralNetModelMaskLarge.pickle', 'rb') as pfile: # neuralNetModelSHLight = pickle.load(pfile) # # meanImage = neuralNetModelSHLight['mean'] # modelType = neuralNetModelSHLight['type'] # param_values = neuralNetModelSHLight['params'] # # testImages = images.reshape([images.shape[0],3,images.shape[1],images.shape[2]]) - meanImage.reshape([1,meanImage.shape[2], meanImage.shape[0],meanImage.shape[1]]).astype(np.float32) # # network = lasagne_nn.load_network(modelType=modelType, param_values=param_values) # maskPredictionFun = lasagne_nn.get_prediction_fun(network) # # maskPredictions = np.zeros([len(testImages), 50*50]) # for start_idx in range(0, len(testImages), nnBatchSize): # maskPredictions[start_idx:start_idx + nnBatchSize] = maskPredictionFun(testImages.astype(np.float32)[start_idx:start_idx + nnBatchSize]) # # maskPredictions = np.reshape(maskPredictions, [len(testImages), 50,50]) # # np.save(trainModelsDirShapeParams + 'maskPredictions.npy', maskPredictions) # # # # #Samples: # # maskSamples = [] # # maskPredNonDetFun = lasagne_nn.get_prediction_fun_nondeterministic(network) # # # # for i in range(100): # # maskPredictionsSamples = np.zeros([len(testImages), 50*50]) # # for start_idx in range(0, len(testImages),nnBatchSize): # # maskPredictionsSamples[start_idx:start_idx + nnBatchSize] = maskPredNonDetFun(testImages.astype(np.float32)[start_idx:start_idx + nnBatchSize]) # # # # maskSamples = maskSamples + [maskPredictionsSamples[:,:][:,:,None]] # # # # maskSamples = np.concatenate(maskSamples, axis=2) # # loadMask = True # # # # gtDirMask = 'groundtruth/train4_occlusion_mask/' # # # # masksDir = gtDirMask + 'masks_occlusion/' # # if loadMask: # # masksGT = loadMasks(masksDir, testSet) # else: # maskPredictions = np.load(trainModelsDirShapeParams + 'maskPredictions.npy')[rangeTests] loadMask = False if loadMask: masksGT = loadMasks(gtDir + '/masks_occlusion/', testSet) print("Finished loading and compiling recognition models") envMapDic = {} SHFilename = 'data/LightSHCoefficients.pickle' with open(SHFilename, 'rb') as pfile: envMapDic = pickle.load(pfile) hdritems = list(envMapDic.items())[:] analyzeSamples = False # def analyzeHue(figurePath, rendererGT, renderer, sampleEl, sampleAz, sampleSH, sampleVColorsPredictions=None, sampleStds=0.1): # global stds # global chAz # global test_i # global samplingMode # # plt.ioff() # fig = plt.figure() # # stds[:] = sampleStds # negLikModel = -ch.sum(generative_models.LogGaussianModel(renderer=renderer, groundtruth=rendererGT, variances=variances)) # negLikModelRobust = -ch.sum(generative_models.LogRobustModel(renderer=renderer, groundtruth=rendererGT, foregroundPrior=globalPrior, variances=variances)) # stds = ch.Ch([initialPixelStdev]) # # models = [negLikModel, negLikModelRobust, negLikModelRobust] # errorFunRobust = models[1] # errorFunGaussian = models[0] # # vColorsPredSamplesHSV = cv2.cvtColor(np.uint8(sampleVColorsPredictions.reshape([1, 100, 3])*255), cv2.COLOR_RGB2HSV)[0,:,0] # # plt.hist(vColorsPredSamplesHSV, bins=30, alpha=0.2) # # hueGT = cv2.cvtColor(np.uint8(testVColorGT[test_i][None,None,:]*255), cv2.COLOR_RGB2HSV)[0,0,0] # huePred = cv2.cvtColor(np.uint8(vColorsPred[test_i][None,None,:]*255), cv2.COLOR_RGB2HSV)[0,0,0] # # chAz[:] = sampleAz # chEl[:] = sampleEl # currentVColors = chVColors.r # currentHSV = cv2.cvtColor(np.uint8(currentVColors[None,None,:]*255), cv2.COLOR_RGB2HSV)[0,0] # # chLightSHCoeffs[:] = sampleSH # # trainingTeapots = [0] # hueRange = np.arange(0,255,5) # # chThErrors = np.zeros([len(trainingTeapots), len(hueRange)]) # # robustErrors = np.array([]) # gaussianErrors = np.array([]) # hues = np.array([]) # for hue_i, hue in enumerate(hueRange): # hues = np.append(hues, hue) # # color = cv2.cvtColor(np.array([hue, currentHSV[1],currentHSV[2]])[None,None,:].astype(np.uint8), cv2.COLOR_HSV2RGB)/255 # chVColors[:] = color # # for idx, renderer_idx in enumerate(trainingTeapots): # renderer_i = renderer_teapots[renderer_idx] # rendererGray = 0.3*renderer_i[:,:,0] + 0.59*renderer_i[:,:,1] + 0.11*renderer_i[:,:,2] # # chThError.opendr_input = rendererGray # # chThErrors[idx, az_i] = chThError.r # # robustErrors = np.append(robustErrors, errorFunRobust.r) # gaussianErrors = np.append(gaussianErrors, errorFunGaussian.r) # # x1,x2,y1,y2 = plt.axis() # # robustErrors = robustErrors - np.min(robustErrors) # gaussianErrors = gaussianErrors - np.min(gaussianErrors) # # chThErrors = chThErrors - np.min(chThErrors) # plt.plot(hues, robustErrors*y2/np.max(robustErrors), c='brown') # plt.plot(hues, gaussianErrors*y2/np.max(gaussianErrors), c='purple') # # chThError.opendr_input = renderer # lineStyles = ['-', '--', '-.', ':'] # # # plt.axvline(np.mod(bestAzNormal*180/np.pi,360), linewidth=2, c='purple') # # plt.axvline(np.mod(bestAzRobust*180/np.pi,360), linewidth=2, c='brown') # # plt.axvline(bestHue, linewidth=2, c='y') # plt.axvline(hueGT, linewidth=2,c='g') # plt.axvline(huePred, linewidth=2,c='r') # # # plt.axvline(np.mod(currentAz*180/np.pi, 360), linewidth=2, linestyle='--',c='b') # # plt.xlabel('Sample') # plt.ylabel('Angular error') # # plt.axis((0,255,y1,y2)) # plt.title('Neuralnet multiple predictions') # fig.savefig(figurePath + 'sample' + '.png', bbox_inches='tight') # plt.close(fig) # # chVColors[:] = currentVColors # cv2.imwrite(figurePath + '_render.png', cv2.cvtColor(np.uint8(lin2srgb(renderer.r.copy())*255), cv2.COLOR_RGB2BGR)) # # def analyzeAz(figurePath, rendererGT, renderer, sampleEl, sampleVColor, sampleSH, sampleAzsPredictions=None, sampleStds=0.1): # global stds # global chAz # global test_i # global samplingMode # # plt.ioff() # fig = plt.figure() # # stds[:] = sampleStds # negLikModel = -ch.sum(generative_models.LogGaussianModel(renderer=renderer, groundtruth=rendererGT, variances=variances)) # negLikModelRobust = -ch.sum(generative_models.LogRobustModel(renderer=renderer, groundtruth=rendererGT, foregroundPrior=globalPrior, variances=variances)) # models = [negLikModel, negLikModelRobust, negLikModelRobust] # errorFunRobust = models[1] # errorFunGaussian = models[0] # # plt.hist(np.mod(sampleAzsPredictions*180/np.pi,360), bins=30, alpha=0.2) # # bestAz = testAzsRel[test_i] # # currentAz = chAz.r.copy() # chEl[:] = sampleEl # chVColors[:] = sampleVColor # chLightSHCoeffs[:] = sampleSH # # trainingTeapots = [0,14,20,25,26,1] # trainingTeapots = [0] # azRange = np.arange(0,2*np.pi,5*np.pi/180) # chThErrors = np.zeros([len(trainingTeapots), len(azRange)]) # # robustErrors = np.array([]) # gaussianErrors = np.array([]) # angles = np.array([]) # for az_i, az in enumerate(azRange): # angles = np.append(angles, az*180/np.pi) # chAz[:] = az # # for idx, renderer_idx in enumerate(trainingTeapots): # # renderer_i = renderer_teapots[renderer_idx] # # rendererGray = 0.3*renderer_i[:,:,0] + 0.59*renderer_i[:,:,1] + 0.11*renderer_i[:,:,2] # # chThError.opendr_input = rendererGray # chThErrors[0, az_i] = chThError.r # # robustErrors = np.append(robustErrors, errorFunRobust.r) # gaussianErrors = np.append(gaussianErrors, errorFunGaussian.r) # # x1,x2,y1,y2 = plt.axis() # # robustErrors = robustErrors - np.min(robustErrors) # gaussianErrors = gaussianErrors - np.min(gaussianErrors) # chThErrors = chThErrors - np.min(chThErrors) # plt.plot(angles, robustErrors*y2/np.max(robustErrors), c='brown') # plt.plot(angles, gaussianErrors*y2/np.max(gaussianErrors), c='purple') # # # chThError.opendr_input = renderer # lineStyles = ['-', '--', '-.', ':'] # for renderer_idx in range(len(trainingTeapots)): # plt.plot(angles, chThErrors[renderer_idx]*y2/np.max(chThErrors[renderer_idx]), linestyle=lineStyles[np.mod(renderer_idx,4)], c='y') # # if len(trainingTeapots) > 1: # prodErrors = np.prod(chThErrors, axis=0) # plt.plot(angles, prodErrors*y2/np.max(prodErrors), linestyle='-', c='black') # meanErrors = np.mean(chThErrors, axis=0) # plt.plot(angles, meanErrors*y2/np.max(meanErrors), linestyle='--', c='black') # # plt.plot(angles, gaussianErrors*robustErrors*y2/np.max(gaussianErrors*robustErrors), linestyle='--', c='black') # # # plt.axvline(np.mod(bestAzNormal*180/np.pi,360), linewidth=2, c='purple') # # plt.axvline(np.mod(bestAzRobust*180/np.pi,360), linewidth=2, c='brown') # plt.axvline(np.mod(bestAz*180/np.pi,360), linewidth=2, c='y') # plt.axvline(testAzsRel[test_i]*180/np.pi, linewidth=2,c='g') # plt.axvline(np.mod(azsPred[test_i]*180/np.pi, 360), linewidth=2,c='r') # # plt.axvline(np.mod(currentAz*180/np.pi, 360), linewidth=2, linestyle='--',c='b') # # plt.xlabel('Sample') # plt.ylabel('Angular error') # # # if samplingMode == False: # # # # scaleAzSamples = np.array(errorFunAzSamples) # # scaleAzSamples = scaleAzSamples - np.min(scaleAzSamples) + 1 # # scaleAzSamples = scaleAzSamples*0.25*y2/np.max(scaleAzSamples) # # for azSample_i, azSample in enumerate(scaleAzSamples): # # plt.plot(np.mod(totalAzSamples[azSample_i]*180/np.pi, 360), azSample, marker='o', ms=20., c='r') # # # # scaleAzSamples = np.array(errorFunGaussianAzSamples) # # scaleAzSamples = scaleAzSamples - np.min(scaleAzSamples) + 1 # # scaleAzSamples = scaleAzSamples*0.4*y2/np.max(scaleAzSamples) # # for azSample_i, azSample in enumerate(scaleAzSamples): # # plt.plot(np.mod(totalAzSamples[azSample_i]*180/np.pi, 360), azSample, marker='o', ms=20., c='g') # # # # scaleAzSamples = np.array(errorFunAzSamplesPred) # # scaleAzSamples = scaleAzSamples - np.min(scaleAzSamples) + 1 # # scaleAzSamples = scaleAzSamples*0.65*y2/np.max(scaleAzSamples) # # for azSample_i, azSample in enumerate(scaleAzSamples): # # plt.plot(np.mod(totalAzSamples[azSample_i]*180/np.pi, 360), azSample, marker='o', ms=20., c='b') # # # # scaleAzSamples = np.array(errorFunGaussianAzSamplesPred) # # scaleAzSamples = scaleAzSamples - np.min(scaleAzSamples) + 1 # # scaleAzSamples = scaleAzSamples*0.75*y2/np.max(scaleAzSamples) # # for azSample_i, azSample in enumerate(scaleAzSamples): # # plt.plot(np.mod(totalAzSamples[azSample_i]*180/np.pi, 360), azSample, marker='o', ms=20., c='y') # # plt.axis((0,360,y1,y2)) # plt.title('Neuralnet multiple predictions') # fig.savefig(figurePath + 'sample' + '.png', bbox_inches='tight') # plt.close(fig) # # chAz[:] = currentAz # cv2.imwrite(figurePath + '_render.png', cv2.cvtColor(np.uint8(lin2srgb(renderer.r.copy())*255), cv2.COLOR_RGB2BGR)) #Fit: print("Fitting predictions") print("Using " + modelsDescr[model]) errorFun = models[model] pixelErrorFun = pixelModels[model] testSamples = 1 if recognitionType == 2: testSamples = numSamples predSamples = 50 chDisplacement[:] = np.array([0.0, 0.0,0.0]) chScale[:] = np.array([1.0,1.0,1.0]) chObjAz[:] = 0 shapeIm = [height, width] #Update all error functions with the right renderers. print("Using " + modelsDescr[model]) negLikModel = -ch.sum(generative_models.LogGaussianModel(renderer=renderer, groundtruth=rendererGT, variances=variances))/ numPixels negLikModelRobust = -ch.sum(generative_models.LogRobustModel(renderer=renderer, groundtruth=rendererGT, foregroundPrior=globalPrior, variances=variances))/ numPixels pixelLikelihoodCh = generative_models.LogGaussianModel(renderer=renderer, groundtruth=rendererGT, variances=variances) pixelLikelihoodRobustCh = generative_models.LogRobustModel(renderer=renderer, groundtruth=rendererGT, foregroundPrior=globalPrior, variances=variances) # pixelLikelihoodRobustSQErrorCh = generative_models.LogRobustSQErrorModel(sqeRenderer, foregroundPrior=globalPrior, variances=variances) post = generative_models.layerPosteriorsRobustCh(rendererGT, renderer, np.array([]), 'FULL', globalPrior, variances)[0] # postSqerror = generative_models.layerPosteriorsRobustSQErrorCh(sqeRenderer, np.array([]), 'MASK', globalPrior, variances)[0].r>0.5 # negLikModelRobustSQError = generative_models.NLLRobustSQErrorModel(sqeRenderer=sqeRenderer, Q=globalPrior.r*np.ones([height, width]),variances=variances) / numPixels models = [negLikModel, negLikModelRobust] pixelModels = [pixelLikelihoodCh, pixelLikelihoodRobustCh] modelsDescr = ["Gaussian Model", "Outlier model" ] errorFun = models[model] testRangeStr = str(testSet[0]) + '-' + str(testSet[-1]) testDescription = 'photorealistic-replicateopendr_server-' + testRangeStr testPrefix = experimentPrefix + '_' + testDescription + '_' + optimizationTypeDescr[optimizationType] + '_' + str(len(testSet)) + 'samples_' testPrefixBase = testPrefix runExp = True shapePenaltyTests = [0,0,0,0] # shapePenaltyTests = [0,0,0,0] stdsTests = [0.03] # stdsTests = [0.03] modelTests = len(stdsTests)*[1] # modelTests = [1] methodTests = len(stdsTests)*[1] maxOptIters = len(stdsTests)*[40] if makeVideo: plt.ioff() import matplotlib.animation as animation Writer = animation.writers['ffmpeg'] writer = Writer(fps=1, metadata=dict(title='Fitting process', artist=''), bitrate=1800) figvid, ((vax1, vax2, vax3, vax4), (vax5, vax6, vax7, vax8)) = plt.subplots(2, 4, figsize=(12, 5)) figvid.delaxes(vax8) vax1.axes.get_xaxis().set_visible(False) vax1.axes.get_yaxis().set_visible(False) vax1.set_title("Ground truth") vax2.axes.get_xaxis().set_visible(False) vax2.axes.get_yaxis().set_visible(False) vax2.set_title("Recognition") vax3.axes.get_xaxis().set_visible(False) vax3.axes.get_yaxis().set_visible(False) vax3.set_title("Fit") vax4.axes.get_xaxis().set_visible(False) vax4.axes.get_yaxis().set_visible(False) vax4.set_title("Posterior") vax5.axes.get_xaxis().set_visible(False) vax5.axes.get_yaxis().set_visible(False) vax5.set_title("Env Map GT") vax6.axes.get_xaxis().set_visible(False) vax6.axes.get_yaxis().set_visible(False) vax6.set_title("Env Map Recognition") vax7.axes.get_xaxis().set_visible(False) vax7.axes.get_yaxis().set_visible(False) vax7.set_title("Env Map Fit") plt.tight_layout() vidImgs = [] nearestNeighbours = False if nearestNeighbours: trainImages = readImages(imagesDir, trainIds, loadFromHdf5) trainImagesR = trainImages.reshape([len(trainImages), -1]) methodsPred = ["Mean Baseline", "Nearest Neighbours", "Recognition", "Fit" ] plotColors = ['k', 'm', 'b', 'r'] segmentVColorError = np.array([]) useSegmentation = True segmentVColorsList = [] global annot_t annot_t = None def cb(_): global t global samplingMode elapsed_time = time.time() - t print("Ended interation in " + str(elapsed_time)) # if samplingMode: # analyzeAz(resultDir + 'az_samples/test' + str(test_i) +'/azNum' + str(sampleAzNum) + '_it' + str(iterat) , rendererGT, renderer, chEl.r, chVColors.r, chLightSHCoeffs.r, azsPredictions[test_i], sampleStds=stds.r) # else: # analyzeAz(resultDir + 'az_samples/test' + str(test_i) +'/min_azNum' + str(sampleAzNum) + '_it' + str(iterat) , rendererGT, renderer, chEl.r, chVColors.r, chLightSHCoeffs.r, azsPredictions[test_i], sampleStds=stds.r) global pixelErrorFun global errorFun global iterat iterat = iterat + 1 print("Callback! " + str(iterat)) print("Sq Error: " + str(errorFun.r)) if chThError is not None: print("Theano Error: " + str(chThError.r)) global negLikModelRobustSmallStd global bestShapeParamsSmallStd global bestRobustSmallStdError # if minimizingShape and useShapeModel: # # if negLikModelRobustSmallStd.r < bestRobustSmallStdError: # # bestRobustSmallStdError = negLikModelRobustSmallStd.r.copy() # # bestShapeParamsSmallStd = chShapeParams.r.copy() # maxShapeSize = 2.5 # largeShapeParams = np.abs(chShapeParams.r) > maxShapeSize # if np.any(largeShapeParams): # print("Warning: found large shape parameters to fix!") # chShapeParams[largeShapeParams] = np.sign(chShapeParams.r[largeShapeParams])*maxShapeSize if getColorFromCRF: global chVColors global Q global color segmentation = np.argmax(Q.r, axis=0).reshape(renderer.r.shape[:2]) if np.sum(segmentation == 0) == 0: vColor = color else: segmentRegion = segmentation == 0 vColor = np.median(rendererGT.reshape([-1, 3])[segmentRegion.ravel()], axis=0) * 1.4 vColor = vColor / max(np.max(vColor), 1.) vis_im = np.array(renderer.indices_image == 1).copy().astype(np.bool) chVColors[:] = vColor colorRegion = np.all(renderer.r != 0,axis=2).ravel() * segmentRegion.ravel() * vis_im.ravel() vColor = vColor * np.mean( rendererGT.r.reshape([-1, 3])[colorRegion] / renderer.r.reshape([-1, 3])[colorRegion]) chVColors[:] = vColor if makeVideo: global im1 global im2 global vidImgs global writer global writer_i global rendererRecognition plt.figure(figvid.number) im1 = vax1.imshow(lin2srgb(rendererGT.r.copy())) bbox_props = dict(boxstyle="round", fc="w", ec="0.5", alpha=0.8) im2 = vax2.imshow(lin2srgb(rendererRecognition.copy())) im3 = vax3.imshow(lin2srgb(renderer.r.copy())) stdsOld = stds.r stds[:] = 0.05 vis_im = np.array(renderer.indices_image==1).copy().astype(np.bool) post = generative_models.layerPosteriorsRobustCh(rendererGT, renderer, vis_im, 'MASK', globalPrior, variances)[0].r>0.5 stds[:] = stdsOld im4 = vax4.imshow(post.copy()) # plt.colorbar(im4, ax=vax4, use_gridspec=True) pEnvMap = SHProjection(envMapTexture, np.concatenate([chLightSHCoeffs.r[:,None], chLightSHCoeffs.r[:,None], chLightSHCoeffs.r[:,None]], axis=1)) approxProjectionFitted = np.sum(pEnvMap, axis=(2,3)) # approxProjectionFitted[approxProjectionFitted<0] = 0 # # approxProjectionGT[approxProjectionGT<0] = 0 # approxProjectionPred[approxProjectionPred<0] = 0 # cv2.imwrite(resultDir + 'approxProjectionGT.jpeg' , 255*np.concatenate([approxProjectionGT[...,None], approxProjectionGT[...,None], approxProjectionGT[...,None]], axis=2)[:,:,[2,1,0]]) # cv2.imwrite(resultDir + 'approxProjectionPred.jpeg' , 255*np.concatenate([approxProjectionPred[...,None], approxProjectionPred[...,None], approxProjectionPred[...,None]], axis=2)[:,:,[2,1,0]]) # cv2.imwrite(resultDir + 'approxProjectionFitted.jpeg' , 255*np.concatenate([approxProjectionFitted[...,None], approxProjectionFitted[...,None], approxProjectionFitted[...,None]], axis=2)[:,:,[2,1,0]]) cv2.imwrite(resultDir + 'approxProjectionGT.jpeg' , 255*np.sum(pEnvMapGT, axis=3)[:,:,[2,1,0]]) cv2.imwrite(resultDir + 'approxProjectionPred.jpeg' , 255*np.sum(pEnvMapPred, axis=3)[:,:,[2,1,0]]) cv2.imwrite(resultDir + 'approxProjectionFitted.jpeg' , 255*np.sum(pEnvMap, axis=3)[:,:,[2,1,0]]) approxProjectionGTlocal = skimage.io.imread(resultDir +'approxProjectionGT.jpeg').astype(np.float32)/255. approxProjectionPredlocal = skimage.io.imread(resultDir +'approxProjectionPred.jpeg').astype(np.float32)/255. approxProjectionFittedlocal = skimage.io.imread(resultDir +'approxProjectionFitted.jpeg').astype(np.float32)/255. approxProjectionGTlocal = skimage.transform.resize(approxProjectionGTlocal, [75,150]) approxProjectionPredlocal = skimage.transform.resize(approxProjectionPredlocal, [75,150]) approxProjectionFittedlocal = skimage.transform.resize(approxProjectionFittedlocal, [75,150]) im5 = vax5.imshow(approxProjectionGTlocal.copy()) im6 = vax6.imshow(approxProjectionPredlocal.copy()) im7 = vax7.imshow(approxProjectionFittedlocal.copy()) # if annot_t is not None: # annot_t.remove() annot_t = vax3.annotate("Fitting iter: " + str(iterat), xy=(1, 0), xycoords='axes fraction', fontsize=16, xytext=(-20, 5), textcoords='offset points', ha='right', va='bottom', bbox=bbox_props) plt.tight_layout() vidImgs.append([im1,im2, im3,im4, im5, im6, im7, annot_t]) if iterat == 1: vidImgs.append([im1,im2, im3,im4, im5, im6, im7, annot_t]) vidImgs.append([im1,im2, im3,im4, im5, im6, im7, annot_t]) # figvid.savefig(resultDir + 'videos/' + 'lastFig.png') # writer_i.grab_frame() global imagegt global gradAz global gradEl global performance global azimuths global elevations global shapeParams t = time.time() replaceExisting = False for testSetting, model in enumerate(modelTests): model = modelTests[testSetting] method = methodTests[testSetting] shapePenalty = shapePenaltyTests[testSetting] stds[:] = stdsTests[testSetting] testPrefix = testPrefixBase + '_method' + str(method) + 'errorFun' + str(model) + '_std' + str(stds.r[0]) + '_shapePen'+ str(shapePenalty) resultDir = 'results/' + testPrefix + '/' if not os.path.exists(resultDir + 'imgs/'): os.makedirs(resultDir + 'results/') if not os.path.exists(resultDir + 'imgs/'): os.makedirs(resultDir + 'imgs/') if not os.path.exists(resultDir + 'imgs/samples/'): os.makedirs(resultDir + 'imgs/samples/') ## NN individual prediction samples analysis. if not os.path.exists(resultDir + 'nn_samples/'): os.makedirs(resultDir + 'nn_samples/') if not os.path.exists(resultDir + 'az_samples/'): os.makedirs(resultDir + 'az_samples/') if not os.path.exists(resultDir + 'hue_samples/'): os.makedirs(resultDir + 'hue_samples/') if makeVideo: if not os.path.exists(resultDir + 'videos/'): os.makedirs(resultDir + 'videos/') azimuths = [] elevations = [] vColors = [] lightCoeffs = [] approxProjections = [] likelihoods = [] shapeParams = [] posteriors = [] startTime = time.time() samplingMode = False fittedVColorsList = [] fittedRelLightCoeffsList = [] approxProjectionsFittedList = [] approxProjectionsGTList = [] approxProjectionsPredList = [] fittedAzs = np.array([]) fittedElevs = np.array([]) fittedRelLightCoeffs = [] fittedShapeParamsList = [] fittedVColors = [] fittedPosteriorsList = [] revertedSamples = np.array([]) errorFunAzSamples = [] errorFunAzSamplesPred = [] errorFunGaussianAzSamplesPred = [] errorFunGaussianAzSamples = [] if nearestNeighbours: azimuthNearestNeighboursList = [] elevationNearestNeighboursList = [] vColorNearestNeighboursList = [] shapeParamsNearestNeighboursList = [] lightCoeffsNearestNeighboursList = [] nearestNeighboursErrorFuns = np.array([]) approxProjectionsNearestNeighbourList = [] nearestNeighboursPosteriorsList = [] errorsShapeVerticesSoFar = np.array([]) predictedErrorFuns = np.array([]) predictedPosteriorsList = [] if includeMeanBaseline: meanBaselineErrorFuns = np.array([]) meanBaselinePosteriorList = [] errorsFittedShapeVertices = np.array([]) fittedErrorFuns = np.array([]) fittedShapeParams = np.array([]) if (computePredErrorFuns and optimizationType == 0) or optimizationType != 0: for test_i in range(len(testAzsRel)): print("************** Minimizing loss of prediction " + str(test_i) + "of " + str(len(testAzsRel))) resultDir = 'results/' + testPrefix + '/results/' testDir = resultDir + str(test_i) + '/' if not os.path.exists(testDir): os.makedirs(testDir) if not replaceExisting and os.path.isfile(testDir + 'fitted_' + 'shapeParams'+ 'npy.npy'): continue bestFittedAz = chAz.r bestFittedEl = chEl.r bestModelLik = np.finfo('f').max bestVColors = chVColors.r bestLightSHCoeffs = chLightSHCoeffs.r if useShapeModel: bestShapeParams = chShapeParams.r testId = testIds[test_i] image = skimage.transform.resize(images[test_i], [height,width]) imageSrgb = image.copy() rendererGT[:] = srgb2lin(image) # sqeRenderer.imageGT = ch.Ch(image) negLikModel = -ch.sum(generative_models.LogGaussianModel(renderer=renderer, groundtruth=rendererGT, variances=variances))/numPixels # negLikModelRobust = -ch.sum(generative_models.LogRobustModel(renderer=renderer, groundtruth=rendererGT, foregroundPrior=globalPrior, variances=variances))/numPixels # negLikModelRobust = generative_models.NLLRobustModel(renderer=renderer, groundtruth=rendererGT, Q=globalPrior.r*np.ones([height, width]),variances=variances) / numPixels negLikModelRobust = -ch.sum(generative_models.LogRobustModel(renderer=renderer, groundtruth=rendererGT, foregroundPrior=globalPrior, variances=variances))/ numPixels # negLikModelRobustSQError = generative_models.NLLRobustSQErrorModel(sqeRenderer=sqeRenderer, Q=globalPrior.r*np.ones([height, width]), # variances=variances) / numPixels models = [negLikModel, negLikModelRobust] stds[:] = stdsTests[testSetting] if makeVideo: writer_i = Writer(fps=1, metadata=dict(title='', artist=''), bitrate=1800) writer_i.setup(figvid, resultDir + 'videos/vid_'+ str(test_i) + '.mp4', dpi=70) vidImgs = [] if nearestNeighbours: onenn_i = one_nn(trainImagesR, imageSrgb.ravel()) nearesTrainImage = trainImages[onenn_i] cv2.imwrite(resultDir + 'imgs/test'+ str(test_i) + '/id' + str(testId) +'_nearestneighbour' + '.png', cv2.cvtColor(np.uint8(nearesTrainImage[0].copy()*255), cv2.COLOR_RGB2BGR)) azimuthNearestNeighbour = trainAzsRel[onenn_i] elevationNearestNeighbour = trainElevsGT[onenn_i] vColorNearestNeighbour = trainVColorGT[onenn_i] if useShapeModel: shapeParamsNearestNeighbour = trainShapeModelCoeffsGT[onenn_i] lightCoeffsNearestNeighbour = trainLightCoefficientsGTRel[onenn_i] azimuthNearestNeighboursList = azimuthNearestNeighboursList + [azimuthNearestNeighbour] elevationNearestNeighboursList = elevationNearestNeighboursList + [elevationNearestNeighbour] vColorNearestNeighboursList = vColorNearestNeighboursList + [vColorNearestNeighbour] shapeParamsNearestNeighboursList = shapeParamsNearestNeighboursList + [shapeParamsNearestNeighbour] lightCoeffsNearestNeighboursList = lightCoeffsNearestNeighboursList + [lightCoeffsNearestNeighbour] chAz[:] = azimuthNearestNeighbour chEl[:] = elevationNearestNeighbour chVColors[:] = vColorNearestNeighbour chLightSHCoeffs[:] = lightCoeffsNearestNeighbour if useShapeModel: chShapeParams[:] = shapeParamsNearestNeighbour nearestNeighboursErrorFuns = np.append(nearestNeighboursErrorFuns, errorFun.r) # nearestNeighboursPosteriorsList = nearestNeighboursPosteriorsList + [np.array(renderer.indices_image==1).copy().astype(np.bool)] if includeMeanBaseline: chAz[:] = meanTrainAzimuthRel[test_i] chEl[:] = meanTrainElevation[test_i] chVColors[:] = meanTrainVColors[test_i] chLightSHCoeffs[:] = meanTrainLightCoefficientsGTRel[test_i] if useShapeModel: chShapeParams[:] = meanTrainShapeParams[test_i] # meanBaselinePosteriorList = meanBaselinePosteriorList + [np.array(renderer.indices_image==1).copy().astype(np.bool)[None,:]] meanBaselineErrorFuns = np.append(meanBaselineErrorFuns, errorFun.r) stdsSmall = ch.Ch([0.01]) variancesSmall = stdsSmall ** 2 negLikModelRobustSmallStd = -ch.sum(generative_models.LogRobustModel(renderer=renderer, groundtruth=rendererGT, foregroundPrior=globalPrior, variances=variancesSmall))/numPixels if not os.path.exists(resultDir + 'imgs/test'+ str(test_i) + '/'): os.makedirs(resultDir + 'imgs/test'+ str(test_i) + '/') if not os.path.exists(resultDir + 'imgs/test'+ str(test_i) + '/crf/'): os.makedirs(resultDir + 'imgs/test'+ str(test_i) + '/crf/') if not os.path.exists(resultDir + 'imgs/crf/'): os.makedirs(resultDir + 'imgs/crf/') if not os.path.exists(resultDir + 'imgs/test'+ str(test_i) + '/SH/'): os.makedirs(resultDir + 'imgs/test'+ str(test_i) + '/SH/') cv2.imwrite(resultDir + 'imgs/test'+ str(test_i) + '/id' + str(testId) +'_groundtruth' + '.png', cv2.cvtColor(np.uint8(lin2srgb(rendererGT.r.copy())*255), cv2.COLOR_RGB2BGR)) for sample in range(testSamples): if recognitionType == 0: #Prediction from (near) ground truth. color = testVColorGT[test_i] + nearGTOffsetVColor az = testAzsRel[test_i] + nearGTOffsetRelAz el = testElevsGT[test_i] + nearGTOffsetEl lightCoefficientsRel = testLightCoefficientsGTRel[test_i] if useShapeModel: shapeParams = testShapeParamsGT[test_i] elif recognitionType == 1 or recognitionType == 2: #Recognition estimate: az = azsPred[test_i] el = min(max(elevsPred[test_i],radians(1)), np.pi/2-radians(1)) color = vColorsPred[test_i] lightCoefficientsRel = relLightCoefficientsPred[test_i] if useShapeModel: shapeParams = shapeParamsPred[test_i] chAz[:] = az chEl[:] = el chVColors[:] = color chLightSHCoeffs[:] = lightCoefficientsRel if useShapeModel: chShapeParams[:] = shapeParams rendererRecognition = renderer.r.copy() cv2.imwrite(resultDir + 'imgs/test' + str(test_i) + '/sample' + str(sample) + '_predicted' + '.png', cv2.cvtColor(np.uint8(lin2srgb(renderer.r.copy()) * 255), cv2.COLOR_RGB2BGR)) np.save(testDir + 'pred_' + 'az' + 'npy', az) np.save(testDir + 'pred_' + 'el'+ 'npy', el) np.save(testDir + 'pred_' + 'color'+ 'npy', color) np.save(testDir + 'pred_' + 'lightCoefficientsRel'+ 'npy', lightCoefficientsRel) np.save(testDir + 'pred_' + 'shapeParams'+ 'npy', shapeParams) if not os.path.exists(resultDir + 'az_samples/'): os.makedirs(resultDir + 'az_samples/') # # analyzeAz(resultDir + 'az_samples/test' + str(test_i), rendererGT, renderer, chEl.r, chVColors.r, chLightSHCoeffs.r, azsPredictions[test_i], sampleStds=stds.r) # Dense CRF prediction of labels: vis_im = np.array(renderer.indices_image==1).copy().astype(np.bool) bound_im = renderer.boundarybool_image.astype(np.bool) # if optimizationTypeDescr[optimizationType] != 'optimize': # import densecrf_model # # # # plt.imsave(resultDir + 'imgs/test'+ str(test_i) + '/crf/renderer', renderer.r) # # # sample_i = test_i # segmentation, Q = densecrf_model.crfInference(rendererGT.r, vis_im, bound_im, [0.75,0.25,0.01], resultDir + 'imgs/crf/Q_' + str(sample_i)) # if np.sum(segmentation==0) == 0: # segmentVColorsList = segmentVColorsList + [color] # else: # # segmentRegion = segmentation==0 # segmentColor = np.median(rendererGT.reshape([-1,3])[segmentRegion.ravel()], axis=0) # segmentVColorsList = segmentVColorsList + [segmentColor] # # optLightSHCoeffs = chLightSHCoeffs.r if evaluateWithGT: hdridx = dataEnvMaps[test_i] envMapTexture = np.zeros([360,180,3]) envMapFound = False for hdrFile, hdrValues in hdritems: if hdridx == hdrValues[0]: envMapCoeffsGT = hdrValues[1] envMapFilename = hdrFile try: envMapTexture = np.array(imageio.imread(envMapFilename))[:,:,0:3] except: envMapTexture = np.zeros([360,180,3]) envMapTexture = cv2.resize(src=envMapTexture, dsize=(360,180)) envMapFound = True break if not envMapFound: ipdb.set_trace() pEnvMapGT = SHProjection(envMapTexture, np.concatenate([testLightCoefficientsGTRel[test_i][:,None], testLightCoefficientsGTRel[test_i][:,None], testLightCoefficientsGTRel[test_i][:,None]], axis=1)) approxProjectionGT = np.sum(pEnvMapGT, axis=(2,3)) approxProjectionsGTList = approxProjectionsGTList + [approxProjectionGT[None,:]] cv2.imwrite(resultDir + 'imgs/test'+ str(test_i) + '/SH/' + str(hdridx) + '_GT.jpeg' , 255*np.sum(pEnvMapGT, axis=3)[:,:,[2,1,0]]) pEnvMapPred = SHProjection(envMapTexture, np.concatenate([relLightCoefficientsPred[test_i][:,None], relLightCoefficientsPred[test_i][:,None], relLightCoefficientsPred[test_i][:,None]], axis=1)) approxProjectionPred = np.sum(pEnvMapPred, axis=(2,3)) approxProjectionsPredList = approxProjectionsPredList + [approxProjectionPred[None,:]] predictedPosteriorsList = predictedPosteriorsList + [np.array(renderer.indices_image==1).copy().astype(np.bool)[None,:]] if nearestNeighbours: pEnvMap = SHProjection(envMapTexture, np.concatenate([lightCoeffsNearestNeighbour.ravel()[:,None], lightCoeffsNearestNeighbour.ravel()[:,None], lightCoeffsNearestNeighbour.ravel()[:,None]], axis=1)) approxProjectionNearestNeighbour = np.sum(pEnvMap, axis=(2,3)) approxProjectionsNearestNeighbourList = approxProjectionsNearestNeighbourList + [approxProjectionNearestNeighbour[None,:]] cv2.imwrite(resultDir + 'imgs/test'+ str(test_i) + '/SH/' + str(hdridx) + '_NearestNeighbour.jpeg' , 255*np.sum(pEnvMap, axis=3)[:,:,[2,1,0]]) cv2.imwrite(resultDir + 'imgs/test'+ str(test_i) + '/SH/' + str(hdridx) + '_Pred.jpeg' , 255*np.sum(pEnvMapPred, axis=3)[:,:,[2,1,0]]) ## RecognitionType =2 : Use samples from neural net to explore the space better. if recognitionType == 2: errorFunAzSamples = [] errorFunAzSamplesPred = [] errorFunGaussianAzSamplesPred = [] errorFunGaussianAzSamples = [] if not os.path.exists(resultDir + 'az_samples/test' + str(test_i) + '/'): os.makedirs(resultDir + 'az_samples/test' + str(test_i) + '/') if not os.path.exists(resultDir + 'hue_samples/test' + str(test_i) + '/'): os.makedirs(resultDir + 'hue_samples/test' + str(test_i) + '/') samplingMode = True cv2.imwrite(resultDir + 'az_samples/test' + str(test_i) + '_gt.png', cv2.cvtColor(np.uint8(lin2srgb(rendererGT.r.copy())*255), cv2.COLOR_RGB2BGR)) # stds[:] = 0.1 # # azSampleStdev = np.std(azsPredictions[test_i]) azSampleStdev = np.sqrt(-np.log(np.min([np.mean(sinAzsPredSamples[test_i])**2 + np.mean(cosAzsPredSamples[test_i])**2,1]))) predAz = chAz.r numSamples = max(int(np.ceil(azSampleStdev*180./(np.pi*25.))),1) azSamples = np.linspace(0, azSampleStdev, numSamples) totalAzSamples = predAz + np.concatenate([azSamples, -azSamples[1:]]) sampleAzNum = 0 # model = 1 # errorFun = models[model] bestPredAz = chAz.r # bestPredEl = chEl.r bestPredEl = min(max(chEl.r.copy(),radians(1)), np.pi/2-radians(1)) bestPredVColors = chVColors.r.copy() bestPredLightSHCoeffs = chLightSHCoeffs.r.copy() if useShapeModel: bestPredShapeParams = chShapeParams.r.copy() bestModelLik = np.finfo('f').max bestPredModelLik = np.finfo('f').max # analyzeAz(resultDir + 'az_samples/test' + str(test_i) +'/pre' , rendererGT, renderer, chEl.r, chVColors.r, chLightSHCoeffs.r, azsPredictions[test_i], sampleStds=stds.r) # analyzeHue(resultDir + 'hue_samples/test' + str(test_i) +'/pre', rendererGT, renderer, chEl.r, chAz.r, chLightSHCoeffs.r, vColorsPredSamples[test_i], sampleStds=stds.r) for sampleAz in totalAzSamples: iterat = 0 sampleAzNum += 1 chAz[:] = sampleAz # chEl[:] = elsample print("Minimizing first step") model = 1 errorFun = models[model] # method = 5 chLightSHCoeffs[:] = lightCoefficientsRel chVColors[:] = color #Todo test with adding chEl. free_variables = [chLightSHCoeffs] options={'disp':False, 'maxiter':2} errorFunAzSamplesPred = errorFunAzSamplesPred + [errorFun.r] errorFunGaussianAzSamplesPred = errorFunGaussianAzSamplesPred + [models[0].r] # analyzeAz(resultDir + 'az_samples/test' + str(test_i) +'/azNum' + str(sampleAzNum), rendererGT, renderer, chEl.r, chVColors.r, chLightSHCoeffs.r, azsPredictions[test_i], sampleStds=stds.r) if models[1].r.copy() < bestPredModelLik: print("Found best angle!") # bestPredModelLik = errorFun.r.copy() bestPredModelLik = models[1].r.copy() bestPredAz = sampleAz bestPredEl = min(max(chEl.r.copy(),radians(1)), np.pi/2-radians(1)) bestPredVColors = chVColors.r.copy() bestPredLightSHCoeffs = chLightSHCoeffs.r.copy() if useShapeModel: bestPredShapeParams = chShapeParams.r.copy() bestModelLik = errorFun.r.copy() # cv2.imwrite(resultDir + 'imgs/test'+ str(test_i) + '/best_predSample' + str(numPredSamples) + '.png', cv2.cvtColor(np.uint8(lin2srgb(renderer.r.copy())*255), cv2.COLOR_RGB2BGR)) errorFunAzSamples = errorFunAzSamples + [errorFun.r] errorFunGaussianAzSamples = errorFunGaussianAzSamples + [models[0].r] color = bestPredVColors lightCoefficientsRel = bestPredLightSHCoeffs az = bestPredAz if useShapeModel: shapeParams = bestShapeParams # previousAngles = np.vstack([previousAngles, np.array([[azsample, elsample],[chAz.r.copy(), chEl.r.copy()]])]) samplingMode = False chAz[:] = az chAz[:] = az chEl[:] = min(max(el,radians(1)), np.pi/2-radians(1)) chVColors[:] = color.copy() chLightSHCoeffs[:] = lightCoefficientsRel.copy() if useShapeModel: chShapeParams[:] = shapeParams.copy() cv2.imwrite(resultDir + 'imgs/test'+ str(test_i) + '/best_sample' + '.png', cv2.cvtColor(np.uint8(lin2srgb(renderer.r.copy())*255), cv2.COLOR_RGB2BGR)) predictedErrorFuns = np.append(predictedErrorFuns, errorFun.r) # analyzeAz(resultDir + 'az_samples/test' + str(test_i) + '_samples', rendererGT, renderer, chEl.r, color, lightCoefficientsRel, # azsPredictions[test_i], sampleStds=stds.r) iterat = 0 sampleAzNum = 0 sys.stdout.flush() ## Finally: Do Fitting! ignore = False reverted = False if optimizationTypeDescr[optimizationType] == 'optimize': #Get VCOLOR using CRF: print("** Minimizing from initial predicted parameters. **") globalPrior[:] = 0.9 errorFun = models[model] vColor = color if useCRFOcclusionPred: priorProbs = [0.75, 0.25, 0.01] test_sample = test_i Q = recognition_models.segmentCRFModel(renderer=renderer, groundtruth=rendererGT, priorProbs=priorProbs, resultDir=resultDir, test_i=test_sample) getColorFromCRF = False segmentation = np.argmax(Q.r, axis=0).reshape(renderer.r.shape[:2]) if np.sum(segmentation == 0) == 0: ignore = True vColor = color else: segmentRegion = segmentation == 0 vColor = np.median(rendererGT.reshape([-1, 3])[segmentRegion.ravel()], axis=0) * 1.4 vColor = vColor / max(np.max(vColor), 1.) vis_im = np.array(renderer.indices_image == 1).copy().astype(np.bool) chVColors[:] = vColor colorRegion = np.all(renderer.r != 0,axis=2).ravel() * segmentRegion.ravel() * vis_im.ravel() if colorRegion.sum() == 0: ignore = True else: vColor = vColor * np.median(rendererGT.r.reshape([-1, 3])[colorRegion] / renderer.r.reshape([-1, 3])[colorRegion]) if not ignore: color = vColor chVColors[:] = vColor ## Bayes Optimization ### Bayesian Optimization tests. if useCRFOcclusionPred: free_variables = [ chAz, chEl] free_variables_app_light = [ chLightSHCoeffs] azSampleStdev = np.sqrt(-np.log(np.min([np.mean(np.sin(azsPredictions)[test_i])**2 + np.mean(np.cos(azsPredictions)[test_i])**2,1]))) azMean = azsPred[test_i] elMean = elevsPred[test_i] elSampleStdev = np.sqrt(-np.log(np.min([np.mean(np.sin(elevsPredictions)[test_i])**2 + np.mean(np.cos(elevsPredictions)[test_i])**2,1]))) elBound = np.min(elSampleStdev, np.pi/4) elMean = min(max(elevsPred[test_i],radians(1)), np.pi/2-radians(1)) shapeStdevs = np.min(np.vstack([2*np.sqrt(np.cov(shapeParamsPredSamples[test_i]).diagonal()), np.ones([latentDim])*3.5]), 0) azBound = min(azSampleStdev, np.pi) optBounds = [(azMean - azBound, azMean + azBound), (min(max(elMean - elBound,0), np.pi/2), min(max(elMean + elBound,0), np.pi/2))] + [(max(shapeParams[shape_i] - shapeStdevs[shape_i],-3.5), min(shapeParams[shape_i] + shapeStdevs[shape_i],3.5)) for shape_i in range(len(shapeStdevs))] optBounds = [(azMean - azBound, azMean + azBound), (min(max(elMean - elBound,0), np.pi/2), min(max(elMean + elBound,0), np.pi/2))] numSamples = max(int(np.ceil(2*azSampleStdev*180./(np.pi*20.))),1) # azSamples = np.linspace(0, azSampleStdev, numSamples) azSamples = np.linspace(0, np.pi, 10) numSamples = max(int(np.ceil(2*elSampleStdev*180./(np.pi*10.))),1) elSamples = np.linspace(0, 2*elSampleStdev, numSamples) elSamples = chEl.r + np.concatenate([elSamples, -elSamples[1:]]) totalAzSamples = chAz.r + np.concatenate([azSamples, -azSamples[1:]]) if len(totalAzSamples) == 0: totalAzSamples = np.array([chAz.r]) if azSampleStdev < 20*np.pi/180: totalAzSamples = np.array([chAz.r]) totalElSamples = np.array([elSample for elSample in elSamples if elSample > 0 and elSample < np.pi/2]) if len(totalElSamples) == 0: totalElSamples = np.array([chEl.r]) totalElSamples = np.array([chEl.r]) # totalElSamples = np.array([testElevsGT[test_i]]) # chShapeParams[:] = testShapeParamsGT[test_i] # chLightSHCoeffs[:] = testLightCoefficientsGTRel[test_i] # chVColors[:] = testVColorGT[test_i] # vColor = testVColorGT[test_i] # lightCoefficientsRel = testLightCoefficientsGTRel[test_i] totalSamples = np.meshgrid(totalAzSamples, totalElSamples) totalSamples = np.hstack([totalSamples[0].reshape([-1,1]),totalSamples[1].reshape([-1,1])]) import diffrender_opt # objFun = diffrender_opt.opendrObjectiveFunction(errorFun, free_variables) # res = diffrender_opt.bayesOpt(objFun, objBounds) methodOpt = methods[method] stds[:] = 0.1 # crfObjFun = diffrender_opt.opendrObjectiveFunctionCRF(free_variables, rendererGT, renderer, vColor, chVColors, chLightSHCoeffs, lightCoefficientsRel, free_variables_app_light, resultDir, test_i, stds.r, methodOpt, False, False) cv2.imwrite( resultDir + 'imgs/test' + str(test_i) + '/sample' + str(sample) + '_predictedCRF' + '.png', cv2.cvtColor(np.uint8(lin2srgb(renderer.r.copy()) * 255), cv2.COLOR_RGB2BGR)) # errorFunCRF = -ch.sum(generative_models.LogCRFModel(renderer=renderer, groundtruth=rendererGT, Q=Q.r.reshape([3,height, width]), # variances=variances)) / numPixels errorFunNLLCRF = generative_models.NLLCRFModel(renderer=renderer, groundtruth=rendererGT, Q=Q.r.reshape([3,height, width]), variances=variances) / numPixels # errorFun = models[1] # samplesEvals = [] # for azSample in totalSamples[:,0]: # chAz[:] = azSample # samplesEvals = samplesEvals + [errorFunNLLCRF.r] # # samplesEvals = np.array(samplesEvals) # # # samplesEvals = crfObjFun(totalSamples) # # #Start Optimiaztion # # # res = diffrender_opt.bayesOpt(crfObjFun, totalSamples, samplesEvals, optBounds) # # # # optAz = res.x_opt[0] # # optEl = res.x_opt[1] # # # optShapeParams = res.x_opt[2:] # bestSample = np.argmin(samplesEvals) # optAz = totalSamples[bestSample][0] # optEl = totalSamples[bestSample][1] # # azfig = plt.figure() # ax = azfig.add_subplot(111) # ax.plot(np.mod(totalSamples[:,0], 2*np.pi), samplesEvals, 'ro') # y1, y2 = ax.get_ylim() # ax.vlines(testAzsRel[test_i], y1, y2, 'g', label = 'Groundtruth Az') # ax.vlines(np.mod(optAz, 2*np.pi), y1, y2, 'r', label = 'Optimum CRF Az') # ax.vlines(np.mod(az, 2*np.pi), y1, y2, 'b', label = 'Recognition Az') # ax.set_xlim((0, np.pi*2)) # legend = ax.legend() # azfig.savefig(resultDir + 'imgs/test' + str(test_i) + '/samplesAzPlot_' + str(int(azSampleStdev*180/np.pi)) + '.png') # plt.close(azfig) # # legend = ax.legend() # # chAz[:] = optAz # chEl[:] = optEl # #### Local search: # # cv2.imwrite( # resultDir + 'imgs/test' + str(test_i) + '/sample' + str(sample) + '_optimizedCRFPose' + '.png', # cv2.cvtColor(np.uint8(lin2srgb(renderer.r.copy()) * 255), cv2.COLOR_RGB2BGR)) free_variables = [chShapeParams, chAz, chEl, chLightSHCoeffs, chVColors] # # stds[:] = 0.03 # shapePenalty = 0.0001 stds[:] = stdsTests[testSetting] shapePenalty = 0.0001 options={'disp':False, 'maxiter':100} # options={'disp':False, 'maxiter':2} minimizingShape = True # method = 5 ch.minimize({'raw': errorFun }, bounds=None, method=methods[method], x0=free_variables, callback=cb, options=options) maxShapeSize = 4 largeShapeParams = np.abs(chShapeParams.r) > maxShapeSize if np.any(largeShapeParams) or chEl.r > np.pi/2 + radians(10) or chEl.r < -radians(15) or np.linalg.norm(chVertices.r - chVerticesMean) >= 3.5: print("Warning: found large shape parameters to fix!") reverted = True # chShapeParams[largeShapeParams] = np.sign(chShapeParams.r[largeShapeParams])*maxShapeSize cv2.imwrite(resultDir + 'imgs/test'+ str(test_i) + '/it1'+ '.png', cv2.cvtColor(np.uint8(lin2srgb(renderer.r.copy())*255), cv2.COLOR_RGB2BGR)) # if reverted: # cv2.imwrite(resultDir + 'imgs/test'+ str(test_i) + '/reverted1'+ '.png', cv2.cvtColor(np.uint8(lin2srgb(renderer.r.copy())*255), cv2.COLOR_RGB2BGR)) # reverted = False # chAz[:] = az # chEl[:] = min(max(el,radians(1)), np.pi/2-radians(1)) # chVColors[:] = color.copy() # chLightSHCoeffs[:] = lightCoefficientsRel.copy() # if useShapeModel: # chShapeParams[:] = shapeParams.copy() # # cv2.imwrite(resultDir + 'imgs/test'+ str(test_i) + '/it1'+ '.png', cv2.cvtColor(np.uint8(lin2srgb(renderer.r.copy())*255), cv2.COLOR_RGB2BGR)) # # free_variables = [chVColors, chLightSHCoeffs] # stds[:] = 0.01 # shapePenalty = 0.0 # options={'disp':False, 'maxiter':50} # # options={'disp':False, 'maxiter':2} # # free_variables = [chShapeParams ] # minimizingShape = False # getColorFromCRF = False # ch.minimize({'raw': errorFunNLLCRF }, bounds=None, method=methods[method], x0=free_variables, callback=cb, options=options) # largeShapeParams = np.abs(chShapeParams.r) > maxShapeSize # if np.any(largeShapeParams) or chEl.r > np.pi/2 + radians(10) or chEl.r < -radians(15) or np.linalg.norm(chVertices.r - chVerticesMean) >= 3.5: # print("Warning: found large shape parameters to fix!") # reverted = True # cv2.imwrite(resultDir + 'imgs/test'+ str(test_i) + '/reverted2'+ '.png', cv2.cvtColor(np.uint8(lin2srgb(renderer.r.copy())*255), cv2.COLOR_RGB2BGR)) # # cv2.imwrite(resultDir + 'imgs/test'+ str(test_i) + '/it2'+ '.png', cv2.cvtColor(np.uint8(lin2srgb(renderer.r.copy())*255), cv2.COLOR_RGB2BGR)) else: ## *** If not using CRF *** free_variables = [chShapeParams, chAz, chEl, chLightSHCoeffs, chVColors] # # shapePenalty = 0.0001 stds[:] = stdsTests[testSetting] shapePenalty = shapePenaltyTests[testSetting] options = {'disp': False, 'maxiter': maxOptIters[testSetting]} # options={'disp':False, 'maxiter':2} minimizingShape = True errorFun = models[1] # errorFunFast = generative_models.NLLRobustModel(renderer=renderer, groundtruth=rendererGT, Q=globalPrior.r*np.ones([height, width]), # variances=variances) / numPixels # # plt.imsave('errorscolors1.png', sqeRenderer.render_image) # # chVColors[:] = testVColorGT[test_i] + nearGTOffsetVColor # chAz[:]= testAzsRel[test_i] + nearGTOffsetRelAz # chEl[:] = testElevsGT[test_i] + nearGTOffsetEl # chLightSHCoeffs[:] = testLightCoefficientsGTRel[test_i] # chShapeParams[:] = testShapeParamsGT[test_i] # # chLightSHCoeffs[:] = 0 # # chLightSHCoeffs[2] = 0.5 # # chLightSHCoeffs[0] = 1 # plt.imsave('render.png', renderer.r) # # imageGT = renderer.r.copy() # # chEl[:] = chEl[:].r + 0.2 # # SQError = (rendererGT.r - renderer)**2 # # sqeRenderer.imageGT = ch.Ch(imageGT) # # plt.imsave('errors.png', sqeRenderer.r) # plt.imsave('errorscolors.png', sqeRenderer.render_image) # # plt.imsave('renderer2.png', sqeRenderer.render_dedx) # # plt.imsave('errorsdx.png', sqeRenderer.render_dedx, cmap=matplotlib.cm.coolwarm, vmin=-1, vmax=1) # drAz = np.sum(sqeRenderer.dr_wrt(chEl).toarray().reshape([height,width, 3]), axis=2) # # drAzErrorFun = -pixelLikelihoodRobustSQErrorCh.dr_wrt(chEl).toarray().reshape([height, width]) # # plt.ioff() # fig = plt.figure() # ax = fig.add_subplot(111, aspect='equal') # ims = ax.imshow(drAz, cmap=matplotlib.cm.coolwarm, vmin=-4, vmax=4) # fig.colorbar(ims) # fig.savefig('drAzfig.png', bbox_inches='tight') # plt.close(fig) # plt.ioff() # # fig = plt.figure() # ax = fig.add_subplot(111, aspect='equal') # ims = ax.imshow(drAzErrorFun, cmap=matplotlib.cm.coolwarm, vmin=-300, vmax=300) # fig.colorbar(ims) # fig.savefig('drAzfigErrFun.png', bbox_inches='tight') # plt.close(fig) # # # plt.imsave('errorsdr_wrtaz.png', drAz, cmap=matplotlib.cm.coolwarm, vmin=drAz.min(), vmax=drAz.max()) # # plt.imsave('SQerrorsdr_wrtaz.png', SQError.dr_wrt(chEl).toarray().reshape([height,width, 3]), cmap=matplotlib.cm.coolwarm, vmin=-4, vmax=4) # plt.imsave('errorsgt.png', sqeRenderer.imageGT.r) # # sys.exit() ch.minimize({'raw': errorFun}, bounds=None, method=methods[method], x0=free_variables, callback=cb, options=options) maxShapeSize = 4 largeShapeParams = np.abs(chShapeParams.r) > maxShapeSize if np.any(largeShapeParams) or chEl.r > np.pi / 2 + radians(10) or chEl.r < -radians(15) or np.linalg.norm( chVertices.r - chVerticesMean) >= 3.5: print("Warning: found large shape parameters to fix!") reverted = True # chShapeParams[largeShapeParams] = np.sign(chShapeParams.r[largeShapeParams])*maxShapeSize cv2.imwrite(resultDir + 'imgs/test' + str(test_i) + '/it1' + '.png', cv2.cvtColor(np.uint8(lin2srgb(renderer.r.copy()) * 255), cv2.COLOR_RGB2BGR)) # if reverted: # cv2.imwrite(resultDir + 'imgs/test'+ str(test_i) + '/reverted1'+ '.png', cv2.cvtColor(np.uint8(lin2srgb(renderer.r.copy())*255), cv2.COLOR_RGB2BGR)) # reverted = False # chAz[:] = az # chEl[:] = min(max(el,radians(1)), np.pi/2-radians(1)) # chVColors[:] = color.copy() # chLightSHCoeffs[:] = lightCoefficientsRel.copy() # if useShapeModel: # chShapeParams[:] = shapeParams.copy() # # # cv2.imwrite(resultDir + 'imgs/test'+ str(test_i) + '/it1'+ '.png', cv2.cvtColor(np.uint8(lin2srgb(renderer.r.copy())*255), cv2.COLOR_RGB2BGR)) # # free_variables = [chVColors] # stds[:] = 0.01 # shapePenalty = 0.0 # options = {'disp': False, 'maxiter': 50} # # options={'disp':False, 'maxiter':2} # # free_variables = [chShapeParams ] # minimizingShape = False # getColorFromCRF = False # ch.minimize({'raw': errorFun}, bounds=None, method=methods[method], x0=free_variables, callback=cb, options=options) # # largeShapeParams = np.abs(chShapeParams.r) > maxShapeSize # if np.any(largeShapeParams) or chEl.r > np.pi/2 + radians(10) or chEl.r < -radians(15) or np.linalg.norm(chVertices.r - chVerticesMean) >= 3.5: # print("Warning: found large shape parameters to fix!") # reverted = True # cv2.imwrite(resultDir + 'imgs/test'+ str(test_i) + '/reverted2'+ '.png', cv2.cvtColor(np.uint8(lin2srgb(renderer.r.copy())*255), cv2.COLOR_RGB2BGR)) # # cv2.imwrite(resultDir + 'imgs/test'+ str(test_i) + '/it2'+ '.png', cv2.cvtColor(np.uint8(lin2srgb(renderer.r.copy())*255), cv2.COLOR_RGB2BGR)) if not reverted: bestFittedAz = chAz.r.copy() bestFittedEl = min(max(chEl.r.copy(),radians(1)), np.pi/2-radians(1)) bestVColors = chVColors.r.copy() bestLightSHCoeffs = chLightSHCoeffs.r.copy() if useShapeModel: bestShapeParams = chShapeParams.r.copy() else: revertedSamples = np.append(revertedSamples, test_i) bestFittedAz = az bestFittedEl = min(max(el,radians(1)), np.pi/2-radians(1)) bestVColors = color.copy() bestLightSHCoeffs = lightCoefficientsRel.copy() if useShapeModel: bestShapeParams = shapeParams.copy() chAz[:] = bestFittedAz chEl[:] = min(max(bestFittedEl,radians(1)), np.pi/2-radians(1)) chVColors[:] = bestVColors chLightSHCoeffs[:] = bestLightSHCoeffs if useShapeModel: chShapeParams[:] = bestShapeParams if makeVideo: if len(vidImgs) > 0: im_ani = animation.ArtistAnimation(figvid, vidImgs, interval=2000, repeat_delay=5000, repeat=True, blit=False) im_ani.save(resultDir + 'videos/fitting_'+ str(testSet[test_i]) + '.mp4', fps=None, writer=writer, codec='mp4') vidImgs[-1][7].remove() writer_i.finish() cv2.imwrite(resultDir + 'imgs/test'+ str(test_i) + '/fitted'+ '.png',cv2.cvtColor(np.uint8(lin2srgb(renderer.r.copy())*255), cv2.COLOR_RGB2BGR)) np.save(testDir + 'fitted_' + 'az' + 'npy', chAz.r) np.save(testDir + 'fitted_' + 'el'+ 'npy', chEl.r) np.save(testDir + 'fitted_' + 'color'+ 'npy', chVColors.r) np.save(testDir + 'fitted_' + 'lightCoefficientsRel'+ 'npy', chLightSHCoeffs.r) np.save(testDir + 'fitted_' + 'shapeParams'+ 'npy', chShapeParams.r) if optimizationTypeDescr[optimizationType] != 'predict': if evaluateWithGT: fittedErrorFuns = np.append(fittedErrorFuns, bestModelLik) fittedAzs = np.append(fittedAzs, bestFittedAz) fittedElevs = np.append(fittedElevs, bestFittedEl) fittedVColorsList = fittedVColorsList + [bestVColors] fittedRelLightCoeffsList = fittedRelLightCoeffsList + [bestLightSHCoeffs] if useShapeModel: fittedShapeParamsList = fittedShapeParamsList + [bestShapeParams] pEnvMap = SHProjection(envMapTexture, np.concatenate([bestLightSHCoeffs[:,None], bestLightSHCoeffs[:,None], bestLightSHCoeffs[:,None]], axis=1)) approxProjectionFitted = np.sum(pEnvMap, axis=(2,3)) approxProjectionsFittedList = approxProjectionsFittedList + [approxProjectionFitted[None,:]] cv2.imwrite(resultDir + 'imgs/test'+ str(test_i) + '/SH/' + str(hdridx) + '_Fitted.jpeg' , 255*np.sum(pEnvMap, axis=3)[:,:,[2,1,0]]) #Best std for posterior recognition. stds[:] = 0.1 vis_im = np.array(renderer.indices_image==1).copy().astype(np.bool) post = generative_models.layerPosteriorsRobustCh(rendererGT, renderer, vis_im, 'MASK', globalPrior, variances)[0].r>0.5 # postSqerror = generative_models.layerPosteriorsRobustSQErrorCh(sqeRenderer, np.array([]), 'MASK', globalPrior, variances)[0].r>0.5 fittedPosteriorsList = fittedPosteriorsList + [post[None,:]] stds[:] = stdsTests[testSetting] plt.imsave(resultDir + 'imgs/test'+ str(test_i) + '/' + str(hdridx) + '_Outlier.jpeg', np.tile(post.reshape(shapeIm[0],shapeIm[1],1), [1,1,3]).astype(np.float32)) #Every now and then (or after the final test case), produce plots to keep track of work accross different levels of occlusion. experimentDic = {'model':model, 'method':method, 'shapePenalty':shapePenalty, 'stds':stds.r, 'dataIds': testIds, 'gtPrefix': gtPrefix, 'trainPrefixPose': trainPrefixPose, 'trainPrefixVColor': trainPrefixVColor, 'trainPrefixLightCoeffs': trainPrefixLightCoeffs, 'trainPrefixShapeParams': trainPrefixShapeParams, 'trainModelsDirAppLight': trainModelsDirAppLight, 'experimentDir': experimentDir, 'testSet': testSet, 'methodsPred': methodsPred, 'testOcclusions': testOcclusions, 'likelihoods': likelihoods, 'testPrefixBase': testPrefixBase, 'parameterRecognitionModels': parameterRecognitionModels, 'azimuths': azimuths, 'elevations': elevations, 'vColors': vColors, 'lightCoeffs': lightCoeffs, 'shapeParams': shapeParams, 'evaluateWithGT':evaluateWithGT} with open(resultDir + 'experiment.pickle', 'wb') as pfile: pickle.dump(experimentDic, pfile) if evaluateWithGT: if np.mod(test_i+1,100) == 0 or test_i + 1 >= len(testSet): if approxProjectionsPredList: approxProjectionsPred = np.vstack(approxProjectionsPredList) if approxProjectionsGTList: approxProjectionsGT = np.vstack(approxProjectionsGTList) if predictedPosteriorsList: predictedPosteriors = np.vstack(predictedPosteriorsList) if nearestNeighbours: approxProjectionsNearestNeighbours = np.vstack(approxProjectionsNearestNeighbourList) if optimizationTypeDescr[optimizationType] != 'predict': if fittedVColorsList: fittedVColors = np.vstack(fittedVColorsList) if fittedRelLightCoeffsList: fittedRelLightCoeffs = np.vstack(fittedRelLightCoeffsList) if approxProjectionsFittedList: approxProjectionsFitted = np.vstack(approxProjectionsFittedList) if fittedShapeParamsList and useShapeModel: fittedShapeParams = np.vstack(fittedShapeParamsList) if fittedPosteriorsList: fittedPosteriors = np.vstack(fittedPosteriorsList) if nearestNeighbours: azimuthNearestNeighbours = np.concatenate(azimuthNearestNeighboursList) elevationNearestNeighbours = np.concatenate(elevationNearestNeighboursList) vColorNearestNeighbours = np.vstack(vColorNearestNeighboursList) if useShapeModel: shapeParamsNearestNeighbours = np.vstack(shapeParamsNearestNeighboursList) lightCoeffsNearestNeighbours = np.vstack(lightCoeffsNearestNeighboursList) # nearestNeighboursPosteriors = np.vstack(nearestNeighboursPosteriorsList) if optimizationTypeDescr[optimizationType] != 'predict': numFitted = range(len(fittedAzs)) else: numFitted = range(test_i+1) if includeMeanBaseline: # meanBaselinePosteriors = np.vstack(meanBaselinePosteriorList) azimuths = [meanTrainAzimuthRel] elevations= [meanTrainElevation] vColors= [meanTrainVColors] lightCoeffs= [meanTrainLightCoefficientsGTRel] if useShapeModel: shapeParams = [meanTrainShapeParams] else: shapeParams = [None] approxProjections= [meanTrainEnvMapProjections] likelihoods = [meanBaselineErrorFuns] segmentations = [None] ipdb.set_trace() else: azimuths = [None] elevations= [None] vColors= [None] lightCoeffs= [None] approxProjections= [None] shapeParams = [None] likelihoods = [None] segmentations = [None] if nearestNeighbours: azimuths = azimuths + [azimuthNearestNeighbours] elevations = elevations + [elevationNearestNeighbours] vColors = vColors + [vColorNearestNeighbours] lightCoeffs = lightCoeffs + [lightCoeffsNearestNeighbours] approxProjections = approxProjections + [approxProjectionsNearestNeighbours] if useShapeModel: shapeParams = shapeParams + [shapeParamsNearestNeighbours] else: shapeParams = shapeParams + [None] segmentations = segmentations + [None] else: azimuths = azimuths + [None] elevations= elevations + [None] vColors= vColors + [None] lightCoeffs= lightCoeffs + [None] approxProjections = approxProjections + [None] likelihoods = likelihoods + [None] shapeParams = shapeParams + [None] segmentations = segmentations + [None] azimuths = azimuths + [azsPred] elevations = elevations + [elevsPred] vColors = vColors + [vColorsPred] lightCoeffs = lightCoeffs + [relLightCoefficientsPred] approxProjections = approxProjections + [approxProjectionsPred] likelihoods = likelihoods + [predictedErrorFuns] segmentations = segmentations + [predictedPosteriors] if useShapeModel: shapeParams = shapeParams + [shapeParamsPred] else: shapeParams = shapeParams + [None] if optimizationTypeDescr[optimizationType] != 'predict': azimuths = azimuths + [fittedAzs] elevations = elevations + [fittedElevs] vColors = vColors + [fittedVColors] lightCoeffs = lightCoeffs + [fittedRelLightCoeffs] approxProjections = approxProjections + [approxProjectionsFitted] if useShapeModel: shapeParams = shapeParams + [fittedShapeParams] else: shapeParams = shapeParams + [None] likelihoods = likelihoods + [fittedErrorFuns] segmentations = segmentations + [fittedPosteriors] else: azimuths = azimuths + [None] elevations = elevations + [None] vColors = vColors + [None] lightCoeffs = lightCoeffs + [None] approxProjections = approxProjections + [None] likelihoods = likelihoods + [None] segmentations = segmentations + [None] shapeParams = shapeParams + [None] errorsPosePredList, errorsLightCoeffsList, errorsShapeParamsList, errorsShapeVerticesList, errorsEnvMapList, errorsLightCoeffsCList, errorsVColorsEList, errorsVColorsCList, errorsVColorsSList, errorsSegmentationList \ = computeErrors(numFitted, azimuths, testAzsRel, elevations, testElevsGT, vColors, testVColorGT, lightCoeffs, testLightCoefficientsGTRel, approxProjections, approxProjectionsGT, shapeParams, testShapeParamsGT, useShapeModel, chShapeParams, chVertices, segmentations, masksGT) meanAbsErrAzsList, meanAbsErrElevsList, meanErrorsLightCoeffsList, meanErrorsShapeParamsList, meanErrorsShapeVerticesList, meanErrorsLightCoeffsCList, meanErrorsEnvMapList, meanErrorsVColorsEList, meanErrorsVColorsCList, meanErrorsVColorsCList, meanErrorsSegmentationList \ = computeErrorAverages(np.mean, numFitted, useShapeModel, errorsPosePredList, errorsLightCoeffsList, errorsShapeParamsList, errorsShapeVerticesList, errorsEnvMapList, errorsLightCoeffsCList, errorsVColorsEList, errorsVColorsCList, errorsVColorsSList, errorsSegmentationList) medianAbsErrAzsList, medianAbsErrElevsList,medianErrorsLightCoeffsList, medianErrorsShapeParamsList, medianErrorsShapeVerticesList, medianErrorsLightCoeffsCList, medianErrorsEnvMapList, medianErrorsVColorsEList, medianErrorsVColorsCList, medianErrorsVColorsCList, medianErrorsSegmentationList \ = computeErrorAverages(np.median, numFitted, useShapeModel, errorsPosePredList, errorsLightCoeffsList, errorsShapeParamsList, errorsShapeVerticesList, errorsEnvMapList, errorsLightCoeffsCList, errorsVColorsEList, errorsVColorsCList, errorsVColorsSList, errorsSegmentationList) #Write statistics to file. with open(resultDir + 'performance.txt', 'w') as expfile: for method_i in range(len(azimuths)): # expfile.write(str(z)) expfile.write("Mean Azimuth Error " + methodsPred[method_i] + " " + str(meanAbsErrAzsList) + '\n') expfile.write("Mean Elevation Error " + methodsPred[method_i] + " " + str(meanAbsErrElevsList[method_i])+ '\n') expfile.write("Mean SH Components Error " + methodsPred[method_i] + " " + str(meanErrorsLightCoeffsList[method_i])+ '\n') expfile.write("Mean SH Components Error " + methodsPred[method_i] + " " + str(meanErrorsLightCoeffsCList[method_i])+ '\n') expfile.write("Mean Vertex Colors Error E " + methodsPred[method_i] + " " + str(meanErrorsVColorsEList[method_i])+ '\n') expfile.write("Mean Vertex Colors Error C " + methodsPred[method_i] + " " + str(meanErrorsVColorsCList[method_i])+ '\n') expfile.write("Mean Vertex Colors Error S " + methodsPred[method_i] + " " + str(meanErrorsVColorsCList[method_i]) + '\n') expfile.write("Mean Shape Error " + methodsPred[method_i] + " " + str(meanErrorsShapeParamsList[method_i])+ '\n') expfile.write("Mean Shape Vertices Error " + methodsPred[method_i] + " " + str(meanErrorsShapeVerticesList[method_i])+ '\n') expfile.write("Mean Segmentation Error " + methodsPred[method_i] + " " + str(meanErrorsSegmentationList[method_i])+ '\n\n') #Write statistics to file. with open(resultDir + 'median-performance.txt', 'w') as expfile: for method_i in range(len(azimuths)): # expfile.write(str(z)) expfile.write("Median Azimuth Error " + methodsPred[method_i] + " " + str(medianAbsErrAzsList) + '\n') expfile.write("Median Elevation Error " + methodsPred[method_i] + " " + str(medianAbsErrElevsList[method_i])+ '\n') expfile.write("Median SH Components Error " + methodsPred[method_i] + " " + str(medianErrorsLightCoeffsList[method_i])+ '\n') expfile.write("Median SH Components Error " + methodsPred[method_i] + " " + str(medianErrorsLightCoeffsCList[method_i])+ '\n') expfile.write("Median Vertex Colors Error E " + methodsPred[method_i] + " " + str(medianErrorsVColorsEList[method_i])+ '\n') expfile.write("Median Vertex Colors Error C " + methodsPred[method_i] + " " + str(medianErrorsVColorsCList[method_i])+ '\n') expfile.write("Mean Vertex Colors Error S " + methodsPred[method_i] + " " + str(meanErrorsVColorsCList[method_i]) + '\n') expfile.write("Median Shape Error " + methodsPred[method_i] + " " + str(medianErrorsShapeParamsList[method_i])+ '\n') expfile.write("Median Shape Vertices Error " + methodsPred[method_i] + " " + str(medianErrorsShapeVerticesList[method_i])+ '\n') expfile.write("Median Segmentation Error " + methodsPred[method_i] + " " + str(medianErrorsSegmentationList[method_i])+ '\n\n') if not os.path.exists(resultDir + 'stats/'): os.makedirs(resultDir + 'stats/') with open(resultDir + 'stats/' + 'performance' + str(test_i) + '.txt', 'w') as expfile: for method_i in range(len(azimuths)): expfile.write("Mean Azimuth Error " + methodsPred[method_i] + " " + str(meanAbsErrAzsList) + '\n') expfile.write("Mean Elevation Error " + methodsPred[method_i] + " " + str(meanAbsErrElevsList[method_i])+ '\n') expfile.write("Mean SH Components Error " + methodsPred[method_i] + " " + str(meanErrorsLightCoeffsList[method_i])+ '\n') expfile.write("Mean SH Components Error " + methodsPred[method_i] + " " + str(meanErrorsLightCoeffsCList[method_i])+ '\n') expfile.write("Mean Vertex Colors Error E " + methodsPred[method_i] + " " + str(meanErrorsVColorsEList[method_i])+ '\n') expfile.write("Mean Vertex Colors Error C " + methodsPred[method_i] + " " + str(meanErrorsVColorsCList[method_i])+ '\n') expfile.write("Mean Shape Error " + methodsPred[method_i] + " " + str(meanErrorsShapeParamsList[method_i])+ '\n') expfile.write("Mean Shape Vertices Error " + methodsPred[method_i] + " " + str(meanErrorsShapeVerticesList[method_i])+ '\n') expfile.write("Mean Segmentation Error " + methodsPred[method_i] + " " + str(meanErrorsSegmentationList[method_i])+ '\n\n') segmentationsDic = {'segmentations':segmentations} with open(resultDir + 'segmentations.pickle', 'wb') as pfile: pickle.dump(segmentationsDic, pfile) np.save(resultDir + 'reverted.npy', revertedSamples) envMapDic = {'approxProjections':approxProjections, 'approxProjectionsGT':approxProjectionGT} with open(resultDir + 'approxProjections.pickle', 'wb') as pfile: pickle.dump(envMapDic, pfile) # ipdb.set_trace() totalTime = time.time() - startTime print("Took " + str(totalTime/test_i) + " time per instance.") experimentErrorsDic = {'errorsPosePredList':errorsPosePredList, 'errorsLightCoeffsList':errorsLightCoeffsList, 'errorsShapeParamsLis':errorsShapeParamsList, 'errorsShapeVerticesList':errorsShapeVerticesList, 'errorsEnvMapList':errorsEnvMapList, 'errorsLightCoeffsCList':errorsLightCoeffsCList, 'errorsVColorsEList':errorsVColorsEList, 'errorsVColorsCList':errorsVColorsCList, 'errorsVColorsCList':errorsVColorsCList, 'errorsSegmentationList':errorsSegmentationList} # with open(resultDir + 'experiment_errors.pickle', 'wb') as pfile: pickle.dump(experimentErrorsDic, pfile)
134,439
47.586917
485
py
inversegraphics
inversegraphics-master/export_groundtruth.py
import save_exr_images from save_exr_images import exportExrImages import os print ("Reading xml ") outputDir = '../data/output/' imgDir = outputDir + "images/" lines = [line.strip() for line in open(outputDir + 'groundtruth.txt')] if not os.path.exists(imgDir): os.makedirs(imgDir) for instance in lines: parts = instance.split(' ') teapot = int(parts[3]) frame = int(parts[4]) sceneNum = int(parts[5]) targetIndex = int(parts[6]) prefix = '' if len(parts) == 17: prefix = parts[16] try: exportExrImages(outputDir, imgDir, teapot, frame, sceneNum, targetIndex, prefix) except Exception as e: print(e)
678
19.575758
88
py
inversegraphics
inversegraphics-master/diffrender_groundtruth_multi.py
__author__ = 'pol' import matplotlib # matplotlib.use('Qt4Agg') import bpy import scene_io_utils import mathutils from math import radians import timeit import time import opendr import chumpy as ch import geometry import image_processing import pdb import numpy as np import cv2 from blender_utils import * import generative_models import matplotlib.pyplot as plt from opendr_utils import * import OpenGL.GL as GL import light_probes import imageio from OpenGL import contextdata from light_probes import SHProjection import collision import copy plt.ion() ######################################### # Initialization starts here ######################################### prefix = 'new_scenes' previousGTPrefix = '' #Main script options: renderFromPreviousGT = False useShapeModel = True useOpenDR = False useBlender = True renderBlender = True captureEnvMapFromBlender = False parseSceneInstantiations = False loadBlenderSceneFile = True useCycles = True unpackModelsFromBlender = False unpackSceneFromBlender = False loadSavedSH = False generateTriplets = False replaceNewGroundtruth = True renderOcclusions = False occlusionMin = 0.0 occlusionMax = 0.9 renderTeapots = True renderMugs = False showMug = False centeredObject = True fixedCamDistance = True useDirectionalLight = False renderBackground = True teapotSceneIndex = 0 mugSceneIndex = 0 if renderTeapots and renderMugs: mugSceneIndex = 1 glModes = ['glfw','mesa'] glMode = glModes[0] width, height = (64, 64) win = -1 coords = np.meshgrid(np.arange(width) - width / 2, np.arange(height) - height / 2) coordsMugX = np.array([0]) coordsMugY = np.array([0]) coordsTeapotX = np.array([0]) coordsTeapotY = np.array([0]) if useOpenDR: if glMode == 'glfw': import glfw #Initialize base GLFW context for the Demo and to share context among all renderers. glfw.init() glfw.window_hint(glfw.CONTEXT_VERSION_MAJOR, 3) glfw.window_hint(glfw.CONTEXT_VERSION_MINOR, 3) glfw.window_hint(glfw.OPENGL_FORWARD_COMPAT, GL.GL_TRUE) glfw.window_hint(glfw.OPENGL_PROFILE, glfw.OPENGL_CORE_PROFILE) glfw.window_hint(glfw.DEPTH_BITS,32) glfw.window_hint(glfw.VISIBLE, GL.GL_FALSE) # win = glfw.create_window(width, height, "Demo", None, None) # glfw.make_context_current(win) angle = 60 * 180 / numpy.pi clip_start = 0.05 clip_end = 10 frustum = {'near': clip_start, 'far': clip_end, 'width': width, 'height': height} camDistance = 0.4 teapotsFile = 'teapots.txt' teapots = [line.strip() for line in open(teapotsFile)] renderTeapotsList = np.arange(len(teapots))[0:1] mugs = [] renderMugsList = np.array([]) if renderMugs: mugs = [line.strip() for line in open('mugs.txt')] renderMugsList = np.arange(len(mugs))[0:1] sceneIdx = 0 replaceableScenesFile = 'data/scene_replaceables_backup_new.txt' sceneNumber, sceneFileName, instances, roomName, roomInstanceNum, targetIndices, targetPositions = scene_io_utils.getSceneInformation(sceneIdx, replaceableScenesFile) sceneDicFile = 'data/scene' + str(sceneNumber) + '.pickle' targetParentIdx = 0 targetIndex = targetIndices[targetParentIdx] targetParentPosition = targetPositions[targetParentIdx] targetPosition = targetParentPosition tex_srgb2lin = True v, f_list, vc, vn, uv, haveTextures_list, textures_list = scene_io_utils.loadSavedScene(sceneDicFile, tex_srgb2lin) removeObjectData(int(targetIndex), v, f_list, vc, vn, uv, haveTextures_list, textures_list) targetModels = [] mugModels = [] blender_teapots = [] blender_mugs = [] if useBlender: selection = [ teapots[i] for i in renderTeapotsList] scene_io_utils.loadTargetsBlendData() for teapotIdx, teapotName in enumerate(selection): teapot = bpy.data.scenes[teapotName[0:63]].objects['teapotInstance' + str(renderTeapotsList[teapotIdx])] teapot.layers[1] = True teapot.layers[2] = True targetModels = targetModels + [teapot] blender_teapots = blender_teapots + [teapot] if renderMugs: selectionMugs = [ mugs[i] for i in renderMugsList] scene_io_utils.loadMugsBlendData() for mugIdx, mugName in enumerate(selectionMugs): mug = bpy.data.scenes[mugName[0:63]].objects['mugInstance' + str(renderMugsList[mugIdx])] mug.layers[1] = True mug.layers[2] = True mugModels = mugModels + [mug] blender_mugs = blender_mugs + [mug] v_teapots, f_list_teapots, vc_teapots, vn_teapots, uv_teapots, haveTextures_list_teapots, textures_list_teapots, vflat, varray, center_teapots = scene_io_utils.loadTeapotsOpenDRData(renderTeapotsList, useBlender, unpackModelsFromBlender, targetModels) if renderMugs: v_mugs, f_list_mugs, vc_mugs, vn_mugs, uv_mugs, haveTextures_list_mugs, textures_list_mugs, vflat_mugs, varray_mugs, center_mugs = scene_io_utils.loadMugsOpenDRData(renderMugsList, useBlender, unpackModelsFromBlender, mugModels) v_mug = v_mugs[0][0] f_list_mug = f_list_mugs[0][0] chVColorsMug = ch.Ch([1,0,0]) vc_mug = [chVColorsMug * np.ones(v_mug[0].shape)] vn_mug = vn_mugs[0][0] uv_mug = uv_mugs[0][0] haveTextures_list_mug = haveTextures_list_mugs[0][0] textures_list_mug = textures_list_mugs[0][0] chObjAz = ch.Ch([0]) chDist = ch.Ch([camDistance]) chObjAzGT = ch.Ch([0]) chAzGT = ch.Ch([0]) chElGT = ch.Ch([0]) chAzRelGT = chAzGT - chObjAzGT chDistGT = ch.Ch([camDistance]) chComponentGT = ch.Ch(np.array([2, 0.25, 0.25, 0.12,-0.17,0.36,0.1,0.,0.])) chObjDistGT = ch.Ch([0]) chObjRotationGT = ch.Ch([0]) chObjAzMug = ch.Ch([0]) chObjDistMug = ch.Ch([0]) chObjRotationMug = ch.Ch([0]) light_colorGT = ch.ones(3) chVColorsGT = ch.Ch([0.8,0.8,0.8]) shCoefficientsFile = 'data/sceneSH' + str(sceneIdx) + '.pickle' chAmbientIntensityGT = ch.Ch([0.1]) clampedCosCoeffs = clampedCosineCoefficients() chAmbientSHGT = ch.zeros([9]) envMapDic = {} SHFilename = 'data/LightSHCoefficients.pickle' with open(SHFilename, 'rb') as pfile: envMapDic = pickle.load(pfile) phiOffset = ch.Ch([0]) totalOffset = phiOffset + chObjAzGT envMapCoeffs = ch.Ch(list(envMapDic.items())[0][1][1]) chLightRadGT = ch.Ch([0.1]) chLightDistGT = ch.Ch([0.5]) chLightIntensityGT = ch.Ch([1]) chLightAzGT = ch.Ch([0]) chLightElGT = ch.Ch([0]) chGlobalConstantGT = ch.Ch([0.5]) angleGT = ch.arcsin(chLightRadGT/chLightDistGT) zGT = chZonalHarmonics(angleGT) shDirLightGT = chZonalToSphericalHarmonics(zGT, np.pi/2 - chLightElGT, chLightAzGT - np.pi/2) envMapCoeffsRotated = ch.Ch(np.dot(light_probes.chSphericalHarmonicsZRotation(totalOffset), envMapCoeffs[[0,3,2,1,4,5,6,7,8]])[[0,3,2,1,4,5,6,7,8]]) envMapCoeffsRotatedRel = ch.Ch(np.dot(light_probes.chSphericalHarmonicsZRotation(phiOffset), envMapCoeffs[[0,3,2,1,4,5,6,7,8]])[[0,3,2,1,4,5,6,7,8]]) shCoeffsRGB = envMapCoeffsRotated shCoeffsRGBRel = envMapCoeffsRotatedRel chShCoeffs = 0.3*shCoeffsRGB[:,0] + 0.59*shCoeffsRGB[:,1] + 0.11*shCoeffsRGB[:,2] chShCoeffsRel = 0.3*shCoeffsRGBRel[:,0] + 0.59*shCoeffsRGBRel[:,1] + 0.11*shCoeffsRGBRel[:,2] chAmbientSHGT = chShCoeffs.ravel() * chAmbientIntensityGT * clampedCosCoeffs chAmbientSHGTRel = chShCoeffsRel.ravel() * chAmbientIntensityGT * clampedCosCoeffs chComponentGT = chAmbientSHGT chComponentGTRel = chAmbientSHGTRel chDisplacementGT = ch.Ch([0.0,0.0,0.0]) chScaleGT = ch.Ch([1, 1.,1.]) currentTeapotModel = 0 currentMugModel = 0 if renderTeapots: addObjectData(v, f_list, vc, vn, uv, haveTextures_list, textures_list, v_teapots[currentTeapotModel][0], f_list_teapots[currentTeapotModel][0], vc_teapots[currentTeapotModel][0], vn_teapots[currentTeapotModel][0], uv_teapots[currentTeapotModel][0], haveTextures_list_teapots[currentTeapotModel][0], textures_list_teapots[currentTeapotModel][0]) if renderMugs: addObjectData(v, f_list, vc, vn, uv, haveTextures_list, textures_list, v_mugs[currentMugModel][0], f_list_mugs[currentMugModel][0], vc_mugs[currentMugModel][0], vn_mugs[currentMugModel][0], uv_mugs[currentMugModel][0], haveTextures_list_mugs[currentMugModel][0], textures_list_mugs[currentMugModel][0]) center = center_teapots[currentTeapotModel] if useOpenDR: rendererGT = createRendererGT(glMode, chAzGT, chElGT, chDistGT, center, v, vc, f_list, vn, light_colorGT, chComponentGT, chVColorsGT, targetPosition[:].copy(), chDisplacementGT, width,height, uv, haveTextures_list, textures_list, frustum, None ) rendererGT.overdraw = False rendererGT.nsamples = 1 rendererGT.msaa = False rendererGT.initGL() rendererGT.initGLTexture() vis_gt = np.array(rendererGT.indices_image!=1).copy().astype(np.bool) vis_mask = np.array(rendererGT.indices_image==1).copy().astype(np.bool) shapeIm = vis_gt.shape numPixels = height * width ## For blender parallell rendering. import multiprocessing numTileAxis = np.ceil(np.sqrt(multiprocessing.cpu_count())/2) numTileAxis = 3 smCenterGT = ch.array([0,0,0.1]) shapeVerticesScaling = 0.09 teapotFilePath = 'data/teapotModel.pkl' if useShapeModel: teapot_i = -1 import shape_model #%% Load data teapotModel = shape_model.loadObject(teapotFilePath) faces = teapotModel['faces'] #%% Sample random shape Params latentDim = np.shape(teapotModel['ppcaW'])[1] shapeParams = np.random.randn(latentDim) chShapeParamsGT = ch.Ch(shapeParams) meshLinearTransform=teapotModel['meshLinearTransform'] W=teapotModel['ppcaW'] b=teapotModel['ppcaB'] chVerticesGT = shape_model.VerticesModel(chShapeParams=chShapeParamsGT,meshLinearTransform=meshLinearTransform,W = W,b=b) chVerticesGT.init() chVerticesGT = ch.dot(geometry.RotateZ(-np.pi/2)[0:3,0:3],chVerticesGT.T).T chNormalsGT = shape_model.chGetNormals(chVerticesGT, faces) smNormalsGT = [chNormalsGT] smFacesGT = [[faces]] smVColorsGT = [chVColorsGT*np.ones(chVerticesGT.shape)] smUVsGT = [ch.Ch(np.zeros([chVerticesGT.shape[0],2]))] smHaveTexturesGT = [[False]] smTexturesListGT = [[None]] chVerticesGT = chVerticesGT - ch.mean(chVerticesGT, axis=0) minZ = ch.min(chVerticesGT[:,2]) chMinZ = ch.min(chVerticesGT[:,2]) zeroZVerts = chVerticesGT[:,2]- chMinZ chVerticesGT = ch.hstack([chVerticesGT[:,0:2] , zeroZVerts.reshape([-1,1])]) chVerticesGT = chVerticesGT*shapeVerticesScaling smCenterGT = ch.array([0,0,0.1]) smVerticesGT = [chVerticesGT] chNormalsGT = shape_model.chGetNormals(chVerticesGT, faces) smNormalsGT = [chNormalsGT] else: latentDim = 1 chShapeParamsGT = ch.array([0]) if useShapeModel: teapotMesh = createMeshFromData('teapotShapeModelMesh', chVerticesGT.r.tolist(), faces.astype(np.int32).tolist()) teapotMesh.layers[0] = True teapotMesh.layers[1] = True teapotMesh.pass_index = 1 targetGroup = bpy.data.groups.new('teapotShapeModelGroup') targetGroup.objects.link(teapotMesh) teapot = bpy.data.objects.new('teapotShapeModel', None) teapot.dupli_type = 'GROUP' teapot.dupli_group = targetGroup teapot.pass_index = 1 mat = makeMaterial('teapotMat', (0, 0, 0), (0, 0, 0), 1) setMaterial(teapotMesh, mat) ### Renderer (only teapot) teapotToRender = -1 # renderTeapotsList = np.arange(len(teapots)) renderTeapotsList = np.arange(len(teapots))[0:1] targetModels = [] chAz = ch.Ch([0]) chObjAz = ch.Ch([0]) chAzRel = chAz - chObjAz chEl = ch.Ch([0.0]) chDist = ch.Ch([camDistance]) #Initialize to a random set of SH coefficients chLightSHCoeffs = ch.Ch(np.array([2, 0.25, 0.25, 0.12,-0.17,0.36,0.1,0.,0.])) # if multiObjects: chObjDist = ch.Ch([0]) chObjRotation = ch.Ch([0]) chObjAzMug = ch.Ch([0]) chObjDistMug = ch.Ch([0]) chObjRotationMug = ch.Ch([0]) chVColorsMug = ch.Ch([1,0,0]) chComponent = chLightSHCoeffs * clampedCosCoeffs light_color = ch.ones(3) chVColors = ch.Ch([0.4,0.4,0.4]) chDisplacement = ch.Ch([0.0, 0.0,0.0]) chScale = ch.Ch([1.0,1.0,1.0]) if useShapeModel: import shape_model #%% Load data shapeParams = np.zeros(latentDim) chShapeParams = ch.Ch(shapeParams.copy()) chVertices = shape_model.VerticesModel(chShapeParams=chShapeParams, meshLinearTransform=meshLinearTransform,W = W,b=b) chVertices.init() chVertices = ch.dot(geometry.RotateZ(-np.pi/2)[0:3,0:3],chVertices.T).T smFaces = [[faces]] smVColors = [chVColors*np.ones(chVertices.shape)] smUVs = ch.Ch(np.zeros([chVertices.shape[0],2])) smHaveTextures = [[False]] smTexturesList = [[None]] chVertices = chVertices - ch.mean(chVertices, axis=0) minZ = ch.min(chVertices[:,2]) chMinZ = ch.min(chVertices[:,2]) zeroZVerts = chVertices[:,2]- chMinZ chVertices = ch.hstack([chVertices[:,0:2] , zeroZVerts.reshape([-1,1])]) chVertices = chVertices*0.09 smCenter = ch.array([0,0,0.1]) smVertices = [chVertices] chNormals = shape_model.chGetNormals(chVertices, faces) smNormals = [chNormals] if useShapeModel: center = smCenter UVs = smUVs v = smVertices vn = smNormals Faces = smFaces VColors = smVColors UVs = smUVs HaveTextures = smHaveTextures TexturesList = smTexturesList else: v, vn = v_teapots[currentTeapotModel][0], vn_teapots[currentTeapotModel][0] Faces = f_list_teapots[currentTeapotModel][0] VColors = vc_teapots[currentTeapotModel][0] UVs = uv_teapots[currentTeapotModel][0] HaveTextures = haveTextures_list_teapots[currentTeapotModel][0] TexturesList = textures_list_teapots[currentTeapotModel][0] v, vn, teapotPosOffset = transformObject(v, vn, chScale, chObjAz, chObjDist, chObjRotation, np.array([0,0,0])) if renderMugs: verticesMug, normalsMug, mugPosOffset = transformObject(v_mug, vn_mug, chScale, chObjAzMug + np.pi / 2, chObjDistMug, chObjRotationMug, np.array([0,0,0])) VerticesB = [v] + [verticesMug] NormalsB = [vn] + [normalsMug] FacesB = [Faces] + [f_list_mug] VColorsB = [VColors] + [vc_mug] UVsB = [UVs] + [uv_mug] HaveTexturesB = [HaveTextures] + [haveTextures_list_mug] TexturesListB = [TexturesList] + [textures_list_mug] renderer = createRendererTarget(glMode, chAz, chEl, chDist, center, VerticesB, VColorsB, FacesB, NormalsB, light_color,chComponent, chVColors, np.array([0,0,0]), chDisplacement, width, height, UVsB, HaveTexturesB, TexturesListB, frustum, None) else: renderer = createRendererTarget(glMode, chAz, chEl, chDist, smCenter, [v], [smVColors], [smFaces], [vn], light_color, chComponent, chVColors, 0, chDisplacement, width,height, [smUVs], [smHaveTextures], [smTexturesList], frustum, None ) renderer.msaa = False renderer.overdraw = True # chShapeParams[:] = np.zeros([latentDim]) chVerticesMean = chVertices.r.copy() # else: # renderer = renderer_teapots[teapotToRender] ######################################################## ####### Initialization ends here ######################################################## print("Creating Ground Truth") trainAzsGT = np.array([]) trainObjAzsGT = np.array([]) trainElevsGT = np.array([]) trainLightAzsGT = np.array([]) trainLightElevsGT = np.array([]) trainLightIntensitiesGT = np.array([]) trainVColorGT = np.array([]) trainScenes = np.array([], dtype=np.uint8) trainTeapotIds = np.array([], dtype=np.uint8) trainEnvMaps = np.array([], dtype=np.uint8) trainOcclusions = np.array([]) trainTargetIndices = np.array([], dtype=np.uint8) trainIds = np.array([], dtype=np.uint32) #Multi trainObjDistGT = np.array([]) trainObjRotationGT = np.array([]) trainObjDistMug = np.array([]) trainObjRotationMug = np.array([]) trainObjAzMug = np.array([]) trainVColorsMug = np.array([]) trainTeapotElRel = np.array([]) trainMugElRel = np.array([]) trainMugPosOffset = np.array([]) trainTeapotPosOffset = np.array([]) trainLightCoefficientsGT = np.array([]).reshape([0,9]) trainLightCoefficientsGTRel = np.array([]).reshape([0,9]) trainAmbientIntensityGT = np.array([]) trainEnvMapPhiOffsets = np.array([]) trainShapeModelCoeffsGT = np.array([]).reshape([0,latentDim]) trainBBMug = np.array([],dtype=np.int8).reshape([0,4]) trainBBTeapot = np.array([],dtype=np.int8).reshape([0,4]) trainTeapotPresent = np.array([], dtype=np.bool) trainMugPresent = np.array([], dtype=np.bool) gtDir = 'groundtruth/' + prefix + '/' if not os.path.exists(gtDir + 'images/'): os.makedirs(gtDir + 'images/') if not os.path.exists(gtDir + 'sphericalharmonics/'): os.makedirs(gtDir + 'sphericalharmonics/') if not os.path.exists(gtDir + 'images_opendr/'): os.makedirs(gtDir + 'images_opendr/') if not os.path.exists(gtDir + 'masks_occlusion/'): os.makedirs(gtDir + 'masks_occlusion/') if generateTriplets: if not os.path.exists(gtDir + 'triplets1/'): os.makedirs(gtDir + 'triplets1/') if not os.path.exists(gtDir + 'triplets2/'): os.makedirs(gtDir + 'triplets2/') print("Generating renders") sceneLines = [line.strip() for line in open(replaceableScenesFile)] scenesToRender = range(len(sceneLines))[:] scenesToRender = range(len(sceneLines))[2:3] # scenesToRender = np.random.shuffle(scenesToRender) trainSize = 10000 renderTeapotsList = np.arange(len(teapots))[0:1] ignoreEnvMaps = np.loadtxt('data/bad_envmaps.txt') hdritems = list(envMapDic.items())[:] hdrstorender = [] phiOffsets = [0, np.pi/2, np.pi, 3*np.pi/2] for hdrFile, hdrValues in hdritems: hdridx = hdrValues[0] envMapCoeffs = hdrValues[1] if hdridx not in ignoreEnvMaps: hdrstorender = hdrstorender + [(hdrFile,hdrValues)] # if not os.path.exists('light_probes/envMap' + str(hdridx)): # os.makedirs('light_probes/envMap' + str(hdridx)) # # for phiOffset in phiOffsets: # # # phiOffset = np.random.uniform(0,2*np.pi, 1) # from numpy.random import choice # objAzGT = np.pi/2 # chObjAzGT[:] = 0 # totalOffset = phiOffset + chObjAzGT.r # envMapCoeffsRotated = np.dot(light_probes.sphericalHarmonicsZRotation(totalOffset), envMapCoeffs[[0,3,2,1,4,5,6,7,8]])[[0,3,2,1,4,5,6,7,8]].copy() # envMapCoeffsRotatedRel = np.dot(light_probes.sphericalHarmonicsZRotation(phiOffset), envMapCoeffs[[0,3,2,1,4,5,6,7,8]])[[0,3,2,1,4,5,6,7,8]].copy() # shCoeffsRGB = envMapCoeffsRotated.copy() # shCoeffsRGBRel = envMapCoeffsRotatedRel.copy() # chShCoeffs = 0.3*shCoeffsRGB[:,0] + 0.59*shCoeffsRGB[:,1] + 0.11*shCoeffsRGB[:,2] # chShCoeffsRel = 0.3*shCoeffsRGBRel[:,0] + 0.59*shCoeffsRGBRel[:,1] + 0.11*shCoeffsRGBRel[:,2] # chAmbientSHGT = chShCoeffs * chAmbientIntensityGT * clampedCosCoeffs # chAmbientSHGTRel = chShCoeffsRel * chAmbientIntensityGT * clampedCosCoeffs # chComponentGT[:] = chAmbientSHGT.r[:].copy() # chComponentGTRel[:] = chAmbientSHGTRel.r[:].copy() # cv2.imwrite('light_probes/envMap' + str(hdridx) + '/opendr_' + str(np.int(180*phiOffset/np.pi)) + '.png' , 255*rendererGT.r[:,:,[2,1,0]]) # sys.exit("") gtDtype = [('trainIds', trainIds.dtype.name), ('trainAzsGT', trainAzsGT.dtype.name),('trainObjAzsGT', trainObjAzsGT.dtype.name),('trainElevsGT', trainElevsGT.dtype.name), ('trainLightAzsGT', trainLightAzsGT.dtype.name),('trainLightElevsGT', trainLightElevsGT.dtype.name),('trainLightIntensitiesGT', trainLightIntensitiesGT.dtype.name), ('trainVColorGT', trainVColorGT.dtype.name, (3,) ),('trainScenes', trainScenes.dtype.name),('trainTeapotIds', trainTeapotIds.dtype.name), ('trainEnvMaps', trainEnvMaps.dtype.name),('trainOcclusions', trainOcclusions.dtype.name),('trainTargetIndices', trainTargetIndices.dtype.name), ('trainLightCoefficientsGT',trainLightCoefficientsGT.dtype, (9,)), ('trainLightCoefficientsGTRel', trainLightCoefficientsGTRel.dtype, (9,)), ('trainAmbientIntensityGT', trainAmbientIntensityGT.dtype), ('trainEnvMapPhiOffsets', trainEnvMapPhiOffsets.dtype), ('trainShapeModelCoeffsGT', trainShapeModelCoeffsGT.dtype, (latentDim,)), ('trainObjDistGT', trainObjDistGT.dtype), ('trainObjRotationGT', trainObjRotationGT.dtype), ('trainObjDistMug', trainObjDistMug.dtype), ('trainObjRotationMug', trainObjRotationMug.dtype), ('trainObjAzMug', trainObjAzMug.dtype), ('trainVColorsMug', trainVColorsMug.dtype, (3,)), ('trainTeapotElRel', trainTeapotElRel.dtype), ('trainMugElRel', trainMugElRel.dtype), ('trainMugPosOffset', trainMugPosOffset.dtype,(3,)), ('trainTeapotPosOffset', trainTeapotPosOffset.dtype,(3,)), ('trainBBMug', trainBBMug.dtype, (4,)), ('trainBBTeapot', trainBBTeapot.dtype, (4,)), ('trainTeapotPresent', trainTeapotPresent.dtype), ('trainMugPresent', trainMugPresent.dtype)] groundTruth = np.array([], dtype = gtDtype) groundTruthFilename = gtDir + 'groundTruth.h5' # gtDirToRender = 'groundtruth/' + previousGTPrefix + '/' # gtDataFileToRender = h5py.File(gtDirToRender + 'groundTruth.h5', 'w') # gtDatasetToRender = gtDataFileToRender.create_dataset(previousGTPrefix, data=groundTruth, maxshape=(None,)) nextId = 0 if not replaceNewGroundtruth: gtDataFile = h5py.File(groundTruthFilename, 'a') try: gtDataset = gtDataFile[prefix] if gtDataset.size > 0: nextId = gtDataset['trainIds'][-1] + 1 except: gtDataset = gtDataFile.create_dataset(prefix, data=groundTruth, maxshape=(None,)) else: gtDataFile = h5py.File(groundTruthFilename, 'w') gtDataset = gtDataFile.create_dataset(prefix, data=groundTruth, maxshape=(None,)) train_i = nextId #Re-producible groundtruth generation. if train_i == 0: np.random.seed(1) # np.random.seed(2) unlinkedObj = None scenesToRenderOcclusions = [] scenes = [] lenScenes = 0 #Compute how many different locations can the teapot be instantiated across all scenes. for sceneIdx in scenesToRender: sceneNumber, sceneFileName, instances, roomName, roomInstanceNum, targetIndices, targetPositions = scene_io_utils.getSceneInformation(sceneIdx, replaceableScenesFile) sceneDicFile = 'data/scene' + str(sceneNumber) + '.pickle' # if renderOcclusions: # targetIndicesNew = [] # occlusionSceneFile = 'data/occlusions/occlusionScene' + str(sceneNumber) + '.pickle' # with open(occlusionSceneFile, 'rb') as pfile: # occlusions = pickle.load(pfile) # # for targetidx, targetIndex in enumerate(targetIndices): # if not occlusions[targetIndex][1]: # print("Scene idx " + str(sceneIdx) + " at index " + str(targetIndex) + " has no proper occlusion.") # else: # targetIndicesNew = targetIndicesNew + [targetIndex] # targetIndices = targetIndicesNew # # collisionSceneFile = 'data/collisions/collisionScene' + str(sceneNumber) + '.pickle' # scenes = scenes + [targetIndices] # with open(collisionSceneFile, 'rb') as pfile: # collisions = pickle.load(pfile) # # for targetidx, targetIndex in enumerate(targetIndices): # if not collisions[targetIndex][1]: # print("Scene idx " + str(sceneIdx) + " at index " + str(targetIndex) + " collides everywhere.") scenes = scenes + [targetIndices] lenScenes += len(targetIndices) groundTruthInfo = {'prefix':prefix, 'previousGTPrefix':previousGTPrefix, 'renderFromPreviousGT':renderFromPreviousGT, 'useShapeModel':useShapeModel, 'renderOcclusions':renderOcclusions, 'useOpenDR':useOpenDR, 'useBlender':useBlender, 'renderBlender':renderBlender, 'captureEnvMapFromBlender':captureEnvMapFromBlender, 'parseSceneInstantiations':parseSceneInstantiations, 'loadBlenderSceneFile':loadBlenderSceneFile, 'useCycles':useCycles, 'unpackModelsFromBlender':unpackModelsFromBlender, 'unpackSceneFromBlender':unpackSceneFromBlender, 'loadSavedSH':loadSavedSH, 'renderTeapots':renderTeapots, 'renderMugs':renderMugs, 'width':width, 'height':height, 'angle':angle, 'clip_start':clip_start, 'clip_end':clip_end, 'camDistance':camDistance, 'renderTeapotsList':renderTeapotsList, 'renderMugsList':renderMugsList, 'replaceableScenesFile':replaceableScenesFile, 'teapotsFile':teapotsFile, 'SHFilename':SHFilename, 'light_colorGT':light_colorGT, 'chDisplacement':chDisplacement, 'chDisplacementGT':chDisplacementGT, 'chScale':chScale, 'chScaleGT':chScaleGT} with open(gtDir + 'gtInfo.pickle', 'wb') as pfile: pickle.dump(groundTruthInfo, pfile) #Generate GT labels before rendering them. if not renderFromPreviousGT: for scene_i, sceneIdx in enumerate(scenesToRender): sceneNumber, sceneFileName, instances, roomName, roomInstanceNum, targetIndicesScene, targetPositions = scene_io_utils.getSceneInformation(sceneIdx, replaceableScenesFile) print("Generating groundtruth for scene: " + str(sceneNumber)) # targetIndices = scenes[scene_i] # if not targetIndices: # continue targetIndices = targetIndicesScene sceneDicFile = 'data/scene' + str(sceneNumber) + '.pickle' collisionSceneFile = 'data/collisions/collisionScene' + str(sceneNumber) + '.pickle' with open(collisionSceneFile, 'rb') as pfile: collisions = pickle.load(pfile) # if renderOcclusions: # occlusionSceneFile = 'data/occlusions/occlusionScene' + str(sceneNumber) + '.pickle' # with open(occlusionSceneFile, 'rb') as pfile: # occlusions = pickle.load(pfile) v2, f_list2, vc2, vn2, uv2, haveTextures_list2, textures_list2 = scene_io_utils.loadSavedScene(sceneDicFile, tex_srgb2lin) if useBlender and not loadBlenderSceneFile: bpy.ops.wm.read_factory_settings() scene = scene_io_utils.loadBlenderScene(sceneIdx, replaceableScenesFile) scene_io_utils.setupScene(scene, roomInstanceNum, scene.world, scene.camera, width, height, 16, useCycles, True) scene.update() # Save barebones scene. elif useBlender and loadBlenderSceneFile: bpy.ops.wm.read_factory_settings() scene_io_utils.loadSceneBlendData(sceneIdx, replaceableScenesFile) scene = bpy.data.scenes['Main Scene'] #bpy.ops.wm.save_as_mainfile(filepath="/home/pol/sceneload.blend") # Configure scene if useBlender: if renderTeapots: targetModels = [] blender_teapots = [] teapots = [line.strip() for line in open('teapots.txt')] selection = [teapots[i] for i in renderTeapotsList] scene_io_utils.loadTargetsBlendData() for teapotIdx, teapotName in enumerate(selection): teapot = bpy.data.scenes[teapotName[0:63]].objects['teapotInstance' + str(renderTeapotsList[teapotIdx])] teapot.layers[1] = True teapot.layers[2] = True targetModels = targetModels + [teapotIdx] blender_teapots = blender_teapots + [teapot] if renderMugs: blender_mugs = [] selectionMugs = [mugs[i] for i in renderMugsList] scene_io_utils.loadMugsBlendData() for mugIdx, mugName in enumerate(selectionMugs): mug = bpy.data.scenes[mugName[0:63]].objects['mugInstance' + str(renderMugsList[mugIdx])] mug.layers[1] = True mug.layers[2] = True mugModels = mugModels + [mug] blender_mugs = blender_mugs + [mug] # setupSceneGroundtruth(scene, width, height, clip_start, 2000, 'CUDA', 'CUDA_MULTI_2') # bpy.ops.wm.save_as_mainfile(filepath="/home/pol/sceneload1.blend") setupSceneGroundtruth(scene, width, height, clip_start, 2000) # bpy.ops.wm.save_as_mainfile(filepath="/home/pol/sceneload2.blend") treeNodes = scene.world.node_tree links = treeNodes.links cubeScene = createCubeScene(scene) setEnviornmentMapStrength(20, cubeScene) unlinkedObj = None unlinkedCubeObj = None envMapFilename = None for targetidx, targetIndex in enumerate(targetIndices): targetPosition = targetPositions[np.where(targetIndex==np.array(targetIndicesScene))[0][0]] # if sceneIdx != currentScene or targetIndex != currentTargetIndex: # targetPosition = targetPositions[np.where(targetIndex == np.array(targetIndicesScene))[0]] v, f_list, vc, vn, uv, haveTextures_list, textures_list = copy.deepcopy(v2), copy.deepcopy(f_list2), copy.deepcopy(vc2), copy.deepcopy(vn2)\ , copy.deepcopy(uv2), copy.deepcopy(haveTextures_list2), copy.deepcopy(textures_list2) removeObjectData(len(v) - 1 - targetIndex, v, f_list, vc, vn, uv, haveTextures_list, textures_list) # if sceneIdx != currentScene or targetIndex != currentTargetIndex: parentIdx = instances[targetIndex][1] if useBlender: if unlinkedObj != None: scene.objects.link(unlinkedObj) cubeScene.objects.link(unlinkedCubeObj) unlinkedObj = scene.objects[str(targetIndex)] unlinkedCubeObj = cubeScene.objects['cube' + str(targetIndex)] scene.objects.unlink(unlinkedObj) cubeScene.objects.unlink(unlinkedCubeObj) parentSupportObj = scene.objects[str(parentIdx)] supportWidthMax, supportWidthMin = modelWidth(parentSupportObj.dupli_group.objects, parentSupportObj.matrix_world) supportDepthMax, supportDepthMin = modelDepth(parentSupportObj.dupli_group.objects, parentSupportObj.matrix_world) # supportRad = min(0.5 * np.sqrt((supportDepthMax - supportDepthMin) ** 2),0.5 * np.min((supportWidthMax - supportWidthMin) ** 2)) supportRad = np.sqrt(0.5 * (supportDepthMax - supportDepthMin) ** 2 + 0.5 * (supportWidthMax - supportWidthMin) ** 2) distRange = min(supportRad, 0.3) sceneCollisions = {} collisionsFile = 'data/collisions/discreteCollisions_scene' + str(sceneNumber) + '_targetIdx' + str(targetIndex) + '.pickle' distInterval = 0.05 rotationRange = 2 * np.pi rotationInterval = 10 * np.pi / 180 if not parseSceneInstantiations and os.path.exists(collisionsFile): with open(collisionsFile, 'rb') as pfile: sceneCollisions = pickle.load(pfile) supportRad = sceneCollisions['supportRad'] instantiationBinsTeapot = sceneCollisions['instantiationBinsTeapot'] distRange = sceneCollisions['distRange'] rotationRange = sceneCollisions['rotationRange'] rotationInterval = sceneCollisions['rotationInterval'] distInterval = sceneCollisions['distInterval'] instantiationBinsMug = sceneCollisions['instantiationBinsMug'] totalBinsTeapot = sceneCollisions['totalBinsTeapot'] totalBinsMug = sceneCollisions['totalBinsMug'] else: assert(useBlender) print("Parsing collisions for scene " + str(sceneNumber)) placeCamera(cubeScene.camera, 0, 45, chDistGT.r[0].copy(), center[:].copy() + targetPosition[:].copy()) scaleZ = 0.2 scaleY = 0.2 scaleX = 0.2 cubeTeapot = createCube(scaleX, scaleY, scaleZ, 'cubeTeapot') cubeTeapot.matrix_world = mathutils.Matrix.Translation(targetPosition) # cubeTeapot = getCubeObj(teapot) # cubeMug = getCubeObj(mug) scaleZ = 0.1 scaleY = 0.1 scaleX = 0.1 cubeMug = createCube(scaleX, scaleY, scaleZ, 'cubeMug') cubeMug.matrix_world = mathutils.Matrix.Translation(targetPosition) cubeScene.objects.link(cubeTeapot) cubeParentSupportObj = cubeScene.objects['cube' + str(parentIdx)] cubeRoomObj = cubeScene.objects['cube' + str(roomInstanceNum)] cubeScene.update() objDisplacementMat = computeHemisphereTransformation(chObjRotationGT, 0, chObjDistGT, np.array([0, 0, 0])) objOffset = objDisplacementMat[0:3, 3] instantiationBinsTeapot, totalBinsTeapot = collision.parseSceneCollisions(gtDir, scene_i, targetIndex, cubeTeapot, cubeScene, objOffset, chObjDistGT, chObjRotationGT, cubeParentSupportObj, cubeRoomObj, distRange, rotationRange, distInterval, rotationInterval) cubeScene.objects.unlink(cubeTeapot) cubeScene.objects.link(cubeMug) instantiationBinsMug, totalBinsMug = collision.parseSceneCollisions(gtDir, scene_i, targetIndex, cubeMug, cubeScene, objOffset, chObjDistGT, chObjRotationGT, cubeParentSupportObj, cubeRoomObj, distRange, rotationRange, distInterval, rotationInterval) cubeScene.objects.unlink(cubeMug) deleteObject(cubeTeapot) deleteObject(cubeMug) sceneCollisions = {'totalBinsTeapot': totalBinsTeapot, 'totalBinsMug': totalBinsMug, 'supportRad': supportRad, 'instantiationBinsTeapot': instantiationBinsTeapot, 'distRange': distRange, 'rotationRange': rotationRange, 'distInterval': distInterval, 'rotationInterval': rotationInterval, 'instantiationBinsMug': instantiationBinsMug} with open(collisionsFile, 'wb') as pfile: pickle.dump(sceneCollisions, pfile) #Go to next target Index or scene if there are no plausible place instantiations for teapot or mug. if instantiationBinsTeapot.sum() < 1 or instantiationBinsMug.sum() < 1: print("This scene position has not place to instantiate the objects.") continue # if useShapeModel for teapot_i in renderTeapotsList: if useShapeModel: teapot_i = -1 else: currentTeapotModel = teapot_i center = center_teapots[teapot_i] print("Ground truth on new teapot" + str(teapot_i)) ##Destroy and create renderer if useOpenDR: rendererGT.makeCurrentContext() rendererGT.clear() contextdata.cleanupContext(contextdata.getContext()) if glMode == 'glfw': glfw.destroy_window(rendererGT.win) del rendererGT if renderTeapots: currentTeapotModel = teapot_i center = center_teapots[teapot_i] if useShapeModel: center = smCenterGT UVs = smUVsGT vGT = smVerticesGT vnGT = smNormalsGT Faces = smFacesGT VColors = smVColorsGT UVs = smUVsGT HaveTextures = smHaveTexturesGT TexturesList = smTexturesListGT else: vGT, vnGT = v_teapots[currentTeapotModel][0], vn_teapots[currentTeapotModel][0] Faces = f_list_teapots[currentTeapotModel][0] VColors = vc_teapots[currentTeapotModel][0] UVs = uv_teapots[currentTeapotModel][0] HaveTextures = haveTextures_list_teapots[currentTeapotModel][0] TexturesList = textures_list_teapots[currentTeapotModel][0] vGT, vnGT, teapotPosOffset = transformObject(vGT, vnGT, chScaleGT, chObjAzGT, chObjDistGT, chObjRotationGT, targetPosition) if renderMugs: verticesMug, normalsMug, mugPosOffset = transformObject(v_mug, vn_mug, chScale, chObjAzMug + np.pi / 2, chObjDistMug, chObjRotationMug, targetPosition) VerticesMug = [verticesMug] NormalsMug = [normalsMug] FacesMug = [f_list_mug] VColorsMug = [vc_mug] UVsMug = [uv_mug] HaveTexturesMug = [haveTextures_list_mug] TexturesListMug = [textures_list_mug] else: VerticesMug = [] NormalsMug = [] FacesMug = [] VColorsMug = [] UVsMug = [] HaveTexturesMug = [] TexturesListMug = [] if renderTeapots: VerticesTeapot = [vGT] NormalsTeapot = [vnGT] FacesTeapot = [Faces] VColorsTeapot = [VColors] UVsTeapot = [UVs] HaveTexturesTeapot = [HaveTextures] TexturesListTeapot = [TexturesList] else: VerticesTeapot = [] NormalsTeapot = [] FacesTeapot = [] VColorsTeapot = [] UVsTeapot = [] HaveTexturesTeapot = [] TexturesListTeapot = [] # addObjectData(v, f_list, vc, vn, uv, haveTextures_list, textures_list, verticesMug, f_list_mug, vc_mug, normalsMug, uv_mug, haveTextures_list_mug, textures_list_mug) # addObjectData(v, f_list, vc, vn, uv, haveTextures_list, textures_list, vGT, Faces, VColors, vnGT, UVs, HaveTextures, TexturesList) VColorsTeapot = [[np.ones_like(VColorsTeapot[0][0]) * chVColorsGT.reshape([1, 3])]] if renderBackground: vc_scene = VColorsTeapot + VColorsMug + vc vn_scene = NormalsTeapot + NormalsMug + vn else: vc_scene = VColorsTeapot + VColorsMug vn_scene = NormalsTeapot + NormalsMug vc_scene_illum = [] for mesh in range(len(vc_scene)): if useDirectionalLight: vc_scene_illum += [computeGlobalAndDirectionalLighting(vn_scene[mesh], vc_scene[mesh], chLightAzGT, chLightElGT, chLightIntensityGT, chGlobalConstantGT)] else: vc_scene_illum += [computeSphericalHarmonics(vn_scene[mesh], vc_scene[mesh], light_colorGT, chComponentGT)] if renderBackground: v_scene = VerticesTeapot + VerticesMug + v f_scene = FacesTeapot + FacesMug + f_list UVsTeapot + UVsMug + uv HaveTexturesTeapot + HaveTexturesMug + haveTextures_list TexturesListTeapot + TexturesListMug + textures_list else: v_scene = VerticesTeapot + VerticesMug f_scene = FacesTeapot + FacesMug uvs_scene = UVsTeapot + UVsMug haveTextures_scene = HaveTexturesTeapot + HaveTexturesMug textures_list_scene = TexturesListTeapot + TexturesListMug rendererGT = createRendererGT(glMode, chAzGT, chElGT, chDistGT, center, v_scene, vc_scene_illum, f_scene, vn_scene, light_colorGT, chComponentGT, chVColorsGT, targetPosition.copy(), chDisplacementGT, width, height, uvs_scene, haveTextures_scene, textures_list_scene, frustum, None) rendererGT.overdraw = True rendererGT.nsamples = 1 rendererGT.msaa = False rendererGT.initGL() rendererGT.initGLTexture() ## Blender: Unlink and link new teapot. if useBlender: if renderTeapots: # if currentScene != -1 and currentTargetIndex != -1 and currentTeapot != -1 and teapot != None: if teapot.name in scene.objects: scene.objects.unlink(teapot) if useShapeModel: deleteInstance(teapot) if not useShapeModel: teapot = blender_teapots[currentTeapotModel] else: teapotMesh = createMeshFromData('teapotShapeModelMesh', chVerticesGT.r.tolist(), faces.astype(np.int32).tolist()) teapotMesh.layers[0] = True teapotMesh.layers[1] = True teapotMesh.pass_index = 1 targetGroup = bpy.data.groups.new('teapotShapeModelGroup') targetGroup.objects.link(teapotMesh) teapot = bpy.data.objects.new('teapotShapeModel', None) teapot.dupli_type = 'GROUP' teapot.dupli_group = targetGroup teapot.pass_index = 1 mat = makeMaterial('teapotMat', (0, 0, 0), (0, 0, 0), 1) setMaterial(teapotMesh, mat) # center = centerOfGeometry(teapot.dupli_group.objects, teapot.matrix_world) placeNewTarget(scene, teapot, targetPosition[:].copy()) teapot.layers[0] = True teapot.layers[1] = True original_matrix_world = teapot.matrix_world.copy() if renderMugs: if mug.name in scene.objects: scene.objects.unlink(mug) # deleteInstance(mug) mug = blender_mugs[currentTeapotModel] placeNewTarget(scene, mug, targetPosition[:].copy()) mug.layers[1] = True mug.layers[0] = True original_matrix_world_mug = mug.matrix_world.copy() # LIGHT RANDOMIZATION hdrchoice = np.random.choice(len(hdrstorender)) hdrchoice = 0 hdrValues = hdrstorender[hdrchoice][1] hdrFile = hdrstorender[hdrchoice][0] hdridx = hdrValues[0] envMapCoeffs = hdrValues[1] envMapFilename = hdrFile # if not useBlender: envMapTexture = np.array(imageio.imread(envMapFilename))[:,:,0:3] numTeapotTrain = 0 numAttempts = 0 #Rejection sampling maxAttempts = max(int(trainSize/(lenScenes*len(renderTeapotsList))),1) exitInstantiationLoop = False while numTeapotTrain < maxAttempts and not exitInstantiationLoop: numAttempts = numAttempts + 1 if numAttempts > 50 and numAttempts/(numTeapotTrain + 1) > 20: exitInstantiationLoop = True ignore = False meanValIntensityOffset = np.random.uniform(-0.5,0.5) meanValIntensityOffset = 0 chAmbientIntensityGTVals = (0.8 + meanValIntensityOffset)/(0.3*envMapCoeffs[0,0] + 0.59*envMapCoeffs[0,1]+ 0.11*envMapCoeffs[0,2]) # chAmbientIntensityGTVals = chGlobalConstantGT.r #Light randomization phiOffsetVals = np.random.uniform(0,2*np.pi, 1) # phiOffsetVals = 0 # phiOffset[:] = 0 from numpy.random import choice chAzGTVals = np.mod(np.random.uniform(0,np.pi, 1) - np.pi/2, 2*np.pi) chElGTVals = np.random.uniform(0.05,np.pi/2, 1) chLightAzGTVals = np.random.uniform(0,2*np.pi, 1) # chLightAzGTVals = chLightAzGT.r chLightElGTVals = np.random.uniform(0,np.pi/2, 1) # chLightElGTVals = chLightElGT.r # chLightIntensityGTVals = chLightIntensityGT.r chLightIntensityGTVals = np.random.uniform(0,1) chGlobalConstantGTVals = np.random.uniform(0.1,0.9) # chAmbientIntensityGTVals = chGlobalConstantGTVals chVColorsGTVals = np.random.uniform(0.0,1.0, [1, 3]) envMapCoeffsRotatedVals = np.dot(light_probes.chSphericalHarmonicsZRotation(totalOffset), envMapCoeffs[[0,3,2,1,4,5,6,7,8]])[[0,3,2,1,4,5,6,7,8]] envMapCoeffsRotatedRelVals = np.dot(light_probes.chSphericalHarmonicsZRotation(phiOffset), envMapCoeffs[[0,3,2,1,4,5,6,7,8]])[[0,3,2,1,4,5,6,7,8]] # SHAPE RANDOMIZATION shapeParams = np.random.randn(latentDim) # shapeParams = np.zeros(latentDim) # shapeParams[0:4] = np.random.randn(4) chShapeParamsGTVals = shapeParams ## Update renderer scene latent variables. # chAmbientIntensityGT[:] = chGlobalConstantGTVals chAmbientIntensityGT[:] = chGlobalConstantGT.r phiOffset[:] = phiOffsetVals chAzGT[:] = chAzGTVals chElGT[:] = chElGTVals chLightAzGT[:] = chLightAzGTVals chLightElGT[:] = chLightElGTVals teapotCamElGT = 0 teapotPosOffsetVals = 0 if useOpenDR: cameraEye = np.linalg.inv(np.r_[rendererGT.camera.view_mtx, np.array([[0, 0, 0, 1]])])[0:3, 3] vecToCenter = targetPosition - cameraEye vecToCenter = vecToCenter / np.linalg.norm(vecToCenter) rightCamVec = np.cross(vecToCenter, np.array([0, 0, 1])) chObjAzGTVals = 0 chObjDistGTVals = 0 chObjRotationGTVals = 0 if renderTeapots: chObjAzGTVals = np.random.uniform(0, np.pi * 2) chObjDistGTVals = np.random.uniform(0, np.minimum(supportRad, 0.3)) chObjAzGT[:] = chObjAzGTVals chVColorsGT[:] = chVColorsGTVals try: chShapeParamsGT[:] = shapeParams except: chShapeParamsGT[:] = np.random.randn(latentDim) #Instantiate such that we always see the handle if only a bit! chObjRotationGTVals = np.random.uniform(0,np.pi*2) if not centeredObject: teapotPlacement = np.random.choice(instantiationBinsTeapot.sum()) chObjDistGTVals = totalBinsTeapot[0].ravel()[instantiationBinsTeapot][teapotPlacement] chObjRotationGTVals = totalBinsTeapot[1].ravel()[instantiationBinsTeapot][teapotPlacement] chObjDistGTVals = np.random.uniform(chObjDistGTVals - distInterval/2, chObjDistGTVals + distInterval/2) chObjRotationGTVals = np.mod(np.random.uniform(chObjRotationGTVals - rotationInterval/2, chObjRotationGTVals + rotationInterval/2), 2*np.pi) else: chObjDistGTVals = 0 chObjRotationGTVals = 0 chObjDistGT[:] = chObjDistGTVals chObjRotationGT[:] = chObjRotationGTVals if not centeredObject: vecToTeapot = targetPosition + teapotPosOffset.r - cameraEye vecToTeapot = vecToTeapot / np.linalg.norm(vecToTeapot) teapotPosRight = np.sign(rightCamVec.dot(vecToTeapot)) angleToTeapot = np.arccos(vecToTeapot.dot(vecToCenter)) if np.isnan(angleToTeapot): angleToMug = 0 ignore = True chObjAzGTRel = chAzGT.r - teapotPosRight * angleToTeapot teapotPosOffsetVals = teapotPosOffset.r chObjDistMugVals = 0 chObjRotationMugVals = 0 chObjAzMugVals = 0 mugCamElGT = 0 mugPosOffsetVals = 0 chVColorsMugVals = np.random.uniform(0.0, 1, [1, 3]) if renderMugs: chObjDistMugVals = np.random.uniform(0,np.minimum(supportRad, 0.4)) chObjRotationMugVals = np.random.uniform(0,np.pi*2) instantiationBinsMugUpdated = instantiationBinsMug if renderTeapots: if not centeredObject: ys = -np.cos(totalBinsMug[1]) * totalBinsMug[0] xs = np.sin(totalBinsMug[1]) * totalBinsMug[0] y = -np.cos(chObjRotationGTVals) * chObjDistGTVals x = np.sin(chObjRotationGTVals) * chObjDistGTVals instantiationBinsMugUpdated = instantiationBinsMug.copy() # instantiationBinsMugUpdated[(totalBinsMug[0] > chObjDistGTVals - 0.15) & (totalBinsMug[0] < chObjDistGTVals + 0.15) & (totalBinsMug[1] > chObjRotationGTVals - 15*np.pi/180) & (totalBinsMug[1] < chObjRotationGTVals + 15*np.pi/180)] = False instantiationBinsMugUpdated[np.sqrt((ys-y)**2 + (xs-x)**2) < 0.15] = False if instantiationBinsMugUpdated.sum() == 0: ignore = True mugPlacement = np.random.choice(instantiationBinsMugUpdated.sum()) chObjDistMugVals = totalBinsMug[0].ravel()[instantiationBinsMugUpdated][mugPlacement] chObjRotationMugVals = totalBinsMug[1].ravel()[instantiationBinsMugUpdated][mugPlacement] chObjDistMugVals = np.random.uniform(chObjDistMugVals - distInterval/2, chObjDistMugVals + distInterval/2) chObjRotationMugVals = np.mod(np.random.uniform(chObjRotationMugVals - rotationInterval/2, chObjRotationMugVals + rotationInterval/2), 2*np.pi) else: chObjDistGTVals = 0 chObjRotationGTVals = 0 chObjDistMug[:] = chObjDistMugVals chObjRotationMug[:] = chObjRotationMugVals vecToMug = targetPosition + mugPosOffset.r - cameraEye vecToMug = vecToMug / np.linalg.norm(vecToMug) mugPosRight = np.sign(rightCamVec.dot(vecToMug)) angleToMug = np.arccos(vecToMug.dot(vecToCenter)) if np.isnan(angleToMug): angleToMug = 0 ignore = True chObjAzMugVals = np.random.uniform(chAzGT.r - mugPosRight*angleToMug - 2 * np.pi / 3, chAzGT.r - mugPosRight * angleToMug + 2 * np.pi / 3) chObjAzMugRel = chAzGT.r - mugPosRight*angleToMug chObjAzMug[:] = chObjAzMugVals chVColorsMug[:] = chVColorsMugVals vecMugToCamGT = cameraEye - (mugPosOffset + center) mugCamElGT = 2 * ch.arctan(ch.norm(ch.array([0, -1, 0]) * ch.norm(vecMugToCamGT) - vecMugToCamGT * ch.norm(ch.array([0, -1, 0]))) / ch.norm(ch.array([0, -1, 0]) * ch.norm(vecMugToCamGT) + ch.norm(ch.array([0, -1, 0])) * vecMugToCamGT)) vecTeapotToCamGT = cameraEye - (teapotPosOffset + center) teapotCamElGT = 2 * ch.arctan(ch.norm(ch.array([0, -1, 0]) * ch.norm(vecTeapotToCamGT) - vecTeapotToCamGT * ch.norm(ch.array([0, -1, 0]))) / ch.norm(ch.array([0, -1, 0]) * ch.norm(vecTeapotToCamGT) + ch.norm( ch.array([0, -1, 0])) * vecTeapotToCamGT)) mugPosOffsetVals = mugPosOffset.r if useBlender and not ignore: updateEnviornmentMap(envMapFilename, scene) rotateEnviornmentMap(totalOffset.r.copy(), scene) placeCamera(scene.camera, -chAzGT.r[:].copy() * 180 / np.pi, chElGT.r[:].copy() * 180 / np.pi, chDistGT.r[0].copy(), center[:].copy() + targetPosition[:].copy()) if renderTeapots: azimuthRot = mathutils.Matrix.Rotation(chObjAzGT.r[:].copy(), 4, 'Z') teapot.matrix_world = mathutils.Matrix.Translation(original_matrix_world.to_translation() + mathutils.Vector(teapotPosOffset.r)) * azimuthRot * (mathutils.Matrix.Translation(-original_matrix_world.to_translation())) * original_matrix_world setObjectDiffuseColor(teapot, chVColorsGT.r.copy()) if useShapeModel: mesh = teapot.dupli_group.objects[0] for vertex_i, vertex in enumerate(mesh.data.vertices): vertex.co = mathutils.Vector(chVerticesGT.r[vertex_i]) if renderMugs: setObjectDiffuseColor(mug, chVColorsMug.r.copy()) azimuthRotMug = mathutils.Matrix.Rotation(chObjAzMug.r[:].copy() - np.pi / 2, 4, 'Z') mug.matrix_world = mathutils.Matrix.Translation(original_matrix_world_mug.to_translation() + mathutils.Vector(mugPosOffset.r)) * azimuthRotMug * (mathutils.Matrix.Translation(-original_matrix_world_mug.to_translation())) * original_matrix_world_mug scene.update() ## Some validation checks: if useOpenDR and not ignore: if renderTeapots: occlusion = getOcclusionFraction(rendererGT, id=teapotSceneIndex) vis_occluded = np.array(rendererGT.indices_image == teapotSceneIndex+1).copy().astype(np.bool) vis_occluded = np.ones_like(vis_occluded) vis_im = np.array(rendererGT.image_mesh_bool([teapotSceneIndex])).copy().astype(np.bool) vis_im = np.ones_like(vis_im) if renderMugs: occlusionMug = getOcclusionFraction(rendererGT, id=mugSceneIndex) vis_occluded_mug = np.array(rendererGT.indices_image == mugSceneIndex+1).copy().astype(np.bool) vis_im_mug = np.array(rendererGT.image_mesh_bool([mugSceneIndex])).copy().astype(np.bool) if renderTeapots: if occlusion < occlusionMin or occlusion > occlusionMax: ignore = True # if occlusion > 0.9 or vis_occluded.sum() < 10 or np.isnan(occlusion): #Cian change: if occlusion > 0.01 or vis_occluded.sum() < 10 or np.isnan(occlusion): ignore = True if np.sum(vis_im[:,0]) > 1 or np.sum(vis_im[0,:]) > 1 or np.sum(vis_im[:,-1]) > 1 or np.sum(vis_im[-1,:]) > 1: ignore = True if renderMugs and not showMug: if occlusionMug > 0.9 or vis_occluded_mug.sum() < 10 or np.isnan(occlusionMug): ignore = True #Check that objects are not only partly in the viewing plane of the camera: if np.sum(vis_im_mug[:,0]) > 1 or np.sum(vis_im_mug[0,:]) > 1 or np.sum(vis_im_mug[:,-1]) > 1 or np.sum(vis_im_mug[-1,:]) > 1: if showMug: ignore = True #Don't take into account "ignore" variable for now... ignore = False # if not ignore: # # Ignore if camera collides with occluding object as there are inconsistencies with OpenDR and Blender. # cameraEye = np.linalg.inv(np.r_[rendererGT.camera.view_mtx, np.array([[0, 0, 0, 1]])])[0:3,3] # # vDists = rendererGT.v.r[rendererGT.f[rendererGT.visibility_image[ # rendererGT.visibility_image != 4294967295].ravel()].ravel()] - cameraEye # # minDistToObjects = 0.2 # maxDistToObjects = 0.6 # # #Ignore when teapot or mug is up to 10 cm to the camera eye, or too far (more than 1 meter). # # if np.min(np.linalg.norm(vDists, axis=1)) <= clip_start: # ignore = True # # if renderTeapots: # vDistsTeapot = rendererGT.v.r[rendererGT.f[rendererGT.visibility_image[vis_occluded].ravel()].ravel()] - cameraEye # if np.min(np.linalg.norm(vDistsTeapot, axis=1)) <= minDistToObjects or np.min(np.linalg.norm(vDistsTeapot, axis=1)) > maxDistToObjects: # ignore = True # # if renderMugs: # vDistsMug = rendererGT.v.r[rendererGT.f[rendererGT.visibility_image[vis_occluded_mug].ravel()].ravel()] - cameraEye # if np.min(np.linalg.norm(vDistsMug, axis=1)) <= minDistToObjects or np.min(np.linalg.norm(vDistsMug, axis=1)) > maxDistToObjects: # ignore = True # # if useBlender: # if renderTeapots: # cubeTeapot = getCubeObj(teapot) # cubeScene.objects.link(cubeTeapot) # # if renderMugs: # cubeMug = getCubeObj(mug) # cubeScene.objects.link(cubeMug) # # cubeParentSupportObj = cubeScene.objects['cube'+str(parentIdx)] # cubeRoomObj = cubeScene.objects['cube' + str(roomInstanceNum)] # cubeScene.update() # # if renderTeapots: # if collision.targetCubeSceneCollision(cubeTeapot, cubeScene, 'cube'+str(roomInstanceNum), cubeParentSupportObj): # print("Teapot intersects with an object.") # ignore = True # # if not ignore and collision.instancesIntersect(mathutils.Matrix.Translation(mathutils.Vector((0, 0, +0.02))), [cubeTeapot], mathutils.Matrix.Identity(4), [cubeParentSupportObj]): # print("Teapot interesects supporting object.") # ignore = True # # # if not ignore and collision.instancesIntersect(mathutils.Matrix.Identity(4), [cubeTeapot], mathutils.Matrix.Identity(4), [cubeRoomObj]): # print("Teapot intersects room") # ignore = True # # if not ignore and not ignore and not collision.instancesIntersect(mathutils.Matrix.Translation(mathutils.Vector((0, 0, -0.02))), [cubeTeapot], mathutils.Matrix.Identity(4), [cubeParentSupportObj]): # print("Teapot not on table.") # ignore = True # # if renderMugs: # if not ignore and collision.targetCubeSceneCollision(cubeMug, cubeScene, 'cube' + str(roomInstanceNum), cubeParentSupportObj): # print("Mug intersects with an object.") # ignore = True # # if not ignore and collision.instancesIntersect(mathutils.Matrix.Translation(mathutils.Vector((0, 0, +0.02))), [cubeMug], mathutils.Matrix.Identity(4), [cubeParentSupportObj]): # print("Mug intersects supporting object") # ignore = True # # if not ignore and collision.instancesIntersect(mathutils.Matrix.Identity(4), [cubeMug], mathutils.Matrix.Identity(4), [cubeRoomObj]): # print("Mug intersects room") # ignore = True # # if not ignore and not collision.instancesIntersect(mathutils.Matrix.Translation(mathutils.Vector((0, 0, -0.02))), [cubeMug], mathutils.Matrix.Identity(4), [cubeParentSupportObj]): # print("Mug not on table.") # ignore = True # # if not ignore: # print("No collision issues") # # if renderTeapots: # cubeScene.objects.unlink(cubeTeapot) # deleteObject(cubeTeapot) # # if renderMugs: # cubeScene.objects.unlink(cubeMug) # deleteObject(cubeMug) ## Environment map update if using Cycles. if not ignore and useBlender: if captureEnvMapFromBlender: envMapCoeffs = captureSceneEnvMap(scene, envMapTexture, roomInstanceNum, totalOffset.r.copy(), links, treeNodes, teapot, center, targetPosition, width, height, 2000, gtDir, train_i) placeCamera(scene.camera, -chAzGT.r[:].copy() * 180 / np.pi, chElGT.r[:].copy() * 180 / np.pi, chDistGT.r[0].copy(), center[:].copy() + targetPosition[:].copy()) if useBlender: envMapCoeffsRotated[:] = np.dot(light_probes.chSphericalHarmonicsZRotation(0), envMapCoeffs[[0, 3, 2, 1, 4, 5, 6, 7, 8]])[[0, 3, 2, 1, 4, 5, 6, 7, 8]] envMapCoeffsRotatedRel[:] = np.dot(light_probes.chSphericalHarmonicsZRotation(-chObjAzGT.r), envMapCoeffs[[0, 3, 2, 1, 4, 5, 6, 7, 8]])[[0, 3, 2, 1, 4, 5, 6, 7, 8]] else: envMapCoeffsRotated[:] = np.dot(light_probes.chSphericalHarmonicsZRotation(totalOffset), envMapCoeffs[[0, 3, 2, 1, 4, 5, 6, 7, 8]])[[0, 3, 2, 1, 4, 5, 6, 7, 8]] envMapCoeffsRotatedRel[:] = np.dot(light_probes.chSphericalHarmonicsZRotation(phiOffset), envMapCoeffs[[0, 3, 2, 1, 4, 5, 6, 7, 8]])[[0, 3, 2, 1, 4, 5, 6, 7, 8]] envMapCoeffsRotatedVals = envMapCoeffsRotated.r envMapCoeffsRotatedRelVals = envMapCoeffsRotatedRel.r if renderBlender and not ignore: bpy.context.screen.scene = scene scene.render.filepath = gtDir + 'images/im' + str(train_i) + '.jpeg' # image = np.array(imageio.imread(scene.render.filepath))[:, :, 0:3] # image[image > 1] = 1 # blenderRender = image # lin2srgb(blenderRender) # cv2.imwrite(gtDir + 'images/im' + str(train_i) + '.jpeg', 255 * blenderRender[:, :, [2, 1, 0]], [int(cv2.IMWRITE_JPEG_QUALITY), 100]) scene.render.image_settings.file_format = 'JPEG' scene.render.filepath = gtDir + 'images/im' + str(train_i) + '.jpeg' scene.render.layers[0].use = True nt = bpy.context.scene.node_tree nt.nodes['Render Layers'].layer = 'RenderLayer' scene.render.layers['RenderLayer'].use_pass_combined = True scene.render.layers['RenderLayer'].use_pass_environment = True backgroundNode = scene.world.node_tree.nodes['WorldBackground'] backgroundNode.inputs[1].default_value = 2 bpy.ops.render.render(write_still=True) image = np.array(imageio.imread(scene.render.filepath))[:, :, 0:3] blenderRender = srgb2lin(image) if captureEnvMapFromBlender and useOpenDR: blenderRenderGray = 0.3 * blenderRender[:, :, 0] + 0.59 * blenderRender[:, :, 1] + 0.11 * blenderRender[:, :, 2] # For some reason I need to correct average intensity in OpenDR a few times before it gets it right: rendererGTGray = 0.3 * rendererGT[:, :, 0].r[:] + 0.59 * rendererGT[:, :, 1].r[:] + 0.11 * rendererGT[:, :, 2].r[:] meanIntensityScale = np.mean(blenderRenderGray[vis_occluded]) / np.mean(rendererGTGray[vis_occluded]).copy() chAmbientIntensityGT[:] = chAmbientIntensityGT.r[:].copy() * meanIntensityScale rendererGTGray = 0.3 * rendererGT[:, :, 0].r[:] + 0.59 * rendererGT[:, :, 1].r[:] + 0.11 * rendererGT[:, :, 2].r[:] meanIntensityScale2 = np.mean(blenderRenderGray[vis_occluded]) / np.mean(rendererGTGray[vis_occluded]).copy() chAmbientIntensityGT[:] = chAmbientIntensityGT.r[:].copy() * meanIntensityScale2 rendererGTGray = 0.3 * rendererGT[:, :, 0].r[:] + 0.59 * rendererGT[:, :, 1].r[:] + 0.11 * rendererGT[:, :, 2].r[:] meanIntensityScale3 = np.mean(blenderRenderGray[vis_occluded]) / np.mean(rendererGTGray[vis_occluded]).copy() chAmbientIntensityGT[:] = chAmbientIntensityGT.r[:].copy() * meanIntensityScale3 rendererGTGray = 0.3 * rendererGT[:, :, 0].r[:] + 0.59 * rendererGT[:, :, 1].r[:] + 0.11 * rendererGT[:, :, 2].r[:] meanIntensityScale4 = np.mean(blenderRenderGray[vis_occluded]) / np.mean(rendererGTGray[vis_occluded]).copy() chAmbientIntensityGT[:] = chAmbientIntensityGT.r[:].copy() * meanIntensityScale4 rendererGTGray = 0.3 * rendererGT[:, :, 0].r[:] + 0.59 * rendererGT[:, :, 1].r[:] + 0.11 * rendererGT[:, :, 2].r[:] meanIntensityScale5 = np.mean(blenderRenderGray[vis_occluded]) / np.mean(rendererGTGray[vis_occluded]).copy() chAmbientIntensityGT[:] = chAmbientIntensityGT.r[:].copy() * meanIntensityScale5 if np.mean(rendererGTGray, axis=(0, 1)) < 0.01: ignore = True if useOpenDR: image = rendererGT.r[:].copy() lin2srgb(image) if not ignore: if useOpenDR: cv2.imwrite(gtDir + 'images_opendr/im' + str(train_i) + '.jpeg', 255 * image[:, :, [2, 1, 0]], [int(cv2.IMWRITE_JPEG_QUALITY), 100]) if renderTeapots: np.save(gtDir + 'masks_occlusion/mask' + str(train_i) + '.npy_occluded', vis_occluded) np.save(gtDir + 'masks_occlusion/mask' + str(train_i) + '.npy', vis_im) coordsTeapotX = coords[1][vis_im] coordsTeapotY = coords[0][vis_im] if renderMugs: np.save(gtDir + 'masks_occlusion/mask' + str(train_i) + '_mug_occluded.npy', vis_occluded_mug) np.save(gtDir + 'masks_occlusion/mask' + str(train_i) + '_mug.npy', vis_im_mug) coordsMugX = coords[1][vis_im_mug] coordsMugY = coords[0][vis_im_mug] #Add groundtruth to arrays trainAzsGT = chAzGTVals trainObjAzsGT = chObjAzGTVals trainElevsGT = chElGTVals trainLightAzsGT = chLightAzGTVals trainLightElevsGT = chLightElGTVals trainLightIntensitiesGT = chLightIntensityGTVals trainVColorGT = chVColorsGTVals lightCoeffs = envMapCoeffsRotatedVals[None, :].copy().squeeze() lightCoeffs = 0.3*lightCoeffs[:,0] + 0.59*lightCoeffs[:,1] + 0.11*lightCoeffs[:,2] trainLightCoefficientsGT = lightCoeffs lightCoeffsRel = envMapCoeffsRotatedRelVals[None, :].copy().squeeze() lightCoeffsRel = 0.3*lightCoeffsRel[:,0] + 0.59*lightCoeffsRel[:,1] + 0.11*lightCoeffsRel[:,2] trainLightCoefficientsGTRel = lightCoeffsRel trainAmbientIntensityGT = chAmbientIntensityGTVals trainEnvMapPhiOffsets = phiOffset trainScenes = sceneNumber trainTeapotIds = teapot_i trainEnvMaps = hdridx trainShapeModelCoeffsGT = chShapeParamsGTVals.copy() trainOcclusions = -1 trainIds = train_i trainTargetIndices = targetIndex trainObjDistGT = chObjDistGTVals trainObjRotationGT = chObjRotationGTVals trainObjDistMug = chObjDistMugVals trainObjRotationMug = chObjRotationMugVals trainObjAzMug = chObjAzMugVals trainVColorsMug = chVColorsMugVals trainMugElRel = mugCamElGT trainTeapotElRel = teapotCamElGT trainMugPosOffset = mugPosOffsetVals trainTeapotPosOffset = teapotPosOffsetVals trainBBMug = np.array([coordsMugX.min(),coordsMugX.max(),coordsMugY.min(),coordsMugY.max(),]) trainBBTeapot = np.array([coordsTeapotX.min(),coordsTeapotX.max(),coordsTeapotY.min(),coordsTeapotY.max(),]) trainTeapotPresent = renderTeapots trainMugPresent = renderMugs gtDataset.resize(gtDataset.shape[0] + 1, axis=0) gtDataset[-1] = np.array([(trainIds, trainAzsGT, trainObjAzsGT, trainElevsGT, trainLightAzsGT, trainLightElevsGT, trainLightIntensitiesGT, trainVColorGT, trainScenes, trainTeapotIds, trainEnvMaps, trainOcclusions, trainTargetIndices, trainLightCoefficientsGT, trainLightCoefficientsGTRel, trainAmbientIntensityGT, phiOffsetVals, trainShapeModelCoeffsGT, trainObjDistGT, trainObjRotationGT, trainObjDistMug, trainObjRotationMug, trainObjAzMug, trainVColorsMug, trainTeapotElRel, trainMugElRel, trainMugPosOffset, trainTeapotPosOffset, trainBBMug, trainBBTeapot, trainTeapotPresent, trainMugPresent )], dtype=gtDtype) gtDataFile.flush() train_i = train_i + 1 numTeapotTrain = numTeapotTrain + 1 if np.mod(train_i, 100) == 0: print("Generated " + str(train_i) + " GT instances.") print("Generating groundtruth. Iteration of " + str(range(int(trainSize/(lenScenes*len(hdrstorender)*len(renderTeapotsList))))) + " teapots") if renderFromPreviousGT: groundTruthFilename = 'groundtruth/' + previousGTPrefix + '/groundTruth.h5' gtDataFileToRender = h5py.File(groundTruthFilename, 'r') groundTruthToRender = gtDataFileToRender[previousGTPrefix] else: exit() currentScene = -1 currentTeapot = -1 currentTargetIndex = -1 teapot = None if renderFromPreviousGT: rangeGT = np.arange(len(groundTruthToRender)) else: rangeGT = np.arange(len(groundTruthToRender)) teapot_i = 0 # experimentPrefix = 'train4_occlusion_shapemodel_10k' # experimentDir = 'experiments/' + experimentPrefix + '/' # subsetToRender = np.load(experimentDir + 'test.npy')[np.arange(0,100)] subsetToRender = np.arange(len(rangeGT)) # subsetToRender = np.arange(len(rangeGT)) if useShapeModel: teapot_i = -1 # addObjectData(v, f_list, vc, vn, uv, haveTextures_list, textures_list, v_teapots[currentTeapotModel][0], f_list_teapots[currentTeapotModel][0], vc_teapots[currentTeapotModel][0], vn_teapots[currentTeapotModel][0], uv_teapots[currentTeapotModel][0], haveTextures_list_teapots[currentTeapotModel][0], textures_list_teapots[currentTeapotModel][0]) idxToRender = rangeGT[:][subsetToRender] sceneIdxsToRender = groundTruthToRender['trainScenes'][idxToRender] sortedSceneIndices = np.argsort(sceneIdxsToRender) sortedSceneAndTargetIdxs = np.arange(len(idxToRender)) for sceneIdx in np.unique(sceneIdxsToRender[sortedSceneIndices]): sortedTargetIdxs = np.argsort(groundTruthToRender['trainTargetIndices'][idxToRender][np.where(sceneIdxsToRender[sortedSceneIndices]==sceneIdx)]) sortedSceneAndTargetIdx = idxToRender[sortedSceneIndices][np.where(sceneIdxsToRender[sortedSceneIndices]==sceneIdx)][sortedTargetIdxs] sortedSceneAndTargetIdxs[np.where(sceneIdxsToRender[sortedSceneIndices]==sceneIdx)] = sortedSceneAndTargetIdx if renderFromPreviousGT: for gtIdx in sortedSceneAndTargetIdxs: if not replaceNewGroundtruth: if gtIdx in gtDataset['trainIds']: continue sceneNumber = groundTruthToRender['trainScenes'][gtIdx] # train_i = np.where(idxToRender==gtIdx)[0][0] train_i = gtIdx sceneIdx = scene_io_utils.getSceneIdx(sceneNumber, replaceableScenesFile) print("Rendering scene: " + str(sceneIdx)) sceneNumber, sceneFileName, instances, roomName, roomInstanceNum, targetIndicesScene, targetPositions = scene_io_utils.getSceneInformation(sceneIdx, replaceableScenesFile) sceneDicFile = 'data/scene' + str(sceneNumber) + '.pickle' if sceneIdx != currentScene: # v, f_list, vc, vn, uv, haveTextures_list, textures_list = sceneimport.loadSavedScene(sceneDicFile) import copy v2, f_list2, vc2, vn2, uv2, haveTextures_list2, textures_list2 = scene_io_utils.loadSavedScene(sceneDicFile, tex_srgb2lin) if sceneIdx != currentScene: if useBlender and not loadBlenderSceneFile: bpy.ops.wm.read_factory_settings() scene = scene_io_utils.loadBlenderScene(sceneIdx, replaceableScenesFile) scene_io_utils.setupScene(scene, roomInstanceNum, scene.world, scene.camera, width, height, 16, useCycles, True) scene.update() #Save barebones scene. elif useBlender and loadBlenderSceneFile: bpy.ops.wm.read_factory_settings() scene_io_utils.loadSceneBlendData(sceneIdx, replaceableScenesFile) scene = bpy.data.scenes['Main Scene'] # Configure scene if useBlender: if renderTeapots: targetModels = [] blender_teapots = [] teapots = [line.strip() for line in open('teapots.txt')] selection = [teapots[i] for i in renderTeapotsList] scene_io_utils.loadTargetsBlendData() for teapotIdx, teapotName in enumerate(selection): teapot = bpy.data.scenes[teapotName[0:63]].objects['teapotInstance' + str(renderTeapotsList[teapotIdx])] teapot.layers[1] = True teapot.layers[2] = True targetModels = targetModels + [teapotIdx] blender_teapots = blender_teapots + [teapot] if renderMugs: blender_mugs = [] selectionMugs = [mugs[i] for i in renderMugsList] scene_io_utils.loadMugsBlendData() for mugIdx, mugName in enumerate(selectionMugs): mug = bpy.data.scenes[mugName[0:63]].objects['mugInstance' + str(renderMugsList[mugIdx])] mug.layers[1] = True mug.layers[2] = True mugModels = mugModels + [mug] blender_mugs = blender_mugs + [mug] # setupSceneGroundtruth(scene, width, height, clip_start, 5000, 'CUDA', 'CUDA_MULTI_2') setupSceneGroundtruth(scene, width, height, clip_start, 5000) treeNodes = scene.world.node_tree links = treeNodes.links unlinkedObj = None envMapFilename = None targetIndex = groundTruthToRender['trainTargetIndices'][gtIdx] if sceneIdx != currentScene and not renderMugs and sceneIdx == 44 and sceneNumber == 114: if len(v) > 1: removeObjectData(1, v, f_list, vc, vn, uv, haveTextures_list, textures_list) if sceneIdx != currentScene or targetIndex != currentTargetIndex: targetPosition = targetPositions[np.where(targetIndex==np.array(targetIndicesScene))[0][0]] import copy v, f_list, vc, vn, uv, haveTextures_list, textures_list = copy.deepcopy(v2), copy.deepcopy(f_list2), copy.deepcopy(vc2), copy.deepcopy(vn2), copy.deepcopy(uv2), copy.deepcopy(haveTextures_list2), copy.deepcopy(textures_list2) removeObjectData(len(v) -1 - targetIndex, v, f_list, vc, vn, uv, haveTextures_list, textures_list) if sceneIdx != currentScene or targetIndex != currentTargetIndex: if useBlender: if unlinkedObj != None: scene.objects.link(unlinkedObj) try: unlinkedObj = scene.objects[str(targetIndex)] scene.objects.unlink(unlinkedObj) except: pass if sceneIdx == 44 and sceneNumber == 114: try: scene.objects.unlink(scene.objects['fc035f7d732166b97b6fd5468f603b31_cleaned']) except: pass teapot_i = groundTruthToRender['trainTeapotIds'][gtIdx] if useShapeModel: teapot_i = -1 if sceneIdx != currentScene or targetIndex != currentTargetIndex or teapot_i != currentTeapot: ##Destroy and create renderer if useOpenDR: rendererGT.makeCurrentContext() rendererGT.clear() contextdata.cleanupContext(contextdata.getContext()) if glMode == 'glfw': glfw.destroy_window(rendererGT.win) del rendererGT if generateTriplets: renderer.makeCurrentContext() renderer.clear() contextdata.cleanupContext(contextdata.getContext()) if glMode == 'glfw': glfw.destroy_window(renderer.win) del renderer if renderTeapots: currentTeapotModel = teapot_i center = center_teapots[teapot_i] if useShapeModel: center = smCenterGT UVs = smUVsGT vGT = smVerticesGT vnGT = smNormalsGT Faces = smFacesGT VColors = smVColorsGT UVs = smUVsGT HaveTextures = smHaveTexturesGT TexturesList = smTexturesListGT else: vGT, vnGT = v_teapots[currentTeapotModel][0], vn_teapots[currentTeapotModel][0] Faces = f_list_teapots[currentTeapotModel][0] VColors = vc_teapots[currentTeapotModel][0] UVs = uv_teapots[currentTeapotModel][0] HaveTextures = haveTextures_list_teapots[currentTeapotModel][0] TexturesList = textures_list_teapots[currentTeapotModel][0] vGT, vnGT, teapotPosOffset = transformObject(vGT, vnGT, chScaleGT, chObjAzGT, chObjDistGT, chObjRotationGT, targetPosition) if renderMugs: verticesMug, normalsMug, mugPosOffset = transformObject(v_mug, vn_mug, chScale, chObjAzMug + np.pi / 2, chObjDistMug, chObjRotationMug, targetPosition) VerticesMug = [verticesMug] NormalsMug = [normalsMug] FacesMug = [f_list_mug] VColorsMug = [vc_mug] UVsMug = [uv_mug] HaveTexturesMug = [haveTextures_list_mug] TexturesListMug = [textures_list_mug] else: VerticesMug = [] NormalsMug = [] FacesMug = [] VColorsMug = [] UVsMug = [] HaveTexturesMug = [] TexturesListMug = [] if renderTeapots: VerticesTeapot = [vGT] NormalsTeapot = [vnGT] FacesTeapot = [Faces] VColorsTeapot = [VColors] UVsTeapot = [UVs] HaveTexturesTeapot = [HaveTextures] TexturesListTeapot = [TexturesList] else: VerticesTeapot = [] NormalsTeapot = [] FacesTeapot = [] VColorsTeapot = [] UVsTeapot = [] HaveTexturesTeapot = [] TexturesListTeapot = [] # addObjectData(v, f_list, vc, vn, uv, haveTextures_list, textures_list, verticesMug, f_list_mug, vc_mug, normalsMug, uv_mug, haveTextures_list_mug, textures_list_mug) # addObjectData(v, f_list, vc, vn, uv, haveTextures_list, textures_list, vGT, Faces, VColors, vnGT, UVs, HaveTextures, TexturesList) rendererGT = createRendererGT(glMode, chAzGT, chElGT, chDistGT, center, VerticesTeapot + VerticesMug + v, VColorsTeapot + VColorsMug + vc, FacesTeapot + FacesMug + f_list, NormalsTeapot + NormalsMug + vn, light_colorGT, chComponentGT, chVColorsGT, targetPosition.copy(), chDisplacementGT, width, height, UVsTeapot + UVsMug + uv, HaveTexturesTeapot + HaveTexturesMug + haveTextures_list, TexturesListTeapot + TexturesListMug + textures_list, frustum, None) rendererGT.overdraw = False rendererGT.nsamples = 1 rendererGT.msaa = False rendererGT.initGL() rendererGT.initGLTexture() rendererGT.makeCurrentContext() if generateTriplets: if useShapeModel: center = smCenter UVs = smUVs v = smVertices vn = smNormals Faces = smFaces VColors = smVColors UVs = smUVs HaveTextures = smHaveTextures TexturesList = smTexturesList else: v, vn = v_teapots[currentTeapotModel][0], vn_teapots[currentTeapotModel][0] Faces = f_list_teapots[currentTeapotModel][0] VColors = vc_teapots[currentTeapotModel][0] UVs = uv_teapots[currentTeapotModel][0] HaveTextures = haveTextures_list_teapots[currentTeapotModel][0] TexturesList = textures_list_teapots[currentTeapotModel][0] v, vn, teapotPosOffset = transformObject(v, vn, chScale, chObjAz, chObjDist, chObjRotation, np.array([0, 0, 0])) if renderMugs: verticesMug, normalsMug, mugPosOffset = transformObject(v_mug, vn_mug, chScale, chObjAzMug + np.pi / 2, chObjDistMug, chObjRotationMug, np.array([0, 0, 0])) VerticesB = [v] + [verticesMug] NormalsB = [vn] + [normalsMug] FacesB = [Faces] + [f_list_mug] VColorsB = [VColors] + [vc_mug] UVsB = [UVs] + [uv_mug] HaveTexturesB = [HaveTextures] + [haveTextures_list_mug] TexturesListB = [TexturesList] + [textures_list_mug] renderer = createRendererTarget(glMode, chAz, chEl, chDist, center, VerticesB, VColorsB, FacesB, NormalsB, light_color, chComponent, chVColors, np.array([0, 0, 0]), chDisplacement, width, height, UVsB, HaveTexturesB, TexturesListB, frustum, None) else: renderer = createRendererTarget(glMode, chAz, chEl, chDist, smCenter, [v], [smVColors], [smFaces], [vn], light_color, chComponent, chVColors, 0, chDisplacement, width, height, [smUVs], [smHaveTextures], [smTexturesList], frustum, None) rendererGT.makeCurrentContext() ## Blender: Unlink and link new teapot. if useBlender: if renderTeapots: # if currentScene != -1 and currentTargetIndex != -1 and currentTeapot != -1 and teapot != None: if teapot.name in scene.objects: scene.objects.unlink(teapot) if useShapeModel: deleteInstance(teapot) if not useShapeModel: teapot = blender_teapots[currentTeapotModel] else: teapotMesh = createMeshFromData('teapotShapeModelMesh', chVerticesGT.r.tolist(), faces.astype(np.int32).tolist()) teapotMesh.layers[0] = True teapotMesh.layers[1] = True teapotMesh.pass_index = 1 targetGroup = bpy.data.groups.new('teapotShapeModelGroup') targetGroup.objects.link(teapotMesh) teapot = bpy.data.objects.new('teapotShapeModel', None) teapot.dupli_type = 'GROUP' teapot.dupli_group = targetGroup teapot.pass_index = 1 mat = makeMaterial('teapotMat', (0, 0, 0), (0, 0, 0), 1) setMaterial(teapotMesh, mat) # center = centerOfGeometry(teapot.dupli_group.objects, teapot.matrix_world) placeNewTarget(scene, teapot, targetPosition[:].copy()) teapot.layers[0] = True teapot.layers[1] = True original_matrix_world = teapot.matrix_world.copy() if renderMugs: if mug.name in scene.objects: scene.objects.unlink(mug) # deleteInstance(mug) mug = blender_mugs[currentTeapotModel] placeNewTarget(scene, mug, targetPosition[:].copy()) mug.layers[0] = True mug.layers[1] = True original_matrix_world_mug = mug.matrix_world.copy() hdridx = groundTruthToRender['trainEnvMaps'][gtIdx] envMapFilename = "" for hdrFile, hdrValues in hdritems: if hdridx == hdrValues[0]: envMapCoeffs[:] = hdrValues[1] envMapFilename = hdrFile # updateEnviornmentMap(envMapFilename, scene) envMapTexture = np.array(imageio.imread(envMapFilename))[:,:,0:3] break assert(envMapFilename != "") print("Render " + str(gtIdx) + "of " + str(len(groundTruthToRender))) ignore = False # chAmbientIntensityGT[:] = groundTruthToRender['trainAmbientIntensityGT'][gtIdx] chAmbientIntensityGT[:] = groundTruthToRender['trainAmbientIntensityGT'][gtIdx] phiOffset[:] = groundTruthToRender['trainEnvMapPhiOffsets'][gtIdx] chObjAzGT[:] = groundTruthToRender['trainObjAzsGT'][gtIdx] chAzGT[:] = groundTruthToRender['trainAzsGT'][gtIdx] chElGT[:] = groundTruthToRender['trainElevsGT'][gtIdx] chLightAzGT[:] = groundTruthToRender['trainLightAzsGT'][gtIdx] chLightElGT[:] = groundTruthToRender['trainLightElevsGT'][gtIdx] # chLightIntensityGT[:] = np.random.uniform(5,10, 1) chVColorsGT[:] = groundTruthToRender['trainVColorGT'][gtIdx] try: chShapeParamsGT[:] = groundTruthToRender['trainShapeModelCoeffsGT'][gtIdx] except: chShapeParamsGT[:] = np.random.randn(latentDim) try: chObjDistGT[:] = groundTruthToRender['trainObjDistGT'][gtIdx] except: chObjDistGT[:] = 0 try: chObjRotationGT[:] = groundTruthToRender['trainObjRotationGT'][gtIdx] except: chObjRotationGT[:] = 0 if renderMugs: chObjDistMug[:] = groundTruthToRender['trainObjDistMug'][gtIdx] chObjRotationMug[:] = groundTruthToRender['trainObjRotationMug'][gtIdx] chObjAzMug[:] = groundTruthToRender['trainObjAzMug'][gtIdx] chVColorsMug[:] = groundTruthToRender['trainVColorsMug'][gtIdx] if captureEnvMapFromBlender and not ignore and useBlender: envMapCoeffs = captureSceneEnvMap(scene, envMapTexture, roomInstanceNum, totalOffset.r.copy(), links, treeNodes, teapot, center, targetPosition, width, height, 2000, gtDir, train_i) if useBlender and captureEnvMapFromBlender: envMapCoeffsRotated[:] = np.dot(light_probes.chSphericalHarmonicsZRotation(0), envMapCoeffs[[0,3,2,1,4,5,6,7,8]])[[0,3,2,1,4,5,6,7,8]] envMapCoeffsRotatedRel[:] = np.dot(light_probes.chSphericalHarmonicsZRotation(-chObjAzGT.r), envMapCoeffs[[0,3,2,1,4,5,6,7,8]])[[0,3,2,1,4,5,6,7,8]] else: envMapCoeffsRotated[:] = np.dot(light_probes.chSphericalHarmonicsZRotation(totalOffset), envMapCoeffs[[0,3,2,1,4,5,6,7,8]])[[0,3,2,1,4,5,6,7,8]] envMapCoeffsRotatedRel[:] = np.dot(light_probes.chSphericalHarmonicsZRotation(phiOffset), envMapCoeffs[[0,3,2,1,4,5,6,7,8]])[[0,3,2,1,4,5,6,7,8]] envMapCoeffsRotatedVals = envMapCoeffsRotated.r envMapCoeffsRotatedRelVals = envMapCoeffsRotatedRel.r if useOpenDR and not ignore: if renderTeapots: occlusion = getOcclusionFraction(rendererGT, id=teapotSceneIndex) vis_occluded = np.array(rendererGT.indices_image == teapotSceneIndex + 1).copy().astype(np.bool) vis_im = np.array(rendererGT.image_mesh_bool([teapotSceneIndex])).copy().astype(np.bool) if renderBlender and useBlender and not ignore: placeCamera(scene.camera, -chAzGT.r[:].copy() * 180 / np.pi, chElGT.r[:].copy() * 180 / np.pi, chDistGT.r[0].copy(), center[:].copy() + targetPosition[:].copy()) if renderTeapots: azimuthRot = mathutils.Matrix.Rotation(chObjAzGT.r[:].copy(), 4, 'Z') teapot.matrix_world = mathutils.Matrix.Translation( original_matrix_world.to_translation() + mathutils.Vector(teapotPosOffset.r)) * azimuthRot * ( mathutils.Matrix.Translation(-original_matrix_world.to_translation())) * original_matrix_world setObjectDiffuseColor(teapot, chVColorsGT.r.copy()) if useShapeModel: mesh = teapot.dupli_group.objects[0] for vertex_i, vertex in enumerate(mesh.data.vertices): vertex.co = mathutils.Vector(chVerticesGT.r[vertex_i]) if renderMugs: setObjectDiffuseColor(mug, chVColorsMug.r.copy()) azimuthRotMug = mathutils.Matrix.Rotation(chObjAzMug.r[:].copy() - np.pi / 2, 4, 'Z') mug.matrix_world = mathutils.Matrix.Translation( original_matrix_world_mug.to_translation() + mathutils.Vector(mugPosOffset.r)) * azimuthRotMug * ( mathutils.Matrix.Translation(-original_matrix_world_mug.to_translation())) * original_matrix_world_mug scene.update() scene.render.filepath = gtDir + 'images/im' + str(train_i) + '.jpeg' bpy.ops.render.render(write_still=True) # image = np.array(imageio.imread(scene.render.filepath))[:, :, 0:3] # image[image > 1] = 1 # blenderRender = image # lin2srgb(blenderRender) # cv2.imwrite(gtDir + 'images/im' + str(train_i) + '.jpeg', 255 * blenderRender[:, :, [2, 1, 0]], [int(cv2.IMWRITE_JPEG_QUALITY), 100]) image = np.array(imageio.imread(scene.render.filepath))[:, :, 0:3] blenderRender = image blenderRenderGray = 0.3*blenderRender[:,:,0] + 0.59*blenderRender[:,:,1] + 0.11*blenderRender[:,:,2] #For some reason I need to correct average intensity in OpenDR a few times before it gets it right: # rendererGTGray = 0.3*rendererGT[:,:,0].r[:] + 0.59*rendererGT[:,:,1].r[:] + 0.11*rendererGT[:,:,2].r[:] # meanIntensityScale = np.mean(blenderRenderGray[vis_occluded])/np.mean(rendererGTGray[vis_occluded]).copy() # chAmbientIntensityGT[:] = chAmbientIntensityGT.r[:].copy()*meanIntensityScale # rendererGTGray = 0.3*rendererGT[:,:,0].r[:] + 0.59*rendererGT[:,:,1].r[:] + 0.11*rendererGT[:,:,2].r[:] # meanIntensityScale2 = np.mean(blenderRenderGray[vis_occluded])/np.mean(rendererGTGray[vis_occluded]).copy() # chAmbientIntensityGT[:] = chAmbientIntensityGT.r[:].copy()*meanIntensityScale2 # rendererGTGray = 0.3*rendererGT[:,:,0].r[:] + 0.59*rendererGT[:,:,1].r[:] + 0.11*rendererGT[:,:,2].r[:] # meanIntensityScale3 = np.mean(blenderRenderGray[vis_occluded])/np.mean(rendererGTGray[vis_occluded]).copy() # chAmbientIntensityGT[:] = chAmbientIntensityGT.r[:].copy()*meanIntensityScale3 # rendererGTGray = 0.3*rendererGT[:,:,0].r[:] + 0.59*rendererGT[:,:,1].r[:] + 0.11*rendererGT[:,:,2].r[:] # meanIntensityScale4 = np.mean(blenderRenderGray[vis_occluded])/np.mean(rendererGTGray[vis_occluded]).copy() # chAmbientIntensityGT[:] = chAmbientIntensityGT.r[:].copy()*meanIntensityScale4 # rendererGTGray = 0.3*rendererGT[:,:,0].r[:] + 0.59*rendererGT[:,:,1].r[:] + 0.11*rendererGT[:,:,2].r[:] # meanIntensityScale5 = np.mean(blenderRenderGray[vis_occluded])/np.mean(rendererGTGray[vis_occluded]).copy() # chAmbientIntensityGT[:] = chAmbientIntensityGT.r[:].copy()*meanIntensityScale5 # lin2srgb(blenderRender) if useOpenDR: image = rendererGT.r[:].copy() lin2srgb(image) if not ignore: if generateTriplets: lightCoeffsRel = envMapCoeffsRotatedRel.r[None, :].copy().squeeze() lightCoeffsRel = 0.3 * lightCoeffsRel[:, 0] + 0.59 * lightCoeffsRel[:, 1] + 0.11 * lightCoeffsRel[:, 2] chLightSHCoeffs[:] = lightCoeffsRel * chAmbientIntensityGT.r chObjAz[:] = 0 chAz[:] = chAzRelGT.r + np.random.choice([-1,1]) * np.random.uniform(0,10) * np.pi / 180 chEl[:] = chElGT.r chVColors[:] = chVColorsGT.r chShapeParams[:] = chShapeParamsGT.r renderer.makeCurrentContext() if useOpenDR: cv2.imwrite(gtDir + 'triplets1/im' + str(train_i) + '.jpeg' , 255*lin2srgb(renderer.r.copy())[:,:,[2,1,0]], [int(cv2.IMWRITE_JPEG_QUALITY), 100]) chAz[:] = chAzRelGT.r + np.random.choice([-1,1]) * np.random.uniform(10,40) * np.pi / 180 if useOpenDR: cv2.imwrite(gtDir + 'triplets2/im' + str(train_i) + '.jpeg', 255 * lin2srgb(renderer.r.copy())[:, :, [2, 1, 0]], [int(cv2.IMWRITE_JPEG_QUALITY), 100]) rendererGT.makeCurrentContext() if useBlender and renderBlender: cv2.imwrite(gtDir + 'images/im' + str(train_i) + '.jpeg' , 255*blenderRender[:,:,[2,1,0]], [int(cv2.IMWRITE_JPEG_QUALITY), 100]) if useOpenDR: cv2.imwrite(gtDir + 'images_opendr/im' + str(train_i) + '.jpeg' , 255*image[:,:,[2,1,0]], [int(cv2.IMWRITE_JPEG_QUALITY), 100]) # cv2.imwrite(gtDir + 'images_opendr/im' + str(train_i) + '.jpeg' , 255*image[:,:,[2,1,0]], [int(cv2.IMWRITE_JPEG_QUALITY), 100]) if useOpenDR: np.save(gtDir + 'masks_occlusion/mask' + str(train_i)+ '.npy', vis_occluded) #Add groundtruth to arrays trainAzsGT = chAzGT.r trainObjAzsGT = chObjAzGT.r trainElevsGT = chElGT.r trainLightAzsGT = chLightAzGT.r trainLightElevsGT = chLightElGT.r trainLightIntensitiesGT = groundTruthToRender['trainLightIntensitiesGT'][gtIdx] trainVColorGT = chVColorsGT.r lightCoeffs = envMapCoeffsRotated.r[None, :].copy().squeeze() lightCoeffs = 0.3*lightCoeffs[:,0] + 0.59*lightCoeffs[:,1] + 0.11*lightCoeffs[:,2] trainLightCoefficientsGT = lightCoeffs lightCoeffsRel = envMapCoeffsRotatedRel.r[None, :].copy().squeeze() lightCoeffsRel = 0.3*lightCoeffsRel[:,0] + 0.59*lightCoeffsRel[:,1] + 0.11*lightCoeffsRel[:,2] trainLightCoefficientsGTRel = lightCoeffsRel trainAmbientIntensityGT = chAmbientIntensityGT.r trainEnvMapPhiOffsets = phiOffset trainScenes = sceneNumber trainTeapotIds = teapot_i trainEnvMaps = hdridx trainShapeModelCoeffsGT = chShapeParamsGT.r.copy() trainOcclusions = occlusion trainIds = train_i trainTargetIndices = targetIndex trainObjDistGT = chObjDistGT.r trainObjRotationGT = chObjRotationGT.r trainObjDistMug = chObjDistMug.r trainObjRotationMug = chObjRotationMug.r trainObjAzMug = chObjAzMug.r trainVColorsMug = chVColorsMug.r try: trainMugElRel = groundTruthToRender['trainMugElRel'][gtIdx] except: trainMugElRel = -1 try: trainTeapotElRel = groundTruthToRender['trainTeapotElRel'][gtIdx] except: trainTeapotElRel = -1 try: trainMugPosOffset = groundTruthToRender['trainMugPosOffset'][gtIdx] except: trainMugPosOffset = np.array([0,0,0]) try: trainTeapotPosOffset = groundTruthToRender['trainTeapotPosOffset'][gtIdx] except: trainTeapotPosOffset = np.array([0,0,0]) try: trainBBMug = groundTruthToRender['trainBBMug'][gtIdx] except: trainBBMug = np.array([0,0,0,0]) try: trainBBTeapot = groundTruthToRender['trainBBTeapot'][gtIdx] except: trainBBTeapot = np.array([0,0,0,0]) trainTeapotPresent = renderTeapots trainMugPresent = renderMugs gtDataset.resize(gtDataset.shape[0] + 1, axis=0) gtDataset[-1] = np.array([(trainIds, trainAzsGT, trainObjAzsGT, trainElevsGT, trainLightAzsGT, trainLightElevsGT, trainLightIntensitiesGT, trainVColorGT, trainScenes, trainTeapotIds, trainEnvMaps, trainOcclusions, trainTargetIndices, trainLightCoefficientsGT, trainLightCoefficientsGTRel, trainAmbientIntensityGT, trainEnvMapPhiOffsets, trainShapeModelCoeffsGT, trainObjDistGT, trainObjRotationGT, trainObjDistMug, trainObjRotationMug, trainObjAzMug, trainVColorsMug, trainTeapotElRel, trainMugElRel, trainMugPosOffset, trainTeapotPosOffset, trainBBMug, trainBBTeapot, trainTeapotPresent, trainMugPresent )], dtype=gtDtype) train_i = train_i + 1 currentScene = sceneIdx currentTargetIndex = targetIndex currentTeapot = teapot_i gtDataFile.flush() # np.savetxt(gtDir + 'data.txt',np.array(np.hstack([trainIds[:,None], trainAzsGT[:,None], trainObjAzsGT[:,None], trainElevsGT[:,None], phiOffsets[:,None], trainOcclusions[:,None]])), fmt="%g") gtDataFile.close() gtDataFileToRender.close()
106,833
48.232258
1,055
py
inversegraphics
inversegraphics-master/utils.py
import numpy as np import os import skimage import skimage.io import h5py import ipdb import scipy.spatial.distance import image_processing import matplotlib __author__ = 'pol' import recognition_models import pickle def joinExperiments(range1, range2, testSet1,methodsPred1,testOcclusions1,testPrefixBase1,parameterRecognitionModels1,azimuths1,elevations1,vColors1,lightCoeffs1,likelihoods1,shapeParams1,segmentations1, approxProjections1, approxProjectionsGT1, testSet2,methodsPred2,testOcclusions2,testPrefixBase2,parameterRecognitionModels2,azimuths2,elevations2,vColors2,lightCoeffs2,likelihoods2,shapeParams2,segmentations2, approxProjections2, approxProjectionsGT2): testSet = np.append(testSet1, testSet2) parameterRecognitionModels = [parameterRecognitionModels1] + [parameterRecognitionModels2] testPrefixBase = [testPrefixBase1] + [testPrefixBase2] methodsPred = methodsPred1 testOcclusions = np.append(testOcclusions1, testOcclusions2) azimuths = [] elevations = [] vColors = [] lightCoeffs = [] shapeParams = [] likelihoods = [] segmentations = [] approxProjections = [] approxProjectionsGT = [] for method in range(len(methodsPred)): if azimuths1[method] is not None and azimuths2[method] is not None: azimuths = azimuths + [np.append(azimuths1[method][range1], azimuths2[method][range2])] else: azimuths = azimuths + [None] if elevations1[method] is not None and elevations2[method] is not None: elevations = elevations + [np.append(elevations1[method][range1], elevations2[method][range2])] else: elevations = elevations + [None] if vColors1[method] is not None and vColors2[method] is not None: vColors = vColors + [np.vstack([vColors1[method][range1], vColors2[method][range2]])] else: vColors = vColors + [None] if lightCoeffs1[method] is not None and lightCoeffs2[method] is not None: lightCoeffs = lightCoeffs + [np.vstack([lightCoeffs1[method][range1], lightCoeffs2[method][range2]])] else: lightCoeffs = lightCoeffs + [None] if shapeParams1[method] is not None and shapeParams2[method] is not None: shapeParams = shapeParams + [np.vstack([shapeParams1[method][range1], shapeParams2[method][range2]])] else: shapeParams = shapeParams + [None] if likelihoods1[method] is not None and likelihoods2[method] is not None: likelihoods = likelihoods + [np.append(likelihoods1[method][range1], likelihoods2[method][range2])] else: likelihoods = likelihoods + [None] if segmentations1[method] is not None and segmentations2[method] is not None: segmentations = segmentations + [np.vstack([segmentations1[method][range1], segmentations2[method][range2]])] else: segmentations = segmentations + [None] if approxProjections1 is not None and approxProjections2 is not None: if approxProjections1[method] is not None and approxProjections2[method] is not None: approxProjections = approxProjections + [np.vstack([approxProjections1[method][range1], approxProjections2[method][range2]])] else: approxProjections = approxProjections + [None] if approxProjectionsGT1 is not None and approxProjectionsGT2 is not None: if approxProjectionsGT1[method] is not None and approxProjectionsGT2[method] is not None: approxProjectionsGT = approxProjectionsGT + [np.vstack([approxProjectionsGT1[method][range1], approxProjectionsGT2[method][range2]])] else: approxProjectionsGT = approxProjectionsGT + [None] return testSet, parameterRecognitionModels, testPrefixBase, methodsPred, testOcclusions, azimuths, elevations, vColors, lightCoeffs, shapeParams, likelihoods, segmentations, approxProjections, approxProjectionsGT def latexify(fig_width=None, fig_height=None, columns=1): """Set up matplotlib's RC params for LaTeX plotting. Call this before plotting a figue. Parameters ---------- fig_width : float, optional, inches fig_height : float, optional, inches columns : {1, 2} """ # code adapted from http://www.scipy.org/Cookbook/Matplotlib/LaTeX_Examples # Width and max height in inches for IEEE journals taken from # computer.org/cms/Computer.org/Journal%20templates/transactions_art_guide.pdf assert(columns in [1,2]) if fig_width is None: fig_width = 3.39 if columns==1 else 6.9 # width in inches if fig_height is None: golden_mean = (np.sqrt(5)-1.0)/2.0 # Aesthetic ratio fig_height = fig_width*golden_mean # height in inches MAX_HEIGHT_INCHES = 8.0 if fig_height > MAX_HEIGHT_INCHES: print("WARNING: fig_height too large:" + fig_height + "so will reduce to" + MAX_HEIGHT_INCHES + "inches.") fig_height = MAX_HEIGHT_INCHES params = {'backend': 'pdf', 'axes.labelsize': 12, # fontsize for x and y labels (was 10) 'axes.titlesize': 12, 'font.size': 12, # was 10 'legend.fontsize': 12, # was 10 'xtick.labelsize': 14, 'ytick.labelsize': 14, 'text.usetex': True, 'figure.figsize': [fig_width,fig_height], 'font.family': 'serif', 'lines.linewidth':2 } matplotlib.rcParams.update(params) def saveScatterPlots(resultDir, testOcclusions, useShapeModel, errorsPosePred, errorsPoseFitted,errorsLightCoeffsC,errorsFittedLightCoeffsC,errorsEnvMap,errorsFittedEnvMap,errorsLightCoeffs,errorsFittedLightCoeffs,errorsShapeParams,errorsFittedShapeParams,errorsShapeVertices,errorsFittedShapeVertices,errorsVColorsE,errorsFittedVColorsE,errorsVColorsC,errorsFittedVColorsC, errorsVColorsS,errorsFittedVColorsS): latexify(columns=2) directory = resultDir + 'pred-azimuth-errors_fitted-azimuth-error' fig = plt.figure() ax = fig.add_subplot(111, aspect='equal') scat = ax.scatter(abs(errorsPosePred[0]), abs(errorsPoseFitted[0]), s=20, vmin=0, vmax=100, c=testOcclusions * 100, cmap=matplotlib.cm.plasma) cbar = fig.colorbar(scat, ticks=[0, 50, 100]) cbar.ax.set_yticklabels(['0%', '50%', '100%']) # vertically oriented colorbar ax.set_xlabel('Predicted azimuth errors') ax.set_ylabel('Fitted azimuth errors') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, max(x2,y2))) ax.set_ylim((0, max(x2,y2))) ax.plot([0, max(x2,y2)], [0, max(x2,y2)], ls="--", c=".3") for test_i in range(len(errorsPosePred[0])): if abs(errorsPoseFitted[0][test_i]) > abs(errorsPosePred[0][test_i]) + 10: plt.annotate( str(test_i), xy=(abs(errorsPosePred[0][test_i]), abs(errorsPoseFitted[0][test_i])), xytext=(0, 15), textcoords='offset points', ha='right', va='bottom', arrowprops=dict(arrowstyle='->', connectionstyle='arc3,rad=0'), size=8) # bbox=dict(boxstyle='round,pad=0.5', fc='yellow', alpha=0.5), ax.set_title('Recognition vs fitted azimuth errors') fig.savefig(directory + '-performance-scatter.pdf', bbox_inches='tight') plt.close(fig) directory = resultDir + 'pred-elevation-errors_fitted-elevation-error' fig = plt.figure() ax = fig.add_subplot(111, aspect='equal') scat = ax.scatter(abs(errorsPosePred[1]), abs(errorsPoseFitted[1]), s=20, vmin=0, vmax=100, c=testOcclusions * 100, cmap=matplotlib.cm.plasma) for test_i in range(len(errorsPosePred[1])): if abs(errorsPoseFitted[1][test_i]) > abs(errorsPosePred[1][test_i]) + 10: plt.annotate( str(test_i), xy=(abs(errorsPosePred[1][test_i]), abs(errorsPoseFitted[1][test_i])), xytext=(15, 15), textcoords='offset points', ha='right', va='bottom', arrowprops=dict(arrowstyle='->', connectionstyle='arc3,rad=0'), size=8) # bbox=dict(boxstyle='round,pad=0.5', fc='yellow', alpha=0.5), cbar = fig.colorbar(scat, ticks=[0, 50, 100]) cbar.ax.set_yticklabels(['0%', '50%', '100%']) # vertically oriented colorbar ax.set_xlabel('Predicted elevation errors') ax.set_ylabel('Fitted elevation errors') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, max(x2,y2))) ax.set_ylim((0, max(x2,y2))) ax.plot([0, max(x2,y2)], [0, max(x2,y2)], ls="--", c=".3") ax.set_title('Recognition vs fitted azimuth errors') fig.savefig(directory + '-performance-scatter.pdf', bbox_inches='tight') plt.close(fig) directory = resultDir + 'shcoeffsserrorsC_fitted-shcoeffsserrorsC' fig = plt.figure() ax = fig.add_subplot(111, aspect='equal') scat = ax.scatter(np.mean(errorsLightCoeffsC, axis=1), np.mean(errorsFittedLightCoeffsC, axis=1), s=20, vmin=0, vmax=100, c=testOcclusions * 100, cmap=matplotlib.cm.plasma) for test_i in range(len(errorsLightCoeffsC)): if np.mean(errorsFittedLightCoeffsC, axis=1)[test_i] > np.mean(errorsLightCoeffsC, axis=1)[test_i] + 0.1: plt.annotate( str(test_i), xy=(np.mean(errorsLightCoeffsC, axis=1)[test_i], np.mean(errorsFittedLightCoeffsC, axis=1)[test_i]), xytext=(0, 15), textcoords='offset points', ha='right', va='bottom', arrowprops=dict(arrowstyle='->', connectionstyle='arc3,rad=0'), size=8) # bbox=dict(boxstyle='round,pad=0.5', fc='yellow', alpha=0.5), cbar = fig.colorbar(scat, ticks=[0, 50, 100]) cbar.ax.set_yticklabels(['0%', '50%', '100%']) # vertically oriented colorbar ax.set_xlabel('Predicted SH coefficients errors') ax.set_ylabel('Fitted SH coefficients errors') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, max(x2,y2))) ax.set_ylim((0, max(x2,y2))) ax.plot([0, max(x2,y2)], [0, max(x2,y2)], ls="--", c=".3") ax.set_title('Recognition vs fitted SH coefficients errors') fig.savefig(directory + '-performance-scatter.pdf', bbox_inches='tight') plt.close(fig) if errorsEnvMap is not None and errorsFittedEnvMap is not None: directory = resultDir + 'shcoeffsserrorsC_fitted-shEnvMap' fig = plt.figure() ax = fig.add_subplot(111, aspect='equal') scat = ax.scatter(errorsEnvMap, errorsFittedEnvMap, s=20, vmin=0, vmax=100, c=testOcclusions * 100, cmap=matplotlib.cm.plasma) cbar = fig.colorbar(scat, ticks=[0, 50, 100]) cbar.ax.set_yticklabels(['0%', '50%', '100%']) # vertically oriented colorbar ax.set_xlabel('Predicted SH coefficients errors') ax.set_ylabel('Fitted SH coefficients errors') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, max(x2,y2))) ax.set_ylim((0, max(x2,y2))) ax.plot([0, max(x2,y2)], [0, max(x2,y2)], ls="--", c=".3") ax.set_title('Recognition vs fitted SH Environment map errors') fig.savefig(directory + '-performance-scatter.pdf', bbox_inches='tight') plt.close(fig) directory = resultDir + 'shcoeffsserrors_fitted-shcoeffsserrors' # Show scatter correlations with occlusions. fig = plt.figure() ax = fig.add_subplot(111, aspect='equal') scat = ax.scatter(np.mean(errorsLightCoeffs, axis=1), np.mean(errorsFittedLightCoeffs, axis=1), s=20, vmin=0, vmax=100, c=testOcclusions * 100, cmap=matplotlib.cm.plasma) cbar = fig.colorbar(scat, ticks=[0, 50, 100]) cbar.ax.set_yticklabels(['0%', '50%', '100%']) # vertically oriented colorbar ax.set_xlabel('Predicted SH coefficients errors') ax.set_ylabel('Fitted SH coefficients errors') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, max(x2,y2))) ax.set_ylim((0, max(x2,y2))) ax.plot([0, max(x2,y2)], [0, max(x2,y2)], ls="--", c=".3") ax.set_title('Recognition vs fitted SH coefficients errors') fig.savefig(directory + '-performance-scatter.pdf', bbox_inches='tight') plt.close(fig) if useShapeModel: directory = resultDir + 'shapeparamserrors_fitted-shapeparamserrors' # Show scatter correlations with occlusions. fig = plt.figure() ax = fig.add_subplot(111, aspect='equal') scat = ax.scatter(np.mean(errorsShapeParams, axis=1), np.mean(errorsFittedShapeParams, axis=1), s=20, vmin=0, vmax=100, c=testOcclusions * 100, cmap=matplotlib.cm.plasma) cbar = fig.colorbar(scat, ticks=[0, 50, 100]) cbar.ax.set_yticklabels(['0%', '50%', '100%']) # vertically oriented colorbar ax.set_xlabel('Predicted shape parameters errors') ax.set_ylabel('Fitted shape parameters errors') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, max(x2,y2))) ax.set_ylim((0, max(x2,y2))) ax.plot([0, max(x2,y2)], [0, max(x2,y2)], ls="--", c=".3") ax.set_title('Recognition vs fitted Shape parameters errors') fig.savefig(directory + '-performance-scatter.pdf', bbox_inches='tight') plt.close(fig) if errorsShapeVertices is not None and errorsFittedShapeVertices is not None: directory = resultDir + 'shapeverticeserrors_fitted-shapeverticesserrors' # Show scatter correlations with occlusions. fig = plt.figure() ax = fig.add_subplot(111, aspect='equal') scat = ax.scatter(errorsShapeVertices, errorsFittedShapeVertices, s=20, vmin=0, vmax=100, c=testOcclusions * 100, cmap=matplotlib.cm.plasma) for test_i in range(len(errorsLightCoeffsC)): if errorsFittedShapeVertices[test_i] > errorsShapeVertices[test_i] + 0.75: plt.annotate( str(test_i), xy=(errorsShapeVertices[test_i], errorsFittedShapeVertices[test_i]), xytext=(15, 15), textcoords='offset points', ha='right', va='bottom', arrowprops=dict(arrowstyle='->', connectionstyle='arc3,rad=0'), size=8) cbar = fig.colorbar(scat, ticks=[0, 50, 100]) cbar.ax.set_yticklabels(['0%', '50%', '100%']) # vertically oriented colorbar ax.set_xlabel('Predicted shape vertices errors') ax.set_ylabel('Fitted shape vertices errors') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, max(x2,y2))) ax.set_ylim((0, max(x2,y2))) ax.plot([0, max(x2,y2)], [0, max(x2,y2)], ls="--", c=".3") ax.set_title('Recognition vs fitted Shape parameters errors') fig.savefig(directory + '-performance-scatter.pdf', bbox_inches='tight') plt.close(fig) directory = resultDir + 'vColorsE_fitted-vColorsE' # Show scatter correlations with occlusions. fig = plt.figure() ax = fig.add_subplot(111, aspect='equal') scat = ax.scatter(errorsVColorsE, errorsFittedVColorsE, s=20, vmin=0, vmax=100, c=testOcclusions * 100, cmap=matplotlib.cm.plasma) cbar = fig.colorbar(scat, ticks=[0, 50, 100]) cbar.ax.set_yticklabels(['0%', '50%', '100%']) # vertically oriented colorbar ax.set_xlabel('Predicted VColor E coefficients errors') ax.set_ylabel('Fitted VColor E coefficients errors') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, max(x2,y2))) ax.set_ylim((0, max(x2,y2))) ax.plot([0, max(x2,y2)], [0, max(x2,y2)], ls="--", c=".3") ax.set_title('Recognition vs fitted vertex color errors') fig.savefig(directory + '-performance-scatter.pdf', bbox_inches='tight') plt.close(fig) directory = resultDir + 'vColorsC_fitted-vColorsC' fig = plt.figure() ax = fig.add_subplot(111, aspect='equal') scat = ax.scatter(errorsVColorsC, errorsFittedVColorsC, s=20, vmin=0, vmax=100, c=testOcclusions * 100, cmap=matplotlib.cm.plasma) cbar = fig.colorbar(scat, ticks=[0, 50, 100]) cbar.ax.set_yticklabels(['0%', '50%', '100%']) # vertically oriented colorbar ax.set_xlabel('Predicted vertex color errors') ax.set_ylabel('Fitted vertex color errors') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, max(x2,y2))) ax.set_ylim((0, max(x2,y2))) ax.plot([0, max(x2,y2)], [0, max(x2,y2)], ls="--", c=".3") ax.set_title('Recognition vs fitted vertex color errors') fig.savefig(directory + '-performance-scatter.pdf', bbox_inches='tight') plt.close(fig) directory = resultDir + 'vColorsS_fitted-vColorsS' fig = plt.figure() ax = fig.add_subplot(111, aspect='equal') scat = ax.scatter(errorsVColorsS, errorsFittedVColorsS, s=20, vmin=0, vmax=100, c=testOcclusions * 100, cmap=matplotlib.cm.plasma) for test_i in range(len(errorsVColorsS)): if errorsFittedVColorsS[test_i] > errorsVColorsS[test_i] + 0.1: plt.annotate( str(test_i), xy=(errorsVColorsS[test_i], errorsFittedVColorsS[test_i]), xytext=(15, 15), textcoords='offset points', ha='right', va='bottom', arrowprops=dict(arrowstyle='->', connectionstyle='arc3,rad=0'), size=8) cbar = fig.colorbar(scat, ticks=[0, 50, 100]) cbar.ax.set_yticklabels(['0%', '50%', '100%']) # vertically oriented colorbar ax.set_xlabel('Predicted VColor coefficients errors') ax.set_ylabel('Fitted VColor coefficients errors') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, max(x2, y2))) ax.set_ylim((0, max(x2, y2))) ax.plot([0, max(x2, y2)], [0, max(x2, y2)], ls="--", c=".3") ax.set_title('Recognition vs fitted vertex color errors') fig.savefig(directory + '-performance-scatter.pdf', bbox_inches='tight') plt.close(fig) def saveScatterPlotsMethodFit(methodFitNum, resultDir, testOcclusions, useShapeModel, errorsPosePred, errorsPoseFitted,errorsLightCoeffsC,errorsFittedLightCoeffsC,errorsEnvMap,errorsFittedEnvMap,errorsLightCoeffs,errorsFittedLightCoeffs,errorsShapeParams,errorsFittedShapeParams,errorsShapeVertices,errorsFittedShapeVertices,errorsVColorsE,errorsFittedVColorsE,errorsVColorsC,errorsFittedVColorsC, errorsVColorsS,errorsFittedVColorsS): latexify(columns=2) directory = resultDir + 'pred-azimuth-errors_fitted-azimuth-error-' + str(methodFitNum) + '-' fig = plt.figure() ax = fig.add_subplot(111, aspect='equal') scat = ax.scatter(abs(errorsPosePred[0]), abs(errorsPoseFitted[0]), s=20, vmin=0, vmax=100, c=testOcclusions * 100, cmap=matplotlib.cm.plasma) cbar = fig.colorbar(scat, ticks=[0, 50, 100]) cbar.ax.set_yticklabels(['0%', '50%', '100%']) # vertically oriented colorbar ax.set_xlabel('Predicted azimuth errors') ax.set_ylabel('Fitted azimuth errors') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, max(x2,y2))) ax.set_ylim((0, max(x2,y2))) ax.plot([0, max(x2,y2)], [0, max(x2,y2)], ls="--", c=".3") for test_i in range(len(errorsPosePred[0])): if abs(errorsPoseFitted[0][test_i]) > abs(errorsPosePred[0][test_i]) + 10: plt.annotate( str(test_i), xy=(abs(errorsPosePred[0][test_i]), abs(errorsPoseFitted[0][test_i])), xytext=(0, 15), textcoords='offset points', ha='right', va='bottom', arrowprops=dict(arrowstyle='->', connectionstyle='arc3,rad=0'), size=8) # bbox=dict(boxstyle='round,pad=0.5', fc='yellow', alpha=0.5), ax.set_title('Recognition vs fitted azimuth errors') fig.savefig(directory + '-performance-scatter.pdf', bbox_inches='tight') plt.close(fig) directory = resultDir + 'pred-elevation-errors_fitted-elevation-error-' + str(methodFitNum) + '-' fig = plt.figure() ax = fig.add_subplot(111, aspect='equal') scat = ax.scatter(abs(errorsPosePred[1]), abs(errorsPoseFitted[1]), s=20, vmin=0, vmax=100, c=testOcclusions * 100, cmap=matplotlib.cm.plasma) for test_i in range(len(errorsPosePred[1])): if abs(errorsPoseFitted[1][test_i]) > abs(errorsPosePred[1][test_i]) + 10: plt.annotate( str(test_i), xy=(abs(errorsPosePred[1][test_i]), abs(errorsPoseFitted[1][test_i])), xytext=(15, 15), textcoords='offset points', ha='right', va='bottom', arrowprops=dict(arrowstyle='->', connectionstyle='arc3,rad=0'), size=8) # bbox=dict(boxstyle='round,pad=0.5', fc='yellow', alpha=0.5), cbar = fig.colorbar(scat, ticks=[0, 50, 100]) cbar.ax.set_yticklabels(['0%', '50%', '100%']) # vertically oriented colorbar ax.set_xlabel('Predicted elevation errors') ax.set_ylabel('Fitted elevation errors') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, max(x2,y2))) ax.set_ylim((0, max(x2,y2))) ax.plot([0, max(x2,y2)], [0, max(x2,y2)], ls="--", c=".3") ax.set_title('Recognition vs fitted azimuth errors') fig.savefig(directory + '-performance-scatter.pdf', bbox_inches='tight') plt.close(fig) directory = resultDir + 'shcoeffsserrorsC_fitted-shcoeffsserrorsC-' + str(methodFitNum) + '-' fig = plt.figure() ax = fig.add_subplot(111, aspect='equal') scat = ax.scatter(np.mean(errorsLightCoeffsC, axis=1), np.mean(errorsFittedLightCoeffsC, axis=1), s=20, vmin=0, vmax=100, c=testOcclusions * 100, cmap=matplotlib.cm.plasma) for test_i in range(len(errorsLightCoeffsC)): if np.mean(errorsFittedLightCoeffsC, axis=1)[test_i] > np.mean(errorsLightCoeffsC, axis=1)[test_i] + 0.1: plt.annotate( str(test_i), xy=(np.mean(errorsLightCoeffsC, axis=1)[test_i], np.mean(errorsFittedLightCoeffsC, axis=1)[test_i]), xytext=(0, 15), textcoords='offset points', ha='right', va='bottom', arrowprops=dict(arrowstyle='->', connectionstyle='arc3,rad=0'), size=8) # bbox=dict(boxstyle='round,pad=0.5', fc='yellow', alpha=0.5), cbar = fig.colorbar(scat, ticks=[0, 50, 100]) cbar.ax.set_yticklabels(['0%', '50%', '100%']) # vertically oriented colorbar ax.set_xlabel('Predicted SH coefficients errors') ax.set_ylabel('Fitted SH coefficients errors') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, max(x2,y2))) ax.set_ylim((0, max(x2,y2))) ax.plot([0, max(x2,y2)], [0, max(x2,y2)], ls="--", c=".3") ax.set_title('Recognition vs fitted SH coefficients errors') fig.savefig(directory + '-performance-scatter.pdf', bbox_inches='tight') plt.close(fig) if errorsEnvMap is not None and errorsFittedEnvMap is not None: directory = resultDir + 'shcoeffsserrorsC_fitted-shEnvMap-' + str(methodFitNum) + '-' fig = plt.figure() ax = fig.add_subplot(111, aspect='equal') scat = ax.scatter(errorsEnvMap, errorsFittedEnvMap, s=20, vmin=0, vmax=100, c=testOcclusions * 100, cmap=matplotlib.cm.plasma) cbar = fig.colorbar(scat, ticks=[0, 50, 100]) cbar.ax.set_yticklabels(['0%', '50%', '100%']) # vertically oriented colorbar ax.set_xlabel('Predicted SH coefficients errors') ax.set_ylabel('Fitted SH coefficients errors') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, max(x2,y2))) ax.set_ylim((0, max(x2,y2))) ax.plot([0, max(x2,y2)], [0, max(x2,y2)], ls="--", c=".3") ax.set_title('Recognition vs fitted SH Environment map errors') fig.savefig(directory + '-performance-scatter.pdf', bbox_inches='tight') plt.close(fig) directory = resultDir + 'shcoeffsserrors_fitted-shcoeffsserrors-' + str(methodFitNum) + '-' # Show scatter correlations with occlusions. fig = plt.figure() ax = fig.add_subplot(111, aspect='equal') scat = ax.scatter(np.mean(errorsLightCoeffs, axis=1), np.mean(errorsFittedLightCoeffs, axis=1), s=20, vmin=0, vmax=100, c=testOcclusions * 100, cmap=matplotlib.cm.plasma) cbar = fig.colorbar(scat, ticks=[0, 50, 100]) cbar.ax.set_yticklabels(['0%', '50%', '100%']) # vertically oriented colorbar ax.set_xlabel('Predicted SH coefficients errors') ax.set_ylabel('Fitted SH coefficients errors') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, max(x2,y2))) ax.set_ylim((0, max(x2,y2))) ax.plot([0, max(x2,y2)], [0, max(x2,y2)], ls="--", c=".3") ax.set_title('Recognition vs fitted SH coefficients errors') fig.savefig(directory + '-performance-scatter.pdf', bbox_inches='tight') plt.close(fig) if useShapeModel: directory = resultDir + 'shapeparamserrors_fitted-shapeparamserrors-' + str(methodFitNum) + '-' # Show scatter correlations with occlusions. fig = plt.figure() ax = fig.add_subplot(111, aspect='equal') scat = ax.scatter(np.mean(errorsShapeParams, axis=1), np.mean(errorsFittedShapeParams, axis=1), s=20, vmin=0, vmax=100, c=testOcclusions * 100, cmap=matplotlib.cm.plasma) cbar = fig.colorbar(scat, ticks=[0, 50, 100]) cbar.ax.set_yticklabels(['0%', '50%', '100%']) # vertically oriented colorbar ax.set_xlabel('Predicted shape parameters errors') ax.set_ylabel('Fitted shape parameters errors') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, max(x2,y2))) ax.set_ylim((0, max(x2,y2))) ax.plot([0, max(x2,y2)], [0, max(x2,y2)], ls="--", c=".3") ax.set_title('Recognition vs fitted Shape parameters errors') fig.savefig(directory + '-performance-scatter.pdf', bbox_inches='tight') plt.close(fig) if errorsShapeVertices is not None and errorsFittedShapeVertices is not None: directory = resultDir + 'shapeverticeserrors_fitted-shapeverticesserrors-' + str(methodFitNum) + '-' # Show scatter correlations with occlusions. fig = plt.figure() ax = fig.add_subplot(111, aspect='equal') scat = ax.scatter(errorsShapeVertices, errorsFittedShapeVertices, s=20, vmin=0, vmax=100, c=testOcclusions * 100, cmap=matplotlib.cm.plasma) for test_i in range(len(errorsLightCoeffsC)): if errorsFittedShapeVertices[test_i] > errorsShapeVertices[test_i] + 0.75: plt.annotate( str(test_i), xy=(errorsShapeVertices[test_i], errorsFittedShapeVertices[test_i]), xytext=(15, 15), textcoords='offset points', ha='right', va='bottom', arrowprops=dict(arrowstyle='->', connectionstyle='arc3,rad=0'), size=8) cbar = fig.colorbar(scat, ticks=[0, 50, 100]) cbar.ax.set_yticklabels(['0%', '50%', '100%']) # vertically oriented colorbar ax.set_xlabel('Predicted shape vertices errors') ax.set_ylabel('Fitted shape vertices errors') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, max(x2,y2))) ax.set_ylim((0, max(x2,y2))) ax.plot([0, max(x2,y2)], [0, max(x2,y2)], ls="--", c=".3") ax.set_title('Recognition vs fitted Shape parameters errors') fig.savefig(directory + '-performance-scatter.pdf', bbox_inches='tight') plt.close(fig) directory = resultDir + 'vColorsE_fitted-vColorsE-' + str(methodFitNum) + '-' # Show scatter correlations with occlusions. fig = plt.figure() ax = fig.add_subplot(111, aspect='equal') scat = ax.scatter(errorsVColorsE, errorsFittedVColorsE, s=20, vmin=0, vmax=100, c=testOcclusions * 100, cmap=matplotlib.cm.plasma) cbar = fig.colorbar(scat, ticks=[0, 50, 100]) cbar.ax.set_yticklabels(['0%', '50%', '100%']) # vertically oriented colorbar ax.set_xlabel('Predicted VColor E coefficients errors') ax.set_ylabel('Fitted VColor E coefficients errors') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, max(x2,y2))) ax.set_ylim((0, max(x2,y2))) ax.plot([0, max(x2,y2)], [0, max(x2,y2)], ls="--", c=".3") ax.set_title('Recognition vs fitted vertex color errors') fig.savefig(directory + '-performance-scatter.pdf', bbox_inches='tight') plt.close(fig) directory = resultDir + 'vColorsC_fitted-vColorsC-' + str(methodFitNum) + '-' fig = plt.figure() ax = fig.add_subplot(111, aspect='equal') scat = ax.scatter(errorsVColorsC, errorsFittedVColorsC, s=20, vmin=0, vmax=100, c=testOcclusions * 100, cmap=matplotlib.cm.plasma) cbar = fig.colorbar(scat, ticks=[0, 50, 100]) cbar.ax.set_yticklabels(['0%', '50%', '100%']) # vertically oriented colorbar ax.set_xlabel('Predicted vertex color errors') ax.set_ylabel('Fitted vertex color errors') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, max(x2,y2))) ax.set_ylim((0, max(x2,y2))) ax.plot([0, max(x2,y2)], [0, max(x2,y2)], ls="--", c=".3") ax.set_title('Recognition vs fitted vertex color errors') fig.savefig(directory + '-performance-scatter.pdf', bbox_inches='tight') plt.close(fig) directory = resultDir + 'vColorsS_fitted-vColorsS-' + str(methodFitNum) + '-' fig = plt.figure() ax = fig.add_subplot(111, aspect='equal') scat = ax.scatter(errorsVColorsS, errorsFittedVColorsS, s=20, vmin=0, vmax=100, c=testOcclusions * 100, cmap=matplotlib.cm.plasma) for test_i in range(len(errorsVColorsS)): if errorsFittedVColorsS[test_i] > errorsVColorsS[test_i] + 0.1: plt.annotate( str(test_i), xy=(errorsVColorsS[test_i], errorsFittedVColorsS[test_i]), xytext=(15, 15), textcoords='offset points', ha='right', va='bottom', arrowprops=dict(arrowstyle='->', connectionstyle='arc3,rad=0'), size=8) cbar = fig.colorbar(scat, ticks=[0, 50, 100]) cbar.ax.set_yticklabels(['0%', '50%', '100%']) # vertically oriented colorbar ax.set_xlabel('Predicted VColor coefficients errors') ax.set_ylabel('Fitted VColor coefficients errors') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, max(x2, y2))) ax.set_ylim((0, max(x2, y2))) ax.plot([0, max(x2, y2)], [0, max(x2, y2)], ls="--", c=".3") ax.set_title('Recognition vs fitted vertex color errors') fig.savefig(directory + '-performance-scatter.pdf', bbox_inches='tight') plt.close(fig) def saveLikelihoodPlots(resultDir, occlusions, methodsPred, plotColors, plotMethodsIndices, meanLikelihoodArr): latexify(columns=2) directory = resultDir + 'predictionLikelihood-FitLikelihood' fig = plt.figure() ax = fig.add_subplot(111) ax.plot(occlusions, meanLikelihoodArr[0], c='b', label='Ground-truth NLL') ax.plot(occlusions, meanLikelihoodArr[1], c='r', label='Robust Fit NLL') legend = ax.legend(loc='best') ax.set_xlabel('Occlusion (\%)') ax.set_ylabel('Mean Likelihood') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() # ax.set_xlim((min(x1), 100)) # ax.set_ylim((-0.0, y2)) ax.set_title('Cumulative prediction per occlusion level') fig.savefig(directory + '-plot.pdf', bbox_inches='tight') plt.close(fig) directory = resultDir + 'predictionLikelihoodGaussian-FitLikelihoodGaussian' fig = plt.figure() ax = fig.add_subplot(111) ax.plot(occlusions, meanLikelihoodArr[2], c='b', label='Groundtruth NLL') ax.plot(occlusions, meanLikelihoodArr[3], c='r', label='Robust Fit NLL') legend = ax.legend(loc='best') ax.set_xlabel('Occlusion (\%)') ax.set_ylabel('Mean likelihood') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() # ax.set_xlim((min(x1), 100)) # ax.set_ylim((-0.0, y2)) ax.set_title('Cumulative prediction per occlusion level') fig.savefig(directory + '-plot.pdf', bbox_inches='tight') plt.close(fig) def saveLikelihoodScatter(resultDirOcclusion, testSet, testOcclusions, likelihoods): # Show scatter correlations with occlusions. directory = resultDirOcclusion + 'robustLikelihood_fitted-robustLikelihood' fig = plt.figure() ax = fig.add_subplot(111, aspect='equal') scat = ax.scatter(likelihoods[0][testSet], likelihoods[1][testSet], s=20, vmin=0, vmax=100, c=testOcclusions * 100, cmap=matplotlib.cm.plasma) cbar = fig.colorbar(scat, ticks=[0, 50, 100]) cbar.ax.set_yticklabels(['0%', '50%', '100%']) # vertically oriented colorbar ax.set_xlabel('Ground-truth NLL (avg)') ax.set_ylabel('Recognition+Fit NLL (avg)') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((min(x1,y1), max(x2,y2))) ax.set_ylim((min(x1,y1), max(x2,y2))) ax.plot([min(x1,y1), max(x2,y2)], [min(x1,y1), max(x2,y2)], ls="--", c=".3") ax.set_title('GT vs Recognition+Robust fit negative log-likelihood') fig.savefig(directory + '-performance-scatter.pdf', bbox_inches='tight') plt.close(fig) # directory = resultDirOcclusion + 'gaussianLikelihood_fitted-gaussianLikelihood' # fig = plt.figure() # ax = fig.add_subplot(111, aspect='equal') # scat = ax.scatter(likelihoods[2][testSet], likelihoods[3][testSet], s=20, vmin=0, vmax=100, c=testOcclusions * 100, # cmap=matplotlib.cm.plasma) # cbar = fig.colorbar(scat, ticks=[0, 50, 100]) # cbar.ax.set_yticklabels(['0%', '50%', '100%']) # vertically oriented colorbar # ax.set_xlabel('Ground-truth avg likelihood') # ax.set_ylabel('Fitted avg likelihood') # x1, x2 = ax.get_xlim() # y1, y2 = ax.get_ylim() # ax.set_xlim((0, max(x2,y2))) # ax.set_ylim((0, max(x2,y2))) # ax.plot([0, max(x2,y2)], [0, max(x2,y2)], ls="--", c=".3") # ax.set_title('GT vs Recognition+Gaussian Fit likelihood') # fig.savefig(directory + '-performance-scatter.pdf', bbox_inches='tight') # plt.close(fig) def saveHistograms(): pass # latexify(columns=2) # directory = resultDir + 'reconition-azimuth' # fig = plt.figure() # ax = fig.add_subplot(111) # ax.hist(errorsPosePredList[2][0], bins=40) # ax.set_xlabel('Azimuth errors') # ax.set_ylabel('Counts') # ax.set_title('Azimuth errors histogram') # ax.set_xlim((-180, 180)) # fig.savefig(directory + '-histogram.pdf', bbox_inches='tight') # plt.close(fig) # # directory = resultDir + 'robust-azimuth' # fig = plt.figure() # ax = fig.add_subplot(111) # ax.hist(errorsPosePredList[4][0], bins=40) # ax.set_xlabel('Azimuth errors') # ax.set_ylabel('Counts') # ax.set_xlim((-180, 180)) # ax.set_title('Azimuth errors histogram') # fig.savefig(directory + '-histogram.pdf', bbox_inches='tight') # plt.close(fig) # # directory = resultDir + 'robust-elevation' # fig = plt.figure() # ax = fig.add_subplot(111) # ax.hist(errorsPosePredList[4][1], bins=100) # ax.set_xlabel('Elevation errors') # ax.set_ylabel('Counts') # x1, x2 = ax.get_xlim() # y1, y2 = ax.get_ylim() # ax.set_xlim((-90, 90)) # ax.set_title('Elevation errors histogram') # fig.savefig(directory + '-histogram.pdf', bbox_inches='tight') # plt.close(fig) # # # directory = resultDir + 'recognition-elevation' # fig = plt.figure() # ax = fig.add_subplot(111) # ax.hist(errorsPosePredList[2][1], bins=30) # ax.set_xlabel('Elevation errors') # ax.set_ylabel('Counts') # ax.set_title('Elevation errors histogram') # ax.set_xlim((-90, 90)) # ax.set_ylim((y1, y2)) # fig.savefig(directory + '-histogram.pdf', bbox_inches='tight') # plt.close(fig) # directory = resultDir + 'robust-elevation-occlusion' # fig = plt.figure() # ax = fig.add_subplot(111) # scat = ax.scatter(testOcclusions*100, abs(errorsPosePredList[4][1]), s=20, vmin=0, vmax=100) # ax.set_xlabel('Occlusion level (\%)') # ax.set_ylabel('Robust fit elevation errors') # x1, x2 = ax.get_xlim() # y1, y2 = ax.get_ylim() # ax.set_xlim((0, max(x2,y2))) # ax.set_ylim((0, max(x2,y2))) # ax.set_title('Fitted azimuth errors') # fig.savefig(directory + '-scatter.pdf', bbox_inches='tight') # plt.close(fig) # # directory = resultDir + 'recognition-elevation-occlusion' # fig = plt.figure() # ax = fig.add_subplot(111) # scat = ax.scatter(testOcclusions*100, abs(errorsPosePredList[2][1]), s=20, vmin=0, vmax=100) # ax.set_ylabel('Robust fit elevation errors') # ax.set_xlabel('Occlusion level (\%)') # x1, x2 = ax.get_xlim() # y1, y2 = ax.get_ylim() # ax.set_xlim((0, max(x2,y2))) # ax.set_ylim((0, max(x2,y2))) # ax.set_title('Fitted azimuth errors') # fig.savefig(directory + '-scatter.pdf', bbox_inches='tight') # plt.close(fig) def saveConditionalHistograms(resultDir, occlusions, methodsPred, variablesDescr, plotMethodsIndices, errorsList): latexify(columns=2) numBins = 20 numErrBins = 40 for variable_i in range(len(variablesDescr)): variableDescr = variablesDescr[variable_i] errors = errorsList[variable_i] directory = resultDir + variableDescr maxErr = np.max(np.array([np.max(errors[method_i]) if errors[method_i] is not None else 0 for method_i in plotMethodsIndices])) for method_i in plotMethodsIndices: condHists = [] if errors[method_i] is not None and len(errors[method_i]) > 0: fig = plt.figure() ax = fig.add_subplot(111) bins = np.linspace(0,0.9,numBins) occlusionBinIndices = np.array([(occlusions < bins[i]) & (occlusions > bins[i-1]) for i in np.arange(1,len(bins))]) # ax.plot(occlusions, errorsPosePredList[method_i], c=plotColors[method_i], linestyle=plotStyles[method_i], label=methodsPred[method_i]) legend = ax.legend(loc='best') ax.set_xlabel('Occlusion (\%)') ax.set_ylabel('Error Cond. Histogram') # x1, x2 = ax.get_xlim() # y1, y2 = ax.get_ylim() if variableDescr == 'Illumination': maxErr = 0.2 for i in range(len(bins)-1): hist, bin_edges = np.histogram(errors[method_i][occlusionBinIndices[i]], numErrBins, range=(0,maxErr), density=True) condHists = condHists + [hist] # ipdb.set_trace() histMat = np.flipud(np.vstack(condHists).T) maxHistProb = np.max(histMat) ax.matshow(histMat, cmap=matplotlib.cm.gray, interpolation='none', extent=[0,90,0,45], vmin=0, vmax=maxHistProb) # locs, labels = ax.xticks() # ax.set_xticks(bins) ax.set_xlim((0, 90)) ax.set_ylim((0, 45)) yticks = np.linspace(0,45,10) ytickslabels = ["{:1.2f}".format(num) for num in np.linspace(0, maxErr, 10)] ax.set_yticks(yticks) ax.set_yticklabels(ytickslabels) # ax.set_xlim((0, 100)) # ax.set_ylim((-0.0, y2)) # ax.set_title('Cumulative segmentation accuracy per occlusion level') ax.tick_params(axis='both', which='major', labelsize=8) ax.tick_params(axis='both', which='minor', labelsize=8) fig.savefig(directory + '-' + methodsPred[method_i] + '-conditional-histogram.pdf', bbox_inches='tight') plt.close(fig) def saveOcclusionPlots(resultDir, prefix, occlusions, methodsPred, plotColors, plotStyles, plotMethodsIndices, useShapeModel, meanAbsErrAzsArr, meanAbsErrElevsArr, meanErrorsVColorsCArr, meanErrorsVColorsEArr, meanErrorsVColorsSArr, meanErrorsLightCoeffsArr, meanErrorsShapeParamsArr, meanErrorsShapeVerticesArr, meanErrorsLightCoeffsCArr, meanErrorsEnvMapArr,meanErrorsSegmentationArr): latexify(columns=2) directory = resultDir + prefix + 'predictionMeanError-Segmentation' fig = plt.figure() ax = fig.add_subplot(111) for method_i in plotMethodsIndices: if meanErrorsSegmentationArr[method_i] is not None and len(meanErrorsSegmentationArr[method_i]) > 0: ax.plot(occlusions, meanErrorsSegmentationArr[method_i], c=plotColors[method_i], linestyle=plotStyles[method_i], label=methodsPred[method_i]) legend = ax.legend(loc='best') ax.set_xlabel('Occlusion (\%)') ax.set_ylabel('Segmentation accuracy') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, 90)) ax.set_ylim((-0.0, y2)) # ax.set_title('Cumulative segmentation accuracy per occlusion level') fig.savefig(directory + '-performance-plot.pdf', bbox_inches='tight') plt.close(fig) directory = resultDir + prefix + 'predictionMeanError-Azimuth' fig = plt.figure() ax = fig.add_subplot(111) for method_i in plotMethodsIndices: if meanAbsErrAzsArr[method_i] is not None and len(meanAbsErrAzsArr[method_i]) > 0: ax.semilogy(occlusions, meanAbsErrAzsArr[method_i], c=plotColors[method_i], linestyle=plotStyles[method_i], label=methodsPred[method_i]) legend = ax.legend(loc='best') ax.set_xlabel('Occlusion (\%)') ax.set_ylabel('Angular error') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, 90)) ax.set_ylim((y1, y2)) # ax.set_title('Cumulative prediction per occlusion level') fig.savefig(directory + '-performance-plot.pdf', bbox_inches='tight') plt.close(fig) directory = resultDir + prefix +'predictionMeanError-Elev' fig = plt.figure() ax = fig.add_subplot(111) for method_i in plotMethodsIndices: if meanAbsErrElevsArr[method_i] is not None and len(meanAbsErrElevsArr[method_i]) > 0: ax.plot(occlusions, meanAbsErrElevsArr[method_i], c=plotColors[method_i], linestyle=plotStyles[method_i], label=methodsPred[method_i]) legend = ax.legend(loc='best') ax.set_xlabel('Occlusion (\%)') ax.set_ylabel('Angular error') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, 90)) ax.set_ylim((-0.0, y2)) # ax.set_title('Cumulative prediction per occlusion level') fig.savefig(directory + '-performance-plot.pdf', bbox_inches='tight') plt.close(fig) directory = resultDir + prefix +'predictionMeanError-VColors-C' fig = plt.figure() ax = fig.add_subplot(111) for method_i in plotMethodsIndices: if meanErrorsVColorsCArr[method_i] is not None and len(meanErrorsVColorsCArr[method_i]) > 0: ax.plot(occlusions, meanErrorsVColorsCArr[method_i], c=plotColors[method_i], linestyle=plotStyles[method_i], label=methodsPred[method_i]) legend = ax.legend(loc='best') ax.set_xlabel('Occlusion (\%)') ax.set_ylabel('Appearance error') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, 90)) ax.set_ylim((-0.0, y2)) # ax.set_title('Cumulative prediction per occlusion level') fig.savefig(directory + '-performance-plot.pdf', bbox_inches='tight') plt.close(fig) directory = resultDir + prefix +'predictionMeanError-VColors-E' fig = plt.figure() ax = fig.add_subplot(111) for method_i in plotMethodsIndices: if meanErrorsVColorsEArr[method_i] is not None and len(meanErrorsVColorsEArr[method_i]) > 0: ax.plot(occlusions, meanErrorsVColorsEArr[method_i], c=plotColors[method_i], linestyle=plotStyles[method_i], label=methodsPred[method_i]) legend = ax.legend(loc='best') ax.set_xlabel('Occlusion (\%)') ax.set_ylabel('Appearance error') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, 90)) ax.set_ylim((-0.0, y2)) # ax.set_title('Cumulative prediction per occlusion level') fig.savefig(directory + '-performance-plot.pdf', bbox_inches='tight') plt.close(fig) directory = resultDir + prefix +'predictionMeanError-VColors-S' fig = plt.figure() ax = fig.add_subplot(111) for method_i in plotMethodsIndices: if meanErrorsVColorsSArr[method_i] is not None and len(meanErrorsVColorsSArr[method_i]) > 0: ax.plot(occlusions, meanErrorsVColorsSArr[method_i], c=plotColors[method_i], linestyle=plotStyles[method_i], label=methodsPred[method_i]) legend = ax.legend(loc='best') ax.set_xlabel('Occlusion (\%)') ax.set_ylabel('Appearance error') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, 90)) ax.set_ylim((-0.0, y2)) # ax.set_title('Cumulative prediction per occlusion level') fig.savefig(directory + '-performance-plot.pdf', bbox_inches='tight') plt.close(fig) directory = resultDir + prefix +'predictionMeanError-SH' fig = plt.figure() ax = fig.add_subplot(111) for method_i in plotMethodsIndices: if meanErrorsLightCoeffsArr[method_i] is not None and len(meanErrorsLightCoeffsArr[method_i]) > 0: ax.plot(occlusions, meanErrorsLightCoeffsArr[method_i], c=plotColors[method_i], linestyle=plotStyles[method_i], label=methodsPred[method_i]) legend = ax.legend(loc='best') ax.set_xlabel('Occlusion (\%)') ax.set_ylabel('Illumination error') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, 90)) ax.set_ylim((-0.0, y2)) # ax.set_title('Cumulative prediction per occlusion level') fig.savefig(directory + '-performance-plot.pdf', bbox_inches='tight') plt.close(fig) if useShapeModel: directory = resultDir + prefix +'predictionMeanError-ShapeParams' fig = plt.figure() ax = fig.add_subplot(111) for method_i in plotMethodsIndices: if meanErrorsShapeParamsArr[method_i] is not None and len(meanErrorsShapeParamsArr[method_i]) > 0: ax.plot(occlusions, meanErrorsShapeParamsArr[method_i], c=plotColors[method_i], linestyle=plotStyles[method_i], label=methodsPred[method_i]) legend = ax.legend(loc='best') ax.set_xlabel('Occlusion (\%)') ax.set_ylabel('Shape parameters error') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, 90)) ax.set_ylim((-0.0, y2)) # ax.set_title('Cumulative prediction per occlusion level') fig.savefig(directory + '-performance-plot.pdf', bbox_inches='tight') plt.close(fig) directory = resultDir + prefix +'predictionMeanError-ShapeVertices' fig = plt.figure() ax = fig.add_subplot(111) for method_i in plotMethodsIndices: if meanErrorsShapeVerticesArr[method_i] is not None and len(meanErrorsShapeVerticesArr[method_i]) > 0: ax.semilogy(occlusions, meanErrorsShapeVerticesArr[method_i], c=plotColors[method_i], linestyle=plotStyles[method_i], label=methodsPred[method_i]) legend = ax.legend(loc='best') ax.set_xlabel('Occlusion (\%)') ax.set_ylabel('Shape vertices error') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, 90)) ax.set_ylim((y1, y2)) # ax.set_title('Cumulative prediction per occlusion level') fig.savefig(directory + '-performance-plot.pdf', bbox_inches='tight') plt.close(fig) directory = resultDir + prefix +'predictionMeanError-SH-C' fig = plt.figure() ax = fig.add_subplot(111) for method_i in plotMethodsIndices: if meanErrorsLightCoeffsCArr[method_i] is not None and len(meanErrorsLightCoeffsCArr[method_i]) > 0: ax.plot(occlusions, meanErrorsLightCoeffsCArr[method_i], c=plotColors[method_i], linestyle=plotStyles[method_i], label=methodsPred[method_i]) legend = ax.legend(loc='best') ax.set_xlabel('Occlusion (\%)') ax.set_ylabel('Illumination error') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, 90)) ax.set_ylim((-0.0, y2)) # ax.set_title('Cumulative prediction per occlusion level') fig.savefig(directory + '-performance-plot.pdf', bbox_inches='tight') plt.close(fig) directory = resultDir + prefix +'predictionMeanError-SH-EnvMap' fig = plt.figure() ax = fig.add_subplot(111) for method_i in plotMethodsIndices: if meanErrorsEnvMapArr[method_i] is not None and len(meanErrorsEnvMapArr[method_i]) > 0: ax.plot(occlusions, meanErrorsEnvMapArr[method_i], c=plotColors[method_i], linestyle=plotStyles[method_i], label=methodsPred[method_i]) legend = ax.legend(loc='best') ax.set_xlabel('Occlusion (\%)') ax.set_ylabel('Illumination error') x1, x2 = ax.get_xlim() y1, y2 = ax.get_ylim() ax.set_xlim((0, 90)) ax.set_ylim((-0.0, y2)) # ax.set_title('Cumulative prediction per occlusion level') fig.savefig(directory + '-performance-plot.pdf', bbox_inches='tight') plt.close(fig) from numpy.core.umath_tests import matrix_multiply def scaleInvariantMSECoeff(x_pred, x_target): #Rows: test samples #Cols: target variables scales = (matrix_multiply(x_pred[:,None,:],x_target[:,:,None])/matrix_multiply(x_pred[:,None,:],x_pred[:,:,None])).ravel() return scales def euclidean(X,Y): """ x: matrix (N,D), each row is an datapoint y: matrix (M,D), each row is an datapoint A MxN matrix of Euclidean distances is returned """ return scipy.spatial.distance.cdist(X, Y[None,:], 'euclidean').T def one_nn(x_train, x_test, distance_f=euclidean): """ """ distances = distance_f(x_train, x_test) nn_indices = np.argmin(distances, 1) return nn_indices def shapeVertexErrors(chShapeParams, chVertices, testShapeParamsGT, shapeParamsPred): oldShapeParams = chShapeParams.r.copy() errorsShapeVertices = np.array([]) for test_i in range(len(testShapeParamsGT)): chShapeParams[:] = testShapeParamsGT[test_i].copy() vertsGT = chVertices.r.copy() chShapeParams[:] = shapeParamsPred[test_i].copy() vertsPred = chVertices.r.copy() errorShapeVertices = np.sqrt(np.sum((vertsPred - vertsGT)**2)) errorsShapeVertices = np.append(errorsShapeVertices,errorShapeVertices) chShapeParams[:] = oldShapeParams return errorsShapeVertices def computeErrors(setTest, azimuths, testAzsRel, elevations, testElevsGT, vColors, testVColorGT, lightCoeffs, testLightCoefficientsGTRel, approxProjections, approxProjectionsGT, shapeParams, testShapeParamsGT, useShapeModel, chShapeParams, chVertices, posteriors, masksGT): errorsPosePredList = [] errorsLightCoeffsList = [] errorsShapeParamsList = [] errorsShapeVerticesList = [] errorsEnvMapList = [] errorsLightCoeffsCList = [] errorsVColorsEList = [] errorsVColorsCList = [] errorsVColorsSList = [] errorsSegmentation = [] for method in range(len(azimuths)): print("Computing errors for method " + str(method)) azsPred = azimuths[method] elevsPred = elevations[method] vColorsPred = vColors[method] posteriorsPred = posteriors[method] relLightCoefficientsPred = lightCoeffs[method] if approxProjectionsGT is not None: approxProjectionsPred = approxProjections[method] if useShapeModel: shapeParamsPred = shapeParams[method] if azsPred is not None: errorsPosePred = recognition_models.evaluatePrediction(testAzsRel[setTest], testElevsGT[setTest], azsPred[setTest], elevsPred[setTest]) errorsPosePredList = errorsPosePredList + [errorsPosePred] else: errorsPosePredList = errorsPosePredList + [None] if relLightCoefficientsPred is not None: errorsLightCoeffs = (testLightCoefficientsGTRel[setTest] - relLightCoefficientsPred[setTest]) ** 2 errorsLightCoeffsList = errorsLightCoeffsList + [errorsLightCoeffs] else: errorsLightCoeffsList = errorsLightCoeffsList + [None] if useShapeModel and shapeParamsPred is not None: errorsShapeParams = (testShapeParamsGT[setTest] - shapeParamsPred[setTest]) ** 2 errorsShapeParamsList = errorsShapeParamsList + [errorsShapeParams] errorsShapeVertices = shapeVertexErrors(chShapeParams, chVertices, testShapeParamsGT[setTest], shapeParamsPred[setTest]) errorsShapeVerticesList = errorsShapeVerticesList + [errorsShapeVertices] else: errorsShapeParamsList = errorsShapeParamsList + [None] errorsShapeVerticesList = errorsShapeVerticesList + [None] if approxProjectionsGT is not None and approxProjectionsPred is not None: envMapProjScaling = scaleInvariantMSECoeff(approxProjectionsPred.reshape([len(approxProjectionsPred), -1])[setTest], approxProjectionsGT.reshape([len(approxProjectionsGT), -1])[setTest]) errorsEnvMap = np.mean((approxProjectionsGT[setTest] - envMapProjScaling[:,None, None]*approxProjectionsPred[setTest])**2, axis=(1,2)) errorsEnvMapList= errorsEnvMapList + [errorsEnvMap] else: errorsEnvMapList = errorsEnvMapList + [None] if relLightCoefficientsPred is not None: envMapScaling = scaleInvariantMSECoeff(relLightCoefficientsPred[setTest], testLightCoefficientsGTRel[setTest]) errorsLightCoeffsC = (testLightCoefficientsGTRel[setTest] - envMapScaling[:,None]* relLightCoefficientsPred[setTest]) ** 2 errorsLightCoeffsCList = errorsLightCoeffsCList + [errorsLightCoeffsC] else: errorsLightCoeffsCList = errorsLightCoeffsCList + [None] if vColorsPred is not None: errorsVColorsE = image_processing.eColourDifference(testVColorGT[setTest], vColorsPred[setTest]) errorsVColorsEList = errorsVColorsEList + [errorsVColorsE] errorsVColorsC = image_processing.cColourDifference(testVColorGT[setTest], vColorsPred[setTest]) errorsVColorsCList = errorsVColorsCList + [errorsVColorsC] errorsVColorsS = image_processing.scaleInvariantColourDifference(testVColorGT[setTest], vColorsPred[setTest]) errorsVColorsSList = errorsVColorsSList + [errorsVColorsS] else: errorsVColorsEList = errorsVColorsEList + [None] errorsVColorsCList = errorsVColorsCList + [None] errorsVColorsSList = errorsVColorsSList + [None] if posteriorsPred is not None: # if method == 3: # ipdb.set_trace() masksCat = np.concatenate([masksGT[:,:,:,None][setTest], posteriorsPred[:,:,:,None][setTest]], axis=3) errorSegmentation = np.sum(np.all(masksCat, axis=3), axis=(1,2)) / np.sum(np.any(masksCat, axis=3), axis=(1,2)) errorsSegmentation = errorsSegmentation + [errorSegmentation] else: errorsSegmentation = errorsSegmentation + [None] return errorsPosePredList, errorsLightCoeffsList, errorsShapeParamsList, errorsShapeVerticesList, errorsEnvMapList, errorsLightCoeffsCList, errorsVColorsEList, errorsVColorsCList, errorsVColorsSList, errorsSegmentation def computeErrorAverages(averageFun, testSet, useShapeModel, errorsPosePredList, errorsLightCoeffsList, errorsShapeParamsList, errorsShapeVerticesList, errorsEnvMapList, errorsLightCoeffsCList, errorsVColorsEList, errorsVColorsCList, errorsVColorsSList, errorsSegmentation): meanAbsErrAzsList = [] meanAbsErrElevsList = [] meanErrorsLightCoeffsList = [] meanErrorsShapeParamsList = [] meanErrorsShapeVerticesList = [] meanErrorsLightCoeffsCList = [] meanErrorsEnvMapList = [] meanErrorsVColorsEList = [] meanErrorsVColorsCList = [] meanErrorsVColorsSList = [] meanErrorsSegmentation = [] for method_i in range(len(errorsPosePredList)): if errorsPosePredList[method_i] is not None: meanAbsErrAzsList = meanAbsErrAzsList + [averageFun(np.abs(errorsPosePredList[method_i][0][testSet]))] else: meanAbsErrAzsList = meanAbsErrAzsList + [None] if errorsPosePredList[method_i] is not None: meanAbsErrElevsList = meanAbsErrElevsList + [averageFun(np.abs(errorsPosePredList[method_i][1][testSet]))] else: meanAbsErrElevsList = meanAbsErrElevsList + [None] if errorsLightCoeffsList[method_i] is not None: meanErrorsLightCoeffsList = meanErrorsLightCoeffsList + [averageFun(averageFun(errorsLightCoeffsList[method_i][testSet],axis=1), axis=0)] else: meanErrorsLightCoeffsList = meanErrorsLightCoeffsList + [None] if useShapeModel: if errorsShapeParamsList[method_i] is not None: meanErrorsShapeParamsList = meanErrorsShapeParamsList + [averageFun(np.mean(errorsShapeParamsList[method_i][testSet],axis=1), axis=0)] else: meanErrorsShapeParamsList = meanErrorsShapeParamsList + [None] if errorsShapeVerticesList[method_i] is not None: meanErrorsShapeVerticesList = meanErrorsShapeVerticesList + [averageFun(errorsShapeVerticesList[method_i][testSet], axis=0)] else: meanErrorsShapeVerticesList = meanErrorsShapeVerticesList + [None] if errorsLightCoeffsCList[method_i] is not None: meanErrorsLightCoeffsCList = meanErrorsLightCoeffsCList + [averageFun(np.mean(errorsLightCoeffsCList[method_i][testSet],axis=1), axis=0)] else: meanErrorsLightCoeffsCList = meanErrorsLightCoeffsCList + [None] if errorsEnvMapList[method_i] is not None: meanErrorsEnvMapList = meanErrorsEnvMapList + [averageFun(errorsEnvMapList[method_i][testSet])] else: meanErrorsEnvMapList = meanErrorsEnvMapList + [None] if errorsVColorsEList[method_i] is not None: meanErrorsVColorsEList = meanErrorsVColorsEList + [averageFun(errorsVColorsEList[method_i][testSet], axis=0)] else: meanErrorsVColorsEList = meanErrorsVColorsEList + [None] if errorsVColorsCList[method_i] is not None: meanErrorsVColorsCList = meanErrorsVColorsCList + [averageFun(errorsVColorsCList[method_i][testSet], axis=0)] else: meanErrorsVColorsCList = meanErrorsVColorsCList + [None] if errorsVColorsSList[method_i] is not None: meanErrorsVColorsSList = meanErrorsVColorsSList + [averageFun(errorsVColorsSList[method_i][testSet], axis=0)] else: meanErrorsVColorsSList = meanErrorsVColorsSList + [None] if errorsSegmentation[method_i] is not None: meanErrorsSegmentation = meanErrorsSegmentation + [averageFun(errorsSegmentation[method_i][testSet], axis=0)] else: meanErrorsSegmentation = meanErrorsSegmentation + [None] return meanAbsErrAzsList, meanAbsErrElevsList, meanErrorsLightCoeffsList, meanErrorsShapeParamsList, meanErrorsShapeVerticesList, meanErrorsLightCoeffsCList, meanErrorsEnvMapList, meanErrorsVColorsEList, meanErrorsVColorsCList, meanErrorsVColorsSList, meanErrorsSegmentation def writeImagesHdf5(imagesDir, writeDir, imageSet, writeGray=False ): print("Writing HDF5 file") image = skimage.io.imread(imagesDir + 'im' + str(imageSet[0]) + '.jpeg') imDtype = image.dtype width = image.shape[1] height = image.shape[0] if not writeGray: gtDataFile = h5py.File(writeDir + 'images.h5', 'w') images = np.array([], dtype = np.dtype('('+ str(height)+','+ str(width) +',3)uint8')) gtDataset = gtDataFile.create_dataset("images", data=images, maxshape=(None,height,width, 3)) # images = np.zeros([len(imageSet), height, width, 3], dtype=np.uint8) else: imageGray = 0.3*image[:,:,0] + 0.59*image[:,:,1] + 0.11*image[:,:,2] grayDtype = imageGray.dtype gtDataFile = h5py.File(writeDir + 'images_gray.h5', 'w') # images = np.zeros([], dtype=np.float32) images = np.array([], dtype = np.dtype('('+ str(height)+','+ str(width) +')f')) gtDataset = gtDataFile.create_dataset("images", data=images, maxshape=(None,height,width)) for imageit, imageid in enumerate(imageSet): gtDataset.resize(gtDataset.shape[0]+1, axis=0) image = skimage.io.imread(imagesDir + 'im' + str(imageid) + '.jpeg').astype(np.uint8) if not writeGray: gtDataset[-1] = image else: image = image.astype(np.float32)/255.0 gtDataset[-1] = 0.3*image[:,:,0] + 0.59*image[:,:,1] + 0.11*image[:,:,2] gtDataFile.flush() gtDataFile.close() print("Ended writing HDF5 file") def loadMasks(imagesDir, maskSet): masks = [] for imageit, imageid in enumerate(maskSet): if os.path.isfile(imagesDir + 'mask' + str(imageid) + '.npy'): masks = masks + [np.load(imagesDir + 'mask' + str(imageid) + '.npy')[None,:,:]] return np.vstack(masks) def loadMasksMug(imagesDir, maskSet): masks = [] for imageit, imageid in enumerate(maskSet): if os.path.isfile(imagesDir + 'mask' + str(imageid) + '_mug.npy'): masks = masks + [np.load(imagesDir + 'mask' + str(imageid) + '_mug.npy')[None,:,:]] return np.vstack(masks) def readImages(imagesDir, imageSet, loadGray=False, loadFromHdf5=False): if loadFromHdf5: if not loadGray: if os.path.isfile(imagesDir + 'images.h5'): gtDataFile = h5py.File(imagesDir + 'images.h5', 'r') boolSet = np.zeros(gtDataFile["images"].shape[0]).astype(np.bool) boolSet[imageSet] = True return gtDataFile["images"][boolSet,:,:,:].astype(np.float32)/255.0 else: if os.path.isfile(imagesDir + 'images_gray.h5'): gtDataFile = h5py.File(imagesDir + 'images_gray.h5', 'r') boolSet = np.zeros(gtDataFile["images"].shape[0]).astype(np.bool) boolSet[imageSet] = True return gtDataFile["images"][boolSet,:,:].astype(np.float32) else: image = skimage.io.imread(imagesDir + 'im' + str(imageSet[0]) + '.jpeg') width = image.shape[1] height = image.shape[0] if not loadGray: images = np.zeros([len(imageSet), height, width, 3], dtype=np.float32) else: images = np.zeros([len(imageSet), height, width], dtype=np.float32) for imageit, imageid in enumerate(imageSet): if os.path.isfile(imagesDir + 'im' + str(imageid) + '.jpeg'): image = skimage.io.imread(imagesDir + 'im' + str(imageid) + '.jpeg') else: print("Image " + str(imageid) + " does not exist!") image = np.zeros_like(image) image = image/255.0 if not loadGray: images[imageit, :, :, :] = image else: images[imageit, :, :] = 0.3*image[:,:,0] + 0.59*image[:,:,1] + 0.11*image[:,:,2] return images def readImagesHdf5(imagesDir, loadGray=False): if not loadGray: if os.path.isfile(imagesDir + 'images.h5'): gtDataFile = h5py.File(imagesDir + 'images.h5', 'r') return gtDataFile["images"] else: if os.path.isfile(imagesDir + 'images_gray.h5'): gtDataFile = h5py.File(imagesDir + 'images_gray.h5', 'r') return gtDataFile["images"] def getTriplets(parameterVals, closeDist = 5*np.pi/180, farDist=15*np.pi/180, normConst = 2*np.pi, chunkSize=10): numTrainSet = len(parameterVals) initIdx = np.random.permutation(np.arange(numTrainSet)) orderedIdx = np.argsort(parameterVals) chunkArange = np.arange(chunkSize) shuffledIdx = np.arange(numTrainSet) for i in np.arange(0, numTrainSet, chunkSize): maxi = min(numTrainSet, i + chunkSize) shuffledIdx[i:i+chunkSize] = np.random.permutation(chunkArange)[:maxi] + i idxDist5 = np.roll(np.arange(numTrainSet), int(numTrainSet * closeDist / normConst) ) idxDist15 = np.roll(np.arange(numTrainSet), int(numTrainSet * farDist / normConst) ) anchorIdx = orderedIdx[initIdx] closeIdx = orderedIdx[shuffledIdx[idxDist5][initIdx]] farIdx = orderedIdx[shuffledIdx[idxDist15][initIdx]] return anchorIdx, closeIdx, farIdx def generateExperiment(size, experimentDir, ratio, seed): np.random.seed(seed) data = np.arange(size) np.random.shuffle(data) train = data[0:np.int(size*ratio)] test = data[np.int(size*ratio)::] if not os.path.exists(experimentDir): os.makedirs(experimentDir) np.save(experimentDir + 'train.npy', train) np.save(experimentDir + 'test.npy', test) # saveScatter(xaxis*180/np.pi, yaxis[1], 'Azimuth error (ground-truth)', Azimuth (predicted), filename) import matplotlib.pyplot as plt def saveScatter(xaxis, yaxis, xlabel, ylabel, filename): plt.ioff() fig = plt.figure() plt.scatter(xaxis, yaxis) plt.xlabel('Elevation (degrees)') plt.ylabel('Angular error') x1,x2,y1,y2 = plt.axis() plt.axis((0,90,-90,90)) plt.title('Performance scatter plot') fig.savefig(filename) plt.close(fig) #Method from https://github.com/adamlwgriffiths/Pyrr/blob/master/pyrr/geometry.py def create_cube(scale=(1.0,1.0,1.0), st=False, rgba=np.array([1.,1.,1.,1.]), dtype='float32', type='triangles'): """Returns a Cube reading for rendering.""" shape = [24, 3] rgba_offset = 3 width, height, depth = scale # half the dimensions width /= 2.0 height /= 2.0 depth /= 2.0 vertices = np.array([ # front # top right ( width, height, depth,), # top left (-width, height, depth,), # bottom left (-width,-height, depth,), # bottom right ( width,-height, depth,), # right # top right ( width, height,-depth), # top left ( width, height, depth), # bottom left ( width,-height, depth), # bottom right ( width,-height,-depth), # back # top right (-width, height,-depth), # top left ( width, height,-depth), # bottom left ( width,-height,-depth), # bottom right (-width,-height,-depth), # left # top right (-width, height, depth), # top left (-width, height,-depth), # bottom left (-width,-height,-depth), # bottom right (-width,-height, depth), # top # top right ( width, height,-depth), # top left (-width, height,-depth), # bottom left (-width, height, depth), # bottom right ( width, height, depth), # bottom # top right ( width,-height, depth), # top left (-width,-height, depth), # bottom left (-width,-height,-depth), # bottom right ( width,-height,-depth), ], dtype=dtype) st_values = None rgba_values = None if st: # default st values st_values = np.tile( np.array([ (1.0, 1.0,), (0.0, 1.0,), (0.0, 0.0,), (1.0, 0.0,), ], dtype=dtype), (6,1,) ) if isinstance(st, bool): pass elif isinstance(st, (int, float)): st_values *= st elif isinstance(st, (list, tuple, np.ndarray)): st = np.array(st, dtype=dtype) if st.shape == (2,2,): # min / max st_values *= st[1] - st[0] st_values += st[0] elif st.shape == (4,2,): # per face st values specified manually st_values[:] = np.tile(st, (6,1,)) elif st.shape == (6,2,): # st values specified manually st_values[:] = st else: raise ValueError('Invalid shape for st') else: raise ValueError('Invalid value for st') shape[-1] += st_values.shape[-1] rgba_offset += st_values.shape[-1] if len(rgba) > 0: # default rgba values rgba_values = np.tile(np.array([1.0, 1.0, 1.0, 1.0], dtype=dtype), (24,1,)) if isinstance(rgba, bool): pass elif isinstance(rgba, (int, float)): # int / float expands to RGBA with all values == value rgba_values *= rgba elif isinstance(rgba, (list, tuple, np.ndarray)): rgba = np.array(rgba, dtype=dtype) if rgba.shape == (3,): rgba_values = np.tile(rgba, (24,1,)) elif rgba.shape == (4,): rgba_values[:] = np.tile(rgba, (24,1,)) elif rgba.shape == (4,3,): rgba_values = np.tile(rgba, (6,1,)) elif rgba.shape == (4,4,): rgba_values = np.tile(rgba, (6,1,)) elif rgba.shape == (6,3,): rgba_values = np.repeat(rgba, 4, axis=0) elif rgba.shape == (6,4,): rgba_values = np.repeat(rgba, 4, axis=0) elif rgba.shape == (24,3,): rgba_values = rgba elif rgba.shape == (24,4,): rgba_values = rgba else: raise ValueError('Invalid shape for rgba') else: raise ValueError('Invalid value for rgba') shape[-1] += rgba_values.shape[-1] data = np.empty(shape, dtype=dtype) data[:,:3] = vertices if st_values is not None: data[:,3:5] = st_values if rgba_values is not None: data[:,rgba_offset:] = rgba_values if type == 'triangles': # counter clockwise # top right -> top left -> bottom left # top right -> bottom left -> bottom right indices = np.tile(np.array([0, 1, 2, 0, 2, 3], dtype='int'), (6,1)) for face in range(6): indices[face] += (face * 4) indices.shape = (-1,) return data, indices
70,535
44.477756
456
py
inversegraphics
inversegraphics-master/extract.py
import matplotlib # matplotlib.use('Agg') import bpy import numpy import matplotlib.pyplot as plt width = 110 height = 110 scene = bpy.data.scenes[0] scene.render.resolution_x = width #perhaps set resolution in code scene.render.resolution_y = height scene.render.resolution_percentage = 100 bpy.data.objects['Cube'].pass_index = 1 scene.render.layers[0].use_pass_object_index = True bpy.context.scene.use_nodes = True tree = bpy.context.scene.node_tree rl = tree.nodes[1] links = tree.links v = tree.nodes.new('CompositorNodeViewer') v.location = 750,210 v.use_alpha = False links.new(rl.outputs['Image'], v.inputs['Image']) # link Image output to Viewer input outnode = tree.nodes.new('CompositorNodeOutputFile') outnode.base_path = 'indexob.png' links.new(rl.outputs['IndexOB'], outnode.inputs['Image']) bpy.data.objects['Cube'].pass_index = 1 bpy.ops.render.render( write_still=False ) blendImage = bpy.data.images['Render Result'] image = numpy.flipud(numpy.array(blendImage.extract_render(scene=scene)).reshape([height,width,4]))[:,:,0:3] # # blendImage2 = bpy.data.images['Viewer Node'] # # image2 = numpy.flipud(numpy.array(blendImage2.extract_render(scene=scene)).reshape([256,256,4]))[:,:,0:3] print("DOne")
1,234
24.204082
108
py
inversegraphics
inversegraphics-master/lasagne_visualize.py
from itertools import product from lasagne.layers import get_output import matplotlib.pyplot as plt import numpy as np import theano import theano.tensor as T def plot_loss(net): train_loss = [row['train_loss'] for row in net.train_history_] valid_loss = [row['valid_loss'] for row in net.train_history_] plt.plot(train_loss, label='train loss') plt.plot(valid_loss, label='valid loss') plt.xlabel('epoch') plt.ylabel('loss') plt.legend(loc='best') def plot_conv_weights(layer, figsize=(6, 6)): """Plot the weights of a specific layer. Only really makes sense with convolutional layers. Parameters ---------- layer : lasagne.layers.Layer """ W = layer.W.get_value() shape = W.shape nrows = np.ceil(np.sqrt(shape[0])).astype(int) ncols = nrows for feature_map in range(shape[1]): figs, axes = plt.subplots(nrows, ncols, figsize=figsize) for ax in axes.flatten(): ax.set_xticks([]) ax.set_yticks([]) ax.axis('off') for i, (r, c) in enumerate(product(range(nrows), range(ncols))): if i >= shape[0]: break axes[r, c].imshow(W[i, feature_map], cmap='gray', interpolation='nearest') def plot_conv_activity(layer, x, figsize=(6, 8)): """Plot the acitivities of a specific layer. Only really makes sense with layers that work 2D data (2D convolutional layers, 2D pooling layers ...). Parameters ---------- layer : lasagne.layers.Layer x : numpy.ndarray Only takes one sample at a time, i.e. x.shape[0] == 1. """ if x.shape[0] != 1: raise ValueError("Only one sample can be plotted at a time.") # compile theano function xs = T.tensor4('xs').astype(theano.config.floatX) get_activity = theano.function([xs], get_output(layer, xs)) activity = get_activity(x) shape = activity.shape nrows = np.ceil(np.sqrt(shape[1])).astype(int) ncols = nrows figs, axes = plt.subplots(nrows + 1, ncols, figsize=figsize) axes[0, ncols // 2].imshow(1 - x[0][0], cmap='gray', interpolation='nearest') axes[0, ncols // 2].set_title('original') for ax in axes.flatten(): ax.set_xticks([]) ax.set_yticks([]) ax.axis('off') for i, (r, c) in enumerate(product(range(nrows), range(ncols))): if i >= shape[1]: break ndim = activity[0][i].ndim if ndim != 2: raise ValueError("Wrong number of dimensions, image data should " "have 2, instead got {}".format(ndim)) axes[r + 1, c].imshow(-activity[0][i], cmap='gray', interpolation='nearest') def occlusion_heatmap(net, x, target, square_length=7): """An occlusion test that checks an image for its critical parts. In this function, a square part of the image is occluded (i.e. set to 0) and then the net is tested for its propensity to predict the correct label. One should expect that this propensity shrinks of critical parts of the image are occluded. If not, this indicates overfitting. Depending on the depth of the net and the size of the image, this function may take awhile to finish, since one prediction for each pixel of the image is made. Currently, all color channels are occluded at the same time. Also, this does not really work if images are randomly distorted by the batch iterator. See paper: Zeiler, Fergus 2013 Parameters ---------- net : NeuralNet instance The neural net to test. x : np.array The input data, should be of shape (1, c, x, y). Only makes sense with image data. target : int The true value of the image. If the net makes several predictions, say 10 classes, this indicates which one to look at. square_length : int (default=7) The length of the side of the square that occludes the image. Must be an odd number. Results ------- heat_array : np.array (with same size as image) An 2D np.array that at each point (i, j) contains the predicted probability of the correct class if the image is occluded by a square with center (i, j). """ if (x.ndim != 4) or x.shape[0] != 1: raise ValueError("This function requires the input data to be of " "shape (1, c, x, y), instead got {}".format(x.shape)) if square_length % 2 == 0: raise ValueError("Square length has to be an odd number, instead " "got {}.".format(square_length)) num_classes = net.layers_[-1].num_units img = x[0].copy() bs, col, s0, s1 = x.shape heat_array = np.zeros((s0, s1)) pad = square_length // 2 + 1 x_occluded = np.zeros((s1, col, s0, s1), dtype=img.dtype) probs = np.zeros((s0, s1, num_classes)) # generate occluded images for i in range(s0): # batch s1 occluded images for faster prediction for j in range(s1): x_pad = np.pad(img, ((0, 0), (pad, pad), (pad, pad)), 'constant') x_pad[:, i:i + square_length, j:j + square_length] = 0. x_occluded[j] = x_pad[:, pad:-pad, pad:-pad] y_proba = net.predict_proba(x_occluded) probs[i] = y_proba.reshape(s1, num_classes) # from predicted probabilities, pick only those of target class for i in range(s0): for j in range(s1): heat_array[i, j] = probs[i, j, target] return heat_array def plot_occlusion(net, X, target, square_length=7, figsize=(9, None)): """Plot which parts of an image are particularly import for the net to classify the image correctly. See paper: Zeiler, Fergus 2013 Parameters ---------- net : NeuralNet instance The neural net to test. X : numpy.array The input data, should be of shape (b, c, 0, 1). Only makes sense with image data. target : list or numpy.array of ints The true values of the image. If the net makes several predictions, say 10 classes, this indicates which one to look at. If more than one sample is passed to X, each of them needs its own target. square_length : int (default=7) The length of the side of the square that occludes the image. Must be an odd number. figsize : tuple (int, int) Size of the figure. Plots ----- Figure with 3 subplots: the original image, the occlusion heatmap, and both images super-imposed. """ if (X.ndim != 4): raise ValueError("This function requires the input data to be of " "shape (b, c, x, y), instead got {}".format(X.shape)) num_images = X.shape[0] if figsize[1] is None: figsize = (figsize[0], num_images * figsize[0] / 3) figs, axes = plt.subplots(num_images, 3, figsize=figsize) for ax in axes.flatten(): ax.set_xticks([]) ax.set_yticks([]) ax.axis('off') for n in range(num_images): heat_img = occlusion_heatmap( net, X[n:n + 1, :, :, :], target[n], square_length ) ax = axes if num_images == 1 else axes[n] img = X[n, :, :, :].mean(0) ax[0].imshow(-img, interpolation='nearest', cmap='gray') ax[0].set_title('image') ax[1].imshow(-heat_img, interpolation='nearest', cmap='Reds') ax[1].set_title('critical parts') ax[2].imshow(-img, interpolation='nearest', cmap='gray') ax[2].imshow(-heat_img, interpolation='nearest', cmap='Reds', alpha=0.6) ax[2].set_title('super-imposed')
7,737
31.512605
78
py
inversegraphics
inversegraphics-master/light_probes.py
from math import sin, cos, ceil, floor, pi import importlib import bpy import numpy as np import mathutils from contextlib import contextmanager from uuid import uuid4 from bpy.utils import register_module, unregister_module from bpy import props as p import json import ipdb import matplotlib.pyplot as plt # bl_info = { # "name": "Lightprobe", # "description": "Gives ability to add light probes to a cycles render. \ # Light probes sample incoming light at that location and generate 9 \ # coefficients that can be used to quickly simluate that lighting in a real-time \ # game engine.", # "category": "Object", # "author": "Andrew Moffat", # "version": (1, 0), # "blender": (2, 7, 1) # } JSON_FILE_NAME = "lightprobes.json" FAILSAFE_OFFSET = 0.00001 import cv2 spherical_harmonics_coeffs = np.array([ 0.282095, 0.488603 , 0.488603 , 0.488603, 1.092548, 1.092548, 0.315392 , 1.092548 , 0.546274 ]) def processSphericalEnvironmentMap(envMapTexture): mask = np.ones([envMapTexture.shape[0],envMapTexture.shape[1]]).astype(np.uint8) mask[np.int(mask.shape[0]/2), np.int(mask.shape[1]/2)] = 0 distMask = cv2.distanceTransform(mask, cv2.DIST_L2, 5) envMapTexture[distMask > mask.shape[0]/2,:] = 0 envMapMean = envMapTexture[distMask <= mask.shape[0]/2].mean() envMapTexture[distMask <= mask.shape[0]/2, :] = envMapTexture[distMask <= mask.shape[0]/2] return envMapTexture, envMapMean def is_lightprobe(ob): return ob.name.startswith("lightprobe-") def all_active_lightprobes(): for ob in bpy.context.scene.objects: if is_lightprobe(ob): yield ob def get_all_lightprobe_data(): all_data = [] scale_by = bpy.context.scene.unit_settings.scale_length for probe in all_active_lightprobes(): coeffs = get_coeff_prop(probe) if not coeffs: continue data = {} data["loc"] = [p*scale_by for p in list(probe.location)] data["name"] = probe.name or None data["coeffs"] = coeffs all_data.append(data) return all_data def get_or_create_probe_file(): if JSON_FILE_NAME not in bpy.data.texts: bpy.data.texts.new(JSON_FILE_NAME) f = bpy.data.texts[JSON_FILE_NAME] return f def write_lightprobe_data(data): f = get_or_create_probe_file() f.clear() data = json.dumps(data, indent=4, sort_keys=True) f.write(data) # def fetch_integration_callback(name): # parts = name.split(".") # fn_name = parts[-1] # module_name = ".".join(parts[:-1]) # # try: # module = importlib.import_module(module_name) # except: # return None # else: # fn = getattr(module, fn_name, None) # return fn # # name, theta_res, phi_res, samples = # # def pre_bake_hook(name, context, probe): # fn = fetch_integration_callback(name) # ret = None # if fn: # ret = fn(context, probe) # return ret # # # def post_bake_hook(name, context, data, pre_bake_data): # fn = fetch_integration_callback(name) # if fn: # fn(context, data, pre_bake_data) def hide_object(ob): """ hides an object from cycles rendering, and returns a function that, when called, will restore the visibility """ v = ob.cycles_visibility names = ("shadow", "camera", "glossy", "diffuse", "transmission", "scatter") for name in names: setattr(v, name, False) def set_coeff_prop(ob, coeffs): """ sets our SH coeffs onto an object datablock. we can't use the python dictionary that we've generated, so we'll flatten it to something that can be stored in a datablock """ ob["lightprobe_coeffs"] = json.dumps(coeffs) def get_coeff_prop(ob): """ retrieve our SH coeffs from our object datablock. this is essentially unserializing it to our original data """ coeffs = ob.get("lightprobe_coeffs", None) if not coeffs: return None data = json.loads(coeffs) return data def setup_material(ob): scene = bpy.context.scene mat = bpy.data.materials.new(ob.name) mat.use_nodes = True tree = mat.node_tree for node in list(tree.nodes): tree.nodes.remove(node) #diffuse = tree.nodes.new("ShaderNodeBsdfGlossy") diffuse = tree.nodes.new("ShaderNodeBsdfDiffuse") diffuse.inputs["Color"].default_value = (1, 1, 1, 1) diffuse.inputs["Roughness"].default_value = 1.0 diffuse_out = diffuse.outputs["BSDF"] output = tree.nodes.new("ShaderNodeOutputMaterial") # output.color_space = 'NONE' output_in = output.inputs["Surface"] tree.links.new(diffuse_out, output_in) bake_node = tree.nodes.new("ShaderNodeTexImage") bake_node.label = bake_node.name bake_node.color_space = 'NONE' bake_out = bake_node.outputs["Color"] texture = create_lightmap_image(ob, 128, 128) bake_node.image = texture bake_node.image.colorspace_settings.name = 'Raw' ob.data.uv_textures["lightmap"].active = True color_uvmap = tree.nodes.new("ShaderNodeUVMap") color_uvmap.uv_map = "lightmap" tree.links.new(color_uvmap.outputs["UV"], bake_node.inputs["Vector"]) ob.data.materials.append(mat) @contextmanager def no_interfere_ctx(): old_selected_objects = bpy.context.selected_objects active_object = bpy.context.active_object yield for obj in bpy.context.selected_objects: obj.select = False for obj in old_selected_objects: obj.select = True bpy.context.scene.objects.active = active_object @contextmanager def active(ob): with no_interfere_ctx(): ob.select = True bpy.context.scene.objects.active = ob yield def override_ctx(**kwargs): ctx = bpy.context.copy() ctx.update(kwargs) return ctx def create_lightmap_image(ob, width, height): name = ob.name bpy.ops.image.new(override_ctx(object=ob), name=name, width=width, height=height, alpha=False, float=False) return bpy.data.images[name] def get_lightmap(ob): return bpy.data.images[ob.name] def add_lightprobe(): with no_interfere_ctx(): bpy.ops.mesh.primitive_cube_add() probe = bpy.context.object for _ in range(len(probe.data.uv_layers)): bpy.ops.mesh.uv_texture_remove() bpy.ops.mesh.uv_texture_add() probe.data.uv_layers[0].name = "lightmap" bpy.ops.uv.lightmap_pack(PREF_CONTEXT="ALL_FACES", PREF_PACK_IN_ONE=True, PREF_NEW_UVLAYER=False, PREF_APPLY_IMAGE=False, PREF_IMG_PX_SIZE=512, PREF_BOX_DIV=12, PREF_MARGIN_DIV=0.1) bpy.ops.object.modifier_add(type="SUBSURF") bpy.context.object.modifiers["Subsurf"].levels = 4 bpy.ops.object.modifier_add(type="TRIANGULATE") bpy.ops.object.convert(target='MESH') probe.scale = mathutils.Vector((0.3, 0.3, 0.3)) bpy.ops.object.shade_smooth() probe.name = "lightprobe-" + uuid4().hex return probe def bake(ob): with active(ob): cycles = bpy.context.scene.cycles old_samples = cycles.samples name, theta_res, phi_res, samples = probeProperties() cycles.samples = samples bpy.ops.object.bake(type="COMBINED") cycles.samples = old_samples def get_lightprobe_coefficients(probe, theta_res, phi_res): probe.data.calc_tessface() bake(probe) lightmap = get_lightmap(probe) return get_all_coefficients(probe, lightmap, theta_res, phi_res) def sample_image(channels, width, height, pixel_data, loc): """ samples a blender location at a particular xy integer location """ x, y = loc[0], loc[1] pix_loc = int((y * width * channels) + x * channels) r = pixel_data[pix_loc + 0] g = pixel_data[pix_loc + 1] b = pixel_data[pix_loc + 2] return mathutils.Color((r, g, b)) def bilinear_interpolate(image, uv): """ performs bilinear interpolation of a blender image using texture-space uv coordinates. the boundary conditions are to extend the edges """ lightmap_size = mathutils.Vector(image.size) width, height = lightmap_size[0], lightmap_size[1] px_x, px_y = 1.0/width, 1.0/height half_px_x, half_px_y = 1.0/(2*width), 1.0/(2*height) left_coord = ceil(width * (uv[0] - half_px_x) - 1) right_coord = ceil(width * (uv[0] + half_px_x) - 1) bottom_coord = ceil(height * (uv[1] - half_px_y) - 1) top_coord = ceil(height * (uv[1] + half_px_y) - 1) # these are asking how much of 1-pixel (in uv space) has our uv coordinate # traversed, starting at the left/bottom pixel boundary lerp_x = (uv[0] - (left_coord + 0.5) / width) / px_x lerp_y = (uv[1] - (bottom_coord + 0.5) / height) / px_y # boundary conditions if right_coord + 1 > width: right_coord = left_coord if left_coord < 0: left_coord = right_coord if top_coord + 1 > height: top_coord = bottom_coord if bottom_coord < 0: bottom_coord = top_coord ll_uv = mathutils.Vector((left_coord, bottom_coord)) lr_uv = mathutils.Vector((right_coord, bottom_coord)) ur_uv = mathutils.Vector((right_coord, top_coord)) ul_uv = mathutils.Vector((left_coord, top_coord)) pixel_data = image.pixels[:] chan = image.channels width, height = image.size lower_left = mathutils.Vector(sample_image(chan, width, height, pixel_data, ll_uv)) lower_right = mathutils.Vector(sample_image(chan, width, height, pixel_data, lr_uv)) upper_right = mathutils.Vector(sample_image(chan, width, height, pixel_data, ur_uv)) upper_left = mathutils.Vector(sample_image(chan, width, height, pixel_data, ul_uv)) del pixel_data top = upper_left.lerp(upper_right, lerp_x) bottom = lower_left.lerp(lower_right, lerp_x) color = mathutils.Color(bottom.lerp(top, lerp_y)) return color # http://en.wikipedia.org/wiki/M%C3%B6ller%E2%80%93Trumbore_intersection_algorithm def triangle_intersection(v1, v2, v3, ray, origin): """ performs moller-trumbore ray-triangle intersection and returns barycentric coordinates if an intersection exists, None otherwise """ epsilon = 0.000001 edge1 = v2 - v1 edge2 = v3 - v1 P = ray.cross(edge2) det = edge1.dot(P) if det > -epsilon and det < epsilon: return None inv_det = 1.0 / det T = origin - v1 u = T.dot(P) * inv_det if u < 0 or u > 1: return None Q = T.cross(edge1) v = ray.dot(Q) * inv_det if v < 0 or u + v > 1: return None t = edge2.dot(Q) * inv_det if t > epsilon: w = 1 - u - v return mathutils.Vector((u, v, w)) return None def get_glsl_coefficients(coeffs): """ a convenience function for testing SH coefficients in the shader provided by the opengl orange book, second edition """ tmpl = "const vec3 L%d%s%d = vec3(%f, %f, %f);" lines = [] for l, mdata in coeffs.items(): for m, color in mdata.items(): sign = "" if m < 0: sign = "m" line = tmpl % (l, sign, abs(m), color[0], color[1], color[2]) lines.append(line) return "\n".join(lines) def get_all_coefficients(ob, lightmap, theta_res, phi_res): """ returns all SH coefficients. theta_res and phi_res are the sampling resolutions for theta (zenith) and phi (azimuth) respectively. theta ranges from 0-pi, while phi ranges from 0-2pi """ mapping = {} for l, m in spherical_harmonics.keys(): color = get_coefficients(ob, lightmap, l, m, theta_res, phi_res) mapping.setdefault(l, {})[m] = color return mapping def get_coefficients(ob, lightmap, l, m, theta_res, phi_res): """ returns the RGB spherical harmonic coefficients for a given l and m """ c = mathutils.Color((0, 0, 0)) harmonic = spherical_harmonics[(l, m)] # plt.imsave("lightmap.png", lightmap) for theta in (pi * y / float(theta_res) for y in range(theta_res)): for phi in (pi * 2 * x / float(phi_res) for x in range(phi_res)): color = sample_icosphere_color(ob, lightmap, theta, phi) c += (color * harmonic(theta, phi) * sin(theta) / (theta_res * phi_res)) return c.r, c.g, c.b # http://cseweb.ucsd.edu/~ravir/papers/envmap/envmap.pdf spherical_harmonics = { (0, 0): lambda theta, phi: 0.282095, (1, 1): lambda theta, phi: 0.488603 * sin(theta) * cos(phi), (1, 0): lambda theta, phi: 0.488603 * cos(theta), (1, -1): lambda theta, phi: 0.488603 * sin(theta) * sin(phi), (2, -2): lambda theta, phi: 1.092548 * sin(theta) * cos(phi) * sin(theta) * sin(phi), (2, -1): lambda theta, phi: 1.092548 * sin(theta) * sin(phi) * cos(theta), (2, 0): lambda theta, phi: 0.315392 * (3 * cos(theta)**2 - 1), (2, 1): lambda theta, phi: 1.092548 * sin(theta) * cos(phi) * cos(theta), (2, 2): lambda theta, phi: 0.546274 * (((sin(theta) * cos(phi)) ** 2) - ((sin(theta) * sin(phi)) ** 2)) } def SHProjection(envMap, shCoefficients): #Backprojected. phis = 2*ch.pi*np.tile((np.roll(np.arange(envMap.shape[1])[::-1], np.int(envMap.shape[1]/2))/envMap.shape[1]).reshape([1,envMap.shape[1],1]), [envMap.shape[0],1,3]) thetas = np.pi*np.tile((np.arange(envMap.shape[0])/envMap.shape[0]).reshape([envMap.shape[0],1,1]), [1,envMap.shape[1],3]) phis = np.mod(phis, 2*np.pi) normalize = 1 pEnvMap = np.zeros([envMap.shape[0],envMap.shape[1], 3, 9]) pEnvMap[:,:,:,0] = shCoefficients[0].reshape([1,1,3]) * spherical_harmonics_coeffs[0] * normalize pEnvMap[:,:,:,1] = shCoefficients[1].reshape([1,1,3]) * spherical_harmonics_coeffs[1]*np.sin(thetas) * np.cos(phis) * normalize pEnvMap[:,:,:,2] = shCoefficients[2].reshape([1,1,3]) * spherical_harmonics_coeffs[2]*np.cos(thetas) * normalize pEnvMap[:,:,:,3] = shCoefficients[3].reshape([1,1,3]) * spherical_harmonics_coeffs[3]*np.sin(thetas)*np.sin(phis) * normalize pEnvMap[:,:,:,4] = shCoefficients[4].reshape([1,1,3]) * spherical_harmonics_coeffs[4]*np.sin(thetas) * np.cos(phis) * np.sin(phis)* np.sin(thetas) * normalize pEnvMap[:,:,:,5] = shCoefficients[5].reshape([1,1,3]) * spherical_harmonics_coeffs[5]*np.sin(thetas) * np.sin(phis) * np.cos(thetas) * normalize pEnvMap[:,:,:,6] = shCoefficients[6].reshape([1,1,3]) * spherical_harmonics_coeffs[6]*(3 * np.cos(thetas)**2 - 1) * normalize pEnvMap[:,:,:,7] = shCoefficients[7].reshape([1,1,3]) * spherical_harmonics_coeffs[7]*np.sin(thetas) * np.cos(phis) * np.cos(thetas) * normalize pEnvMap[:,:,:,8] = shCoefficients[8].reshape([1,1,3]) * spherical_harmonics_coeffs[8]*(((np.sin(thetas) * np.cos(phis)) ** 2) - ((np.sin(thetas) * np.sin(phis)) ** 2)) * normalize return pEnvMap def sphericalHarmonicsZRotation(angle): return np.array([[1,0,0,0,0,0,0,0,0],[0, np.cos(angle), 0, np.sin(angle), 0,0,0,0,0],[0,0,1,0,0,0,0,0,0],[0, -np.sin(angle), 0, np.cos(angle), 0,0,0,0,0],[0,0,0,0,np.cos(2*angle),0,0,0,np.sin(2*angle)],[0,0,0,0,0,np.cos(angle), 0, np.sin(angle),0],[0,0,0,0,0,0,1,0,0],[0,0,0,0,0, -np.sin(angle),0, np.cos(angle),0],[0,0,0,0,-np.sin(2*angle),0,0,0,np.cos(2*angle)]]) import chumpy as ch from chumpy import depends_on, Ch class SphericalHarmonicsZRotation(Ch): dterms = 'angle' def compute_r(self): return np.array([[1,0,0,0,0,0,0,0,0],[0, np.cos(self.angle), 0, np.sin(self.angle), 0,0,0,0,0],[0,0,1,0,0,0,0,0,0],[0, -np.sin(self.angle), 0, np.cos(self.angle), 0,0,0,0,0],[0,0,0,0,np.cos(2*self.angle),0,0,0,np.sin(2*self.angle)],[0,0,0,0,0,np.cos(self.angle), 0, np.sin(self.angle),0],[0,0,0,0,0,0,1,0,0],[0,0,0,0,0, -np.sin(self.angle),0, np.cos(self.angle),0],[0,0,0,0,-np.sin(2*self.angle),0,0,0,np.cos(2*self.angle)]]) #No need to make it differentiable for now. def compute_dr_wrt(self, wrt): return None import chumpy as ch def chSphericalHarmonicsZRotation(angle): return SphericalHarmonicsZRotation(angle=angle) def getEnvironmentMapCoefficients(envMap, normalize, phiOffset, type): """ returns the RGB spherical harmonic coefficients for a given l and m """ if type == 'equirectangular': phis = 2*np.pi*np.tile((np.roll(np.arange(envMap.shape[1])[::-1], np.int(envMap.shape[1]/2))/envMap.shape[1]).reshape([1,envMap.shape[1],1]), [envMap.shape[0],1,3]) thetas = np.pi*np.tile((np.arange(envMap.shape[0])/envMap.shape[0]).reshape([envMap.shape[0],1,1]), [1,envMap.shape[1],3]) elif type == 'spherical': vcoords = (-envMap.shape[0]/2 + np.tile(np.arange(envMap.shape[0]).reshape([envMap.shape[0], 1,1]), [1,envMap.shape[1],3]))/(envMap.shape[0]/2) ucoords = (-envMap.shape[1]/2 + np.tile(np.arange(envMap.shape[1]).reshape([1,envMap.shape[1],1]), [envMap.shape[0],1,3]))/(envMap.shape[1]/2) thetas=np.arctan2(vcoords,ucoords) phis=pi*np.sqrt(ucoords*ucoords + vcoords*vcoords) else: print("Environment map format not recognized") return phis = np.mod(phis + phiOffset, 2*np.pi) L = np.zeros([9,3]) # ipdb.set_trace() L[0] = np.sum(envMap * spherical_harmonics_coeffs[0] *np.sin(thetas) , axis=(0,1)) L[1] = np.sum(envMap *spherical_harmonics_coeffs[1]*np.sin(thetas)*np.sin(thetas) * np.cos(phis), axis=(0,1)) L[2] = np.sum(envMap *spherical_harmonics_coeffs[2]*np.sin(thetas)*np.cos(thetas), axis=(0,1)) L[3] = np.sum(envMap *spherical_harmonics_coeffs[3]*np.sin(thetas)*np.sin(thetas)*np.sin(phis), axis=(0,1)) L[4] = np.sum(envMap *spherical_harmonics_coeffs[4]*np.sin(thetas)*np.sin(thetas) * np.cos(phis) * np.sin(phis)* np.sin(thetas), axis=(0,1)) L[5] = np.sum(envMap *spherical_harmonics_coeffs[5]*np.sin(thetas)*np.sin(thetas) * np.sin(phis) * np.cos(thetas) , axis=(0,1)) L[6] = np.sum(envMap *spherical_harmonics_coeffs[6]*np.sin(thetas)*(3 * np.cos(thetas)**2 - 1), axis=(0,1)) L[7] = np.sum(envMap *spherical_harmonics_coeffs[7]*np.sin(thetas) * np.cos(phis) * np.cos(thetas)* np.sin(thetas), axis=(0,1)) L[8] = np.sum(envMap *spherical_harmonics_coeffs[8]*np.sin(thetas)*(((np.sin(thetas) * np.cos(phis)) ** 2) - ((np.sin(thetas) * np.sin(phis)) ** 2)), axis=(0,1)) L = 2*np.pi*L*np.pi/(envMap.shape[0]*envMap.shape[1]*normalize) return L def sample_icosphere_color(ob, lightmap, theta, phi): """ takes a theta and phi and casts a ray out from the center of an icosphere, bilinearly sampling the surface where the ray intersects """ ray = angle_to_ray(theta, phi) # we extend the ray arbitrarily so it's guaranteed to intersect with the # icosphere, instead of falling short ray *= 100 face, location = find_intersecting_face(ob, ray) # it is possible that we couldn't find an intersecting face if the ray # we shot aligns perfectly with a vertex. in this case, we'll offset the # ray slightly and try again. this should not fail a second time. if face is None: ray.x += FAILSAFE_OFFSET ray.y += FAILSAFE_OFFSET ray.z += FAILSAFE_OFFSET face, location = find_intersecting_face(ob, ray) assert(face is not None) color = sample_lightmap(ob, lightmap, face, location) return color def angle_to_ray(theta, phi): """ converts a spherical coordinate to cartesian coordinate """ x = sin(theta) * cos(phi) y = sin(theta) * sin(phi) z = cos(theta) ray = mathutils.Vector((x, y, z)).normalized() return ray def find_intersecting_face(ob, ray): """ finds the face where a ray from the center of an icosphere intersects """ mesh = ob.data origin = mathutils.Vector() # we'll use scale to ensure that our transform applies to our vertices. # assume scale is uniform and just use x scale scale = ob.scale[0] for face in mesh.tessfaces: v = face.vertices v1 = mesh.vertices[v[0]].co * scale v2 = mesh.vertices[v[1]].co * scale v3 = mesh.vertices[v[2]].co * scale intersection = triangle_intersection(v1, v2, v3, ray, origin) if intersection: return face, intersection # we should never get here, but we may in the case of a ray aligning # perfectly with a vertex. in this case, we'll catch this error up at # the caller return None, None def sample_lightmap(ob, lightmap, face, loc): """ """ mesh = ob.data uvs = mesh.tessface_uv_textures[0].data[face.index] location_uv = loc[0] * uvs.uv1 + loc[1] * uvs.uv2 + loc[2] * uvs.uv3 return bilinear_interpolate(lightmap, location_uv) def bakeOp(context): probe = context.object name, theta_res, phi_res, samples = probeProperties() # ret = pre_bake_hook(scene_settings.pre_bake_hook, context, probe) coeffs = get_lightprobe_coefficients(probe, theta_res, phi_res) set_coeff_prop(probe, coeffs) lp_data = get_all_lightprobe_data() write_lightprobe_data(lp_data) # post_bake_hook(scene_settings.post_bake_hook, context, lp_data, ret) return lp_data # class BakeAllOp(context): # override = context.copy() # for probe in all_active_lightprobes(): # override["object"] = probe # bpy.ops.object.bake_lightprobe(override) # # def resizeAll(context): size = p.FloatProperty(name="Units", default=1) dimensions = mathutils.Vector((self.size, self.size, self.size)) for probe in all_active_lightprobes(): probe.dimensions = dimensions def lightProbeOp(context): probe = add_lightprobe() hide_object(probe) probe.show_x_ray = True with active(probe): setup_material(probe) def probeProperties(): name = "" theta_res = 20 phi_res = 40 samples = 1000 return name, theta_res, phi_res, samples
22,272
33.639191
433
py
inversegraphics
inversegraphics-master/collision.py
__author__ = 'pol' #From http://blender.stackexchange.com/a/9080 import bpy import bmesh from blender_utils import * import blender_utils def bmesh_copy_from_object(obj, objTransf, transform=True, triangulate=True, apply_modifiers=False): assert (obj.type == 'MESH') if apply_modifiers and obj.modifiers: me = obj.to_mesh(bpy.context.scene, True, 'PREVIEW', calc_tessface=False) bm = bmesh.new() bm.from_mesh(me) bpy.data.meshes.remove(me) else: me = obj.data if obj.mode == 'EDIT': bm_orig = bmesh.from_edit_mesh(me) bm = bm_orig.copy() else: bm = bmesh.new() bm.from_mesh(me) # Remove custom data layers to save memory for elem in (bm.faces, bm.edges, bm.verts, bm.loops): for layers_name in dir(elem.layers): if not layers_name.startswith("_"): layers = getattr(elem.layers, layers_name) for layer_name, layer in layers.items(): layers.remove(layer) if transform: bm.transform(objTransf * obj.matrix_world) if triangulate: bmesh.ops.triangulate(bm, faces=bm.faces) return bm def aabb_intersect(matrix_world1, instanceObjs1, matrix_world2, instanceObjs2): minX1, maxX1 = blender_utils.modelWidth(instanceObjs1, matrix_world1) minY1, maxY1 = blender_utils.modelDepth(instanceObjs1, matrix_world1) minZ1, maxZ1 = blender_utils.modelHeight(instanceObjs1, matrix_world1) minX2, maxX2 = blender_utils.modelWidth(instanceObjs2, matrix_world2) minY2, maxY2 = blender_utils.modelDepth(instanceObjs2, matrix_world2) minZ2, maxZ2 = blender_utils.modelHeight(instanceObjs2, matrix_world2) return ((maxX1 > minX2) and (minX1 < maxX2) and (maxY1 > minY2) and (minY1 < maxY2) and (maxZ1 > minZ2) and (minZ1 < maxZ2)) def bmesh_check_intersect_objects(obj, objTransf, obj2, obj2Transf, selectface=False): """ Check if any faces intersect with the other object returns a boolean """ from mathutils.bvhtree import BVHTree assert(obj != obj2) # Triangulate bm = bmesh_copy_from_object(obj, objTransf, transform=True, triangulate=True) bm2 = bmesh_copy_from_object(obj2, obj2Transf, transform=True, triangulate=True) intersect = False BMT1 = BVHTree.FromBMesh(bm) BMT2 = BVHTree.FromBMesh(bm2) overlap_pairs = BMT1.overlap(BMT2) if len(overlap_pairs) > 0: intersect = True if selectface: # deselect everything for both objects bpy.context.scene.objects.active = obj bpy.ops.object.mode_set(mode='EDIT', toggle=False) bpy.ops.mesh.select_all(action='DESELECT') bpy.ops.object.mode_set(mode='OBJECT', toggle=False) bpy.context.scene.objects.active = obj2 bpy.ops.object.mode_set(mode='EDIT', toggle=False) bpy.ops.mesh.select_all(action='DESELECT') bpy.ops.object.mode_set(mode='OBJECT', toggle=False) for each in overlap_pairs: obj.data.polygons[each[0]].select = True obj.update_from_editmode() obj2.data.polygons[each[1]].select = True obj2.update_from_editmode() bm.free() bm2.free() return intersect def instancesIntersect(matrix_world1, instanceObjs1, matrix_world2, instanceObjs2): if aabb_intersect(matrix_world1, instanceObjs1, matrix_world2, instanceObjs2): # print ("AABB intersection!") for mesh1 in instanceObjs1: for mesh2 in instanceObjs2: if bmesh_check_intersect_objects(mesh1, matrix_world1, mesh2, matrix_world2): # print ("There's a MESH intersection!") return True # else: # print ("There's NO intersection!") # print("There's NO intersection!") return False def targetSceneCollision(target, scene, roomName, targetParentInstance): # bpy.ops.mesh.primitive_cube_add for sceneInstance in scene.objects: if sceneInstance.type == 'EMPTY' and sceneInstance != target and sceneInstance.name != roomName and sceneInstance != targetParentInstance: if instancesIntersect(target.matrix_world, target.dupli_group.objects, sceneInstance.matrix_world, sceneInstance.dupli_group.objects): return True return False def targetCubeSceneCollision(target, scene, roomName, targetParentInstance): # bpy.ops.mesh.primitive_cube_add for sceneInstance in scene.objects: if sceneInstance.type == 'MESH' and sceneInstance != target and sceneInstance.name != roomName and sceneInstance != targetParentInstance: if instancesIntersect(mathutils.Matrix.Identity(4), [target], mathutils.Matrix.Identity(4), [sceneInstance]): return True return False def parseSceneCollisions(gtDir, scene_i, target_i, target, scene, targetPosOffset, chObjDistGT, chObjRotationGT, targetParentInstance, roomObj, distRange, rotationRange, distInterval, rotationInterval): scene.cycles.samples = 500 original_matrix_world = target.matrix_world.copy() distBins = np.linspace(distInterval, distRange, distRange/distInterval) rotBins = np.linspace(0, rotationRange, rotationRange / rotationInterval) totalBins = np.meshgrid(distBins, rotBins) totalBins[0] = np.append([0], totalBins[0].ravel()) totalBins[1] = np.append([0], totalBins[1].ravel()) boolBins = np.zeros(len(totalBins[0].ravel())).astype(np.bool) scene.update() for bin_i in range(len(totalBins[0].ravel())): dist = totalBins[0].ravel()[bin_i] rot = totalBins[1].ravel()[bin_i] chObjDistGT[:]= dist chObjRotationGT[:]= rot ignore = False azimuthRot = mathutils.Matrix.Rotation(0, 4, 'Z') target.matrix_world = mathutils.Matrix.Translation(original_matrix_world.to_translation() + mathutils.Vector(targetPosOffset.r)) * azimuthRot * (mathutils.Matrix.Translation(-original_matrix_world.to_translation())) * original_matrix_world if targetCubeSceneCollision(target, scene, roomObj.name, targetParentInstance): # print("Teapot intersects with an object.") ignore = True # pass if not instancesIntersect(mathutils.Matrix.Translation(mathutils.Vector((0, 0, -0.01))), [target], mathutils.Matrix.Identity(4), [targetParentInstance]): # print("Teapot not on table.") ignore = True if instancesIntersect(mathutils.Matrix.Translation(mathutils.Vector((0, 0, +0.02))), [target], mathutils.Matrix.Identity(4), [targetParentInstance]): # print("Teapot interesects supporting object.") ignore = True # ipdb.set_trace() if instancesIntersect(mathutils.Matrix.Identity(4), [target], mathutils.Matrix.Identity(4), [roomObj]): # print("Teapot intersects room") ignore = True boolBins[bin_i] = not ignore # bpy.ops.render.render(write_still=True) # # import imageio # # image = np.array(imageio.imread(scene.render.filepath))[:, :, 0:3] # # from blender_utils import lin2srgb # blender_utils.lin2srgb(image) # # cv2.imwrite(gtDir + 'images/scene' + str(scene_i) + '_' + 'targetidx' + str(target_i) + 'bin' + str(bin_i) + '_ignore' + str(ignore) + '.jpeg', 255 * image[:, :, [2, 1, 0]], [int(cv2.IMWRITE_JPEG_QUALITY), 100]) return boolBins, totalBins
7,497
37.649485
247
py
inversegraphics
inversegraphics-master/blender_utils.py
import bpy import bpy_extras import numpy import numpy as np import mathutils from math import radians import h5py import scipy.io import cv2 import sys import io import os import light_probes import imageio try: import cPickle as pickle except: import pickle import ipdb import re from collision import instancesIntersect inchToMeter = 0.0254 def createMeshFromData(name, verts, faces): # Create mesh and object me = bpy.data.meshes.new(name+'Mesh') ob = bpy.data.objects.new(name, me) # Create mesh from given verts, faces. me.from_pydata(verts, [], faces) # Update mesh with new data me.update() return ob def addLamp(scene, lightAz, lightEl, lightDist, center, lightIntensity): #Add directional light to match spherical harmonics lamp_data = bpy.data.lamps.new(name="point", type='POINT') lamp = bpy.data.objects.new(name="point", object_data=lamp_data) lamp.layers[1] = True lamp.layers[2] = True lampLoc = getRelativeLocation(lightAz, lightAz, lightEl, center) lamp.location = mathutils.Vector((lampLoc[0],lampLoc[1],lampLoc[2])) lamp.data.cycles.use_multiple_importance_sampling = True lamp.data.use_nodes = True lamp.data.node_tree.nodes['Emission'].inputs[1].default_value = lightIntensity scene.objects.link(lamp) def loadData(): #data # fdata = h5py.File('../data/data-all-flipped-cropped-512.mat','r') # data = fdata["data"] data = scipy.io.loadmat('../data/data-all-flipped-cropped-512-scipy.mat')['data'] images = h5py.File('../data/images-all-flipped-cropped-512-color.mat','r') # images = f["images"] # imgs = numpy.array(images) # N = imgs.shape[0] # imgs = imgs.transpose(0,2,3,1) # f = h5py.File('../data/all-flipped-cropped-512-crossval6div2_py-experiment.mat') # experiments = f["experiments_data"] # f = h5py.File('../data/all-flipped-cropped-512-crossval6div2_py-experiment.mat') experiments = scipy.io.loadmat('../data/all-flipped-cropped-512-crossval6all2-experiment.mat') return data, images, experiments['experiments_data'] def loadGroundTruth(rendersDir): lines = [line.strip() for line in open(rendersDir + 'groundtruth.txt')] groundTruthLines = [] imageFiles = [] segmentFiles = [] segmentSingleFiles = [] unoccludedFiles = [] prefixes = [] for instance in lines: parts = instance.split(' ') framestr = '{0:04d}'.format(int(parts[4])) prefix = '' az = float(parts[0]) objAz = float(parts[1]) el = float(parts[2]) objIndex = int(parts[3]) frame = int(parts[4]) sceneNum = int(parts[5]) targetIndex = int(parts[6]) spoutPosX = int(float(parts[7])) spoutPosY = int(float(parts[8])) handlePosX = int(float(parts[9])) handlePosY = int(float(parts[10])) tipPosX = int(float(parts[11])) tipPosY = int(float(parts[12])) spoutOccluded = int(float(parts[13])) handleOccluded = int(float(parts[14])) tipOccluded = int(float(parts[15])) if len(parts) == 17: prefix = parts[16] outfilename = "render" + prefix + "_obj" + str(objIndex) + "_scene" + str(sceneNum) + '_target' + str(targetIndex) + '_' + framestr outfilenamesingle = "render" + prefix + "_obj" + str(objIndex) + "_scene" + str(sceneNum) + '_target' + str(targetIndex) + '_single_' + framestr outfilenameunoccluded = "render" + prefix + "_obj" + str(objIndex) + "_scene" + str(sceneNum) + '_target' + str(targetIndex) + '_unoccluded' + framestr imageFile = rendersDir + "images/" + outfilename + ".png" segmentFile = rendersDir + "images/" + outfilename + "_segment.png" segmentFileSingle = rendersDir + "images/" + outfilenamesingle + "_segment.png" unoccludedFile = rendersDir + "images/" + outfilenameunoccluded + ".png" if os.path.isfile(imageFile): imageFiles = imageFiles + [imageFile] segmentFiles = segmentFiles + [segmentFile] segmentSingleFiles = segmentSingleFiles + [segmentFileSingle] unoccludedFiles = unoccludedFiles + [unoccludedFile] prefixes = prefixes + [prefix] groundTruthLines = groundTruthLines + [[az, objAz, el, objIndex, frame, 0.0, sceneNum, targetIndex, spoutPosX, spoutPosY, handlePosX, handlePosY, tipPosX, tipPosY, spoutOccluded, handleOccluded, tipOccluded]] # groundTruth = numpy.zeros([len(groundTruthLines), 5]) groundTruth = numpy.array(groundTruthLines) # groundTruth = numpy.hstack((groundTruth,numpy.zeros((groundTruth.shape[0],1)))) lines = [line.strip() for line in open(rendersDir + 'occlusions.txt')] for instance in lines: parts = instance.split(' ') prefix = '' if len(parts) == 6: prefix = parts[5] eqPrefixes = [ x==y for (x,y) in zip(prefixes, [prefix]*len(prefixes))] try: index = numpy.where((groundTruth[:, 3] == int(parts[0])) & (groundTruth[:, 4] == int(parts[1])) & (groundTruth[:,6] == int(parts[2])) & (groundTruth[:,7] == int(parts[3])) & (eqPrefixes))[0][0] groundTruth[index, 5] = float(parts[4]) except: print("Problem!") return groundTruth, imageFiles, segmentFiles, segmentSingleFiles, unoccludedFiles, prefixes def modifySpecular(scene, delta): for model in scene.objects: if model.type == 'MESH': for mat in model.data.materials: mat.specular_shader = 'PHONG' mat.specular_intensity = mat.specular_intensity + delta mat.specular_hardness = mat.specular_hardness / 4.0 def makeMaterial(name, diffuse, specular, alpha): mat = bpy.data.materials.new(name) mat.diffuse_color = diffuse mat.diffuse_shader = 'LAMBERT' mat.diffuse_intensity = 1.0 mat.specular_color = specular mat.specular_shader = 'COOKTORR' mat.specular_intensity = 0.5 mat.alpha = alpha mat.ambient = 1 return mat def setMaterial(ob, mat): me = ob.data me.materials.append(mat) def look_at(obj_camera, point): loc_camera = obj_camera.location vecPoint = mathutils.Vector((point[0],point[1],point[2])) direction = vecPoint - loc_camera # point the cameras '-Z' and use its 'Y' as up rot_quat = direction.to_track_quat('-Z', 'Y') # assume we're using euler rotation obj_camera.rotation_euler = rot_quat.to_euler() def look_atZ(obj_camera, point): loc_camera = obj_camera.location vecPoint = mathutils.Vector((point[0],point[1],point[2])) direction = vecPoint - loc_camera # point the cameras '-Z' and use its 'Y' as up rot_quat = direction.to_track_quat('Y', 'Z') # assume we're using euler rotation obj_camera.rotation_euler = rot_quat.to_euler() def modelHeight(objects, transform): maxZ = -999999; minZ = 99999; for model in objects: if model.type == 'MESH': for v in model.data.vertices: if (transform * model.matrix_world * v.co).z > maxZ: maxZ = (transform * model.matrix_world * v.co).z if (transform * model.matrix_world * v.co).z < minZ: minZ = (transform * model.matrix_world * v.co).z return minZ, maxZ def modelDepth(objects, transform): maxY = -999999; minY = 99999; for model in objects: if model.type == 'MESH': for v in model.data.vertices: if (transform * model.matrix_world * v.co).y > maxY: maxY = (transform * model.matrix_world * v.co).y if (transform * model.matrix_world * v.co).y < minY: minY = (transform * model.matrix_world * v.co).y return minY, maxY def modelWidth(objects, transform): maxX = -999999; minX = 99999; for model in objects: if model.type == 'MESH': for v in model.data.vertices: if (transform * model.matrix_world * v.co).x > maxX: maxX = (transform * model.matrix_world * v.co).x if (transform * model.matrix_world * v.co).x < minX: minX = (transform * model.matrix_world * v.co).x return minX, maxX def centerOfGeometry(objects, transform): center = mathutils.Vector((0.0,0.0,0.0)) numVertices = 0.0 for model in objects: if model.type == 'MESH': numVertices = numVertices + len(model.data.vertices) for v in model.data.vertices: center = center + (transform * model.matrix_world * v.co) return center/numVertices def setEulerRotation(scene, eulerVectorRotation): for model in scene.objects: if model.type == 'MESH': model.rotation_euler = eulerVectorRotation scene.update() def rotateMatrixWorld(scene, rotationMat): for model in scene.objects: if model.type == 'MESH': model.matrix_world = rotationMat * model.matrix_world scene.update() def AutoNodeOff(): mats = bpy.data.materials for cmat in mats: cmat.use_nodes=False def AutoNode(): mats = bpy.data.materials for cmat in mats: #print(cmat.name) cmat.use_nodes=True TreeNodes=cmat.node_tree links = TreeNodes.links shader='' for n in TreeNodes.nodes: if n.type == 'ShaderNodeTexImage' or n.type == 'RGBTOBW': TreeNodes.nodes.remove(n) if n.type == 'OUTPUT_MATERIAL': shout = n if n.type == 'BACKGROUND': shader=n if n.type == 'BSDF_DIFFUSE': shader=n if n.type == 'BSDF_GLOSSY': shader=n if n.type == 'BSDF_GLASS': shader=n if n.type == 'BSDF_TRANSLUCENT': shader=n if n.type == 'BSDF_TRANSPARENT': shader=n if n.type == 'BSDF_VELVET': shader=n if n.type == 'EMISSION': shader=n if n.type == 'HOLDOUT': shader=n if cmat.raytrace_mirror.use and cmat.raytrace_mirror.reflect_factor>0.001: print("MIRROR") if shader: if not shader.type == 'BSDF_GLOSSY': print("MAKE MIRROR SHADER NODE") TreeNodes.nodes.remove(shader) shader = TreeNodes.nodes.new('BSDF_GLOSSY') # RGB node shader.location = 0,450 #print(shader.glossy) links.new(shader.outputs[0],shout.inputs[0]) if not shader: shader = TreeNodes.nodes.new('BSDF_DIFFUSE') # RGB node shader.location = 0,450 shout = TreeNodes.nodes.new('OUTPUT_MATERIAL') shout.location = 200,400 links.new(shader.outputs[0],shout.inputs[0]) if shader: textures = cmat.texture_slots for tex in textures: if tex: if tex.texture.type=='IMAGE': img = tex.texture.image #print(img.name) shtext = TreeNodes.nodes.new('ShaderNodeTexImage') shtext.location = -200,400 shtext.image=img if tex.use_map_color_diffuse: links.new(shtext.outputs[0],shader.inputs[0]) if tex.use_map_normal: t = TreeNodes.nodes.new('RGBTOBW') t.location = -0,300 links.new(t.outputs[0],shout.inputs[2]) links.new(shtext.outputs[0],t.inputs[0]) def cleanBPYScene(scene): for blenderScene in bpy.data.scenes: if blenderScene != scene: if len(blenderScene.name) > 7 and blenderScene.name[0:7] == 'teapots': bpy.data.scenes.remove(blenderScene) def addEnvironmentMapWorld(scene): scene.world.use_nodes = True treeNodes=scene.world.node_tree envTextureNode = treeNodes.nodes.new('ShaderNodeTexEnvironment') mappingNode = treeNodes.nodes.new('ShaderNodeMapping') links = treeNodes.links links.new(mappingNode.outputs[0],envTextureNode.inputs[0]) texCoordNode = treeNodes.nodes.new('ShaderNodeTexCoord') links.new(texCoordNode.outputs[0],mappingNode.inputs[0]) rgbToBWNode = treeNodes.nodes.new('ShaderNodeRGBToBW') links.new(envTextureNode.outputs[0],rgbToBWNode.inputs[0]) colorBackgroundNode = treeNodes.nodes.new('ShaderNodeBackground') colorBackgroundNode.name = "WorldBackground" colorBackgroundNode.inputs[0].default_value[0:3] = (1.0,1.0,1.0) links.new(rgbToBWNode.outputs[0],colorBackgroundNode.inputs[0]) # links.new(envTextureNode.outputs[0],colorBackgroundNode.inputs[0]) output_node = treeNodes.nodes['World Output'] links.new(colorBackgroundNode.outputs["Background"],output_node.inputs["Surface"]) # # links.new(rgbToBWNode.outputs[0],treeNodes.nodes['Background'].inputs[0]) # links.new(envTextureNode.outputs[0],treeNodes.nodes['Background'].inputs[0]) # mixShaderNode = treeNodes.nodes.new('ShaderNodeMixShader') # mixShaderNode.name = 'mixShaderNode' # light_pathNode = treeNodes.nodes.new('ShaderNodeLightPath') # light_pathNode.name = 'lightPathNode' # links.new(treeNodes.nodes['Background'].outputs[0],mixShaderNode.inputs[1]) # links.new(colorBackgroundNode.outputs[0],mixShaderNode.inputs[2]) # links.new(light_pathNode.outputs[0],mixShaderNode.inputs[0]) # mixShaderNode.inputs[0].default_value = 0 envTextureNode.color_space="NONE" def setEnviornmentMapStrength(strength, scene): backgroundNode = scene.world.node_tree.nodes['Background'] backgroundNode.inputs[1].default_value = strength def updateEnviornmentMap(envMapFilename, scene): envTextureNode = scene.world.node_tree.nodes['Environment Texture'] if envTextureNode.image != None: envTextureNode.image.user_clear() bpy.data.images.remove(envTextureNode.image) image = bpy.data.images.load(envMapFilename) envTextureNode.image = image envTextureNode.color_space="NONE" def rotateEnviornmentMap(angle, scene): mappingNode = scene.world.node_tree.nodes['Mapping'] mappingNode.rotation[2] = angle def cameraLookingInsideRoom(cameraAzimuth): if cameraAzimuth > 270 and cameraAzimuth < 90: return True return False def deleteInstance(instance): for mesh in instance.dupli_group.objects: mesh.user_clear() bpy.data.objects.remove(mesh) instance.dupli_group.user_clear() bpy.data.groups.remove(instance.dupli_group) instance.user_clear() bpy.data.objects.remove(instance) def deleteObject(object): object.user_clear() bpy.data.objects.remove(object) def placeNewTarget(scene, target, targetPosition): target.layers[1] = True target.layers[2] = True scene.objects.link(target) target.matrix_world = mathutils.Matrix.Translation(targetPosition) # center = centerOfGeometry(target.dupli_group.objects, target.matrix_world) # original_matrix_world = target.matrix_world.copy() # camera = scene.camera # scene.update() # look_at(camera, center) scene.update() def placeCamera(camera, azimuth, elevation, camDistance, center): location = getRelativeLocation(azimuth, elevation, camDistance, center) camera.location = location look_at(camera, center) def getRelativeLocation(azimuth, elevation, distance, center): azimuthRot = mathutils.Matrix.Rotation(radians(-azimuth), 4, 'Z') elevationRot = mathutils.Matrix.Rotation(radians(-elevation), 4, 'X') originalLoc = mathutils.Vector((0,-distance, 0)) location = center + azimuthRot * elevationRot * originalLoc return location def srgb2lin(im): cond1 = im <= 0.04045 im[cond1] = im[cond1]/12.92 im[~cond1] = ((im[~cond1]+0.055)/1.055)** 2.4 return im def lin2srgb(im): cond1 = im <= 0.0031308 im[cond1] = im[cond1]*12.92 im[~cond1] = (im[~cond1]**(1/2.4)*1.055) - 0.055 return im def sortface(f): if len(f) != 4: return f v=[mathutils.Vector(list(p)) for p in f] v2m0=v[2]-v[0] # The normal of the plane v1m0 = v[1] - v[0] n=v1m0.cross(v2m0) #k=DotVecs(v[0],n) #if DotVecs(v[3],n) != k: # raise ValueError("Not Coplanar") # Well, the above test would be a good hint to make triangles. # Get a vector pointing along the plane perpendicular to v[0]-v[2] n2=n.cross(v2m0) # Get the respective distances along that line k=[p.dot(n2) for p in v[1:]] # Check if the vertices are on the proper side cmp = lambda x, y: (x > y) - (x < y) if cmp(k[1],k[0]) == cmp(k[1],k[2]): #print "Bad",v f.v=[f[0],f[2],f[3],f[1]] def createCube(scaleX, scaleY, scaleZ, name): bpy.ops.mesh.primitive_cube_add() cubeObj = bpy.context.object cubeObj.name = name cubeScale = mathutils.Matrix([[scaleX / 2, 0, 0, 0], [0, scaleY / 2, 0, 0], [0, 0, scaleZ / 2, 0], [0, 0, 0, 1]]) cubeObj.data.transform(cubeScale * mathutils.Matrix.Translation(mathutils.Vector((0, 0, 1)))) bpy.context.scene.objects.unlink(cubeObj) return cubeObj def getCubeObj(instance): minX1, maxX1 = modelWidth(instance.dupli_group.objects, mathutils.Matrix.Identity(4)) minY1, maxY1 = modelDepth(instance.dupli_group.objects, mathutils.Matrix.Identity(4)) minZ1, maxZ1 = modelHeight(instance.dupli_group.objects, mathutils.Matrix.Identity(4)) ob = instance.dupli_group.objects[0] bbox = [ob.matrix_world * mathutils.Vector(corner) for corner in ob.bound_box] # bpy.ops.mesh.primitive_cube_add() # cubeObj2 = bpy.context.object # cubeObj2.name = 'cube' + instance.name # # cubeScale = mathutils.Matrix([[(maxX1 - minX1)/ 2, 0, 0, 0], [0, (maxY1 - minY1)/ 2, 0, 0], [0, 0, (maxZ1 - minZ1)/ 2, 0], [0, 0, 0, 1]]) # cubeObj2.data.transform(cubeScale * mathutils.Matrix.Translation(mathutils.Vector((0,0,1)))) # # cubeObj2.data.update() # cubeObj2.data.show_double_sided = True # # cubeObj2.matrix_world = instance.matrix_world mesh = bpy.data.meshes.new('meshCube' + instance.name) cubeObj = bpy.data.objects.new('cube' + instance.name, mesh) # cubeObj.name = 'cube' + instance.name import bmesh bm = bmesh.new() # bm.from_mesh(cubeObj.data) # fill it in from a Mesh bm.verts.ensure_lookup_table() # for v_co in bm.verts.new((v_co)) # bm.verts.new(bbox[0]) bm.verts.new(bbox[1]) bm.verts.new(bbox[2]) bm.verts.new(bbox[3]) bm.verts.new(bbox[4]) bm.verts.new(bbox[5]) bm.verts.new(bbox[6]) bm.verts.new(bbox[7]) faces = [(3, 7, 6, 2), (6, 5, 4, 7), (5, 1, 0, 4), (4, 7, 3, 0), (5, 6, 2, 1), (1, 3, 2, 0), ] # faces = [(1, 2, 3, 0), # (5, 6, 7, 4), # (1, 2, 6, 5), # (0, 3, 7, 4), # (1, 0, 4, 5), # (2, 3, 7, 6), # ] bm.verts.ensure_lookup_table() bm.faces.ensure_lookup_table() for f_idx in faces: bm.faces.new([bm.verts[i] for i in f_idx]) # ipdb.set_trace() # Finish up, write the bmesh back to the mesh bm.to_mesh(cubeObj.data) cubeObj.matrix_world = instance.matrix_world cubeObj.data.update() cubeObj.data.show_double_sided = True return cubeObj def createCubeScene(scene): bpy.ops.scene.new(type='EMPTY') bpy.context.scene.name = "CubeScene" cubeScene = bpy.context.scene cubeScene.camera = scene.camera cubeScene.world = scene.world for sceneInstanceIdx, sceneInstance in enumerate(scene.objects): if sceneInstance.type == 'EMPTY': cubeObj = getCubeObj(sceneInstance) cubeScene.objects.link(cubeObj) cubeScene.update() return cubeScene def captureSceneEnvMap(scene, envMapTexture, roomInstanceNum, rotationOffset, links, treeNodes, teapot, center, targetPosition, width, height, cyclesSamples=3000, gtDir='', train_i=0): bpy.context.screen.scene = scene envMapTexture = cv2.resize(src=envMapTexture, dsize=(360, 180)) # envMapTexture = skimage.transform.resize(images[test_i], [height,width]) envMapGray = 0.3 * envMapTexture[:, :, 0] + 0.59 * envMapTexture[:, :, 1] + 0.11 * envMapTexture[:, :, 2] envMapGrayMean = np.mean(envMapGray, axis=(0, 1)) envMapGrayRGB = np.concatenate([envMapGray[..., None], envMapGray[..., None], envMapGray[..., None]], axis=2) / envMapGrayMean envMapCoeffsNew = light_probes.getEnvironmentMapCoefficients(envMapGrayRGB, 1, 0, 'equirectangular') pEnvMap = light_probes.SHProjection(envMapTexture, envMapCoeffsNew) # pEnvMap = SHProjection(envMapGrayRGB, envMapCoeffs) approxProjection = np.sum(pEnvMap, axis=3).astype(np.float32) # envMapCoeffsNewRE = light_probes.getEnvironmentMapCoefficients(approxProjectionRE, 1, 0, 'equirectangular') # pEnvMapRE = SHProjection(envMapTexture, envMapCoeffsNewRE) # # pEnvMap = SHProjection(envMapGrayRGB, envMapCoeffs) # approxProjectionRE = np.sum(pEnvMapRE, axis=3).astype(np.float32) approxProjection[approxProjection < 0] = 0 cv2.imwrite(gtDir + 'im.exr', approxProjection) # updateEnviornmentMap(envMapFilename, scene) updateEnviornmentMap(gtDir + 'im.exr', scene) rotateEnviornmentMap(-rotationOffset, scene) cv2.imwrite(gtDir + 'sphericalharmonics/envMapProjOr' + str(train_i) + '.jpeg', 255 * approxProjection[:, :, [2, 1, 0]]) cv2.imwrite(gtDir + 'sphericalharmonics/envMapGrayOr' + str(train_i) + '.jpeg', 255 * envMapGrayRGB[:, :, [2, 1, 0]]) links.remove(treeNodes.nodes['lightPathNode'].outputs[0].links[0]) scene.world.cycles_visibility.camera = True scene.camera.data.type = 'PANO' scene.camera.data.cycles.panorama_type = 'EQUIRECTANGULAR' scene.render.resolution_x = 360 # perhaps set resolution in code scene.render.resolution_y = 180 roomInstance = scene.objects[str(roomInstanceNum)] roomInstance.cycles_visibility.camera = False roomInstance.cycles_visibility.shadow = False teapot.cycles_visibility.camera = False #NOt sure if should set to True or False teapot.cycles_visibility.shadow = False # image = cv2.imread(scene.render.filepath) # image = np.float64(cv2.cvtColor(image, cv2.COLOR_BGR2RGB))/255.0 scene.render.image_settings.file_format = 'OPEN_EXR' scene.render.filepath = gtDir + 'sphericalharmonics/envMap' + str(train_i) + '.exr' # bpy.context.user_preferences.system.compute_device_type = 'NONE' # bpy.context.user_preferences.system.compute_device = 'CPU' scene.cycles.samples = 1000 scene.camera.up_axis = 'Z' # placeCamera(scene.camera, 0, 0, 1, ) scene.camera.location = center[:].copy() + targetPosition[:].copy() look_at(scene.camera, center[:].copy() + targetPosition[:].copy() + mathutils.Vector((1, 0, 0))) scene.update() bpy.ops.render.render(write_still=True) imageEnvMap = np.array(imageio.imread(scene.render.filepath))[:, :, 0:3] cv2.imwrite(gtDir + 'sphericalharmonics/envMapCycles' + str(train_i) + '.jpeg', 255 * imageEnvMap[:, :, [2, 1, 0]]) envMapCoeffs = light_probes.getEnvironmentMapCoefficients(imageEnvMap, 1, 0, 'equirectangular') pEnvMap = light_probes.SHProjection(envMapTexture, envMapCoeffs) approxProjection = np.sum(pEnvMap, axis=3) cv2.imwrite(gtDir + 'sphericalharmonics/envMapCyclesProjection' + str(train_i) + '.jpeg', 255 * approxProjection[:, :, [2, 1, 0]]) links.new(treeNodes.nodes['lightPathNode'].outputs[0], treeNodes.nodes['mixShaderNode'].inputs[0]) scene.cycles.samples = cyclesSamples scene.render.filepath = 'opendr_blender.exr' roomInstance.cycles_visibility.camera = True scene.render.image_settings.file_format = 'OPEN_EXR' scene.render.resolution_x = width # perhaps set resolution in code scene.render.resolution_y = height scene.camera.data.type = 'PERSP' scene.world.cycles_visibility.camera = True scene.camera.data.cycles.panorama_type = 'FISHEYE_EQUISOLID' teapot.cycles_visibility.camera = True teapot.cycles_visibility.shadow = True # updateEnviornmentMap(envMapFilename, scene) return envMapCoeffs def setupSceneGroundtruth(scene, width, height, clip_start, cyclesSamples, device_type=None, compute_device=None): scene.render.resolution_x = width # perhaps set resolution in code scene.render.resolution_y = height scene.render.tile_x = height scene.render.tile_y = width scene.cycles.samples = cyclesSamples bpy.context.screen.scene = scene addEnvironmentMapWorld(scene) bpy.context.scene.render.image_settings.quality = 100 scene.render.image_settings.file_format = 'JPEG' # scene.render.filepath = 'image.jpeg' scene.sequencer_colorspace_settings.name = 'sRGB' # scene.display_settings.display_device = 'None' #or sRGB # bpy.context.user_preferences.filepaths.render_cache_directory = '/disk/scratch1/pol/.cache/' #Uncomment if you can use any of these: if device_type is not None: bpy.context.user_preferences.system.compute_device_type = device_type if compute_device is not None: bpy.context.user_preferences.system.compute_device = compute_Device # bpy.context.user_preferences.system.compute_device = compute_device bpy.ops.wm.save_userpref() scene.world.horizon_color = mathutils.Color((1.0, 1.0, 1.0)) scene.camera.data.clip_start = clip_start scene.render.layers['RenderLayer'].use_pass_combined = True scene.layers[0] = True scene.camera.layers[0] = True scene.render.layers.active = scene.render.layers['RenderLayer'] nt = bpy.context.scene.node_tree nt.nodes['Render Layers'].layer = 'RenderLayer' def setupScene(scene, roomInstanceNum, world, camera, width, height, numSamples, useCycles, useGPU, device_type=None, compute_device=None): if useCycles: #Switch Engine to Cycles scene.render.engine = 'CYCLES' if useGPU: bpy.context.scene.cycles.device = 'GPU' bpy.context.user_preferences.system.compute_device_type = device_type bpy.context.user_preferences.system.compute_device = compute_device scene.use_nodes = True AutoNode() cycles = bpy.context.scene.cycles cycles.samples = 2000 cycles.max_bounces = 36 cycles.min_bounces = 4 cycles.caustics_reflective = False cycles.caustics_refractive = False cycles.diffuse_bounces = 36 cycles.glossy_bounces = 12 cycles.transmission_bounces = 2 cycles.volume_bounces = 12 cycles.transparent_min_bounces = 2 cycles.transparent_max_bounces = 2 world.cycles_visibility.camera = True world.use_nodes = True world.cycles.sample_as_light = True world.cycles.sample_map_resolution = 2048 scene.render.threads = 4 scene.render.tile_x = height/2 scene.render.tile_y = width/2 scene.render.image_settings.compression = 0 scene.render.resolution_x = width #perhaps set resolution in code scene.render.resolution_y = height scene.render.resolution_percentage = 100 scene.camera = camera # scene.objects.link(camera) camera.up_axis = 'Y' camera.data.angle = 60 * 180 / numpy.pi camera.data.clip_start = 0.001 camera.data.clip_end = 10 roomInstance = scene.objects[str(roomInstanceNum)] if useCycles: roomInstance.cycles_visibility.shadow = False scene.world = world scene.world.light_settings.distance = 0.1 if not useCycles: scene.render.use_raytrace = False scene.render.use_shadows = False # scene.view_settings.exposure = 5 # scene.view_settings.gamma = 0.5 scene.world.light_settings.use_ambient_occlusion = True scene.world.light_settings.ao_blend_type = 'ADD' scene.world.light_settings.use_indirect_light = True scene.world.light_settings.indirect_bounces = 1 scene.world.light_settings.use_cache = True scene.world.light_settings.ao_factor = 1 scene.world.light_settings.indirect_factor = 1 scene.world.light_settings.gather_method = 'APPROXIMATE' world.light_settings.use_environment_light = False world.light_settings.environment_energy = 0.0 world.horizon_color = mathutils.Color((1.0,1.0,1.0)) # world.light_settings.samples = 20 # world.light_settings.use_ambient_occlusion = False # # world.light_settings.ao_factor = 1 # world.exposure = 1.1 # world.light_settings.use_indirect_light = True scene.sequencer_colorspace_settings.name = 'Linear' scene.display_settings.display_device = 'None' scene.update() bpy.ops.scene.render_layer_add() bpy.ops.scene.render_layer_add() camera.layers[1] = True scene.render.layers[0].use_pass_combined = True scene.render.layers[1].use_pass_object_index = True scene.render.layers[1].use_pass_combined = True camera.layers[2] = True scene.layers[1] = False scene.layers[2] = False scene.layers[0] = True scene.render.layers[0].use = True scene.render.layers[1].use = False scene.render.layers[2].use = False scene.render.use_sequencer = False def addAmbientLightingScene(scene, useCycles): roomName = '' for model in scene.objects: reg = re.compile('(room[0-9]+)') res = reg.match(model.name) if res: roomName = res.groups()[0] roomInstance = scene.objects[roomName] ceilMinX, ceilMaxX = modelWidth(roomInstance.dupli_group.objects, roomInstance.matrix_world) ceilWidth = (ceilMaxX - ceilMinX) ceilMinY, ceilMaxY = modelDepth(roomInstance.dupli_group.objects, roomInstance.matrix_world) ceilDepth = (ceilMaxY - ceilMinY) ceilMinZ, ceilMaxZ = modelHeight(roomInstance.dupli_group.objects, roomInstance.matrix_world) ceilPos = mathutils.Vector(((ceilMaxX + ceilMinX) / 2.0, (ceilMaxY + ceilMinY) / 2.0 , ceilMaxZ)) numLights = int(numpy.floor((ceilWidth-0.2)/1.2)) lightInterval = ceilWidth/numLights for light in range(numLights): lightXPos = light*lightInterval + lightInterval/2.0 lamp_data = bpy.data.lamps.new(name="Rect", type='AREA') lamp = bpy.data.objects.new(name="Rect", object_data=lamp_data) lamp.data.size = 0.2 lamp.data.size_y = ceilDepth - 0.2 lamp.data.shape = 'RECTANGLE' lamp.location = mathutils.Vector((ceilPos.x - ceilWidth/2.0 + lightXPos, ceilPos.y, ceilMaxZ)) lamp.data.energy = 0.0025 if useCycles: lamp.data.cycles.use_multiple_importance_sampling = True lamp.data.use_nodes = True lamp.data.node_tree.nodes['Emission'].inputs[1].default_value = 30 scene.objects.link(lamp) lamp.layers[1] = True lamp.layers[2] = True def view_plane(camd, winx, winy, xasp, yasp): #/* fields rendering */ ycor = yasp / xasp use_fields = False if (use_fields): ycor *= 2 def BKE_camera_sensor_size(p_sensor_fit, sensor_x, sensor_y): #/* sensor size used to fit to. for auto, sensor_x is both x and y. */ if (p_sensor_fit == 'VERTICAL'): return sensor_y; return sensor_x; if (camd.type == 'ORTHO'): #/* orthographic camera */ #/* scale == 1.0 means exact 1 to 1 mapping */ pixsize = camd.ortho_scale else: #/* perspective camera */ sensor_size = BKE_camera_sensor_size(camd.sensor_fit, camd.sensor_width, camd.sensor_height) pixsize = (sensor_size * camd.clip_start) / camd.lens #/* determine sensor fit */ def BKE_camera_sensor_fit(p_sensor_fit, sizex, sizey): if (p_sensor_fit == 'AUTO'): if (sizex >= sizey): return 'HORIZONTAL' else: return 'VERTICAL' return p_sensor_fit sensor_fit = BKE_camera_sensor_fit(camd.sensor_fit, xasp * winx, yasp * winy) if (sensor_fit == 'HORIZONTAL'): viewfac = winx else: viewfac = ycor * winy pixsize /= viewfac #/* extra zoom factor */ pixsize *= 1 #params->zoom #/* compute view plane: # * fully centered, zbuffer fills in jittered between -.5 and +.5 */ xmin = -0.5 * winx ymin = -0.5 * ycor * winy xmax = 0.5 * winx ymax = 0.5 * ycor * winy #/* lens shift and offset */ dx = camd.shift_x * viewfac # + winx * params->offsetx dy = camd.shift_y * viewfac # + winy * params->offsety xmin += dx ymin += dy xmax += dx ymax += dy #/* fields offset */ #if (params->field_second): # if (params->field_odd): # ymin -= 0.5 * ycor # ymax -= 0.5 * ycor # else: # ymin += 0.5 * ycor # ymax += 0.5 * ycor #/* the window matrix is used for clipping, and not changed during OSA steps */ #/* using an offset of +0.5 here would give clip errors on edges */ xmin *= pixsize xmax *= pixsize ymin *= pixsize ymax *= pixsize return xmin, xmax, ymin, ymax def projection_matrix(camd, scene): r = scene.render left, right, bottom, top = view_plane(camd, r.resolution_x, r.resolution_y, 1, 1) farClip, nearClip = camd.clip_end, camd.clip_start Xdelta = right - left Ydelta = top - bottom Zdelta = farClip - nearClip mat = [[0]*4 for i in range(4)] mat[0][0] = nearClip * 2 / Xdelta mat[1][1] = nearClip * 2 / Ydelta mat[2][0] = (right + left) / Xdelta #/* note: negate Z */ mat[2][1] = (top + bottom) / Ydelta mat[2][2] = -(farClip + nearClip) / Zdelta mat[2][3] = -1 mat[3][2] = (-2 * nearClip * farClip) / Zdelta # ipdb.set_trace() # return sum([c for c in mat], []) projMat = mathutils.Matrix(mat) return projMat.transposed() def image_projection(scene, point): p4d = mathutils.Vector.Fill(4, 1) p4d.x = point.x p4d.y = point.y p4d.z = point.z projectionMat = projection_matrix(scene.camera.data, scene) proj = projectionMat * scene.camera.matrix_world.inverted() * p4d return [scene.render.resolution_x*(proj.x/proj.w + 1)/2, scene.render.resolution_y*(proj.y/proj.w + 1)/2] def image_project(scene, camera, point): co_2d = bpy_extras.object_utils.world_to_camera_view(scene, camera, point) # print("2D Coords:", co_2d) # If you want pixel coords render_scale = scene.render.resolution_percentage / 100 render_size = ( int(scene.render.resolution_x * render_scale), int(scene.render.resolution_y * render_scale)) return (round(co_2d.x * render_size[0]), round(co_2d.y * render_size[1])) #Need to verify! def closestCameraIntersection(scene, point): for instance in scene.objects: if instance.type == 'EMPTY' and instance.dupli_type == 'GROUP': instanceLoc = numpy.array(instance.location) camLoc = numpy.array(scene.camera.location) pointLoc = numpy.array(point) invInstanceTransf = instance.matrix_world.inverted() localCamTmp = invInstanceTransf * scene.camera.location if numpy.linalg.norm(instanceLoc - camLoc) < numpy.linalg.norm(pointLoc - camLoc) and (instanceLoc - camLoc).dot(pointLoc - camLoc) > 0: for mesh in instance.dupli_group.objects: if mesh.type == 'MESH': invMeshTransf = mesh.matrix_world.inverted() localCam = invMeshTransf * localCamTmp localPoint = invMeshTransf * invInstanceTransf * point location, normal, index = mesh.ray_cast(localCam, localPoint) if index != -1: #Success. return True return False def sceneIntersection(scene, point): result, object, matrix, location, normal = scene.ray_cast(scene.camera.location, point) return result # def flattenMesh(mesh, transform): # return # def flattenInstance(instance, transform): def setupBlender(teapot, width, height, angle, clip_start, clip_end, camDistance): cam = bpy.data.cameras.new("MainCamera") camera = bpy.data.objects.new("MainCamera", cam) world = bpy.data.worlds.new("MainWorld") bpy.ops.scene.new() bpy.context.scene.name = 'Main Scene' scene = bpy.context.scene scene.objects.link(teapot) scene.camera = camera camera.up_axis = 'Y' camera.data.angle = angle camera.data.clip_start = clip_start camera.data.clip_end = clip_end scene.world = world world.light_settings.use_environment_light = False world.light_settings.environment_energy = 0.0 world.horizon_color = mathutils.Color((0.0,0.0,0.0)) scene.render.resolution_x = width #perhaps set resolution in code scene.render.resolution_y = height scene.render.resolution_percentage = 100 scene.update() lamp_data2 = bpy.data.lamps.new(name="LampBotData", type='POINT') lamp2 = bpy.data.objects.new(name="LampBot", object_data=lamp_data2) lamp2.location = mathutils.Vector((0,0,1.5)) lamp2.data.energy = 1 scene.objects.link(lamp2) lamp_data2 = bpy.data.lamps.new(name="LampBotData2", type='POINT') lamp2 = bpy.data.objects.new(name="LampBot2", object_data=lamp_data2) lamp2.location = mathutils.Vector((0,0,-1.5)) lamp2.data.energy = 0.5 scene.objects.link(lamp2) bpy.context.screen.scene = scene # teapot.matrix_world = mathutils.Matrix.Translation(mathutils.Vector((0,0,0))) center = centerOfGeometry(teapot.dupli_group.objects, teapot.matrix_world) azimuth = 90 elevation = 25 azimuthRot = mathutils.Matrix.Rotation(radians(-azimuth), 4, 'Z') elevationRot = mathutils.Matrix.Rotation(radians(-elevation), 4, 'X') originalLoc = mathutils.Vector((0,-camDistance, 0)) location = center + azimuthRot * elevationRot * originalLoc camera.location = location scene.update() look_at(camera, center) scene.update() return scene ########################################################### # # Round values of the 3D vector # ########################################################### def r3d(v): return round(v[0],6), round(v[1],6), round(v[2],6) ########################################################### # # Round values of the 2D vector # ########################################################### def r2d(v): return round(v[0],6), round(v[1],6) ########################################################### # # Convert object name to be suitable for C definition # ########################################################### def clearName(name): tmp=name.upper() ret="" for i in tmp: if (i in " ./\-+#$%^!@"): ret=ret+"_" else: ret=ret+i return ret ########################################################### # # Build data for each object (MESH) # ########################################################### def getDrWrtAzimuth(SqE_raw, rotation): rot, rot_dr = cv2.Rodrigues(np.array(rotation)) a = mathutils.Matrix(rot).to_euler()[2] a2 = np.arctan(rot[1,0]/rot[0,0]) b = mathutils.Matrix(rot).to_euler()[1] b2 = np.arctan(-rot[2,0]/np.sqrt(rot[2,1]**2 + rot[2,2]**2)) g = mathutils.Matrix(rot).to_euler()[0] g2 = np.arctan(rot[2,1]/rot[2,2]) dra11 = -np.sin(a)*np.cos(b) dra12 = -np.sin(a)*np.sin(b)*np.sin(g) - np.sin(a)*np.cos(g) dra13 = -np.sin(a)*np.sin(b)*np.cos(g) + np.cos(a)*np.sin(g) dra21 = np.cos(a)*np.cos(b) dra22 = np.cos(a)*np.sin(b)*np.sin(g) - np.sin(a)*np.cos(g) dra23 = np.cos(a)*np.sin(b)*np.cos(g) + np.sin(a)*np.cos(g) dra31 = 0 dra32 = 0 dra33 = 0 rotwrtaz = np.array([[dra11,dra12,dra13], [dra21,dra22,dra23], [dra31,dra32,dra33]]) # ipdb.set_trace() drazimuth = np.dot(SqE_raw.dr_wrt(rotation), np.dot(rot_dr , rotwrtaz.ravel()))/(2*400*400) return drazimuth, np.array([a,g,b]) from matplotlib import pyplot as plt from mpl_toolkits.mplot3d import Axes3D from matplotlib.patches import FancyArrowPatch from mpl_toolkits.mplot3d import proj3d class Arrow3D(FancyArrowPatch): def __init__(self, xs, ys, zs, *args, **kwargs): FancyArrowPatch.__init__(self, (0,0), (0,0), *args, **kwargs) self._verts3d = xs, ys, zs def draw(self, renderer): xs3d, ys3d, zs3d = self._verts3d xs, ys, zs = proj3d.proj_transform(xs3d, ys3d, zs3d, renderer.M) self.set_positions((xs[0],ys[0]),(xs[1],ys[1])) FancyArrowPatch.draw(self, renderer) def setObjectDiffuseColor(object, color): for mesh in object.dupli_group.objects: if mesh.type == 'MESH': for mat in mesh.data.materials: mat.diffuse_color = color
41,350
35.529152
220
py
inversegraphics
inversegraphics-master/shape_model.py
import numpy as np import pickle import chumpy as ch import ipdb from chumpy import depends_on, Ch import scipy.sparse as sp #%% Helper functions def longToPoints3D(pointsLong): nPointsLong = np.size(pointsLong) return np.reshape(pointsLong, [nPointsLong/3, 3]) def shapeParamsToVerts(shapeParams, teapotModel): landmarksLong = shapeParams.dot(teapotModel['ppcaW'].T) + teapotModel['ppcaB'] landmarks = longToPoints3D(landmarksLong) vertices = teapotModel['meshLinearTransform'].dot(landmarks) return vertices def chShapeParamsToVerts(landmarks, meshLinearTransform): vertices = ch.dot(meshLinearTransform,landmarks) return vertices class VerticesModel(Ch): terms = 'meshLinearTransform', 'W', 'b' dterms = 'chShapeParams' def init(self): self.jac = self.meshLinearTransform.dot(self.W.reshape([self.meshLinearTransform.shape[1], -1, len(self.chShapeParams)]).transpose((1,0,2))).reshape([-1,len(self.chShapeParams)]) def compute_r(self): landmarks = np.dot(self.chShapeParams.r,self.W.T) + self.b landmarks = landmarks.reshape([-1,3]) return np.dot(self.meshLinearTransform, landmarks) def compute_dr_wrt(self,wrt): if self.chShapeParams is wrt: # ipdb.set_trace() return self.jac return None def chShapeParamsToNormals(N, landmarks, linT): T = ch.dot(linT,landmarks) invT = [] nLandmarks = landmarks.r.shape[0] for i in range(nLandmarks): R = T[4*i:4*i+3,:3].T invR = ch.linalg.inv(R.T) invT = invT + [invR] invT = ch.vstack(invT) newNormals = ch.dot(N, invT) import opendr.geometry n = opendr.geometry.NormalizedNx3(newNormals) return newNormals def getT(targetPoints, linT): T = linT.dot(targetPoints) return T def shapeParamsToNormals(shapeParams, teapotModel): landmarksLong = shapeParams.dot(teapotModel['ppcaW'].T) + teapotModel['ppcaB'] landmarks = longToPoints3D(landmarksLong) T = getT(landmarks, teapotModel['linT']) nLandmarks = np.shape(landmarks)[0] invT = np.empty([3*nLandmarks, 3]) for i in range(nLandmarks): R = T[4*i:4*i+3,:3].T invR = np.linalg.inv(R) invT[3*i:3*(i+1),:] = invR newNormals = np.array(teapotModel['N'].dot(invT)) normalize_v3(newNormals) return newNormals def saveObj(vertices, faces, normals, filePath): with open(filePath, 'w') as f: f.write("# OBJ file\n") for v in vertices: f.write("v %.4f %.4f %.4f\n" % (v[0], v[1], v[2])) for n in normals: f.write("vn %.4f %.4f %.4f\n" % (n[0], n[1], n[2])) for p in faces: f.write("f") for i in p: f.write(" %d" % (i + 1)) f.write("\n") def loadObject(fileName): with open(fileName, 'rb') as inpt: return pickle.load(inpt) def normalize_v3(arr): lens = np.sqrt( arr[:,0]**2 + arr[:,1]**2 + arr[:,2]**2 ) arr[:,0] /= lens arr[:,1] /= lens arr[:,2] /= lens return arr def getNormals(vertices, faces): norm = np.zeros( vertices.shape, dtype=vertices.dtype ) tris = vertices[faces] n = np.cross( tris[::,1 ] - tris[::,0] , tris[::,2 ] - tris[::,0] ) normalize_v3(n) norm[ faces[:,0] ] += n norm[ faces[:,1] ] += n norm[ faces[:,2] ] += n normalize_v3(norm) return norm def chGetNormals(vertices, faces): import opendr.geometry return opendr.geometry.VertNormals(vertices, faces).reshape((-1,3))
3,538
28.491667
187
py
inversegraphics
inversegraphics-master/opendr_utils.py
__author__ = 'pol' from utils import * import opendr import chumpy as ch import geometry import bpy import mathutils import numpy as np from math import radians from opendr.camera import ProjectPoints from opendr.renderer import TexturedRenderer from opendr.lighting import SphericalHarmonics from opendr.lighting import LambertianPointLight from opendr.renderer import ResidualRendererOpenDR import ipdb import light_probes import scene_io_utils from blender_utils import * import imageio import chumpy as ch from chumpy import depends_on, Ch import scipy.sparse as sp class TheanoFunOnOpenDR(Ch): terms = 'opendr_input_gt' dterms = 'opendr_input' initialized = False def compileFunctions(self, theano_output, theano_input, dim_output, theano_input_gt, theano_output_gt): import theano import theano.tensor as T self.prediction_fn = theano.function([theano_input], theano_output) # self.J, updates = theano.scan(lambda i, y,x : T.grad(y[i], x), sequences=T.arange(y.shape[0]), non_sequences=[y,x]) # self.J, updates = theano.scan(lambda i, y,x : T.grad(y[i], x), sequences=T.arange(y.shape[0]), non_sequences=[y,x]) self.prediction_fn_gt = theano.function([theano_input_gt], theano_output_gt) x = theano_input gt_output = T.vector('gt_output') # self.error = T.sum(theano_output + gt_output) self.error = T.sum(T.pow(theano_output.ravel() - theano_output_gt.ravel(),2)) self.errorGrad = T.grad(self.error, x) from theano.compile.nanguardmode import NanGuardMode self.errorGrad_fun = theano.function([theano_input, theano_input_gt], self.errorGrad) # self.error_fun = theano.function([theano_input, theano_input_gt], self.error, mode=NanGuardMode(nan_is_error=True, inf_is_error=True, big_is_error=True)) self.error_fun = theano.function([theano_input, theano_input_gt], self.error) # self.grad = theano.function([x], self.J, updates=updates, mode='FAST_RUN') self.initialized = True def compute_r(self): if not self.initialized: self.compileFunctions() n_channels = 1 if len(self.opendr_input.r.shape) == 3: n_channels = self.opendr_input.r.shape[2] h = self.opendr_input.r.shape[1] w = self.opendr_input.r.shape[0] x = self.opendr_input.r.reshape([1,n_channels,h,w]).astype(np.float32) x_gt = self.opendr_input_gt.r.reshape([1,n_channels,h,w]).astype(np.float32) # out_gt = self.predict_input_gt().ravel().astype(np.float32) output = np.array(self.error_fun(x, x_gt)) return output.ravel() def predict_input(self): n_channels = 1 if len(self.opendr_input.r.shape) == 3: n_channels = self.opendr_input.r.shape[2] h = self.opendr_input.r.shape[1] w = self.opendr_input.r.shape[0] x = self.opendr_input.r.reshape([1,n_channels,h,w]).astype(np.float32) output = self.prediction_fn(x) return output.ravel() def predict_input_gt(self): n_channels = 1 if len(self.opendr_input.r.shape) == 3: n_channels = self.opendr_input.r.shape[2] h = self.opendr_input.r.shape[1] w = self.opendr_input.r.shape[0] x_gt = self.opendr_input_gt.r.reshape([1, n_channels, h, w]).astype(np.float32) output = self.prediction_fn_gt(x_gt) return output.ravel() def compute_dr_wrt(self,wrt): if self.opendr_input is wrt: if not self.initialized: self.compileFunctions() n_channels = 1 if len(self.opendr_input.r.shape) == 3: n_channels = self.opendr_input.r.shape[2] h = self.opendr_input.r.shape[1] w = self.opendr_input.r.shape[0] x = self.opendr_input.r.reshape([1, n_channels, h, w]).astype(np.float32) x_gt = self.opendr_input_gt.r.reshape([1, n_channels, h, w]).astype(np.float32) jac = np.array(self.errorGrad_fun(x,x_gt)).squeeze().reshape([1,self.opendr_input.r.size]) return sp.csr.csr_matrix(jac) return None def old_grads(self): import theano import theano.tensor as T self.grad_fns = [theano.function([self.theano_input], theano.gradient.grad(self.theano_output.flatten()[grad_i], self.theano_input)) for grad_i in range(self.dim_output)] x = self.opendr_input.r jac = [sp.lil_matrix(np.array(grad_fun(x[None,None, :,:].astype(np.float32))).ravel()) for grad_fun in self.grad_fns] return sp.vstack(jac).tocsr() class TheanoFunFiniteDiff(Ch): terms = 'opendr_input_gt' dterms = 'opendr_input' initialized = False def compileFunctions(self, theano_output, theano_input, dim_output, theano_input_gt, theano_output_gt): import theano import theano.tensor as T self.prediction_fn = theano.function([theano_input], theano_output) # self.J, updates = theano.scan(lambda i, y,x : T.grad(y[i], x), sequences=T.arange(y.shape[0]), non_sequences=[y,x]) # self.J, updates = theano.scan(lambda i, y,x : T.grad(y[i], x), sequences=T.arange(y.shape[0]), non_sequences=[y,x]) self.prediction_fn_gt = theano.function([theano_input_gt], theano_output_gt) # x = theano_input # gt_output = T.vector('gt_output') # # self.error = T.sum(theano_output + gt_output) # self.error = T.sum(T.pow(theano_output.ravel() - theano_output_gt.ravel(),2)) # # self.errorGrad = T.grad(self.error, x) # # from theano.compile.nanguardmode import NanGuardMode # self.errorGrad_fun = theano.function([theano_input, theano_input_gt], self.errorGrad) # # self.error_fun = theano.function([theano_input, theano_input_gt], self.error, mode=NanGuardMode(nan_is_error=True, inf_is_error=True, big_is_error=True)) # self.error_fun = theano.function([theano_input, theano_input_gt], self.error) # # self.grad = theano.function([x], self.J, updates=updates, mode='FAST_RUN') self.initialized = True def compute_r(self): if not self.initialized: self.compileFunctions() x = self.opendr_input.r.copy() x_gt = self.opendr_input_gt.r.copy() xs = skimage.transform.resize(x, [self.imSize, self.imSize])[None,None,:,:].astype(np.float32) x_gts = skimage.transform.resize(x_gt, [self.imSize, self.imSize])[None,None,:,:].astype(np.float32) emb = self.prediction_fn(xs) emb_gt = self.prediction_fn_gt(x_gts) diff = np.sum((emb - emb_gt)**2) return diff def predict_input(self): x = self.opendr_input.r[None,None,:,:].astype(np.float32) output = self.prediction_fn(x) return output.ravel() def predict_input_gt(self): x = self.opendr_input_gt.r[None,None,:,:].astype(np.float32) output = self.prediction_fn_gt(x) return output.ravel() def compute_dr_wrt(self,wrt): if self.opendr_input is wrt: if not self.initialized: self.compileFunctions() delta = 0.01 x = self.opendr_input.r[None,None,:,:].astype(np.float32) x_gt = self.opendr_input_gt.r[None,None,:,:].astype(np.float32) # jac = np.array(self.errorGrad_fun(x,x_gt)).squeeze().reshape([1,self.opendr_input.r.size]) #Finite differences: approxjacs = [] for idx, freevar in enumerate(vars): f0 = self.error_fun(x, x_gt) oldvar = freevar.r[:].copy() freevar[:] = freevar.r[:] + delta f1 = self.error_fun(x, x_gt) diff = (f1 - f0) / np.abs(delta) freevar[:] = oldvar.copy() approxjacs = approxjacs + [diff] approxjacs = np.concatenate(approxjacs) return sp.csr.csr_matrix(approxjacs) return None def old_grads(self): import theano import theano.tensor as T self.grad_fns = [theano.function([self.theano_input], theano.gradient.grad(self.theano_output.flatten()[grad_i], self.theano_input)) for grad_i in range(self.dim_output)] x = self.opendr_input.r jac = [sp.lil_matrix(np.array(grad_fun(x[None,None, :,:].astype(np.float32))).ravel()) for grad_fun in self.grad_fns] return sp.vstack(jac).tocsr() def recoverAmbientIntensities(hdritems, gtDataset, clampedCosCoeffs): hdrscoeffs = np.zeros([100, 9,3]) for hdrFile, hdrValues in hdritems: hdridx = hdrValues[0] hdrscoeffs[hdridx] = hdrValues[1] trainEnvMapCoeffs = hdrscoeffs[gtDataset['trainEnvMaps']] trainTotaloffsets = gtDataset['trainEnvMapPhiOffsets'] + gtDataset['trainObjAzsGT'] rotations = np.vstack([light_probes.sphericalHarmonicsZRotation(trainTotaloffset)[None,:] for trainTotaloffset in trainTotaloffsets]) rotationsRel = np.vstack([light_probes.sphericalHarmonicsZRotation(trainEnvMapPhiOffsets)[None,:] for trainEnvMapPhiOffsets in gtDataset['trainEnvMapPhiOffsets']]) envMapCoeffsRotated = np.vstack([np.dot(rotations[i], trainEnvMapCoeffs[i,[0,3,2,1,4,5,6,7,8]])[[0,3,2,1,4,5,6,7,8]][None,:] for i in range(len(rotations))]) envMapCoeffsRotatedRel = np.vstack([np.dot(rotationsRel[i], trainEnvMapCoeffs[i,[0,3,2,1,4,5,6,7,8]])[[0,3,2,1,4,5,6,7,8]][None,:] for i in range(len(rotations))]) trainAmbientIntensityGT = gtDataset['trainComponentsGT'][:,0]/((0.3*envMapCoeffsRotated[:,0,0] + 0.59*envMapCoeffsRotated[:,0,1] + 0.11*envMapCoeffsRotated[:,0,2])*clampedCosCoeffs[0]) trainAmbientIntensityGT1 = gtDataset['trainComponentsGT'][:,1]/((0.3*envMapCoeffsRotated[:,1,0] + 0.59*envMapCoeffsRotated[:,1,1] + 0.11*envMapCoeffsRotated[:,1,2])*clampedCosCoeffs[1]) if np.any(trainAmbientIntensityGT != trainAmbientIntensityGT1): print("Problem with recovery of intensities") return trainAmbientIntensityGT def SHSpherePlot(): #Visualize plots ignoreEnvMaps = np.loadtxt('data/bad_envmaps.txt') envMapDic = {} SHFilename = 'data/LightSHCoefficients.pickle' with open(SHFilename, 'rb') as pfile: envMapDic = pickle.load(pfile) hdritems = list(envMapDic.items())[0:10] pEnvMapsList = [] envMapsList = [] width = 600 height = 300 for hdrFile, hdrValues in hdritems: hdridx = hdrValues[0] if hdridx not in ignoreEnvMaps: if not os.path.exists('light_probes/envMap' + str(hdridx)): os.makedirs('light_probes/envMap' + str(hdridx)) envMapCoeffs = hdrValues[1] envMap = np.array(imageio.imread(hdrFile))[:,:,0:3] # normalize = envMap.mean()*envMap.shape[0]*envMap.shape[1] # envMap = 0.3*envMap[:,:,0] + 0.59*envMap[:,:,1] + 0.11*envMap[:,:,2] # if envMap.shape[0] != height or envMap.shape[1] != width: envMap = cv2.resize(src=envMap, dsize=(width,height)) print("Processing hdridx" + str(hdridx)) # envMapCoeffs = 0.3*envMapCoeffs[:,0] + 0.59*envMapCoeffs[:,1] + 0.11*envMapCoeffs[:,2] # envMapCoeffs *= normalize # pEnvMap = SHProjection(envMap, envMapCoeffs) envMapMean = envMap.mean() envMapCoeffs2 = light_probes.getEnvironmentMapCoefficients(envMap, 1, 0, 'equirectangular') # envMapCoeffs2 = 0.3*envMapCoeffs2[:,0] + 0.59*envMapCoeffs2[:,1] + 0.11*envMapCoeffs2[:,2] pEnvMap = SHProjection(envMap, envMapCoeffs2) tm = cv2.createTonemapDrago(gamma=2.2) tmEnvMap = tm.process(envMap) cv2.imwrite('light_probes/envMap' + str(hdridx) + '/texture.jpeg' , 255*tmEnvMap[:,:,[2,1,0]]) approxProjections = [] for coeffApprox in np.arange(9) + 1: approxProjection = np.sum(pEnvMap[:,:,:,0:coeffApprox], axis=3) approxProjectionPos = approxProjection.copy() approxProjectionPos[approxProjection<0] = 0 # tm = cv2.createTonemapDrago(gamma=2.2) # tmApproxProjection = tm.process(approxProjection) approxProjections = approxProjections + [approxProjection[None, :,:,:,None]] cv2.imwrite('light_probes/envMap' + str(hdridx) + '/approx' + str(coeffApprox-1) + '.jpeg', 255 * approxProjectionPos[:,:,[2,1,0]]) approxProjections = np.concatenate(approxProjections, axis=4) pEnvMapsList = pEnvMapsList + [approxProjections] envMapsList = envMapsList + [envMap[None,:,:,:]] pEnvMaps = np.concatenate(pEnvMapsList, axis=0) envMaps = np.concatenate(envMapsList, axis=0) #Total sum of squares envMapsMean = np.mean(envMaps, axis=0) pEnvMapsMean = np.mean(pEnvMaps, axis=0) tsq = np.sum((envMaps - envMapsMean)**2)/(envMapsMean.shape[0]*envMapsMean.shape[1]) ess = np.sum((pEnvMaps - envMapsMean[None,:,:,:,None])**2, axis=(0,1,2,3))/(envMapsMean.shape[0]*envMapsMean.shape[1]) uess = np.sum((pEnvMaps - envMaps[:,:,:,:,None])**2, axis=(0,1,2,3))/(envMapsMean.shape[0]*envMapsMean.shape[1]) explainedVar = ess/tsq tsq2 = uess + ess ipdb.set_trace() #Do something with standardized residuals. stdres = (pEnvMaps - pEnvMapsMean)/np.sqrt(np.sum((pEnvMaps - pEnvMapsMean)**2,axis=0)) return explainedVar def exportEnvMapSHImages(shCoeffsRGB, useBlender, scene, width, height, rendererGT): import glob for hdridx, hdrFile in enumerate(glob.glob("data/hdr/dataset/*")): envMapFilename = hdrFile envMapTexture = np.array(imageio.imread(envMapFilename))[:,:,0:3] phiOffset = 0 phiOffsets = [0, np.pi/2, np.pi, 3*np.pi/2] # phiOffsets = [0, np.pi/2, np.pi, 3*np.pi/2] if not os.path.exists('light_probes/envMap' + str(hdridx)): os.makedirs('light_probes/envMap' + str(hdridx)) cv2.imwrite('light_probes/envMap' + str(hdridx) + '/texture.png' , 255*envMapTexture[:,:,[2,1,0]]) for phiOffset in phiOffsets: envMapMean = envMapTexture.mean() envMapCoeffs = light_probes.getEnvironmentMapCoefficients(envMapTexture, envMapMean, -phiOffset, 'equirectangular') shCoeffsRGB[:] = envMapCoeffs if useBlender: updateEnviornmentMap(envMapFilename, scene) setEnviornmentMapStrength(0.3/envMapMean, scene) rotateEnviornmentMap(phiOffset, scene) scene.render.filepath = 'light_probes/envMap' + str(hdridx) + '/blender_' + str(np.int(180*phiOffset/np.pi)) + '.png' bpy.ops.render.render(write_still=True) cv2.imwrite('light_probes/envMap' + str(hdridx) + '/opendr_' + str(np.int(180*phiOffset/np.pi)) + '.png' , 255*rendererGT.r[:,:,[2,1,0]]) def exportEnvMapSHLightCoefficients(): import glob envMapDic = {} hdrs = glob.glob("data/hdr/dataset/*") hdrstorender = hdrs sphericalMap = False for hdridx, hdrFile in enumerate(hdrstorender): print("Processing env map" + str(hdridx)) envMapFilename = hdrFile envMapTexture = np.array(imageio.imread(envMapFilename))[:,:,0:3] phiOffset = 0 if not os.path.exists('light_probes/envMap' + str(hdridx)): os.makedirs('light_probes/envMap' + str(hdridx)) # cv2.imwrite('light_probes/envMap' + str(hdridx) + '/texture.png' , 255*envMapTexture[:,:,[2,1,0]]) if sphericalMap: envMapTexture, envMapMean = light_probes.processSphericalEnvironmentMap(envMapTexture) envMapCoeffs = light_probes.getEnvironmentMapCoefficients(envMapTexture, 1, 0, 'spherical') else: envMapGray = 0.3*envMapTexture[:,:,0] + 0.59*envMapTexture[:,:,1] + 0.11*envMapTexture[:,:,2] envMapGrayMean = np.mean(envMapGray, axis=(0,1)) envMapTexture = envMapTexture/envMapGrayMean # envMapTexture = 4*np.pi*envMapTexture/np.sum(envMapTexture, axis=(0,1)) envMapCoeffs = light_probes.getEnvironmentMapCoefficients(envMapTexture, 1, 0, 'equirectangular') pEnvMap = SHProjection(envMapTexture, envMapCoeffs) approxProjection = np.sum(pEnvMap, axis=3) # envMapCoeffs = light_probes.getEnvironmentMapCoefficients(envMapTexture, envMapMean, phiOffset, 'equirectangular') envMapDic[hdrFile] = [hdridx, envMapCoeffs] SHFilename = 'data/LightSHCoefficients.pickle' with open(SHFilename, 'wb') as pfile: pickle.dump(envMapDic, pfile) def transformObject(v, vn, chScale, chObjAz, chObjDisplacement, chObjRotation, targetPosition): scaleMat = geometry.Scale(x=chScale[0], y=chScale[1],z=chScale[2])[0:3,0:3] chRotAzMat = geometry.RotateZ(a=-chObjAz)[0:3,0:3] transformation = ch.dot(chRotAzMat, scaleMat) invTranspModel = ch.transpose(ch.linalg.inv(transformation)) objDisplacementMat = computeHemisphereTransformation(chObjRotation, 0, chObjDisplacement, np.array([0,0,0])) newPos = objDisplacementMat[0:3, 3] vtransf = [] vntransf = [] for mesh_i, mesh in enumerate(v): vtransf = vtransf + [ch.dot(v[mesh_i], transformation) + newPos + targetPosition] # ipdb.set_trace() # vtransf = vtransf + [v[mesh_i] + chPosition] vndot = ch.dot(vn[mesh_i], invTranspModel) vndot = vndot/ch.sqrt(ch.sum(vndot**2,1))[:,None] vntransf = vntransf + [vndot] return vtransf, vntransf, newPos def createRendererTarget(glMode, chAz, chEl, chDist, center, v, vc, f_list, vn, light_color, chComponent, chVColors, targetPosition, chDisplacement, width,height, uv, haveTextures_list, textures_list, frustum, win ): renderer = TexturedRenderer() renderer.set(glMode=glMode) vflat = [item for sublist in v for item in sublist] rangeMeshes = range(len(vflat)) vnflat = [item for sublist in vn for item in sublist] vcflat = [item for sublist in vc for item in sublist] # vcch = [np.ones_like(vcflat[mesh])*chVColors.reshape([1,3]) for mesh in rangeMeshes] vc_list = computeSphericalHarmonics(vnflat, vcflat, light_color, chComponent) if len(vflat)==1: vstack = vflat[0] else: vstack = ch.vstack(vflat) camera, modelRotation, _ = setupCamera(vstack, chAz, chEl, chDist, center + targetPosition + chDisplacement, width, height) setupTexturedRenderer(renderer, vstack, vflat, f_list, vc_list, vnflat, uv, haveTextures_list, textures_list, camera, frustum, win) return renderer def createNewRendererTarget(glMode, chAz, chEl, chDist, center, v, vc, f_list, vn, light_color, chComponent, chVColors, targetPosition, chDisplacement, width,height, uv, haveTextures_list, textures_list, frustum, win ): renderer = ResidualRendererOpenDR() renderer.set(glMode=glMode) vflat = [item for sublist in v for item in sublist] rangeMeshes = range(len(vflat)) vnflat = [item for sublist in vn for item in sublist] vcflat = [item for sublist in vc for item in sublist] # vcch = [np.ones_like(vcflat[mesh])*chVColors.reshape([1,3]) for mesh in rangeMeshes] vc_list = computeSphericalHarmonics(vnflat, vcflat, light_color, chComponent) if len(vflat)==1: vstack = vflat[0] else: vstack = ch.vstack(vflat) camera, modelRotation, _ = setupCamera(vstack, chAz, chEl, chDist, center + targetPosition + chDisplacement, width, height) setupTexturedRenderer(renderer, vstack, vflat, f_list, vc_list, vnflat, uv, haveTextures_list, textures_list, camera, frustum, win) return renderer def getCubeData(scale=(2,2,2), st=False, rgb=np.array([0.0, 0.0, 0.0])): dataCube, facesCube = create_cube(scale=(2,2,2), st=False, rgba=np.array([rgb[0], rgb[1], rgb[2], 1.0]), dtype='float32', type='triangles') verticesCube = ch.Ch(dataCube[:,0:3]) UVsCube = ch.Ch(np.zeros([verticesCube.shape[0],2])) facesCube = facesCube.reshape([-1,3]) import shape_model normalsCube = -shape_model.chGetNormals(verticesCube, facesCube) haveTexturesCube = [[False]] texturesListCube = [[None]] vColorsCube = ch.Ch(dataCube[:,3:6]) return verticesCube, facesCube, normalsCube, vColorsCube, texturesListCube, haveTexturesCube def createRendererGT(glMode, chAz, chEl, chDist, center, v, vc, f_list, vn, light_color, chComponent, chVColors, targetPosition, chDisplacement, width,height, uv, haveTextures_list, textures_list, frustum, win ): renderer = TexturedRenderer() renderer.set(glMode=glMode) vflat = [item for sublist in v for item in sublist] rangeMeshes = range(len(vflat)) # vch = [ch.array(vflat[mesh]) for mesh in rangeMeshes] vch = vflat if len(vch)==1: vstack = vch[0] else: vstack = ch.vstack(vch) camera, modelRotation, _ = setupCamera(vstack, chAz, chEl, chDist, center + targetPosition + chDisplacement, width, height) vnflat = [item for sublist in vn for item in sublist] # vnch = [ch.array(vnflat[mesh]) for mesh in rangeMeshes] # vnchnorm = [vnch[mesh]/ch.sqrt(vnch[mesh][:,0]**2 + vnch[mesh][:,1]**2 + vnch[mesh][:,2]**2).reshape([-1,1]) for mesh in rangeMeshes] vcflat = [item for sublist in vc for item in sublist] vcch = [vcflat[mesh] for mesh in rangeMeshes] # vc_list = computeGlobalAndPointLighting(vch, vnch, vcch, lightPosGT, chGlobalConstantGT, light_colorGT) setupTexturedRenderer(renderer, vstack, vch, f_list, vcch, vnflat, uv, haveTextures_list, textures_list, camera, frustum, win) return renderer #Old method def generateSceneImages(width, height, envMapFilename, envMapMean, phiOffset, chAzGT, chElGT, chDistGT, light_colorGT, chComponentGT, glMode): replaceableScenesFile = '../databaseFull/fields/scene_replaceables_backup.txt' sceneLines = [line.strip() for line in open(replaceableScenesFile)] for sceneIdx in np.arange(len(sceneLines)): sceneNumber, sceneFileName, instances, roomName, roomInstanceNum, targetIndices, targetPositions = scene_io_utils.getSceneInformation(sceneIdx, replaceableScenesFile) sceneDicFile = 'data/scene' + str(sceneNumber) + '.pickle' bpy.ops.wm.read_factory_settings() scene_io_utils.loadSceneBlendData(sceneIdx, replaceableScenesFile) scene = bpy.data.scenes['Main Scene'] bpy.context.screen.scene = scene scene_io_utils.setupScene(scene, roomInstanceNum, scene.world, scene.camera, width, height, 16, True, False) scene.update() scene.render.resolution_x = width #perhaps set resolution in code scene.render.resolution_y = height scene.render.tile_x = height/2 scene.render.tile_y = width scene.cycles.samples = 100 addEnvironmentMapWorld(envMapFilename, scene) setEnviornmentMapStrength(0.3/envMapMean, scene) rotateEnviornmentMap(phiOffset, scene) if not os.path.exists('scenes/' + str(sceneNumber)): os.makedirs('scenes/' + str(sceneNumber)) for targetIdx, targetIndex in enumerate(targetIndices): targetPosition = targetPositions[targetIdx] rendererGT.clear() del rendererGT v, f_list, vc, vn, uv, haveTextures_list, textures_list = scene_io_utils.loadSavedScene(sceneDicFile) # removeObjectData(targetIndex, v, f_list, vc, vn, uv, haveTextures_list, textures_list) # addObjectData(v, f_list, vc, vn, uv, haveTextures_list, textures_list, v_teapots[currentTeapotModel][0], f_list_teapots[currentTeapotModel][0], vc_teapots[currentTeapotModel][0], vn_teapots[currentTeapotModel][0], uv_teapots[currentTeapotModel][0], haveTextures_list_teapots[currentTeapotModel][0], textures_list_teapots[currentTeapotModel][0]) vflat = [item for sublist in v for item in sublist] rangeMeshes = range(len(vflat)) vch = [ch.array(vflat[mesh]) for mesh in rangeMeshes] # vch[0] = ch.dot(vch[0], scaleMatGT) + targetPosition if len(vch)==1: vstack = vch[0] else: vstack = ch.vstack(vch) cameraGT, modelRotationGT = setupCamera(vstack, chAzGT, chElGT, chDistGT, targetPosition, width, height) # cameraGT, modelRotationGT = setupCamera(vstack, chAzGT, chElGT, chDistGT, center + targetPosition, width, height) vnflat = [item for sublist in vn for item in sublist] vnch = [ch.array(vnflat[mesh]) for mesh in rangeMeshes] vnchnorm = [vnch[mesh]/ch.sqrt(vnch[mesh][:,0]**2 + vnch[mesh][:,1]**2 + vnch[mesh][:,2]**2).reshape([-1,1]) for mesh in rangeMeshes] vcflat = [item for sublist in vc for item in sublist] vcch = [ch.array(vcflat[mesh]) for mesh in rangeMeshes] vc_list = computeSphericalHarmonics(vnchnorm, vcch, light_colorGT, chComponentGT) rendererGT = TexturedRenderer() rendererGT.set(glMode=glMode) setupTexturedRenderer(rendererGT, vstack, vch, f_list, vc_list, vnchnorm, uv, haveTextures_list, textures_list, cameraGT, frustum, win) cv2.imwrite('scenes/' + str(sceneNumber) + '/opendr_' + str(targetIndex) + '.png' , 255*rendererGT.r[:,:,[2,1,0]]) placeCamera(scene.camera, -chAzGT[0].r*180/np.pi, chElGT[0].r*180/np.pi, chDistGT, targetPosition) scene.update() scene.render.filepath = 'scenes/' + str(sceneNumber) + '/blender_' + str(targetIndex) + '.png' bpy.ops.render.render(write_still=True) def getOcclusionFraction(renderer, id=0): vis_occluded = np.array(renderer.indices_image==id+1).copy().astype(np.bool) vis_im = np.array(renderer.image_mesh_bool([id])).copy().astype(np.bool) # return 1. - np.sum(vis_occluded)/np.sum(vis_im) return 0. #From http://www.ppsloan.org/publications/StupidSH36.pdf def chZonalHarmonics(a): zl0 = -ch.sqrt(ch.pi)*(-1.0 + ch.cos(a)) zl1 = 0.5*ch.sqrt(3.0*ch.pi)*ch.sin(a)**2 zl2 = -0.5*ch.sqrt(5.0*ch.pi)*ch.cos(a)*(-1.0 + ch.cos(a))*(ch.cos(a)+1.0) z = [zl0, zl1, zl2] return ch.concatenate(z) # http://cseweb.ucsd.edu/~ravir/papers/envmap/envmap.pdf chSpherical_harmonics = { (0, 0): lambda theta, phi: ch.Ch([0.282095]), (1, -1): lambda theta, phi: 0.488603 * ch.sin(theta) * ch.sin(phi), (1, 0): lambda theta, phi: 0.488603 * ch.cos(theta), (1, 1): lambda theta, phi: 0.488603 * ch.sin(theta) * ch.cos(phi), (2, -2): lambda theta, phi: 1.092548 * ch.sin(theta) * ch.cos(phi) * ch.sin(theta) * ch.sin(phi), (2, -1): lambda theta, phi: 1.092548 * ch.sin(theta) * ch.sin(phi) * ch.cos(theta), (2, 0): lambda theta, phi: 0.315392 * (3 * ch.cos(theta)**2 - 1), (2, 1): lambda theta, phi: 1.092548 * ch.sin(theta) * ch.cos(phi) * ch.cos(theta), (2, 2): lambda theta, phi: 0.546274 * (((ch.sin(theta) * ch.cos(phi)) ** 2) - ((ch.sin(theta) * ch.sin(phi)) ** 2)) } #From http://www.ppsloan.org/publications/StupidSH36.pdf def chZonalToSphericalHarmonics(z, theta, phi): sphCoeffs = [] for l in np.arange(len(z)): for m in np.arange(np.int(-(l*2+1)/2),np.int((l*2+1)/2) + 1): ylm_d = chSpherical_harmonics[(l,m)](theta,phi) sh = np.sqrt(4*np.pi/(2*l + 1))*z[l]*ylm_d sphCoeffs = sphCoeffs + [sh] #Correct order in band l=1. sphCoeffs[1],sphCoeffs[3] = sphCoeffs[3],sphCoeffs[1] chSphCoeffs = ch.concatenate(sphCoeffs) return chSphCoeffs #From http://www.ppsloan.org/publications/StupidSH36.pdf def clampedCosineCoefficients(): constants = [] for l in np.arange(3): for m in np.arange(np.int(-(l*2+1)/2),np.int((l*2+1)/2) + 1): normConstant = np.pi if l > 1 and l % 1 == 0: normConstant = 0 if l == 1: normConstant = 2*np.pi/3 if l > 1 and l % 2 == 0: normConstant = 2*np.pi*(((-1)**(l/2.-1.))/((l+2)*(l-1)))*(np.math.factorial(l)/((2**(l))*(np.math.factorial(l/2)**2))) # normConstant = 0.785398 constants = constants + [normConstant] return np.array(constants) def setupCamera(v, chAz, chEl, chDist, objCenter, width, height): chCamModelWorld = computeHemisphereTransformation(chAz, chEl, chDist, objCenter) chMVMat = ch.dot(chCamModelWorld, np.array(mathutils.Matrix.Rotation(radians(270), 4, 'X'))) chInvCam = ch.linalg.inv(chMVMat) modelRotation = chInvCam[0:3,0:3] chRod = opendr.geometry.Rodrigues(rt=modelRotation).reshape(3) chTranslation = chInvCam[0:3,3] translation, rotation = (chTranslation, chRod) camera = ProjectPoints(v=v, rt=rotation, t=translation, f= 1.12*ch.array([width,width]), c=ch.array([width,height])/2.0, k=ch.zeros(5)) camera.openglMat = np.array(mathutils.Matrix.Rotation(radians(180), 4, 'X')) return camera, modelRotation, chMVMat def computeHemisphereTransformation(chAz, chEl, chDist, objCenter): chDistMat = geometry.Translate(x=ch.Ch(0), y=-chDist, z=ch.Ch(0)) chToObjectTranslate = geometry.Translate(x=objCenter[0], y=objCenter[1], z=objCenter[2]) chRotAzMat = geometry.RotateZ(a=chAz) chRotElMat = geometry.RotateX(a=-chEl) chCamModelWorld = ch.dot(chToObjectTranslate, ch.dot(chRotAzMat, ch.dot(chRotElMat,chDistMat))) return chCamModelWorld def computeSphericalHarmonics(vn, vc, light_color, components, ignoreFirstMesh = True): # vnflat = [item for sublist in vn for item in sublist] # vcflat = [item for sublist in vc for item in sublist] A_list = [] rangeMeshes = range(len(vn)) for mesh in rangeMeshes: if ignoreFirstMesh and len(vn) > 1: if mesh > 0: A_list += [SphericalHarmonics(vn=vn[mesh],components=components,light_color=light_color)] else: A_list += [1.] else: A_list += [SphericalHarmonics(vn=vn[mesh],components=components,light_color=light_color)] vc_list = [A_list[mesh]*vc[mesh] for mesh in rangeMeshes] return vc_list def computeGlobalAndDirectionalLighting(vn, vc, chLightAzimuth, chLightElevation, chLightIntensity, chGlobalConstant): # Construct point light source rangeMeshes = range(len(vn)) vc_list = [] chRotAzMat = geometry.RotateZ(a=chLightAzimuth)[0:3,0:3] chRotElMat = geometry.RotateX(a=chLightElevation)[0:3,0:3] chLightVector = -ch.dot(chRotAzMat, ch.dot(chRotElMat, np.array([0,0,-1]))) for mesh in rangeMeshes: l1 = ch.maximum(ch.dot(vn[mesh], chLightVector).reshape((-1,1)), 0.) vcmesh = vc[mesh] * (chLightIntensity * l1 + chGlobalConstant) vc_list = vc_list + [vcmesh] return vc_list def computeGlobalAndPointLighting(v, vn, vc, light_pos, globalConstant, light_color): # Construct point light source rangeMeshes = range(len(vn)) vc_list = [] for mesh in rangeMeshes: l1 = LambertianPointLight( v=v[mesh], vn=vn[mesh], num_verts=len(v[mesh]), light_pos=light_pos, vc=vc[mesh], light_color=light_color) vcmesh = vc[mesh]*(l1 + globalConstant) vc_list = vc_list + [vcmesh] return vc_list def setupTexturedRenderer(renderer, vstack, vch, f_list, vc_list, vnch, uv, haveTextures_list, textures_list, camera, frustum, sharedWin=None): f = [] f_listflat = [item for sublist in f_list for item in sublist] lenMeshes = 0 for mesh_i, mesh in enumerate(f_listflat): polygonLen = 0 for polygons in mesh: f = f + [polygons + lenMeshes] polygonLen += len(polygons) lenMeshes += len(vch[mesh_i]) fstack = np.vstack(f) # vnflat = [item for sublist in vnch for item in sublist] if len(vnch)==1: vnstack = vnch[0] else: vnstack = ch.vstack(vnch) # vc_listflat = [item for sublist in vc_list for item in sublist] if len(vc_list)==1: vcstack = vc_list[0] else: vcstack = ch.vstack(vc_list) uvflat = [item for sublist in uv for item in sublist] ftstack = np.vstack(uvflat) texturesch = [] textures_listflat = [item for sublist in textures_list for item in sublist] for texture_list in textures_listflat: if texture_list is not None: for texture in texture_list: if texture is not None: texturesch = texturesch + [ch.array(texture)] if len(texturesch) == 0: texture_stack = ch.Ch([]) elif len(texturesch) == 1: texture_stack = texturesch[0].ravel() else: texture_stack = ch.concatenate([tex.ravel() for tex in texturesch]) haveTextures_listflat = [item for sublist in haveTextures_list for item in sublist] renderer.set(camera=camera, frustum=frustum, v=vstack, f=fstack, vn=vnstack, vc=vcstack, ft=ftstack, texture_stack=texture_stack, v_list=vch, f_list=f_listflat, vc_list=vc_list, ft_list=uvflat, textures_list=textures_listflat, haveUVs_list=haveTextures_listflat, bgcolor=ch.ones(3), overdraw=True) renderer.msaa = True renderer.sharedWin = sharedWin # renderer.clear() # renderer.useShaderErrors = True # renderer.initGL() # renderer.initGLTexture() def addObjectData(v, f_list, vc, vn, uv, haveTextures_list, textures_list, vmod, fmod_list, vcmod, vnmod, uvmod, haveTexturesmod_list, texturesmod_list): v.insert(0,vmod) f_list.insert(0,fmod_list) vc.insert(0,vcmod) vn.insert(0,vnmod) uv.insert(0,uvmod) haveTextures_list.insert(0,haveTexturesmod_list) textures_list.insert(0,texturesmod_list) def addObjectDataLast(v, f_list, vc, vn, uv, haveTextures_list, textures_list, vmod, fmod_list, vcmod, vnmod, uvmod, haveTexturesmod_list, texturesmod_list): v.insert(-1,vmod) f_list.insert(-1,fmod_list) vc.insert(-1,vcmod) vn.insert(-1,vnmod) uv.insert(-1,uvmod) haveTextures_list.insert(-1,haveTexturesmod_list) textures_list.insert(-1,texturesmod_list) def removeObjectData(objIdx, v, f_list, vc, vn, uv, haveTextures_list, textures_list): del v[objIdx] del f_list[objIdx] del vc[objIdx] del vn[objIdx] del uv[objIdx] del haveTextures_list[objIdx] del textures_list[objIdx] def transformObject2(v, vn, chScale, chObjAz, chPosition): if chScale.size == 1: scaleMat = geometry.Scale(x=chScale[0], y=chScale[0],z=chScale[0])[0:3,0:3] elif chScale.size == 2: scaleMat = geometry.Scale(x=chScale[0], y=chScale[0], z=chScale[1])[0:3, 0:3] else: scaleMat = geometry.Scale(x=chScale[0], y=chScale[1], z=chScale[2])[0:3, 0:3] chRotAzMat = geometry.RotateZ(a=chObjAz)[0:3,0:3] chRotAzMatX = geometry.RotateX(a=0)[0:3,0:3] # transformation = scaleMat transformation = ch.dot(ch.dot(chRotAzMat,chRotAzMatX), scaleMat) invTranspModel = ch.transpose(ch.linalg.inv(transformation)) vtransf = [] vntransf = [] for mesh_i, mesh in enumerate(v): vtransf = vtransf + [ch.dot(v[mesh_i], transformation) + chPosition] # ipdb.set_trace() # vtransf = vtransf + [v[mesh_i] + chPosition] vndot = ch.dot(vn[mesh_i], invTranspModel) vndot = vndot/ch.sqrt(ch.sum(vndot**2,1))[:,None] vntransf = vntransf + [vndot] return vtransf, vntransf def getCubeData(scale=(2,2,2), st=False, rgb=np.array([1.0, 1.0, 1.0])): dataCube, facesCube = create_cube(scale=(1,1,1), st=False, rgba=np.array([rgb[0], rgb[1], rgb[2], 1.0]), dtype='float32', type='triangles') verticesCube = ch.Ch(dataCube[:,0:3]) UVsCube = ch.Ch(np.zeros([verticesCube.shape[0],2])) facesCube = facesCube.reshape([-1,3]) normalsCube = geometry.chGetNormals(verticesCube, facesCube) haveTexturesCube = [[False]] texturesListCube = [[None]] vColorsCube = ch.Ch(dataCube[:,3:6]) return verticesCube, facesCube, normalsCube, vColorsCube, texturesListCube, haveTexturesCube
36,041
42.11244
359
py
inversegraphics
inversegraphics-master/image_processing.py
__author__ = 'pol' from skimage.feature import hog from skimage import data, color, exposure import numpy as np import ipdb import skimage.color from numpy.core.umath_tests import matrix_multiply # def xyz2labCh(xyz, illuminant="D65", observer="2"): # """XYZ to CIE-LAB color space conversion. # Parameters # ---------- # xyz : array_like # The image in XYZ format, in a 3- or 4-D array of shape # ``(.., ..,[ ..,] 3)``. # illuminant : {"A", "D50", "D55", "D65", "D75", "E"}, optional # The name of the illuminant (the function is NOT case sensitive). # observer : {"2", "10"}, optional # The aperture angle of the observer. # Returns # ------- # out : ndarray # The image in CIE-LAB format, in a 3- or 4-D array of shape # ``(.., ..,[ ..,] 3)``. # Raises # ------ # ValueError # If `xyz` is not a 3-D array of shape ``(.., ..,[ ..,] 3)``. # ValueError # If either the illuminant or the observer angle is unsupported or # unknown. # Notes # ----- # By default Observer= 2A, Illuminant= D65. CIE XYZ tristimulus values # x_ref=95.047, y_ref=100., z_ref=108.883. See function `get_xyz_coords` for # a list of supported illuminants. # References # ---------- # .. [1] http://www.easyrgb.com/index.php?X=MATH&H=07#text7 # .. [2] http://en.wikipedia.org/wiki/Lab_color_space # Examples # -------- # # """ # arr = xyz # # xyz_ref_white = skimage.color.get_xyz_coords(illuminant, observer) # # # scale by CIE XYZ tristimulus values of the reference white point # arr = arr / xyz_ref_white # # # Nonlinear distortion and linear transformation # mask = arr > 0.008856 # notmask = arr <= 0.008856 # arr[mask] = ch.power(arr[mask], 1. / 3.) # arr[notmask] = 7.787 * arr[notmask] + 16. / 116. # # x, y, z = arr[..., 0], arr[..., 1], arr[..., 2] # # # Vector scaling # L = (116. * y) - 16. # a = 500.0 * (x - y) # b = 200.0 * (y - z) # # return ch.concatenate([x[..., np.newaxis] for x in [L, a, b]], axis=-1) # # def rgb2xyz(rgb): # """RGB to XYZ color space conversion. # Parameters # ---------- # rgb : array_like # The image in RGB format, in a 3- or 4-D array of shape # ``(.., ..,[ ..,] 3)``. # Returns # ------- # out : ndarray # The image in XYZ format, in a 3- or 4-D array of shape # ``(.., ..,[ ..,] 3)``. # Raises # ------ # ValueError # If `rgb` is not a 3- or 4-D array of shape ``(.., ..,[ ..,] 3)``. # Notes # ----- # The CIE XYZ color space is derived from the CIE RGB color space. Note # however that this function converts from sRGB. # References # ---------- # .. [1] http://en.wikipedia.org/wiki/CIE_1931_color_space # Examples # -------- # # """ # # Follow the algorithm from http://www.easyrgb.com/index.php # # except we don't multiply/divide by 100 in the conversion # arr = rgb # mask = arr > 0.04045 # notmask = arr <= 0.04045 # arr[mask] = ch.power((arr[mask] + 0.055) / 1.055, 2.4) # arr[notmask] /= 12.92 # return _convert(xyz_from_rgb, arr) def scaleInvariantMSECoeff(x_pred, x_target): #Rows: test samples #Cols: target variables scales = (matrix_multiply(x_pred[:,None,:],x_target[:,:,None])/matrix_multiply(x_pred[:,None,:],x_pred[:,:,None])).ravel() return scales def scaleInvariantColourDifference(rgbGT, rgbPred): scaleRGB = scaleInvariantMSECoeff(rgbPred, rgbGT) dist = np.sqrt(np.sum((rgbGT - scaleRGB[:,None]*rgbPred)**2, axis=1)) return dist def cColourDifference(rgb1, rgb2): lab1 = cv2.cvtColor(np.uint8(rgb1.reshape([-1,1,3])*255), cv2.COLOR_RGB2LAB)/255 lab2 = cv2.cvtColor(np.uint8(rgb2.reshape([-1,1,3])*255), cv2.COLOR_RGB2LAB)/255 dist = np.sqrt(np.sum((lab1[:,0,1:] - lab2[:,0,1:])**2, axis=1)) return dist def eColourDifference(rgb1, rgb2): lab1 = cv2.cvtColor(np.uint8(rgb1.reshape([-1,1,3])*255), cv2.COLOR_RGB2LAB)/255 lab2 = cv2.cvtColor(np.uint8(rgb2.reshape([-1,1,3])*255), cv2.COLOR_RGB2LAB)/255 dist = np.sqrt(np.sum((lab1[:,0] - lab2[:,0])**2, axis=1)) return dist # conf.cellSize = cellSize; # conf.numOrientations = 9; def computeHoG(image): image = color.rgb2gray(image) hog_descr, hog_image = hog(image, orientations=9, pixels_per_cell=(8, 8),cells_per_block=(1, 1), visualise=True) # fig, ( ax1, ax2) = plt.subplots(1, 2, figsize=(8, 4)) # # ax1.axis('off') # ax1.imshow(image, cmap=plt.cm.gray) # ax1.set_title('Input image') # # Rescale histogram for better display # hog_image_rescaled = exposure.rescale_intensity(hog_image, in_range=(0, 0.02)) # # ax2.axis('off') # ax2.imshow(hog_image_rescaled, cmap=plt.cm.gray) # ax2.set_title('Histogram of Oriented Gradients') # plt.show() return hog_descr def computeHoGFeatures(images): hogs = [] for image in images: features = computeHoG(image) hogs = hogs + [features[None,:] ] hogfeats = np.vstack(hogs) return hogfeats def computeIllumFeatures(images, numFreq): illum = [] win = 40 for image in images: features = featuresIlluminationDirection(image, win)[:numFreq,:].ravel() illum = illum + [features[None,:]] illumfeats = np.vstack(illum) return illumfeats def featuresIlluminationDirection(image,win): image = color.rgb2gray(image) coeffs = np.fft.fft2(image[image.shape[0]/2-win:image.shape[0]/2+win,image.shape[1]/2-win:image.shape[1]/2+win]) magnitudes = np.sqrt(coeffs.real**2 + coeffs.imag**2) phases = np.angle(coeffs) return np.hstack([magnitudes.ravel()[:,None], phases.ravel()[:,None]]) from chumpy import depends_on, Ch import chumpy as ch from math import radians import cv2 import scipy import matplotlib.pyplot as plt def dr_wrt_convolution(x, filter): print("Computing convolution gradients") widthRolls = x.shape[1] heightRolls = x.shape[0] tmpShape = [x.shape[0]+filter.shape[0], x.shape[1]+filter.shape[1]] template = np.zeros(tmpShape) template[0:filter.shape[0], 0:filter.shape[1]] = filter jacs = [] for i in range(heightRolls): for j in range(widthRolls): templateRolled = np.roll(template, shift=i, axis=0) templateRolled = np.roll(templateRolled, shift=j, axis=1) templateGrad = templateRolled[tmpShape[0] - x.shape[0] - np.int(filter.shape[0]/2): tmpShape[0] - np.int(filter.shape[0]/2), tmpShape[1] - x.shape[1] - np.int(filter.shape[1]/2): tmpShape[1] - np.int(filter.shape[1]/2)] jacs = jacs + [scipy.sparse.coo_matrix(templateGrad.ravel())] return scipy.sparse.vstack(jacs).tocsc() class convolve2D(Ch): terms = 'filter' dterms = 'x' def compute_r(self): # convolved = scipy.signal.convolve2d(self.x, self.filter, mode='same') convolved = scipy.ndimage.convolve(self.x, self.filter, mode='reflect') # return convolved[np.int((convolved.shape[0]-self.x.shape[0])/2):np.int((convolved.shape[0]-self.x.shape[0])/2) + self.x.shape[1], np.int((convolved.shape[1]-self.x.shape[1])/2):np.int((convolved.shape[1]-self.x.shape[1])/2) + self.x.shape[1]] return convolved def compute_dr_wrt(self, wrt): if wrt is self.x: return self.convolve2DDr else: return None class HogImage(Ch): terms = 'numOrient', 'cwidth', 'cheight' dterms = 'image', 'hog' def compute_r(self): from skimage import draw sy,sx, _ = self.image.shape radius = min(self.cwidth, self.cheight) // 2 - 1 orientations_arr = np.arange(self.numOrient) dx_arr = radius * np.cos(orientations_arr / self.numOrient * np.pi) dy_arr = radius * np.sin(orientations_arr / self.numOrient * np.pi) cr2 = self.cheight + self.cheight cc2 = self.cwidth + self.cwidth hog_image = np.zeros((sy, sx), dtype=float) n_cellsx = int(np.floor(sx // self.cwidth)) # number of cells in x n_cellsy = int(np.floor(sy // self.cheight)) # number of cells in y for x in range(n_cellsx): for y in range(n_cellsy): for o, dx, dy in zip(orientations_arr, dx_arr, dy_arr): centre = tuple([y * cr2 // 2, x * cc2 // 2]) rr, cc = draw.line(int(centre[0] + dy), int(centre[1] + dx), int(centre[0] - dy), int(centre[1] - dx)) hog_image[rr, cc] += self.hog[y, x, o] return hog_image def compute_dr_wrt(self, wrt): return None import skimage def diffHog(image, drconv=None, numOrient = 9, cwidth=8, cheight=8): imagegray = 0.3*image[:,:,0] + 0.59*image[:,:,1] + 0.11*image[:,:,2] sy,sx = imagegray.shape # gx = ch.empty(imagegray.shape, dtype=np.double) gx = imagegray[:, 2:] - imagegray[:, :-2] gx = ch.hstack([np.zeros([sy,1]), gx, np.zeros([sy,1])]) gy = imagegray[2:, :] - imagegray[:-2, :] # gy = imagegray[:, 2:] - imagegray[:, :-2] gy = ch.vstack([np.zeros([1,sx]), gy, np.zeros([1,sx])]) gx += 1e-5 # gy = imagegray[:-2,1:-1] - imagegray[2:,1:-1] + 0.00001 # gx = imagegray[1:-1,:-2] - imagegray[1:-1, 2:] + 0.00001 distFilter = np.ones([2*cheight,2*cwidth], dtype=np.uint8) distFilter[np.int(2*cheight/2), np.int(2*cwidth/2)] = 0 distFilter = (cv2.distanceTransform(distFilter, cv2.DIST_L2, 3)- np.max(cv2.distanceTransform(distFilter, cv2.DIST_L2, 3)))/(-np.max(cv2.distanceTransform(distFilter, cv2.DIST_L2, 3))) magn = ch.sqrt(gy**2 + gx**2)*180/np.sqrt(2) angles = ch.arctan(gy/gx)*180/np.pi + 90 # meanOrient = np.linspace(0, 180, numOrient) orientations_arr = np.arange(numOrient) meanOrient = orientations_arr / numOrient * 180 fb_resttmp = 1 - ch.abs(ch.expand_dims(angles[:,:],2) - meanOrient[1:].reshape([1,1,numOrient-1]))*numOrient/180 zeros_rest = np.zeros([sy,sx, numOrient-1, 1]) fb_rest = ch.max(ch.concatenate([fb_resttmp[:,:,:,None], zeros_rest],axis=3), axis=3) chMinOrient0 = ch.min(ch.concatenate([ch.abs(ch.expand_dims(angles[:,:],2) - meanOrient[0].reshape([1,1,1]))[:,:,:,None], ch.abs(180 - ch.expand_dims(angles[:,:],2) - meanOrient[0].reshape([1,1,1]))[:,:,:,None]], axis=3), axis=3) zeros_fb0 = np.zeros([sy,sx, 1]) fb0_tmp = ch.concatenate([1 - chMinOrient0[:,:]*numOrient/180, zeros_fb0],axis=2) fb_0 = ch.max(fb0_tmp,axis=2) fb = ch.concatenate([fb_0[:,:,None], fb_rest],axis=2) # fb[:,:,0] = ch.max(1 - ch.abs(ch.expand_dims(angles,2) - meanOrient.reshape([1,1,numOrient]))*numOrient/180,0) # fb = 1./(1. + ch.exp(1 - ch.abs(ch.expand_dims(angles,2) - meanOrient.reshape([1,1,numOrient]))*numOrient/180)) Fb = ch.expand_dims(magn,2)*fb if drconv is None: drconv = dr_wrt_convolution(Fb[:,:,0], distFilter) Fs_list = [convolve2D(x=Fb[:,:,Fbi], filter=distFilter, convolve2DDr=drconv).reshape([Fb.shape[0], Fb.shape[1],1]) for Fbi in range(numOrient)] # Fs_list = [scipy.signal.convolve2d(Fb[:,:,Fbi], distFilter).reshape([Fb.shape[0], Fb.shape[1],1]) for Fbi in range(numOrient)] Fs = ch.concatenate(Fs_list, axis=2) # cellCols = np.arange(start=cwidth/2, stop=Fs.shape[1]-cwidth/2 , step=cwidth) # cellRows = np.arange(start=cheight/2, stop=Fs.shape[0]-cheight/2 , step=cheight) Fcells = Fs[0:Fs.shape[0] :cheight,0:Fs.shape[1] :cwidth,:] epsilon = 1e-5 v = Fcells/ch.sqrt(ch.sum(Fcells**2) + epsilon) # v = Fcells # hog, hogim = skimage.feature.hog(imagegray, orientations=numOrient, pixels_per_cell=(cheight, cwidth), visualise=True) hog_image = HogImage(image=image, hog=Fcells, numOrient=numOrient, cwidth=cwidth, cheight=cheight) # plt.imshow(hog_image) # plt.figure() # plt.imshow(hogim) # ipdb.set_trace() return v, hog_image, drconv import zernike def zernikeProjection(images, numCoeffs, win=50): croppedImages = images[:,images.shape[1]/2-win:images.shape[1]/2+win,images.shape[2]/2-win:images.shape[2]/2+win, :] zpolys = zernikePolynomials(imageSize=(croppedImages.shape[1],croppedImages.shape[2]), numCoeffs=numCoeffs) coeffs = np.sum(np.sum(croppedImages[:,:,:,:,None]*zpolys.reshape([1,zpolys.shape[0], zpolys.shape[1], 1, -1]), axis=2), axis=1) return coeffs def zernikeProjectionGray(images, numCoeffs, win=50): if images.shape[3] == 3: images = 0.3*images[:,:,:,0] + 0.59*images[:,:,:,1] + 0.11*images[:,:,:,2] croppedImages = images[:,images.shape[1]/2-win:images.shape[1]/2+win,images.shape[2]/2-win:images.shape[2]/2+win] zpolys = zernikePolynomials(imageSize=(croppedImages.shape[1],croppedImages.shape[2]), numCoeffs=numCoeffs) coeffs = np.sum(np.sum(croppedImages[:,:,:,None]*zpolys.reshape([1,zpolys.shape[0], zpolys.shape[1], -1]), axis=2), axis=1) return coeffs def chZernikeProjection(image, numCoeffs=20, win=50): croppedImage = image[image.shape[0]/2-win:image.shape[0]/2+win,image.shape[1]/2-win:image.shape[1]/2+win, :] zpolys = zernikePolynomials(imageSize=(croppedImage.shape[0],croppedImage.shape[1]), numCoeffs=numCoeffs) imageProjections = croppedImage[:,:,:,None]*zpolys.reshape([zpolys.shape[0], zpolys.shape[1], 1, -1]) coeffs = ch.sum(ch.sum(imageProjections, axis=0), axis=0) return coeffs, imageProjections def zernikePolynomials(imageSize=(100,100), numCoeffs=20): sy,sx = imageSize # distFilter = np.ones([sy,sx], dtype=np.uint8) # distFilter[np.int(sy/2), np.int(sx/2)] = 0 # distFilter = cv2.distanceTransform(distFilter, cv2.DIST_L2, 3) # distFilter /= np.max(distFilter) # np.arange() ones = np.ones([sy,sx], dtype=np.bool) imgind = np.where(ones) dy = imgind[0] - int(sy/2) dx = imgind[1] - int(sx/2) pixaz = np.arctan2(dy,dx) pixrad = np.sqrt(dy**2 + dx**2) imaz = np.zeros([sy,sx]) imrad = np.zeros([sy,sx]) imaz[imgind] = pixaz imrad[imgind] = pixrad outcircle = imrad>=sy/2 imrad[outcircle] = 0 imaz[outcircle] = 0 imrad/=np.max(imrad) zpolys = [zernike.zernikel(j, imrad, imaz)[:,:,None] for j in range(numCoeffs)] zpolys = np.concatenate(zpolys, axis=2) return zpolys
14,461
35.428212
252
py
inversegraphics
inversegraphics-master/generative_models.py
import cv2 import numpy as np import matplotlib.pyplot as plt import ipdb import scipy import chumpy as ch from chumpy.ch import MatVecMult, Ch, depends_on def scoreImage(img, template, method, methodParams): score = 0 if method == 'chamferModelToData': sqDists = chamferDistanceModelToData(img, template, methodParams['minThresImage'], methodParams['maxThresImage'], methodParams['minThresTemplate'],methodParams['maxThresTemplate']) score = np.sum(sqDists) elif method == 'robustChamferModelToData': sqDists = np.sum(chamferDistanceModelToData(img, template, methodParams['minThresImage'], methodParams['maxThresImage'], methodParams['minThresTemplate'],methodParams['maxThresTemplate'])) score = robustDistance(sqDists, methodParams['scale']) elif method == 'chamferDataToModel': sqDists = chamferDistanceDataToModel(img, template, methodParams['minThresImage'], methodParams['maxThresImage'], methodParams['minThresTemplate'],methodParams['maxThresTemplate']) score = np.sum(sqDists) elif method == 'robustChamferDataToModel': sqDists = np.sum(chamferDistanceDataToModel(img, template, methodParams['minThresImage'], methodParams['maxThresImage'], methodParams['minThresTemplate'],methodParams['maxThresTemplate'])) score = robustDistance(sqDists, methodParams['scale']) elif method == 'sqDistImages': sqDists = sqDistImages(img, template) score = np.sum(sqDists) / template.size elif method == 'ignoreSqDistImages': sqDists = sqDistImages(img, template) score = np.sum(sqDists * (template > 0)) / np.sum(template > 0) elif method == 'robustSqDistImages': sqDists = sqDistImages(img, template) score = robustDistance(sqDists, methodParams['scale']) elif method == 'negLogLikelihoodRobust': score = -modelLogLikelihoodRobust(img, template, methodParams['testMask'], methodParams['backgroundModel'], methodParams['layerPrior'], methodParams['variances']) elif method == 'negLogLikelihood': score = -modelLogLikelihood(img, template, methodParams['testMask'], methodParams['backgroundModel'], methodParams['variances']) return score def chamferDistanceModelToData(img, template, minThresImage, maxThresImage, minThresTemplate, maxThresTemplate): imgEdges = cv2.Canny(np.uint8(img*255), minThresImage,maxThresImage) tempEdges = cv2.Canny(np.uint8(template*255), minThresTemplate, maxThresTemplate) bwEdges1 = cv2.distanceTransform(~imgEdges, cv2.DIST_L2, 5) score = np.sum(np.multiply(tempEdges/255, bwEdges1))/np.sum(tempEdges/255.0) return score def chamferDistanceDataToModel(img, template, minThresImage, maxThresImage, minThresTemplate, maxThresTemplate): imgEdges = cv2.Canny(np.uint8(img*255), minThresImage,maxThresImage) tempEdges = cv2.Canny(np.uint8(template*255), minThresTemplate, maxThresTemplate) bwEdges1 = cv2.distanceTransform(~tempEdges, cv2.DIST_L2, 5) score = np.multiply(imgEdges/255.0, bwEdges1)/np.sum(imgEdges/255.0) return score def sqDistImages(img, template): sqResiduals = np.square(img - template) return sqResiduals def computeVariances(sqResiduals): return np.sum(sqResiduals, axis=3)/sqResiduals.shape[-1] def pixelLayerPriors(masks): return np.sum(masks, axis=2) / masks.shape[-1] def globalLayerPrior(masks): return np.sum(masks) / masks.size def modelLogLikelihoodRobust(image, template, testMask, backgroundModel, layerPriors, variances): likelihood = pixelLikelihoodRobust(image, template, testMask, backgroundModel, layerPriors, variances) liksum = np.sum(np.log(likelihood)) return liksum def modelLogLikelihoodRobustCh(image, template, testMask, backgroundModel, layerPriors, variances): likelihood = pixelLikelihoodRobustCh(image, template, testMask, backgroundModel, layerPriors, variances) liksum = ch.sum(ch.log(likelihood)) return liksum def modelLogLikelihoodRobustRegionCh(image, template, testMask, backgroundModel, layerPriors, variances): likelihood = pixelLikelihoodRobustRegionCh(image, template, testMask, backgroundModel, layerPriors, variances) liksum = ch.sum(ch.log(likelihood)) return liksum def modelLogLikelihood(image, template, testMask, backgroundModel, variances): likelihood = pixelLikelihood(image, template, testMask, backgroundModel, variances) liksum = np.sum(np.log(likelihood)) def modelLogLikelihoodCh(image, template, testMask, backgroundModel, variances): logLikelihood = logPixelLikelihoodCh(image, template, testMask, backgroundModel, variances) return ch.sum(logLikelihood) def pixelLikelihoodRobust(image, template, testMask, backgroundModel, layerPrior, variances): sigma = np.sqrt(variances) mask = testMask if backgroundModel == 'FULL': mask = np.ones(image.shape[0:2]) # mask = np.repeat(mask[..., np.newaxis], 3, 2) repPriors = np.tile(layerPrior, image.shape[0:2]) # sum = np.sum(np.log(layerPrior * scipy.stats.norm.pdf(image, location = template, scale=np.sqrt(variances) ) + (1 - repPriors))) # uniformProbs = np.ones(image.shape) foregroundProbs = np.prod(1/(sigma * np.sqrt(2 * np.pi)) * np.exp( - (image - template)**2 / (2 * variances)) * layerPrior, axis=2) + (1 - repPriors) return foregroundProbs * mask + (1-mask) def pixelLikelihoodRobustSQErrorCh(sqeRenderer, testMask, backgroundModel, layerPrior, variances): sigma = ch.sqrt(variances) mask = testMask if backgroundModel == 'FULL': mask = np.ones(sqeRenderer.r.shape[0:2]) # mask = np.repeat(mask[..., np.newaxis], 3, 2) repPriors = ch.tile(layerPrior, sqeRenderer.r.shape[0:2]) # sum = np.sum(np.log(layerPrior * scipy.stats.norm.pdf(image, location = template, scale=np.sqrt(variances) ) + (1 - repPriors))) # uniformProbs = np.ones(image.shape) probs = ch.exp( - (sqeRenderer) / (2 * variances)) * (1./(sigma * np.sqrt(2 * np.pi))) foregroundProbs = (probs[:,:,0] * probs[:,:,1] * probs[:,:,2]) * layerPrior + (1 - repPriors) return foregroundProbs * mask + (1-mask) def pixelLikelihoodRobustCh(image, template, testMask, backgroundModel, layerPrior, variances): sigma = ch.sqrt(variances) mask = testMask if backgroundModel == 'FULL': mask = np.ones(image.shape[0:2]) # mask = np.repeat(mask[..., np.newaxis], 3, 2) repPriors = ch.tile(layerPrior, image.shape[0:2]) # sum = np.sum(np.log(layerPrior * scipy.stats.norm.pdf(image, location = template, scale=np.sqrt(variances) ) + (1 - repPriors))) # uniformProbs = np.ones(image.shape) probs = ch.exp( - (image - template)**2 / (2 * variances)) * (1./(sigma * np.sqrt(2 * np.pi))) foregroundProbs = (probs[:,:,0] * probs[:,:,1] * probs[:,:,2]) * layerPrior + (1 - repPriors) return foregroundProbs * mask + (1-mask) def pixelLikelihoodRobustRegionCh(image, template, testMask, backgroundModel, layerPrior, variances): sigma = ch.sqrt(variances) mask = testMask if backgroundModel == 'FULL': mask = np.ones(image.shape[0:2]) # mask = np.repeat(mask[..., np.newaxis], 3, 2) repPriors = ch.tile(layerPrior, image.shape[0:2]) # sum = np.sum(np.log(layerPrior * scipy.stats.norm.pdf(image, location = template, scale=np.sqrt(variances) ) + (1 - repPriors))) # uniformProbs = np.ones(image.shape) imshape = image.shape from opendr.filters import filter_for from opendr.filters import GaussianKernel2D blur_mtx = filter_for(imshape[0], imshape[1], imshape[2] if len(imshape)>2 else 1, kernel = GaussianKernel2D(3, 1)) blurred_image = MatVecMult(blur_mtx, image).reshape(imshape) blurred_template = MatVecMult(blur_mtx, template).reshape(imshape) probs = ch.exp( - (blurred_image - template)**2 / (2 * variances)) * (1./(sigma * np.sqrt(2 * np.pi))) foregroundProbs = (probs[:,:,0] * probs[:,:,1] * probs[:,:,2]) * layerPrior + (1 - repPriors) return foregroundProbs * mask + (1-mask) import chumpy as ch from chumpy import depends_on, Ch class EdgeFilter(Ch): dterms = ['renderer', 'rendererGT'] def compute_r(self): return self.blurredDiff() def compute_dr_wrt(self, wrt): if wrt is self.renderer: return self.blurredDiff().dr_wrt(self.renderer) def blurredDiff(self): edges = self.renderer.boundarybool_image imshape = self.renderer.shape rgbEdges = np.tile(edges.reshape([imshape[0],imshape[1],1]),[1,1,3]).astype(np.bool) from opendr.filters import filter_for from opendr.filters import GaussianKernel2D blur_mtx = filter_for(imshape[0], imshape[1], imshape[2] if len(imshape)>2 else 1, kernel = GaussianKernel2D(3, 1)) blurred_diff = MatVecMult(blur_mtx, self.renderer - self.rendererGT).reshape(imshape) return blurred_diff[rgbEdges] class NLLRobustModel(Ch): terms = ['Q', 'variances'] dterms = ['renderer', 'groundtruth'] def compute_r(self): return -np.sum(np.log(self.prob)) def compute_dr_wrt(self, wrt): if wrt is self.renderer: fgMask = np.array(self.renderer.image_mesh_bool([0])).astype(np.bool) dr = (-1./(self.prob) * fgMask * self.fgProb[:,:,0]*self.fgProb[:,:,1]*self.fgProb[:,:,2] * self.Q[:, :])[:, :, None] * ((self.groundtruth.r - self.renderer.r)/self.variances.r) return dr.ravel() @depends_on(dterms) def fgProb(self): return np.exp(- (self.renderer.r - self.groundtruth.r) ** 2 / (2 * self.variances.r)) * (1. / (np.sqrt(self.variances.r) * np.sqrt(2 * np.pi))) @depends_on(dterms) def prob(self): h = self.renderer.r.shape[0] w = self.renderer.r.shape[1] occProb = np.ones([h, w]) bgProb = np.ones([h, w]) fgMask = np.array(self.renderer.image_mesh_bool([0])).astype(np.bool) errorFun = fgMask[:, :]*(self.Q[:, :] * self.fgProb[:,:,0]*self.fgProb[:,:,1]*self.fgProb[:,:,2] + (1-self.Q[:, :]))+ (1- fgMask[:, :]) return errorFun # @depends_on(dterms) # def prob(self): # h = self.renderer.r.shape[0] # w = self.renderer.r.shape[1] # # occProb = np.ones([h, w]) # bgProb = np.ones([h, w]) # # fgMask = np.array(self.renderer.image_mesh_bool([0])).astype(np.bool) # # errorFun = fgMask[:, :, None] * ((self.Q[0][:, :, None] * self.fgProb) + (self.Q[1] * occProb + self.Q[2] * bgProb)[:, :, None]) + (1 - fgMask[:, :, None]) # # return errorFun class NLLRobustSQErrorModel(Ch): terms = ['Q', 'variances'] dterms = ['sqeRenderer'] def compute_r(self): return -np.sum(np.log(self.prob)) def compute_dr_wrt(self, wrt): if wrt is self.sqeRenderer: fgMask = np.array(self.sqeRenderer.image_mesh_bool([0])).astype(np.bool) # dr = (-1./(self.prob) * fgMask * self.fgProb[:,:,0]*self.fgProb[:,:,1]*self.fgProb[:,:,2] * self.Q[:, :])[:, :, None] * ((self.sqeRenderer.imageGT.r - self.sqeRenderer.render_image)/self.variances.r) dr = np.tile((-1./(self.prob) * fgMask * self.fgProb[:,:,0]*self.fgProb[:,:,1]*self.fgProb[:,:,2] * self.Q[:, :])[:, :, None], [1,1,3]) * (-0.5/self.variances.r) # dr = (-1./(self.prob[:, :,None]) * fgMask[:, :,None] * self.fgProb * self.Q[:, :,None]) * (-0.5/self.variances.r) return dr.ravel() @depends_on(dterms) def fgProb(self): return np.exp(- (self.sqeRenderer.r) / (2 * self.variances.r)) * (1. / (np.sqrt(self.variances.r) * np.sqrt(2 * np.pi))) @depends_on(dterms) def prob(self): h = self.sqeRenderer.r.shape[0] w = self.sqeRenderer.r.shape[1] occProb = np.ones([h, w]) bgProb = np.ones([h, w]) fgMask = np.array(self.sqeRenderer.image_mesh_bool([0])).astype(np.bool) errorFun = fgMask[:, :]*(self.Q[:, :] * self.fgProb[:,:,0]*self.fgProb[:,:,1]*self.fgProb[:,:,2] + (1-self.Q[:, :]))+ (1- fgMask[:, :]) return errorFun class NLLCRFModel(Ch): terms = ['Q', 'variances'] dterms = ['renderer', 'groundtruth'] def compute_r(self): return -np.sum(np.log(self.prob)) def compute_dr_wrt(self, wrt): if wrt is self.renderer: fgMask = np.array(self.renderer.image_mesh_bool([0])).astype(np.bool) dr = -1./(self.prob) * fgMask[:, :, None] * self.fgProb * self.Q[0][:, :, None] * ((self.groundtruth.r - self.renderer.r)/self.variances.r) return dr.ravel() @depends_on(dterms) def fgProb(self): return np.exp(- (self.renderer.r - self.groundtruth.r) ** 2 / (2 * self.variances.r)) * (1. / (np.sqrt(self.variances.r) * np.sqrt(2 * np.pi))) @depends_on(dterms) def prob(self): h = self.renderer.r.shape[0] w = self.renderer.r.shape[1] occProb = np.ones([h, w]) bgProb = np.ones([h, w]) fgMask = np.array(self.renderer.image_mesh_bool([0])).astype(np.bool) errorFun = fgMask[:, :, None]*((self.Q[0][:, :, None] * self.fgProb) + (self.Q[1] * occProb + self.Q[2] * bgProb)[:, :, None])[:, :, None] + (1- fgMask[:, :, None]) return errorFun @depends_on(dterms) def prob(self): h = self.renderer.r.shape[0] w = self.renderer.r.shape[1] occProb = np.ones([h, w]) bgProb = np.ones([h, w]) fgMask = np.array(self.renderer.image_mesh_bool([0])).astype(np.bool) errorFun = fgMask[:, :, None] * ((self.Q[0][:, :, None] * self.fgProb) + (self.Q[1] * occProb + self.Q[2] * bgProb)[:, :, None]) + (1 - fgMask[:, :, None]) return errorFun class LogCRFModel(Ch): terms = ['groundtruth', 'Q', 'variances'] dterms = ['renderer'] def compute_r(self): return self.logProb() def compute_dr_wrt(self, wrt): if wrt is self.renderer: return self.logProb().dr_wrt(self.renderer) def logProb(self): fgProb = ch.exp(- (self.renderer - self.groundtruth.r) ** 2 / (2 * self.variances.r)) * (1. / (np.sqrt(self.variances.r)* np.sqrt(2 * np.pi))) h = self.renderer.r.shape[0] w = self.renderer.r.shape[1] occProb = np.ones([h, w]) bgProb = np.ones([h, w]) fgMask = np.array(self.renderer.image_mesh_bool([0])).astype(np.bool) errorFun = ch.log(fgMask[:, :, None]*((self.Q[0][:, :, None] * fgProb) + (self.Q[1] * occProb + self.Q[2] * bgProb)[:, :, None]) + (1- fgMask[:, :, None])) return errorFun class LogRobustModelOld(Ch): dterms = ['renderer', 'groundtruth', 'foregroundPrior', 'variances'] def compute_r(self): return self.logProb() def compute_dr_wrt(self, wrt): if wrt is self.renderer: return self.logProb().dr_wrt(self.renderer) def logProb(self): visibility = self.renderer.visibility_image visible = visibility != 4294967295 visible = np.array(self.renderer.image_mesh_bool([0])).copy().astype(np.bool) return ch.log(pixelLikelihoodRobustCh(self.groundtruth, self.renderer, visible, 'MASK', self.foregroundPrior, self.variances)) class LogRobustModel(Ch): terms = ['useMask'] dterms = ['renderer', 'groundtruth', 'foregroundPrior', 'variances'] def compute_r(self): return self.logProb() def compute_dr_wrt(self, wrt): if wrt is self.renderer: return self.logProb().dr_wrt(self.renderer) def logProb(self): # visibility = self.renderer.visibility_image # visible = visibility != 4294967295 try: self.useMask except: self.useMask = False if self.useMask: visible = self.renderer.indices_image != 0 else: visible = np.ones_like(self.renderer.indices_image.astype(np.bool)) # visible = np.array(self.renderer.image_mesh_bool([0])).copy().astype(np.bool) return ch.log(pixelLikelihoodRobustCh(self.groundtruth, self.renderer, visible, 'MASK', self.foregroundPrior, self.variances)) class LogRobustSQErrorModel(Ch): dterms = ['sqeRenderer', 'foregroundPrior', 'variances'] def compute_r(self): return self.logProb() def compute_dr_wrt(self, wrt): if wrt is self.sqeRenderer: return self.logProb().dr_wrt(self.sqeRenderer) def logProb(self): visibility = self.sqeRenderer.visibility_image # visible = visibility != 4294967295 visible = np.array(self.sqeRenderer.image_mesh_bool([0])).copy().astype(np.bool) return ch.log(pixelLikelihoodRobustSQErrorCh(self.sqeRenderer, visible, 'MASK', self.foregroundPrior, self.variances)) class LogRobustModelRegion(Ch): dterms = ['renderer', 'groundtruth', 'foregroundPrior', 'variances'] def compute_r(self): return self.logProb() def compute_dr_wrt(self, wrt): if wrt is self.renderer: return self.logProb().dr_wrt(self.renderer) def logProb(self): visibility = self.renderer.visibility_image visible = visibility != 4294967295 visible = np.array(self.renderer.image_mesh_bool([0])).copy().astype(np.bool) return ch.log(pixelLikelihoodRobustRegionCh(self.groundtruth, self.renderer, visible, 'MASK', self.foregroundPrior, self.variances)) class LogGaussianModelOld(Ch): dterms = ['renderer', 'groundtruth', 'variances'] def compute_r(self): return self.logProb() def compute_dr_wrt(self, wrt): if wrt is self.renderer: return self.logProb().dr_wrt(self.renderer) def logProb(self): visibility = self.renderer.visibility_image visible = visibility != 4294967295 visible = np.array(self.renderer.image_mesh_bool([0])).copy().astype(np.bool) return logPixelLikelihoodCh(self.groundtruth, self.renderer, visible, 'MASK', self.variances) class LogGaussianModel(Ch): terms = ['useMask'] dterms = ['renderer', 'groundtruth', 'variances'] def compute_r(self): return self.logProb() def compute_dr_wrt(self, wrt): if wrt is self.renderer: return self.logProb().dr_wrt(self.renderer) def logProb(self): # visibility = self.renderer.visibility_image # visible = visibility != 4294967295 try: self.useMask except: self.useMask = False if self.useMask: visible = self.renderer.indices_image != 0 # assumes the first mesh is the background cube. else: visible = np.ones_like(self.renderer.indices_image.astype(np.bool)) # visible = np.array(self.renderer.image_mesh_bool([0])).copy().astype(np.bool) return logPixelLikelihoodCh(self.groundtruth, self.renderer, visible, 'MASK', self.variances) def pixelLikelihood(image, template, testMask, backgroundModel, variances): sigma = np.sqrt(variances) # sum = np.sum(np.log(layerPrior * scipy.stats.norm.pdf(image, location = template, scale=np.sqrt(variances) ) + (1 - repPriors))) mask = testMask if backgroundModel == 'FULL': mask = np.ones(image.shape[0:2]) # mask = np.repeat(mask[..., np.newaxis], 3, 2) uniformProbs = np.ones(image.shape[0:2]) normalProbs = np.prod((1/(sigma * np.sqrt(2 * np.pi)) * np.exp( - (image - template)**2 / (2 * variances))),axis=2) return normalProbs * mask + (1-mask) def logPixelLikelihoodCh(image, template, testMask, backgroundModel, variances): sigma = ch.sqrt(variances) # sum = np.sum(np.log(layerPrior * scipy.stats.norm.pdf(image, location = template, scale=np.sqrt(variances) ) + (1 - repPriors))) mask = testMask if backgroundModel == 'FULL': mask = np.ones(image.shape[0:2]) # mask = np.repeat(mask[..., np.newaxis], 3, 2) uniformProbs = np.ones(image.shape[0:2]) logprobs = (-(image - template)**2 / (2. * variances)) - ch.log((sigma * np.sqrt(2.0 * np.pi))) pixelLogProbs = logprobs[:,:,0] + logprobs[:,:,1] + logprobs[:,:,2] return pixelLogProbs * mask def pixelLikelihoodCh(image, template, testMask, backgroundModel, layerPrior, variances): sigma = ch.sqrt(variances) mask = testMask if backgroundModel == 'FULL': mask = np.ones(image.shape[0:2]) # mask = np.repeat(mask[..., np.newaxis], 3, 2) repPriors = ch.tile(layerPrior, image.shape[0:2]) # sum = np.sum(np.log(layerPrior * scipy.stats.norm.pdf(image, location = template, scale=np.sqrt(variances) ) + (1 - repPriors))) # uniformProbs = np.ones(image.shape) probs = ch.exp( - (image - template)**2 / (2 * variances)) * (1./(sigma * np.sqrt(2 * np.pi))) foregroundProbs = (probs[:,:,0] * probs[:,:,1] * probs[:,:,2]) return foregroundProbs * mask + (1-mask) def layerPosteriorsRobust(image, template, testMask, backgroundModel, layerPrior, variances): sigma = np.sqrt(variances) mask = testMask if backgroundModel == 'FULL': mask = np.ones(image.shape[0:2]) # mask = np.repeat(mask[..., np.newaxis], 3, 2) repPriors = np.tile(layerPrior, image.shape[0:2]) foregroundProbs = np.prod(1/(sigma * np.sqrt(2 * np.pi)) * np.exp( - (image - template)**2 / (2 * variances)) * layerPrior, axis=2) backgroundProbs = np.ones(image.shape) outlierProbs = (1-repPriors) lik = pixelLikelihoodRobust(image, template, testMask, backgroundModel, layerPrior, variances) # prodlik = np.prod(lik, axis=2) # return np.prod(foregroundProbs*mask, axis=2)/prodlik, np.prod(outlierProbs*mask, axis=2)/prodlik return foregroundProbs*mask/lik, outlierProbs*mask/lik def layerPosteriorsRobustCh(image, template, testMask, backgroundModel, layerPrior, variances): sigma = ch.sqrt(variances) mask = testMask if backgroundModel == 'FULL': mask = np.ones(image.shape[0:2]) # mask = np.repeat(mask[..., np.newaxis], 3, 2) repPriors = ch.tile(layerPrior, image.shape[0:2]) probs = ch.exp( - (image - template)**2 / (2 * variances)) * (1/(sigma * np.sqrt(2 * np.pi))) foregroundProbs = probs[:,:,0] * probs[:,:,1] * probs[:,:,2] * layerPrior backgroundProbs = np.ones(image.shape) outlierProbs = ch.Ch(1-repPriors) lik = pixelLikelihoodRobustCh(image, template, testMask, backgroundModel, layerPrior, variances) # prodlik = np.prod(lik, axis=2) # return np.prod(foregroundProbs*mask, axis=2)/prodlik, np.prod(outlierProbs*mask, axis=2)/prodlik return foregroundProbs*mask/lik, outlierProbs*mask/lik def layerPosteriorsRobustSQErrorCh(sqeRenderer, testMask, backgroundModel, layerPrior, variances): sigma = ch.sqrt(variances) mask = testMask if backgroundModel == 'FULL': mask = np.ones(sqeRenderer.r.shape[0:2]) # mask = np.repeat(mask[..., np.newaxis], 3, 2) repPriors = ch.tile(layerPrior, sqeRenderer.r.shape[0:2]) probs = ch.exp( - (sqeRenderer) / (2 * variances)) * (1/(sigma * np.sqrt(2 * np.pi))) foregroundProbs = probs[:,:,0] * probs[:,:,1] * probs[:,:,2] * layerPrior backgroundProbs = np.ones(sqeRenderer.r.shape) outlierProbs = ch.Ch(1-repPriors) lik = pixelLikelihoodRobustSQErrorCh(sqeRenderer, testMask, backgroundModel, layerPrior, variances) # prodlik = np.prod(lik, axis=2) # return np.prod(foregroundProbs*mask, axis=2)/prodlik, np.prod(outlierProbs*mask, axis=2)/prodlik return foregroundProbs*mask/lik, outlierProbs*mask/lik def robustDistance(sqResiduals, scale): return np.sum(sqResiduals/(sqResiduals + scale**2)) def testImageMatching(): minThresTemplate = 10 maxThresTemplate = 100 methodParams = {'scale': 85000, 'minThresImage': minThresTemplate, 'maxThresImage': maxThresTemplate, 'minThresTemplate': minThresTemplate, 'maxThresTemplate': maxThresTemplate} teapots = ["test/teapot1", "test/teapot2","test/teapot3","test/teapot4","test/teapot5","test/teapot6"] images = [] edges = [] for teapot in teapots: im = cv2.imread(teapot + ".png") can = cv2.Canny(im, minThresTemplate,maxThresTemplate) images.append(im) edges.append(can) cv2.imwrite(teapot + "_can.png", can) confusion = np.zeros([6,6]) for tp1 in np.arange(1,7): for tp2 in np.arange(tp1,7): dist = distance = scoreImage(images[tp1-1], images[tp2-1], 'robustSqDistImages', methodParams) print(dist) confusion[tp1-1, tp2-1] = dist plt.matshow(confusion) plt.colorbar() plt.savefig('test/confusion.png') # elif method == 'chamferDataToModel': # sqDists = chamferDistanceDataToModel(img, template, methodParams['minThresImage'], methodParams['maxThresImage'], methodParams['minThresTemplate'],methodParams['maxThresTemplate']) # score = np.sum(sqDists) # elif method == 'robustChamferDataToModel': # sqDists = np.sum(chamferDistanceDataToModel(img, template, methodParams['minThresImage'], methodParams['maxThresImage'], methodParams['minThresTemplate'],methodParams['maxThresTemplate'])) # score = robustDistance(sqDists, methodParams['robustScale']) # elif method == 'sqDistImages': # sqDists = sqDistImages(img, template) # score = np.sum(sqDists) # elif method == 'robustSqDistImages':
25,302
38.910095
213
py
inversegraphics
inversegraphics-master/var_inf.py
__author__ = 'pol' import ipdb import matplotlib matplotlib.use('Qt4Agg') from math import radians import chumpy as ch import numpy as np import cv2 import matplotlib.pyplot as plt from sklearn import mixture from numpy.random import choice plt.ion() image = cv2.imread('opendr_GT.png') image = np.float64(cv2.cvtColor(image, cv2.COLOR_BGR2RGB))/255.0 nComps = 5 nRecComps = 4 gmm = mixture.GMM(n_components=nComps, covariance_type='spherical') win = 40 colors = image[image.shape[0]/2-win:image.shape[0]/2+win,image.shape[1]/2-win:image.shape[1]/2+win,:].reshape([4*win*win,3]) gmm.fit(colors) imshape = [win*2,win*2,3] numPixels = win*2 * win*2 chInput = ch.Ch(colors) numVars = chInput.size recSoftmaxW = ch.Ch(np.random.uniform(0,1, [nRecComps,numVars])/numVars) chRecLogistic = ch.exp(ch.dot(recSoftmaxW,chInput.reshape([numVars,1]))) chRecSoftmax = chRecLogistic.ravel()/ch.sum(chRecLogistic) chZRecComps = ch.zeros([numVars, nRecComps]) chZ = ch.zeros([numVars]) recMeans = ch.Ch(np.random.uniform(0,1, [3,nRecComps])) recCovars = 0.2 chRecLogLikelihoods = - 0.5*(chZ.reshape([numPixels,3, 1]) - ch.tile(recMeans, [numPixels, 1,1]))**2 - ch.log((2 * recCovars) * (1/(ch.sqrt(recCovars) * np.sqrt(2 * np.pi)))) genZCovars = 0.2 chGenComponentsProbs = ch.Ch(gmm.weights_) chCompMeans = ch.zeros([nComps, 3]) for comp in range(nComps): chCompMeans[comp, :] = gmm.means_[comp] chPZComp = ch.exp( - (ch.tile(chZ.reshape([numPixels,3,1]), [1, 1, nComps]) - chCompMeans.reshape([1,3, nComps]))**2 / (2 * genZCovars)) * (1/(ch.sqrt(genZCovars) * np.sqrt(2 * np.pi))) chPZ = ch.dot(chGenComponentsProbs.reshape([1,nComps]), chPZComp.reshape([5, numVars])) prec = 0.5 covars = np.eye(colors.size, colors.size) # for row in np.arange(covars.shape[0]): # cols = [max(row-1, 0), min(row+1,colors.size-1)] # covars[row, cols[0]] = prec # covars[row, cols[1]] = prec # covars[cols[0], row] = prec # covars[cols[1], row] = prec # covars = np.linalg.inv(covars) detCov = 1 # detCov = np.linalg.det(covars) # covars = [ch.Ch([covar] for covar in gmm.covars_)] chResiduals = chInput.ravel() - chZ.ravel() covar = 0.2 ipdb.set_trace() chLogJoint = ch.log(chPZ.ravel()) - 0.5*covar*ch.dot(chResiduals,chResiduals) - 0.5*(ch.log(detCov) + numVars*ch.log((2 * np.pi))) # chGenMarginal = ch.prod(chLikelihoods) ipdb.set_trace() # likelihoodsZ = [chGenComponentsProbs[comp]*ch.exp( - (chInput - chZ)**2 / (2 * covars)) * (1/(ch.sqrt(covars) * np.sqrt(2 * np.pi))) for comp in range(nComps)] # chLikelihoodsZ = ch.concatenate(likelihoods) # chGenMarginalZ = ch.exp(ch.sum(ch.log(chLikelihoodsZ))) gmmRec = mixture.GMM(n_components=nRecComps, covariance_type='spherical') gmmRec.covars_=gmm.covars_.copy() #Update the mean of the gaussians and update the mixing weights. methods=['dogleg', 'minimize', 'BFGS', 'L-BFGS-B', 'Nelder-Mead'] free_vars = [recMeans.ravel(), recSoftmaxW] print("Beginning optimization.") while True: gmmRec.weights_=np.array(chRecSoftmax.r) gmmRec.means_=np.array(ch.concatenate(recMeans)) epsilon = np.random.randn(numVars) u = choice(nRecComps, size=1, p=chRecSoftmax.r) chZ[:] = chZRecComps[:,u].r.ravel() + recCovars*epsilon.ravel() pu = chRecSoftmax L = ch.log(pu[u]) + ch.sum(chLogJoint.ravel()) - ch.sum(chRecLogLikelihoods[:,:,u].ravel()) drL = L.dr_wrt(recMeans)/numPixels alpha = 0.1 recSoftmaxW[:] = recSoftmaxW.r[:] + alpha*L.dr_wrt(recSoftmaxW).reshape(recSoftmaxW.shape)/numPixels ipdb.set_trace() chZ[:] = chZ.r[:] + alpha*L.dr_wrt(chZ).reshape(chZ.r.shape)/numPixels chZRecComps[:,u] = chZ.r[:] # ch.minimize({'raw': -L}, bounds=None, method=methods[1], x0=free_vars, callback=None, options={'disp':False, 'maxiter':1})
3,755
32.837838
186
py
inversegraphics
inversegraphics-master/geometry.py
import chumpy as ch from chumpy import depends_on, Ch import cv2 import numpy as np import scipy.sparse as sp from chumpy.utils import row, col from opendr.geometry import Rodrigues class RotateZ(Ch): dterms = 'a' def compute_r(self): return np.array([[np.cos(self.a.r), -np.sin(self.a.r), 0, 0], [np.sin(self.a.r), np.cos(self.a.r), 0, 0], [0, 0, 1, 0], [0,0,0,1]]) def compute_dr_wrt(self, wrt): if wrt is not self.a: return if wrt is self.a: return np.array([[-np.sin(self.a.r)[0], -np.cos(self.a.r)[0], 0, 0], [np.cos(self.a.r)[0], -np.sin(self.a.r)[0], 0, 0], [0, 0, 0, 0], [0,0,0,0]]).reshape(16,1) class RotateX(Ch): dterms = 'a' def compute_r(self): return np.array([[1, 0, 0, 0],[0, np.cos(self.a.r), -np.sin(self.a.r), 0], [0, np.sin(self.a.r), np.cos(self.a.r),0],[0,0,0,1]]) def compute_dr_wrt(self, wrt): # if wrt is not self.a: # return if wrt is self.a: return np.array([[0, 0, 0, 0],[0, -np.sin(self.a.r)[0], -np.cos(self.a.r)[0], 0], [0, np.cos(self.a.r)[0], -np.sin(self.a.r)[0],0],[0,0,0,0]]).reshape(16,1) class Scale(Ch): dterms = 'x', 'y','z' def compute_r(self): return np.array([[self.x.r, 0, 0, 0],[0, self.y.r, 0, 0], [0, 0, self.z.r,0],[0,0,0,1]]) def compute_dr_wrt(self, wrt): return def compute_dr_wrt(self, wrt): if wrt is not self.x and wrt is not self.y and wrt is not self.z: return if wrt is self.x: return np.array([[1, 0, 0, 0],[0, 0, 0, 0], [0, 0, 0,0],[0,0,0,0]]).reshape(16,1) if wrt is self.y: return np.array([[0, 0, 0, 0],[0, 1, 0, 0], [0, 0, 0,0],[0,0,0,0]]).reshape(16,1) if wrt is self.z: return np.array([[0, 0, 0, 0],[0, 0, 0, 0], [0, 0, 1, 0],[0,0,0,0]]).reshape(16,1) class Translate(Ch): dterms = 'x', 'y', 'z' def compute_r(self): return np.array([[1, 0, 0, self.x.r],[0, 1, 0, self.y.r], [0, 0, 1,self.z.r],[0,0,0,1]]) def compute_dr_wrt(self, wrt): if wrt is not self.x and wrt is not self.y and wrt is not self.z: return if wrt is self.x: return np.array([[0, 0, 0, 1],[0, 0, 0, 0], [0, 0, 0,0],[0,0,0,0]]).reshape(16,1) if wrt is self.y: return np.array([[0, 0, 0, 0],[0, 0, 0, 1], [0, 0, 0,0],[0,0,0,0]]).reshape(16,1) if wrt is self.z: return np.array([[0, 0, 0, 0],[0, 0, 0, 0], [0, 0, 0, 1],[0,0,0,0]]).reshape(16,1) def getNormals(vertices, faces): norm = np.zeros( vertices.shape, dtype=vertices.dtype ) tris = vertices[faces] n = np.cross( tris[::,1 ] - tris[::,0] , tris[::,2 ] - tris[::,0] ) normalize_v3(n) norm[ faces[:,0] ] += n norm[ faces[:,1] ] += n norm[ faces[:,2] ] += n normalize_v3(norm) return norm def chGetNormals(vertices, faces): import opendr.geometry return opendr.geometry.VertNormals(vertices, faces).reshape((-1,3))
3,023
30.831579
171
py
inversegraphics
inversegraphics-master/diffrender_analyze.py
__author__ = 'pol' import matplotlib matplotlib.use('Qt4Agg') import scene_io_utils from math import radians import timeit import time import opendr import chumpy as ch import geometry import image_processing import numpy as np import cv2 import generative_models import recognition_models import matplotlib.pyplot as plt from opendr_utils import * from utils import * import OpenGL.GL as GL import glfw useShapeModel = True ######################################### # OpenDR Initialization starts here ######################################### #Main script options: glModes = ['glfw','mesa'] glMode = glModes[0] width, height = (150, 150) win = -1 if glMode == 'glfw': #Initialize base GLFW context for the Demo and to share context among all renderers. glfw.init() glfw.window_hint(glfw.CONTEXT_VERSION_MAJOR, 3) glfw.window_hint(glfw.CONTEXT_VERSION_MINOR, 3) # glfw.window_hint(glfw.OPENGL_FORWARD_COMPAT, GL.GL_TRUE) glfw.window_hint(glfw.OPENGL_PROFILE, glfw.OPENGL_CORE_PROFILE) glfw.window_hint(glfw.DEPTH_BITS,32) glfw.window_hint(glfw.VISIBLE, GL.GL_FALSE) win = glfw.create_window(width, height, "Demo", None, None) glfw.make_context_current(win) angle = 60 * 180 / np.pi clip_start = 0.05 clip_end = 10 frustum = {'near': clip_start, 'far': clip_end, 'width': width, 'height': height} camDistance = 0.4 teapots = [line.strip() for line in open('teapots.txt')] # renderTeapotsList = np.arange(len(teapots)) renderTeapotsList = np.arange(len(teapots))[0:1] targetModels = [] v_teapots, f_list_teapots, vc_teapots, vn_teapots, uv_teapots, haveTextures_list_teapots, textures_list_teapots, vflat, varray, center_teapots = scene_io_utils.loadTeapotsOpenDRData(renderTeapotsList, False, False, targetModels) azimuth = np.pi chCosAz = ch.Ch([np.cos(azimuth)]) chSinAz = ch.Ch([np.sin(azimuth)]) chAz = 2*ch.arctan(chSinAz/(ch.sqrt(chCosAz**2 + chSinAz**2) + chCosAz)) chAz = ch.Ch([0]) chObjAz = ch.Ch([0]) chAzRel = chAz - chObjAz elevation = 0 chLogCosEl = ch.Ch(np.log(np.cos(elevation))) chLogSinEl = ch.Ch(np.log(np.sin(elevation))) chEl = 2*ch.arctan(ch.exp(chLogSinEl)/(ch.sqrt(ch.exp(chLogCosEl)**2 + ch.exp(chLogSinEl)**2) + ch.exp(chLogCosEl))) chEl = ch.Ch([0.95993109]) chDist = ch.Ch([camDistance]) chLightSHCoeffs = ch.Ch(np.array([2, 0.25, 0.25, 0.12,-0.17,0.36,0.1,0.,0.])) clampedCosCoeffs = clampedCosineCoefficients() chComponent = chLightSHCoeffs * clampedCosCoeffs chPointLightIntensity = ch.Ch([1]) chLightAz = ch.Ch([0.0]) chLightEl = ch.Ch([np.pi/2]) chLightDist = ch.Ch([0.5]) light_color = ch.ones(3)*chPointLightIntensity chVColors = ch.Ch([0.4,0.4,0.4]) chDisplacement = ch.Ch([0.0, 0.0,0.0]) chScale = ch.Ch([1.0,1.0,1.0]) # vcch[0] = np.ones_like(vcflat[0])*chVColorsGT.reshape([1,3]) renderer_teapots = [] for teapot_i in range(len(renderTeapotsList)): vmod = v_teapots[teapot_i] fmod_list = f_list_teapots[teapot_i] vcmod = vc_teapots[teapot_i] vnmod = vn_teapots[teapot_i] uvmod = uv_teapots[teapot_i] haveTexturesmod_list = haveTextures_list_teapots[teapot_i] texturesmod_list = textures_list_teapots[teapot_i] centermod = center_teapots[teapot_i] vmod, vnmod, _ = transformObject(vmod, vnmod, chScale, chObjAz, ch.Ch([0]), ch.Ch([0]), np.array([0, 0, 0])) renderer = createRendererTarget(glMode, chAz, chEl, chDist, centermod, vmod, vcmod, fmod_list, vnmod, light_color, chComponent, chVColors, 0, chDisplacement, width,height, uvmod, haveTexturesmod_list, texturesmod_list, frustum, win ) renderer.overdraw = True renderer.nsamples = 8 renderer.msaa = False renderer.initGL() renderer.initGLTexture() # renderer.initGL_AnalyticRenderer() renderer.imageGT = None renderer.r renderer_teapots = renderer_teapots + [renderer] currentTeapotModel = 0 center = center_teapots[currentTeapotModel] if useShapeModel: import shape_model #%% Load data filePath = 'data/teapotModel.pkl' teapotModel = shape_model.loadObject(filePath) faces = teapotModel['faces'] #%% Sample random shape Params latentDim = np.shape(teapotModel['ppcaW'])[1] shapeParams = np.random.randn(latentDim) chShapeParams = ch.Ch(shapeParams) meshLinearTransform=teapotModel['meshLinearTransform'] W=teapotModel['ppcaW'] b=teapotModel['ppcaB'] chVertices = shape_model.VerticesModel(chShapeParams=chShapeParams,meshLinearTransform=meshLinearTransform,W = W,b=b) chVertices.init() chVertices = ch.dot(geometry.RotateZ(-np.pi/2)[0:3,0:3],chVertices.T).T chNormals = shape_model.chGetNormals(chVertices, faces) smNormals = [chNormals] smFaces = [[faces]] smVColors = [chVColors*np.ones(chVertices.shape)] smUVs = [ch.Ch(np.zeros([chVertices.shape[0],2]))] smHaveTextures = [[False]] smTexturesList = [[None]] chVertices = chVertices - ch.mean(chVertices, axis=0) minZ = ch.min(chVertices[:,2]) chMinZ = ch.min(chVertices[:,2]) zeroZVerts = chVertices[:,2]- chMinZ chVertices = ch.hstack([chVertices[:,0:2] , zeroZVerts.reshape([-1,1])]) chVertices = chVertices*0.09 smCenter = ch.array([0,0,0.1]) smVertices = [chVertices] chNormals = shape_model.chGetNormals(chVertices, faces) smNormals = [chNormals] smVertices, smNormals, _ = transformObject(smVertices, smNormals, chScale, chObjAz, ch.Ch([0]), ch.Ch([0]), np.array([0, 0, 0])) renderer = createRendererTarget(glMode, chAz, chEl, chDist, smCenter, [smVertices], [smVColors], [smFaces], [smNormals], light_color, chComponent, chVColors, 0, chDisplacement, width,height, [smUVs], [smHaveTextures], [smTexturesList], frustum, win ) renderer.overdraw = True renderer.nsamples = 8 renderer.msaa = False renderer.initGL() renderer.initGLTexture() # renderer.initGL_AnalyticRenderer() renderer.imageGT = None renderer.r chShapeParams[:] = np.zeros([latentDim]) chVerticesMean = chVertices.r.copy() else: renderer = renderer_teapots[0] ##GT Renderer chObjAzGT = ch.Ch([0]) chAzGT = ch.Ch([0]) chAzRelGT = chAzGT - chObjAzGT chElGT = ch.Ch(chEl.r[0]) chDistGT = ch.Ch([camDistance]) phiOffset = ch.Ch([0]) totalOffset = phiOffset + chObjAzGT chVColorsGT = ch.Ch([0.8,0.8,0.8]) chAmbientIntensityGT = ch.Ch([0.025]) clampedCosCoeffs = clampedCosineCoefficients() SHFilename = 'data/LightSHCoefficients.pickle' with open(SHFilename, 'rb') as pfile: envMapDic = pickle.load(pfile) hdritems = list(envMapDic.items()) envMapCoeffs = ch.Ch(list(envMapDic.items())[0][1][1]) envMapCoeffsRotated = ch.Ch(np.dot(light_probes.chSphericalHarmonicsZRotation(totalOffset), envMapCoeffs[[0,3,2,1,4,5,6,7,8]])[[0,3,2,1,4,5,6,7,8]]) envMapCoeffsRotatedRel = ch.Ch(np.dot(light_probes.chSphericalHarmonicsZRotation(phiOffset), envMapCoeffs[[0,3,2,1,4,5,6,7,8]])[[0,3,2,1,4,5,6,7,8]]) shCoeffsRGB = envMapCoeffsRotated shCoeffsRGBRel = envMapCoeffsRotatedRel chShCoeffs = 0.3*shCoeffsRGB[:,0] + 0.59*shCoeffsRGB[:,1] + 0.11*shCoeffsRGB[:,2] chShCoeffsRel = 0.3*shCoeffsRGBRel[:,0] + 0.59*shCoeffsRGBRel[:,1] + 0.11*shCoeffsRGBRel[:,2] chAmbientSHGT = chShCoeffs.ravel() * chAmbientIntensityGT * clampedCosCoeffs chAmbientSHGTRel = chShCoeffsRel.ravel() * chAmbientIntensityGT * clampedCosCoeffs chComponentGT = chAmbientSHGT chComponentGTRel = chAmbientSHGTRel # shapeParams = np.random.randn(latentDim) if useShapeModel: shapeParams = np.random.randn(latentDim) chShapeParamsGT = ch.Ch(shapeParams) chVerticesGT = shape_model.VerticesModel(chShapeParams =chShapeParamsGT,meshLinearTransform=meshLinearTransform,W=W,b=b) chVerticesGT.init() chVerticesGT = ch.dot(geometry.RotateZ(-np.pi/2)[0:3,0:3],chVerticesGT.T).T # chNormalsGT = shape_model.chShapeParamsToNormals(teapotModel['N'], landmarks, teapotModel['linT']) # chNormalsGT = shape_model.shapeParamsToNormals(shapeParams, teapotModel) chNormalsGT = shape_model.chGetNormals(chVerticesGT, faces) smNormalsGT = [chNormalsGT] smFacesGT = [[faces]] smVColorsGT = [chVColorsGT*np.ones(chVerticesGT.shape)] smUVsGT = [ch.Ch(np.zeros([chVerticesGT.shape[0],2]))] smHaveTexturesGT = [[False]] smTexturesListGT = [[None]] smCenterGT = ch.mean(chVerticesGT, axis=0) chVerticesGT = chVerticesGT - ch.mean(chVerticesGT, axis=0) minZ = ch.min(chVerticesGT[:,2]) chMinZ = ch.min(chVerticesGT[:,2]) zeroZVerts = chVerticesGT[:,2]- chMinZ chVerticesGT = ch.hstack([chVerticesGT[:,0:2] , zeroZVerts.reshape([-1,1])]) chVerticesGT = chVerticesGT*0.09 smCenterGT = ch.array([0,0,0.1]) smVerticesGT = [chVerticesGT] ######################################### # Initialization ends here ######################################### ######################################### # Generative model set up ######################################### rendererGT = ch.Ch(renderer.r.copy()) numPixels = width*height E_raw = renderer - rendererGT SE_raw = ch.sum(E_raw*E_raw, axis=2) SSqE_raw = ch.SumOfSquares(E_raw)/numPixels initialPixelStdev = 0.01 reduceVariance = False # finalPixelStdev = 0.05 stds = ch.Ch([initialPixelStdev]) variances = stds ** 2 globalPrior = ch.Ch([0.9]) negLikModel = -ch.sum(generative_models.LogGaussianModel(renderer=renderer, groundtruth=rendererGT, variances=variances)) negLikModelRobust = -ch.sum(generative_models.LogRobustModel(renderer=renderer, groundtruth=rendererGT, foregroundPrior=globalPrior, variances=variances)) pixelLikelihoodCh = generative_models.LogGaussianModel(renderer=renderer, groundtruth=rendererGT, variances=variances) pixelLikelihoodRobustCh = generative_models.LogRobustModel(renderer=renderer, groundtruth=rendererGT, foregroundPrior=globalPrior, variances=variances) # negLikModel = -generative_models.modelLogLikelihoodCh(rendererGT, renderer, np.array([]), 'FULL', variances)/numPixels # negLikModelRobust = -generative_models.modelLogLikelihoodRobustCh(rendererGT, renderer, np.array([]), 'FULL', globalPrior, variances)/numPixels # pixelLikelihoodCh = generative_models.logPixelLikelihoodCh(rendererGT, renderer, np.array([]), 'FULL', variances) # pixelLikelihoodRobustCh = ch.log(generative_models.pixelLikelihoodRobustCh(rendererGT, renderer, np.array([]), 'FULL', globalPrior, variances)) post = generative_models.layerPosteriorsRobustCh(rendererGT, renderer, np.array([]), 'MASK', globalPrior, variances)[0] # modelLogLikelihoodRobustRegionCh = -ch.sum(generative_models.LogRobustModelRegion(renderer=renderer, groundtruth=rendererGT, foregroundPrior=globalPrior, variances=variances))/numPixels # # pixelLikelihoodRobustRegionCh = generative_models.LogRobustModelRegion(renderer=renderer, groundtruth=rendererGT, foregroundPrior=globalPrior, variances=variances) # models = [negLikModel, negLikModelRobust, hogError] models = [negLikModel, negLikModelRobust] pixelModels = [pixelLikelihoodCh, pixelLikelihoodRobustCh] modelsDescr = ["Gaussian Model", "Outlier model" ] model = 1 pixelErrorFun = pixelModels[model] errorFun = models[model] ############ # Experiments ############ seed = 1 np.random.seed(seed) gtPrefix = 'train4_occlusion_shapemodel_photorealistic_10K_test100-1100' # gtPrefix = 'train4_occlusion_shapemodel' gtDir = 'groundtruth/' + gtPrefix + '/' featuresDir = gtDir experimentPrefix = 'train4_occlusion_shapemodel_10k' experimentDir = 'experiments/' + experimentPrefix + '/' ignore = [] if os.path.isfile(gtDir + 'ignore.npy'): ignore = np.load(gtDir + 'ignore.npy') groundTruthFilename = gtDir + 'groundTruth.h5' gtDataFile = h5py.File(groundTruthFilename, 'r') rangeTests = np.arange(100,1100) idsInRange = np.arange(len(rangeTests)) testSet = np.load(experimentDir + 'test.npy')[rangeTests][idsInRange] shapeGT = gtDataFile[gtPrefix].shape boolTestSet = np.array([np.any(num == testSet) for num in gtDataFile[gtPrefix]['trainIds']]) dataIds = gtDataFile[gtPrefix][boolTestSet]['trainIds'] dataIdsTestIndices = np.array([np.where(dataIds==num)[0][0] for num in testSet]) groundTruth = gtDataFile[gtPrefix][boolTestSet][dataIdsTestIndices] dataTeapotIdsTest = groundTruth['trainTeapotIds'] test = np.arange(len(testSet)) testSet = testSet[test] print("Reading experiment.") dataAzsGT = groundTruth['trainAzsGT'] dataObjAzsGT = groundTruth['trainObjAzsGT'] dataElevsGT = groundTruth['trainElevsGT'] dataLightAzsGT = groundTruth['trainLightAzsGT'] dataLightElevsGT = groundTruth['trainLightElevsGT'] dataLightIntensitiesGT = groundTruth['trainLightIntensitiesGT'] dataVColorGT = groundTruth['trainVColorGT'] dataScenes = groundTruth['trainScenes'] dataTeapotIds = groundTruth['trainTeapotIds'] dataEnvMaps = groundTruth['trainEnvMaps'] dataOcclusions = groundTruth['trainOcclusions'] dataTargetIndices = groundTruth['trainTargetIndices'] dataLightCoefficientsGT = groundTruth['trainLightCoefficientsGT'] dataLightCoefficientsGTRel = groundTruth['trainLightCoefficientsGTRel'] dataEnvMapPhiOffsets = groundTruth['trainEnvMapPhiOffsets'] dataAmbientIntensityGT = groundTruth['trainAmbientIntensityGT'] dataIds = groundTruth['trainIds'] gtDtype = groundTruth.dtype testSetFixed = testSet whereBad = [] for test_it, test_id in enumerate(testSet): if test_id in ignore: bad = np.where(testSetFixed==test_id) testSetFixed = np.delete(testSetFixed, bad) whereBad = whereBad + [bad] # testSet = testSetFixed loadFromHdf5 = False useShapeModel = True syntheticGroundtruth = False synthPrefix = '_cycles' if syntheticGroundtruth: synthPrefix = '' if syntheticGroundtruth: imagesDir = gtDir + 'images_opendr/' else: imagesDir = gtDir + 'images/' images = readImages(imagesDir, dataIds, loadFromHdf5) # testSet = np.arange(len(images))[0:10] testAzsGT = dataAzsGT testObjAzsGT = dataObjAzsGT testElevsGT = dataElevsGT testLightAzsGT = dataLightAzsGT testLightElevsGT = dataLightElevsGT testLightIntensitiesGT = dataLightIntensitiesGT testVColorGT = dataVColorGT testOcclusions = dataOcclusions if useShapeModel: dataShapeModelCoeffsGT = groundTruth['trainShapeModelCoeffsGT'] testShapeParamsGT = dataShapeModelCoeffsGT testLightCoefficientsGTRel = dataLightCoefficientsGTRel * dataAmbientIntensityGT[:,None] testAzsRel = np.mod(testAzsGT - testObjAzsGT, 2*np.pi) ##### Load Results data: # testPrefix = 'train4_occlusion_shapemodel_10k_ECCVNEW-JOINED-ALL2018' testPrefix = 'train4_occlusion_shapemodel_10k_ECCV-PHOTOREALISTIC-JOINT2018' resultDir = 'results/' + testPrefix + '/' with open(resultDir + 'experiment.pickle', 'rb') as pfile: experimentDic = pickle.load(pfile) testSet = experimentDic['testSet'] methodsPred = experimentDic['methodsPred'] testOcclusions = experimentDic['testOcclusions'] # testOcclusions = testOcclusions[dataIdsTestIndices] #Change depending on the source of GT. # import ipdb; ipdb.set_trace() testPrefixBase = experimentDic[ 'testPrefixBase'] parameterRecognitionModels = experimentDic[ 'parameterRecognitionModels'] azimuths = experimentDic[ 'azimuths'] azimuths = [azimuths[method][dataIdsTestIndices] if azimuths[method] is not None else None for method in range(len(azimuths))] elevations = experimentDic[ 'elevations'] elevations = [elevations[method][dataIdsTestIndices] if elevations[method] is not None else None for method in range(len(elevations))] vColors = experimentDic[ 'vColors'] vColors = [vColors[method][dataIdsTestIndices] if vColors[method] is not None else None for method in range(len(vColors))] lightCoeffs = experimentDic[ 'lightCoeffs'] lightCoeffs = [lightCoeffs[method][dataIdsTestIndices] if lightCoeffs[method] is not None else None for method in range(len(lightCoeffs))] likelihoods = [] # likelihoods = experimentDic['likelihoods'] # likelihoods = [likelihoods[method][dataIdsTestIndices] if likelihoods[method] is not None else None for method in range(len(likelihoods))] shapeParams = experimentDic['shapeParams'] shapeParams = [shapeParams[method][dataIdsTestIndices] if shapeParams[method] is not None else None for method in range(len(shapeParams))] if os.path.isfile(resultDir + 'segmentations.pickle'): with open(resultDir + 'segmentations.pickle', 'rb') as pfile: segmentationsDic = pickle.load(pfile) segmentations = segmentationsDic['segmentations'] segmentations = [segmentations[method][dataIdsTestIndices] if segmentations[method] is not None else None for method in range(len(segmentations))] else: segmentations = [None]*len(methodsPred) testOcclusionsFull = testOcclusions.copy() loadMask = True if loadMask: masksGT = loadMasks(gtDir + '/masks_occlusion/', dataIds) # # with open(resultDir + 'approxProjections.pickle', 'rb') as pfile: # approxProjectionsDic = pickle.load(pfile) # # approxProjections = approxProjectionsDic['approxProjections'] # approxProjectionsGT = approxProjectionsDic['approxProjectionsGT'] # # envMapTexture = np.zeros([180,360,3]) # approxProjections = [] # for method in range(len(methodsPred)): # print("Approx projection on method " + str(method)) # approxProjectionsFittedList = [] # for test_i in range(len(testSet)): # pEnvMap = SHProjection(envMapTexture, np.concatenate([lightCoeffs[4][test_i][:,None], lightCoeffs[4][test_i][:,None], lightCoeffs[4][test_i][:,None]], axis=1)) # approxProjection = np.sum(pEnvMap, axis=(2,3)) # approxProjectionsFittedList = approxProjectionsFittedList + [approxProjection[None,:]] # approxProjections = approxProjections + [np.vstack(approxProjectionsFittedList)] # # # approxProjectionsGTList = [] # for test_i in range(len(testSet)): # pEnvMap = SHProjection(envMapTexture, np.concatenate([testLightCoefficientsGTRel[test_i][:,None], testLightCoefficientsGTRel[test_i][:,None], testLightCoefficientsGTRel[test_i][:,None]], axis=1)) # approxProjectionGT = np.sum(pEnvMap, axis=(2,3)) # approxProjectionsGTList = approxProjectionsGTList + [approxProjectionGT[None,:]] # approxProjectionsGT = np.vstack(approxProjectionsGTList) approxProjections = None approxProjectionsGT = None # # # # ##### Load Results data: # # # # testPrefix2 = 'train4_occlusion_shapemodel_10k_ECCV-PHOTREALISTIC-MEANBASELINE-2530-17944_predict_1000samples__method1errorFun1_std0.05_shapePen0' # resultDir2 = 'results/' + testPrefix2 + '/' # with open(resultDir2 + 'experiment.pickle', 'rb') as pfile: # experimentDic2 = pickle.load(pfile) # testSet2 = experimentDic2['testSet'] # methodsPred2 = experimentDic2['methodsPred'] # testOcclusions2 = experimentDic2[ 'testOcclusions'] # testPrefixBase2 = experimentDic2[ 'testPrefixBase'] # parameterRecognitionModels2 = experimentDic2[ 'parameterRecognitionModels'] # azimuths2 = experimentDic2['azimuths'] # elevations2 = experimentDic2[ 'elevations'] # vColors2 = experimentDic2[ 'vColors'] # lightCoeffs2 = experimentDic2[ 'lightCoeffs'] # likelihoods2 = experimentDic2['likelihoods'] # shapeParams2 = experimentDic2['shapeParams'] # if os.path.isfile(resultDir2 + 'segmentations.pickle'): # with open(resultDir2 + 'segmentations.pickle', 'rb') as pfile: # segmentationsDic2 = pickle.load(pfile) # segmentations2 = segmentationsDic2['segmentations'] # else: # segmentations2 = [None]*len(methodsPred2) # # testOcclusionsFull2 = testOcclusions2.copy() # # # testPrefix3 = 'train4_occlusion_shapemodel_10k_ECCV-SYNTH-FIX-2530-17944_optimize_1000samples__method1errorFun1_std0.03_shapePen0' # resultDir3 = 'results/' + testPrefix3 + '/' # with open(resultDir3 + 'experiment.pickle', 'rb') as pfile: # experimentDic3 = pickle.load(pfile) # testSet3 = experimentDic3['testSet'] # methodsPred3 = experimentDic3['methodsPred'] # testOcclusions3 = experimentDic3[ 'testOcclusions'] # testPrefixBase3 = experimentDic3[ 'testPrefixBase'] # parameterRecognitionModels3 = experimentDic3[ 'parameterRecognitionModels'] # azimuths3 = experimentDic3['azimuths'] # elevations3 = experimentDic3[ 'elevations'] # vColors3 = experimentDic3[ 'vColors'] # lightCoeffs3 = experimentDic3[ 'lightCoeffs'] # likelihoods3 = experimentDic3['likelihoods'] # shapeParams3 = experimentDic3['shapeParams'] # if os.path.isfile(resultDir3 + 'segmentations.pickle'): # with open(resultDir3 + 'segmentations.pickle', 'rb') as pfile: # segmentationsDic3 = pickle.load(pfile) # segmentations3 = segmentationsDic3['segmentations'] # else: # segmentations3 = [None]*len(methodsPred3) # # testOcclusionsFull3 = testOcclusions3.copy() # # # with open(resultDir2 + 'approxProjections.pickle', 'rb') as pfile: # # approxProjectionsDic2 = pickle.load(pfile) # # # # approxProjections2 = approxProjectionsDic2['approxProjections'] # # approxProjectionsGT2 = approxProjectionsDic2['approxProjectionsGT'] # # approxProjections2 = None # approxProjectionsGT2 = None # # testPrefix3 = 'train4_occlusion_shapemodel_10k_ECCV-SYNTHETIC-GAUSSIAN11664-17944_optimize_300samples__method1errorFun1_std0.05_shapePen0' # resultDir3 = 'results/' + testPrefix3 + '/' # with open(resultDir3 + 'experiment.pickle', 'rb') as pfile: # experimentDic3 = pickle.load(pfile) # testSet3 = experimentDic3['testSet'] # methodsPred3 = experimentDic3['methodsPred'] # testOcclusions3 = experimentDic3[ 'testOcclusions'] # testPrefixBase3 = experimentDic3[ 'testPrefixBase'] # parameterRecognitionModels3 = experimentDic3[ 'parameterRecognitionModels'] # azimuths3 = experimentDic3['azimuths'] # elevations3 = experimentDic3[ 'elevations'] # vColors3 = experimentDic3[ 'vColors'] # lightCoeffs3 = experimentDic3[ 'lightCoeffs'] # likelihoods3 = experimentDic3['likelihoods'] # shapeParams3 = experimentDic3['shapeParams'] # if os.path.isfile(resultDir3 + 'segmentations.pickle'): # with open(resultDir3 + 'segmentations.pickle', 'rb') as pfile: # segmentationsDic3 = pickle.load(pfile) # segmentations3 = segmentationsDic3['segmentations'] # else: # segmentations3 = [None]*len(methodsPred3) # # testOcclusionsFull3 = testOcclusions3.copy() # # # with open(resultDir3 + 'approxProjections.pickle', 'rb') as pfile: # # approxProjectionsDic3 = pickle.load(pfile) # # # # approxProjections3 = approxProjectionsDic3['approxProjections'] # # approxProjectionsGT3 = approxProjectionsDic3['approxProjectionsGT'] # # approxProjections3 = None # approxProjectionsGT3 = None # # ipdb.set_trace() # # range1 = np.arange(len(azimuths2[3])) # range2 = np.arange(-1000 + len(azimuths2[3]) + len(azimuths3[3]),len(azimuths3[3])) # # testSet, parameterRecognitionModels, testPrefixBase, methodsPred, testOcclusions, azimuths, elevations, vColors, lightCoeffs, shapeParams, likelihoods, segmentations, approxProjections, approxProjectionsGT = \ # joinExperiments(range1, range2, testSet2,methodsPred2,testOcclusions2,testPrefixBase2,parameterRecognitionModels2,azimuths2,elevations2,vColors2,lightCoeffs2,likelihoods2,shapeParams2,segmentations2, approxProjections2, approxProjectionsGT3, testSet3,methodsPred3,testOcclusions3,testPrefixBase3,parameterRecognitionModels3,azimuths3,elevations3,vColors3,lightCoeffs3,likelihoods3,shapeParams3,segmentations3, approxProjections3, approxProjectionsGT3) # # testSet = testSet2 # testOcclusions = testOcclusions2 # testPrefix = 'train4_occlusion_shapemodel_10k_ECCV-SYNTHETIC-GAUSSIAN2530-17944_optimize_1000samples__method1errorFun1_std0.05_shapePen0-FINAL' # resultDir = 'results/' + testPrefix + '/' # # experimentDic = {'testSet':testSet, 'methodsPred':methodsPred, 'testOcclusions':testOcclusions, 'likelihoods':likelihoods, 'testPrefixBase':testPrefixBase, 'parameterRecognitionModels':parameterRecognitionModels, 'azimuths':azimuths, 'elevations':elevations, 'vColors':vColors, 'lightCoeffs':lightCoeffs, 'shapeParams':shapeParams} # # with open(resultDir + 'experiment.pickle', 'wb') as pfile: # pickle.dump(experimentDic, pfile) # # # # # # methodsPred = methodsPred + ['Gaussian (OpenGL)'] # azimuths = azimuths + [azimuths2[3]] # elevations = elevations + [elevations2[3]] # vColors = vColors + [vColors2[3]] # lightCoeffs = lightCoeffs + [lightCoeffs2[3]] # likelihoods = likelihoods + [likelihoods2[3]] # shapeParams = shapeParams + [shapeParams2[3]] # segmentations = segmentations + [segmentations2[3]] # # methodsPred[0] = methodsPred2[0] # azimuths[0] = azimuths2[0] # elevations[0] = elevations2[0] # vColors[0] = vColors2[0] # lightCoeffs[0] = lightCoeffs2[0] # likelihoods[0] = likelihoods2[0] # shapeParams[0] = shapeParams2[0] # segmentations[0] = segmentations2[0] rendererGT = None chPointLightIntensityGT = ch.Ch([1]) light_colorGT = ch.ones(3)*chPointLightIntensityGT chDisplacementGT = ch.Ch([0.0,0.0,0.0]) chScaleGT = ch.Ch([1, 1.,1.]) replaceableScenesFile = '../databaseFull/fields/scene_replaceables_backup.txt' # likelihoods = [np.array([]), np.array([])] from OpenGL import contextdata # # # segmentations[2] = np.zeros([len(testSet), 150,150]) # segmentations[3] = np.zeros([len(testSet), 150,150]) # segmentations[4] = np.zeros([len(testSet), 150,150]) # segmentations[5] = np.zeros([len(testSet), 150,150]) # # # for test_i in range(len(testSet)): # # # sceneNumber = dataScenes[test_i] # # sceneIdx = scene_io_utils.getSceneIdx(sceneNumber, replaceableScenesFile) # # sceneNumber, sceneFileName, instances, roomName, roomInstanceNum, targetIndicesScene, targetPositions = scene_io_utils.getSceneInformation(sceneIdx, replaceableScenesFile) # # # sceneNumber, sceneFileName, instances, roomName, roomInstanceNum, targetIndices, targetPositions = scene_io_utils.getSceneInformation(sceneIdx, replaceableScenesFile) # # # # targetIndex = dataTargetIndices[test_i] # # sceneDicFile = 'data/scene' + str(sceneNumber) + '.pickle' # # v, f_list, vc, vn, uv, haveTextures_list, textures_list = scene_io_utils.loadSavedScene(sceneDicFile, True) # # # # removeObjectData(len(v) -1 - targetIndex, v, f_list, vc, vn, uv, haveTextures_list, textures_list) # # # # addObjectData(v, f_list, vc, vn, uv, haveTextures_list, textures_list, smVerticesGT, smFacesGT, smVColorsGT, smNormalsGT, smUVsGT, smHaveTexturesGT, smTexturesListGT) # # # # if rendererGT is not None: # # rendererGT.makeCurrentContext() # # rendererGT.clear() # # contextdata.cleanupContext(contextdata.getContext()) # # if glMode == 'glfw': # # glfw.destroy_window(rendererGT.win) # # del rendererGT # # # # targetPosition = targetPositions[np.where(targetIndex==np.array(targetIndicesScene))[0]] # # # # rendererGT = createRendererGT(glMode, chAzGT, chObjAzGT, chElGT, chDistGT, center, v, vc, f_list, vn, light_colorGT, chComponentGT, chVColorsGT, targetPosition.copy(), chDisplacementGT, chScaleGT, width,height, uv, haveTextures_list, textures_list, frustum, None ) # # # # for hdrFile, hdrValues in hdritems: # # hdridx = hdrValues[0] # # envMapCoeffs = hdrValues[1] # # if hdridx == dataEnvMaps[test_i]: # # break # # envMapFilename = hdrFile # # # # phiOffset[:] = dataEnvMapPhiOffsets[test_i] # # chObjAzGT[:] = testObjAzsGT[test_i] # # chAzGT[:] = testAzsGT[test_i] # # chElGT[:] = testElevsGT[test_i] # # chVColorsGT[:] = testVColorGT[test_i] # # envMapCoeffsRotated[:] = np.dot(light_probes.chSphericalHarmonicsZRotation(totalOffset), envMapCoeffs[[0,3,2,1,4,5,6,7,8]])[[0,3,2,1,4,5,6,7,8]] # # envMapCoeffsRotatedRel[:] = np.dot(light_probes.chSphericalHarmonicsZRotation(phiOffset), envMapCoeffs[[0,3,2,1,4,5,6,7,8]])[[0,3,2,1,4,5,6,7,8]] # # chShapeParamsGT[:] = dataShapeModelCoeffsGT[test_i] # # im = images[test_i] # rendererGT = srgb2lin(im.copy()) # # chLightSHCoeffs[:] = testLightCoefficientsGTRel[test_i] # chObjAz[:] = 0 # chAz[:] = testAzsRel[test_i] # chEl[:] = testElevsGT[test_i] # chVColors[:] = testVColorGT[test_i] # chShapeParams[:] = testShapeParamsGT[test_i] # # stds[:] = 0.05 # # negLikModelRobust = -ch.sum(generative_models.LogRobustModel(renderer=renderer, groundtruth=rendererGT, foregroundPrior=globalPrior, variances=variances))/numPixels # # # likelihoods[0] = np.append(likelihoods[0], negLikModelRobust.r) # # chLightSHCoeffs[:] = lightCoeffs[3][idsInRange[test_i]] # chObjAz[:] = 0 # chAz[:] = azimuths[3][idsInRange[test_i]] # chEl[:] = elevations[3][idsInRange[test_i]] # chVColors[:] = vColors[3][idsInRange[test_i]] # chShapeParams[:] = shapeParams[3][idsInRange[test_i]] # # # likelihoods[1] = np.append(likelihoods[1], negLikModelRobust.r) # # #masksGT # #render[~mask*vis_im] = np.concatenate([np.ones([1000,1000])[:,:,None], np.zeros([1000,1000])[:,:,None],np.zeros([1000,1000])[:,:,None]], axis=2)[~mask*vis_im] # # vis_im = np.array(renderer.indices_image==1).copy().astype(np.bool) # post = generative_models.layerPosteriorsRobustCh(rendererGT, renderer, vis_im, 'MASK', globalPrior, variances)[0].r>0.5 # render = ~post.copy() # mask = masksGT[test_i] # render[~vis_im] = 1 # # segmentations[3][test_i] = post # # # chLightSHCoeffs[:] = lightCoeffs[2][idsInRange[test_i]] # chObjAz[:] = 0 # chAz[:] = azimuths[2][idsInRange[test_i]] # chEl[:] = elevations[2][idsInRange[test_i]] # chVColors[:] = vColors[2][idsInRange[test_i]] # chShapeParams[:] = shapeParams[2][idsInRange[test_i]] # # # likelihoods[2] = np.append(likelihoods[2], negLikModelRobust.r) # # # masksGT # # render[~mask*vis_im] = np.concatenate([np.ones([1000,1000])[:,:,None], np.zeros([1000,1000])[:,:,None],np.zeros([1000,1000])[:,:,None]], axis=2)[~mask*vis_im] # # post = np.array(renderer.indices_image == 1).copy().astype(np.bool) # # segmentations[2][test_i] = post # # # chLightSHCoeffs[:] = lightCoeffs[5][idsInRange[test_i]] # chObjAz[:] = 0 # chAz[:] = azimuths[5][idsInRange[test_i]] # chEl[:] = elevations[5][idsInRange[test_i]] # chVColors[:] = vColors[5][idsInRange[test_i]] # chShapeParams[:] = shapeParams[5][idsInRange[test_i]] # # # likelihoods[1] = np.append(likelihoods[1], negLikModelRobust.r) # # # masksGT # # render[~mask*vis_im] = np.concatenate([np.ones([1000,1000])[:,:,None], np.zeros([1000,1000])[:,:,None],np.zeros([1000,1000])[:,:,None]], axis=2)[~mask*vis_im] # # post = np.array(renderer.indices_image == 1).copy().astype(np.bool) # # segmentations[5][test_i] = post # # # # chLightSHCoeffs[:] = lightCoeffs[4][idsInRange[test_i]] # # chObjAz[:] = 0 # # chAz[:] = azimuths[4][idsInRange[test_i]] # # chEl[:] = elevations[4][idsInRange[test_i]] # # chVColors[:] = vColors[4][idsInRange[test_i]] # # chShapeParams[:] = shapeParams[4][idsInRange[test_i]] # # # likelihoods[1] = np.append(likelihoods[1], negLikModelRobust.r) # # # masksGT # # render[~mask*vis_im] = np.concatenate([np.ones([1000,1000])[:,:,None], np.zeros([1000,1000])[:,:,None],np.zeros([1000,1000])[:,:,None]], axis=2)[~mask*vis_im] # # # vis_im = np.array(renderer.indices_image == 1).copy().astype(np.bool) # # post = generative_models.layerPosteriorsRobustCh(rendererGT, renderer, vis_im, 'MASK', globalPrior, variances)[0].r > 0.5 # # render = ~post.copy() # # mask = masksGT[test_i] # # render[~vis_im] = 1 # # # # segmentations[4][test_i] = post # # # renderRGB = np.concatenate([render[:,:,None], render[:,:,None], render[:,:,None]], axis=2) # # # cv2.imwrite('tmp/SH/renderergt' + str(test_i) + '.jpeg' , 255*lin2srgb(rendererGT[:,:,[2,1,0]]), [int(cv2.IMWRITE_JPEG_QUALITY), 100]) # # plt.imsave('tmp/SH/mask' + str(test_i) + '.jpeg', mask) # # cv2.imwrite('tmp/SH/post' + str(test_i) + '.jpeg' , 255*post, [int(cv2.IMWRITE_JPEG_QUALITY), 100]) # # cv2.imwrite('tmp/SH/post' + str(test_i) + '.jpeg' , 255*render, [int(cv2.IMWRITE_JPEG_QUALITY), 100]) # # plt.imsave('tmp/SH/renderer' + str(test_i) + '.png' , lin2srgb(renderer.r.copy())) # # ipdb.set_trace() # # # # # # methodsPred[3] = 'Robust Fit' # # methodsPred[4] = 'Robust (OpenGL)' # # methodsPred[5] = 'Recognition (OpenGL)' # # methodsPred[6] = 'Gaussian Fit' # # segmentations[4] = segmentations3[3] # # errorsPosePredList, errorsLightCoeffsList, errorsShapeParamsList, errorsShapeVerticesList, errorsEnvMapList, errorsLightCoeffsCList, errorsVColorsEList, errorsVColorsCList, errorsVColorsSList, errorsSegmentationList \ # = computeErrors(np.arange(len(rangeTests)), azimuths, testAzsRel, elevations, testElevsGT, vColors, testVColorGT, lightCoeffs, testLightCoefficientsGTRel, approxProjections, approxProjectionsGT, shapeParams, testShapeParamsGT, useShapeModel, chShapeParams, chVertices, segmentations, masksGT) # envMapTexture = np.zeros([180,360,3]) # approxProjectionsFittedList = [] # for test_i in range(len(testSet)): # pEnvMap = SHProjection(envMapTexture, np.concatenate([lightCoeffs[4][test_i][:,None], lightCoeffs[4][test_i][:,None], lightCoeffs[4][test_i][:,None]], axis=1)) # approxProjection = np.sum(pEnvMap, axis=(2,3)) # approxProjectionsFittedList = approxProjectionsFittedList + [approxProjection[None,:]] # approxProjectionsFitted = np.vstack(approxProjectionsFittedList) # # envMapTexture = np.zeros([180,360,3]) # approxProjectionsGTList = [] # for test_i in range(len(testSet)): # pEnvMap = SHProjection(envMapTexture, np.concatenate([testLightCoefficientsGTRel[test_i][:,None], testLightCoefficientsGTRel[test_i][:,None], testLightCoefficientsGTRel[test_i][:,None]], axis=1)) # approxProjectionGT = np.sum(pEnvMap, axis=(2,3)) # approxProjectionsGTList = approxProjectionsGTList + [approxProjectionGT[None,:]] # approxProjectionsGT = np.vstack(approxProjectionsGTList) # with open(resultDir + 'experiment_errors.pickle', 'rb') as pfile: # experimentErrorsDic = pickle.load(pfile) # # errorsPosePredList = experimentErrorsDic['errorsPosePredList'] # errorsLightCoeffsList = experimentErrorsDic['errorsLightCoeffsList'] # errorsShapeParamsList = experimentErrorsDic['errorsShapeParamsLis'] # errorsShapeVerticesList = experimentErrorsDic['errorsShapeVerticesList'] # errorsEnvMapList = experimentErrorsDic['errorsEnvMapList'] # errorsLightCoeffsCList = experimentErrorsDic['errorsLightCoeffsCList'] # errorsVColorsEList = experimentErrorsDic['errorsVColorsEList'] # errorsVColorsCList = experimentErrorsDic['errorsVColorsCList'] # errorsVColorsSList = experimentErrorsDic['errorsVColorsSList'] # errorsSegmentationList = experimentErrorsDic['errorsSegmentationList'] testPrefix = 'train4_occlusion_shapemodel_10k_ECCV-PHOTOREALISTIC-JOINT2018' resultDir = 'results/' + testPrefix + '/' # experimentDic = {'testSet':testSet, 'methodsPred':methodsPred, 'testOcclusions':testOcclusions, 'likelihoods':likelihoods, 'testPrefixBase':testPrefixBase, 'parameterRecognitionModels':parameterRecognitionModels, 'azimuths':azimuths, 'elevations':elevations, 'vColors':vColors, 'lightCoeffs':lightCoeffs, 'shapeParams':shapeParams} # # with open(resultDir + 'experiment.pickle', 'wb') as pfile: # pickle.dump(experimentDic, pfile) # # # experimentErrorsDic = {'errorsPosePredList':errorsPosePredList, 'errorsLightCoeffsList':errorsLightCoeffsList, 'errorsShapeParamsLis':errorsShapeParamsList, 'errorsShapeVerticesList':errorsShapeVerticesList, 'errorsEnvMapList':errorsEnvMapList, 'errorsLightCoeffsCList':errorsLightCoeffsCList, 'errorsVColorsEList':errorsVColorsEList, 'errorsVColorsCList':errorsVColorsCList, 'errorsVColorsSList':errorsVColorsSList,'errorsSegmentationList':errorsSegmentationList} # # # with open(resultDir + 'experiment_errors.pickle', 'wb') as pfile: # pickle.dump(experimentErrorsDic, pfile) with open(resultDir + 'experiment_errors.pickle', 'rb') as pfile: experimentErrorsDic = pickle.load(pfile) errorsPosePredList = experimentErrorsDic['errorsPosePredList'] errorsLightCoeffsList = experimentErrorsDic['errorsLightCoeffsList'] errorsShapeParamsList = experimentErrorsDic['errorsShapeParamsLis'] errorsShapeVerticesList = experimentErrorsDic['errorsShapeVerticesList'] errorsEnvMapList = experimentErrorsDic['errorsEnvMapList'] errorsLightCoeffsCList = experimentErrorsDic['errorsLightCoeffsCList'] errorsVColorsEList = experimentErrorsDic['errorsVColorsEList'] errorsVColorsCList = experimentErrorsDic['errorsVColorsCList'] errorsVColorsSList = experimentErrorsDic['errorsVColorsSList'] errorsSegmentationList = experimentErrorsDic['errorsSegmentationList'] meanAbsErrAzsList, meanAbsErrElevsList, meanErrorsLightCoeffsList, meanErrorsShapeParamsList, meanErrorsShapeVerticesList, meanErrorsLightCoeffsCList, meanErrorsEnvMapList, meanErrorsVColorsEList, meanErrorsVColorsCList, meanErrorsVColorsCList, meanErrorsSegmentation \ = computeErrorAverages(np.mean, np.arange(len(rangeTests)), useShapeModel, errorsPosePredList, errorsLightCoeffsList, errorsShapeParamsList, errorsShapeVerticesList, errorsEnvMapList, errorsLightCoeffsCList, errorsVColorsEList, errorsVColorsCList,errorsVColorsSList, errorsSegmentationList) nearestNeighbours = False if 'Nearest Neighbours' in set(methodsPred): nearestNeighbours = True plotColors = ['k'] if nearestNeighbours: # methodsPred = methodsPred + ["Nearest Neighbours"] plotColors = plotColors + ['m'] plotColors = plotColors + ['b'] plotColors = plotColors + ['r'] plotColors = plotColors + ['r'] plotColors = plotColors + ['b'] plotColors = plotColors + ['g'] plotColors = plotColors + ['g'] plotStyles = ['solid'] if nearestNeighbours: # methodsPred = methodsPred + ["Nearest Neighbours"] plotStyles = plotStyles + ['solid'] plotStyles = plotStyles + ['solid'] plotStyles = plotStyles + ['solid'] plotStyles = plotStyles + ['dashed'] plotStyles = plotStyles + ['dashed'] plotStyles = plotStyles + ['solid'] plotStyles = plotStyles + ['dashed'] plotMethodsIndices = [2,5,6, 7, 3,4] # plotMethodsIndices = [0,2,6,3] import ipdb; ipdb.set_trace() recognitionIdx = 2 robustIdx = 3 errorsAzimuthList = [errorsPosePredList[i][0] if errorsPosePredList[i] is not None else None for i in range(len(methodsPred))] errorsElevationList = [errorsPosePredList[i][1] if errorsPosePredList[i] is not None else None for i in range(len(methodsPred))] variablesDescr = ['Azimuth', 'Elevation', 'VColor', 'Illumination', 'Shape', 'Segmentation'] errorsList = [errorsAzimuthList,errorsElevationList, errorsVColorsCList, errorsLightCoeffsCList, errorsShapeVerticesList, errorsSegmentationList] saveConditionalHistograms(resultDir, testOcclusions, methodsPred, variablesDescr, plotMethodsIndices, errorsList) print("Printing occlusin-likelihood plots!") # meanLikelihoodArr = [np.array([]), np.array([]), np.array([]), np.array([])] # occlusions = [] # for occlusionLevel in range(100): # # setUnderOcclusionLevel = testOcclusionsFull * 100 < occlusionLevel # # if np.any(setUnderOcclusionLevel): # occlusions = occlusions + [occlusionLevel] # testOcclusions = testOcclusionsFull[setUnderOcclusionLevel] # # if likelihoods[0] is not None: # meanLikelihoodArr[0] = np.append(meanLikelihoodArr[0], np.mean(likelihoods[0][setUnderOcclusionLevel])) # if likelihoods[1] is not None: # meanLikelihoodArr[1] = np.append(meanLikelihoodArr[1], np.mean(likelihoods[1][setUnderOcclusionLevel])) # if likelihoods[2] is not None: # meanLikelihoodArr[2] = np.append(meanLikelihoodArr[2], np.mean(likelihoods[2][setUnderOcclusionLevel])) # if likelihoods[3] is not None: # meanLikelihoodArr[3] = np.append(meanLikelihoodArr[3], np.mean(likelihoods[3][setUnderOcclusionLevel])) # saveLikelihoodPlots(resultDir, occlusions, methodsPred, plotColors, plotMethodsIndices, meanLikelihoodArr) # # print("Computing means!") # meanAbsErrAzsArr = [] meanAbsErrElevsArr = [] meanErrorsLightCoeffsArr = [] meanErrorsEnvMapArr = [] meanErrorsShapeParamsArr = [] meanErrorsShapeVerticesArr = [] meanErrorsLightCoeffsCArr = [] meanErrorsVColorsEArr = [] meanErrorsVColorsCArr = [] meanErrorsVColorsSArr = [] meanErrorsSegmentationArr = [] for method_i in range(len(methodsPred)): meanAbsErrAzsArr = meanAbsErrAzsArr + [np.array([])] meanAbsErrElevsArr = meanAbsErrElevsArr + [np.array([])] meanErrorsLightCoeffsArr = meanErrorsLightCoeffsArr + [np.array([])] meanErrorsShapeParamsArr = meanErrorsShapeParamsArr + [np.array([])] meanErrorsShapeVerticesArr = meanErrorsShapeVerticesArr + [np.array([])] meanErrorsLightCoeffsCArr = meanErrorsLightCoeffsCArr + [np.array([])] meanErrorsVColorsEArr = meanErrorsVColorsEArr + [np.array([])] meanErrorsVColorsCArr = meanErrorsVColorsCArr + [np.array([])] meanErrorsVColorsSArr = meanErrorsVColorsSArr + [np.array([])] meanErrorsEnvMapArr = meanErrorsEnvMapArr + [np.array([])] meanErrorsSegmentationArr = meanErrorsSegmentationArr + [np.array([])] occlusions = [] print("Printing occlusin-error plots!") for occlusionLevel in range(100): setUnderOcclusionLevel = testOcclusionsFull * 100 < occlusionLevel if np.any(setUnderOcclusionLevel): occlusions = occlusions + [occlusionLevel] testOcclusions = testOcclusionsFull[setUnderOcclusionLevel] colors = matplotlib.cm.plasma(testOcclusions) for method_i in range(len(methodsPred)): if errorsPosePredList[method_i] is not None: meanAbsErrAzsArr[method_i] = np.append(meanAbsErrAzsArr[method_i], np.mean(np.abs(errorsPosePredList[method_i][0][setUnderOcclusionLevel]))) meanAbsErrElevsArr[method_i] = np.append(meanAbsErrElevsArr[method_i], np.mean(np.abs(errorsPosePredList[method_i][1][setUnderOcclusionLevel]))) if errorsLightCoeffsList[method_i] is not None: meanErrorsLightCoeffsArr[method_i] = np.append(meanErrorsLightCoeffsArr[method_i],np.mean(np.mean(errorsLightCoeffsList[method_i][setUnderOcclusionLevel], axis=1), axis=0)) if errorsLightCoeffsCList[method_i] is not None: meanErrorsLightCoeffsCArr[method_i] = np.append(meanErrorsLightCoeffsCArr[method_i],np.mean(np.mean(errorsLightCoeffsCList[method_i][setUnderOcclusionLevel], axis=1), axis=0)) if useShapeModel: if errorsShapeParamsList[method_i] is not None: meanErrorsShapeParamsArr[method_i] = np.append(meanErrorsShapeParamsArr[method_i],np.mean(np.mean(errorsShapeParamsList[method_i][setUnderOcclusionLevel], axis=1), axis=0)) if errorsShapeVerticesList[method_i] is not None: meanErrorsShapeVerticesArr[method_i] = np.append(meanErrorsShapeVerticesArr[method_i], np.mean(errorsShapeVerticesList[method_i][setUnderOcclusionLevel], axis=0)) if errorsEnvMapList[method_i] is not None: meanErrorsEnvMapArr[method_i] = np.append(meanErrorsEnvMapArr[method_i], np.mean(errorsEnvMapList[method_i][setUnderOcclusionLevel])) if errorsVColorsEList[method_i] is not None: meanErrorsVColorsEArr[method_i] = np.append(meanErrorsVColorsEArr[method_i], np.mean(errorsVColorsEList[method_i][setUnderOcclusionLevel], axis=0)) if errorsVColorsCList[method_i] is not None: meanErrorsVColorsCArr[method_i] = np.append(meanErrorsVColorsCArr[method_i], np.mean(errorsVColorsCList[method_i][setUnderOcclusionLevel], axis=0)) if errorsVColorsSList[method_i] is not None: meanErrorsVColorsSArr[method_i] = np.append(meanErrorsVColorsSArr[method_i], np.mean(errorsVColorsSList[method_i][setUnderOcclusionLevel], axis=0)) if errorsSegmentationList[method_i] is not None: meanErrorsSegmentationArr[method_i] = np.append(meanErrorsSegmentationArr[method_i], np.mean(errorsSegmentationList[method_i][setUnderOcclusionLevel], axis=0)) else: meanErrorsSegmentationArr[method_i] = None print("Printing occlusin-error plots - median!") saveOcclusionPlots(resultDir, 'mean',occlusions, methodsPred, plotColors, plotStyles, plotMethodsIndices, useShapeModel, meanAbsErrAzsArr, meanAbsErrElevsArr, meanErrorsVColorsCArr, meanErrorsVColorsEArr, meanErrorsVColorsSArr, meanErrorsLightCoeffsArr, meanErrorsShapeParamsArr, meanErrorsShapeVerticesArr, meanErrorsLightCoeffsCArr, meanErrorsEnvMapArr, meanErrorsSegmentationArr) medianAbsErrAzsArr = [] medianAbsErrElevsArr = [] medianErrorsLightCoeffsArr = [] medianErrorsEnvMapArr = [] medianErrorsShapeParamsArr = [] medianErrorsShapeVerticesArr = [] medianErrorsLightCoeffsCArr = [] medianErrorsVColorsEArr = [] medianErrorsVColorsCArr = [] medianErrorsVColorsSArr = [] medianErrorsSegmentationArr = [] for method_i in range(len(methodsPred)): medianAbsErrAzsArr = medianAbsErrAzsArr + [np.array([])] medianAbsErrElevsArr = medianAbsErrElevsArr + [np.array([])] medianErrorsLightCoeffsArr = medianErrorsLightCoeffsArr + [np.array([])] medianErrorsShapeParamsArr = medianErrorsShapeParamsArr + [np.array([])] medianErrorsShapeVerticesArr = medianErrorsShapeVerticesArr + [np.array([])] medianErrorsLightCoeffsCArr = medianErrorsLightCoeffsCArr + [np.array([])] medianErrorsVColorsEArr = medianErrorsVColorsEArr + [np.array([])] medianErrorsVColorsCArr = medianErrorsVColorsCArr + [np.array([])] medianErrorsVColorsSArr = medianErrorsVColorsSArr + [np.array([])] medianErrorsEnvMapArr = medianErrorsEnvMapArr + [np.array([])] medianErrorsSegmentationArr = medianErrorsSegmentationArr + [np.array([])] occlusions = [] print("Printing occlusin-error plots!") for occlusionLevel in range(100): setUnderOcclusionLevel = testOcclusionsFull * 100 < occlusionLevel if np.any(setUnderOcclusionLevel): occlusions = occlusions + [occlusionLevel] testOcclusions = testOcclusionsFull[setUnderOcclusionLevel] colors = matplotlib.cm.plasma(testOcclusions) for method_i in range(len(methodsPred)): if errorsPosePredList[method_i] is not None: medianAbsErrAzsArr[method_i] = np.append(medianAbsErrAzsArr[method_i], np.median(np.abs(errorsPosePredList[method_i][0][setUnderOcclusionLevel]))) medianAbsErrElevsArr[method_i] = np.append(medianAbsErrElevsArr[method_i], np.median(np.abs(errorsPosePredList[method_i][1][setUnderOcclusionLevel]))) if errorsLightCoeffsList[method_i] is not None: medianErrorsLightCoeffsArr[method_i] = np.append(medianErrorsLightCoeffsArr[method_i],np.median(np.mean(errorsLightCoeffsList[method_i][setUnderOcclusionLevel], axis=1), axis=0)) if errorsLightCoeffsCList[method_i] is not None: medianErrorsLightCoeffsCArr[method_i] = np.append(medianErrorsLightCoeffsCArr[method_i],np.median(np.mean(errorsLightCoeffsCList[method_i][setUnderOcclusionLevel], axis=1), axis=0)) if useShapeModel: if errorsShapeParamsList[method_i] is not None: medianErrorsShapeParamsArr[method_i] = np.append(medianErrorsShapeParamsArr[method_i],np.median(np.median(errorsShapeParamsList[method_i][setUnderOcclusionLevel], axis=1), axis=0)) if errorsShapeVerticesList[method_i] is not None: medianErrorsShapeVerticesArr[method_i] = np.append(medianErrorsShapeVerticesArr[method_i], np.median(errorsShapeVerticesList[method_i][setUnderOcclusionLevel], axis=0)) if errorsEnvMapList[method_i] is not None: medianErrorsEnvMapArr[method_i] = np.append(medianErrorsEnvMapArr[method_i], np.median(errorsEnvMapList[method_i][setUnderOcclusionLevel])) if errorsVColorsEList[method_i] is not None: medianErrorsVColorsEArr[method_i] = np.append(medianErrorsVColorsEArr[method_i], np.median(errorsVColorsEList[method_i][setUnderOcclusionLevel], axis=0)) if errorsVColorsCList[method_i] is not None: medianErrorsVColorsCArr[method_i] = np.append(medianErrorsVColorsCArr[method_i], np.median(errorsVColorsCList[method_i][setUnderOcclusionLevel], axis=0)) if errorsVColorsSList[method_i] is not None: medianErrorsVColorsSArr[method_i] = np.append(medianErrorsVColorsSArr[method_i], np.median(errorsVColorsSList[method_i][setUnderOcclusionLevel], axis=0)) if errorsSegmentationList[method_i] is not None: medianErrorsSegmentationArr[method_i] = np.append(medianErrorsSegmentationArr[method_i], np.median(errorsSegmentationList[method_i][setUnderOcclusionLevel], axis=0)) else: medianErrorsSegmentationArr[method_i] = None saveOcclusionPlots(resultDir, 'median', occlusions,methodsPred, plotColors, plotStyles, plotMethodsIndices, useShapeModel, medianAbsErrAzsArr, medianAbsErrElevsArr, medianErrorsVColorsCArr, medianErrorsVColorsEArr, medianErrorsVColorsSArr, medianErrorsLightCoeffsArr, medianErrorsShapeParamsArr, medianErrorsShapeVerticesArr, medianErrorsLightCoeffsCArr, medianErrorsEnvMapArr, medianErrorsSegmentationArr) SHModel = "" for occlusionLevel in [25,75,100]: resultDirOcclusion = 'results/' + testPrefix + '/occlusion' + str(occlusionLevel) + '/' if not os.path.exists(resultDirOcclusion): os.makedirs(resultDirOcclusion) setUnderOcclusionLevel = testOcclusionsFull * 100 < occlusionLevel testOcclusions = testOcclusionsFull[setUnderOcclusionLevel] # # errorsPosePred = [errorsPosePredList[recognitionIdx][0][setUnderOcclusionLevel], errorsPosePredList[recognitionIdx][1][setUnderOcclusionLevel]] errorsLightCoeffs = errorsLightCoeffsList[recognitionIdx][setUnderOcclusionLevel] errorsShapeParams = errorsShapeParamsList[recognitionIdx][setUnderOcclusionLevel] errorsShapeVertices= errorsShapeVerticesList[recognitionIdx][setUnderOcclusionLevel] if errorsEnvMapList[recognitionIdx] is not None: errorsEnvMap= errorsEnvMapList[recognitionIdx][setUnderOcclusionLevel] else: errorsEnvMap = None errorsLightCoeffsC= errorsLightCoeffsCList[recognitionIdx][setUnderOcclusionLevel] errorsVColorsE= errorsVColorsEList[recognitionIdx][setUnderOcclusionLevel] errorsVColorsC= errorsVColorsCList[recognitionIdx][setUnderOcclusionLevel] errorsVColorsS= errorsVColorsSList[recognitionIdx][setUnderOcclusionLevel] # errorsSegmentation = errorsSegmentationList[recognitionIdx][setUnderOcclusionLevel] errorsSegmentation = None errorsPoseFitted = [errorsPosePredList[robustIdx][0][setUnderOcclusionLevel], errorsPosePredList[robustIdx][1][setUnderOcclusionLevel]] errorsFittedLightCoeffs = errorsLightCoeffsList[robustIdx][setUnderOcclusionLevel] errorsFittedShapeParams = errorsShapeParamsList[robustIdx][setUnderOcclusionLevel] errorsFittedShapeVertices= errorsShapeVerticesList[robustIdx][setUnderOcclusionLevel] if errorsEnvMapList[recognitionIdx] is not None: errorsFittedEnvMap = errorsEnvMapList[robustIdx][setUnderOcclusionLevel] else: errorsFittedEnvMap = None errorsFittedLightCoeffsC= errorsLightCoeffsCList[robustIdx][setUnderOcclusionLevel] errorsFittedVColorsE = errorsVColorsEList[robustIdx][setUnderOcclusionLevel] errorsFittedVColorsC = errorsVColorsCList[robustIdx][setUnderOcclusionLevel] errorsFittedVColorsS = errorsVColorsSList[robustIdx][setUnderOcclusionLevel] # errorsFittedSegmentation = errorsSegmentationList[recognitionIdx][setUnderOcclusionLevel] errorsFittedSegmentation = None saveScatterPlots(resultDirOcclusion, testOcclusions, useShapeModel, errorsPosePred, errorsPoseFitted,errorsLightCoeffsC,errorsFittedLightCoeffsC,errorsEnvMap,errorsFittedEnvMap,errorsLightCoeffs,errorsFittedLightCoeffs,errorsShapeParams,errorsFittedShapeParams,errorsShapeVertices,errorsFittedShapeVertices,errorsVColorsE,errorsFittedVColorsE,errorsVColorsC,errorsFittedVColorsC, errorsVColorsS,errorsFittedVColorsS) errorsPoseFitted = [errorsPosePredList[5][0][setUnderOcclusionLevel], errorsPosePredList[5][1][setUnderOcclusionLevel]] errorsFittedLightCoeffs = errorsLightCoeffsList[5][setUnderOcclusionLevel] errorsFittedShapeParams = errorsShapeParamsList[5][setUnderOcclusionLevel] errorsFittedShapeVertices= errorsShapeVerticesList[5][setUnderOcclusionLevel] if errorsEnvMapList[recognitionIdx] is not None: errorsFittedEnvMap = errorsEnvMapList[5][setUnderOcclusionLevel] else: errorsFittedEnvMap = None errorsFittedLightCoeffsC= errorsLightCoeffsCList[5][setUnderOcclusionLevel] errorsFittedVColorsE= errorsVColorsEList[5][setUnderOcclusionLevel] errorsFittedVColorsC= errorsVColorsCList[5][setUnderOcclusionLevel] errorsFittedVColorsS= errorsVColorsSList[5][setUnderOcclusionLevel] errorsFittedSegmentation = errorsSegmentationList[5][setUnderOcclusionLevel] saveScatterPlotsMethodFit(4, resultDirOcclusion, testOcclusions, useShapeModel, errorsPosePred, errorsPoseFitted,errorsLightCoeffsC,errorsFittedLightCoeffsC,errorsEnvMap,errorsFittedEnvMap,errorsLightCoeffs,errorsFittedLightCoeffs,errorsShapeParams,errorsFittedShapeParams,errorsShapeVertices,errorsFittedShapeVertices,errorsVColorsE,errorsFittedVColorsE,errorsVColorsC,errorsFittedVColorsC, errorsVColorsS,errorsFittedVColorsS) # saveLikelihoodScatter(resultDirOcclusion, setUnderOcclusionLevel, testOcclusions, likelihoods) # if len(stdevsFull) > 0: # stdevs = stdevsFull[setUnderOcclusionLevel] colors = matplotlib.cm.plasma(testOcclusions) meanAbsErrAzsList, meanAbsErrElevsList, meanErrorsLightCoeffsList, meanErrorsShapeParamsList, meanErrorsShapeVerticesList, meanErrorsLightCoeffsCList, meanErrorsEnvMapList, meanErrorsVColorsEList, meanErrorsVColorsCList, meanErrorsVColorsSList, meanErrorsSegmentationList \ = computeErrorAverages(np.mean, setUnderOcclusionLevel, useShapeModel, errorsPosePredList, errorsLightCoeffsList, errorsShapeParamsList, errorsShapeVerticesList, errorsEnvMapList, errorsLightCoeffsCList, errorsVColorsEList, errorsVColorsCList, errorsVColorsSList, errorsSegmentationList) medianAbsErrAzsList, medianAbsErrElevsList, medianErrorsLightCoeffsList, medianErrorsShapeParamsList, medianErrorsShapeVerticesList, medianErrorsLightCoeffsCList, medianErrorsEnvMapList, medianErrorsVColorsEList, medianErrorsVColorsCList, medianErrorsVColorsSList, medianErrorsSegmentationList \ = computeErrorAverages(np.median, setUnderOcclusionLevel, useShapeModel, errorsPosePredList, errorsLightCoeffsList, errorsShapeParamsList, errorsShapeVerticesList, errorsEnvMapList, errorsLightCoeffsCList, errorsVColorsEList, errorsVColorsCList, errorsVColorsSList, errorsSegmentationList) # Write statistics to file. import tabulate headers = ["Errors"] + methodsPred table = [["Azimuth"] + meanAbsErrAzsList, ["Elevation"] + meanAbsErrElevsList, ["VColor C"] + meanErrorsVColorsCList, ["VColor S"] + meanErrorsVColorsSList, ["SH Light"] + meanErrorsLightCoeffsList, ["SH Light C"] + meanErrorsLightCoeffsCList, ["SH Env Map"] + meanErrorsEnvMapList, ["Shape Params"] + meanErrorsShapeParamsList, ["Shape Vertices"] + meanErrorsShapeVerticesList, ["Segmentation"] + meanErrorsSegmentationList ] performanceTable = tabulate.tabulate(table, headers=headers, tablefmt="latex", floatfmt=".3f") with open(resultDirOcclusion + 'performance.tex', 'w') as expfile: expfile.write(performanceTable) table = [["Azimuth"] + medianAbsErrAzsList, ["Elevation"] + medianAbsErrElevsList, ["VColor C"] + medianErrorsVColorsCList, ["VColor S"] + medianErrorsVColorsSList, ["SH Light"] + medianErrorsLightCoeffsList, ["SH Light C"] + medianErrorsLightCoeffsCList, ["SH Env Map"] + medianErrorsEnvMapList, ["Shape Params"] + medianErrorsShapeParamsList, ["Shape Vertices"] + medianErrorsShapeVerticesList, ["Segmentation"] + medianErrorsSegmentationList ] performanceTable = tabulate.tabulate(table, headers=headers, tablefmt="latex", floatfmt=".3f") with open(resultDirOcclusion + 'median-performance.tex', 'w') as expfile: expfile.write(performanceTable) headers = ["", "l=0", "SH $l=0,m=-1$", "SH $l=1,m=0$", "SH $l=1,m=1$", "SH $l=1,m=-2$", "SH $l=2,m=-1$", "SH $l=2,m=0$", "SH $l=2,m=1$", "SH $l=2,m=2$"] for method_i in plotMethodsIndices: SMSE_SH = np.mean(errorsLightCoeffsList[method_i][setUnderOcclusionLevel], axis=0) table = [ [SHModel, SMSE_SH[0], SMSE_SH[1], SMSE_SH[2], SMSE_SH[3], SMSE_SH[4], SMSE_SH[5], SMSE_SH[6], SMSE_SH[7], SMSE_SH[8]], ] performanceTable = tabulate.tabulate(table, headers=headers, tablefmt="latex", floatfmt=".3f") with open(resultDir + 'performance_SH_' + methodsPred[method_i] + '.tex', 'w') as expfile: expfile.write(performanceTable) # expfile.write(performanceTable) SMSE_SH = np.mean(errorsLightCoeffsCList[method_i][setUnderOcclusionLevel], axis=0) table = [ [SHModel, SMSE_SH[0], SMSE_SH[1], SMSE_SH[2], SMSE_SH[3], SMSE_SH[4], SMSE_SH[5], SMSE_SH[6], SMSE_SH[7], SMSE_SH[8]], ] performanceTable = tabulate.tabulate(table, headers=headers, tablefmt="latex", floatfmt=".3f") with open(resultDir + 'performance_SH_C_' + methodsPred[method_i] + '.tex', 'w') as expfile: expfile.write(performanceTable) if useShapeModel: SMSE_SHAPE_PARAMS = np.mean(errorsShapeParamsList[method_i][setUnderOcclusionLevel], axis=0) table = [[SHModel, SMSE_SHAPE_PARAMS[0], SMSE_SHAPE_PARAMS[1], SMSE_SHAPE_PARAMS[2], SMSE_SHAPE_PARAMS[3], SMSE_SHAPE_PARAMS[4], SMSE_SHAPE_PARAMS[5], SMSE_SHAPE_PARAMS[6], SMSE_SHAPE_PARAMS[7], SMSE_SHAPE_PARAMS[8], SMSE_SHAPE_PARAMS[9]]] performanceTable = tabulate.tabulate(table, headers=headers, tablefmt="latex", floatfmt=".3f") with open(resultDir + 'performance_ShapeParams_' + methodsPred[method_i] + '.tex', 'w') as expfile: expfile.write(performanceTable) plt.ion() print("Finished.")
58,258
45.127474
466
py
inversegraphics
inversegraphics-master/export_collisions.py
#!/usr/bin/env python3.4m import scene_io_utils import re from blender_utils import * from collision import * numpy.random.seed(1) inchToMeter = 0.0254 outputDir = 'data/' width = 150 height = 150 numSamples = 100 useCycles = False distance = 0.75 scene_io_utils.loadTargetsBlendData() sceneCollisions = {} replaceableScenesFile = '../databaseFull/fields/scene_replaceables_backup_new.txt' sceneLines = [line.strip() for line in open(replaceableScenesFile)] teapots = [line.strip() for line in open('teapots.txt')] renderTeapotsList = np.arange(len(teapots)) scaleZ = 0.265 scaleY = 0.18 scaleX = 0.35 cubeScale = mathutils.Matrix([[scaleX/2, 0,0,0], [0,scaleY/2,0 ,0] ,[0,0,scaleZ/2,0],[0,0,0,1]]) for sceneIdx in range(len(sceneLines))[:]: sceneLine = sceneLines[sceneIdx] sceneParts = sceneLine.split(' ') sceneFile = sceneParts[0] sceneNumber = int(re.search('.+?scene([0-9]+)\.txt', sceneFile, re.IGNORECASE).groups()[0]) scene_io_utils.loadTargetsBlendData() bpy.ops.wm.read_factory_settings() scene_io_utils.loadSceneBlendData(sceneIdx, replaceableScenesFile) scene = bpy.data.scenes['Main Scene'] scene.render.resolution_x = width #perhaps set resolution in code scene.render.resolution_y = height scene.render.tile_x = height/2 scene.render.tile_y = width/2 scene.cycles.samples = 10 sceneNumber, sceneFileName, instances, roomName, roomInstanceNum, targetIndices, targetPositions = scene_io_utils.getSceneInformation(sceneIdx, replaceableScenesFile) roomName = '' for model in scene.objects: reg = re.compile('(room[0-9]+)') res = reg.match(model.name) if res: roomName = res.groups()[0] cubeTarget = bpy.data.scenes['Scene'].objects['Cube'] try: bpy.data.scenes['Main Scene'].objects.unlink(bpy.data.scenes['Scene'].objects['Cube']) except: pass scene.objects.link(cubeTarget) targetCollisions = {} for targetIdx in range(len(targetIndices)): targetParentIndex = targetIndices[targetIdx] targetParentPosition = np.array(targetPositions[targetIdx]) scene.objects.unlink(scene.objects[str(targetParentIndex)]) director = outputDir cubeTarget.layers[1] = True # objAzimuths = numpy.arange(0,360, 5) # Map it to non colliding rotations. objAzimuths = numpy.array([]) intersections = [] translationMat = mathutils.Matrix.Translation(targetParentPosition) azimuthRot = mathutils.Matrix.Rotation(radians(0), 4, 'Z') cubeTarget.matrix_world = translationMat * azimuthRot * cubeScale * mathutils.Matrix.Translation(mathutils.Vector((0,0,1))) scene.render.filepath = 'scenes/' + str(sceneNumber) + '/collisionCubeExample_' + str(targetIdx) + '.png' placeCamera(scene.camera, -90, 25, 0.75, targetParentPosition) bpy.ops.render.render(write_still=True) collisionInterval = 5 for objAzimuth in numpy.arange(0,360,collisionInterval): azimuthRot = mathutils.Matrix.Rotation(radians(-objAzimuth), 4, 'Z') cubeTarget.matrix_world = translationMat * azimuthRot * cubeScale * mathutils.Matrix.Translation(mathutils.Vector((0,0,1))) # scene.render.filepath = 'scenes/' + str(sceneNumber) + '/' + str(objAzimuth) + 'collisionCubeExample_' + str(targetIdx) + '.png' # bpy.ops.render.render(write_still=True) intersect = False for sceneInstanceIdx, sceneInstance in enumerate(scene.objects): if sceneInstance.type == 'EMPTY' and sceneInstance != cubeTarget and sceneInstance.name != str(roomInstanceNum) and sceneInstance.name != str(instances[targetParentIndex][1]): intersect = instancesIntersect(mathutils.Matrix.Identity(4), [cubeTarget], sceneInstance.matrix_world, sceneInstance.dupli_group.objects) if intersect: print("Intersects with " + sceneInstance.name) break intersections = intersections + [[objAzimuth, intersect]] startInterval = True intervals = [] initInterval = 0 endInterval = 0 for idx, intersection in enumerate(intersections): if not intersection[1]: if startInterval: initInterval = intersection[0] startInterval = False else: if not startInterval: if idx >= 1 and intersections[idx-1][0] != initInterval: endInterval = intersection[0] - collisionInterval intervals = intervals + [[initInterval, endInterval]] startInterval = True if not intersection[1]: endInterval = intersection[0] if not intersections[0][1]: intervals = intervals + [[initInterval, endInterval+collisionInterval]] else: intervals = intervals + [[initInterval, endInterval]] targetCollisions[targetParentIndex] = (targetParentPosition, intervals) sceneCollisions[sceneNumber] = targetCollisions with open('data/collisions_new/collisionScene' + str(sceneNumber) + '.pickle', 'wb') as pfile: pickle.dump(targetCollisions, pfile) print("Collision detection ended.") # with open('data/collisions/collisions.pickle', 'wb') as pfile: # pickle.dump(sceneCollisions, pfile) # Cleanup # for scene in bpy.data.scenes: # # for objnum, obji in enumerate(scene.objects): # # # # obji.user_clear() # # bpy.data.objects.remove(obji) # # scene = bpy.data.scenes['Main Scene'] # if scene.name != 'Scene': # scene.user_clear() # bpy.data.scenes.remove(scene) # bpy.ops.scene.delete()
5,875
37.657895
191
py
inversegraphics
inversegraphics-master/scene_io_utils.py
from blender_utils import * from sklearn.preprocessing import normalize from collections import OrderedDict def loadTeapotsOpenDRData(renderTeapotsList, useBlender, unpackModelsFromBlender, targetModels): v_teapots = [] f_list_teapots = [] vc_teapots = [] vn_teapots = [] uv_teapots = [] haveTextures_list_teapots = [] textures_list_teapots = [] blender_teapots = [] center_teapots = [] for teapotIdx, teapotNum in enumerate(renderTeapotsList): objectDicFile = 'data/target' + str(teapotNum) + '.pickle' if useBlender: teapot = targetModels[teapotIdx] teapot.layers[1] = True teapot.layers[2] = True blender_teapots = blender_teapots + [teapot] if unpackModelsFromBlender: vmod, fmod_list, vcmod, vnmod, uvmod, haveTexturesmod_list, texturesmod_list = unpackBlenderObject(teapot, objectDicFile, True) else: vmod, fmod_list, vcmod, vnmod, uvmod, haveTexturesmod_list, texturesmod_list = loadSavedObject(objectDicFile) v_teapots = v_teapots + [vmod] f_list_teapots = f_list_teapots + [fmod_list] vc_teapots = vc_teapots + [vcmod] vn_teapots = vn_teapots + [vnmod] uv_teapots = uv_teapots + [uvmod] haveTextures_list_teapots = haveTextures_list_teapots + [haveTexturesmod_list] textures_list_teapots = textures_list_teapots + [texturesmod_list] vflat = [item for sublist in vmod for item in sublist] varray = np.vstack(vflat) center_teapots = center_teapots + [np.sum(varray, axis=0)/len(varray)] return v_teapots, f_list_teapots, vc_teapots, vn_teapots, uv_teapots, haveTextures_list_teapots, textures_list_teapots, vflat, varray, center_teapots def loadMugsOpenDRData(mugFiles, useBlender, unpackModelsFromBlender, mugModels=None): v_mugs = [] f_list_mugs = [] vc_mugs = [] vn_mugs = [] uv_mugs = [] haveTextures_list_mugs = [] textures_list_mugs = [] blender_mugs = [] center_mugs = [] for mugIdx, mugNum in enumerate(mugFiles): objectDicFile = 'data/mug' + str(mugNum) + '.pickle' if useBlender: mug = mugModels[mugIdx] mug.layers[1] = True mug.layers[2] = True blender_mugs = blender_mugs + [mug] if unpackModelsFromBlender: vmod, fmod_list, vcmod, vnmod, uvmod, haveTexturesmod_list, texturesmod_list = unpackBlenderObject(mug, objectDicFile, True) else: vmod, fmod_list, vcmod, vnmod, uvmod, haveTexturesmod_list, texturesmod_list = loadSavedObject(objectDicFile) v_mugs = v_mugs + [vmod] f_list_mugs = f_list_mugs + [fmod_list] vc_mugs = vc_mugs + [vcmod] vn_mugs = vn_mugs + [vnmod] uv_mugs = uv_mugs + [uvmod] haveTextures_list_mugs = haveTextures_list_mugs + [haveTexturesmod_list] textures_list_mugs = textures_list_mugs + [texturesmod_list] vflat = [item for sublist in vmod for item in sublist] varray = np.vstack(vflat) center_mugs = center_mugs + [np.sum(varray, axis=0)/len(varray)] return v_mugs, f_list_mugs, vc_mugs, vn_mugs, uv_mugs, haveTextures_list_mugs, textures_list_mugs, vflat, varray, center_mugs def getSceneInstancesInfo(sceneFile): sceneLines = [line.strip() for line in open(sceneFile)] numModels = sceneLines[2].split()[1] instances = [] for line in sceneLines: parts = line.split() if parts[0] == 'newModel': modelId = parts[2] if parts[0] == 'parentContactPosition': parentContactPosition = mathutils.Vector([float(parts[1])*inchToMeter, float(parts[2])*inchToMeter, float(parts[3])*inchToMeter]) if parts[0] == 'parentIndex': parentIndex = int(parts[1]) if parts[0] == 'transform': transform = mathutils.Matrix([[float(parts[1]), float(parts[2]), float(parts[3]), float(parts[4])], [float(parts[5]), float(parts[6]), float(parts[7]), float(parts[8])], [ float(parts[9]), float(parts[10]), float(parts[11]), float(parts[12])], [float(parts[13]), float(parts[14]), float(parts[15]), float(parts[16])]]).transposed() # ipdb.set_trace() transform[0][3] = transform[0][3]*inchToMeter transform[1][3] = transform[1][3]*inchToMeter transform[2][3] = transform[2][3]*inchToMeter # ipdb.set_trace() instances.append([modelId, parentIndex, parentContactPosition, transform]) return instances def composeScene(modelInstances): bpy.context.scene.name = 'Main Scene' scene = bpy.context.scene scene.unit_settings.system = 'METRIC' modelNum = 0 for instanceidx, instance in enumerate(modelInstances): # if modelNum != targetIndex: instance.name = str(instanceidx) scene.objects.link(instance) # modelNum = modelNum + 1 return scene def importBlenderScenes(instances, completeScene): baseDir = '../COLLADA/' blenderScenes = [] modelInstances = [] modelNum = 0 for instance in instances: modelId = instance[0] reg = re.compile('(room[0-9]+)') isRoom = reg.match(modelId) if completeScene or isRoom: transform = instance[3] modelPath = baseDir + modelId + '_cleaned.obj' print('Importing ' + modelPath ) # if modelNum != targetIndex: bpy.ops.scene.new() bpy.context.scene.name = modelId scene = bpy.context.scene # scene.unit_settings.system = 'METRIC' # bpy.utils.collada_import(modelPath) bpy.ops.import_scene.obj(filepath=modelPath, split_mode='OFF', use_split_objects=True, use_split_groups=False) # ipdb.set_trace() sceneGroup = bpy.data.groups.new(modelId) scene.update() scaleMat = mathutils.Matrix.Scale(inchToMeter, 4) # xrotation = mathutils.Matrix.Rotation(-90,4, 'X') for mesh in scene.objects: if mesh.type == 'MESH': sceneGroup.objects.link(mesh) # ipdb.set_trace() # mesh_transform = mesh.matrix_world # mesh.matrix_world = transform * mesh.matrix_world mesh.pass_index = 0 # mesh.matrix_world[0][3] = mesh.matrix_world[0][3]*inchToMeter # mesh.matrix_world[1][3] = mesh.matrix_world[1][3]*inchToMeter # mesh.matrix_world[2][3] = mesh.matrix_world[2][3]*inchToMeter # mesh.matrix_world = scaleMat * mesh.matrix_world # ipdb.set_trace() # mesh.data.show_double_sided = True mesh.data.update() modelInstance = bpy.data.objects.new(modelId, None) modelInstance.dupli_type = 'GROUP' modelInstance.dupli_group = sceneGroup modelInstance.matrix_world = transform modelInstance.pass_index = 0 modelInstances.append(modelInstance) modelNum = modelNum + 1 # ipdb.set_trace() scene.update() blenderScenes.append(scene) return blenderScenes, modelInstances import os.path def createTargetsBlendFile(): bpy.ops.wm.read_factory_settings() # bpy.ops.wm.read_homefile(load_ui=False) teapots = [line.strip() for line in open('teapots.txt')] renderTeapotsList = np.arange(len(teapots)) [targetScenes, targetModels, transformations] = loadTargetModels(renderTeapotsList) # bpy.data.scenes.remove(bpy.data.scenes['Scene']) bpy.ops.file.pack_all() bpy.ops.wm.save_as_mainfile(filepath='data/targets.blend') bpy.ops.wm.read_factory_settings() def createMugsBlendFile(): bpy.ops.wm.read_factory_settings() # bpy.ops.wm.read_homefile(load_ui=False) mugs = [line.strip() for line in open('mugs.txt')] renderMugsList = np.arange(len(mugs)) [targetScenes, targetModels, transformations] = loadMugsModels(renderMugsList) # bpy.data.scenes.remove(bpy.data.scenes['Scene']) bpy.ops.file.pack_all() bpy.ops.wm.save_as_mainfile(filepath='data/mugs.blend') bpy.ops.wm.read_factory_settings() def createSceneBlendFiles(overwrite=True): replaceableScenesFile = '../databaseFull/fields/scene_replaceables_backup.txt' sceneLines = [line.strip() for line in open(replaceableScenesFile)] for sceneIdx in range(len(sceneLines)): sceneNumber, sceneFileName, instances, roomName, roomInstanceNum, targetIndices, targetPositions = getSceneInformation(sceneIdx, replaceableScenesFile) blendFilename = 'data/scene' + str(sceneNumber) + '.blend' if overwrite or not os.path.isfile(blendFilename): print("Importing Scene " + sceneFileName) scene = loadBlenderScene(sceneIdx, replaceableScenesFile) setupScene(scene, roomInstanceNum, scene.world, scene.camera, 100, 100, 16, True, False) scene.update() # bpy.data.scenes.remove(bpy.data.scenes['Scene']) bpy.ops.file.pack_all() bpy.ops.wm.save_as_mainfile(filepath=blendFilename) bpy.ops.wm.read_factory_settings() # bpy.ops.wm.read_homefile(load_ui=False) # bpy.ops.wm.open_mainfile(filepath='data/Scene.blend') def createSceneOpenDRFiles(overwrite=True): replaceableScenesFile = '../databaseFull/fields/scene_replaceables_backup.txt' sceneLines = [line.strip() for line in open(replaceableScenesFile)] for sceneIdx in range(len(sceneLines)): sceneNumber, sceneFileName, instances, roomName, roomInstanceNum, targetIndices, targetPositions = getSceneInformation(sceneIdx, replaceableScenesFile) pickleFilename = 'data/scene' + str(sceneNumber) + '.pickle' if overwrite or not os.path.isfile(pickleFilename): print("Unpacking Scene " + str(sceneNumber)) loadSceneBlendData(sceneIdx, replaceableScenesFile) scene = bpy.data.scenes['Main Scene'] v, f_list, vc, vn, uv, haveTextures_list, textures_list = unpackBlenderScene(scene, pickleFilename, True) bpy.ops.wm.read_factory_settings() def loadSceneBlendData(sceneIdx, replaceableScenesFile): replaceableScenesFile = replaceableScenesFile sceneLines = [line.strip() for line in open(replaceableScenesFile)] sceneLine = sceneLines[sceneIdx] sceneParts = sceneLine.split(' ') sceneFile = sceneParts[0] sceneNumber = int(re.search('.+?scene([0-9]+)\.txt', sceneFile, re.IGNORECASE).groups()[0]) sceneFilename = 'data/scene' + str(sceneNumber) + '.blend' with bpy.data.libraries.load(filepath=sceneFilename) as (data_from, data_to): for attr in dir(data_to): setattr(data_to, attr, getattr(data_from, attr)) bpy.context.screen.scene = bpy.data.scenes['Main Scene'] def loadTargetsBlendData(): targetsFilename = 'data/targets.blend' with bpy.data.libraries.load(filepath=targetsFilename) as (data_from, data_to): for attr in dir(data_to): setattr(data_to, attr, getattr(data_from, attr)) return data_to def loadMugsBlendData(): targetsFilename = 'data/mugs.blend' with bpy.data.libraries.load(filepath=targetsFilename) as (data_from, data_to): for attr in dir(data_to): setattr(data_to, attr, getattr(data_from, attr)) return data_to def loadTargetModels(experimentTeapots): teapots = [line.strip() for line in open('teapots.txt')] targetModels = [] baseDir = '../databaseFull/models/' targetInstances = [] blenderTeapots = [] transformations = [] modelNum = 0 selection = [ teapots[i] for i in experimentTeapots] for teapotidx, teapot in enumerate(selection): targetGroup = bpy.data.groups.new(teapot) fullTeapot = baseDir + teapot + '.obj' modelPath = fullTeapot bpy.ops.scene.new() bpy.context.scene.name = teapot scene = bpy.context.scene scene.unit_settings.system = 'METRIC' print("Importing " + modelPath) # bpy.utils.collada_import(modelPath) bpy.ops.import_scene.obj(filepath=modelPath, split_mode='OFF', use_split_objects=True, use_split_groups=False) scene.update() # modifySpecular(scene, 0.3) matrix_world = mathutils.Matrix.Identity(4) minZ, maxZ = modelHeight(scene.objects, mathutils.Matrix.Identity(4)) minY, maxY = modelDepth(scene.objects, mathutils.Matrix.Identity(4)) scaleZ = 0.265/(maxZ-minZ) scaleY = 0.18/(maxY-minY) ratio = (maxZ-minZ)/(maxY-minY) if ratio > 0.265/0.18: scale = scaleZ else: scale = scaleY # scale = min(scaleZ, scaleY) scaleMat = mathutils.Matrix.Scale(scale, 4) for mesh in scene.objects: if mesh.type == 'MESH': mesh.matrix_world = scaleMat * mesh.matrix_world mesh.data.update() matrix_world = scaleMat * matrix_world rot = mathutils.Matrix.Rotation(radians(90), 4, 'Z') rotateMatrixWorld(scene, rot ) matrix_world = rot * matrix_world minZ, maxZ = modelHeight(scene.objects, mathutils.Matrix.Identity(4)) center = centerOfGeometry(scene.objects, mathutils.Matrix.Identity(4)) for mesh in scene.objects: if mesh.type == 'MESH': mesh.matrix_world = mathutils.Matrix.Translation(-center) * mesh.matrix_world matrix_world = mathutils.Matrix.Translation(-center) * matrix_world minZ, maxZ = modelHeight(scene.objects, mathutils.Matrix.Identity(4)) for mesh in scene.objects: if mesh.type == 'MESH': mesh.matrix_world = mathutils.Matrix.Translation(mathutils.Vector((0,0,-minZ))) * mesh.matrix_world matrix_world = mathutils.Matrix.Translation(mathutils.Vector((0,0,-minZ))) * matrix_world transformations = transformations + [matrix_world] for mesh in scene.objects: # mesh.update() targetGroup.objects.link(mesh) mesh.pass_index = 1 targetInstance = bpy.data.objects.new(teapot, None) targetInstance.dupli_type = 'GROUP' targetInstance.dupli_group = targetGroup targetInstance.pass_index = 1 targetInstances.append(targetInstance) targetInstance.name = 'teapotInstance' + str(experimentTeapots[teapotidx]) scene.objects.link(targetInstance) scene.update() blenderTeapots.append(scene) # ipdb.set_trace() return blenderTeapots, targetInstances, transformations def loadMugsModels(experimentMugs): mugs = [line.strip() for line in open('mugs.txt')] targetModels = [] baseDir = '../databaseFull/models/' targetInstances = [] blenderMugs = [] transformations = [] modelNum = 0 selection = [ mugs[i] for i in experimentMugs] for mugidx, mug in enumerate(selection): targetGroup = bpy.data.groups.new(mug) fullTeapot = baseDir + mug + '.obj' modelPath = fullTeapot bpy.ops.scene.new() bpy.context.scene.name = mug scene = bpy.context.scene scene.unit_settings.system = 'METRIC' print("Importing " + modelPath) # bpy.utils.collada_import(modelPath) bpy.ops.import_scene.obj(filepath=modelPath, split_mode='OFF', use_split_objects=True, use_split_groups=False) scene.update() # modifySpecular(scene, 0.3) matrix_world = mathutils.Matrix.Identity(4) minZ, maxZ = modelHeight(scene.objects, mathutils.Matrix.Identity(4)) minY, maxY = modelDepth(scene.objects, mathutils.Matrix.Identity(4)) scaleZ = 0.1/(maxZ-minZ) ratio = (maxZ-minZ)/(maxY-minY) scale = scaleZ # scale = min(scaleZ, scaleY) scaleMat = mathutils.Matrix.Scale(scale, 4) for mesh in scene.objects: if mesh.type == 'MESH': mesh.matrix_world = scaleMat * mesh.matrix_world mesh.data.update() matrix_world = scaleMat * matrix_world rot = mathutils.Matrix.Rotation(radians(90), 4, 'Z') rotateMatrixWorld(scene, rot ) matrix_world = rot * matrix_world minZ, maxZ = modelHeight(scene.objects, mathutils.Matrix.Identity(4)) center = centerOfGeometry(scene.objects, mathutils.Matrix.Identity(4)) for mesh in scene.objects: if mesh.type == 'MESH': mesh.matrix_world = mathutils.Matrix.Translation(-center) * mesh.matrix_world matrix_world = mathutils.Matrix.Translation(-center) * matrix_world minZ, maxZ = modelHeight(scene.objects, mathutils.Matrix.Identity(4)) for mesh in scene.objects: if mesh.type == 'MESH': mesh.matrix_world = mathutils.Matrix.Translation(mathutils.Vector((0,0,-minZ))) * mesh.matrix_world matrix_world = mathutils.Matrix.Translation(mathutils.Vector((0,0,-minZ))) * matrix_world transformations = transformations + [matrix_world] for mesh in scene.objects: # mesh.update() targetGroup.objects.link(mesh) mesh.pass_index = 1 targetInstance = bpy.data.objects.new(mug, None) targetInstance.dupli_type = 'GROUP' targetInstance.dupli_group = targetGroup targetInstance.pass_index = 1 targetInstances.append(targetInstance) targetInstance.name = 'mugInstance' + str(experimentMugs[mugidx]) scene.objects.link(targetInstance) scene.update() blenderMugs.append(scene) # ipdb.set_trace() return blenderMugs, targetInstances, transformations def getSceneInformation(sceneIdx, scenesFile): replaceableScenesFile = scenesFile sceneLines = [line.strip() for line in open(replaceableScenesFile)] sceneLineNums = numpy.arange(len(sceneLines)) sceneNum = sceneLineNums[sceneIdx] sceneLine = sceneLines[sceneNum] sceneParts = sceneLine.split(' ') sceneFile = sceneParts[0] sceneNumber = int(re.search('.+?scene([0-9]+)\.txt', sceneFile, re.IGNORECASE).groups()[0]) sceneFileName = re.search('.+?(scene[0-9]+\.txt)', sceneFile, re.IGNORECASE).groups()[0] instances = getSceneInstancesInfo('../databaseFull/scenes/' + sceneFileName) targetPositions = [] targetIndices = [] for targetIndex in sceneParts[1::]: targetIndex = int(targetIndex) targetParentPosition = instances[targetIndex][2] targetIndices = targetIndices + [targetIndex] targetPositions = targetPositions + [np.array(targetParentPosition)] roomName = '' roomInstanceNum = 0 for modelIdx, model in enumerate(instances): reg = re.compile('(room[0-9]+)') res = reg.match(model[0]) if res: roomName = res.groups()[0] roomInstanceNum = modelIdx return sceneNumber, sceneFileName, instances, roomName, roomInstanceNum, targetIndices, targetPositions def getSceneIdx(sceneNumber, scenesFile): replaceableScenesFile = scenesFile sceneLines = [line.strip() for line in open(replaceableScenesFile)] sceneLineNums = numpy.arange(len(sceneLines)) for sceneNum in sceneLineNums: sceneLine = sceneLines[sceneNum] sceneParts = sceneLine.split(' ') sceneFile = sceneParts[0] sceneNumInFile = int(re.search('.+?scene([0-9]+)\.txt', sceneFile, re.IGNORECASE).groups()[0]) if sceneNumInFile == sceneNumber: return sceneNum return -2 def loadBlenderScene(sceneIdx, replaceableScenesFile): sceneNumber, sceneFileName, instances, roomName, roomInstanceNum, targetIndices, targetPositions = getSceneInformation(sceneIdx, replaceableScenesFile) # targetParentPosition = instances[targetIndex][2] # targetParentIndex = instances[targetIndex][1] cam = bpy.data.cameras.new("MainCamera") camera = bpy.data.objects.new("MainCamera", cam) world = bpy.data.worlds.new("MainWorld") [blenderScenes, modelInstances] = importBlenderScenes(instances, True) # targetParentInstance = modelInstances[targetParentIndex] scene = composeScene(modelInstances) scene.world = world scene.camera = camera # roomInstance = scene.objects[roomName] # roomInstance.layers[2] = True # targetParentInstance.layers[2] = True return scene def loadSavedScene(sceneDicFile, tex_srgb2lin): with open(sceneDicFile, 'rb') as pfile: sceneDic = pickle.load(pfile) v = sceneDic['v'] f_list = sceneDic['f_list'] vc = sceneDic['vc'] uv = sceneDic['uv'] haveTextures_list = sceneDic['haveTextures_list'] vn = sceneDic['vn'] textures_list = sceneDic['textures_list'] if tex_srgb2lin: textures_listflat = [item for sublist in textures_list for item in sublist] for texture_list in textures_listflat: if texture_list is not None: for texture in texture_list: if texture is not None: srgb2lin(texture) print("Loaded serialized scene!") return v, f_list, vc, vn, uv, haveTextures_list, textures_list def unpackBlenderScene(scene, sceneDicFile, serializeScene): # bpy.ops.render.render( write_still=True ) # ipdb.set_trace() # v,f_list, vc, vn, uv, haveTextures_list, textures_list = unpackObjects(teapot) v = [] f_list = [] vc = [] vn = [] uv = [] haveTextures_list = [] textures_list = [] print("Unpacking blender data for OpenDR.") for modelInstance in scene.objects: if modelInstance.dupli_group != None: vmod,f_listmod, vcmod, vnmod, uvmod, haveTextures_listmod, textures_listmod = unpackBlenderObject(modelInstance, '', False) # gray = np.dot(np.array([0.3, 0.59, 0.11]), vcmod[0].T).T # sat = 0.5 # vcmod[0][:,0] = vcmod[0][:,0] * sat + (1-sat) * gray # vcmod[0][:,1] = vcmod[0][:,1] * sat + (1-sat) * gray # vcmod[0][:,2] = vcmod[0][:,2] * sat + (1-sat) * gray v = v + vmod f_list = f_list + f_listmod vc = vc + vcmod vn = vn + vnmod uv = uv + uvmod haveTextures_list = haveTextures_list + haveTextures_listmod textures_list = textures_list + textures_listmod #Serialize if serializeScene: sceneDic = {'v':v,'f_list':f_list,'vc':vc,'uv':uv,'haveTextures_list':haveTextures_list,'vn':vn,'textures_list': textures_list} with open(sceneDicFile, 'wb') as pfile: pickle.dump(sceneDic, pfile) print("Serialized scene!") return v, f_list, vc, vn, uv, haveTextures_list, textures_list def loadSavedObject(objectDicFile): with open(objectDicFile, 'rb') as pfile: targetDic = pickle.load(pfile) v = targetDic['v'] f_list = targetDic['f_list'] vc = targetDic['vc'] uv = targetDic['uv'] haveTextures_list = targetDic['haveTextures_list'] vn = targetDic['vn'] textures_list = targetDic['textures_list'] print("Loaded serialized target!") return [v], [f_list], [vc], [vn], [uv], [haveTextures_list], [textures_list] def unpackBlenderObject(object, objectDicFile, saveData): f_list = [] v = [] vc = [] vn = [] uv = [] haveTextures = [] textures_list = [] vertexMeshIndex = 0 for mesh in object.dupli_group.objects: if mesh.type == 'MESH': # mesh.data.validate(verbose=True, clean_customdata=True) fmesh, vmesh, vcmesh, nmesh, uvmesh, haveTexture, textures = buildData(mesh.data) f_list = f_list + [fmesh] vc = vc + [vcmesh] transf = np.array(np.dot(object.matrix_world, mesh.matrix_world)) vmesh = np.hstack([vmesh, np.ones([vmesh.shape[0],1])]) vmesh = ( np.dot(transf , vmesh.T)).T[:,0:3] v = v + [vmesh] transInvMat = np.linalg.inv(transf).T nmesh = np.hstack([nmesh, np.ones([nmesh.shape[0],1])]) nmesh = (np.dot(transInvMat , nmesh.T)).T[:,0:3] vn = vn + [normalize(nmesh, axis=1)] uv = uv + [uvmesh] haveTextures_list = haveTextures + [haveTexture] textures_list = textures_list + [textures] vertexMeshIndex = vertexMeshIndex + len(vmesh) #Serialize if saveData: targetDic = {'v':v,'f_list':f_list,'vc':vc,'uv':uv,'haveTextures_list':haveTextures_list,'vn':vn,'textures_list': textures_list} with open(objectDicFile, 'wb') as pfile: pickle.dump(targetDic, pfile) print("Serialized object!") return [v],[f_list],[vc],[vn], [uv], [haveTextures_list], [textures_list] def buildData (msh): lvdic = {} # local dictionary lfl = [] # lcoal faces index list lvl = [] # local vertex list lvcl = [] lnl = [] # local normal list luvl = [] # local uv list lvcnt = 0 # local vertices count isSmooth = False texdic = {} # local dictionary msh.calc_tessface() # if len(msh.tessfaces) == 0 or msh.tessfaces is None: # msh.calc_tessface() textureNames = [] haveUVs = [] for i,f in enumerate(msh.polygons): isSmooth = f.use_smooth tmpfaces = [] hasUV = False # true by default, it will be verified below texture = None texname = None if (len(msh.tessface_uv_textures)>0): activeUV = msh.tessface_uv_textures.active.data if msh.tessface_uv_textures.active.data[i].image is not None: # ipdb.set_trace() texname = msh.tessface_uv_textures.active.data[i].image.name hasUV = True texture = texdic.get(texname) if (texture is None): # vertex not found # print("Image: " + texname) # print("Clamp x: " + str(msh.tessface_uv_textures.active.data[i].image.use_clamp_x)) # print("Clamp y: " + str(msh.tessface_uv_textures.active.data[i].image.use_clamp_y)) # print("Tile x: " + str(msh.tessface_uv_textures.active.data[i].image.tiles_x)) # print("Tile y: " + str(msh.tessface_uv_textures.active.data[i].image.tiles_y)) texture = np.flipud(np.array(msh.tessface_uv_textures.active.data[i].image.pixels).reshape([msh.tessface_uv_textures.active.data[i].image.size[1],msh.tessface_uv_textures.active.data[i].image.size[0],4])[:,:,:3]) texture = srgb2lin(texture) if np.any(np.isnan(texture)) or np.any(texture<0) or np.any(texture>1) or texture.size == 0: print("Problem with texture from Blender") texture = np.flipud(np.array(msh.tessface_uv_textures.active.data[i].image.pixels).reshape([msh.tessface_uv_textures.active.data[i].image.size[1],msh.tessface_uv_textures.active.data[i].image.size[0],4])[:,:,:3]) hasUV = False texture = None texname = None if hasUV: texdic[texname] = texture textureNames = textureNames + [texname] haveUVs = haveUVs + [hasUV] for j,v in enumerate(f.vertices): vec = msh.vertices[v].co vec = r3d(vec) if (isSmooth): # use vertex normal nor = msh.vertices[v].normal else: # use face normal nor = f.normal vcolor = msh.materials[f.material_index].diffuse_color[:] if vcolor == (0.0,0.0,0.0) and msh.materials[f.material_index].specular_color[:] != (0.0,0.0,0.0): vcolor = msh.materials[f.material_index].specular_color[:] # print("Using specular!") nor = r3d(nor) co = (0.0, 0.0) if hasUV: co = activeUV[i].uv[j] co = r2d(co) vcolor = (1.0,1.0,1.0) key = vec, nor, co vinx = lvdic.get(key) if (vinx is None): # vertex not found lvdic[key] = lvcnt lvl.append(vec) lnl.append(nor) lvcl.append(vcolor) luvl.append(co) tmpfaces.append(lvcnt) lvcnt+=1 else: inx = lvdic[key] tmpfaces.append(inx) if (len(tmpfaces)==3): lfl.append(tmpfaces) else: lfl.append([tmpfaces[0], tmpfaces[1], tmpfaces[2]]) lfl.append([tmpfaces[0], tmpfaces[2], tmpfaces[3]]) # vtx.append(lvdic) textures = [] haveTextures = [] f_list = [] orderedtexs = OrderedDict(sorted(texdic.items(), key=lambda t: t[0])) for texname, texture in orderedtexs.items(): fidxs = [lfl[idx] for idx in range(len(lfl)) if textureNames[idx] == texname] f_list = f_list + [np.vstack(fidxs)] textures = textures + [texture] haveTextures = haveTextures + [True] try: fidxs = [lfl[idx] for idx in range(len(lfl)) if haveUVs[idx] == False] except: ipdb.set_trace() if fidxs != None and fidxs != []: f_list = f_list + [np.vstack(fidxs)] textures = textures + [None] haveTextures = haveTextures + [False] #update global lists and dictionaries v = np.vstack(lvl) vc = np.vstack(lvcl) n = np.vstack(lnl) uv = np.vstack(luvl) return f_list, v, vc, n, uv, haveTextures, textures
29,860
39.905479
343
py
inversegraphics
inversegraphics-master/torch_nn.py
import ipdb import PyTorchAug import PyTorch nn = PyTorch.Nn() lua = PyTorchAug.lua lua.getGlobal("require") lua.pushString('modules/LinearCR') lua.call(1, 0) lua = PyTorchAug.lua lua.getGlobal("require") lua.pushString('modules/Reparametrize') lua.call(1, 0) lua = PyTorchAug.lua lua.getGlobal("require") lua.pushString('modules/SelectiveOutputClamp') lua.call(1, 0) lua = PyTorchAug.lua lua.getGlobal("require") lua.pushString('modules/SelectiveGradientFilter') lua.call(1, 0) dim_hidden = 200 feature_maps = 96 filter_size = 5 colorchaPyTorchAugels = 1 # ipdb.set_trace() encoder = PyTorchAug.Sequential() encoder.add(PyTorchAug.SpatialConvolution(colorchaPyTorchAugels,feature_maps,filter_size,filter_size)) encoder.add(PyTorchAug.SpatialMaxPooling(2,2,2,2)) encoder.add(PyTorchAug.Threshold(0,1e-6)) encoder.add(PyTorchAug.SpatialConvolution(feature_maps,feature_maps/2,filter_size,filter_size)) encoder.add(PyTorchAug.SpatialMaxPooling(2,2,2,2)) encoder.add(PyTorchAug.Threshold(0,1e-6)) encoder.add(PyTorchAug.SpatialConvolution(feature_maps/2,feature_maps/4,filter_size,filter_size)) encoder.add(PyTorchAug.SpatialMaxPooling(2,2,2,2)) encoder.add(PyTorchAug.Threshold(0,1e-6)) encoder.add(PyTorchAug.Reshape((feature_maps/4)*15*15)) z = PyTorchAug.ConcatTable() mu = PyTorchAug.Sequential() mu.add(PyTorchAug.LinearCR((feature_maps/4)*15*15, dim_hidden)) mu.add(PyTorchAug.SelectiveGradientFilter()) mu.add(PyTorchAug.SelectiveOutputClamp()) z.add(mu) sigma = PyTorchAug.Sequential() sigma.add(PyTorchAug.LinearCR((feature_maps/4)*15*15, dim_hidden)) sigma.add(PyTorchAug.SelectiveGradientFilter()) sigma.add(PyTorchAug.SelectiveOutputClamp()) z.add(sigma) encoder.add(z) decoder = PyTorchAug.Sequential() decoder.add(PyTorchAug.LinearCR(dim_hidden, (feature_maps/4)*15*15 )) decoder.add(PyTorchAug.Threshold(0,1e-6)) decoder.add(PyTorchAug.Reshape((feature_maps/4),15,15)) decoder.add(PyTorchAug.SpatialUpSamplingNearest(2)) decoder.add(PyTorchAug.SpatialConvolution(feature_maps/4,feature_maps/2, 7, 7)) decoder.add(PyTorchAug.Threshold(0,1e-6)) decoder.add(PyTorchAug.SpatialUpSamplingNearest(2)) decoder.add(PyTorchAug.SpatialConvolution(feature_maps/2,feature_maps,7,7)) decoder.add(PyTorchAug.Threshold(0,1e-6)) decoder.add(PyTorchAug.SpatialUpSamplingNearest(2)) decoder.add(PyTorchAug.SpatialConvolution(feature_maps,feature_maps,7,7)) decoder.add(PyTorchAug.Threshold(0,1e-6)) decoder.add(PyTorchAug.SpatialUpSamplingNearest(2)) decoder.add(PyTorchAug.SpatialConvolution(feature_maps,1,7,7)) decoder.add(PyTorchAug.Sigmoid()) model = PyTorchAug.Sequential() model.add(encoder) model.add(PyTorchAug.Reparametrize(dim_hidden)) model.add(decoder) model.cuda() nn.collectgarbage()
2,724
25.980198
102
py
inversegraphics
inversegraphics-master/diffrender_experiment.py
__author__ = 'pol' import matplotlib matplotlib.use('Qt4Agg') from math import radians import timeit import time import numpy as np from utils import * import matplotlib.pyplot as plt plt.ion() import h5py import ipdb import pickle ######################################### # Initialization ends here ######################################### seed = 1 np.random.seed(seed) gtPrefix = 'objectnet3d_teapots' experimentPrefix = 'objectnet3d_teapots' experimentDescr = 'The real teapots' gtDir = 'groundtruth/' + gtPrefix + '/' experimentDir = 'experiments/' + experimentPrefix + '/' # groundTruthFilename = gtDir + 'groundTruth.h5' # gtDataFile = h5py.File(groundTruthFilename, 'r') # # onlySynthetic = False # # # print("Reading experiment data.") # # shapeGT = gtDataFile[gtPrefix].shape # # groundTruth = gtDataFile[gtPrefix] # # dataAzsGT = groundTruth['trainAzsGT'] # dataObjAzsGT = groundTruth['trainObjAzsGT'] # dataElevsGT = groundTruth['trainElevsGT'] # dataLightAzsGT = groundTruth['trainLightAzsGT'] # dataLightElevsGT = groundTruth['trainLightElevsGT'] # dataLightIntensitiesGT = groundTruth['trainLightIntensitiesGT'] # dataVColorGT = groundTruth['trainVColorGT'] # dataScenes = groundTruth['trainScenes'] # dataTeapotIds = groundTruth['trainTeapotIds'] # dataEnvMaps = groundTruth['trainEnvMaps'] # dataOcclusions = groundTruth['trainOcclusions'] # dataTargetIndices = groundTruth['trainTargetIndices'] # dataLightCoefficientsGT = groundTruth['trainLightCoefficientsGT'] # dataLightCoefficientsGTRel = groundTruth['trainLightCoefficientsGTRel'] # dataAmbientIntensityGT = groundTruth['trainAmbientIntensityGT'] # dataIds = groundTruth['trainIds'] # # gtDtype = groundTruth.dtype # # allDataIds = gtDataFile[gtPrefix]['trainIds'] ########## Check if there is anything wrong with the renders: # print("Reading images.") # # images = readImages(imagesDir, trainSet, False, loadFromHdf5) # writeHdf5 = False # writeGray = False # if writeHdf5: # writeImagesHdf5(gtDir, gtDir, allDataIds, writeGray) # if onlySynthetic: # imagesDir = gtDir + 'images_opendr/' # else: # imagesDir = gtDir + 'images/' # # loadGray = True # imagesAreH5 = False # loadGrayFromHdf5 = False # # if not imagesAreH5: # grayImages = readImages(imagesDir, allDataIds, loadGray, loadGrayFromHdf5) # else: # grayImages = h5py.File(imagesDir + 'images_gray.h5', 'r')["images"] # # badImages = np.where(np.mean(grayImages, (1,2)) < 0.01)[0] # # for id, badImage in enumerate(grayImages[badImages]): # print("There are bad images!") # plt.imsave('tmp/check/badImage' + str(badImages[id]) + '.png', np.tile(badImage[:,:,None], [1,1,3])) # size = 375 # if not os.path.isfile(experimentDir + 'train.npy'): np.random.seed(seed) data = np.arange(size) np.random.shuffle(data) generateExperiment(size, experimentDir, 0, 1) # if not os.path.exists(experimentDir): # os.makedirs(experimentDir) # # np.save(experimentDir + 'train.npy', train) # np.save(experimentDir + 'test.npy', test) ########## Out of sample selections. # testSamplesIds= [2,4] # trainSamplesIds = [0,14,20,25,26,1] # # dataIdx = np.arange(shapeGT[0]) # train = np.array([],dtype=np.uint16) # test = np.array([],dtype=np.uint16) # for testId in testSamplesIds: # test = np.append(test, np.where(dataTeapotIds == testId)) # # for trainId in trainSamplesIds: # train = np.append(train, np.where(dataTeapotIds == trainId)) # # # boolTrainSet = np.ones(shapeGT[0]).astype(np.bool) # # boolTrainSet[test] = False # # train = dataIdx[boolTrainSet] # # np.random.shuffle(train) # np.random.shuffle(test) # # if not os.path.exists(experimentDir): # os.makedirs(experimentDir) # # np.save(experimentDir + 'train.npy', train) # np.save(experimentDir + 'test.npy', test) with open(experimentDir + 'description.txt', 'w') as expfile: expfile.write(experimentDescr)
3,854
26.733813
106
py
inversegraphics
inversegraphics-master/diffrender_train.py
__author__ = 'pol' import matplotlib matplotlib.use('Qt4Agg') from math import radians import timeit import time import image_processing import numpy as np import cv2 from utils import * import generative_models import matplotlib.pyplot as plt plt.ion() import recognition_models import skimage import h5py import ipdb import pickle import lasagne_nn import lasagne import theano ######################################### # Initialization ends here ######################################### seed = 1 np.random.seed(seed) gtPrefix = 'train4_occlusion_shapemodel_triplets' experimentPrefix = 'train4_occlusion_shapemodel_10k' gtDir = 'groundtruth/' + gtPrefix + '/' experimentDir = 'experiments/' + experimentPrefix + '/' groundTruthFilename = gtDir + 'groundTruth.h5' gtDataFile = h5py.File(groundTruthFilename, 'r') allDataIds = gtDataFile[gtPrefix]['trainIds'] onlySynthetic = True # trainSet = np.load(experimentDir + 'train.npy')[:12800] trainSet = np.load(experimentDir + 'train.npy')[:] print("Reading experiment data.") shapeGT = gtDataFile[gtPrefix].shape boolTestSet = np.zeros(shapeGT).astype(np.bool) boolTestSet[trainSet] = True trainGroundTruth = gtDataFile[gtPrefix][boolTestSet] groundTruth = np.zeros(shapeGT, dtype=trainGroundTruth.dtype) groundTruth[boolTestSet] = trainGroundTruth groundTruth = groundTruth[trainSet] dataAzsGT = groundTruth['trainAzsGT'] dataObjAzsGT = groundTruth['trainObjAzsGT'] dataElevsGT = groundTruth['trainElevsGT'] # dataLightAzsGT = groundTruth['trainLightAzsGT'] # dataLightElevsGT = groundTruth['trainLightElevsGT'] # dataLightIntensitiesGT = groundTruth['trainLightIntensitiesGT'] dataVColorGT = groundTruth['trainVColorGT'] # dataScenes = groundTruth['trainScenes'] dataTeapotIds = groundTruth['trainTeapotIds'] # dataEnvMaps = groundTruth['trainEnvMaps'] dataOcclusions = groundTruth['trainOcclusions'] # dataTargetIndices = groundTruth['trainTargetIndices'] # dataLightCoefficientsGT = groundTruth['trainLightCoefficientsGT'] dataLightCoefficientsGTRel = groundTruth['trainLightCoefficientsGTRel'] dataAmbientIntensityGT = groundTruth['trainAmbientIntensityGT'] dataShapeModelCoeffsGT = groundTruth['trainShapeModelCoeffsGT'] dataIds = groundTruth['trainIds'] gtDtype = groundTruth.dtype # gtDtype = [('trainIds', trainIds.dtype.name), ('trainAzsGT', trainAzsGT.dtype.name),('trainObjAzsGT', trainObjAzsGT.dtype.name),('trainElevsGT', trainElevsGT.dtype.name),('trainLightAzsGT', trainLightAzsGT.dtype.name),('trainLightElevsGT', trainLightElevsGT.dtype.name),('trainLightIntensitiesGT', trainLightIntensitiesGT.dtype.name),('trainVColorGT', trainVColorGT.dtype.name, (3,) ),('trainScenes', trainScenes.dtype.name),('trainTeapotIds', trainTeapotIds.dtype.name),('trainEnvMaps', trainEnvMaps.dtype.name),('trainOcclusions', trainOcclusions.dtype.name),('trainTargetIndices', trainTargetIndices.dtype.name), ('trainComponentsGT', trainComponentsGT.dtype, (9,)),('trainComponentsGTRel', trainComponentsGTRel.dtype, (9,))] # Create experiment simple sepratation. # trainAzsRel = np.mod(dataAzsGT - dataObjAzsGT, 2*np.pi) trainElevsGT = dataElevsGT # trainComponentsGTRel = dataComponentsGTRel trainVColorGT = dataVColorGT loadFromHdf5 = False print("Reading images.") if onlySynthetic: imagesDir = gtDir + 'images_opendr/' else: imagesDir = gtDir + 'images/' loadGray = False imagesAreH5 = False loadGrayFromHdf5 = False trainTriplets = True filter = np.ones(len(trainSet)).astype(np.bool) filter = np.array(tuple(dataOcclusions >= 0) and tuple(dataOcclusions < 0.9)) # filter = dataOcclusions < 0.3 trainSet = trainSet[filter] trainLightCoefficientsGTRel = dataLightCoefficientsGTRel[filter] trainShapeModelCoeffsGT = dataShapeModelCoeffsGT[filter] trainAzsRel=trainAzsRel[filter] trainElevsGT =trainElevsGT[filter] trainVColorGT=trainVColorGT[filter] trainAmbientIntensityGT = dataAmbientIntensityGT[filter] grayImages = readImages(imagesDir, trainSet, True, loadGrayFromHdf5) numTrainSet = len(grayImages) if trainTriplets: imagesDir_t1 = gtDir + 'triplets1/' imagesDir_t2 = gtDir + 'triplets2/' grayImages_t1 = readImages(imagesDir_t1, trainSet, True, loadGrayFromHdf5) grayImages_t2 = readImages(imagesDir_t2, trainSet, True, loadGrayFromHdf5) grayImages_75 = np.zeros([numTrainSet, 75,75]) grayImages_t1_75 = np.zeros([numTrainSet, 75,75]) grayImages_t2_75 = np.zeros([numTrainSet, 75, 75]) for imidx in range(numTrainSet): grayImages_75[imidx] = skimage.transform.resize(grayImages[imidx], [75, 75]) grayImages_t1_75[imidx] = skimage.transform.resize(grayImages_t1[imidx], [75, 75]) grayImages_t2_75[imidx] = skimage.transform.resize(grayImages_t2[imidx], [75, 75]) # images = readImages(imagesDir, trainSet, False, False) # grayImages = 0.3*images[:,:,:,0] + 0.59*images[:,:,:,1] + 0.11*images[:,:,:,2] # backprojectionsDir = gtDir + 'backprojections/' # backprojections = readImages(backprojectionsDir, trainSet, True, False) loadMask = False gtDirMask = 'groundtruth/train4_occlusion_mask/' masksDir = gtDirMask + 'masks_occlusion/' if loadMask: masksGT = loadMasks(masksDir, trainSet) loadHogFeatures = False loadFourierFeatures = False loadZernikeFeatures = False synthPrefix = '_cycles' if onlySynthetic: synthPrefix = '' # if loadHogFeatures: # hogfeatures = np.load(gtDir + 'hog' + synthPrefix + '.npy') # else: # print("Extracting Hog features .") # hogfeatures = image_processing.computeHoGFeatures(images) # np.save(gtDir + 'hog' + synthPrefix + '.npy', hogfeatures) # if loadIllumFeatures: # illumfeatures = np.load(experimentDir + 'illum.npy') # else: # print("Extracting Illumination features (FFT.") # illumfeatures = image_processing.computeIllumFeatures(images, images[0].size/12) # np.save(experimentDir + 'illum.npy', illumfeatures) # print("Extracting Zernike features.") # # numCoeffs=200 # # numTrees = 400 # win=40 # if loadZernikeFeatures: # trainZernikeCoeffs = np.load(gtDir + 'zernike_numCoeffs' + str(numCoeffs) + '_win' + str(win) + '.npy')[trainSet] # else: # print("Extracting Zernike features.") # batchSize = 1000 # # trainZernikeCoeffs = np.empty([images.shape[0], numCoeffs]) # for batch in range(int(images.shape[0]/batchSize)): # trainZernikeCoeffs[batchSize*batch:batchSize*batch + batchSize] = image_processing.zernikeProjectionGray(images[batchSize*batch:batchSize*batch + batchSize], numCoeffs=numCoeffs, win=win) # np.save(gtDir + 'zernike_numCoeffs' + str(numCoeffs) + '_win' + str(win) + synthPrefix + '.npy', trainZernikeCoeffs) # trainZernikeCoeffs = trainZernikeCoeffs[trainSet] # # trainHogfeatures = hogfeatures[trainSet] # trainIllumfeatures = illumfeatures[trainSet] parameterTrainSet = set(['azimuthsRF', 'elevationsRF', 'vcolorsRF']) # parameterTrainSet = set(['vcolorsRF', 'spherical_harmonicsRF']) # parameterTrainSet = set(['spherical_harmonicsZernike']) # parameterTrainSet = set(['spherical_harmonicsZernike']) # parameterTrainSet = set(['spherical_harmonicsNN']) parameterTrainSet = set(['appearanceAndLightNN']) parameterTrainSet = set(['appearanceNN']) parameterTrainSet = set(['neuralNetModelLight']) parameterTrainSet = set(['neuralNetModelShape']) parameterTrainSet = set(['poseNN']) parameterTrainSet = set(['azimuthTripletNN']) # parameterTrainSet = set(['appearanceNN']) # parameterTrainSet = set(['poseNN']) print("Training recognition models.") if 'azimuthTripletNN' in parameterTrainSet: modelType = 'cnn_pose_embedding' network = lasagne_nn.load_network(modelType=modelType, param_values=[], imgSize=75) print("Training NN Pose Components") validRatio = 0.9 trainValSet = np.arange(len(trainSet))[:np.uint(len(trainSet)*validRatio)] validSet = np.arange(len(trainSet))[np.uint(len(trainSet)*validRatio)::] # modelPath = experimentDir + 'neuralNetModelRelSHComponents.npz' param_values = [] fineTune = False pretrainedExperimentDir = 'experiments/train3_test/' if fineTune: pretrainedModelFile = pretrainedExperimentDir + 'neuralNetModelAzimuthTriplet.pickle' with open(pretrainedModelFile, 'rb') as pfile: neuralNetTripletAzimuth = pickle.load(pfile) meanImage = neuralNetTripletAzimuth['mean'] # ipdb.set_trace() modelType = neuralNetTripletAzimuth['type'] param_values = neuralNetTripletAzimuth['params'] else: meanImage = np.mean(grayImages, axis=0) meanImage = np.zeros(grayImages.shape) modelPath=experimentDir + 'neuralNetModelAzimuthTriplet.pickle' poseGT = np.hstack([np.cos(trainAzsRel)[:,None] , np.sin(trainAzsRel)[:,None], np.cos(trainElevsGT)[:,None], np.sin(trainElevsGT)[:,None]]) grayImagesR = grayImages.reshape([grayImages.shape[0], 1, grayImages.shape[1], grayImages.shape[2]]) grayImages_t1_R = grayImages_t1.reshape([grayImages_t1.shape[0], 1, grayImages_t1.shape[1], grayImages_t1.shape[2]]) grayImages_t2_R = grayImages_t2.reshape([grayImages_t2.shape[0], 1, grayImages_t2.shape[1], grayImages_t2.shape[2]]) grayImages_75_R = grayImages_75.reshape([grayImages_75.shape[0], 1, grayImages_75.shape[1], grayImages_75.shape[2]]) grayImages_t1_75_R = grayImages_t1_75.reshape([grayImages_t1_75.shape[0], 1, grayImages_t1_75.shape[1], grayImages_t1_75.shape[2]]) grayImages_t2_75_R = grayImages_t2_75.reshape([grayImages_t2_75.shape[0], 1, grayImages_t2_75.shape[1], grayImages_t2_75.shape[2]]) tripletsModel = lasagne_nn.train_triplets_h5(grayImages_75_R, grayImages_t1_75_R, grayImages_t2_75_R, len(trainValSet), meanImage=meanImage, network=network, modelType=modelType, num_epochs=150, saveModelAtEpoch=True, modelPath=modelPath, param_values=param_values) # poseNNmodel = lasagne_nn.train_nn(grayImages, trainSet, validSet, len(trainValSet), poseGT[trainValSet].astype(np.float32), poseGT[validSet].astype(np.float32), meanImage=meanImage, network=network, modelType=modelType, num_epochs=10, saveModelAtEpoch=True, modelPath=modelPath, param_values=param_values) # np.savez(modelPath, *SHNNparams) with open(modelPath, 'wb') as pfile: pickle.dump(tripletsModel, pfile) if 'poseNN' in parameterTrainSet: modelType = 'cnn_pose' network = lasagne_nn.load_network(modelType=modelType, param_values=[]) print("Training NN Pose Components") validRatio = 0.9 trainValSet = np.arange(len(trainSet))[:np.uint(len(trainSet)*validRatio)] validSet = np.arange(len(trainSet))[np.uint(len(trainSet)*validRatio)::] # modelPath = experimentDir + 'neuralNetModelRelSHComponents.npz' # grayTrainImages = grayImages[trainValSet][:,:,:] # grayValidImages = grayImages[validSet][:,:,:] # grayTrainImages = grayTrainImages[:,None, :,:] # grayValidImages = grayValidImages[:,None, :,:] # import sys # sys.exit("NN") param_values = [] fineTune = False pretrainedExperimentDir = 'experiments/train3_test/' if fineTune: pretrainedModelFile = pretrainedExperimentDir + 'neuralNetModelPose.pickle' with open(pretrainedModelFile, 'rb') as pfile: neuralNetModelPose = pickle.load(pfile) meanImage = neuralNetModelPose['mean'] # ipdb.set_trace() modelType = neuralNetModelPose['type'] param_values = neuralNetModelPose['params'] else: meanImage = np.mean(grayImages, axis=0) modelPath=experimentDir + 'neuralNetModelPose3.pickle' poseGT = np.hstack([np.cos(trainAzsRel)[:,None] , np.sin(trainAzsRel)[:,None], np.cos(trainElevsGT)[:,None], np.sin(trainElevsGT)[:,None]]) poseNNmodel = lasagne_nn.train_nn_h5(grayImages.reshape([grayImages.shape[0],1,grayImages.shape[1],grayImages.shape[2]]), len(trainValSet), poseGT[trainValSet].astype(np.float32), poseGT[validSet].astype(np.float32), meanImage=meanImage, network=network, modelType=modelType, num_epochs=150, saveModelAtEpoch=True, modelPath=modelPath, param_values=param_values) # poseNNmodel = lasagne_nn.train_nn(grayImages, trainSet, validSet, len(trainValSet), poseGT[trainValSet].astype(np.float32), poseGT[validSet].astype(np.float32), meanImage=meanImage, network=network, modelType=modelType, num_epochs=10, saveModelAtEpoch=True, modelPath=modelPath, param_values=param_values) # np.savez(modelPath, *SHNNparams) with open(modelPath, 'wb') as pfile: pickle.dump(poseNNmodel, pfile) if 'poseNNColor' in parameterTrainSet: modelType = 'cnn_pose_color' network = lasagne_nn.load_network(modelType=modelType, param_values=[]) print("Training NN Pose Components") validRatio = 0.9 trainValSet = np.arange(len(trainSet))[:np.uint(len(trainSet)*validRatio)] validSet = np.arange(len(trainSet))[np.uint(len(trainSet)*validRatio)::] # modelPath = experimentDir + 'neuralNetModelRelSHComponents.npz' # grayTrainImages = grayImages[trainValSet][:,:,:] # grayValidImages = grayImages[validSet][:,:,:] # grayTrainImages = grayTrainImages[:,None, :,:] # grayValidImages = grayValidImages[:,None, :,:] # import sys # sys.exit("NN") param_values = [] fineTune = False pretrainedExperimentDir = 'experiments/train3_test/' if fineTune: pretrainedModelFile = pretrainedExperimentDir + 'neuralNetModelPoseColor.pickle' with open(pretrainedModelFile, 'rb') as pfile: neuralNetModelPose = pickle.load(pfile) meanImage = neuralNetModelPose['mean'] # ipdb.set_trace() modelType = neuralNetModelPose['type'] param_values = neuralNetModelPose['params'] else: meanImage = np.mean(images, axis=0) modelPath=experimentDir + 'neuralNetModelPoseColor.pickle' poseGT = np.hstack([np.cos(trainAzsRel)[:,None] , np.sin(trainAzsRel)[:,None], np.cos(trainElevsGT)[:,None], np.sin(trainElevsGT)[:,None]]) poseNNmodel = lasagne_nn.train_nn_h5(images.reshape([images.shape[0],3,images.shape[1],images.shape[2]]), len(trainValSet), poseGT[trainValSet].astype(np.float32), poseGT[validSet].astype(np.float32), meanImage=meanImage, network=network, modelType=modelType, num_epochs=150, saveModelAtEpoch=True, modelPath=modelPath, param_values=param_values) # poseNNmodel = lasagne_nn.train_nn(grayImages, trainSet, validSet, len(trainValSet), poseGT[trainValSet].astype(np.float32), poseGT[validSet].astype(np.float32), meanImage=meanImage, network=network, modelType=modelType, num_epochs=10, saveModelAtEpoch=True, modelPath=modelPath, param_values=param_values) # np.savez(modelPath, *SHNNparams) with open(modelPath, 'wb') as pfile: pickle.dump(poseNNmodel, pfile) if 'appearanceAndLightNN' in parameterTrainSet: modelType = 'cnn_appLight' network = lasagne_nn.load_network(modelType=modelType, param_values=[]) print("Training NN Appereance and Light Components") validRatio = 0.9 trainValSet = np.arange(len(trainSet))[:np.uint(len(trainSet)*validRatio)] validSet = np.arange(len(trainSet))[np.uint(len(trainSet)*validRatio)::] # modelPath = experimentDir + 'neuralNetModelRelSHComponents.npz' # grayTrainImages = grayImages[trainValSet][:,:,:] # grayValidImages = grayImages[validSet][:,:,:] # grayTrainImages = grayTrainImages[:,None, :,:] # grayValidImages = grayValidImages[:,None, :,:] # import sys # sys.exit("NN") param_values = [] fineTune = True pretrainedExperimentDir = 'experiments/train4_occlusion_10k/' if fineTune: pretrainedModelFile = pretrainedExperimentDir + 'neuralNetModelAppLight.pickle' with open(pretrainedModelFile, 'rb') as pfile: neuralNetModelAppLight = pickle.load(pfile) meanImage = neuralNetModelAppLight['mean'] # ipdb.set_trace() modelType = neuralNetModelAppLight['type'] param_values = neuralNetModelAppLight['params'] else: meanImage = np.mean(images, axis=0) modelPath=experimentDir + 'neuralNetModelAppLight.pickle' appLightGT = np.hstack([trainLightCoefficientsGTRel*trainAmbientIntensityGT[:,None] , trainVColorGT]) appLightNNmodel = lasagne_nn.train_nn_h5(images.reshape([images.shape[0],3,images.shape[1],images.shape[2]]), len(trainValSet), appLightGT[trainValSet].astype(np.float32), appLightGT[validSet].astype(np.float32), meanImage=meanImage, network=network, modelType=modelType, num_epochs=150, saveModelAtEpoch=True, modelPath=modelPath, param_values=param_values) # poseNNmodel = lasagne_nn.train_nn(grayImages, trainSet, validSet, len(trainValSet), poseGT[trainValSet].astype(np.float32), poseGT[validSet].astype(np.float32), meanImage=meanImage, network=network, modelType=modelType, num_epochs=10, saveModelAtEpoch=True, modelPath=modelPath, param_values=param_values) # np.savez(modelPath, *SHNNparams) with open(modelPath, 'wb') as pfile: pickle.dump(appLightNNmodel, pfile) if 'appearanceNN' in parameterTrainSet: modelType = 'cnn_app' network = lasagne_nn.load_network(modelType=modelType, param_values=[]) print("Training NN Appeareance Components") validRatio = 0.9 trainValSet = np.arange(len(trainSet))[:np.uint(len(trainSet)*validRatio)] validSet = np.arange(len(trainSet))[np.uint(len(trainSet)*validRatio)::] # modelPath = experimentDir + 'neuralNetModelRelSHComponents.npz' # grayTrainImages = grayImages[trainValSet][:,:,:] # grayValidImages = grayImages[validSet][:,:,:] # grayTrainImages = grayTrainImages[:,None, :,:] # grayValidImages = grayValidImages[:,None, :,:] # import sys # sys.exit("NN") param_values = [] fineTune = False pretrainedExperimentDir = 'experiments/train3_test/' if fineTune: pretrainedModelFile = pretrainedExperimentDir + 'neuralNetModelAppearance.pickle' with open(pretrainedModelFile, 'rb') as pfile: neuralNetModelAppearance = pickle.load(pfile) meanImage = neuralNetModelAppearance['mean'] # ipdb.set_trace() modelType = neuralNetModelAppearance['type'] param_values = neuralNetModelAppearance['params'] else: meanImage = np.mean(images, axis=0) modelPath=experimentDir + 'neuralNetModelAppearance.pickle' appNNmodel = lasagne_nn.train_nn_h5(images.reshape([images.shape[0],3,images.shape[1],images.shape[2]]), len(trainValSet), trainVColorGT[trainValSet].astype(np.float32), trainVColorGT[validSet].astype(np.float32), meanImage=meanImage, network=network, modelType=modelType, num_epochs=150, saveModelAtEpoch=True, modelPath=modelPath, param_values=param_values) # poseNNmodel = lasagne_nn.train_nn(grayImages, trainSet, validSet, len(trainValSet), poseGT[trainValSet].astype(np.float32), poseGT[validSet].astype(np.float32), meanImage=meanImage, network=network, modelType=modelType, num_epochs=10, saveModelAtEpoch=True, modelPath=modelPath, param_values=param_values) # np.savez(modelPath, *SHNNparams) with open(modelPath, 'wb') as pfile: pickle.dump(appNNmodel, pfile) if 'maskNN' in parameterTrainSet: modelType = 'cnn_mask_large' network = lasagne_nn.load_network(modelType=modelType, param_values=[]) print("Training NN Appeareance Components") validRatio = 0.9 trainValSet = np.arange(len(trainSet))[:np.uint(len(trainSet)*validRatio)] validSet = np.arange(len(trainSet))[np.uint(len(trainSet)*validRatio)::] # modelPath = experimentDir + 'neuralNetModelRelSHComponents.npz' # grayTrainImages = grayImages[trainValSet][:,:,:] # grayValidImages = grayImages[validSet][:,:,:] # grayTrainImages = grayTrainImages[:,None, :,:] # grayValidImages = grayValidImages[:,None, :,:] # import sys # sys.exit("NN") param_values = [] fineTune = True pretrainedExperimentDir = experimentDir if fineTune: pretrainedModelFile = pretrainedExperimentDir + 'neuralNetModelMaskLarge.pickle' with open(pretrainedModelFile, 'rb') as pfile: neuralNetModelMask = pickle.load(pfile) meanImage = neuralNetModelMask['mean'] # ipdb.set_trace() modelType = neuralNetModelMask['type'] param_values = neuralNetModelMask['params'] else: meanImage = np.mean(images, axis=0) modelPath=experimentDir + 'neuralNetModelMaskLarge.pickle' # masksGT = masksGT.reshape([-1, 150,150]) rsMasksGt = np.zeros([len(masksGT), 50,50]) for mask_i, mask in enumerate(masksGT): rsMasksGt[mask_i] = skimage.transform.resize(mask, [50,50]) meanImage = np.mean(images, axis=0) rsMasksGt = rsMasksGt.reshape([rsMasksGt.shape[0], 50*50]) maskNNmodel = lasagne_nn.train_nn_h5(images.reshape([images.shape[0],3,images.shape[1],images.shape[2]]), len(trainValSet), rsMasksGt[trainValSet].astype(np.float32), rsMasksGt[validSet].astype(np.float32), meanImage=meanImage, network=network, modelType=modelType, num_epochs=150, saveModelAtEpoch=True, modelPath=modelPath, param_values=param_values) # poseNNmodel = lasagne_nn.train_nn(grayImages, trainSet, validSet, len(trainValSet), poseGT[trainValSet].astype(np.float32), poseGT[validSet].astype(np.float32), meanImage=meanImage, network=network, modelType=modelType, num_epochs=10, saveModelAtEpoch=True, modelPath=modelPath, param_values=param_values) # np.savez(modelPath, *SHNNparams) with open(modelPath, 'wb') as pfile: pickle.dump(maskNNmodel, pfile) if 'azimuthsRF' in parameterTrainSet: print("Training RFs Cos Azs") randForestModelCosAzs = recognition_models.trainRandomForest(trainHogfeatures, np.cos(trainAzsRel)) trainedModel = {'randForestModelCosAzs':randForestModelCosAzs} with open(experimentDir + 'randForestModelCosAzs05.pickle', 'wb') as pfile: pickle.dump(trainedModel, pfile) print("Training RFs Sin Azs") randForestModelSinAzs = recognition_models.trainRandomForest(trainHogfeatures, np.sin(trainAzsRel)) trainedModel = {'randForestModelSinAzs':randForestModelSinAzs} with open(experimentDir + 'randForestModelSinAzs05.pickle', 'wb') as pfile: pickle.dump(trainedModel, pfile) if 'elevationsRF' in parameterTrainSet: print("Training RFs Cos Elevs") randForestModelCosElevs = recognition_models.trainRandomForest(trainHogfeatures, np.cos(trainElevsGT)) trainedModel = {'randForestModelCosElevs':randForestModelCosElevs} with open(experimentDir + 'randForestModelCosElevs05.pickle', 'wb') as pfile: pickle.dump(trainedModel, pfile) print("Training RFs Sin Elevs") randForestModelSinElevs = recognition_models.trainRandomForest(trainHogfeatures, np.sin(trainElevsGT)) trainedModel = {'randForestModelSinElevs':randForestModelSinElevs} with open(experimentDir + 'randForestModelSinElevs05.pickle', 'wb') as pfile: pickle.dump(trainedModel, pfile) if 'neuralNetModelLight' in parameterTrainSet: modelType = 'cnn_light' network = lasagne_nn.load_network(modelType=modelType, param_values=[]) print("Training NN SH Components") validRatio = 0.9 trainValSet = np.arange(len(trainSet))[:np.uint(len(trainSet)*validRatio)] validSet = np.arange(len(trainSet))[np.uint(len(trainSet)*validRatio)::] # modelPath = experimentDir + 'neuralNetModelRelSHComponents.npz' param_values = [] fineTune = False pretrainedExperimentDir = 'experiments/train3_test/' if fineTune: pretrainedModelFile = pretrainedExperimentDir + 'neuralNetModelLight.pickle' with open(pretrainedModelFile, 'rb') as pfile: neuralNetModelLight = pickle.load(pfile) meanImage = neuralNetModelLight['mean'] # ipdb.set_trace() modelType = neuralNetModelLight['type'] param_values = neuralNetModelLight['params'] else: meanImage = np.mean(images, axis=0) lightGT = trainLightCoefficientsGTRel*trainAmbientIntensityGT[:,None] modelPath=experimentDir + 'neuralNetModelLight.pickle' lightNNmodel = lasagne_nn.train_nn_h5(images.reshape([images.shape[0],3,images.shape[1],images.shape[2]]), len(trainValSet), lightGT[trainValSet].astype(np.float32), lightGT[validSet].astype(np.float32), meanImage=meanImage, network=network, modelType=modelType, num_epochs=150, saveModelAtEpoch=True, modelPath=modelPath, param_values=param_values) # np.savez(modelPath, *SHNNparams) with open(modelPath, 'wb') as pfile: pickle.dump(lightNNmodel, pfile) if 'neuralNetModelShape' in parameterTrainSet: modelType = 'cnn_shape' network = lasagne_nn.load_network(modelType=modelType, param_values=[]) print("Training NN Shape Components") validRatio = 0.9 trainValSet = np.arange(len(trainSet))[:np.uint(len(trainSet)*validRatio)] validSet = np.arange(len(trainSet))[np.uint(len(trainSet)*validRatio)::] # modelPath = experimentDir + 'neuralNetModelRelSHComponents.npz' param_values = [] fineTune = True pretrainedExperimentDir = 'experiments/train4_occlusion_shapemodel_10k/' if fineTune: pretrainedModelFile = pretrainedExperimentDir + 'neuralNetModelShape.pickle' with open(pretrainedModelFile, 'rb') as pfile: neuralNetModelShape = pickle.load(pfile) meanImage = neuralNetModelShape['mean'] # ipdb.set_trace() modelType = neuralNetModelShape['type'] param_values = neuralNetModelShape['params'] else: meanImage = np.mean(grayImages, axis=0) modelPath=experimentDir + 'neuralNetModelShape.pickle' shapeNNmodel = lasagne_nn.train_nn_h5(grayImages.reshape([grayImages.shape[0],1,grayImages.shape[1],grayImages.shape[2]]), len(trainValSet), trainShapeModelCoeffsGT[trainValSet].astype(np.float32), trainShapeModelCoeffsGT[validSet].astype(np.float32), meanImage=meanImage, network=network, modelType=modelType, num_epochs=150, saveModelAtEpoch=True, modelPath=modelPath, param_values=param_values) # np.savez(modelPath, *SHNNparams) with open(modelPath, 'wb') as pfile: pickle.dump(shapeNNmodel, pfile) if 'spherical_harmonicsZernike' in parameterTrainSet: print("Training on Zernike features") # linRegModelZernikeSH = recognition_models.trainLinearRegression(trainZernikeCoeffs,dataLightCoefficientsGTRel * dataAmbientIntensityGT[:,None]) # with open(experimentDir + 'linRegModelZernike' + str(numCoeffs) +'_win' + str(win) + '.pickle', 'wb') as pfile: # pickle.dump(linRegModelZernikeSH, pfile) trainZernikeCoeffs[trainZernikeCoeffs >= 1000] = 35 randForestModelRelZernikeSH = recognition_models.trainRandomForest(trainZernikeCoeffs, trainLightCoefficientsGTRel * trainAmbientIntensityGT[:,None]) with open(experimentDir + 'randomForestModelZernike400' + str(numCoeffs) + '_win' + str(win) + '.pickle', 'wb') as pfile: pickle.dump(randForestModelRelZernikeSH, pfile) if 'vcolorsRF' in parameterTrainSet: print("Training RF on Vertex Colors") numTrainSet = images.shape[0] colorWindow = 30 image = images[0] croppedImages = images[:,image.shape[0]/2-colorWindow:image.shape[0]/2+colorWindow,image.shape[1]/2-colorWindow:image.shape[1]/2+colorWindow,:] randForestModelVColor = recognition_models.trainRandomForest(croppedImages.reshape([numTrainSet,-1]), trainVColorGT) with open(experimentDir + 'randForestModelVColor05.pickle', 'wb') as pfile: pickle.dump(randForestModelVColor, pfile) if 'vcolorsLR' in parameterTrainSet: print("Training LR on Vertex Colors") numTrainSet = images.shape[0] colorWindow = 30 image = images[0] croppedImages = images[:,image.shape[0]/2-colorWindow:image.shape[0]/2+colorWindow,image.shape[1]/2-colorWindow:image.shape[1]/2+colorWindow,:] linRegModelVColor = recognition_models.trainLinearRegression(croppedImages.reshape([numTrainSet,-1]), trainVColorGT) with open(experimentDir + 'linearRegressionModelVColor.pickle', 'wb') as pfile: pickle.dump(linRegModelVColor, pfile) # # imagesStack = np.vstack([image.reshape([1,-1]) for image in images]) # randForestModelLightIntensity = recognition_models.trainRandomForest(imagesStack, trainLightIntensitiesGT) # # # # # print("Training LR") # # # linRegModelCosAzs = recognition_models.trainLinearRegression(hogfeatures, np.cos(trainAzsGT)) # # # linRegModelSinAzs = recognition_models.trainLinearRegression(hogfeatures, np.sin(trainAzsGT)) # # # linRegModelCosElevs = recognition_models.trainLinearRegression(hogfeatures, np.cos(trainElevsGT)) # # # linRegModelSinElevs = recognition_models.trainLinearRegression(hogfeatures, np.sin(trainElevsGT)) # # # # print("Finished training recognition models.")
28,634
44.524642
730
py
inversegraphics
inversegraphics-master/zernike.py
""" @file py102-example2-zernike.py @brief Fitting a surface in Python example for Python 102 lecture @author Tim van Werkhoven (t.i.m.vanwerkhoven@gmail.com) @url http://python101.vanwerkhoven.org @date 20111012 Created by Tim van Werkhoven (t.i.m.vanwerkhoven@xs4all.nl) on 2011-10-12 Copyright (c) 2011 Tim van Werkhoven. All rights reserved. This file is licensed under the Creative Commons Attribution-Share Alike license versions 3.0 or higher, see http://creativecommons.org/licenses/by-sa/3.0/ """ ### Libraries import numpy as N from scipy.misc import factorial as fac ### Init functions def zernike_rad(m, n, rho): """ Calculate the radial component of Zernike polynomial (m, n) given a grid of radial coordinates rho. """ if (n < 0 or m < 0 or abs(m) > n): raise ValueError if ((n-m) % 2): return rho*0.0 pre_fac = lambda k: (-1.0)**k * fac(n-k) / ( fac(k) * fac( (n+m)/2.0 - k ) * fac( (n-m)/2.0 - k ) ) return sum(pre_fac(k) * rho**(n-2.0*k) for k in range(int((n-m)/2+1))) def zernike(m, n, rho, phi): """ Calculate Zernike polynomial (m, n) given a grid of radial coordinates rho and azimuthal coordinates phi. """ if (m > 0): return zernike_rad(m, n, rho) * N.cos(m * phi) if (m < 0): return zernike_rad(-m, n, rho) * N.sin(-m * phi) return zernike_rad(0, n, rho) def zernikel(j, rho, phi): """ Calculate Zernike polynomial with Noll coordinate j given a grid of radial coordinates rho and azimuthal coordinates phi. """ n = 0 while (j > n): n += 1 j -= n m = -n+2*j return zernike(m, n, rho, phi)
1,649
27.448276
103
py
inversegraphics
inversegraphics-master/save_exr_images.py
#!/usr/bin/python import OpenEXR import Imath from PIL import Image import sys import numpy as np def exportExrImages(annotationdir, imgdir, numTeapot, frame, sceneNum, target, prefix): framestr = '{0:04d}'.format(frame) outfilename = "render" + prefix + "_obj" + str(numTeapot) + "_scene" + str(sceneNum) + "_target" + str(target) + "_" + framestr exrfile = OpenEXR.InputFile(annotationdir + outfilename + ".exr") pt = Imath.PixelType(Imath.PixelType.FLOAT) dw = exrfile.header()['dataWindow'] size = (dw.max.x - dw.min.x + 1, dw.max.y - dw.min.y + 1) quality_val = 100 # rgbf = [Image.fromstring("F", size, exrfile.channel("shNormal." + c, pt)) for c in "RGB"] # extrema = [im.getextrema() for im in rgbf] # darkest = min([lo for (lo,hi) in extrema]) # lighest = max([hi for (lo,hi) in extrema]) # scale = 255 / (lighest - darkest) # def normalize_0_255(v): # return (v * scale) + darkest # rgb8 = [im.point(normalize_0_255).convert("L") for im in rgbf] # Image.merge("RGB", rgb8).save(jpgfilename + "_normal.jpg", "JPEG") rgbf = [Image.fromstring("F", size, exrfile.channel("RenderLayer.Combined." + c, pt)) for c in "RGB"] pix = [np.array(im) for im in rgbf] pix[0][pix[0]>1] = 1 pix[1][pix[1]>1] = 1 pix[2][pix[2]>1] = 1 pix[0] = pix[0]*255 pix[1] = pix[1]*255 pix[2] = pix[2]*255 imgr = Image.fromarray(pix[0].astype('uint8')) imgg = Image.fromarray(pix[1].astype('uint8')) imgb = Image.fromarray(pix[2].astype('uint8')) finalimg = Image.merge("RGB", (imgr, imgg, imgb)) # finalimg.save(imgdir + outfilename + ".png", "PNG", quality=quality_val) distancestr = exrfile.channel('RenderLayer.IndexOB.X', pt) distance = Image.fromstring("F", size, distancestr) shapeIndexstr = exrfile.channel('RenderLayer.IndexOB.X', pt) shapeIndex = Image.fromstring("F", size, shapeIndexstr) segment = np.array(shapeIndex) sumComplete = np.sum(segment) segmentimg = Image.fromarray(segment.astype('uint8')*255) segmentimg.save(imgdir + outfilename + "_segment.png", "PNG", quality=quality_val) singlefile = "render" + prefix + "_obj" + str(numTeapot) + "_scene" + str(sceneNum) + "_target" + str(target) + "_single_" + framestr exrfile = OpenEXR.InputFile(annotationdir + singlefile + ".exr") rgbf = [Image.fromstring("F", size, exrfile.channel("RenderLayer.001.Combined." + c, pt)) for c in "RGB"] pix = [np.array(im) for im in rgbf] pix[0][pix[0]>1] = 1 pix[1][pix[1]>1] = 1 pix[2][pix[2]>1] = 1 pix[0] = pix[0]*255 pix[1] = pix[1]*255 pix[2] = pix[2]*255 imgr = Image.fromarray(pix[0].astype('uint8')) imgg = Image.fromarray(pix[1].astype('uint8')) imgb = Image.fromarray(pix[2].astype('uint8')) finalimg = Image.merge("RGB", (imgr, imgg, imgb)) finalimg.save(imgdir + singlefile + ".png", "PNG", quality=quality_val) shapeIndexstrsingle = exrfile.channel('RenderLayer.001.IndexOB.X', pt) shapeIndexsingle = Image.fromstring("F", size, shapeIndexstrsingle) segmentsingle = np.array(shapeIndexsingle) segmentimgsingle = Image.fromarray(segmentsingle.astype('uint8')*255) segmentimgsingle.save(imgdir + singlefile + "_segment.png", "PNG", quality=quality_val) sumSingle = np.sum(segmentsingle) print "Sum Complete " + str(sumComplete) print "Sum Single " + str(sumSingle) with open(annotationdir + 'occlusions' + ".txt", "a") as myfile: myfile.write(str(numTeapot) + ' ' + str(frame) + ' ' + str(sceneNum) + " " + str(target) + " " + str(sumComplete/sumSingle) + ' ' + prefix + "\n") return
3,659
40.590909
154
py
inversegraphics
inversegraphics-master/diffrender_demo.py
__author__ = 'pol' import matplotlib matplotlib.use('Qt4Agg') import bpy import scene_io_utils import mathutils from math import radians import timeit import time import opendr import chumpy as ch import geometry import image_processing import numpy as np import cv2 from blender_utils import * import glfw import generative_models import matplotlib.pyplot as plt from opendr_utils import * import OpenGL.GL as GL import light_probes import imageio from utils import * plt.ion() #__GL_THREADED_OPTIMIZATIONS #Main script options:r useBlender = False loadBlenderSceneFile = False groundTruthBlender = False useShapeModel = True datasetGroundtruth = False syntheticGroundtruth = True useCycles = False demoMode = False showSubplots = False unpackModelsFromBlender = False unpackSceneFromBlender = False loadSavedSH = False useGTasBackground = False refreshWhileMinimizing = True computePerformance = False savePerformance = False glModes = ['glfw','mesa'] glMode = glModes[0] sphericalMap = False np.random.seed(1) width, height = (150, 150) win = -1 if glMode == 'glfw': #Initialize base GLFW context for the Demo and to share context among all renderers. glfw.init() glfw.window_hint(glfw.CONTEXT_VERSION_MAJOR, 3) glfw.window_hint(glfw.CONTEXT_VERSION_MINOR, 3) # glfw.window_hint(glfw.OPENGL_FORWARD_COMPAT, GL.GL_TRUE) glfw.window_hint(glfw.OPENGL_PROFILE, glfw.OPENGL_CORE_PROFILE) glfw.window_hint(glfw.DEPTH_BITS,32) if demoMode: glfw.window_hint(glfw.VISIBLE, GL.GL_TRUE) else: glfw.window_hint(glfw.VISIBLE, GL.GL_FALSE) win = glfw.create_window(width, height, "Demo", None, None) glfw.make_context_current(win) angle = 60 * 180 / numpy.pi clip_start = 0.05 clip_end = 10 frustum = {'near': clip_start, 'far': clip_end, 'width': width, 'height': height} camDistance = 0.4 gtPrefix = 'train4_occlusion_shapemodel' gtDirPref = 'train4_occlusion_shapemodel' gtDir = 'groundtruth/' + gtDirPref + '/' groundTruthFilename = gtDir + 'groundTruth.h5' gtDataFile = h5py.File(groundTruthFilename, 'r') groundTruth = gtDataFile[gtPrefix] dataAzsGT = groundTruth['trainAzsGT'] dataObjAzsGT = groundTruth['trainObjAzsGT'] dataElevsGT = groundTruth['trainElevsGT'] # dataLightAzsGT = groundTruth['trainLightAzsGT'] # dataLightElevsGT = groundTruth['trainLightElevsGT'] # dataLightIntensitiesGT = groundTruth['trainLightIntensities'] dataVColorGT = groundTruth['trainVColorGT'] dataScenes = groundTruth['trainScenes'] dataTeapotIds = groundTruth['trainTeapotIds'] dataEnvMaps = groundTruth['trainEnvMaps'] dataOcclusions = groundTruth['trainOcclusions'] dataTargetIndices = groundTruth['trainTargetIndices'] dataIds = groundTruth['trainIds'] dataLightCoefficientsGT = groundTruth['trainLightCoefficientsGT'] dataLightCoefficientsGTRel = groundTruth['trainLightCoefficientsGTRel'] dataAmbientIntensityGT = groundTruth['trainAmbientIntensityGT'] dataEnvMapPhiOffsets = groundTruth['trainEnvMapPhiOffsets'] dataShapeModelCoeffsGT = groundTruth['trainShapeModelCoeffsGT'] readDataId = 10045 readDataId = 6102 import shape_model if useShapeModel: import shape_model #%% Load data filePath = 'data/teapotModel.pkl' teapotModel = shape_model.loadObject(filePath) faces = teapotModel['faces'] #%% Sample random shape Params latentDim = np.shape(teapotModel['ppcaW'])[1] shapeParams = np.random.randn(latentDim) shapeParams = dataShapeModelCoeffsGT[readDataId] chShapeParams = ch.Ch(shapeParams) meshLinearTransform=teapotModel['meshLinearTransform'] W=teapotModel['ppcaW'] b=teapotModel['ppcaB'] teapots = [line.strip() for line in open('teapots.txt')] renderTeapotsList = np.arange(len(teapots))[0:1] sceneNumber = dataScenes[readDataId] replaceableScenesFile = '../databaseFull/fields/scene_replaceables_backup.txt' sceneIdx = scene_io_utils.getSceneIdx(sceneNumber, replaceableScenesFile) sceneNumber, sceneFileName, instances, roomName, roomInstanceNum, targetIndices, targetPositions = scene_io_utils.getSceneInformation(sceneIdx, replaceableScenesFile) sceneDicFile = 'data/scene' + str(sceneNumber) + '.pickle' targetParentIdx = 0 targetIndex = targetIndices[targetParentIdx] targetIndex = dataTargetIndices[readDataId] for targetParentIdx, targetParentIndex in enumerate(targetIndices): if targetParentIndex == targetIndex: #Now targetParentIdx has the right idx of the list of parent indices. break targetParentPosition = targetPositions[targetParentIdx] targetPosition = targetParentPosition if useBlender and not loadBlenderSceneFile: scene = scene_io_utils.loadBlenderScene(sceneIdx, replaceableScenesFile) scene_io_utils.setupScene(scene, roomInstanceNum, scene.world, scene.camera, width, height, 16, useCycles, False) scene.update() targetPosition = np.array(targetPosition) #Save barebones scene. elif useBlender and loadBlenderSceneFile: scene_io_utils.loadSceneBlendData(sceneIdx, replaceableScenesFile) scene = bpy.data.scenes['Main Scene'] scene.render.resolution_x = width #perhaps set resolution in code scene.render.resolution_y = height scene.render.tile_x = height/2 scene.render.tile_y = width/2 scene.cycles.samples = 1024 scene.sequencer_colorspace_settings.name = 'Linear' scene.display_settings.display_device = 'None' bpy.context.screen.scene = scene tex_srgb2lin = True if unpackSceneFromBlender: v, f_list, vc, vn, uv, haveTextures_list, textures_list = scene_io_utils.unpackBlenderScene(scene, sceneDicFile, True) else: v, f_list, vc, vn, uv, haveTextures_list, textures_list = scene_io_utils.loadSavedScene(sceneDicFile, tex_srgb2lin) removeObjectData(len(v) -1 - targetIndex, v, f_list, vc, vn, uv, haveTextures_list, textures_list) targetModels = [] if useBlender and not loadBlenderSceneFile: [targetScenes, targetModels, transformations] = scene_io_utils.loadTargetModels(renderTeapotsList) elif useBlender: teapots = [line.strip() for line in open('teapots.txt')] selection = [ teapots[i] for i in renderTeapotsList] scene_io_utils.loadTargetsBlendData() for teapotIdx, teapotName in enumerate(selection): targetModels = targetModels + [bpy.data.scenes[teapotName[0:63]].objects['teapotInstance' + str(renderTeapotsList[teapotIdx])]] v_teapots, f_list_teapots, vc_teapots, vn_teapots, uv_teapots, haveTextures_list_teapots, textures_list_teapots, vflat, varray, center_teapots = scene_io_utils.loadTeapotsOpenDRData(renderTeapotsList, useBlender, unpackModelsFromBlender, targetModels) mugs = [line.strip() for line in open('mugs.txt')] renderMugsList = np.arange(len(mugs))[:] # [mugScenes, mugModels, transformations] = scene_io_utils.loadMugsModels(renderMugsList) v_mugs, f_list_mugs, vc_mugs, vn_mugs, uv_mugs, haveTextures_list_mugs, textures_list_mugs, vflat, varray, center_mugs = scene_io_utils.loadMugsOpenDRData(renderMugsList, False, False, None) azimuth = np.pi chCosAz = ch.Ch([np.cos(azimuth)]) chSinAz = ch.Ch([np.sin(azimuth)]) chAz = 2*ch.arctan(chSinAz/(ch.sqrt(chCosAz**2 + chSinAz**2) + chCosAz)) elevation = 0 chLogCosEl = ch.Ch(np.log(np.cos(elevation))) chLogSinEl = ch.Ch(np.log(np.sin(elevation))) chEl = 2*ch.arctan(ch.exp(chLogSinEl)/(ch.sqrt(ch.exp(chLogCosEl)**2 + ch.exp(chLogSinEl)**2) + ch.exp(chLogCosEl))) chDist = ch.Ch([camDistance]) chPointLightIntensity = ch.Ch([1]) chPointLightIntensityGT = ch.Ch([1]) chLightAz = ch.Ch([0.0]) chLightEl = ch.Ch([0]) chLightDist = ch.Ch([0.5]) chLightDistGT = ch.Ch([0.5]) chLightAzGT = ch.Ch([0.0]) chLightElGT = ch.Ch([np.pi/4]) ligthTransf = computeHemisphereTransformation(chLightAz, chLightEl, chLightDist, targetPosition) ligthTransfGT = computeHemisphereTransformation(chLightAzGT, chLightElGT, chLightDistGT, targetPosition) lightPos = ch.dot(ligthTransf, ch.Ch([0.,0.,0.,1.]))[0:3] lightPos = ch.Ch([targetPosition[0]+0.5,targetPosition[1],targetPosition[2] + 0.5]) lightPosGT = ch.dot(ligthTransfGT, ch.Ch([0.,0.,0.,1.]))[0:3] chGlobalConstant = ch.Ch([0.5]) chGlobalConstantGT = ch.Ch([0.5]) light_color = ch.ones(3)*chPointLightIntensity light_colorGT = ch.ones(3)*chPointLightIntensityGT chVColors = ch.Ch(dataVColorGT[readDataId]) chVColorsGT = ch.Ch(dataVColorGT[readDataId]) shCoefficientsFile = 'data/sceneSH' + str(sceneIdx) + '.pickle' clampedCosCoeffs = clampedCosineCoefficients() # envMapFilename = 'data/hdr/dataset/TropicalRuins_3k.hdr' SHFilename = 'data/LightSHCoefficients.pickle' with open(SHFilename, 'rb') as pfile: envMapDic = pickle.load(pfile) hdritems = list(envMapDic.items()) hdrstorender = [] phiOffsets = [0, np.pi/2, np.pi, 3*np.pi/2] for hdrFile, hdrValues in hdritems: hdridx = hdrValues[0] envMapCoeffs = hdrValues[1] if hdridx == dataEnvMaps[readDataId]: break envMapFilename = hdrFile envMapTexture = np.array(imageio.imread(envMapFilename))[:,:,0:3] envMapGray = 0.3*envMapTexture[:,:,0] + 0.59*envMapTexture[:,:,1] + 0.11*envMapTexture[:,:,2] envMapGrayMean = np.mean(envMapGray, axis=(0,1)) # if sphericalMap: # envMapTexture, envMapMean = light_probes.processSphericalEnvironmentMap(envMapTexture) # envMapCoeffsGT = light_probes.getEnvironmentMapCoefficients(envMapTexture, 1, 0, 'spherical') # else: # envMapMean = np.mean(envMapTexture,axis=(0,1))[None,None,:] # envMapGray = 0.3*envMapTexture[:,:,0] + 0.59*envMapTexture[:,:,1] + 0.11*envMapTexture[:,:,2] # envMapGrayMean = np.mean(envMapGray, axis=(0,1)) # envMapTexture = envMapTexture/envMapGrayMean # # # envMapTexture = 4*np.pi*envMapTexture/np.sum(envMapTexture, axis=(0,1)) # envMapCoeffsGT = light_probes.getEnvironmentMapCoefficients(envMapTexture, 1, 0, 'equirectangular') # pEnvMap = SHProjection(envMapTexture, envMapCoeffsGT) # approxProjection = np.sum(pEnvMap, axis=3) # imageio.imwrite("tmp.exr", approxProjection) envMapCoeffsGT = ch.Ch(envMapCoeffs) # rotation = ch.Ch([0.0]) phiOffsetGT = ch.Ch(dataEnvMapPhiOffsets[readDataId]) phiOffset = ch.Ch(dataEnvMapPhiOffsets[readDataId]) chObjAzGT = ch.Ch(dataObjAzsGT[readDataId]) # chObjAzGT[:] = 0 chAzGT = ch.Ch(dataAzsGT[readDataId]) # chAzGT[:] = 0 chAzRelGT = chAzGT - chObjAzGT chElGT = ch.Ch(dataElevsGT[readDataId]) # chElGT[:] = 0 chDistGT = ch.Ch([camDistance]) totalOffsetGT = phiOffsetGT + chObjAzGT chAmbientIntensityGT = ch.Ch(dataAmbientIntensityGT[readDataId]) # chAmbientIntensityGT = ch.Ch([0.125]) shCoeffsRGBGT = ch.dot(light_probes.chSphericalHarmonicsZRotation(totalOffsetGT), envMapCoeffsGT[[0,3,2,1,4,5,6,7,8]])[[0,3,2,1,4,5,6,7,8]] shCoeffsRGBGTRel = ch.dot(light_probes.chSphericalHarmonicsZRotation(phiOffsetGT), envMapCoeffsGT[[0,3,2,1,4,5,6,7,8]])[[0,3,2,1,4,5,6,7,8]] chShCoeffsGT = 0.3*shCoeffsRGBGT[:,0] + 0.59*shCoeffsRGBGT[:,1] + 0.11*shCoeffsRGBGT[:,2] chShCoeffsGTRel = 0.3*shCoeffsRGBGTRel[:,0] + 0.59*shCoeffsRGBGTRel[:,1] + 0.11*shCoeffsRGBGTRel[:,2] chAmbientSHGT = chShCoeffsGT.ravel() * chAmbientIntensityGT * clampedCosCoeffs chAmbientSHGTRel = chShCoeffsGTRel.ravel() * chAmbientIntensityGT * clampedCosCoeffs chLightRadGT = ch.Ch([0.1]) chLightDistGT = ch.Ch([0.5]) chLightIntensityGT = ch.Ch([0]) chLightAzGT = ch.Ch([0]) chLightElGT = ch.Ch([0]) angleGT = ch.arcsin(chLightRadGT/chLightDistGT) zGT = chZonalHarmonics(angleGT) shDirLightGTOriginal = np.array(chZonalToSphericalHarmonics(zGT, np.pi/2 - chLightElGT, chLightAzGT - np.pi/2).r[:]).copy() shDirLightGT = ch.Ch(shDirLightGTOriginal.copy()) chComponentGTOriginal = ch.array(np.array(chAmbientSHGT + shDirLightGT*chLightIntensityGT * clampedCosCoeffs).copy()) # chComponentGT = chAmbientSHGT + shDirLightGT*chLightIntensityGT * clampedCosCoeffs chComponentGT = chAmbientSHGT # chComponentGT = ch.Ch([0.2,0,0,0,0,0,0,0,0]) chAz = ch.Ch(dataAzsGT[readDataId]) - ch.Ch(dataObjAzsGT[readDataId]) chObjAz = 0 chEl = ch.Ch(dataElevsGT[readDataId]) chAzRel = chAz - chObjAz totalOffset = phiOffset + chObjAz chAmbientIntensity = ch.Ch(dataAmbientIntensityGT[readDataId]) shCoeffsRGB = ch.dot(light_probes.chSphericalHarmonicsZRotation(totalOffset), envMapCoeffs[[0,3,2,1,4,5,6,7,8]])[[0,3,2,1,4,5,6,7,8]] shCoeffsRGBRel = ch.dot(light_probes.chSphericalHarmonicsZRotation(phiOffset), envMapCoeffs[[0,3,2,1,4,5,6,7,8]])[[0,3,2,1,4,5,6,7,8]] chShCoeffs = 0.3*shCoeffsRGB[:,0] + 0.59*shCoeffsRGB[:,1] + 0.11*shCoeffsRGB[:,2] chShCoeffs = ch.Ch(0.3*shCoeffsRGB.r[:,0] + 0.59*shCoeffsRGB.r[:,1] + 0.11*shCoeffsRGB.r[:,2]) chShCoeffsRel = 0.3*shCoeffsRGBRel[:,0] + 0.59*shCoeffsRGBRel[:,1] + 0.11*shCoeffsRGBRel[:,2] chAmbientSH = chShCoeffs.ravel() * chAmbientIntensity * clampedCosCoeffs chLightRad = ch.Ch([0.1]) chLightDist = ch.Ch([0.5]) chLightIntensity = ch.Ch([0]) chLightAz = ch.Ch([np.pi/2]) chLightEl = ch.Ch([0]) angle = ch.arcsin(chLightRad/chLightDist) z = chZonalHarmonics(angle) shDirLight = chZonalToSphericalHarmonics(z, np.pi/2 - chLightEl, chLightAz - np.pi/2) * clampedCosCoeffs chComponent = chAmbientSH + shDirLight*chLightIntensity # chComponent = chComponentGT if useBlender: addEnvironmentMapWorld(scene) updateEnviornmentMap(envMapFilename, scene) setEnviornmentMapStrength(1./envMapGrayMean, scene) rotateEnviornmentMap(-totalOffset, scene) chDisplacement = ch.Ch([0.0, 0.0,0.0]) chDisplacementGT = ch.Ch([0.0,0.0,0.0]) chScale = ch.Ch([1.0,1.0,1.0]) chScaleGT = ch.Ch([1, 1.,1.]) # vcch[0] = np.ones_like(vcflat[0])*chVColorsGT.reshape([1,3]) renderer_teapots = [] blender_teapots = [] teapots = [line.strip() for line in open('teapots.txt')] selection = [ teapots[i] for i in renderTeapotsList] scene_io_utils.loadTargetsBlendData() for teapotIdx, teapotName in enumerate(selection): teapot = bpy.data.scenes[teapotName[0:63]].objects['teapotInstance' + str(renderTeapotsList[teapotIdx])] teapot.layers[1] = True teapot.layers[2] = True targetModels = targetModels + [teapot] blender_teapots = blender_teapots + [teapot] setObjectDiffuseColor(teapot, chVColorsGT.r.copy()) # for teapot_i in range(len(renderTeapotsList)): # if useBlender: # teapot = blender_teapots[teapot_i] # teapot.matrix_world = mathutils.Matrix.Translation(targetPosition) # # vmod = v_teapots[teapot_i] # fmod_list = f_list_teapots[teapot_i] # vcmod = vc_teapots[teapot_i] # vnmod = vn_teapots[teapot_i] # uvmod = uv_teapots[teapot_i] # haveTexturesmod_list = haveTextures_list_teapots[teapot_i] # texturesmod_list = textures_list_teapots[teapot_i] # centermod = center_teapots[teapot_i] # # vmod, vnmod, _ = transformObject(vmod, vnmod, chScale, np.pi/2, ch.Ch([0]), ch.Ch([0]), targetPosition) # renderer = createRendererTarget(glMode, False, chAz, chEl, chDist, centermod, vmod, vcmod, fmod_list, vnmod, light_color, chComponent, chVColors, targetPosition, chDisplacement, width,height, uvmod, haveTexturesmod_list, texturesmod_list, frustum, win ) # # renderer.msaa = True # renderer.overdraw = True # renderer.r # renderer_teapots = renderer_teapots + [renderer] if useShapeModel: shapeParams = np.random.randn(latentDim) shapeParams = dataShapeModelCoeffsGT[readDataId] chShapeParams = ch.Ch(shapeParams) # landmarksLong = ch.dot(chShapeParams,teapotModel['ppcaW'].T) + teapotModel['ppcaB'] # landmarks = landmarksLong.reshape([-1,3]) # chVertices = shape_model.chShapeParamsToVerts(landmarks, teapotModel['meshLinearTransform']) chVertices = shape_model.VerticesModel(chShapeParams =chShapeParams,meshLinearTransform=meshLinearTransform,W=W,b=b) chVertices.init() chVertices = ch.dot(geometry.RotateZ(-np.pi/2)[0:3,0:3],chVertices.T).T # teapotNormals = teapotModel['N'] # chNormals = shape_model.chShapeParamsToNormals(teapotNormals, landmarks, teapotModel['linT']) # rot = mathutils.Matrix.Rotation(radians(90), 4, 'X') # chNormals= ch.dot(np.array(rot)[0:3, 0:3], chNormals.T).T # chNormals2 = ch.array(shape_model.shapeParamsToNormals(shapeParams, teapotModel)) chNormals = shape_model.chGetNormals(chVertices, faces) smNormals = [chNormals] smFaces = [[faces]] smVColors = [chVColors*np.ones(chVertices.shape)] smUVs = ch.Ch(np.zeros([chVertices.shape[0],2])) smHaveTextures = [[False]] smTexturesList = [[None]] chVertices = chVertices - ch.mean(chVertices, axis=0) minZ = ch.min(chVertices[:,2]) chMinZ = ch.min(chVertices[:,2]) # chVertices[:,2] = chVertices[:,2] - minZ zeroZVerts = chVertices[:,2]- chMinZ chVertices = ch.hstack([chVertices[:,0:2] , zeroZVerts.reshape([-1,1])]) chVertices = chVertices*0.09 smCenter = ch.array([0,0,0.1]) smVertices = [chVertices] smFacesB = [smFaces] smVColorsB = [smVColors] smUVsB = [[smUVs]] smHaveTexturesB = [smHaveTextures] smTexturesListB = [smTexturesList] smVertices, smNormals, _ = transformObject(smVertices, smNormals, chScale, chObjAz, ch.Ch([0.0]), ch.Ch([0]), np.array([0,0,0])) v_mug = v_mugs[0][0] f_list_mug = f_list_mugs[0] vc_mug = [[np.array([1,0,0])*np.ones(v_mug[0][0].shape)]] vn_mug = vn_mugs[0][0] uv_mug = uv_mugs[0] haveTextures_list_mug = haveTextures_list_mugs[0] textures_list_mug = textures_list_mugs[0] verticesMug, normalsMug, _ = transformObject(v_mug, vn_mug, chScale, chObjAz - np.pi/2, ch.Ch([0.2]), ch.Ch([0]), np.array([0,0,0])) if False: VerticesB = [smVertices ] + [verticesMug] NormalsB = [smNormals] + [normalsMug] FacesB = smFacesB + f_list_mug VColorsB = smVColorsB + vc_mug UVsB = smUVsB + uv_mug HaveTexturesB = smHaveTexturesB + haveTextures_list_mug TexturesListB = smTexturesListB + textures_list_mug else: VerticesB = [smVertices ] NormalsB = [smNormals] FacesB = smFacesB VColorsB = smVColorsB UVsB = smUVsB HaveTexturesB = smHaveTexturesB TexturesListB = smTexturesListB # renderer = createRendererTarget(glMode, True, chAz, chEl, chDist, smCenter, [smVertices], smVColorsB, smFacesB, [smNormals], light_color, chComponent, chVColors, targetPosition, chDisplacement, width,height,smUVsB, smHaveTexturesB, smTexturesListB, frustum, win ) renderer = createRendererTarget(glMode, chAz, chEl, chDist, smCenter, VerticesB, VColorsB, FacesB, NormalsB, light_color, chComponent, chVColors, np.array([0,0,0]), chDisplacement, width,height, UVsB, HaveTexturesB, TexturesListB, frustum, win ) renderer.overdraw = True renderer.nsamples = 8 renderer.msaa = True renderer.initGL() renderer.initGLTexture() chAz2 = ch.Ch(chAz.r[:] - 0.3) chEl2 = ch.Ch(chEl.r[:]- 0.3) chVColors2 = ch.Ch(np.array([1,0,0.5])) light_color2 = light_color + np.array([0.2,-0.4,-0.4]) renderer2 = createRendererTarget(glMode, chAz2, chEl2, chDist, smCenter, VerticesB, VColorsB, FacesB, NormalsB, light_color2, chComponent, chVColors2, np.array([0,0,0]), chDisplacement, width,height, UVsB, HaveTexturesB, TexturesListB, frustum, win ) renderer2.overdraw = True renderer2.nsamples = 8 renderer2.msaa = True renderer2.initGL() renderer2.initGLTexture() # plt.imsave('errors.png', sqeRenderer.r) # plt.imsave('errorscolors.png', sqeRenderer.render_image) # # plt.imsave('renderer2.png', sqeRenderer.render_dedx) # plt.imsave('errorsdx.png', sqeRenderer.render_dedx,cmap=matplotlib.cm.coolwarm, vmin=-1, vmax=1) # plt.imsave('errorsdy.png', sqeRenderer.render_dedy,cmap=matplotlib.cm.coolwarm, vmin=-1, vmax=1) # sys.exit() # # # Funky theano stuff # import lasagne_nn # import lasagne # import theano # import theano.tensor as T # with open('experiments/train4/neuralNetModelRelSHLight.pickle', 'rb') as pfile: # neuralNetModelSHLight = pickle.load(pfile) # meanImage = neuralNetModelSHLight['mean'] # modelType = neuralNetModelSHLight['type'] # param_values = neuralNetModelSHLight['params'] # rendererGray = 0.3*renderer[:,:,0] + 0.59*renderer[:,:,1] + 0.11*renderer[:,:,2] # input = rendererGray.r[None,None, :,:] # input_var = T.tensor4('inputs') # network = lasagne_nn.build_cnn(input_var) # network_small = lasagne_nn.build_cnn_small(input_var) # lasagne.layers.set_all_param_values(network, param_values) # prediction = lasagne.layers.get_output(network) # chThFun = TheanoFunOnOpenDR(theano_input=input_var, theano_output=prediction, opendr_input=renderer, dim_output = 9) # sys.exit(0) currentTeapotModel = 0 if not useShapeModel: renderer = renderer_teapots[currentTeapotModel] # shapeParams = np.random.randn(latentDim) if useShapeModel: shapeParams = np.random.randn(latentDim) shapeParams = dataShapeModelCoeffsGT[readDataId] chShapeParamsGT = ch.Ch(shapeParams) chVerticesGT = shape_model.VerticesModel(chShapeParams =chShapeParamsGT,meshLinearTransform=meshLinearTransform,W=W,b=b) chVerticesGT.init() chVerticesGT = ch.dot(geometry.RotateZ(-np.pi/2)[0:3,0:3],chVerticesGT.T).T # chNormalsGT = shape_model.chShapeParamsToNormals(teapotModel['N'], landmarks, teapotModel['linT']) # chNormalsGT = shape_model.shapeParamsToNormals(shapeParams, teapotModel) chNormalsGT = shape_model.chGetNormals(chVerticesGT, faces) smNormalsGT = [chNormalsGT] smFacesGT = [[faces]] smVColorsGT = [chVColorsGT*np.ones(chVerticesGT.shape)] smUVsGT = [ch.Ch(np.zeros([chVerticesGT.shape[0],2]))] smHaveTexturesGT = [[False]] smTexturesListGT = [[None]] smCenterGT = ch.mean(chVerticesGT, axis=0) chVerticesGT = chVerticesGT - ch.mean(chVerticesGT, axis=0) minZ = ch.min(chVerticesGT[:,2]) chMinZ = ch.min(chVerticesGT[:,2]) zeroZVerts = chVerticesGT[:,2]- chMinZ chVerticesGT = ch.hstack([chVerticesGT[:,0:2] , zeroZVerts.reshape([-1,1])]) chVerticesGT = chVerticesGT*0.09 smCenterGT = ch.array([0,0,0.1]) smVerticesGT = [chVerticesGT] smVerticesGT, smNormalsGT, _ = transformObject(smVerticesGT, smNormalsGT, chScaleGT, chObjAzGT, ch.Ch([0]), ch.Ch([0]), targetPosition) if useShapeModel: addObjectData(v, f_list, vc, vn, uv, haveTextures_list, textures_list, smVerticesGT, smFacesGT, smVColorsGT, smNormalsGT, smUVsGT, smHaveTexturesGT, smTexturesListGT) else: addObjectData(v, f_list, vc, vn, uv, haveTextures_list, textures_list, v_teapots[currentTeapotModel][0], f_list_teapots[currentTeapotModel][0], vc_teapots[currentTeapotModel][0], vn_teapots[currentTeapotModel][0], uv_teapots[currentTeapotModel][0], haveTextures_list_teapots[currentTeapotModel][0], textures_list_teapots[currentTeapotModel][0]) center = center_teapots[currentTeapotModel] if not useShapeModel: smCenterGT = ch.array([0, 0, 0.1]) rendererGT = createRendererGT(glMode, chAzGT, chElGT, chDistGT, smCenterGT, v, vc, f_list, vn, light_colorGT, chComponentGT, chVColorsGT, targetPosition, chDisplacementGT, width,height, uv, haveTextures_list, textures_list, frustum, win ) rendererGT.useShaderErrors = False rendererGT.msaa = True rendererGT.nsamples = 8 rendererGT.overdraw = True rendererGT.initGL() rendererGT.initGLTexture() cv2.imwrite('renderergt' + str(readDataId) + '.jpeg' , 255*lin2srgb(rendererGT.r[:,:,[2,1,0]]), [int(cv2.IMWRITE_JPEG_QUALITY), 100]) ipdb.set_trace() render = renderer.r.copy() red = np.zeros(render.shape) red[:,:,0] = 1 vis_occluded = np.array(rendererGT.indices_image == 0 + 1).copy().astype(np.bool) vis_im = np.array(rendererGT.image_mesh_bool([0])).copy().astype(np.bool) occlusion = (vis_im & (1 - vis_occluded)).astype(np.bool) render[occlusion] = render[occlusion]*0.1 + red[occlusion] *0.9 cv2.imwrite('renderer' + str(readDataId) + '.jpeg' , 255*lin2srgb(render[:,:,[2,1,0]]), [int(cv2.IMWRITE_JPEG_QUALITY), 100]) ipdb.set_trace() if useGTasBackground: for teapot_i in range(len(renderTeapotsList)): renderer = renderer_teapots[teapot_i] renderer.set(background_image=rendererGT.r) currentTeapotModel = 0 if not useShapeModel: renderer = renderer_teapots[currentTeapotModel] import differentiable_renderer paramsList = [chAz, chEl] # diffRenderer = differentiable_renderer.DifferentiableRenderer(renderer=renderer, params_list=paramsList, params=ch.concatenate(paramsList)) diffRenderer = renderer vis_gt = np.array(rendererGT.indices_image!=1).copy().astype(np.bool) vis_mask = np.array(rendererGT.indices_image==1).copy().astype(np.bool) vis_im = np.array(renderer.indices_image!=1).copy().astype(np.bool) oldChAz = chAz[0].r oldChEl = chEl[0].r # Show it shapeIm = vis_gt.shape numPixels = shapeIm[0] * shapeIm[1] shapeIm3D = [vis_im.shape[0], vis_im.shape[1], 3] if useBlender: center = centerOfGeometry(teapot.dupli_group.objects, teapot.matrix_world) # addLamp(scene, center, chLightAzGT.r, chLightElGT.r, chLightDistGT, chLightIntensityGT.r) #Add ambient lighting to scene (rectangular lights at even intervals). # addAmbientLightingScene(scene, useCycles) teapot = blender_teapots[currentTeapotModel] teapotGT = blender_teapots[currentTeapotModel] placeNewTarget(scene, teapot, targetPosition) teapot.layers[1]=True # scene.layers[0] = False # scene.layers[1] = True scene.objects.unlink(scene.objects[str(targetIndex)]) placeCamera(scene.camera, -chAzGT.r[:].copy()*180/np.pi, chElGT.r[:].copy()*180/np.pi, chDistGT, center) azimuthRot = mathutils.Matrix.Rotation(chObjAzGT.r[:].copy(), 4, 'Z') original_matrix_world = teapot.matrix_world.copy() teapot.matrix_world = mathutils.Matrix.Translation(original_matrix_world.to_translation()) * azimuthRot * (mathutils.Matrix.Translation(-original_matrix_world.to_translation())) * original_matrix_world scene.update() scene.render.image_settings.file_format = 'OPEN_EXR' scene.render.filepath = 'opendr_blender.exr' # bpy.ops.file.pack_all() # bpy.ops.wm.save_as_mainfile(filepath='data/scene' + str(sceneIdx) + '_complete.blend') # scene.render.filepath = 'blender_envmap_render.exr' def imageGT(): global groundTruthBlender global rendererGT global blenderRender global datasetGroundtruth if datasetGroundtruth: return np.copy(np.array(imageDataset)).astype(np.float64) if groundTruthBlender: return blenderRender else: return np.copy(np.array(rendererGT.r)).astype(np.float64) global datasetGroundtruth if syntheticGroundtruth: imagesDir = gtDir + 'images_opendr/' else: imagesDir = gtDir + 'images/' import utils import skimage.transform image = utils.readImages(imagesDir, [readDataId], False)[0] if image.shape[0] != height or image.shape[1] != width: image = skimage.transform.resize(image, [height,width]) imageDataset = srgb2lin(image) imagegt = imageGT() chImage = ch.array(imagegt) # E_raw_simple = renderer - rendererGT negVisGT = ~vis_gt imageWhiteMask = imagegt.copy() imageWhiteMask[np.tile(negVisGT.reshape([shapeIm[0],shapeIm[1],1]),[1,1,3]).astype(np.bool)] = 1 chImageWhite = ch.Ch(imageWhiteMask) E_raw = renderer - rendererGT SE_raw = ch.sum(E_raw*E_raw, axis=2) SSqE_raw = ch.SumOfSquares(E_raw)/numPixels initialPixelStdev = 0.01 reduceVariance = False # finalPixelStdev = 0.05 stds = ch.Ch([initialPixelStdev]) variances = stds ** 2 globalPrior = ch.Ch([0.8]) negLikModel = -ch.sum(generative_models.LogGaussianModel(renderer=renderer, groundtruth=rendererGT, variances=variances))/numPixels negLikModelRobust = -ch.sum(generative_models.LogRobustModel(renderer=renderer, groundtruth=rendererGT, foregroundPrior=globalPrior, variances=variances))/numPixels modelLogLikelihoodRobustRegionCh = -ch.sum(generative_models.LogRobustModelRegion(renderer=renderer, groundtruth=rendererGT, foregroundPrior=globalPrior, variances=variances))/numPixels pixelLikelihoodRobustRegionCh = generative_models.LogRobustModelRegion(renderer=renderer, groundtruth=rendererGT, foregroundPrior=globalPrior, variances=variances) pixelLikelihoodCh = generative_models.LogGaussianModel(renderer=renderer, groundtruth=rendererGT, variances=variances) pixelLikelihoodRobustCh = generative_models.LogRobustModel(renderer=renderer, groundtruth=rendererGT, foregroundPrior=globalPrior, variances=variances) post = generative_models.layerPosteriorsRobustCh(rendererGT, renderer, vis_im, 'FULL', globalPrior, variances)[0] vis_im = np.array(renderer.indices_image==1).copy().astype(np.bool) # hogGT, hogImGT, drconv = image_processing.diffHog(rendererGT) # hogRenderer, hogImRenderer, _ = image_processing.diffHog(renderer, drconv) # # hogE_raw = hogGT - hogRenderer # hogCellErrors = ch.sum(hogE_raw*hogE_raw, axis=2) # hogError = -ch.dot(hogGT.ravel(),hogRenderer.ravel())/(ch.sqrt(ch.SumOfSquares(hogGT))*ch.sqrt(ch.SumOfSquares(hogGT))) # import opendr.filters # robPyr = opendr.filters.gaussian_pyramid(renderer - rendererGT, n_levels=6, normalization=None)/numPixels # robPyrSum = -ch.sum(ch.log(ch.exp(-0.5*robPyr**2/variances) + 1)) # # edgeErrorPixels = generative_models.EdgeFilter(rendererGT=rendererGT, renderer=renderer)**2 # edgeError = ch.sum(edgeErrorPixels) models = [negLikModel, negLikModelRobust] pixelModels = [pixelLikelihoodCh, pixelLikelihoodRobustCh] modelsDescr = ["Gaussian Model", "Outlier model"] # , negLikModelPyr, negLikModelRobustPyr, SSqE_raw # negLikModel2 = -generative_models.modelLogLikelihoodCh(rendererGT, diffRenderer, vis_im, 'FULL', variances)/numPixels # # negLikModelRobust2 = -generative_models.modelLogLikelihoodRobustCh(rendererGT, diffRenderer, vis_im, 'FULL', globalPrior, variances)/numPixels # # pixelLikelihoodCh2 = generative_models.logPixelLikelihoodCh(rendererGT, diffRenderer, vis_im, 'FULL', variances) # # pixelLikelihoodRobustCh2 = ch.log(generative_models.pixelLikelihoodRobustCh(rendererGT, diffRenderer, vis_im, 'FULL', globalPrior, variances)) # # post2 = generative_models.layerPosteriorsRobustCh(rendererGT, diffRenderer, vis_im, 'FULL', globalPrior, variances)[0] # pixelModels2 = [pixelLikelihoodCh2, pixelLikelihoodRobustCh2] # models2 = [negLikModel2, negLikModelRobust2] model = 1 pixelErrorFun = pixelModels[model] errorFun = models[model] # pixelErrorFun2 = pixelModels2[model] # errorFun2 = models2[model] # zpolys = image_processing.zernikePolynomials(image=rendererGT.r.copy(), numCoeffs=20) iterat = 0 changedGT = False refresh = True drawSurf = False makeVideo = True updateErrorFunctions = True pendingCyclesRender = True performance = {} elevations = {} azimuths = {} gradEl = {} gradAz = {} performanceSurf = {} elevationsSurf = {} azimuthsSurf = {} gradElSurf = {} gradAzSurf = {} gradFinElSurf = {} gradFinAzSurf = {} ims = [] # free_variables = [chCosAz, chSinAz, chLogCosEl, chLogSinEl] free_variables = [chAz, chEl, chVColors, chShCoeffs] # free_variables = [chShapeParams] azVar = 1 elVar = 1 vColorVar = 0.00001 shCoeffsVar = 0.00001 df_vars = np.concatenate([azVar*np.ones(chAz.shape), elVar*np.ones(chEl.shape), vColorVar*np.ones(chVColors.r.shape), shCoeffsVar*np.ones(chShCoeffs.r.shape)]) if useShapeModel: df_vars = np.concatenate([np.ones(chShapeParams.shape)]) maxiter = 20 method=1 options={'disp':False, 'maxiter':maxiter} mintime = time.time() boundEl = (0, np.pi/2.0) boundAz = (0, None) boundscomponents = (0,None) bounds = [boundAz,boundEl] bounds = [(None , None ) for sublist in free_variables for item in sublist] methods=['dogleg', 'minimize', 'BFGS', 'L-BFGS-B', 'Nelder-Mead', 'SGDMom', 'probLineSearch'] exit = False minimize = False plotMinimization = False changeRenderer = False printStatsBool = False beginTraining = False createGroundTruth = False beginTesting = False exploreSurfaceBool = False newTeapotAsGT = False global chAzSaved global chElSaved global chComponentSaved chAzSaved = chAz.r[0] chElSaved = chEl.r[0] # chComponentSaved = chComponent.r[0] if useShapeModel: chShapeParamsSaved = chShapeParams.r[:] if showSubplots: f, ((ax1, ax2), (ax3, ax4), (ax5,ax6)) = plt.subplots(3, 2, subplot_kw={'aspect':'equal'}, figsize=(9, 12)) pos1 = ax1.get_position() pos5 = ax5.get_position() pos5.x0 = pos1.x0 ax5.set_position(pos5) f.tight_layout() ax1.set_title("Ground Truth") ax2.set_title("Backprojection") rendererIm = lin2srgb(renderer.r.copy()) pim2 = ax2.imshow(rendererIm) edges = renderer.boundarybool_image gtoverlay = imageGT().copy() gtoverlay = lin2srgb(gtoverlay) gtoverlay[np.tile(edges.reshape([shapeIm[0],shapeIm[1],1]),[1,1,3]).astype(np.bool)] = 1 pim1 = ax1.imshow(gtoverlay) # # extent = ax1.get_window_extent().transformed(f.dpi_scale_trans.inverted()) # f.savefig('ax1_figure.png', bbox_inches=extent) # ax3.set_title("Pixel negative log probabilities") # pim3 = ax3.imshow(-pixelErrorFun.r) # cb3 = plt.colorbar(pim3, ax=ax3,use_gridspec=True) # cb3.mappable = pim3 # # ax4.set_title("Posterior probabilities") # pim4 = ax4.imshow(np.tile(post.reshape(shapeIm[0],shapeIm[1],1), [1,1,3])) # cb4 = plt.colorbar(pim4, ax=ax4,use_gridspec=True) paramWrt1 = chAz paramWrt2 = chEl if useShapeModel: paramWrt1 = chShapeParams[0] paramWrt2 = chShapeParams[1] diffAz = -ch.optimization.gradCheckSimple(pixelErrorFun, paramWrt1, 0.1) diffEl = -ch.optimization.gradCheckSimple(pixelErrorFun, paramWrt2, 0.1) ax3.set_title("Dr wrt. Azimuth Checkgrad") drazsum = np.sign(-diffAz.reshape(shapeIm[0],shapeIm[1],1))*pixelErrorFun.dr_wrt(paramWrt1).reshape(shapeIm[0],shapeIm[1],1) drazsumnobnd = -pixelErrorFun.dr_wrt(paramWrt1).reshape(shapeIm[0],shapeIm[1],1)*(1-renderer.boundarybool_image.reshape(shapeIm[0],shapeIm[1],1)) drazsumbnd = -pixelErrorFun.dr_wrt(paramWrt1).reshape(shapeIm[0],shapeIm[1],1)*(renderer.boundarybool_image.reshape(shapeIm[0],shapeIm[1],1)) drazsumnobnddiff = diffAz.reshape(shapeIm[0],shapeIm[1],1)*(1-renderer.boundarybool_image.reshape(shapeIm[0],shapeIm[1],1)) drazsumbnddiff = diffAz.reshape(shapeIm[0],shapeIm[1],1)*(renderer.boundarybool_image.reshape(shapeIm[0],shapeIm[1],1)) img3 = ax3.imshow(drazsum.squeeze(),cmap=matplotlib.cm.coolwarm, vmin=-1, vmax=1) cb3 = plt.colorbar(img3, ax=ax3,use_gridspec=True) cb3.mappable = img3 ax4.set_title("Dr wrt. param 2 Checkgrad") drazsum = np.sign(-diffEl.reshape(shapeIm[0],shapeIm[1],1))*pixelErrorFun.dr_wrt(paramWrt2).reshape(shapeIm[0],shapeIm[1],1) img4 = ax4.imshow(drazsum.squeeze(),cmap=matplotlib.cm.coolwarm, vmin=-1, vmax=1) cb4 = plt.colorbar(img4, ax=ax4,use_gridspec=True) cb4.mappable = img4 ax5.set_title("Dr wrt. param 1") drazsum = -pixelErrorFun.dr_wrt(paramWrt1).reshape(shapeIm[0],shapeIm[1],1).reshape(shapeIm[0],shapeIm[1],1) img5 = ax5.imshow(drazsum.squeeze(),cmap=matplotlib.cm.coolwarm, vmin=-1, vmax=1) cb5 = plt.colorbar(img5, ax=ax5,use_gridspec=True) cb5.mappable = img5 ax6.set_title("Dr wrt. param 2") drazsum = -pixelErrorFun.dr_wrt(paramWrt2).reshape(shapeIm[0],shapeIm[1],1).reshape(shapeIm[0],shapeIm[1],1) img6 = ax6.imshow(drazsum.squeeze(),cmap=matplotlib.cm.coolwarm, vmin=-1, vmax=1) cb6 = plt.colorbar(img6, ax=ax6,use_gridspec=True) cb6.mappable = img6 pos1 = ax1.get_position() pos5 = ax5.get_position() pos5.x0 = pos1.x0 ax5.set_position(pos5) pos1 = ax1.get_position() pos5 = ax5.get_position() pos5.x0 = pos1.x0 ax5.set_position(pos5) plt.show() plt.pause(0.01) t = time.time() if makeVideo: import matplotlib.animation as animation Writer = animation.writers['ffmpeg'] writer = Writer(fps=1, metadata=dict(title='', artist=''), bitrate=1800) figvid, (vax1, vax2) = plt.subplots(1, 2, sharey=True, subplot_kw={'aspect':'equal'}, figsize=(12, 6)) vax1.axes.get_xaxis().set_visible(False) vax1.axes.get_yaxis().set_visible(False) vax1.set_title("Ground truth") vax2.axes.get_xaxis().set_visible(False) vax2.axes.get_yaxis().set_visible(False) vax2.set_title("Backprojection") plt.tight_layout() if computePerformance: from mpl_toolkits.mplot3d import Axes3D global figperf figperf = plt.figure() global axperf axperf = figperf.add_subplot(111, projection='3d') from matplotlib.font_manager import FontProperties fontP = FontProperties() fontP.set_size('small') # x1,x2,y1,y2 = plt.axis() # plt.axis((0,360,0,90)) performance[(model, chAzGT.r[0], chElGT.r[0])] = np.array([]) azimuths[(model, chAzGT.r[0], chElGT.r[0])] = np.array([]) elevations[(model, chAzGT.r[0], chElGT.r[0])] = np.array([]) gradAz[(model, chAzGT.r[0], chElGT.r[0])] = np.array([]) gradEl[(model, chAzGT.r[0], chElGT.r[0])] = np.array([]) performanceSurf[(model, chAzGT.r[0], chElGT.r[0])] = np.array([]) azimuthsSurf[(model, chAzGT.r[0], chElGT.r[0])] = np.array([]) elevationsSurf[(model, chAzGT.r[0], chElGT.r[0])] = np.array([]) gradElSurf[(model, chAzGT.r[0], chElGT.r[0])] = np.array([]) gradAzSurf[(model, chAzGT.r[0], chElGT.r[0])] = np.array([]) gradFinAzSurf[(model, chAzGT.r[0], chElGT.r[0])] = np.array([]) gradFinElSurf[(model, chAzGT.r[0], chElGT.r[0])] = np.array([]) def refreshSubplots(): #Other subplots visualizing renders and its pixel derivatives edges = renderer.boundarybool_image imagegt = imageGT() gtoverlay = imageGT().copy() gtoverlay = lin2srgb(gtoverlay) gtoverlay[np.tile(edges.reshape([shapeIm[0],shapeIm[1],1]),[1,1,3]).astype(np.bool)] = 1 pim1.set_data(gtoverlay) rendererIm = lin2srgb(renderer.r.copy()) pim2.set_data(rendererIm) paramWrt1 = chAz paramWrt2 = chEl if useShapeModel: paramWrt1 = chShapeParams[0] paramWrt2 = chShapeParams[1] global model if model != 2: diffAz = -ch.optimization.gradCheckSimple(pixelErrorFun, paramWrt1, 0.1) diffEl = -ch.optimization.gradCheckSimple(pixelErrorFun, paramWrt2, 0.1) # ax3.set_title("Pixel negative log probabilities") # pim3 = ax3.imshow(-pixelErrorFun.r) # cb3.mappable = pim3 # cb3.update_normal(pim3) ax3.set_title("Dr wrt. Parameter 0 Checkgrad") drazsum = -np.sign(diffAz.reshape(shapeIm[0],shapeIm[1],1))*pixelErrorFun.dr_wrt(paramWrt1).reshape(shapeIm[0],shapeIm[1],1) # drazsum = drazsum*(renderer.boundarybool_image.reshape(shapeIm[0],shapeIm[1],1)) img3 = ax3.imshow(drazsum.squeeze(),cmap=matplotlib.cm.coolwarm, vmin=-1, vmax=1) cb3.mappable = img3 # else: # ax3.set_title("HoG Image GT") # pim3 = ax3.imshow(hogImGT.r) # cb3.mappable = pim3 # cb3.update_normal(pim3) # ax4.set_title("Posterior probabilities") # ax4.imshow(np.tile(post.reshape(shapeIm[0],shapeIm[1],1), [1,1,3])) ax4.set_title("Dr wrt. Parameter 1 Checkgrad") drazsum = -np.sign(diffEl.reshape(shapeIm[0],shapeIm[1],1))*pixelErrorFun.dr_wrt(paramWrt2).reshape(shapeIm[0],shapeIm[1],1).reshape(shapeIm[0],shapeIm[1],1) img4 = ax4.imshow(drazsum.squeeze(),cmap=matplotlib.cm.coolwarm, vmin=-1, vmax=1) cb4.mappable = img4 # else: # ax4.set_title("HoG image renderer") # ax4.set_title("HoG Image GT") # pim4 = ax4.imshow(hogImRenderer.r) # cb4.mappable = pim4 # cb4.update_normal(pim4) sdy, sdx = pixelErrorFun.shape drazsum = -pixelErrorFun.dr_wrt(paramWrt1).reshape(sdy,sdx,1).reshape(sdy,sdx,1) # drazsum = drazsum*(renderer.boundarybool_image.reshape(shapeIm[0],shapeIm[1],1)) img5 = ax5.imshow(drazsum.squeeze(),cmap=matplotlib.cm.coolwarm, vmin=-1, vmax=1) cb5.mappable = img5 cb5.update_normal(img5) drazsum = -pixelErrorFun.dr_wrt(paramWrt2).reshape(sdy, sdx,1).reshape(sdy,sdx,1) img6 = ax6.imshow(drazsum.squeeze(),cmap=matplotlib.cm.coolwarm, vmin=-1, vmax=1) cb6.mappable = img6 cb6.update_normal(img6) f.canvas.draw() plt.pause(0.01) def plotSurface(model): global figperf global axperf global surf global line global drawSurf global computePerformance global plotMinimization global chAz global chEl global chDist global chAzGT global chElGT global chDistGT global scene if not plotMinimization and not drawSurf: figperf.clear() global axperf axperf = figperf.add_subplot(111, projection='3d') plt.figure(figperf.number) axperf.clear() from matplotlib.font_manager import FontProperties fontP = FontProperties() fontP.set_size('small') scaleSurfGrads = 0 if drawSurf: print("Drawing gardient surface.") from scipy.interpolate import griddata x1 = np.linspace((azimuthsSurf[(model, chAzGT.r[0], chElGT.r[0])]*180./np.pi).min(), (azimuthsSurf[(model, chAzGT.r[0], chElGT.r[0])]*180./np.pi).max(), len((azimuthsSurf[(model, chAzGT.r[0], chElGT.r[0])]*180./np.pi))) y1 = np.linspace((elevationsSurf[(model, chAzGT.r[0], chElGT.r[0])]*180./np.pi).min(), (elevationsSurf[(model, chAzGT.r[0], chElGT.r[0])]*180./np.pi).max(), len((elevationsSurf[(model, chAzGT.r[0], chElGT.r[0])]*180./np.pi))) x2, y2 = np.meshgrid(x1, y1) z2 = griddata(((azimuthsSurf[(model, chAzGT.r[0], chElGT.r[0])]*180./np.pi), (elevationsSurf[(model, chAzGT.r[0], chElGT.r[0])]*180./np.pi)), performanceSurf[(model, chAzGT.r[0], chElGT.r[0])], (x2, y2), method='cubic') from matplotlib import cm, colors surf = axperf.plot_surface(x2, y2, z2, rstride=3, cstride=3, cmap=cm.coolwarm, linewidth=0.1, alpha=0.85) # scaleSurfGrads = 5./avgSurfGradMagnitudes for point in range(len(performanceSurf[(model, chAzGT.r[0], chElGT.r[0])])): perfi = performanceSurf[(model, chAzGT.r[0], chElGT.r[0])][point] azi = azimuthsSurf[(model, chAzGT.r[0], chElGT.r[0])][point] eli = elevationsSurf[(model, chAzGT.r[0], chElGT.r[0])][point] gradAzi = -gradAzSurf[(model, chAzGT.r[0], chElGT.r[0])][point] gradEli = -gradElSurf[(model, chAzGT.r[0], chElGT.r[0])][point] scaleGrad = np.sqrt(gradAzi**2+gradEli**2) / 5 arrowGrad = Arrow3D([azi*180./np.pi, azi*180./np.pi + gradAzi/scaleGrad], [eli*180./np.pi, eli*180./np.pi + gradEli/scaleGrad], [perfi, perfi], mutation_scale=10, lw=1, arrowstyle="-|>", color="b") axperf.add_artist(arrowGrad) diffAzi = -gradFinAzSurf[(model, chAzGT.r[0], chElGT.r[0])][point] diffEli = -gradFinElSurf[(model, chAzGT.r[0], chElGT.r[0])][point] scaleDiff = np.sqrt(diffAzi**2+diffEli**2) / 5 colorArrow = 'g' if diffAzi * gradAzi + diffEli * gradEli < 0: colorArrow = 'r' arrowGradDiff = Arrow3D([azi*180./np.pi, azi*180./np.pi + diffAzi/scaleDiff], [eli*180./np.pi, eli*180./np.pi + diffEli/scaleDiff], [perfi, perfi], mutation_scale=10, lw=1, arrowstyle="-|>", color=colorArrow) # axperf.add_artist(arrowGradDiff) axperf.plot([chAzGT.r[0]*180./np.pi, chAzGT.r[0]*180./np.pi], [chElGT.r[0]*180./np.pi,chElGT.r[0]*180./np.pi], [z2.min(), z2.max()], 'b--', linewidth=1) errorFun = models[model] axperf.plot(chAz.r*180./np.pi, chEl.r*180./np.pi, errorFun.r[0], 'yD') import scipy.sparse as sp if sp.issparse(errorFun.dr_wrt(chAz)): drAz = -errorFun.dr_wrt(chAz).toarray()[0][0] else: drAz = -errorFun.dr_wrt(chAz)[0][0] if sp.issparse(errorFun.dr_wrt(chEl)): drEl = -errorFun.dr_wrt(chEl).toarray()[0][0] else: drEl = -errorFun.dr_wrt(chEl)[0][0] scaleDr = np.sqrt(drAz**2+drEl**2) / 5 chAzOldi = chAz.r[0] chElOldi = chEl.r[0] diffAz = -ch.optimization.gradCheckSimple(errorFun, chAz, 0.01745) diffEl = -ch.optimization.gradCheckSimple(errorFun, chEl, 0.01745) scaleDiff = np.sqrt(diffAz**2+diffEl**2) / 5 chAz[0] = chAzOldi chEl[0] = chElOldi arrowGrad = Arrow3D([chAz.r[0]*180./np.pi, chAz.r[0]*180./np.pi + drAz/scaleDr], [chEl.r[0]*180./np.pi, chEl.r[0]*180./np.pi + drEl/scaleDr], [errorFun.r[0], errorFun.r[0]], mutation_scale=10, lw=1, arrowstyle="-|>", color="b") axperf.add_artist(arrowGrad) colorArrow = 'g' if diffAz * drAz + diffEl * drEl < 0: colorArrow = 'r' arrowGradDiff = Arrow3D([chAz.r[0]*180./np.pi, chAz.r[0]*180./np.pi + diffAz/scaleDiff], [chEl.r[0]*180./np.pi, chEl.r[0]*180./np.pi + diffEl/scaleDiff], [errorFun.r[0], errorFun.r[0]], mutation_scale=10, lw=1, arrowstyle="-|>", color=colorArrow) axperf.add_artist(arrowGradDiff) if plotMinimization: if azimuths.get((model, chAzGT.r[0], chElGT.r[0])) != None: axperf.plot(azimuths[(model, chAzGT.r[0], chElGT.r[0])]*180./np.pi, elevations[(model, chAzGT.r[0], chElGT.r[0])]*180./np.pi, performance[(model, chAzGT.r[0], chElGT.r[0])], color='g', linewidth=1.5) axperf.plot(azimuths[(model, chAzGT.r[0], chElGT.r[0])]*180./np.pi, elevations[(model, chAzGT.r[0], chElGT.r[0])]*180./np.pi, performance[(model, chAzGT.r[0], chElGT.r[0])], 'rD') axperf.set_xlabel('Azimuth (degrees)') axperf.set_ylabel('Elevation (degrees)') if model == 2: axperf.set_zlabel('Squared Error') plt.title('Model: ' + modelsDescr[model]) plt.pause(0.01) plt.draw() def printStats(): print("**** Statistics ****" ) print("Relative Azimuth: " + str(chAzRel)) print("GT Relative Azimuth: " + str(chAzRelGT)) print("GT Cam Azimuth: " + str(chAzGT)) print("Cam Azimuth: " + str(chAz)) print("GT Cam Elevation: " + str(chElGT)) print("Cam Elevation: " + str(chEl)) print("Dr wrt cam Azimuth: " + str(errorFun.dr_wrt(chAz))) print("Dr wrt cam Elevation: " + str(errorFun.dr_wrt(chEl))) if useShapeModel: print("Dr wrt Shape Param 0: " + str(errorFun.dr_wrt(chShapeParams[0]))) print("Dr wrt Shape Param 1: " + str(errorFun.dr_wrt(chShapeParams[1]))) # print("Dr wrt Distance: " + str(errorFun.dr_wrt(chDist))) print("Occlusion is " + str(getOcclusionFraction(rendererGT)*100) + " %") if drawSurf: avgError = np.mean(np.sqrt((gradAzSurf[(model, chAzGT.r[0], chElGT.r[0])] - gradFinAzSurf[(model, chAzGT.r[0], chElGT.r[0])])**2 + (gradElSurf[(model, chAzGT.r[0], chElGT.r[0])] - gradFinElSurf[(model, chAzGT.r[0], chElGT.r[0])])**2)) print("** Approx gradients - finite differenes." ) print("Avg Eucl. distance :: " + str(avgError)) norm2Grad = np.sqrt((gradAzSurf[(model, chAzGT.r[0], chElGT.r[0])])**2 + (gradElSurf[(model, chAzGT.r[0], chElGT.r[0])])**2) norm2Diff = np.sqrt((gradFinAzSurf[(model, chAzGT.r[0], chElGT.r[0])])**2 + (gradFinElSurf[(model, chAzGT.r[0], chElGT.r[0])])**2) avgAngle = np.arccos((gradFinAzSurf[(model, chAzGT.r[0], chElGT.r[0])]*gradAzSurf[(model, chAzGT.r[0], chElGT.r[0])] + gradFinElSurf[(model, chAzGT.r[0], chElGT.r[0])]*gradElSurf[(model, chAzGT.r[0], chElGT.r[0])])/(norm2Grad*norm2Diff)) print("Avg Angle.: " + str(np.mean(avgAngle))) print("Num opposite (red) gradients: " + str(np.sum((gradFinAzSurf[(model, chAzGT.r[0], chElGT.r[0])]*gradAzSurf[(model, chAzGT.r[0], chElGT.r[0])] + gradFinElSurf[(model, chAzGT.r[0], chElGT.r[0])]*gradElSurf[(model, chAzGT.r[0], chElGT.r[0])]) < 0))) idxmin = np.argmin(performanceSurf[(model, chAzGT.r[0], chElGT.r[0])]) azDiff = np.arctan2(np.arcsin(chAzGT - azimuthsSurf[(model, chAzGT.r[0], chElGT.r[0])][idxmin]), np.arccos(chAzGT - azimuthsSurf[(model, chAzGT.r[0], chElGT.r[0])][idxmin])) elDiff = np.arctan2(np.arcsin(chElGT - elevationsSurf[(model, chAzGT.r[0], chElGT.r[0])][idxmin]), np.arccos(chElGT - elevationsSurf[(model, chAzGT.r[0], chElGT.r[0])][idxmin])) print("Minimum Azimuth difference of " + str(azDiff*180/np.pi)) print("Minimum Elevation difference of " + str(elDiff*180/np.pi)) azDiff = np.arctan2(np.arcsin(chAzGT - chAz.r[0]), np.arccos(chAzGT - chAz.r[0])) elDiff = np.arctan2(np.arcsin(chElGT - chEl.r[0]), np.arccos(chElGT - chEl.r[0])) # print("Current Azimuth difference of " + str(azDiff*180/np.pi)) # print("Current Elevation difference of " + str(elDiff*180/np.pi)) global printStatsBool printStatsBool = False def cb(errorFunMin): # global t # elapsed_time = time.time() - t # print("Ended interation in " + str(elapsed_time)) # # global pixelErrorFun # global errorFun # global iterat # iterat = iterat + 1 print("Callback! " ) # print("Model Log Likelihood: " + str(errorFunMin.r)) # global imagegt # global renderer # global gradAz # global gradEl # global performance # global azimuths # global elevations # t = time.time() def cb2(_): global t elapsed_time = time.time() - t print("Ended interation in " + str(elapsed_time)) global pixelErrorFun global errorFun global iterat iterat = iterat + 1 print("Callback! " + str(iterat)) print("Model Log Likelihood: " + str(errorFun.r)) global imagegt global renderer global gradAz global gradEl global performance global azimuths global elevations if reduceVariance: stds[:] = stds.r[:]*0.9 if demoMode and refreshWhileMinimizing: refreshSubplots() if makeVideo: plt.figure(figvid.number) im1 = vax1.imshow(gtoverlay) bbox_props = dict(boxstyle="round", fc="w", ec="0.5", alpha=0.8) t = vax1.annotate("Minimization iteration: " + str(iterat), xy=(1, 0), xycoords='axes fraction', fontsize=16, xytext=(-20, 5), textcoords='offset points', ha='right', va='bottom', bbox=bbox_props) im2 = vax2.imshow(renderer.r) ims.append([im1, im2, t]) if computePerformance and demoMode: if performance.get((model, chAzGT.r[0], chElGT.r[0])) == None: performance[(model, chAzGT.r[0], chElGT.r[0])] = np.array([]) azimuths[(model, chAzGT.r[0], chElGT.r[0])] = np.array([]) elevations[(model, chAzGT.r[0], chElGT.r[0])] = np.array([]) gradAz[(model, chAzGT.r[0], chElGT.r[0])] = np.array([]) gradEl[(model, chAzGT.r[0], chElGT.r[0])] = np.array([]) performance[(model, chAzGT.r[0], chElGT.r[0])] = numpy.append(performance[(model, chAzGT.r[0], chElGT.r[0])], errorFun.r) azimuths[(model, chAzGT.r[0], chElGT.r[0])] = numpy.append(azimuths[(model, chAzGT.r[0], chElGT.r[0])], chAz.r) elevations[(model, chAzGT.r[0], chElGT.r[0])] = numpy.append(elevations[(model, chAzGT.r[0], chElGT.r[0])], chEl.r) import scipy.sparse as sp if sp.issparse(errorFun.dr_wrt(chAz)): drAz = errorFun.dr_wrt(chAz).toarray()[0][0] else: drAz = errorFun.dr_wrt(chAz)[0][0] if sp.issparse(errorFun.dr_wrt(chEl)): drEl = errorFun.dr_wrt(chEl).toarray()[0][0] else: drEl = errorFun.dr_wrt(chEl)[0][0] gradAz[(model, chAzGT.r[0], chElGT.r[0])] = numpy.append(gradAz[(model, chAzGT.r[0], chElGT.r[0])], drAz) gradEl[(model, chAzGT.r[0], chElGT.r[0])] = numpy.append(gradEl[(model, chAzGT.r[0], chElGT.r[0])], drEl) plotSurface(model) if drawSurf and demoMode and refreshWhileMinimizing: # plt.pause(0.1) plt.show() plt.draw() plt.pause(0.01) t = time.time() def readKeys(window, key, scancode, action, mods): print("Reading keys...") global exit global refresh global chAz global chEl global phiOffsetGT global chComponent global changedGT refresh = False if mods!=glfw.MOD_SHIFT and key == glfw.KEY_ESCAPE and action == glfw.RELEASE: glfw.set_window_should_close(window, True) exit = True if mods!=glfw.MOD_SHIFT and key == glfw.KEY_LEFT and action == glfw.RELEASE: refresh = True chAz[:] = chAz.r[0] - radians(5) azimuth = chAz.r[0] - radians(5) # chCosAz[:] = np.cos(azimuth) # chSinAz[:] = np.sin(azimuth) if mods!=glfw.MOD_SHIFT and key == glfw.KEY_RIGHT and action == glfw.RELEASE: refresh = True chAz[:] = chAz.r[0] + radians(5) azimuth = chAz.r[0] + radians(5) # chCosAz[:] = np.cos(azimuth) # chSinAz[:] = np.sin(azimuth) if mods!=glfw.MOD_SHIFT and key == glfw.KEY_DOWN and action == glfw.RELEASE: refresh = True chEl[:] = chEl[0].r - radians(5) elevation = chEl[0].r - radians(5) if elevation <= 0: elevation = 0.0001 chLogCosEl[:] = np.log(np.cos(elevation)) chLogSinEl[:] = np.log(np.sin(elevation)) refresh = True if mods!=glfw.MOD_SHIFT and key == glfw.KEY_UP and action == glfw.RELEASE: refresh = True chEl[:] = chEl[0].r + radians(5) elevation = chEl[0].r + radians(5) if elevation >= np.pi/2 - 0.0001: elevation = np.pi/2 - 0.0002 chLogCosEl[:] = np.log(np.cos(elevation)) chLogSinEl[:] = np.log(np.sin(elevation)) if mods==glfw.MOD_SHIFT and key == glfw.KEY_LEFT and action == glfw.RELEASE: refresh = True chAz[:] = chAz.r[0] - radians(1) # azimuth = chAz.r[0] - radians(1) # chCosAz[:] = np.cos(azimuth) # chSinAz[:] = np.sin(azimuth) if mods==glfw.MOD_SHIFT and key == glfw.KEY_RIGHT and action == glfw.RELEASE: refresh = True chAz[:] = chAz.r[0] + radians(1) # azimuth = chAz.r[0] + radians(1) # chCosAz[:] = np.cos(azimuth) # chSinAz[:] = np.sin(azimuth) # if mods==glfw.MOD_SHIFT and key == glfw.KEY_LEFT and action == glfw.RELEASE: # print("Left modifier!") # refresh = True # chAzGT[0] = chAzGT[0].r - radians(1) # if mods==glfw.MOD_SHIFT and key == glfw.KEY_RIGHT and action == glfw.RELEASE: # refresh = True # # chAz[0] = chAz[0].r + radians(1) # rotation[:] = rotation.r[0] + np.pi/4 # # rotation[:] = np.pi/2 # shCoeffsRGBGT[:] = np.dot(light_probes.sphericalHarmonicsZRotation(totalOffsetGT.r[:]), envMapCoeffs[[0,3,2,1,4,5,6,7,8]])[[0,3,2,1,4,5,6,7,8]] # chAzGT[:] = rotation.r[:] # # ipdb.set_trace() # shOriginal = chComponentGTOriginal[[0,3,2,1,4,5,6,7,8]] # shOriginalDir = shDirLightGTOriginal[[0,3,2,1,4,5,6,7,8]] # # chComponentGT[:] = np.dot(light_probes.sphericalHarmonicsZRotation(rotation), shOriginal)[[0,3,2,1,4,5,6,7,8]] # # shDirLightGT[:] = np.dot(light_probes.sphericalHarmonicsZRotation(rotation), shOriginalDir)[[0,3,2,1,4,5,6,7,8]] # # shDirLightGT[:] = np.dot(shDirLightGTOriginal.T, light_probes.sphericalHarmonicsZRotation(rotation)).T[:] # # shDirLightGT[:] = np.sum(np.array(light_probes.sphericalHarmonicsZRotation(rotation) * shDirLightGTOriginal[:,None]), axis=1) # # shDirLightGT[:] = chZonalToSphericalHarmonics(zGT, np.pi/2 - chLightElGT, chLightAzGT + rotation[:] - np.pi/2).r[:] # print("Original: " + str(shDirLightGTOriginal)) # print(str(shCoeffsRGBGT.r)) # print(str(shDirLightGT.r)) # print(str(rendererGT.tn[0])) if mods==glfw.MOD_SHIFT and key == glfw.KEY_DOWN and action == glfw.RELEASE: refresh = True chEl[0] = chEl[0].r - radians(1) refresh = True if mods==glfw.MOD_SHIFT and key == glfw.KEY_UP and action == glfw.RELEASE: refresh = True chEl[0] = chEl[0].r + radians(1) # if mods!=glfw.MOD_SHIFT and key == glfw.KEY_X and action == glfw.RELEASE: # refresh = True # chScale[0] = chScale[0].r + 0.05 # # if mods==glfw.MOD_SHIFT and key == glfw.KEY_X and action == glfw.RELEASE: # refresh = True # chScale[0] = chScale[0].r - 0.05 # if mods!=glfw.MOD_SHIFT and key == glfw.KEY_Y and action == glfw.RELEASE: # refresh = True # chScale[1] = chScale[1].r + 0.05 # # if mods==glfw.MOD_SHIFT and key == glfw.KEY_Y and action == glfw.RELEASE: # refresh = True # chScale[1] = chScale[1].r - 0.05 # if mods!=glfw.MOD_SHIFT and key == glfw.KEY_Z and action == glfw.RELEASE: # refresh = True # chScale[2] = chScale[2].r + 0.05 # # if mods==glfw.MOD_SHIFT and key == glfw.KEY_Z and action == glfw.RELEASE: # refresh = True # chScale[2] = chScale[2].r - 0.05 global errorFun if mods != glfw.MOD_SHIFT and key == glfw.KEY_C and action == glfw.RELEASE: print("Azimuth grad check: ") jacs, approxjacs, check = ch.optimization.gradCheck(errorFun, [chAz], [1.49e-08]) print("Grad check jacs: " + "%.2f" % jacs) print("Grad check fin jacs: " + "%.2f" % approxjacs) print("Grad check check: " + "%.2f" % check) # print("Scipy grad check: " + "%.2f" % ch.optimization.scipyGradCheck({'raw': errorFun}, [chAz])) print("Elevation grad check: ") jacs, approxjacs, check = ch.optimization.gradCheck(errorFun, [chEl], [1]) print("Grad check jacs: " + "%.2f" % jacs) print("Grad check fin jacs: " + "%.2f" % approxjacs) print("Grad check check: " + "%.2f" % check) # print("Scipy grad check: " + "%.2f" % ch.optimization.scipyGradCheck({'raw': errorFun}, [chEl])) print("Red VColor grad check: ") jacs, approxjacs, check = ch.optimization.gradCheck(errorFun, [chVColors[0]], [0.01]) print("Grad check jacs: " + "%.2f" % jacs) print("Grad check fin jacs: " + "%.2f" % approxjacs) print("Grad check check: " + "%.2f" % check) # print("Scipy grad check: " + "%.2f" % ch.optimization.scipyGradCheck({'raw': errorFun}, [chVColors])) if key == glfw.KEY_D: refresh = True # chComponent[0] = chComponent[0].r + 0.1 if mods == glfw.MOD_SHIFT and glfw.KEY_D: refresh = True # chComponent[0] = chComponent[0].r - 0.1 global drawSurf global model global models global updateErrorFunctions if key == glfw.KEY_G and action == glfw.RELEASE: refresh = True changedGT = True updateErrorFunctions = True global targetPosition global center global cameraGT global rendererGT global renderer global teapotGT global teapot global newTeapotAsGT if mods==glfw.MOD_SHIFT and key == glfw.KEY_G and action == glfw.RELEASE: newTeapotAsGT = True global groundTruthBlender global blenderRender if mods != glfw.MOD_SHIFT and key == glfw.KEY_B and action == glfw.RELEASE: if useBlender: updateErrorFunctions = True groundTruthBlender = not groundTruthBlender # changedGT = True # if groundTruthBlender: # bpy.ops.render.render( write_still=True ) # image = cv2.imread(scene.render.filepath) # image = np.float64(cv2.cvtColor(image, cv2.COLOR_BGR2RGB))/255.0 # blenderRender = image refresh = True #Compute in order to plot the surface neighouring the azimuth/el of the gradients and error function. global exploreSurfaceBool if key == glfw.KEY_E and action == glfw.RELEASE: print("Key E pressed?") exploreSurfaceBool = True if key == glfw.KEY_P and action == glfw.RELEASE: ipdb.set_trace() refresh = True if key == glfw.KEY_N and action == glfw.RELEASE: print("Back to GT!") chAz[:] = chAzGT.r[:] chEl[:] = chElGT.r[:] chShapeParams[:] = chShapeParamsGT.r[:] # chComponent[:] = chComponentGT.r[:] refresh = True global chAzSaved global chElSaved global chComponentSaved global chShapeParamsSaved if key == glfw.KEY_Z and action == glfw.RELEASE: print("Saved!") chAzSaved = chAz.r[0] chElSaved = chEl.r[0] chShapeParamsSaved = chShapeParams.r[:] # chComponentSaved = chComponent.r[0] if key == glfw.KEY_X and action == glfw.RELEASE: print("Back to Saved!") chAz[0] = chAzSaved chEl[0] = chElSaved # chComponent[0] = chComponentSaved chShapeParams[:] = chShapeParamsSaved refresh = True global printStatsBool if mods!=glfw.MOD_SHIFT and key == glfw.KEY_S and action == glfw.RELEASE: printStatsBool = True if key == glfw.KEY_V and action == glfw.RELEASE: global ims if makeVideo: im_ani = animation.ArtistAnimation(figvid, ims, interval=2000, repeat_delay=3000, repeat=False, blit=True) im_ani.save('minimization_demo.mp4', fps=None, writer=writer, codec='mp4') ims = [] global stds global globalPrior global plotMinimization if mods != glfw.MOD_CONTROL and mods != glfw.MOD_SHIFT and key == glfw.KEY_KP_1 and action == glfw.RELEASE: stds[:] = stds.r[0]/1.5 print("New standard devs of " + str(stds.r)) refresh = True drawSurf = False plotMinimization = False if mods != glfw.MOD_CONTROL and mods != glfw.MOD_SHIFT and key == glfw.KEY_KP_2 and action == glfw.RELEASE: stds[:] = stds.r[0]*1.5 print("New standard devs of " + str(stds.r)) refresh = True drawSurf = False plotMinimization = False if mods != glfw.MOD_CONTROL and mods != glfw.MOD_SHIFT and key == glfw.KEY_KP_4 and action == glfw.RELEASE: globalPrior[0] = globalPrior.r[0] - 0.05 print("New foreground prior of" + str(globalPrior.r)) refresh = True drawSurf = False plotMinimization = False if mods != glfw.MOD_CONTROL and mods != glfw.MOD_SHIFT and key == glfw.KEY_KP_5 and action == glfw.RELEASE: globalPrior[0] = globalPrior.r[0] + 0.05 print("New foreground prior of " + str(globalPrior.r)) refresh = True drawSurf = False plotMinimization = False global changeRenderer global currentTeapotModel changeRenderer = False if mods != glfw.MOD_SHIFT and key == glfw.KEY_KP_7 and action == glfw.RELEASE: currentTeapotModel = (currentTeapotModel - 1) % len(renderTeapotsList) changeRenderer = True if mods != glfw.MOD_SHIFT and key == glfw.KEY_KP_8 and action == glfw.RELEASE: currentTeapotModel = (currentTeapotModel + 1) % len(renderTeapotsList) changeRenderer = True global renderer global chShapeParams if mods == glfw.MOD_SHIFT and key == glfw.KEY_KP_1 and action == glfw.RELEASE: refresh = True chShapeParams[1] = chShapeParams.r[1] + 0.2 if mods == glfw.MOD_SHIFT and key == glfw.KEY_KP_2 and action == glfw.RELEASE: refresh = True chShapeParams[2] = chShapeParams.r[2] + 0.2 if mods == glfw.MOD_SHIFT and key == glfw.KEY_KP_3 and action == glfw.RELEASE: chShapeParams[3] = chShapeParams.r[3] + 0.2 refresh = True if mods == glfw.MOD_SHIFT and key == glfw.KEY_KP_4 and action == glfw.RELEASE: chShapeParams[4] = chShapeParams.r[4] + 0.2 refresh = True if mods == glfw.MOD_SHIFT and key == glfw.KEY_KP_5 and action == glfw.RELEASE: chShapeParams[5] = chShapeParams.r[5] + 0.2 refresh = True if mods == glfw.MOD_SHIFT and mods == glfw.MOD_SHIFT and key == glfw.KEY_KP_6 and action == glfw.RELEASE: chShapeParams[6] = chShapeParams.r[6] + 0.2 refresh = True if mods == glfw.MOD_SHIFT and key == glfw.KEY_KP_8 and action == glfw.RELEASE: chShapeParams[7] = chShapeParams.r[7] + 0.2 refresh = True if mods == glfw.MOD_SHIFT and key == glfw.KEY_KP_9 and action == glfw.RELEASE: chShapeParams[9] = chShapeParams.r[9] + 0.2 refresh = True if mods == glfw.MOD_SHIFT and key == glfw.KEY_KP_8 and action == glfw.RELEASE: chShapeParams[8] = chShapeParams.r[8] + 0.2 refresh = True if mods == glfw.MOD_SHIFT and key == glfw.KEY_KP_0 and action == glfw.RELEASE: chShapeParams[0] = chShapeParams.r[0] + 0.2 refresh = True if mods == glfw.MOD_CONTROL and key == glfw.KEY_KP_1 and action == glfw.RELEASE: refresh = True chShapeParams[1] = chShapeParams.r[1] - 0.2 if mods == glfw.MOD_CONTROL and key == glfw.KEY_KP_2 and action == glfw.RELEASE: refresh = True chShapeParams[2] = chShapeParams.r[2] - 0.2 if mods == glfw.MOD_CONTROL and key == glfw.KEY_KP_3 and action == glfw.RELEASE: chShapeParams[3] = chShapeParams.r[3] - 0.2 refresh = True if mods == glfw.MOD_CONTROL and key == glfw.KEY_KP_4 and action == glfw.RELEASE: chShapeParams[4] = chShapeParams.r[4] - 0.2 refresh = True if mods == glfw.MOD_CONTROL and key == glfw.KEY_KP_5 and action == glfw.RELEASE: chShapeParams[5] = chShapeParams.r[5] - 0.2 refresh = True if mods == glfw.MOD_CONTROL and mods == glfw.MOD_CONTROL and key == glfw.KEY_KP_6 and action == glfw.RELEASE: chShapeParams[6] = chShapeParams.r[6] - 0.2 refresh = True if mods == glfw.MOD_CONTROL and key == glfw.KEY_KP_8 and action == glfw.RELEASE: chShapeParams[7] = chShapeParams.r[7] - 0.2 refresh = True if mods == glfw.MOD_CONTROL and key == glfw.KEY_KP_9 and action == glfw.RELEASE: chShapeParams[9] = chShapeParams.r[9] - 0.2 refresh = True if mods == glfw.MOD_CONTROL and key == glfw.KEY_KP_8 and action == glfw.RELEASE: chShapeParams[8] = chShapeParams.r[8] - 0.2 refresh = True if mods == glfw.MOD_CONTROL and key == glfw.KEY_KP_0 and action == glfw.RELEASE: chShapeParams[0] = chShapeParams.r[0] - 0.2 refresh = True if key == glfw.KEY_R and action == glfw.RELEASE: refresh = True if mods == glfw.MOD_CONTROL and key == glfw.KEY_R and action == glfw.RELEASE: refresh = True chShapeParams[:] = np.random.randn(latentDim) chShapeParamsGT[:] = np.random.randn(latentDim) global pixelErrorFun global pixelErrorFun2 global errorFun2 global models2 global pixelModels2 global model global models global modelsDescr global pixelModels global reduceVariance if key == glfw.KEY_O and action == glfw.RELEASE: # drawSurf = False model = (model + 1) % len(models) print("Using " + modelsDescr[model]) errorFun = models[model] pixelErrorFun = pixelModels[model] # errorFun2 = models2[model] # pixelErrorFun2 = pixelModels2[model] # if model == 2: # reduceVariance = True # else: # reduceVariance = False refresh = True global method global methods global options global maxiter if key == glfw.KEY_1 and action == glfw.RELEASE: method = 0 options={'disp':False, 'maxiter':maxiter} print("Changed to minimizer: " + methods[method]) if key == glfw.KEY_2 and action == glfw.RELEASE: method = 1 maxiter = 20 options={'disp':False, 'maxiter':maxiter} print("Changed to minimizer: " + methods[method]) if key == glfw.KEY_3 and action == glfw.RELEASE: method = 2 options={'disp':False, 'maxiter':maxiter} print("Changed to minimizer: " + methods[method]) if key == glfw.KEY_4 and action == glfw.RELEASE: print("Changed to minimizer: " + methods[method]) method = 3 options={'disp':False, 'maxiter':maxiter} if key == glfw.KEY_5 and action == glfw.RELEASE: method = 4 maxiter = 1000 options={'disp':False, 'maxiter':maxiter} print("Changed to minimizer: " + methods[method]) if key == glfw.KEY_6 and action == glfw.RELEASE: method = 5 maxiter = 200 options = {'disp':False, 'maxiter':maxiter, 'lr':0.01, 'momentum':0.5, 'decay':0.99} print("Changed to minimizer: " + methods[method]) if key == glfw.KEY_7 and action == glfw.RELEASE: maxiter = 50 method = 6 options={'disp':False, 'maxiter':maxiter, 'df_vars':df_vars} print("Changed to minimizer: " + methods[method]) global minimize global free_variables global df_vars if mods==glfw.MOD_SHIFT and key == glfw.KEY_M and action == glfw.RELEASE: # free_variables = [renderer.v.a.a] minimize = True if mods!=glfw.MOD_SHIFT and key == glfw.KEY_M and action == glfw.RELEASE: # free_variables = [chAz, chEl, chVColors, chShCoeffs] minimize = True def timeRendering(iterations): t = time.time() for i in range(iterations): chAz[:] = chAz.r[0] + radians(0.001) renderer.r print("Per iteration time of " + str((time.time() - t)/iterations)) def timeGradients(iterations): t = time.time() for i in range(iterations): chAz[:] = chAz.r[0] + radians(0.001) errorFun.dr_wrt(chAz) print("Per iteration time of " + str((time.time() - t)/iterations)) def exploreSurface(): global drawSurf global errorFun global refresh global model if computePerformance: print("Estimating cost function surface and gradients...") drawSurf = True chAzOld = chAz.r[0] chElOld = chEl.r[0] for model_num, errorFun in enumerate(models): performanceSurf[(model_num, chAzGT.r[0], chElGT.r[0])] = np.array([]) azimuthsSurf[(model_num, chAzGT.r[0], chElGT.r[0])] = np.array([]) elevationsSurf[(model_num, chAzGT.r[0], chElGT.r[0])] = np.array([]) gradAzSurf[(model_num, chAzGT.r[0], chElGT.r[0])] = np.array([]) gradElSurf[(model_num, chAzGT.r[0], chElGT.r[0])] = np.array([]) gradFinAzSurf[(model_num, chAzGT.r[0], chElGT.r[0])] = np.array([]) gradFinElSurf[(model_num, chAzGT.r[0], chElGT.r[0])] = np.array([]) for chAzi in np.linspace(max(chAzGT.r[0]-np.pi/3.,0), min(chAzGT.r[0] + np.pi/3., 2.*np.pi), num=20): for chEli in np.linspace(max(chElGT.r[0]-np.pi/2,0), min(chElGT.r[0]+np.pi/2, np.pi/2), num=10): for model_num, errorFun in enumerate(models): chAz[:] = chAzi chEl[:] = chEli performanceSurf[(model_num, chAzGT.r[0], chElGT.r[0])] = numpy.append(performanceSurf[(model_num, chAzGT.r[0], chElGT.r[0])], errorFun.r) azimuthsSurf[(model_num, chAzGT.r[0], chElGT.r[0])] = numpy.append(azimuthsSurf[(model_num, chAzGT.r[0], chElGT.r[0])], chAzi) elevationsSurf[(model_num, chAzGT.r[0], chElGT.r[0])] = numpy.append(elevationsSurf[(model_num, chAzGT.r[0], chElGT.r[0])], chEli) import scipy.sparse as sp if sp.issparse(errorFun.dr_wrt(chAz)): drAz = errorFun.dr_wrt(chAz).toarray()[0][0] else: drAz = errorFun.dr_wrt(chAz)[0][0] if sp.issparse(errorFun.dr_wrt(chEl)): drEl = errorFun.dr_wrt(chEl).toarray()[0][0] else: drEl = errorFun.dr_wrt(chEl)[0][0] gradAzSurf[(model_num, chAzGT.r[0], chElGT.r[0])] = numpy.append(gradAzSurf[(model_num, chAzGT.r[0], chElGT.r[0])], drAz) gradElSurf[(model_num, chAzGT.r[0], chElGT.r[0])] = numpy.append(gradElSurf[(model_num, chAzGT.r[0], chElGT.r[0])], drEl) chAzOldi = chAz.r[0] chElOldi = chEl.r[0] diffAz = ch.optimization.gradCheckSimple(errorFun, chAz, 0.01745) diffEl = ch.optimization.gradCheckSimple(errorFun, chEl, 0.01745) chAz[:] = chAzOldi chEl[:] = chElOldi gradFinAzSurf[(model_num, chAzGT.r[0], chElGT.r[0])] = numpy.append(gradFinAzSurf[(model_num, chAzGT.r[0], chElGT.r[0])], diffAz) gradFinElSurf[(model_num, chAzGT.r[0], chElGT.r[0])] = numpy.append(gradFinElSurf[(model_num, chAzGT.r[0], chElGT.r[0])], diffEl) errorFun = models[model] chAz[:] = chAzOld chEl[:] = chElOld refresh = True if savePerformance: def writeStats(model): with open('stats/statistics.txt', 'a') as statsFile: statsFile.write("**** Statistics for " + modelsDescr[model] + " ****" + '\n') if drawSurf: avgError = np.mean(np.sqrt((gradAzSurf[(model, chAzGT.r[0], chElGT.r[0])] - gradFinAzSurf[(model, chAzGT.r[0], chElGT.r[0])])**2 + (gradElSurf[(model, chAzGT.r[0], chElGT.r[0])] - gradFinElSurf[(model, chAzGT.r[0], chElGT.r[0])])**2)) statsFile.write("** Approx gradients - finite differenes." + '\n') statsFile.write("Avg Eucl. distance :: " + str(avgError) + '\n') norm2Grad = np.sqrt((gradAzSurf[(model, chAzGT.r[0], chElGT.r[0])])**2 + (gradElSurf[(model, chAzGT.r[0], chElGT.r[0])])**2) norm2Diff = np.sqrt((gradFinAzSurf[(model, chAzGT.r[0], chElGT.r[0])])**2 + (gradFinElSurf[(model, chAzGT.r[0], chElGT.r[0])])**2) avgAngle = np.arccos((gradFinAzSurf[(model, chAzGT.r[0], chElGT.r[0])]*gradAzSurf[(model, chAzGT.r[0], chElGT.r[0])] + gradFinElSurf[(model, chAzGT.r[0], chElGT.r[0])]*gradElSurf[(model, chAzGT.r[0], chElGT.r[0])])/(norm2Grad*norm2Diff)) statsFile.write("Avg Angle.: " + str(np.mean(avgAngle)) + '\n') statsFile.write("Num opposite (red) gradients: " + str(np.sum((gradFinAzSurf[(model, chAzGT.r[0], chElGT.r[0])]*gradAzSurf[(model, chAzGT.r[0], chElGT.r[0])] + gradFinElSurf[(model, chAzGT.r[0], chElGT.r[0])]*gradElSurf[(model, chAzGT.r[0], chElGT.r[0])]) < 0)) + '\n') idxmin = np.argmin(performanceSurf[(model, chAzGT.r[0], chElGT.r[0])]) azDiff = np.arctan2(np.arcsin(chAzGT - azimuthsSurf[(model, chAzGT.r[0], chElGT.r[0])][idxmin]), np.arccos(chAzGT - azimuthsSurf[(model, chAzGT.r[0], chElGT.r[0])][idxmin])) elDiff = np.arctan2(np.arcsin(chElGT - elevationsSurf[(model, chAzGT.r[0], chElGT.r[0])][idxmin]), np.arccos(chElGT - elevationsSurf[(model, chAzGT.r[0], chElGT.r[0])][idxmin])) statsFile.write("Minimum Azimuth difference of " + str(azDiff*180/np.pi) + '\n') statsFile.write("Minimum Elevation difference of " + str(elDiff*180/np.pi) + '\n') azDiff = np.arctan2(np.arcsin(chAzGT - chAz.r[0]), np.arccos(chAzGT - chAz.r[0])) elDiff = np.arctan2(np.arcsin(chElGT - chEl.r[0]), np.arccos(chElGT - chEl.r[0])) statsFile.write("Current Azimuth difference of " + str(azDiff*180/np.pi) + '\n') statsFile.write("Current Elevation difference of " + str(elDiff*180/np.pi) + '\n\n') for model_idx, model_i in enumerate(models): writeStats(model_idx) plotSurface(model_idx) plt.savefig('stats/surfaceModel' + modelsDescr[model_idx] + '.png') print("Finshed estimating.") if demoMode: glfw.set_key_callback(win, readKeys) while not exit: # Poll for and process events glfw.make_context_current(win) glfw.poll_events() if newTeapotAsGT: rendererGT.makeCurrentContext() rendererGT.clear() del rendererGT removeObjectData(len(v) - targetIndex - 1, v, f_list, vc, vn, uv, haveTextures_list, textures_list) addObjectData(v, f_list, vc, vn, uv, haveTextures_list, textures_list, v_teapots[currentTeapotModel][0], f_list_teapots[currentTeapotModel][0], vc_teapots[currentTeapotModel][0], vn_teapots[currentTeapotModel][0], uv_teapots[currentTeapotModel][0], haveTextures_list_teapots[currentTeapotModel][0], textures_list_teapots[currentTeapotModel][0]) rendererGT = createRendererGT(glMode, chAzGT, chObjAzGT, chElGT, chDistGT, center, v, vc, f_list, vn, light_colorGT, chComponentGT, chVColorsGT, targetPosition, chDisplacementGT, chScaleGT, width,height, uv, haveTextures_list, textures_list, frustum, win ) updateErrorFunctions = True refresh = True changedGT = True #Unlink and place the new teapot for Blender. if useBlender: scene.objects.unlink(teapotGT) teapot.matrix_world = mathutils.Matrix.Translation(targetPosition) teapotGT = blender_teapots[currentTeapotModel] placeNewTarget(scene, teapotGT, targetPosition) placeCamera(scene.camera, -chAzGT.r[:].copy()*180/np.pi, chElGT.r[:].copy()*180/np.pi, chDistGT, center) scene.update() newTeapotAsGT = False if printStatsBool: printStats() if changedGT: drawSurf = False plotMinimization = False imagegt = imageGT() chImage[:,:,:] = imagegt[:,:,:] chAzGT[:] = chAz.r[:] chElGT[:] = chEl.r[:] chDistGT[:] = chDist.r[:] # chComponentGT[:] = chComponent.r[:] # chVColorsGT[:] = chVColors.r[:] if makeVideo: ims = [] performance[(model, chAzGT.r[0], chElGT.r[0])] = np.array([]) azimuths[(model, chAzGT.r[0], chElGT.r[0])] = np.array([]) elevations[(model, chAzGT.r[0], chElGT.r[0])] = np.array([]) gradAz[(model, chAzGT.r[0], chElGT.r[0])] = np.array([]) gradEl[(model, chAzGT.r[0], chElGT.r[0])] = np.array([]) performanceSurf[(model, chAzGT.r[0], chElGT.r[0])] = np.array([]) azimuthsSurf[(model, chAzGT.r[0], chElGT.r[0])] = np.array([]) elevationsSurf[(model, chAzGT.r[0], chElGT.r[0])] = np.array([]) gradElSurf[(model, chAzGT.r[0], chElGT.r[0])] = np.array([]) gradAzSurf[(model, chAzGT.r[0], chElGT.r[0])] = np.array([]) gradFinAzSurf[(model, chAzGT.r[0], chElGT.r[0])] = np.array([]) gradFinElSurf[(model, chAzGT.r[0], chElGT.r[0])] = np.array([]) if useBlender: print("Updating Ground Truth blender camera!") scene.update() center = centerOfGeometry(teapotGT.dupli_group.objects, teapotGT.matrix_world) placeCamera(scene.camera, -chAzGT.r[:].copy()*180/np.pi, chElGT.r[:].copy()*180/np.pi, chDistGT, center) scene.update() if useBlender: pendingCyclesRender = True if useGTasBackground: for teapot_i in range(len(renderTeapotsList)): renderer_i = renderer_teapots[teapot_i] renderer_i.set(background_image=rendererGT.r) changedGT = False if exploreSurfaceBool: print("Explores surface is true??") exploreSurface() exploreSurfaceBool = False if groundTruthBlender and pendingCyclesRender: scene.update() # scene.layers[0] = False # scene.layers[1] = True bpy.ops.render.render( write_still=True ) # image = cv2.imread(scene.render.filepath) # image = np.float64(cv2.cvtColor(image, cv2.COLOR_BGR2RGB))/255.0 image = np.array(imageio.imread(scene.render.filepath))[:,:,0:3] plt.imsave('blenderImage.png', lin2srgb(image)) image[image>1]=1 blenderRender = image # blenderRenderGray = 0.3*blenderRender[:,:,0] + 0.59*blenderRender[:,:,1] + 0.11*blenderRender[:,:,2] # rendererGTGray = 0.3*rendererGT[:,:,0] + 0.59*rendererGT[:,:,1] + 0.11*rendererGT[:,:,2] # chAmbientIntensityGT[:] = chAmbientIntensityGT.r*(np.mean(blenderRenderGray,axis=(0,1))/np.mean(rendererGTGray.r,axis=(0,1))) pendingCyclesRender = False if changeRenderer: print("New teapot model " + str(currentTeapotModel)) drawSurf = False plotMinimization = False refresh = True renderer = renderer_teapots[currentTeapotModel] updateErrorFunctions = True if useBlender: teapot = blender_teapots[currentTeapotModel] changeRenderer = False if updateErrorFunctions: # currentGT = rendererGT # if useBlender and groundTruthBlender: # image = np.array(imageio.imread(scene.render.filepath))[:,:,0:3] # image[image>1]=1 # blenderRender = image # currentGT = blenderRender currentGT = ch.Ch(imageGT()) negLikModel = -ch.sum(generative_models.LogGaussianModel(renderer=renderer, groundtruth=currentGT, variances=variances))/numPixels negLikModelRobust = -ch.sum(generative_models.LogRobustModel(renderer=renderer, groundtruth=currentGT, foregroundPrior=globalPrior, variances=variances))/numPixels pixelLikelihoodCh = generative_models.LogGaussianModel(renderer=renderer, groundtruth=currentGT, variances=variances) pixelLikelihoodRobustCh = generative_models.LogRobustModel(renderer=renderer, groundtruth=currentGT, foregroundPrior=globalPrior, variances=variances) modelLogLikelihoodRobustRegionCh = -ch.sum(generative_models.LogRobustModelRegion(renderer=renderer, groundtruth=currentGT, foregroundPrior=globalPrior, variances=variances))/numPixels pixelLikelihoodRobustRegionCh = generative_models.LogRobustModelRegion(renderer=renderer, groundtruth=currentGT, foregroundPrior=globalPrior, variances=variances) post = generative_models.layerPosteriorsRobustCh(currentGT, renderer, vis_im, 'FULL', globalPrior, variances)[0] edgeErrorPixels = generative_models.EdgeFilter(rendererGT=currentGT, renderer=renderer)**2 edgeError = ch.sum(edgeErrorPixels) # hogGT, hogImGT, _ = image_processing.diffHog(currentGT, drconv) # hogRenderer, hogImRenderer, _ = image_processing.diffHog(renderer, drconv) # # hogE_raw = hogGT - hogRenderer # hogCellErrors = ch.sum(hogE_raw*hogE_raw, axis=2) # hogError = -ch.dot(hogGT.ravel(),hogRenderer.ravel())/(ch.sqrt(ch.SumOfSquares(hogGT))*ch.sqrt(ch.SumOfSquares(hogGT))) import opendr.filters robPyr = opendr.filters.gaussian_pyramid(renderer - currentGT, n_levels=6, normalization=None)/numPixels robPyrSum = -ch.sum(ch.log(ch.exp(-0.5*robPyr**2/variances) + 1)) # edgeErrorPixels = generative_models.EdgeFilter(rendererGT=currentGT, renderer=renderer)**2 # edgeError = ch.sum(edgeErrorPixels) models = [negLikModel, negLikModelRobust, robPyrSum] pixelModels = [pixelLikelihoodCh, pixelLikelihoodRobustCh, robPyr] modelsDescr = ["Gaussian Model", "Outlier model", "Region Robust", "Pyr Error" ] pixelErrorFun = pixelModels[model] errorFun = models[model] updateErrorFunctions = False if minimize: iterat = 0 print("Minimizing with method " + methods[method]) ch.minimize({'raw': errorFun}, bounds=bounds, method=methods[method], x0=free_variables, callback=cb2, options=options) plotMinimization = True minimize = False if refresh: print("Model Log Likelihood: " + str(errorFun.r)) if showSubplots: refreshSubplots() if computePerformance and drawSurf: plotSurface(model) # if demoMode or drawSurf: # plt.pause(0.1) # plt.draw() refresh = False refreshSubplots()
84,241
42.002552
357
py
sa-nmt
sa-nmt-master/Loss.py
""" This file handles the details of the loss function during training. This includes: loss criterion, training statistics, and memory optimizations. """ from __future__ import division import time import sys import math import torch import torch.nn as nn def nmt_criterion(vocab_size, pad_id=0): """ Construct the standard NMT Criterion """ weight = torch.ones(vocab_size) weight[pad_id] = 0 crit = nn.NLLLoss(weight, size_average=False) return crit class Statistics: """ Training loss function statistics. """ def __init__(self, loss=0, n_words=0, n_correct=0): self.loss = loss self.n_words = n_words self.n_correct = n_correct self.start_time = time.time() def update(self, stat): self.loss += stat.loss self.n_words += stat.n_words self.n_correct += stat.n_correct def accuracy(self): return 100 * (self.n_correct / self.n_words) def ppl(self): print(self.loss, self.n_words) return math.exp(min(self.loss / self.n_words, 100)) def elapsed_time(self): return time.time() - self.start_time def output(self, epoch, uidx, max_updates, start): t = self.elapsed_time() print(("Epoch %2d, %5d/%5d; acc: %6.2f; ppl: %6.2f; " + "%3.0f tgt tok/s; %6.0f s elapsed") % (epoch, uidx, max_updates, self.accuracy(), self.ppl(), self.n_words / (t + 1e-5), time.time() - start)) sys.stdout.flush() def log(self, prefix, experiment, optim): t = self.elapsed_time() experiment.add_scalar_value(prefix + "_ppl", self.ppl()) experiment.add_scalar_value(prefix + "_accuracy", self.accuracy()) experiment.add_scalar_value(prefix + "_tgtper", self.n_words / t) experiment.add_scalar_value(prefix + "_lr", optim.lr) @staticmethod def score(loss, scores, targ, pad): pred = scores.max(1)[1] non_padding = targ.ne(pad) num_correct = pred.eq(targ) \ .masked_select(non_padding).int() \ .sum().item() return Statistics(loss, non_padding.int().sum().item(), num_correct) def filter_gen_state(state): for k, v in state.items(): if v is not None: yield k, v def new_split(x, size): xs = [] for u in torch.split(x, size): v = u.detach() if u.requires_grad: v.requires_grad_(True) xs += [v] return tuple(xs) def shards(state, shard_size, eval=False): if eval: yield state else: non_none = dict(filter_gen_state(state)) keys, values = zip(*((k, new_split(v, shard_size)) for k, v in non_none.items())) for shard_tensors in zip(*values): yield dict(zip(keys, shard_tensors)) # Assumed backprop'd variables = [] for i, k in enumerate(keys): dv = [v.grad for v in values[i] if v.grad is not None] if dv: dv = torch.cat(dv) variables += [(state[k], dv)] inputs, grads = zip(*variables) torch.autograd.backward(inputs, grads) class LossCompute: def __init__(self, generator, crit): self.generator = generator self.crit = crit def make_loss_batch(self, outputs, targets): """ Create all the variables that need to be sharded. This needs to match compute loss exactly. """ return {"out": outputs, "target": targets} def compute_loss(self, out, target): def bottle(v): return v.view(-1, v.size(2)) target = target.view(-1) # Standard generator. scores = self.generator(bottle(out)) loss = self.crit(scores, target) scores_data = scores.detach() target = target.clone() # Coverage loss term. stats = Statistics.score(loss.item(), scores_data, target, 0) return loss, stats
4,092
27.227586
77
py
sa-nmt
sa-nmt-master/Iterator.py
import numpy import random import pickle as pkl import gzip from tempfile import mkstemp import os import string def fopen(filename, mode='r'): if filename.endswith('.gz'): return gzip.open(filename, mode) return open(filename, mode) class TextIterator: """Simple Bitext iterator.""" def __init__(self, source_file, target_file, source_dict, target_dict, batch_size=32, batch_tokens=0, max_seq_length=50, src_vocab_size=-1, tgt_vocab_size=-1): self.source_file = source_file self.target_file = target_file self.shuffle([source_file, target_file]) self.source = fopen(source_file + '.shuf', 'r') self.target = fopen(target_file + '.shuf', 'r') with open(source_dict, 'rb') as f: self.source_dict = pkl.load(f) with open(target_dict, 'rb') as f: self.target_dict = pkl.load(f) self.src_punct = set() for p in set(string.punctuation): if p in self.source_dict: self.src_punct.add(self.source_dict[p]) self.batch_tokens = batch_tokens self.batch_size = batch_size self.max_seq_length = max_seq_length self.src_vocab_size = src_vocab_size self.tgt_vocab_size = tgt_vocab_size if src_vocab_size < 0: self.src_vocab_size = len(self.source_dict) if tgt_vocab_size < 0: self.tgt_vocab_size = len(self.target_dict) self.source_buffer = [] self.target_buffer = [] self.k = batch_size * 20 self.end_of_data = False def __iter__(self): return self @staticmethod def shuffle(files): print('shuffle %s | %s' % (files[0], files[1])) tf_os, tpath = mkstemp() tf = open(tpath, 'w') fds = [open(ff) for ff in files] punct = set(string.punctuation) def punct_count(ws): p = [w for w in ws if w in punct] return len(p) for l in fds[0]: lines = [l.strip()] + [ff.readline().strip() for ff in fds[1:]] tokens = [l.split() for l in lines] lengths = [len(t) for t in tokens] if min(lengths) * 2 < max(lengths): continue if min(lengths) < 1: continue # remove lines that have many punctuations n_punct = punct_count(tokens[0]) if n_punct * 4 > lengths[0]: continue print("|||".join(lines), file=tf) [ff.close() for ff in fds] tf.close() tf = open(tpath, 'r') lines = tf.readlines() random.shuffle(lines) fds = [open(ff + '.shuf', 'w') for ff in files] for l in lines: s = l.strip().split('|||') for ii, fd in enumerate(fds): print(s[ii], file=fd) [ff.close() for ff in fds] os.remove(tpath) return def reset(self): self.shuffle([self.source_file, self.target_file]) self.source.seek(0) self.target.seek(0) def __next__(self): if self.end_of_data: self.end_of_data = False self.reset() raise StopIteration source = [] target = [] batch_tokens = 0 # fill buffer, if it's empty assert len(self.source_buffer) == len(self.target_buffer) if len(self.source_buffer) == 0: for k_ in range(self.k): ss = self.source.readline() if ss == "": break tt = self.target.readline() if tt == "": break self.source_buffer.append(ss.strip().split()) self.target_buffer.append(tt.strip().split()) # sort by target buffer tlen = numpy.array([len(t) for t in self.target_buffer]) tidx = tlen.argsort() _sbuf = [self.source_buffer[i] for i in tidx] _tbuf = [self.target_buffer[i] for i in tidx] self.source_buffer = _sbuf self.target_buffer = _tbuf if len(self.source_buffer) == 0 or len(self.target_buffer) == 0: self.end_of_data = False self.reset() raise StopIteration try: # actual work here while True: # read from source file and map to word index try: ss = self.source_buffer.pop() except IndexError: break ss = [self.source_dict.get(w, 1) for w in ss] if self.src_vocab_size > 0: ss = [w if w < self.src_vocab_size else 1 for w in ss] # read from source file and map to word index tt = self.target_buffer.pop() # append bos and eos to target tt = [2] + [self.target_dict.get(w, 1) for w in tt] + [3] if self.tgt_vocab_size > 0: tt = [w if w < self.tgt_vocab_size else 1 for w in tt] if len(ss) > self.max_seq_length or \ len(tt) > self.max_seq_length: continue source.append(ss) target.append(tt) batch_tokens += len(ss) + len(tt) if self.batch_tokens > 0: if batch_tokens >= self.batch_tokens: break else: if len(source) >= self.batch_size: break except IOError: self.end_of_data = True if len(source) <= 0 or len(target) <= 0: self.end_of_data = False self.reset() raise StopIteration return source, target
5,911
29.474227
75
py
sa-nmt
sa-nmt-master/opts.py
import argparse def model_opts(parser): """ These options are passed to the construction of the model. Be careful with these as they will be used during translation. """ # Model options # Embedding Options parser.add_argument('-word_vec_size', type=int, default=512, help='Word embedding for both.') parser.add_argument('-share_decoder_embeddings', action='store_true', help='Share the word and out embeddings for decoder.') # RNN Options parser.add_argument('-encoder_type', type=str, default='brnn', choices=['brnn', 'sabrnn', 'fabrnn'], help="""Type of encoder layer to use.""") parser.add_argument('-decoder_type', type=str, default='rnn', choices=['rnn'], help='Type of decoder layer to use.') parser.add_argument('-layers', type=int, default=2, help='Number of layers in enc/dec.') parser.add_argument('-rnn_size', type=int, default=512, help='Size of LSTM hidden states') parser.add_argument('-input_feed', type=int, default=1, help="""Feed the context vector at each time step as additional input (via concatenation with the word embeddings) to the decoder.""") parser.add_argument('-rnn_type', type=str, default='LSTM', choices=['LSTM'], help="""The gate type to use in the RNNs""") parser.add_argument('-encode_multi_key', action='store_true', help="""Using multi keys encoding of source""") parser.add_argument('-share_attn', action='store_true', help="""sharing attention weights""") def train_opts(parser): parser.add_argument('-save_model', default='model', help="""Model filename (the model will be saved as <save_model>_epochN_PPL.pt where PPL is the validation perplexity""") parser.add_argument('-train_from', default='', type=str, help="""If training from a checkpoint then this is the path to the pretrained model's state_dict.""") # GPU parser.add_argument('-gpuid', default=[], nargs='+', type=int, help="Use CUDA on the listed devices.") parser.add_argument('-seed', type=int, default=42, help="""Random seed used for the experiments reproducibility.""") parser.add_argument('-max_updates', type=int, default=1000000, help="""max number of updates""") # Init options parser.add_argument('-start_epoch', type=int, default=1, help='The epoch from which to start') parser.add_argument('-param_init', type=float, default=0.1, help="""Parameters are initialized over uniform distribution with support (-param_init, param_init). Use 0 to not use initialization""") # Optimization options parser.add_argument('-batch_size', type=int, default=32, help='Maximum batch size') parser.add_argument('-max_generator_batches', type=int, default=32, help="""Maximum batches of words in a sequence to run the generator on in parallel. Higher is faster, but uses more memory.""") parser.add_argument('-epochs', type=int, default=13, help='Number of training epochs') parser.add_argument('-optim', default='adam', choices=['sgd', 'adagrad', 'adadelta', 'adam'], help="""Optimization method.""") parser.add_argument('-max_grad_norm', type=float, default=5, help="""If the norm of the gradient vector exceeds this, renormalize it to have the norm equal to max_grad_norm""") parser.add_argument('-dropout', type=float, default=0.3, help="Dropout probability; applied in LSTM stacks.") # learning rate parser.add_argument('-learning_rate', type=float, default=1.0, help="""Starting learning rate. If adagrad/adadelta/adam is used, then this is the global learning rate. Recommended settings: sgd = 1, adagrad = 0.1, adadelta = 1, adam = 0.001""") parser.add_argument('-learning_rate_decay', type=float, default=0.5, help="""If update_learning_rate, decay learning rate by this much if (i) perplexity does not decrease on the validation set or (ii) epoch has gone past start_decay_at""") parser.add_argument('-max_decay_step', type=int, default=5, help='maximum number of decay step for learning rate') parser.add_argument('-start_checkpoint_at', type=int, default=0, help="""Start checkpointing every epoch after and including this epoch""") parser.add_argument('-start_eval_checkpoint_at', type=int, default=0, help="""Start checkpointing every eval after and including this eval""") parser.add_argument('-report_every', type=int, default=50, help="Print stats at this interval.") parser.add_argument('-eval_every', type=int, default=10000, help="Evaluate at this interval.") parser.add_argument('-min_thres', type=float, default=-5.0, help="clip the value lower than this point.") parser.add_argument('-max_thres', type=float, default=7.0, help="clip the value larger than this point.") parser.add_argument('-hard', action='store_true', help="using hard structured attention.") def preprocess_opts(parser): # Dictionary Options parser.add_argument('-src_vocab_size', type=int, default=-1, help="Size of the source vocabulary") parser.add_argument('-tgt_vocab_size', type=int, default=-1, help="Size of the target vocabulary") # Truncation options parser.add_argument('-max_seq_length', type=int, default=100, help="Maximum sequence length") # Data processing options parser.add_argument('-shuffle', type=int, default=1, help="Shuffle data") def add_md_help_argument(parser): parser.add_argument('-md', action=MarkdownHelpAction, help='print Markdown-formatted help text and exit.') # MARKDOWN boilerplate # Copyright 2016 The Chromium Authors. All rights reserved. # Use of this source code is governed by a BSD-style license that can be # found in the LICENSE file. class MarkdownHelpFormatter(argparse.HelpFormatter): """A really bare-bones argparse help formatter that generates valid markdown. This will generate something like: usage # **section heading**: ## **--argument-one** ``` argument-one help text ``` """ def _format_usage(self, usage, actions, groups, prefix): usage_text = super(MarkdownHelpFormatter, self)._format_usage( usage, actions, groups, prefix) return '\n```\n%s\n```\n\n' % usage_text def format_help(self): self._root_section.heading = '# %s' % self._prog return super(MarkdownHelpFormatter, self).format_help() def start_section(self, heading): super(MarkdownHelpFormatter, self).start_section('## **%s**' % heading) def _format_action(self, action): lines = [] action_header = self._format_action_invocation(action) lines.append('### **%s** ' % action_header) if action.help: lines.append('') lines.append('```') help_text = self._expand_help(action) lines.extend(self._split_lines(help_text, 80)) lines.append('```') lines.extend(['', '']) return '\n'.join(lines) class MarkdownHelpAction(argparse.Action): def __init__(self, option_strings, dest=argparse.SUPPRESS, default=argparse.SUPPRESS, **kwargs): super(MarkdownHelpAction, self).__init__( option_strings=option_strings, dest=dest, default=default, nargs=0, **kwargs) def __call__(self, parser, namespace, values, option_string=None): parser.formatter_class = MarkdownHelpFormatter parser.print_help() parser.exit()
8,871
44.968912
84
py
sa-nmt
sa-nmt-master/translate.py
import argparse import torch import modelx as models import infer import string # build args parser parser = argparse.ArgumentParser(description='Training NMT') parser.add_argument('-checkpoint', required=True, help='saved checkpoit.') parser.add_argument('-input', required=True, help='Text file to translate.') parser.add_argument('-output', default='trans.bpe', help='output file') parser.add_argument('-beam_size', default=5, type=int, help="Beam size.") parser.add_argument('-gpuid', default=[], nargs='+', type=int, help="Use CUDA") opt = parser.parse_args() use_gpu = len(opt.gpuid) > 0 if torch.cuda.is_available() and not use_gpu: print("so you should probably run with -gpus 0") checkpoint = torch.load(opt.checkpoint) train_opt = checkpoint['opt'] print('| train configuration') train_opt.min_thres = -5.0 train_opt.max_thres = 7.0 #if train_opt.hard is None: #train_opt.hard = False print(train_opt) model = models.make_base_model(train_opt, use_gpu, checkpoint) if train_opt.encoder_type == "sabrnn": punct_idx = set() for p in set(string.punctuation): if p in model.dicts[0]: punct_idx.add(model.dicts[0][p]) model.encoder.punct(punct_idx) # over-write some options train_opt.beam_size = opt.beam_size agent = infer.Beam(train_opt, model) agent.translate(opt.input, opt.output)
1,416
28.520833
71
py
sa-nmt
sa-nmt-master/Utils.py
def aeq(*args): """ Assert all arguments have the same value """ arguments = (arg for arg in args) first = next(arguments) assert all(arg == first for arg in arguments), \ "Not all arguments have the same value: " + str(args) def use_gpu(opt): return (hasattr(opt, 'gpuid') and len(opt.gpuid) > 0) or \ (hasattr(opt, 'gpu') and opt.gpu > -1)
388
26.785714
62
py
sa-nmt
sa-nmt-master/extract_tree.py
import argparse import torch from torch.autograd import Variable import modelx as models import networkx as nx from networkx.algorithms.tree import maximum_spanning_arborescence import string # build args parser parser = argparse.ArgumentParser(description='Training NMT') parser.add_argument('-checkpoint', required=True, help='saved checkpoit.') parser.add_argument('-input', required=True, help='Text file to translate.') parser.add_argument('-output', default='tree.txt', help='output file') parser.add_argument('-gpuid', default=[], nargs='+', type=int, help="Use CUDA") opt = parser.parse_args() use_gpu = True if torch.cuda.is_available() and not use_gpu: print("so you should probably run with -gpus 0") checkpoint = torch.load(opt.checkpoint) train_opt = checkpoint['opt'] print('| train configuration') print(train_opt) use_gpu = len(train_opt.gpuid) > 0 model = models.make_base_model(train_opt, use_gpu, checkpoint) if train_opt.encoder_type in ["sabrnn", "fabrnn"]: punct_idx = set() for p in set(string.punctuation): if p in model.dicts[0]: punct_idx.add(model.dicts[0][p]) print('Add punctuation constraint') model.encoder.punct(punct_idx) # get the encoder encoder = model.encoder dicts = model.dicts tt = torch.cuda if use_gpu else torch def encode_string(ss): ss = ss.split() ss = [dicts[0].get(w, 1) for w in ss] ss = Variable(tt.LongTensor(ss).view(-1, 1), volatile=True) return ss def collapse_bpe(s, score): """Collapse BPEed tokens Args: s: a bped sentence score: beta in the paper, a 2D tensor of p(z|s), sum over the last dimension should be 1 """ punct = set(string.punctuation) # (1) identify bpe tokens = s.split() bpe = [] indices = [] punct_idx = [] for i, w in enumerate(tokens): if w in punct: punct_idx += [i] if w.endswith("@@"): bpe += [i] else: if len(bpe) == 0: indices += [[i]] else: bpe += [i] # add the last bped token indices += [bpe] bpe = [] # collapsing from here s_ = [] for bpe in indices: # (1) collapse heads s_.append(score[:, bpe].sum(1).view(-1, 1)) s_ = torch.cat(s_, 1) collapsed_score = [] for bpe in indices: # (2) collapse childs collapsed_score += [s_[bpe, :].sum(0).view(1, -1)] collapsed_score = torch.cat(collapsed_score, 0) s = s.replace("@@ ", "") # the original string return s, collapsed_score def build_graph(score): """Build graph from potential score matrix Args: score: FloatTensor (n, n), score.sum(1) = 1 Returns: a graph object """ # return a list of (parent, child, weight) arcs = [] n = score.size(0) # find the root first val, idx = score.diag().max(0) arcs.append((0, idx[0] + 1, val[0])) for i in range(n): for j in range(n): if i == j: # root on the diagonal continue # arcs.append([0, i+1, score[i, j]]) else: arcs.append([j+1, i+1, score[i, j]]) g = nx.DiGraph() g.add_weighted_edges_from(arcs) return g def mst(score): """Get spaning tree from the adjacent matrix""" g = build_graph(score) mst = maximum_spanning_arborescence(g) tree = [] for e in mst.edges(): head, child = e tree.append('%s->%s' % (head, child)) return ' '.join(tree) def renorm(m): x = m.exp() x = x / x.sum(1, keepdim=True) return x.log() def collapse(): fw = open(opt.output, 'w') for line in open(opt.input): x = encode_string(line) model.encoder(x) if train_opt.encoder_type == 'sabrnn': score = model.encoder.tree_attn.edge_score.squeeze(0) else: score = model.encoder.attn.score.squeeze(0).log() s, score = collapse_bpe(line, score.data) try: tree = mst(score) out = '%s ||| %s\n' % (s.strip(), tree) fw.write(out) except: pass fw.close() collapse()
4,276
26.242038
70
py
sa-nmt
sa-nmt-master/models.py
import torch import torch.nn as nn from torch.autograd import Variable from attention import GlobalAttention, SelfAttention from Utils import aeq from torch.nn.utils.rnn import pack_padded_sequence as pack from torch.nn.utils.rnn import pad_packed_sequence as unpack import math class EncoderBase(nn.Module): """ EncoderBase class for sharing code among various encoder. """ def _check_args(self, input, lengths=None, hidden=None): s_len, n_batch = input.size() if lengths is not None: n_batch_ = len(lengths) # lengths.size() aeq(n_batch, n_batch_) def forward(self, input, lengths=None, hidden=None): """ Args: input (LongTensor): len x batch x nfeat. lengths (LongTensor): batch hidden: Initial hidden state. Returns: hidden_t (Variable): Pair of layers x batch x rnn_size - final encoder state outputs (FloatTensor): len x batch x rnn_size - Memory bank """ raise NotImplementedError class Encoder(EncoderBase): """ The standard RNN encoder. """ def __init__(self, word_vocab_size, embedding_size, word_padding_idx, num_layers, hidden_size, dropout, bidirectional=True): super(Encoder, self).__init__() num_directions = 2 if bidirectional else 1 assert hidden_size % num_directions == 0 hidden_size = hidden_size // num_directions self.embeddings = nn.Embedding(word_vocab_size, embedding_size, padding_idx=word_padding_idx) self.rnn = nn.LSTM(input_size=embedding_size, hidden_size=hidden_size, num_layers=num_layers, dropout=dropout, bidirectional=bidirectional) def forward(self, input, lengths=None, hidden=None): """ See EncoderBase.forward() for description of args and returns.""" self._check_args(input, lengths, hidden) emb = self.embeddings(input) s_len, batch, emb_dim = emb.size() self.word_embs = emb # for accessing later packed_emb = emb if lengths is not None: packed_emb = pack(emb, lengths) outputs, hidden = self.rnn(packed_emb, hidden) if lengths is not None: outputs = unpack(outputs)[0] return outputs, hidden # Structured Attention and our models class MatrixTree(nn.Module): """Implementation of the matrix-tree theorem for computing marginals of non-projective dependency parsing. This attention layer is used in the paper "Learning Structured Text Representations." """ def __init__(self, eps=1e-5, device=torch.device('cpu')): self.eps = eps super(MatrixTree, self).__init__() self.device = device def forward(self, input, lengths=None): laplacian = input.exp() output = input.clone() output.data.fill_(0) for b in range(input.size(0)): lx = lengths[b] if lengths is not None else input.size(1) input_b = input[b, :lx, :lx] lap = laplacian[b, :lx, :lx].masked_fill( torch.eye(lx).to(self.device).ne(0), 0) lap = -lap + torch.diag(lap.sum(0)) # store roots on diagonal lap[0] = input_b.diag().exp() inv_laplacian = lap.inverse() factor = inv_laplacian.diag().unsqueeze(1)\ .expand(lx, lx).transpose(0, 1) term1 = input_b.exp().mul(factor).clone() term2 = input_b.exp().mul(inv_laplacian.transpose(0, 1)).clone() term1[:, 0] = 0 term2[0] = 0 output_b = term1 - term2 roots_output = input_b.diag().exp().mul( inv_laplacian.transpose(0, 1)[0]) output[b, :lx, :lx] = output_b + torch.diag(roots_output) return output class TreeAttention(nn.Module): """Structured attention class""" def __init__(self, dim, min_thres=-5, max_thres=7, hard=False, device=torch.device('cpu')): super(TreeAttention, self).__init__() self.q = nn.Linear(dim, dim, bias=False) self.k = nn.Linear(dim, dim, bias=False) self.v = nn.Linear(dim, dim, bias=False) self.root_query = nn.Parameter(torch.randn(dim)) self.scale = math.sqrt(1 / dim) self.dtree = MatrixTree() self.min_thres = min_thres self.max_thres = max_thres self.hard = hard self.device = device def forward(self, input, punct_mask=None, lengths=None): s_len, batch, dim = input.size() input = input.contiguous().transpose(0, 1) \ .contiguous().view(-1, dim) q = self.q(input).view(batch, s_len, -1) # (batch, s_len, dim) k = self.k(input).view(batch, s_len, -1) # (batch, s_len, dim) v = self.v(input).view(batch, s_len, -1) # (batch, s_len, dim) _score = torch.bmm(q, k.transpose(1, 2)) # (batch, s_len, s_len) # compute root r_ = self.root_query.view(1, -1, 1).expand(batch, dim, 1) root = torch.bmm(k, r_).squeeze(-1) # (batch, s_len) mask = torch.eye(s_len).to(self.device) score = _score.clone() for b in range(batch): score[b] = _score[b] * mask + torch.diag(root[b]) # normalized score *= self.scale if punct_mask is not None: punct_mask = punct_mask.transpose(0, 1) punct_mask = punct_mask[:, None, :].expand(batch, s_len, s_len) \ .transpose(1, 2) score.data.masked_fill_(punct_mask, -math.inf) score = score.clamp(self.min_thres, self.max_thres) self.edge_score = score.transpose(1, 2) edge_score = self.dtree(score, lengths).transpose(1, 2) # edge_score.sum(2) == 1 if self.hard: y = edge_score.data.new(edge_score.size()).fill_(0) _, max_idx = edge_score.data.max(2) y.scatter_(2, max_idx[:, :, None], 1) hard_edge = (Variable(y) - edge_score).detach() + edge_score edge_score = hard_edge return torch.bmm(edge_score, v) class SAEncoder(Encoder): """ The structured attention RNN encoder. """ def __init__(self, word_vocab_size, embedding_size, word_padding_idx, num_layers, hidden_size, dropout, bidirectional=True, encode_multi_key=True, min_thres=-5, max_thres=7, hard=False): super(SAEncoder, self).__init__(word_vocab_size, embedding_size, word_padding_idx, num_layers, hidden_size, dropout, bidirectional=True) self.tree_attn = TreeAttention(hidden_size, min_thres, max_thres, hard) self.encode_multi_key = encode_multi_key if not self.encode_multi_key: self.linear = nn.Linear(hidden_size, hidden_size, bias=False) self.punct_idx = None def punct(self, punct_idx): self.punct_idx = punct_idx def forward(self, input, lengths=None): outputs, hidden = Encoder.forward(self, input, lengths) # find puncts punct_mask = None if self.punct_idx is not None: punct = set(input.data.contiguous().view(-1).tolist()) punct &= self.punct_idx if len(punct) > 0: punct_mask = 0 for p in punct: punct_mask += input.data.eq(p) tree_outputs = self.tree_attn(outputs, punct_mask, lengths) # tree_outputs has size (batch_size, s_len, hidden_size) if not self.encode_multi_key: # compute gate syntax z = self.linear(outputs.view(-1, outputs.size(2))).sigmoid() gtree = z.view_as(outputs) * tree_outputs.transpose(0, 1) outputs = outputs + gtree return outputs, hidden return (outputs, tree_outputs), hidden class FAEncoder(Encoder): """ The flat attention RNN encoder. """ def __init__(self, word_vocab_size, embedding_size, word_padding_idx, num_layers, hidden_size, dropout, bidirectional=True, encode_multi_key=True): super(FAEncoder, self).__init__(word_vocab_size, embedding_size, word_padding_idx, num_layers, hidden_size, dropout, bidirectional=True) self.attn = SelfAttention(hidden_size) self.encode_multi_key = encode_multi_key def punct(self, punct_idx): self.punct_idx = punct_idx def forward(self, input, lengths=None): mask = input.data.eq(0).t() outputs, hidden = Encoder.forward(self, input, lengths) punct_mask = None if self.punct_idx is not None: punct = set(input.data.contiguous().view(-1).tolist()) punct &= self.punct_idx if len(punct) > 0: punct_mask = 0 for p in punct: punct_mask += input.data.eq(p) self_attn_outputs = self.attn(outputs, mask, punct_mask) if not self.encode_multi_key: outputs = outputs + self_attn_outputs.transpose(0, 1) return outputs, hidden return (outputs, self_attn_outputs), hidden class StackedLSTM(nn.Module): """ Our own implementation of stacked LSTM. Needed for the decoder, because we do input feeding. """ def __init__(self, num_layers, input_size, rnn_size, dropout): super(StackedLSTM, self).__init__() self.dropout = nn.Dropout(dropout) self.num_layers = num_layers self.layers = nn.ModuleList() for i in range(num_layers): self.layers.append(nn.LSTMCell(input_size, rnn_size)) input_size = rnn_size def forward(self, input, hidden): h_0, c_0 = hidden h_1, c_1 = [], [] for i, layer in enumerate(self.layers): h_1_i, c_1_i = layer(input, (h_0[i], c_0[i])) input = h_1_i if i + 1 != self.num_layers: input = self.dropout(input) h_1 += [h_1_i] c_1 += [c_1_i] h_1 = torch.stack(h_1) c_1 = torch.stack(c_1) return input, (h_1, c_1) class Decoder(nn.Module): def __init__(self, word_vocab_size, embedding_size, word_padding_idx, num_layers, hidden_size, dropout, multi_key=False, shared_attn=False): input_size = embedding_size + hidden_size self.hidden_size = hidden_size super(Decoder, self).__init__() self.embeddings = nn.Embedding(word_vocab_size, embedding_size, padding_idx=word_padding_idx) self.rnn = StackedLSTM(num_layers, input_size, hidden_size, dropout) self.attn = GlobalAttention(hidden_size, multi_key, shared_attn) self.dropout = nn.Dropout(dropout) def forward(self, input, hidden, context, mask=None, init_output=None): emb = self.embeddings(input) batch_size = input.size(1) h_size = (batch_size, self.hidden_size) outputs = [] if init_output is None: output = Variable(emb.data.new(*h_size).zero_(), requires_grad=False) else: output = init_output attns = [] # set mask if mask is not None: self.attn.apply_mask(mask) for i, emb_t in enumerate(emb.split(1)): emb_t = emb_t.squeeze(0) emb_t = torch.cat([emb_t, output], 1) rnn_output, hidden = self.rnn(emb_t, hidden) attn_output, attn = self.attn(rnn_output, context) output = self.dropout(attn_output) outputs += [output] attns.append(attn) attns = torch.stack(attns) outputs = torch.stack(outputs) return outputs, hidden, attns class NMT(nn.Module): def __init__(self, encoder, decoder): super(NMT, self).__init__() self.encoder = encoder self.decoder = decoder def _fix_enc_hidden(self, h): """ The encoder hidden is (layers*directions) x batch x dim. We need to convert it to layers x batch x (directions*dim). """ return torch.cat([h[0:h.size(0):2], h[1:h.size(0):2]], 2) def _init_hidden(self, enc_hidden): hidden = (self._fix_enc_hidden(enc_hidden[0]), self._fix_enc_hidden(enc_hidden[1])) return hidden def forward(self, src, tgt, src_lengths): context, enc_hidden = self.encoder(src, src_lengths) if isinstance(context, tuple): context_ = (context[0].transpose(0, 1), context[1]) else: context_ = context.transpose(0, 1) enc_hidden = self._init_hidden(enc_hidden) src_pad_mask = src.data.eq(0).t() out, dec_hidden, attn = self.decoder(tgt, enc_hidden, context_, src_pad_mask) return out def make_encoder(opt): """ Various encoder dispatcher function. Args: opt: the option in current environment. """ if opt.encoder_type == "sabrnn": return SAEncoder(opt.src_vocab_size, opt.word_vec_size, 0, opt.layers, opt.rnn_size, opt.dropout, bidirectional=True, encode_multi_key=opt.encode_multi_key, min_thres=opt.min_thres, max_thres=opt.max_thres, hard=opt.hard) elif opt.encoder_type == "fabrnn": return FAEncoder(opt.src_vocab_size, opt.word_vec_size, 0, opt.layers, opt.rnn_size, opt.dropout, bidirectional=True, encode_multi_key=opt.encode_multi_key) else: return Encoder(opt.src_vocab_size, opt.word_vec_size, 0, opt.layers, opt.rnn_size, opt.dropout) def make_decoder(opt): """ Various decoder dispatcher function. Args: opt: the option in current environment. embeddings (Embeddings): vocab embeddings for this decoder. """ return Decoder(opt.tgt_vocab_size, opt.word_vec_size, 0, opt.layers, opt.rnn_size, opt.dropout, opt.encode_multi_key, opt.share_attn) def make_base_model(model_opt, checkpoint=None): """ Args: model_opt: the option loaded from checkpoint. fields: `Field` objects for the model. gpu: Boolean: whether to use gpu. checkpoint: the snapshot model. Returns: the NMTModel. """ # Make encoder. encoder = make_encoder(model_opt) decoder = make_decoder(model_opt) # Make NMT (= encoder + decoder). model = NMT(encoder, decoder) # Make Generator. generator = nn.Sequential( nn.Linear(model_opt.rnn_size, model_opt.tgt_vocab_size), nn.LogSoftmax(dim=-1)) if model_opt.share_decoder_embeddings: generator[0].weight = decoder.embeddings.weight model.generator = generator # Load the model states from checkpoint. if checkpoint is not None: print('Loading model') model.load_state_dict(checkpoint['model']) model.dicts = checkpoint['dicts'] return model
15,669
36.488038
79
py
sa-nmt
sa-nmt-master/infer.py
import torch from torch.autograd import Variable import pickle as pkl import math # TODO: documentation of functions class Beam(object): r"""Beam search class for NMT. This is a simple beam search object. It takes model, which can be used to compute the next probable output and dictionaries that will be used to map from word indices to the real strings. opt: opt (argpaser): that contains path to dictionaries model (pytorch network): this should be creat ed before """ def __init__(self, opt, model): self.opt = opt self.tt = torch.cuda if len(opt.gpuid) > 0 else torch self.model = model self.model.eval() self.dicts = model.dicts # create an inverse map from int->word for target side self.idx2w = {} for w, idx in self.dicts[1].items(): self.idx2w[idx] = w self.bos_idx = self.dicts[1]['<bos>'] self.eos_idx = self.dicts[1]['<eos>'] self.pad_idx = self.dicts[1]['<pad>'] def encode_string(self, ss): ss = ss.split() ss = [self.dicts[0].get(w, 1) for w in ss] if self.opt.src_vocab_size > 0: ss = [w if w < self.opt.src_vocab_size else 1 for w in ss] ss = Variable(self.tt.LongTensor(ss).view(-1, 1), volatile=True) return ss def decode_string(self, tidx): ts = [] for i in list(tidx): if i == self.eos_idx: break else: ts += [self.idx2w[i]] return ' '.join(ts) def beam_search(self, input): """ Beam search function. opt: trained optiongs input: Tensor (bptt x 1) """ k = self.opt.beam_size completed_hyps = [] input = input.expand(input.size(0), k) max_len = int(input.size(0) * 1.5) hypos = self.tt.LongTensor(max_len, k).fill_(2) init_target = self.tt.LongTensor(1, k).fill_(2) init_target = Variable(init_target, volatile=True) scores = self.tt.FloatTensor(k).fill_(-math.inf) scores[0] = 0 #lengths = [input.size(0) for i in range(k)] context, enc_hidden = self.model.encoder(input) init_hidden = self.model._init_hidden(enc_hidden) # alias decoder = self.model.decoder generator = self.model.generator init_output = None if isinstance(context, tuple): context = (context[0].transpose(0, 1), context[1]) else: context = context.transpose(0, 1) decoder.attn.mask = None for t in range(max_len): out, dec_hidden, attn = decoder(init_target, init_hidden, context, init_output=init_output) log_probs = generator(out.squeeze(0)).data scores_t, idx_t = log_probs.topk(k, 1) scores_t = scores_t + scores.view(-1, 1).expand_as(scores_t) scores, k_idx = scores_t.view(-1).topk(k) next_hp = k_idx.div(k) next_ys = idx_t.view(-1).index_select(0, k_idx) done_y = next_ys.eq(self.eos_idx) if done_y.sum() > 0 and t > 0: for i in range(k): if next_ys[i] == self.eos_idx: j = next_hp[i] text = self.decode_string(hypos[0:t, j]) completed_hyps.append((text, scores[i] / (t+1))) k -= 1 if k > 0: cont_y = next_ys.ne(self.eos_idx) next_ys = next_ys.masked_select(cont_y) next_hp = next_hp.masked_select(cont_y) if isinstance(context, tuple): context = (context[0][:k], context[1][:k]) else: context = context[:k] scores = scores.masked_select(cont_y) if k == 0: break hypos = hypos.index_select(1, next_hp) hypos[t] = next_ys init_target = Variable(next_ys.view(1, -1), volatile=True) next_hp = Variable(next_hp) init_output = out.squeeze(0).index_select(0, next_hp) init_hidden = [h.index_select(1, next_hp) for h in dec_hidden] if len(completed_hyps) > 0: completed_hyps.sort(key=lambda tup: tup[1]) best_h = completed_hyps.pop() return best_h[0] else: best_s, idx = scores.topk(1) best_h = hypos.index_select(1, idx).view(-1) return self.decode_string(best_h) def translate(self, text_file, out_file='output.txt'): fw = open(out_file, 'w') for line in open(text_file): src_idx = self.encode_string(line) s = self.beam_search(src_idx) fw.write(s + '\n') fw.close()
4,938
35.316176
78
py
sa-nmt
sa-nmt-master/attention.py
import torch import torch.nn as nn from Utils import aeq import math import torch.nn.functional as F class SelfAttention(nn.Module): """Self attention class""" def __init__(self, dim): super(SelfAttention, self).__init__() self.q = nn.Linear(dim, dim, bias=False) self.k = nn.Linear(dim, dim, bias=False) self.v = nn.Linear(dim, dim, bias=False) self.scale = math.sqrt(1 / dim) def forward(self, input, mask=None, punct_mask=None): """ input (FloatTensor): s_len, batch, dim mask (ByteTensor): batch x s_len """ s_len, batch, dim = input.size() input = input.contiguous().transpose(0, 1) \ .contiguous().view(-1, dim) q = self.q(input).view(batch, s_len, -1) k = self.k(input).view(batch, s_len, -1) v = self.v(input).view(batch, s_len, -1) score = torch.bmm(q, k.transpose(1, 2)) * self.scale # now (batch, s_len, s_len) FloatTensor if mask is not None: mask = mask[:, None, :].expand(batch, s_len, s_len) score.data.masked_fill_(mask, -math.inf) if punct_mask is not None: punct_mask = punct_mask.transpose(0, 1) # batch, s_len punct_mask = punct_mask[:, None, :].expand(batch, s_len, s_len) score.data.masked_fill_(punct_mask, -math.inf) # need proper masking attn = F.softmax(score.view(-1, s_len), dim=-1).view(-1, s_len, s_len) self.score = attn return torch.bmm(attn, v) class GlobalAttention(nn.Module): """ Luong Attention. This implement general attention Concrete distribution: The Concrete Distribution: A Continuous Relaxation of Discrete Random Variables """ def __init__(self, dim, multi_key=False, share_attn=False): """ dim (Int): dimension of input vector multi_key (Boolean): using multi keys encoder share_attn (Boolean): sharing attention weights between semantic and syntactic annotations """ super(GlobalAttention, self).__init__() # make a local copy of hyper-parameters self.dim = dim self.share_attn = share_attn self.multi_key = multi_key self.linear_in = nn.Linear(dim, dim, bias=False) self.share_attn = share_attn if multi_key: if not share_attn: # using a separate attention self.linear_sa = nn.Linear(dim, dim, bias=False) self.linear_out = nn.Linear(dim*3, dim, bias=False) self.gate = nn.Linear(dim, dim) else: self.linear_out = nn.Linear(dim*2, dim, bias=False) self.mask = None def apply_mask(self, mask): self.mask = mask def score(self, h_t, h_s, sa_attn=False): """ h_t (FloatTensor): batch x tgt_len x dim h_s (FloatTensor): batch x src_len x dim returns scores (FloatTensor): batch x tgt_len x src_len: raw attention scores for each src index sa_attn (Boolean): using a separate attention for syntax context """ # Check input sizes src_batch, src_len, src_dim = h_s.size() tgt_batch, tgt_len, tgt_dim = h_t.size() aeq(src_batch, tgt_batch) aeq(src_dim, tgt_dim) aeq(self.dim, src_dim) h_t_ = h_t.view(tgt_batch*tgt_len, tgt_dim) if sa_attn: h_t_ = self.linear_sa(h_t_) else: h_t_ = self.linear_in(h_t_) h_t = h_t_.view(tgt_batch, tgt_len, tgt_dim) h_s_ = h_s.transpose(1, 2) # (batch, t_len, d) x (batch, d, s_len) --> (batch, t_len, s_len) return torch.bmm(h_t, h_s_) def forward(self, input, context): """ input (FloatTensor): batch x tgt_len x dim: decoder's rnn's output. context (FloatTensor): batch x src_len x dim: src hidden states """ # one step input if isinstance(context, tuple): context, tree_context = context if input.dim() == 2: one_step = True input = input.unsqueeze(1) else: one_step = False batch, sourceL, dim = context.size() batch_, targetL, dim_ = input.size() aeq(batch, batch_) aeq(dim, dim_) aeq(self.dim, dim) # compute attention scores, as in Luong et al. align = self.score(input, context) if self.mask is not None: mask_ = self.mask[:, None, :] align.data.masked_fill_(mask_, -math.inf) # Softmax to normalize attention weights align_vectors = F.softmax(align, dim=-1) # each context vector c_t is the weighted average # over all the source hidden states c = torch.bmm(align_vectors, context) if self.multi_key: # sharing attention weight if self.share_attn: sc = torch.bmm(align_vectors, tree_context) else: # computing attention scores for syntax tree_align = self.score(input, tree_context, True) if self.mask is not None: tree_align.data.masked_fill_(self.mask[:, None, :], -math.inf) tree_align_vectors = F.softmax(tree_align, dim=-1) sc = torch.bmm(tree_align_vectors, tree_context) z = F.sigmoid(self.gate(input)) # batch x tgt_len x dim self.z = z # for visualization sc = sc * z concat_c = torch.cat([c, input, sc], 2).view(batch*targetL, dim*3) else: concat_c = torch.cat([c, input], 2).view(batch*targetL, dim*2) attn_h = self.linear_out(concat_c).view(batch, targetL, dim) attn_h = F.tanh(attn_h) if one_step: attn_h = attn_h.squeeze(1) align_vectors = align_vectors.squeeze(1) # Check output sizes batch_, dim_ = attn_h.size() aeq(batch, batch_) aeq(dim, dim_) batch_, sourceL_ = align_vectors.size() aeq(batch, batch_) aeq(sourceL, sourceL_) else: attn_h = attn_h.transpose(0, 1).contiguous() align_vectors = align_vectors.transpose(0, 1).contiguous() # Check output sizes targetL_, batch_, dim_ = attn_h.size() aeq(targetL, targetL_) aeq(batch, batch_) aeq(dim, dim_) targetL_, batch_, sourceL_ = align_vectors.size() aeq(targetL, targetL_) aeq(batch, batch_) aeq(sourceL, sourceL_) return attn_h, align_vectors
6,737
35.032086
78
py
sa-nmt
sa-nmt-master/train.py
import argparse import torch from Iterator import TextIterator import models from itertools import zip_longest import random import Loss import opts import os import math import subprocess from infer import Beam import re from torch.optim.lr_scheduler import ReduceLROnPlateau parser = argparse.ArgumentParser(description='train.py') # Data and loading options parser.add_argument('-datasets', required=True, default=[], nargs='+', type=str, help='source_file target_file.') parser.add_argument('-valid_datasets', required=True, default=[], nargs='+', type=str, help='valid_source valid target files.') parser.add_argument('-beam_size', default=12, type=int, help="beam size") # dictionaries parser.add_argument('-dicts', required=True, default=[], nargs='+', help='source_vocab.pkl target_vocab.pkl files.') # opts.py opts.add_md_help_argument(parser) opts.model_opts(parser) opts.train_opts(parser) opts.preprocess_opts(parser) opt = parser.parse_args() device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") # for reproducibility torch.manual_seed(opt.seed) random.seed(opt.seed) if torch.cuda.is_available(): torch.cuda.manual_seed(opt.seed) print(opt) # batch preparation def prepare_data(seqs_x, seqs_y): mb = [(seqs_x[i], seqs_y[i]) for i in range(len(seqs_x))] mb.sort(key=lambda x: len(x[0]), reverse=True) xs = torch.LongTensor( list(zip_longest(*map(lambda x: x[0], mb), fillvalue=0))).to(device) ys = torch.LongTensor( list(zip_longest(*map(lambda x: x[1], mb), fillvalue=0))).to(device) lengths_x = [len(x[0]) for x in mb] return xs, ys, lengths_x def eval(model, criterion, valid_data): stats = Loss.Statistics() model.eval() loss = Loss.LossCompute(model.generator, criterion) for src, tgt in valid_data: src, tgt, src_lengths = prepare_data(src, tgt, True) outputs = model(src, tgt[:-1], src_lengths) gen_state = loss.make_loss_batch(outputs, tgt[1:]) _, batch_stats = loss.compute_loss(**gen_state) stats.update(batch_stats) model.train() return stats def init_uniform(model, init_range=0.04): """Simple uniform initialization of all the weights""" for p in model.parameters(): p.data.uniform_(-init_range, init_range) def tally_parameters(model): n_params = sum([p.nelement() for p in model.parameters()]) print('* number of parameters: %d' % n_params) enc = 0 dec = 0 for name, param in model.named_parameters(): if 'encoder' in name: enc += param.nelement() elif 'decoder' or 'generator' in name: dec += param.nelement() print('encoder: ', enc) print('decoder: ', dec) def check_model_path(): save_model_path = os.path.abspath(opt.save_model) model_dirname = os.path.dirname(save_model_path) if not os.path.exists(model_dirname): os.makedirs(model_dirname) def train(opt): print('| build data iterators') train = TextIterator(*opt.datasets, *opt.dicts, src_vocab_size=opt.src_vocab_size, tgt_vocab_size=opt.tgt_vocab_size, batch_size=opt.batch_size, max_seq_length=opt.max_seq_length) valid = TextIterator(*opt.valid_datasets, *opt.dicts, src_vocab_size=opt.src_vocab_size, tgt_vocab_size=opt.tgt_vocab_size, batch_size=opt.batch_size, max_seq_length=opt.max_seq_length) if opt.src_vocab_size < 0: opt.src_vocab_size = len(train.source_dict) if opt.tgt_vocab_size < 0: opt.tgt_vocab_size = len(train.target_dict) print('| vocabulary size. source = %d; target = %d' % (opt.src_vocab_size, opt.tgt_vocab_size)) dicts = [train.source_dict, train.target_dict] crit = Loss.nmt_criterion(opt.tgt_vocab_size, 0).to(device) if opt.train_from != '': print('| Load trained model!') checkpoint = torch.load(opt.train_from) model = models.make_base_model(opt, checkpoint) else: model = models.make_base_model(opt) init_uniform(model) model.to(device) if opt.encoder_type in ["sabrnn", "fabrnn"]: print('Add punctuation constrain!') model.encoder.punct(train.src_punct) print(model) model.dicts = dicts check_model_path() tally_parameters(model) optimizer = torch.optim.Adam(model.parameters(), lr=opt.learning_rate) scheduler = ReduceLROnPlateau(optimizer, 'min', factor=opt.learning_rate_decay, patience=0) uidx = 0 # number of updates estop = False min_lr = opt.learning_rate * math.pow(opt.learning_rate_decay, 5) best_bleu = -1 for eidx in range(1, opt.epochs + 1): closs = Loss.LossCompute(model.generator, crit) tot_loss = 0 total_stats = Loss.Statistics() report_stats = Loss.Statistics() for x, y in train: model.zero_grad() src, tgt, lengths_x = prepare_data(x, y) out = model(src, tgt[:-1], lengths_x) gen_state = closs.make_loss_batch(out, tgt[1:]) shard_size = opt.max_generator_batches batch_size = len(lengths_x) batch_stats = Loss.Statistics() for shard in Loss.shards(gen_state, shard_size): loss, stats = closs.compute_loss(**shard) loss.div(batch_size).backward() batch_stats.update(stats) tot_loss += loss.item() torch.nn.utils.clip_grad_norm_(model.parameters(), opt.max_grad_norm) optimizer.step() total_stats.update(batch_stats) report_stats.update(batch_stats) uidx += 1 if uidx % opt.report_every == 0: report_stats.output(eidx, uidx, opt.max_updates, total_stats.start_time) report_stats = Loss.Statistics() if uidx % opt.eval_every == 0: valid_stats = eval(model, crit, valid) # maybe adjust learning rate scheduler.step(valid_stats.ppl()) cur_lr = optimizer.param_groups[0]['lr'] print('Validation perplexity %d: %g' % (uidx, valid_stats.ppl())) print('Learning rate: %g' % cur_lr) if cur_lr < min_lr: print('Reaching minimum learning rate. Stop training!') estop = True break model_state_dict = model.state_dict() if eidx >= opt.start_checkpoint_at: checkpoint = { 'model': model_state_dict, 'opt': opt, 'dicts': dicts } # evaluate with BLEU score inference = Beam(opt, model) output_bpe = opt.save_model + '.bpe' output_txt = opt.save_model + '.txt' inference.translate(opt.valid_datasets[0], output_bpe) model.train() subprocess.call("sed 's/@@ //g' {:s} > {:s}" .format(output_bpe, output_txt), shell=True) ref = opt.valid_datasets[1][:-4] subprocess.call("sed 's/@@ //g' {:s} > {:s}" .format(opt.valid_datasets[1], ref), shell=True) cmd = "perl data/multi-bleu.perl {} < {}" \ .format(ref, output_txt) p = subprocess.Popen(cmd, shell=True, stdout=subprocess.PIPE) \ .stdout.read().decode('utf-8') bleu = re.search("[\d]+.[\d]+", p) bleu = float(bleu.group()) print('Validation BLEU %d: %g' % (uidx, bleu)) if bleu > best_bleu: best_bleu = bleu torch.save(checkpoint, '%s_best.pt' % opt.save_model) print('Saved model: %d | BLEU %.2f' % (uidx, bleu)) if uidx >= opt.max_updates: print('Finishing after {:d} iterations!'.format(uidx)) estop = True break if estop: break train(opt)
8,797
36.598291
77
py
sa-nmt
sa-nmt-master/data/filter.py
import os import sys import random from tempfile import mkstemp from subprocess import call def main(files): tf_os, tpath = mkstemp() tf = open(tpath, 'w') fds = [open(ff) for ff in files] for l in fds[0]: lines = [l.strip()] + [ff.readline().strip() for ff in fds[1:]] lengths = [len(l.split()) for l in lines] if min(lengths) * 2 < max(lengths): continue if max(lengths) > 50 or min(lengths) < 2: continue print >>tf, "|||".join(lines) [ff.close() for ff in fds] tf.close() tf = open(tpath, 'r') lines = tf.readlines() random.shuffle(lines) fds = [open(ff+'.shuf','w') for ff in files] for l in lines: s = l.strip().split('|||') for ii, fd in enumerate(fds): print >>fd, s[ii] [ff.close() for ff in fds] os.remove(tpath) if __name__ == '__main__': main(sys.argv[1:])
933
18.458333
71
py
sa-nmt
sa-nmt-master/data/shuffle.py
from __future__ import print_function import os import sys import random from tempfile import mkstemp from subprocess import call def main(files): tf_os, tpath = mkstemp() tf = open(tpath, 'w') fds = [open(ff) for ff in files] for l in fds[0]: lines = [l.strip()] + [ff.readline().strip() for ff in fds[1:]] lengths = [len(l.split()) for l in lines] if min(lengths) * 2 < max(lengths): continue if max(lengths) > 50 or min(lengths) < 1: continue print("|||".join(lines), file=tf) [ff.close() for ff in fds] tf.close() tf = open(tpath, 'r') lines = tf.readlines() random.shuffle(lines) fds = [open(ff+'.shuf','w') for ff in files] for l in lines: s = l.strip().split('|||') for ii, fd in enumerate(fds): print(s[ii], file=fd) [ff.close() for ff in fds] os.remove(tpath) if __name__ == '__main__': main(sys.argv[1:])
982
17.903846
71
py
sa-nmt
sa-nmt-master/data/shuffle.bak.py
from __future__ import print_function import os import sys import random from tempfile import mkstemp from subprocess import call def main(files): tf_os, tpath = mkstemp() tf = open(tpath, 'w') fds = [open(ff) for ff in files] for l in fds[0]: lines = [l.strip()] + [ff.readline().strip() for ff in fds[1:]] print("|||".join(lines), file=tf) [ff.close() for ff in fds] tf.close() tf = open(tpath, 'r') lines = tf.readlines() random.shuffle(lines) fds = [open(ff+'.shuf','w') for ff in files] for l in lines: s = l.strip().split('|||') for ii, fd in enumerate(fds): print(s[ii], file=fd) [ff.close() for ff in fds] os.remove(tpath) if __name__ == '__main__': main(sys.argv[1:])
800
16.042553
71
py
sa-nmt
sa-nmt-master/data/scan_example.py
from __future__ import print_function import numpy import theano from theano import tensor # some numbers n_steps = 10 n_samples = 5 dim = 10 input_dim = 20 output_dim = 2 # one step function that will be used by scan def oneStep(x_t, h_tm1, W_x, W_h, W_o): h_t = tensor.tanh(tensor.dot(x_t, W_x) + tensor.dot(h_tm1, W_h)) o_t = tensor.dot(h_t, W_o) return h_t, o_t # spawn theano tensor variable, our symbolic input # a 3D tensor (n_steps, n_samples, dim) x = tensor.tensor3(dtype='float32') # initial state of our rnn init_state = tensor.alloc(0., n_samples, dim) # create parameters that we will use, # note that, parameters are theano shared variables # parameters for input to hidden states W_x_ = numpy.random.randn(input_dim, dim).astype('float32') W_x = theano.shared(W_x_) # parameters for hidden state transition W_h_ = numpy.random.randn(dim, dim).astype('float32') W_h = theano.shared(W_h_) # parameters from hidden state to output W_o_ = numpy.random.randn(dim, output_dim).astype('float32') W_o = theano.shared(W_o_) # scan function ([h_vals, o_vals], updates) = theano.scan( fn=oneStep, sequences=[x], outputs_info=[init_state, None], non_sequences=[W_x, W_h, W_o], n_steps=n_steps, strict=True) # let us now compile a function to get the output f = theano.function([x], [h_vals, o_vals]) # now we will call the compiled function with actual input actual_input = numpy.random.randn( n_steps, n_samples, input_dim).astype('float32') h_vals_, o_vals_ = f(actual_input) # print the shapes print('shape of input :', actual_input.shape) print('shape of h_vals:', h_vals_.shape) print('shape of o_vals:', o_vals_.shape)
1,709
23.782609
60
py
sa-nmt
sa-nmt-master/data/strip_sgml.py
from __future__ import print_function import sys import re def main(): fin = sys.stdin fout = sys.stdout for l in fin: line = l.strip() text = re.sub('<[^<]+>', "", line).strip() if len(text) == 0: continue print(text, file=fout) if __name__ == "__main__": main()
346
15.52381
50
py
sa-nmt
sa-nmt-master/data/build_dictionary.py
from __future__ import print_function import numpy import pickle as pkl import sys from collections import OrderedDict def main(): for filename in sys.argv[1:]: print('Processing', filename) word_freqs = OrderedDict() with open(filename, 'r') as f: for line in f: words_in = line.strip().split(' ') for w in words_in: if w not in word_freqs: word_freqs[w] = 0 word_freqs[w] += 1 words = list(word_freqs.keys()) freqs = list(word_freqs.values()) sorted_idx = numpy.argsort(freqs) sorted_words = [words[ii] for ii in sorted_idx[::-1]] worddict = OrderedDict() worddict['<pad>'] = 0 worddict['<unk>'] = 1 worddict['<bos>'] = 2 worddict['<eos>'] = 3 for ii, ww in enumerate(sorted_words): worddict[ww] = ii+4 with open('%s.pkl' % filename, 'wb') as f: pkl.dump(worddict, f) print('Done') if __name__ == '__main__': main()
1,087
24.904762
61
py
sa-nmt
sa-nmt-master/data/extract_files.py
#!/usr/bin/python import argparse import logging import os import tarfile TRAIN_DATA_URL = 'http://www.statmt.org/europarl/v7/fr-en.tgz' VALID_DATA_URL = 'http://matrix.statmt.org/test_sets/newstest2011.tgz' parser = argparse.ArgumentParser( description=""" This script donwloads parallel corpora given source and target pair language indicators. Adapted from, https://github.com/orhanf/blocks-examples/tree/master/machine_translation """, formatter_class=argparse.RawTextHelpFormatter) parser.add_argument("-s", "--source", type=str, help="Source language", default="fr") parser.add_argument("-t", "--target", type=str, help="Target language", default="en") parser.add_argument("--source-dev", type=str, default="newstest2011.fr", help="Source language dev filename") parser.add_argument("--target-dev", type=str, default="newstest2011.en", help="Target language dev filename") parser.add_argument("--outdir", type=str, default=".", help="Output directory") def extract_tar_file_to(file_to_extract, extract_into, names_to_look): extracted_filenames = [] try: logger.info("Extracting file [{}] into [{}]" .format(file_to_extract, extract_into)) tar = tarfile.open(file_to_extract, 'r') src_trg_files = [ff for ff in tar.getnames() if any([ff.find(nn) > -1 for nn in names_to_look])] if not len(src_trg_files): raise ValueError("[{}] pair does not exist in the archive!" .format(src_trg_files)) for item in tar: # extract only source-target pair if item.name in src_trg_files: file_path = os.path.join(extract_into, item.path) if not os.path.exists(file_path): logger.info("...extracting [{}] into [{}]" .format(item.name, file_path)) tar.extract(item, extract_into) else: logger.info("...file exists [{}]".format(file_path)) extracted_filenames.append( os.path.join(extract_into, item.path)) except Exception as e: logger.error("{}".format(str(e))) return extracted_filenames def main(): train_data_file = os.path.join(args.outdir, 'train_data.tgz') valid_data_file = os.path.join(args.outdir, 'valid_data.tgz') # Download europarl v7 and extract it extract_tar_file_to( train_data_file, os.path.dirname(train_data_file), ["{}-{}".format(args.source, args.target)]) # Download development set and extract it extract_tar_file_to( valid_data_file, os.path.dirname(valid_data_file), [args.source_dev, args.target_dev]) if __name__ == "__main__": logging.basicConfig(level=logging.INFO) logger = logging.getLogger('prepare_data') args = parser.parse_args() main()
2,999
36.974684
76
py
sa-nmt
sa-nmt-master/data/length.py
from __future__ import print_function import numpy import sys for name in sys.argv[1:]: lens = [] with open(name, 'r') as f: for ll in f: lens.append(len(ll.strip().split(' '))) print(name, ' max ', numpy.max(lens), ' min ', numpy.min(lens), ' average ', numpy.mean(lens))
310
19.733333
98
py
qsft
qsft-master/synt_exp/run-tests-complexity-vs-size.py
import numpy as np import sys import pandas as pd import uuid pd.set_option('display.max_rows', 500) pd.set_option('display.max_columns', 504) pd.set_option('display.width', 1000) sys.path.append("..") sys.path.append("../src") import argparse from pathlib import Path from synt_src.synthetic_helper import SyntheticHelper from qsft.parallel_tests import run_tests from synt_src.synthetic_signal import generate_signal_w if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument('--debug', type=bool, default=False) parser.add_argument('--num_subsample', type=int, nargs="+") parser.add_argument('--num_repeat', type=int, nargs="+") parser.add_argument('--b', type=int, nargs="+") parser.add_argument('--a', type=int) parser.add_argument('--snr', type=float) parser.add_argument('--n', type=int, nargs="+") parser.add_argument('--q', type=int) parser.add_argument('--t', type=int) parser.add_argument('--sparsity', type=int) parser.add_argument('--iters', type=int, default=1) parser.add_argument('--subsampling', type=int, default=True) parser.add_argument('--jobid', type=int) args = parser.parse_args() debug = args.debug if debug: args.num_subsample = [2] args.num_repeat = [2] args.b = [6] args.a = 1 args.n = np.linspace(40, 40, num=1, dtype=int) args.q = 3 args.t = 5 args.sparsity = 10 args.snr = 50 args.iters = 1 args.jobid = "debug-" + str(uuid.uuid1())[:8] args.subsampling = True if debug: exp_dir_base = Path(f"results/{str(args.jobid)}") else: exp_dir_base = Path(f"/global/scratch/users/erginbas/qsft/synt-exp-results/{str(args.jobid)}") exp_dir_base.mkdir(parents=True, exist_ok=True) (exp_dir_base / "figs").mkdir(exist_ok=True) print("Parameters :", args, flush=True) methods = ["qsft_coded"] dataframes = [] print("Starting the tests...", flush=True) for n_idx in range(len(args.n)): n = int(args.n[n_idx]) noise_sd = np.sqrt((args.sparsity * args.a ** 2) / (10 ** (args.snr / 10))) print(fr"n = {n}, N = {args.q ** n:.2e}, sigma = {noise_sd}") b_valid = [b for b in args.b if b <= n] subsampling_args = { "num_subsample": max(args.num_subsample), "num_repeat": max(args.num_repeat), "b": max(b_valid), "all_bs": b_valid } test_args = { "n_samples": 200000 } for it in range(args.iters): exp_dir = exp_dir_base / f"n{n}_i{it}" exp_dir.mkdir(parents=True, exist_ok=True) _, locq, strengths = generate_signal_w(n, args.q, args.sparsity, args.a, args.a, full=False, max_weight=args.t) signal_args = { "n": n, "q": args.q, "t": args.t, "locq": locq, "strengths": strengths, } helper = SyntheticHelper(signal_args=signal_args, methods=methods, subsampling=args.subsampling, exp_dir=exp_dir, subsampling_args=subsampling_args, test_args=test_args) for method in methods: if method == "lasso" and args.q ** n > 8000: pass else: dataframes.append(run_tests(method, helper, 1, args.num_subsample, args.num_repeat, b_valid, [noise_sd], parallel=False)) results_df = pd.concat(dataframes, ignore_index=True) results_df.to_pickle(exp_dir_base / "result.pkl")
3,745
31.017094
109
py
qsft
qsft-master/synt_exp/plot-complexity-vs-size.py
import numpy as np import matplotlib.pyplot as plt import sys import pandas as pd from matplotlib import ticker import matplotlib from mpl_toolkits.axes_grid1 import make_axes_locatable pd.set_option('display.max_rows', 500) pd.set_option('display.max_columns', 504) pd.set_option('display.width', 1000) sys.path.append("..") sys.path.append("../src") from pathlib import Path if __name__ == '__main__': font = {'family': 'sans', 'size': 12} matplotlib.rc('font', **font) exp_dir = Path(f"results/debug-510ff35e") results_df = pd.read_pickle(exp_dir / "result.pkl") group_by = ["method", "n", "q", "num_subsample", "num_repeat", "b", "noise_sd"] means = results_df.groupby(group_by, as_index=False).mean() stds = results_df.groupby(group_by, as_index=False).std() print(means) q = results_df["q"][0] # fig, ax = plt.subplots(2, 2) colorMap = plt.get_cmap('cividis_r') sampleAlpha = 0.6 timeAlpha = 0.8 methods = ["qsft_coded"] method_names = ["q-SFT Coded"] sample_comp = [[] for _ in methods] time_comp = [[] for _ in methods] for i in means.index: mean_row = means.iloc[i] m = np.where(mean_row["method"] == np.array(methods))[0][0] if m != -1: sample_comp[m].append((mean_row["n"], mean_row["n_samples"], mean_row["nmse"])) time_comp[m].append((mean_row["n"], mean_row["runtime"], mean_row["nmse"])) min_samples = np.min(np.array(sample_comp[0]), axis=0)[1] max_samples = np.max(np.array(sample_comp[0]), axis=0)[1] sample_bin_count = 10 ns = np.unique(np.array(sample_comp[0])[:, 0]) sample_bins = np.linspace(np.log10(min_samples), np.log10(max_samples), sample_bin_count + 1)[1:] (exp_dir / "figs").mkdir(exist_ok=True) for m in range(len(methods)): fig, ax = plt.subplots(1, 1, figsize=(4.3, 3.2), dpi=300) sample_comp_m = np.array(sample_comp[m]) sample_bin_totals = np.zeros((len(ns), sample_bin_count)) sample_bin_counts = np.zeros((len(ns), sample_bin_count)) for row in sample_comp_m: sample_bin = np.where(sample_bins >= np.log10(row[1]))[0][0] n_bin = np.where(int(row[0]) == ns)[0][0] sample_bin_totals[n_bin, sample_bin] += row[2] sample_bin_counts[n_bin, sample_bin] += 1 sample_bin_avg = sample_bin_totals / sample_bin_counts for n_idx in range(len(ns)): sample_row = sample_bin_avg[n_idx] if len(np.where(~np.isnan(sample_row))[0]) > 0: first_non_nan = np.where(~np.isnan(sample_row))[0][0] sample_row[:first_non_nan] = sample_row[first_non_nan] last_non_nan = np.where(~np.isnan(sample_row))[0][-1] sample_row[last_non_nan+1:] = sample_row[last_non_nan] sample_bin_avg[n_idx] = sample_row sample_bin_avg = sample_bin_avg.T sample_bin_avg = np.minimum(sample_bin_avg, 1) masked_sample_bin_avg = np.ma.array(sample_bin_avg, mask=np.isnan(sample_bin_avg)) colorMap.set_bad('lightgrey') divider = make_axes_locatable(ax) cax = divider.append_axes('right', size='5%', pad=0.05) data = ax.pcolormesh(ns, 10 ** sample_bins, masked_sample_bin_avg, cmap=colorMap) ax.set_yscale("log") ax.set_xlabel("n\nN") ax.set_ylabel('Sample Complexity') ax.set_xticks(ns[1::3]) # ax.set_yticks([10, 10**2, 10**3, 10**4, 10**5]) ax.xaxis.set_major_formatter(ticker.FormatStrFormatter(rf"${{%d}}$")) ax.xaxis.set_label_coords(-0.03, -.04) secax = ax.secondary_xaxis('bottom', functions=(lambda x: x, lambda x: x)) secax.xaxis.set_major_formatter(ticker.FormatStrFormatter("".join(["\n", rf"${q}^{{%d}}$"]))) secax.set_xticks(ns[1::3]) cbar = fig.colorbar(data, cax=cax, orientation='vertical') cbar.ax.set_yticks([0, 0.2, 0.4, 0.6, 0.8, 1.0]) cbar.set_label('NMSE', rotation=270, labelpad=15) plt.tight_layout() plt.savefig(exp_dir / f'figs/complexity-vs-n-{methods[m]}.pdf', bbox_inches='tight', transparent="True", pad_inches=0) plt.show() _, ax = plt.subplots(1, 1, figsize=(4, 3.2), dpi=300) for m in range(len(methods)): time_comp_m = np.array(time_comp[m]) total_success_time = np.zeros(len(ns)) count_success_time = np.zeros(len(ns)) for row in time_comp_m: if row[2] < 0.1: n_bin = np.where(int(row[0]) == ns)[0][0] total_success_time[n_bin] += row[1] count_success_time[n_bin] += 1 avg_success_time = total_success_time / count_success_time ax.plot(ns, avg_success_time, "o-", label=method_names[m]) ax.set_xlabel("n\nN") ax.set_ylabel('Runtime Complexity (sec)', fontsize=11) ax.set_yscale("log") ax.grid(True) ax.set_yticks([0.01, 0.1, 1, 10, 100]) ax.set_xticks(ns[1::3]) ax.xaxis.set_major_formatter(ticker.FormatStrFormatter(rf"${{%d}}$")) ax.xaxis.set_label_coords(-0.03, -.04) secax = ax.secondary_xaxis('bottom', functions=(lambda x: x, lambda x: x)) secax.xaxis.set_major_formatter(ticker.FormatStrFormatter("".join(["\n", rf"${q}^{{%d}}$"]))) secax.set_xticks(ns[1::3]) ax.legend() plt.tight_layout() plt.savefig(exp_dir / f'figs/complexity-vs-n-runtime.pdf', bbox_inches='tight', transparent="True", pad_inches=0.1) plt.show()
5,591
36.033113
101
py
qsft
qsft-master/synt_exp/run-tests-nmse-vs-snr.py
import numpy as np import sys import pandas as pd import uuid pd.set_option('display.max_rows', 500) pd.set_option('display.max_columns', 504) pd.set_option('display.width', 1000) sys.path.append("..") import argparse from pathlib import Path from synt_exp.synt_src.synthetic_helper import SyntheticHelper from qsft.parallel_tests import run_tests from synt_exp.synt_src.synthetic_signal import generate_signal_w if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument('--debug', type=bool, default=False) parser.add_argument('--num_subsample', type=int, nargs="+") parser.add_argument('--num_repeat', type=int, nargs="+") parser.add_argument('--b', type=int, nargs="+") parser.add_argument('--a', type=int) parser.add_argument('--snr', type=float, nargs="+") parser.add_argument('--n', type=int) parser.add_argument('--q', type=int) parser.add_argument('--sparsity', type=int, nargs="+") parser.add_argument('--iters', type=int, default=1) parser.add_argument('--subsampling', type=int, default=True) parser.add_argument('--jobid', type=int) args = parser.parse_args() debug = args.debug if debug: args.num_subsample = [3] args.num_repeat = [1] args.b = [7] args.a = 1 args.n = 20 args.q = 4 args.sparsity = [100, 250, 1000] args.iters = 5 args.jobid = "debug-" + str(uuid.uuid1())[:8] args.subsampling = True args.snr = np.linspace(-25, 5, num=30) if debug: exp_dir_base = Path(f"results/{str(args.jobid)}") else: exp_dir_base = Path(f"/global/scratch/users/erginbas/qsft/synt-exp-results/{str(args.jobid)}") exp_dir_base.mkdir(parents=True, exist_ok=True) (exp_dir_base / "figs").mkdir(exist_ok=True) print("Parameters :", args, flush=True) methods = ["qsft"] dataframes = [] print("Starting the tests...", flush=True) subsampling_args = { "num_subsample": max(args.num_subsample), "num_repeat": max(args.num_repeat), "b": max(args.b), "all_bs": args.b } test_args = { "n_samples": 50000 } print() print("n = {}, N = {:.2e}".format(args.n, args.q ** args.n)) for s in range(len(args.sparsity)): sparsity = args.sparsity[s] noise_sd = np.sqrt((sparsity * args.a ** 2) / (args.q ** args.n * 10 ** (np.array(args.snr) / 10))) for it in range(args.iters): exp_dir = exp_dir_base / f"s{sparsity}_i{it}" exp_dir.mkdir(parents=True, exist_ok=True) _, locq, strengths = generate_signal_w(args.n, args.q, sparsity, args.a, args.a, full=False) signal_args = { "n": args.n, "q": args.q, "locq": locq, "strengths": strengths, } helper = SyntheticHelper(signal_args=signal_args, methods=methods, subsampling=args.subsampling, exp_dir=exp_dir, subsampling_args=subsampling_args, test_args=test_args) for method in methods: run_df = run_tests(method, helper, 1, args.num_subsample, args.num_repeat, args.b, noise_sd, parallel=False) run_df["sparsity"] = sparsity dataframes.append(run_df) results_df = pd.concat(dataframes, ignore_index=True) results_df['snr'] = 10 * np.log10((results_df['sparsity'] / (results_df["q"] ** results_df["n"])) * \ (args.a ** 2 / results_df["noise_sd"] ** 2)) print(results_df) results_df.to_pickle(exp_dir_base / "result.pkl")
3,712
32.45045
109
py
qsft
qsft-master/synt_exp/qsft-sample-vs-nmse.py
#!/usr/bin/env python # coding: utf-8 # In[1]: import numpy as np import matplotlib.pyplot as plt import sys import pandas as pd import uuid sys.path.append("..") sys.path.append("../src") from qsft.utils import best_convex_underestimator import argparse from pathlib import Path from synt_exp.synt_src.synthetic_helper import SyntheticHelper from qsft.parallel_tests import run_tests parser = argparse.ArgumentParser() parser.add_argument('--debug', type=bool, default=False) parser.add_argument('--num_subsample', type=int, nargs="+") parser.add_argument('--num_repeat', type=int, nargs="+") parser.add_argument('--b', type=int, nargs="+") parser.add_argument('--a', type=int) parser.add_argument('--noise_sd', type=float, nargs="+") parser.add_argument('--n', type=int) parser.add_argument('--q', type=int) parser.add_argument('--sparsity', type=int) parser.add_argument('--iters', type=int, default=1) parser.add_argument('--subsampling', type=int, default=True) parser.add_argument('--jobid', type=int) args = parser.parse_args() debug = args.debug if debug: args.num_subsample = [3] args.num_repeat = [2] args.b = [4] args.a = 1 args.n = 6 args.q = 4 args.sparsity = 1 args.noise_sd = [1e-3] args.iters = 1 args.jobid = "debug-" + str(uuid.uuid1())[:8] args.subsampling = True if debug: exp_dir = Path(f"results/{str(args.jobid)}") else: exp_dir = Path(f"/global/scratch/users/erginbas/qsft/synt-exp-results/{str(args.jobid)}") print("Parameters :", args, flush=True) query_args = { "query_method": "complex", "delays_method": "nso", "num_subsample": max(args.num_subsample), "num_repeat": max(args.num_repeat), "b": max(args.b), "all_bs": args.b } methods = ["qsft", "lasso"] colors = ["red", "blue"] test_args = { "n_samples": 50000 } print("Loading/Calculating data...", flush=True) exp_dir.mkdir(parents=True, exist_ok=True) (exp_dir / "figs").mkdir(exist_ok=True) helper = SyntheticHelper(args.n, args.q, noise_sd=args.noise_sd[0], sparsity=args.sparsity, a_min=args.a, a_max=args.a, baseline_methods=methods, subsampling=args.subsampling, exp_dir=exp_dir, query_args=query_args, test_args=test_args) print("n = {}, N = {:.2e}".format(args.n, args.q ** args.n)) print("Starting the tests...", flush=True) fig, ax = plt.subplots() for m in range(len(methods)): # Test QSFT with different parameters # Construct a grid of parameters. For each entry, run multiple test rounds. # Compute the average for each parameter selection. results_df = run_tests(methods[m], helper, args.iters, args.num_subsample, args.num_repeat, args.b, args.noise_sd, parallel=False) # results_df.to_csv(f'results/{str(args.jobid)}/results_df_{methods[m]}.csv') means = results_df.groupby(["num_subsample", "num_repeat", "b", "noise_sd"], as_index=False).mean() stds = results_df.groupby(["num_subsample", "num_repeat", "b", "noise_sd"], as_index=False).std() pd.set_option('display.max_rows', None) pd.set_option('display.max_columns', None) print(results_df) x_values = [] y_values = [] labels = [] all_points = [] for i in means.index: mean_row = means.iloc[i] std_row = stds.iloc[i] ax.errorbar(mean_row['n_samples'], mean_row['nmse'], xerr=std_row['n_samples'], yerr=std_row['nmse'], fmt="o", color=colors[m]) all_points.append([mean_row['n_samples'], mean_row['nmse']]) label = f'({int(mean_row["b"])},{int(mean_row["num_subsample"])},{int(mean_row["num_repeat"])})' labels.append(label) for i in range(len(all_points)): ax.annotate(labels[i], xy=all_points[i], xycoords='data', xytext=(20, 10), textcoords='offset points', arrowprops=dict(arrowstyle="->", shrinkA=0, shrinkB=5, connectionstyle="arc3,rad=0.4", color='blue'), ) try: if len(all_points) > 3: bcue = best_convex_underestimator(np.array(all_points)) ax.plot(bcue[:, 0], bcue[:, 1], 'r--', lw=1.5, label="Best Cvx Underest.") except: pass ax.set_xlabel('Total Samples') ax.set_ylabel('Test NMSE') plt.legend() plt.grid() plt.savefig(exp_dir / f'figs/nmse-vs-sample.png') plt.show()
4,482
30.131944
104
py
qsft
qsft-master/synt_exp/plot-nmse-vs-snr.py
import numpy as np import matplotlib.pyplot as plt import sys import pandas as pd pd.set_option('display.max_rows', 500) pd.set_option('display.max_columns', 504) pd.set_option('display.width', 1000) sys.path.append("..") sys.path.append("../src") from pathlib import Path if __name__ == '__main__': exp_dir = Path(f"results_final/debug-e51c0b3c") results_df = pd.read_pickle(exp_dir / "result.pkl") group_by = ["method", "n", "q", "num_subsample", "num_repeat", "b", "noise_sd", "sparsity"] means = results_df.groupby(group_by, as_index=False).mean() stds = results_df.groupby(group_by, as_index=False).std() _, ax = plt.subplots(1, 1, figsize=(3.5, 3), dpi=300) sparsity_list = np.unique(results_df["sparsity"]) snr_values = [[] for _ in range(len(sparsity_list))] for i in means.index: mean_row = means.iloc[i] s_idx = np.where(mean_row["sparsity"] == sparsity_list)[0][0] snr_values[s_idx].append((mean_row["snr"], mean_row["nmse"])) for s in range(len(sparsity_list)): ax.plot(*zip(*snr_values[s]), "o-", label=f"S = {sparsity_list[s]}", markersize=4) ax.set_xlabel('SNR (dB)') ax.set_ylabel('NMSE') ax.legend() ax.grid() (exp_dir / "figs").mkdir(exist_ok=True) plt.tight_layout() plt.savefig(exp_dir / f'figs/nmse-vs-snr.pdf', bbox_inches='tight', transparent="True", pad_inches=0) plt.show()
1,435
27.156863
95
py
qsft
qsft-master/synt_exp/quick_example.py
import numpy as np from qsft.qsft import QSFT from qsft.query import get_reed_solomon_dec from synt_exp.synt_src.synthetic_signal import get_random_subsampled_signal if __name__ == '__main__': np.random.seed(20) q = 3 n = 40 N = q ** n sparsity = 100 a_min = 1 a_max = 1 b = 4 noise_sd = 1 num_subsample = 3 num_repeat = 1 t = 4 decoder = get_reed_solomon_dec(n, t, q) delays_method_source = "identity" delays_method_channel = "identity" query_args = { "query_method": "complex", "num_subsample": num_subsample, "delays_method_source": delays_method_source, "subsampling_method": "qsft", "delays_method_channel": delays_method_channel, "num_repeat": num_repeat, "b": b, "t": t } qsft_args = { "num_subsample": num_subsample, "num_repeat": num_repeat, "reconstruct_method_source": delays_method_source, "reconstruct_method_channel": delays_method_channel, "b": b, "noise_sd": noise_sd, "source_decoder": decoder } ''' Generate a Signal Object ''' test_signal = get_random_subsampled_signal(n=n, q=q, sparsity=sparsity, a_min=a_min, a_max=a_max, noise_sd=noise_sd, query_args=query_args, max_weight=t) ''' Create a QSFT instance and perform the transformation ''' sft = QSFT(**qsft_args) result = sft.transform(test_signal, verbosity=10, timing_verbose=True, report=True, sort=True) ''' Display the Reported Results ''' gwht = result.get("gwht") loc = result.get("locations") n_used = result.get("n_samples") peeled = result.get("locations") avg_hamming_weight = result.get("avg_hamming_weight") max_hamming_weight = result.get("max_hamming_weight") print("found non-zero indices QSFT: ") print(peeled) print("True non-zero indices: ") print(test_signal.locq.T) print("Total samples = ", n_used) print("Total sample ratio = ", n_used / q ** n) signal_w_diff = test_signal.signal_w.copy() for key in gwht.keys(): signal_w_diff[key] = signal_w_diff.get(key, 0) - gwht[key] print("NMSE SPRIGHT= ", np.sum(np.abs(list(signal_w_diff.values())) ** 2) / np.sum(np.abs(list(test_signal.signal_w.values())) ** 2)) print("AVG Hamming Weight of Nonzero Locations = ", avg_hamming_weight) print("Max Hamming Weight of Nonzero Locations = ", max_hamming_weight)
2,810
32.86747
118
py
qsft
qsft-master/synt_exp/__init__.py
0
0
0
py
qsft
qsft-master/synt_exp/synt_src/synthetic_signal.py
import numpy as np from qsft.utils import igwht_tensored, random_signal_strength_model, qary_vec_to_dec, sort_qary_vecs from qsft.input_signal import Signal from qsft.input_signal_subsampled import SubsampledSignal from qsft.utils import dec_to_qary_vec from multiprocess import Pool import time def generate_signal_w(n, q, sparsity, a_min, a_max, noise_sd=0, full=True, max_weight=None): """ Generates a sparse fourier transform """ max_weight = n if max_weight is None else max_weight N = q ** n if max_weight == n: locq = sort_qary_vecs(np.random.randint(q, size=(n, sparsity)).T).T else: non_zero_idx_vals = np.random.randint(q-1, size=(max_weight, sparsity))+1 non_zero_idx_pos = np.random.choice(a=n, size=(sparsity, max_weight)) locq = np.zeros((n, sparsity), dtype=int) for i in range(sparsity): locq[non_zero_idx_pos[i, :], i] = non_zero_idx_vals[:, i] locq = sort_qary_vecs(locq.T).T loc = qary_vec_to_dec(locq, q) strengths = random_signal_strength_model(sparsity, a_min, a_max) if full: wht = np.zeros((N,), dtype=complex) for l, s in zip(loc, strengths): wht[l] = s signal_w = wht + np.random.normal(0, noise_sd, size=(N, 2)).view(np.complex).reshape(N) return np.reshape(signal_w, [q] * n), locq, strengths else: signal_w = dict(zip(list(map(tuple, locq.T)), strengths)) return signal_w, locq, strengths def get_random_signal(n, q, noise_sd, sparsity, a_min, a_max): """ Computes a full random time-domain signal, which is sparse in the fequency domain. This function is only suitable for small n since for large n, storing all q^n symbols is not tractable. """ signal_w, locq, strengths = generate_signal_w(n, q, noise_sd, sparsity, a_min, a_max, full=True) signal_t = igwht_tensored(signal_w, q, n) signal_params = { "n": n, "q": q, "noise_sd": noise_sd, "signal_t": signal_t, "signal_w": signal_w, "folder": "test_data" } return SyntheticSignal(locq, strengths, **signal_params) class SyntheticSignal(Signal): """ This is essentially just a signal object, except the strengths and locations of the non-zero indicies are known, and included as attributes """ def __init__(self, locq, strengths, **kwargs): super().__init__(**kwargs) self.locq = locq self.strengths = strengths def get_random_subsampled_signal(n, q, noise_sd, sparsity, a_min, a_max, query_args, max_weight=None): """ Similar to get_random_signal, but instead of returning a SyntheticSignal object, it returns a SyntheticSubsampledSignal object. The advantage of this is that a subsampled signal does not compute the time domain signal on creation, but instead, creates it on the fly. This should be used (1) when n is large or (2) when sampling is expensive. """ start_time = time.time() signal_w, locq, strengths = generate_signal_w(n, q, sparsity, a_min, a_max, noise_sd, full=False, max_weight=max_weight) signal_params = { "n": n, "q": q, "query_args": query_args, } print(f"Generation Time:{time.time() - start_time}", flush=True) return SyntheticSubsampledSignal(signal_w=signal_w, locq=locq, strengths=strengths, noise_sd=noise_sd, **signal_params) class SyntheticSubsampledSignal(SubsampledSignal): """ This is a Subsampled signal object, except it implements the unimplemented 'subsample' function. """ def __init__(self, **kwargs): self.q = kwargs["q"] self.n = kwargs["n"] self.locq = kwargs["locq"] self.noise_sd = kwargs["noise_sd"] freq_normalized = 2j * np.pi * kwargs["locq"] / kwargs["q"] strengths = kwargs["strengths"] def sampling_function(query_batch): query_indices_qary_batch = np.array(dec_to_qary_vec(query_batch, self.q, self.n)).T return np.exp(query_indices_qary_batch @ freq_normalized) @ strengths self.sampling_function = sampling_function super().__init__(**kwargs) def subsample(self, query_indices): """ Computes the signal/function values at the queried indicies on the fly """ batch_size = 10000 res = [] query_indices_batches = np.array_split(query_indices, len(query_indices)//batch_size + 1) with Pool() as pool: for new_res in pool.imap(self.sampling_function, query_indices_batches): res = np.concatenate((res, new_res)) return res def get_MDU(self, ret_num_subsample, ret_num_repeat, b, trans_times=False): """ wraps get_MDU method from SubsampledSignal to add synthetic noise """ mdu = super().get_MDU(ret_num_subsample, ret_num_repeat, b, trans_times) for i in range(len(mdu[2])): for j in range(len(mdu[2][i])): size = np.array(mdu[2][i][j]).shape nu = self.noise_sd / np.sqrt(2 * self.q ** b) mdu[2][i][j] += np.random.normal(0, nu, size=size + (2,)).view(np.complex).reshape(size) return mdu
5,257
39.137405
124
py
qsft
qsft-master/synt_exp/synt_src/synthetic_helper.py
from qsft.test_helper import TestHelper from synt_exp.synt_src.synthetic_signal import SyntheticSubsampledSignal class SyntheticHelper(TestHelper): def generate_signal(self, signal_args): return SyntheticSubsampledSignal(**signal_args)
249
34.714286
72
py
qsft
qsft-master/rna_exp/run-tests-complexity-vs-size.py
import numpy as np import pandas as pd import uuid import argparse from pathlib import Path from rna_exp.rna_src.rna_helper import RNAHelper from qsft.parallel_tests import run_tests pd.set_option('display.max_rows', 500) pd.set_option('display.max_columns', 504) pd.set_option('display.width', 1000) if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument('--debug', type=bool, default=False) parser.add_argument('--num_subsample', type=int, nargs="+") parser.add_argument('--num_repeat', type=int, nargs="+") parser.add_argument('--b', type=int, nargs="+") parser.add_argument('--noise_sd', type=float, nargs="+") parser.add_argument('--n', type=int, nargs="+") parser.add_argument('--iters', type=int, default=1) parser.add_argument('--subsampling', type=int, default=True) parser.add_argument('--jobid', type=int) args = parser.parse_args() debug = args.debug if debug: args.num_subsample = [4] args.num_repeat = [3] args.b = [7] args.n = [15] args.iters = 1 args.jobid = "debug-" + str(uuid.uuid1())[:8] args.subsampling = True args.noise_sd = np.logspace(-3, -5, num=10) if debug: exp_dir_base = Path(f"results/{str(args.jobid)}") else: exp_dir_base = Path(f"/global/scratch/users/erginbas/qsft/synt-exp-results/{str(args.jobid)}") args.q = 4 exp_dir_base.mkdir(parents=True, exist_ok=True) (exp_dir_base / "figs").mkdir(exist_ok=True) print("Parameters :", args, flush=True) methods = ["qsft"] dataframes = [] print("Starting the tests...", flush=True) for n_idx in range(len(args.n)): n = args.n[n_idx] b_valid = [b for b in args.b if b <= n] subsampling_args = { "num_subsample": max(args.num_subsample), "num_repeat": max(args.num_repeat), "b": max(b_valid), "all_bs": b_valid } test_args = { "n_samples": 5000 } print() print(fr"n = {n}, N = {args.q ** n}, sigma = {args.noise_sd}") for it in range(args.iters): exp_dir = exp_dir_base / f"i{it}" exp_dir.mkdir(parents=True, exist_ok=True) signal_args = { "n": n, "q": args.q } helper = RNAHelper(signal_args=signal_args, methods=methods, subsampling=args.subsampling, exp_dir=exp_dir, subsampling_args=subsampling_args, test_args=test_args) for method in methods: if method == "lasso" and args.q ** n > 3000: pass else: dataframes.append(run_tests(method, helper, 1, args.num_subsample, args.num_repeat, b_valid, args.noise_sd, parallel=False)) results_df = pd.concat(dataframes, ignore_index=True) print() print(results_df) results_df.to_pickle(exp_dir_base / "result.pkl")
3,056
29.57
103
py
qsft
qsft-master/rna_exp/plot-complexity-vs-size.py
import numpy as np import matplotlib.pyplot as plt import sys import pandas as pd from scipy import interpolate pd.set_option('display.max_rows', 500) pd.set_option('display.max_columns', 504) pd.set_option('display.width', 1000) sys.path.append("..") sys.path.append("../src") from pathlib import Path from qsft.utils import best_convex_underestimator import scipy if __name__ == '__main__': exp_nums = ["13844137", "13844138", "13846802", "13846805"] dfs = [] for exp_id in exp_nums: exp_dir = Path(f"results_final/{exp_id}") dfs.append(pd.read_pickle(exp_dir / "result.pkl")) results_df = pd.concat(dfs, ignore_index=True) group_by = ["method", "n", "q", "num_subsample", "num_repeat", "b"] results_df = results_df.groupby(group_by, as_index=False).min() # fig, ax = plt.subplots(2, 2) colorMap = plt.get_cmap('cividis_r') sampleAlpha = 0.6 timeAlpha = 0.8 methods = ["qsft"] _, ax = plt.subplots(1, 1, figsize=(4, 3)) ns = results_df["n"].unique() for n in ns: df_n = results_df.loc[results_df['n'] == n] xs = df_n["n_samples"] / 10 ys = 10 * df_n["nmse"] / np.log(df_n["n_samples"]) all_points = list(zip(xs, ys)) model = lambda t, a, b: a + b * np.exp(-t) a, _ = scipy.optimize.curve_fit(model, np.log10(xs), ys) polyline_points = np.logspace(2.6, 7, 50) plt.plot(polyline_points, [model(np.log10(p), a[0], a[1]) for p in polyline_points]) # if len(all_points) > 3: # bcue = best_convex_underestimator(np.array(all_points)) # ax.plot(bcue[:, 0], bcue[:, 1], lw=1.5, label=f"n={n}") ax.scatter(*zip(*all_points), s=10) ax.set_xscale("log") ax.set_xlabel('Sample Complexity') ax.set_ylabel('Test NMSE') ax.legend() ax.grid() # ax.set_xticks(ns) # for m in range(len(methods)): # time_comp_m = np.array(time_comp[m]) # total_success_time = np.zeros(len(ns)) # count_success_time = np.zeros(len(ns)) # for row in time_comp_m: # if row[2] < 0.1: # n_bin = np.where(int(row[0]) == ns)[0][0] # total_success_time[n_bin] += row[1] # count_success_time[n_bin] += 1 # # avg_success_time = total_success_time / count_success_time # # print(total_success_time) # # ax[-1].plot(ns, avg_success_time, "o-") # ax[-1].set_xlabel('n') # ax[-1].set_ylabel('Runtime Complexity (sec)') # ax[-1].set_yscale("log") # ax[-1].grid(True) # ax[-1].set_xticks(ns) plt.tight_layout() Path(f"figs/").mkdir(exist_ok=True) plt.savefig('figs/complexity-vs-n-rna.pdf', bbox_inches='tight', transparent="True", pad_inches=0.1) plt.show()
2,831
30.120879
92
py
qsft
qsft-master/rna_exp/qspright-sample-vs-nmse.py
#!/usr/bin/env python # coding: utf-8 # In[1]: import numpy as np import matplotlib.pyplot as plt import sys sys.path.append("..") sys.path.append("../src") import pandas as pd import uuid from rna_exp.rna_src.rna_helper import RNAHelper from qsft.utils import best_convex_underestimator from qsft.parallel_tests import run_tests import argparse from pathlib import Path parser = argparse.ArgumentParser() parser.add_argument('--debug', type=bool, default=False) parser.add_argument('--num_subsample', type=int, nargs="+") parser.add_argument('--num_repeat', type=int, nargs="+") parser.add_argument('--b', type=int, nargs="+") parser.add_argument('--noise_sd', type=float, nargs="+") parser.add_argument('--n', type=int) parser.add_argument('--iters', type=int, default=1) parser.add_argument('--subsampling', type=int, default=True) parser.add_argument('--jobid', type=int) args = parser.parse_args() debug = args.debug if debug: args.num_subsample = [2, 3, 4] args.num_repeat = [4, 6] # args.b = [7, 8] # args.n = 15 # args.noise_sd = np.logspace(-3.7, -4.3, num=3) args.b = [2, 3, 4] args.n = 8 args.noise_sd = np.logspace(-2, -3.5, num=5) args.iters = 1 args.jobid = "debug-" + str(uuid.uuid1())[:8] args.subsampling = True exp_dir = Path(f"results/{str(args.jobid)}") else: exp_dir = Path(f"/global/scratch/users/erginbas/qsft/rna-exp-results/{str(args.jobid)}") print(exp_dir) print("Parameters :", args, flush=True) query_args = { "query_method": "complex", "delays_method": "nso", "num_subsample": max(args.num_subsample), "num_repeat": max(args.num_repeat), "b": max(args.b), "all_bs": args.b } methods = ["qsft", "lasso"] colors = ["red", "blue", "green", "purple"] test_args = { "n_samples": 50000 } print("Loading/Calculating data...", flush=True) exp_dir.mkdir(parents=True, exist_ok=True) (exp_dir / "figs").mkdir(exist_ok=True) print(exp_dir) helper = RNAHelper(args.n, baseline_methods=methods, subsampling=args.subsampling, query_args=query_args, test_args=test_args, exp_dir=exp_dir) n = helper.n q = 4 print("n = {}, N = {:.2e}".format(n, q ** n)) print("Starting the tests...", flush=True) fig, ax = plt.subplots() for m in range(len(methods)): # Test QSFT with different parameters # Construct a grid of parameters. For each entry, run multiple test rounds. # Compute the average for each parameter selection. results_df = run_tests(methods[m], helper, args.iters, args.num_subsample, args.num_repeat, args.b, args.noise_sd, parallel=False) # results_df.to_csv(f'results/{str(args.jobid)}/results_df_{methods[m]}.csv') means = results_df.groupby(["num_subsample", "num_repeat", "b", "noise_sd"], as_index=False).mean() stds = results_df.groupby(["num_subsample", "num_repeat", "b", "noise_sd"], as_index=False).std() pd.set_option('display.max_rows', None) pd.set_option('display.max_columns', None) print(results_df) x_values = [] y_values = [] labels = [] all_points = [] for i in means.index: mean_row = means.iloc[i] std_row = stds.iloc[i] ax.errorbar(mean_row['n_samples'], mean_row['nmse'], xerr=std_row['n_samples'], yerr=std_row['nmse'], fmt="o", color=colors[m]) all_points.append([mean_row['n_samples'], mean_row['nmse']]) label = f'({int(mean_row["b"])},{int(mean_row["num_subsample"])},{int(mean_row["num_repeat"])})' labels.append(label) for i in range(len(all_points)): ax.annotate(labels[i], xy=all_points[i], xycoords='data', xytext=(20, 10), textcoords='offset points', arrowprops=dict(arrowstyle="->", shrinkA=0, shrinkB=5, connectionstyle="arc3,rad=0.4", color='blue'), ) try: if len(all_points) > 3: bcue = best_convex_underestimator(np.array(all_points)) ax.plot(bcue[:, 0], bcue[:, 1], 'r--', lw=1.5, label="Best Cvx Underest.") except: pass ax.set_xlabel('Total Samples') ax.set_ylabel('Test NMSE') plt.legend() plt.grid() plt.savefig(exp_dir / f'figs/nmse-vs-sample.png') plt.show()
4,317
30.289855
104
py
qsft
qsft-master/rna_exp/__init__.py
0
0
0
py
qsft
qsft-master/rna_exp/rna_src/query_iterator.py
import numpy as np from qsft.utils import dec_to_qary_vec class QueryIterator(object): nucs = np.array(["A", "U", "C", "G"]) q = 4 def __init__(self, base_seq, positions, query_indices, q): self.base_seq = np.array(list(base_seq)) self.positions = positions self.full = self.base_seq.copy() self.length = len(query_indices) self.q = q self.query_indices = query_indices self.i = 0 def __iter__(self): return self def __next__(self): if self.q == 4: return self._next_q4() elif self.q == 2: return self._next_q2() def _next_q4(self): if self.i == self.length: raise StopIteration idx = self.query_indices[self.i][0] idx_dec = self.query_indices[self.i][1] seq = self.nucs[idx] self.full[:] = self.base_seq self.full[self.positions] = seq self.i += 1 return idx_dec, "".join(self.full) def _next_q2(self): if self.i == self.length: raise StopIteration idx_bin = self.query_indices[self.i][0] idx_dec = self.query_indices[self.i][1] idx = dec_to_qary_vec([idx_dec], 4, len(idx_bin)//2)[:, 0] seq = self.nucs[idx] self.full[:] = self.base_seq self.full[self.positions] = seq self.i += 1 return idx_dec, "".join(self.full) def __len__(self): return self.length
1,466
28.34
66
py
qsft
qsft-master/rna_exp/rna_src/utils.py
import numpy as np from tqdm import tqdm from itertools import chain, combinations def divisors(num): """Returns all divisors of a given integer""" divs = [] for x in range (1, num): if (num % x) == 0: divs.append(x) return divs def powerset(iterable): """Returns the powerset of a given set""" s = list(iterable) return chain.from_iterable(combinations(s, r) for r in range(len(s)+1)) def get_all_interactions(L, index_1=False): """ Returns a list of all epistatic interactions for a given sequence length. This sets of the order used for beta coefficients throughout the code. If index_1=True, then returns epistatic interactions corresponding to 1-indexing. """ if index_1: pos = range(1, L+1) else: pos = range(L) all_U = list(powerset(pos)) return all_U def complete_graph_evs(q): """ Returns a set of eigenvectors of complete graph of size q as column vectors of a matrix """ x = np.ones(q) y = np.eye(q)[0] v = x - np.linalg.norm(x, ord=2) * y w = v / np.linalg.norm(v, ord=2) w = w.reshape(q, 1) P = np.eye(q) - 2*np.dot(w, w.T) return P def fourier_basis_recursive(L, q): """ Recursively constructs the Fourier basis corresponding to the H(L, q) hamming graph (i.e. graphs of sequences of length L with alphabet size q). Hamming graphs are the Cartesian product H(L+1, q) = H(L, q) x Kq, and the eigevectors of a cartesian product (A x B) are the Kronecker product of the eigenvectors of A and B. This method is much faster than constructing the the Fourier representation for each sequence individually (as with fourier_for_seqs, below) but the resulting basis less interpretable (i.e. it is difficult to associate rows with particular sequences and columns with particular epistatic interactions). In all of this code, we enforce that beta is ordered according to epistatic interactions, so DO NOT multiply the basis resulting from this method by a sample of beta. """ Pq = complete_graph_evs(q) phi = np.copy(Pq) for i in range(L-1): phi = np.kron(Pq, phi) return phi def get_encodings(qs): """ Returns a length L list of arrays containing the encoding vectors corresponding to each alphabet element at each position in sequence, given the alphabet size at each position. """ encodings = [] Pqs = [] L = len(qs) for i in range(L): qi = qs[i] Pq = complete_graph_evs(qi) * np.sqrt(qi) Pqs.append(Pq) enc_i = Pq[:, 1:] encodings.append(enc_i) return encodings def fourier_for_seq(int_seq, encodings): """ Returns an M x 1 array containing the Fourier encoding of a sequence, given the integer representation of the sequence and the encodings returned by get_encodings, where M = prod(qs) and qs is the alphabet size at each position. """ L = len(int_seq) all_U = get_all_interactions(L) all_U = [list(U) for U in all_U] epi_encs = [] enc_1 = encodings[0][int_seq[0]] for U in all_U: if len(U) > 0 and 0 == U[0]: U_enc = enc_1 U.pop(0) else: U_enc = np.array([1]) epi_encs.append(U_enc) for l in range(1, L): enc_l = encodings[l][int_seq[l]] for k, U in enumerate(all_U): U_enc = epi_encs[k] if len(U) > 0 and l==U[0]: U_enc = np.kron(U_enc,enc_l) U.pop(0) epi_encs[k] = U_enc all_enc = np.concatenate(epi_encs) return all_enc def fourier_from_seqs(int_seqs, qs): """ Returns an N x M array containing the Fourier encodings of a given list of N sequences with alphabet sizes qs. """ if type(qs) == int: qs = [qs]*len(int_seqs[0]) M = np.prod(qs) N = len(int_seqs) encodings = get_encodings(qs) phi = np.zeros((N, M)) for i, seq in enumerate(tqdm(int_seqs)): phi[i] = fourier_for_seq(seq, encodings) / np.sqrt(M) return phi def convert_01_bin_seqs(bin_seqs): """Converts a numpy array of {0, 1} binary sequences to {-1, 1} sequences""" assert type(bin_seqs) == np.ndarray bin_seqs[bin_seqs == 0] = -1 return bin_seqs def walsh_hadamard_from_seqs(bin_seqs): """ Returns an N x 2^L array containing the Walsh-Hadamard encodings of a given list of N binary ({0,1}) sequences. This will return the same array as fourier_from_seqs(bin_seqs, [2]*L), but is much faster. """ bin_seqs_ = convert_01_bin_seqs(np.array(bin_seqs)) L = len(bin_seqs_[0]) all_U = get_all_interactions(L) M = 2**L N = len(bin_seqs) X = np.zeros((N, len(all_U))) for i, U in enumerate(all_U): if len(U) == 0: X[:, i] = 1 else: X[:, i] = np.prod(bin_seqs_[:, U], axis=1) X = X / np.sqrt(M) return X def calc_frac_var_explained(beta): """ Calculates the fraction of variance explained by the largest elements in a given vector beta. """ beta_sq = np.abs(beta)**2 beta_sq /= np.sum(beta_sq) bm_sorted = sorted(beta_sq, reverse=True) return np.cumsum(bm_sorted) def calc_frac_var_explained_from_beta_var(beta_var, samples=1000, up_to=None): """ Numerically calculates the fraction variance explained by the largest elements in samples of a normally distributed random vector with variances given by beta_var. Let S be the number of largest elements to consider. For each value of S between 0 and up_to (default up_to is len(beta_var)), this method returns the mean and std. dev. of the fraction variance explained by the largest S elements in samples of the random vector. """ beta_var_nz = beta_var[np.nonzero(beta_var)] M = len(beta_var_nz) if up_to is None: up_to = M fv = np.zeros(up_to) fv_std = np.zeros(up_to) samples = np.random.randn(samples, M) * np.sqrt(beta_var_nz).reshape(1, M) sorted_beta_sq = -np.sort(-samples**2, axis=1) mags = np.sum(sorted_beta_sq, axis=1) for i in range(up_to): if i > M: fv[i] = 1 fv_std[i] = 1 if i == 0: continue all_fv = np.sum(sorted_beta_sq[:, :i], axis=1) / mags fv[i] = np.mean(all_fv) fv_std[i] = np.std(all_fv) return fv, fv_std
6,463
30.842365
92
py
qsft
qsft-master/rna_exp/rna_src/data_utils.py
import pandas as pd import numpy as np import itertools from Bio import PDB from tqdm import tqdm from sklearn.linear_model import Lasso from rna_exp.rna_src import utils from rna_exp.rna_src import gnk_model from rna_exp.rna_src import structure_utils """ Utility functions for loading and processing empirical fitness function data """ # global variables MTAGBFP_POSITIONS = [20, 45, 63, 127, 143, 158, 168, 172, 174, 197, 206, 207, 227] HIS3P_BIG_QS = [2, 2, 3, 2, 2, 3, 3, 4, 2, 4, 4] HIS3P_POSITIONS = [145, 147, 148, 151, 152, 154, 164, 165, 168, 169, 170] def load_mtagbfp_data(): """ Loads mTagBFP2 blue flourescence fitness data from Poelwijk et. al. (2019). Returns the data as (X, y), where X is a matrix of Walsh-Hadamard encodings of sequences and y is an array of corresponding fitness values. Raw data is taken from Supplementary Data 3 of Poelwijk et. al. (2019) at https://www.nature.com/articles/s41467-019-12130-8#Sec20 """ df = pd.read_csv("../data/mtagbfp_raw_data.csv") y = np.array(df['brightness.1'][1:]).astype(float) bin_seqs_ = list(df['binary'][1:]) bin_seqs = [] for s_ in bin_seqs_: s_ = s_[1:-1] s = [] for si in s_: if si == '0': s.append(0) else: s.append(1) bin_seqs.append(s) bin_seqs = np.array(bin_seqs) L = len(bin_seqs[0]) X = utils.walsh_hadamard_from_seqs(bin_seqs) return X, y def load_his3p_small_data(): """ Loads His3p fitness data from Pokusaeva et. al. (2019) for binary sequences. Returns the data as (X, y), where X is a matrix of Walsh-Hadamard encodings of sequences and y is an array of corresponding fitness values. Raw data is taken from https://github.com/Lcarey/HIS3InterspeciesEpistasis/tree/master/Data. """ df = pd.read_csv("../data/his3p_raw_data.csv") extract = lambda x: "".join([x[i-1] for i in HIS3P_POSITIONS]) mut_seqs = df['seq'].apply(extract) # extract subsequences seqs_split = mut_seqs.str.split(pat ="\s*", expand = True) seqs_split = seqs_split.iloc[:, 1:-1] # find the two most frequently occuring alphabet elements at each position best = [] for i in range(1, 12): pos = seqs_split[i] cts = pos.value_counts() best.append(list(cts.iloc[:2].index)) bin_seqs_ = list(itertools.product((0, 1), repeat=11)) best_combos = [] for bs in bin_seqs_: seq = "".join([best[i][bs[i]] for i in range(len(bs))]) best_combos.append(seq) num = 0 bin_match =[] y_match = [] # collect fitness values for i, s in enumerate(best_combos): matches = mut_seqs.loc[mut_seqs == s] if len(matches) == 0: continue fitness = np.log(np.mean(np.exp(df['log_fitness'].loc[matches.index]))) # calculate mean if there are multiple fitness values y_match.append(fitness) bin_match.append(bin_seqs_[i]) X = utils.walsh_hadamard_from_seqs(np.array(bin_match)) return X, np.array(y_match) ####################### ### His3p(big) data ### ####################### """ Loading and processing the His3p(big) data takes much more time and memory than the smaller datasets above, so the pipeline is split into multiple functions, whose results are saved for further use. In order to load this data from scratch, one must run find_his3p_big_sequences(), followed by build_his3p_fourier(), and then load_his3p_big_fourier(). """ def find_his3p_big_sequences(save=False): """ Searches through the His3p raw data to find sequences with fitness data that correspond to combinations of extant amino acids at each position (i.e. the most frequently occuring amino acids at each position in the data). See page 4 of Pokusaeva et. al. (2019) for more information about these extant sequences. Returns a dictionary containing the found sequences, corresponding fitness values, and indices in the raw data. This takes some time to run, so there is an option to save the dictionary into the results fold by setting save=True. """ df = pd.read_csv("../data/his3p_raw_data.csv") qs = HIS3P_BIG_QS extract = lambda x: "".join([x[i-1] for i in HIS3P_POSITIONS]) mut_seqs = df['seq'].apply(extract) seqs_split = mut_seqs.str.split(pat ="\s*", expand = True) seqs_split = seqs_split.iloc[:, 1:-1] best = [] for i in range(1, 12): pos = seqs_split[i] cts = pos.value_counts() if i == 6: best.append(list(cts.iloc[[0,1,3]].index)) else: best.append(list(cts.iloc[:qs[i-1]].index)) sizes = [list(range(q)) for q in qs] int_seqs = list(itertools.product(*sizes)) best_combos = [] for bs in int_seqs: seq = [] for i in range(len(bs)): seq.append(best[i][bs[i]]) seq = "".join(seq) best_combos.append(seq) num = 0 int_match =[] y_match = [] idx_match = [] print("Finding sequences in data...") for i, s in enumerate(tqdm(best_combos)): matches = mut_seqs.loc[mut_seqs == s] if len(matches) == 0: continue fitness = np.log(np.mean(np.exp(df['log_fitness'].loc[matches.index]))) y_match.append(fitness) int_match.append(int_seqs[i]) idx_match.append(i) num += 1 out_dict = {"seq": int_match, "y": y_match, "idx": idx_match} if save: np.save("../results/his3p_big_data.npy", out_dict) return out_dict def build_his3p_big_fourier(): """ Converts the His3p(big) sequences into Fourier encodings and returns the matrix Will try to load the dict resulting from find_his3p_big_sequences, but will otherwise run that method. """ qs = HIS3P_BIG_QS try: save_dict = np.load("../results/his3p_big_data.npy",allow_pickle=True).item() except FileNotFoundError: save_dict = find_his3p_big_sequences(save=save) int_seqs = save_dict['seq'] M = np.prod(qs) N = len(int_seqs) phi = np.zeros((N, M)) encodings = utils.get_encodings(qs) print("Calculating Fourier encoding for each sequence...") for i, seq in enumerate(tqdm(int_seqs)): phi[i] = utils.fourier_for_seq(seq, encodings) / np.sqrt(M) return phi def load_his3p_big_data(save=False): """ Loads His3p(big) fitness data from Pokusaeva, et. al. (2019). Returns the data as (X, y), where X is a matrix of Fourier encodings of sequences and y is an array of corresponding fitness values. Will try to load dictionary from find_his3p_big_sequences, but will otherwise run that method (if save=True, then the dictionary will be saved for future use). """ try: save_dict = np.load("../results/his3p_big_data.npy",allow_pickle=True).item() except FileNotFoundError: save_dict = find_his3p_big_sequences(save=save) y = np.array(save_dict['y']) try: X = np.load("../results/his3p_big_fourier.npy") except FileNotFoundError: X = build_his3p_big_fourier(save=save) return X, y def _get_contact_map(which_data): """ Returns the contact map corresponding to either the TagBFP (which_data='mtagbfp') or His3p data (which_data='his3p'). """ if which_data == 'mtagbfp': name = '3m24' pos = MTAGBFP_POSITIONS elif which_data == 'his3p': name = 'his3_itasser' pos = HIS3P_POSITIONS structure = PDB.PDBParser().get_structure(name, '../data/%s.pdb' % name) chains = structure.get_chains() chain1 = next(chains) contact_map, resid2idx = structure_utils.calc_min_dist_contact_map(chain1) pos_in_cm = [resid2idx[p] for p in pos] cm_sub = contact_map[pos_in_cm][:, pos_in_cm] return cm_sub def _get_binarized_contact_map(which_data, threshold=4.5): """ Returns the binarized contact map corresponding to either the mTagBFP (which_data='mtagbfp') or His3p data (which_data='his3p'). """ cm_sub = _get_contact_map(which_data) bin_cm = structure_utils.binarize_contact_map(cm_sub, threshold=threshold) return bin_cm def get_mtagbfp_contact_map(): """ Returns the contact map of the TagBFP structure, for the positions in the empirical fitness function of Poelwijk, et. al. (2019). """ return _get_contact_map('mtagbfp') def get_his3p_contact_map(): """ Returns the contact map of the His3p I-TASSER predicted structure, for the positions in the empirical fitness function of Pokusaeva, et. al. (2019). """ return _get_contact_map('his3p') def get_mtagbfp_binarized_contact_map(threshold=4.5): """ Returns the binarized contact map of the TagBFP structure, for the positions in the empirical fitness function of Poelwijk, et. al. (2019). """ return _get_binarized_contact_map('mtagbfp', threshold=threshold) def get_his3p_binarized_contact_map(threshold=4.5): """ Returns the binarized contact map of the His3p I-TASSER predicted structure, for the positions in the empirical fitness function of Pokusaeva, et. al. (2019). """ return _get_binarized_contact_map('his3p', threshold=threshold) def _calculate_wh_coefficients_complete(which_data, save=False): """ Calculate the WH coefficients of the complete mTagBFP (which_data='mtagbfp'), His3p(small) (which_data='his3p_small') or His3p(big) (which_data='his3p_big') empirical fitness functions. """ alpha = 1e-12 if which_data == 'mtagbfp': X, y = load_mtagbfp_data() elif which_data == 'his3p_small': X, y = load_his3p_small_data() elif which_data == 'his3p_big': try: beta = np.load("../results/his3p_big_beta.npy") print("Loaded saved beta array.") return beta except FileNotFoundError: X, y = load_his3p_big_data() alpha = 1e-10 # slightly higher because data is less complete than others model = Lasso(alpha=alpha) print("Fitting Fourier coefficients (this may take some time)...") model.fit(X, y) beta = model.coef_ beta[0] = model.intercept_ if which_data == 'his3p_big' and save: np.save("../results/his3p_big_beta.npy", beta) return beta def calculate_mtagbfp_wh_coefficients(): """ Calculate the WH coefficients of the mTagBFP2 fitness functions. """ return _calculate_wh_coefficients_complete('mtagbfp') def calculate_his3p_small_wh_coefficients(): """ Calculate the Walsh-Hadamard coefficients of the His3p(small) fitness functions. """ return _calculate_wh_coefficients_complete('his3p_small') def calculate_his3p_big_fourier_coefficients(save=False): """ Calculate the Fourier coefficients of the His3p(big) fitness functions. Since this takes a long time, there is an option to save the beta coefficients in the results folder. """ return _calculate_wh_coefficients_complete('his3p_big', save=save) def calculate_mtagbfp_gnk_wh_coefficient_vars(return_neighborhoods=False): """ Returns the variances of WH coefficients in GNK fitness functions with Structural neighborhoods corresponding to the TagBFP structure. If return_neighborhoods is True then the Structural neighborhoods are also returned """ L = 13 q = 2 bin_cm = get_mtagbfp_binarized_contact_map() V = structure_utils.contact_map_to_neighborhoods(bin_cm) gnk_beta_var = gnk_model.calc_beta_var(L, q, V) if return_neighborhoods: return gnk_beta_var, V else: return gnk_beta_var def calculate_his3p_small_gnk_wh_coefficient_vars(return_neighborhoods=False): """ Returns the variances of WH coefficients in GNK fitness functions with Structural neighborhoods corresponding to the His3p structure. """ L = 11 q = 2 bin_cm = get_his3p_binarized_contact_map() V = structure_utils.contact_map_to_neighborhoods(bin_cm) gnk_beta_var = gnk_model.calc_beta_var(L, q, V) if return_neighborhoods: return gnk_beta_var, V else: return gnk_beta_var
12,224
34.537791
134
py
qsft
qsft-master/rna_exp/rna_src/rna_helper.py
import numpy as np import json from qsft.test_helper import TestHelper from qsft.utils import NpEncoder from rna_exp.rna_src.input_rna_signal_subsampled import RnaSubsampledSignal from rna_exp.rna_src.rna_utils import get_rna_base_seq class RNAHelper(TestHelper): mfe_base = 0 base_seq_list = None positions = None def __init__(self, signal_args, methods, subsampling_args, test_args, exp_dir, subsampling=False): config_path = exp_dir / "config.json" config_exists = config_path.is_file() if config_exists: with open(config_path) as f: config_dict = json.load(f) subsampling_args = config_dict["subsampling_args"] signal_args = config_dict["signal_args"] else: positions = list(np.sort(np.random.choice(len(get_rna_base_seq()), size=signal_args["n"], replace=False))) signal_args.update({ "base_seq": get_rna_base_seq(), "positions": positions }) config_dict = {"signal_args": signal_args, "subsampling_args": subsampling_args} with open(config_path, "w") as f: json.dump(config_dict, f, cls=NpEncoder) print("Positions: ", signal_args["positions"]) super().__init__(signal_args, methods, subsampling_args, test_args, exp_dir, subsampling) def generate_signal(self, signal_args): return RnaSubsampledSignal(**signal_args)
1,466
36.615385
118
py
qsft
qsft-master/rna_exp/rna_src/rna_utils.py
import RNA import itertools import utils import linecache import tracemalloc """ Utility functions for loading and processing the quasi-empirical RNA fitness function. """ def dna_to_rna(seq): """ Converts DNA sequences to RNA sequences. """ rs = [] for s in seq: if s == 'T': rs.append('U') else: rs.append(s) return "".join(rs) def insert(base_seq, positions, sub_seq): """ Inserts a subsequence into a base sequence """ new_seq = list(base_seq) for i, p in enumerate(positions): new_seq[p-1] = sub_seq[i] return "".join(new_seq) def get_rna_base_seq(): """ Returns the sequence of RFAM: AANN01066007.1 """ # base_seq = "CTGAGCCGTTACCTGCAGCTGATGAGCTCCAAAAAGAGCGAAACCTGCTAGGTCCTGCAGTACTGGCTTAAGAGGCT" base_seq = "CTGAGCCGTTACCTGCAGCTGATGAGCTCCAAAAAGA" return dna_to_rna(base_seq) def sample_structures_and_find_pairs(base_seq, positions, samples=10000): """ Samples secondary structures from the Boltzmann distribution and finds pairs of positions that are paired in any of the sampled strutures. """ md = RNA.md() md.uniq_ML = 1 fc = RNA.fold_compound(base_seq, md) (ss, mfe) = fc.mfe() fc.exp_params_rescale(mfe) fc.pf() important_pairs = set() for s in fc.pbacktrack(10000): pairs = find_pairs(s) for p in pairs: if p[0] in positions and p[1] in positions: if p[0] > p[1]: print(p, s) important_pairs.add(tuple(p)) return important_pairs def pairs_to_neighborhoods(positions, pairs): """ Converts a list of pairs of interacting positions into a set of neighborhoods. """ V = [] for i, p in enumerate(positions): Vp = [i+1] for pair in pairs: if pair[0] == p: Vp.append(positions.index(pair[1]) + 1) elif pair[1] == p: Vp.append(positions.index(pair[0]) + 1) V.append(sorted(Vp)) return V def find_pairs(ss): """ Finds interacting pairs in a RNA secondary structure """ pairs = [] op = [] N = len(ss) for i in range(N): if ss[i] == '(': op.append(i) elif ss[i] == ')': pair = (op.pop(), i) pairs.append(pair) return pairs def generate_householder_matrix(positions, n): nucs = ["A", "U", "C", "G"] nucs_idx = {nucs[i]: i for i in range(len(nucs))} seqs_as_list = list(itertools.product(nucs, repeat=len(positions))) int_seqs = [[nucs_idx[si] for si in s] for s in seqs_as_list] print("Constructing Fourier matrix...") X = utils.fourier_from_seqs(int_seqs, [4] * n) return X ''' def calculate_rna_gnk_wh_coefficient_vars(pairs_from_scratch=False, return_neighborhoods=False): """ Returns the variances of WH coefficients in GNK fitness functions with Structural neighborhoods corresponding to RNA secondary structure. If pairs_from_scratch is True, then structures are sampled to find paired positions, otherwise pre-calculated pairs are used. """ L = 8 q = 4 if pairs_from_scratch: important_pairs = sample_structures_and_find_pairs(data_utils.get_rna_base_seq(), positions, samples=10000) # uncomment to calculate from scratch else: important_pairs = {(21, 52), (20, 44), (20, 52), (20, 43)} # pre-calculated # add adjacent pairs important_pairs = important_pairs.union({(20, 21), (43, 44)}) V = pairs_to_neighborhoods(self.positions, important_pairs) gnk_beta_var = gnk_model.calc_beta_var(L, q, V) if return_neighborhoods: return gnk_beta_var, V else: return gnk_beta_var ''' def display_top(snapshot, key_type='lineno', limit=10): snapshot = snapshot.filter_traces(( tracemalloc.Filter(False, "<frozen importlib._bootstrap>"), tracemalloc.Filter(False, "<unknown>"), )) top_stats = snapshot.statistics(key_type) print("----------------------") print("Top %s lines" % limit) for index, stat in enumerate(top_stats[:limit], 1): frame = stat.traceback[0] print("#%s: %s:%s: %.1f KiB" % (index, frame.filename, frame.lineno, stat.size / 1024)) line = linecache.getline(frame.filename, frame.lineno).strip() if line: print(' %s' % line) other = top_stats[limit:] if other: size = sum(stat.size for stat in other) print("%s other: %.1f KiB" % (len(other), size / 1024)) total = sum(stat.size for stat in top_stats) print("Total allocated size: %.1f KiB" % (total / 1024)) print("----------------------")
4,803
27.093567
122
py
qsft
qsft-master/rna_exp/rna_src/__init__.py
0
0
0
py
qsft
qsft-master/rna_exp/rna_src/input_rna_signal_subsampled.py
from qsft.input_signal_subsampled import SubsampledSignal import numpy as np from multiprocessing import Pool from qsft.utils import dec_to_qary_vec, qary_vec_to_dec import RNA class RnaSubsampledSignal(SubsampledSignal): nucs = np.array(["A", "U", "C", "G"]) def __init__(self, **kwargs): self.base_seq = kwargs.get("base_seq") self.positions = kwargs.get("positions") self.sampling_function = kwargs.get("sampling_function") self.q = kwargs.get("q") self.n = kwargs.get("n") (_, mfe_base) = RNA.fold("".join(self.base_seq)) self.mfe_base = mfe_base self.pool = Pool(initializer=worker_init, initargs=(self.base_seq, self.positions, self.mfe_base)) super().__init__(**kwargs) self.pool.close() def subsample(self, query_indices): batch_size = 250 res = np.zeros(len(query_indices)) counter = 0 # pbar = tqdm(total=len(query_indices), miniters=batch_size, position=0) query_batches = np.array_split(query_indices, len(query_indices)//batch_size) for new_res in self.pool.imap(sampling_function, query_batches): res[counter: counter+len(new_res)] = new_res counter += len(new_res) # pbar.update(len(new_res)) return res def worker_init(base_seq, positions_input, mfe_base_input): global base_seq_list global positions global mfe_base base_seq_list = np.array(list(base_seq)) mfe_base = mfe_base_input positions = positions_input def sampling_function(query_batch): global base_seq_list global positions global mfe_base query_batch_q = np.array(dec_to_qary_vec(query_batch, 4, len(positions))).T y = [] for query_index in query_batch_q: full = next_q4(base_seq_list, positions, query_index) fc = RNA.fold_compound(full) (_, mfe) = fc.mfe() y.append(mfe - mfe_base) return y nucs = np.array(["A", "U", "C", "G"]) def next_q4(base_seq_list, positions, query_index): seq = nucs[query_index] base_seq_list[positions] = seq return "".join(base_seq_list) def next_q2(base_seq_list, positions, query_index): idx = dec_to_qary_vec([qary_vec_to_dec(query_index, 2)], 4, len(query_index)//2)[:, 0] seq = nucs[idx] base_seq_list[positions] = seq return "".join(base_seq_list)
2,382
28.060976
106
py
qsft
qsft-master/rna_exp/rna_src/structure_utils.py
import numpy as np from Bio import PDB """ Various utility functions for working with PDB structures. """ def binarize_contact_map(contact_map, threshold=8.0): """Returns binary version of contact map.""" return np.less(contact_map, threshold) def calc_min_dist(res1, res2): """Returns the minimum distance between any two atoms in res1 and res2.""" md = -1 for atom1 in res1: for atom2 in res2: dist = atom1-atom2 # biopython defines this operator if md == -1 or dist < md: md = dist return md def calc_min_dist_contact_map(chain1, chain2=None): """ Calculates the minimum distance between any two atoms in every pair of residues in chain1 and (optionally) chain2. Returns the array of distances along with a dictionary that maps residue IDs to indices in the array. """ resid_to_idx1 = {} if chain2 is None: chain2 = chain1 one_chain = True else: one_chain = False resid_to_idx2 = {} l1 = len(chain1) l2 = len(chain2) contact_map = np.ones((l1, l2)) * -1 for i, res1 in enumerate(chain1): resid_to_idx1[int(res1.id[1])] = i for j, res2 in enumerate(chain2): if one_chain: if j < i: continue contact_map[i, j] = calc_min_dist(res1, res2) if one_chain: contact_map[j, i] = contact_map[i, j] resid_to_idx1[int(res2.id[1])] = j else: resid_to_idx2[int(res2.id[1])] = j # Set missing distances larger than maximum distance in structure. missing_value = np.amax(contact_map) + 1.0 contact_map[np.less(contact_map, 0)] = missing_value if one_chain: return contact_map, resid_to_idx1 else: return contact_map, resid_to_idx1, resid_to_idx2 def contact_map_to_neighborhoods(contact_map): """Convert a binarized contact map to a set of neighborhoods""" L = contact_map.shape[0] V = [] for i in range(L): Vi = [i+1] for j in range(L): if i == j: continue elif contact_map[i, j]: Vi.append(j+1) V.append(sorted(Vi)) return V
2,281
28.636364
84
py
qsft
qsft-master/rna_exp/rna_src/gnk_model.py
import numpy as np import itertools from scipy.special import binom from math import factorial from rna_exp.rna_src import utils def get_neighborhood_powerset(V): """Returns the union of powersets of a set of neighborhoods""" Vs = [sorted(Vk) for Vk in V] powersets = [tuple(utils.powerset(Vs[i])) for i in range(len(Vs))] T = set().union(*powersets) return T def calculate_sparsity(L, q, V): """Calculates sparsity given any neighborhoods V=[V1,V2,...,VL]""" T = get_neighborhood_powerset(V) sparsity = 0 for U in T: sparsity += (q-1)**len(U) return sparsity def calc_bn_sparsity(L, q, K): """Calculates sparsity of Block Beighborhood scheme""" sparsity = (L/K)*(q**K - 1) +1 return sparsity def calc_an_sparsity(L, q, K): """Calculates sparsity of Adjacent Neighborhood scheme""" return 1 + L*(q-1)*q**(K-1) def _calc_set_prob(r, L, K): """Calculates p(r) for a set of size r, for use with 'calc_mean_rn_sparsity'""" if r == 0 or r == 1: return 1 else: ar = (factorial(K-1) / factorial(L-1)) * (factorial(L-r) / factorial(K-r)) br = ar * ((K-r) / (L-r)) term1 = (1-ar)**r term2 = (1-br)**(L-r) return 1-term1*term2 def calc_mean_rn_sparsity(L, q, K): """Calculates expected sparsity of Random Neighborhood scheme""" sparsity = 0 for r in range(K+1): pr = _calc_set_prob(r, L, K) sparsity += binom(L, r)*pr *(q-1)**r return sparsity def calc_max_rn_sparsity(L, q, K): """Calculates an upper bound on the sparsity of the Random Neighborhood scheme""" bd = 1+L*(q-1) for r in range(2, K+1): bd += L * binom(K, r) * (q-1)**r return bd def build_adj_neighborhoods(L, K, symmetric=True): """Build Adjacent Neighborhoods with periodic boundary conditions""" V = [] M = (K-1)/2 for i in range(L): if symmetric: start = np.floor(i-M) else: start = i Vi = [int(((start + j) % L)+1) for j in range(K)] V.append(Vi) return V def build_block_neighborhoods(L, K): """Build neighborhoods according to the Block Neighborhood scheme""" assert L % K == 0 V = [] block_size = int(L/K) for j in range(L): val = int(K*np.floor(j / K)) Vj = list(range(val+1, val+K+1)) V.append(Vj) return V def sample_random_neighborhoods(L, K): """Sample neighborhoods according to the Random Neighborhood scheme""" V = [] for i in range(L): indices = [j+1 for j in range(L) if j != i] Vi = list(np.random.choice(indices, size=K-1, replace=False)) Vi.append(i+1) V.append(sorted(Vi)) return V def calc_beta_var(L, qs, V): """ Calculates the variance of beta coefficients for a given sequence length, L, list of alphabet sizes, and neighborhoods V. The returned coefficients are ordered by degree of epistatic interaction. """ if type(qs) is int: qs = [qs]*L all_U = utils.get_all_interactions(L, index_1=True) # index by 1 to match neighborhoods z = np.prod(qs) beta_var_U = [] facs = [] for j, Vj in enumerate(V): fac = 1 for k in Vj: fac *= 1/qs[k-1] facs.append(fac) for i, U in enumerate(all_U): sz = np.prod([qs[k-1]-1 for k in U]) bv = 0 for j, Vj in enumerate(V): Uset = set(U) Vj_set = set(Vj) if Uset.issubset(Vj_set): bv += facs[j] bv *= z bv_expand = bv*np.ones(int(sz)) beta_var_U.append(bv_expand) beta_var = np.concatenate(beta_var_U) return beta_var def sample_gnk_fitness_function(L, qs, V='random', K=None): """ Sample a GNK fitness function given the sequence length, alphabet sizes and neighborhoods. If V='random', V='block', or V='adjacent', then the neighborhoods will be set to the corresponding standard neighborhood scheme. Otherwise, V must be a list of neighborhoods. """ if type(V) is str: assert K is not None if V == 'random': V = sample_random_neighborhoods(L, K) elif V == 'adjacent': V = build_adj_neighborhoods(L, K) elif V == 'block': V = build_block_neighborhoods(L, K) beta_var = calc_beta_var(L, qs, V) use_wh = False if type(qs) is int: if qs == 2: use_wh = True qs = [qs]*L alphs = [list(range(q)) for q in qs] seqs = list(itertools.product(*alphs)) if use_wh: phi = utils.walsh_hadamard_from_seqs(seqs) else: phi = utils.fourier_from_seqs(seqs, qs) beta = np.random.randn(len(beta_var))*np.sqrt(beta_var) f = np.dot(phi, beta) return f
4,787
28.018182
91
py
qsft
qsft-master/rna_exp/rna_src/input_rna_signal.py
from qsft.input_signal import Signal import numpy as np import itertools from rna_exp.rna_src.rna_utils import insert from multiprocessing import Pool from tqdm import tqdm from functools import partial tqdm = partial(tqdm, position=0, leave=True) class RnaSignal(Signal): def __init__(self, **kwargs): self.base_seq = kwargs.get("base_seq") self.positions = kwargs.get("positions") self.n = len(self.positions) self.sampling_function = kwargs.get("sampling_function") super().__init__(**kwargs) def sample(self): nucs = ["A", "U", "C", "G"] seqs_as_list = list(itertools.product(nucs, repeat=len(self.positions))) seqs = np.array(["".join(s) for s in seqs_as_list]) query = [] for i, s in enumerate(seqs): full = insert(self.base_seq, self.positions, s) query.append(full) raise NotImplementedError("things need to be changed") with Pool() as pool: y = list(tqdm(pool.imap(self.sampling_function, query), total=len(seqs))) self._signal_t = np.array(y)
1,107
31.588235
85
py
qsft
qsft-master/qsft/spright.py
''' SPRIGHT decoding main file. Logic flow: 1. Generate a signal from src/input_signal.py 2. Subsample from src/query.py 3. Peel using src/reconstruct.py ''' import numpy as np import galois import sys import tqdm import time sys.path.append("../src") from archive.qsft_rand import dec_to_bin, bin_to_dec, qary_vec_to_dec, binary_ints, qary_ints from archive.qsft_rand import compute_delayed_wht, compute_delayed_gwht, get_Ms, get_b, get_D from archive.qsft_rand import singleton_detection class SPRIGHT: ''' Class to store encoder/decoder configurations and carry out encoding/decoding. Attributes ---------- query_method : str The method to generate the sparsity coefficient and the subsampling matrices. Currently implemented methods: "simple" : choose some predetermined matrices based on problem size. delays_method : str The method to generate the matrix of delays. Currently implemented methods: "identity_like" : return a zero row and the identity matrix vertically stacked together. "random" : make a random set of delays. reconstruct_method : str The method to detect singletons. Currently implemented methods: "noiseless" : decode according to [2], section 4.2, with the assumption the signal is noiseless. "mle" : naive noisy decoding; decode by taking the maximum-likelihood singleton that could be at that bin. ''' def __init__(self, query_method, delays_method, reconstruct_method): self.query_method = query_method self.delays_method = delays_method self.reconstruct_method = reconstruct_method def transform(self, signal, verbose=False, report=False): ''' Full SPRIGHT encoding and decoding. Implements Algorithms 1 and 2 from [2]. (numbers) in the comments indicate equation numbers in [2]. Arguments --------- signal : Signal object. The signal to be transformed / compared to. verbose : boolean Whether to print intermediate steps. Returns ------- wht : ndarray The WHT constructed by subsampling and peeling. ''' GF = galois.GF(signal.q) # check the condition for p_failure > eps # upper bound on the number of peeling rounds, error out after that point num_peeling = 0 result = [] wht = np.zeros_like(signal.signal_t) b = get_b(signal, method=self.query_method) peeling_max = 2 ** b Ms = get_Ms(signal.n, b, signal.q, method=self.query_method) Us, Ss = [], [] singletons = {} multitons = [] if report: used = set() if self.delays_method != "nso": num_delays = signal.n + 1 else: num_delays = signal.n * int(np.log2(signal.n)) # idk if signal.q == 2: K = binary_ints(signal.n) else: K = GF(qary_ints(signal.n, signal.q)) # subsample, make the observation [U] and offset signature [S] matrices for M in Ms: D = get_D(signal.n, method=self.delays_method, num_delays=num_delays, q=signal.q) if verbose: print("------") print("a delay matrix") print(D) if signal.q == 2: U, used_i = compute_delayed_wht(signal, M, D) else: U, used_i = compute_delayed_gwht(signal, M, D, signal.q) Us.append(U) Ss.append((-1) ** (D @ K)) # offset signature matrix if report: used = used.union(used_i) cutoff = 2 * signal.noise_sd ** 2 * (2 ** (signal.n - b)) * num_delays # noise threshold if verbose: print('cutoff: {}'.format(cutoff)) # K is the binary representation of all integers from 0 to 2 ** n - 1. if signal.q == 2: select_froms = np.array([[bin_to_dec(row) for row in M.T.dot(K).T] for M in Ms]) else: select_froms = np.array([qary_vec_to_dec(-M.T @ K, signal.q) for M in Ms]) # `select_froms` is the collection of 'j' values and associated indices # so that we can quickly choose from the coefficient locations such that M.T @ k = j as in (20) # example: ball j goes to bin at "select_froms[i][j]"" in stage i # begin peeling # index convention for peeling: 'i' goes over all M/U/S values # i.e. it refers to the index of the subsampling group (zero-indexed - off by one from the paper). # 'j' goes over all columns of the WHT subsample matrix, going from 0 to 2 ** b - 1. # e.g. (i, j) = (0, 2) refers to subsampling group 0, and aliased bin 2 (10 in binary) # which in the example of section 3.2 is the multiton X[0110] + X[1010] + W1[10] # a multiton will just store the (i, j)s in a list # a singleton will map from the (i, j)s to the true (binary) values k. # e.g. the singleton (0, 0), which in the example of section 3.2 is X[0100] + W1[00] # would be stored as the dictionary entry (0, 0): array([0, 1, 0, 0]). there_were_multitons = True while there_were_multitons and num_peeling < peeling_max: if verbose: print('-----') print('the measurement matrix') for U in Us: print(U) # first step: find all the singletons and multitons. singletons = {} # dictionary from (i, j) values to the true index of the singleton, k. multitons = [] # list of (i, j) values indicating where multitons are. if signal.q > 2: # TODO selection = 0 else: for i, (U, S, select_from) in enumerate(zip(Us, Ss, select_froms)): for j, col in enumerate(U.T): # note that np.inner(x, x) is used as norm-squared: marginally faster than taking norm and squaring if np.inner(col, col) > cutoff: selection = np.where(select_from == j)[0] # pick all the k such that M.T @ k = j k, sgn = singleton_detection( col, method=self.reconstruct_method, selection=selection, S_slice=S[:, selection], n=signal.n ) # find the best fit singleton k_dec = bin_to_dec(k) rho = np.dot(S[:,k_dec], col)*sgn/len(col) residual = col - sgn * rho * S[:,k_dec] if verbose: print((i, j), np.inner(residual, residual)) if np.inner(residual, residual) > cutoff: multitons.append((i, j)) else: # declare as singleton singletons[(i, j)] = (k, rho, sgn) if verbose: print('amplitude: {}'.format(rho)) # all singletons and multitons are discovered if verbose: print('singletons:') for ston in singletons.items(): print("\t{0} {1}\n".format(ston, bin_to_dec(ston[1][0]))) print("Multitons : {0}\n".format(multitons)) # raise RuntimeError("stop") # WARNING: this is not a correct thing to do # in the last iteration of peeling, everything will be singletons and there # will be no multitons if len(multitons) == 0: # no more multitons, and can construct final WHT there_were_multitons = False # balls to peel balls_to_peel = set() ball_values = {} ball_sgn = {} for (i, j) in singletons: k, rho, sgn = singletons[(i, j)] ball = bin_to_dec(k) balls_to_peel.add(ball) ball_values[ball] = rho ball_sgn[ball] = sgn if verbose: print('these balls will be peeled') print(balls_to_peel) # peel for ball in balls_to_peel: num_peeling += 1 k = dec_to_bin(ball, signal.n) potential_peels = [(l, bin_to_dec(M.T.dot(k))) for l, M in enumerate(Ms)] result.append((k, ball_sgn[ball]*ball_values[ball])) for peel in potential_peels: signature_in_stage = Ss[peel[0]][:,ball] to_subtract = ball_sgn[ball] * ball_values[ball] * signature_in_stage Us[peel[0]][:,peel[1]] -= to_subtract if verbose: print('this is subtracted:') print(to_subtract) print("Peeled ball {0} off bin {1}".format(bin_to_dec(k), peel)) loc = set() for k, value in result: # iterating over (i, j)s idx = bin_to_dec(k) # converting 'k's of singletons to decimals loc.add(idx) if wht[idx] == 0: wht[idx] = value else: wht[idx] = (wht[idx] + value) / 2 # average out noise; e.g. in the example in 3.2, U1[11] and U2[11] are the same singleton, # so averaging them reduces the effect of noise. wht /= 2 ** (signal.n - b) if not report: return wht else: return wht, len(used), loc def method_test(self, signal, num_runs=10): ''' Tests a method on a signal and reports its average execution time and sample efficiency. ''' time_start = time.time() samples = 0 successes = 0 for i in tqdm.trange(num_runs): wht, num_samples, loc = self.transform(signal, report=True) if loc == set(signal.loc): successes += 1 samples += num_samples return (time.time() - time_start) / num_runs, successes / num_runs, samples / (num_runs * 2 ** signal.n) def method_report(self, signal, num_runs=10): ''' Reports the results of a method_test. ''' print( "Testing SPRIGHT with query method {0}, delays method {1}, reconstruct method {2}." .format(self.query_method, self.delays_method, self.reconstruct_method) ) t, s, sam = self.method_test(signal, num_runs) print("Average time in seconds: {}".format(t)) print("Success ratio: {}".format(s)) print("Average sample ratio: {}".format(sam)) if __name__ == "__main__": np.random.seed(10) from archive.qsft_rand import Signal test_signal = Signal(8, [4, 6, 10, 15, 24, 37, 48, 54], q=2, strengths=[2, 4, 1, 1, 1, 3, 8, 1], noise_sd=0) test_signal_complex = Signal(3, [4, 6, 10, 15, 24, 37, 48, 54], q=4, strengths=[2, 4, 1, 1, 1, 3, 8, 1], noise_sd=0) test_one_method = False if test_one_method: spright = SPRIGHT( query_method="simple", delays_method="complex", reconstruct_method="noiseless" ) residual = spright.transform(test_signal, verbose=False, report=False) - test_signal.signal_w print("Residual energy: {0}".format(np.inner(residual, residual))) else: configs = [ {"query_method" : "simple", "delays_method" : "identity_like", "reconstruct_method" : "mle"}, {"query_method" : "simple", "delays_method" : "random" , "reconstruct_method" : "mle"}, {"query_method" : "simple", "delays_method" : "nso" , "reconstruct_method" : "mle"}, {"query_method" : "simple", "delays_method" : "nso" , "reconstruct_method" : "nso"} ] for config in configs: try: SPRIGHT(**config).method_report(test_signal) except KeyboardInterrupt: continue
12,417
42.118056
123
py
qsft
qsft-master/qsft/input_signal_subsampled.py
from qsft.utils import qary_ints, qary_vec_to_dec, gwht, load_data, save_data from qsft.input_signal import Signal from qsft.query import get_Ms_and_Ds from pathlib import Path from math import floor from tqdm import tqdm import numpy as np import random import time class SubsampledSignal(Signal): """ A shell Class for input signal/functions that are too large and cannot be stored in their entirety. In addition to the signal itself, this must also contain information about the M and D matricies that are used for subsampling Notable attributes are included below. Attributes --------- query_args : dict These are the parameters that determine the structure of the Ms and Ds needed for subsampling. It contains the following sub-parameters: b : int The max dimension of subsampling (i.e., we will subsample functions with b inputs, or equivalently a signal of length q^b) all_bs : list, (optional) List of all the b values that should be subsampled. This is most useful when you want to repeat an experiment many times with different values of b to see which is most efficient For a description of the "delays_method_channel", "delays_method_source", "num_repeat" and "num_subsample", see the docstring of the QSFT class. subsampling_method If set to "simple" the M matricies are generated according to the construction in Appendix C, i.e., a block-wise identity structure. If set to "complex" the elements of the M matricies are uniformly populated from integers from 0 to q-1. It should be noted that these matricies are not checked to be full rank (w.r.t. the module where arithemtic is over the integer quotient ring), and so it is possible that the actual dimension of subsampling may be lower. For large enough n and b this isn't a problem, since w.h.p. the matricies are full rank. L : np.array An array that enumerates all q^b q-ary vectors of length b foldername : str If set, and the file {foldername}/Ms_and_Ds.pickle exists, the Ms and Ds are read directly from the file. Furthermore, if the transforms for all the bs are in {foldername}/transforms/U{i}_b{b}.pickle, the transforms can be directly loaded into memory. """ def _set_params(self, **kwargs): self.n = kwargs.get("n") self.q = kwargs.get("q") self.N = self.q ** self.n self.signal_w = kwargs.get("signal_w") self.query_args = kwargs.get("query_args") self.b = self.query_args.get("b") self.all_bs = self.query_args.get("all_bs", [self.b]) # all b values to sample/transform at self.num_subsample = self.query_args.get("num_subsample") if "num_repeat" not in self.query_args: self.query_args["num_repeat"] = 1 self.num_repeat = self.query_args.get("num_repeat") self.subsampling_method = self.query_args.get("subsampling_method") self.delays_method_source = self.query_args.get("delays_method_source") self.delays_method_channel = self.query_args.get("delays_method_channel") self.L = None # List of all length b qary vectors self.foldername = kwargs.get("folder") def _init_signal(self): if self.subsampling_method == "uniform": self._subsample_uniform() elif self.subsampling_method == "qsft": self._set_Ms_and_Ds_qsft() self._subsample_qsft() else: self._set_Ms_and_Ds_qsft() self._subsample_qsft() def _check_transforms_qsft(self): """ Returns ------- True if the transform is already computed and saved for all values of b, else False """ if self.foldername: Path(f"{self.foldername}/transforms/").mkdir(exist_ok=True) for b in self.all_bs: for i in range(len(self.Ms)): Us_path = Path(f"{self.foldername}/transforms/U{i}_b{b}.pickle") if not Us_path.is_file(): return False return True else: return False def _set_Ms_and_Ds_qsft(self): """ Sets the values of Ms and Ds, either by loading from folder if exists, otherwise it loaded from query_args """ if self.foldername: Path(f"{self.foldername}").mkdir(exist_ok=True) Ms_and_Ds_path = Path(f"{self.foldername}/Ms_and_Ds.pickle") if Ms_and_Ds_path.is_file(): self.Ms, self.Ds = load_data(Ms_and_Ds_path) else: self.Ms, self.Ds = get_Ms_and_Ds(self.n, self.q, **self.query_args) save_data((self.Ms, self.Ds), f"{self.foldername}/Ms_and_Ds.pickle") else: self.Ms, self.Ds = get_Ms_and_Ds(self.n, self.q, **self.query_args) def _subsample_qsft(self): """ Subsamples and computes the sparse fourier transform for each subsampling group if the samples are not already present in the folder """ self.Us = [[{} for j in range(len(self.Ds[i]))] for i in range(len(self.Ms))] self.transformTimes = [[{} for j in range(len(self.Ds[i]))] for i in range(len(self.Ms))] if self.foldername: Path(f"{self.foldername}/samples").mkdir(exist_ok=True) Path(f"{self.foldername}/transforms/").mkdir(exist_ok=True) pbar = tqdm(total=0, position=0) for i in range(len(self.Ms)): for j in range(len(self.Ds[i])): transform_file = Path(f"{self.foldername}/transforms/U{i}_{j}.pickle") if self.foldername and transform_file.is_file(): self.Us[i][j], self.transformTimes[i][j] = load_data(transform_file) pbar.total = len(self.Ms) * len(self.Ds[0]) * len(self.Us[i][j]) pbar.update(len(self.Us[i][j])) else: sample_file = Path(f"{self.foldername}/samples/M{i}_D{j}.pickle") if self.foldername and sample_file.is_file(): samples = load_data(sample_file) pbar.total = len(self.Ms) * len(self.Ds[0]) * len(samples) pbar.update(len(samples)) else: query_indices = self._get_qsft_query_indices(self.Ms[i], self.Ds[i][j]) block_length = len(query_indices[0]) samples = np.zeros((len(query_indices), block_length), dtype=np.complex) pbar.total = len(self.Ms) * len(self.Ds[0]) * len(query_indices) if block_length > 10000: for k in range(len(query_indices)): samples[k] = self.subsample(query_indices[k]) pbar.update() else: all_query_indices = np.concatenate(query_indices) all_samples = self.subsample(all_query_indices) for k in range(len(query_indices)): samples[k] = all_samples[k * block_length: (k+1) * block_length] pbar.update() if self.foldername: save_data(samples, sample_file) for b in self.all_bs: start_time = time.time() self.Us[i][j][b] = self._compute_subtransform(samples, b) self.transformTimes[i][j][b] = time.time() - start_time if self.foldername: save_data((self.Us[i][j], self.transformTimes[i][j]), transform_file) def _subsample_uniform(self): """ Uniformly subsamples the signal. Useful when you are solving via LASSO """ if self.foldername: Path(f"{self.foldername}").mkdir(exist_ok=True) sample_file = Path(f"{self.foldername}/signal_t.pickle") if self.foldername and sample_file.is_file(): signal_t = load_data(sample_file) else: query_indices = self._get_random_query_indices(self.query_args["n_samples"]) samples = self.subsample(query_indices) signal_t = dict(zip(query_indices, samples)) if self.foldername: save_data(signal_t, sample_file) self.signal_t = signal_t def get_all_qary_vectors(self): if self.L is None: self.L = np.array(qary_ints(self.b, self.q)) # List of all length b qary vectors return self.L def subsample(self, query_indices): raise NotImplementedError def _get_qsft_query_indices(self, M, D_sub): """ Gets the indicies to be queried for a given M and D Parameters ---------- M D_sub Returns ------- base_inds_dec : list The i-th element in the list is the affine space {Mx + d_i, forall x}, but in a decimal index, because it is more efficient, where d_i is the i-th row of D_sub. """ b = M.shape[1] L = self.get_all_qary_vectors() ML = (M @ L) % self.q base_inds = [(ML + np.outer(d, np.ones(self.q ** b, dtype=int))) % self.q for d in D_sub] base_inds = np.array(base_inds) base_inds_dec = [] for i in range(len(base_inds)): base_inds_dec.append(qary_vec_to_dec(base_inds[i], self.q)) return base_inds_dec def _get_random_query_indices(self, n_samples): """ Returns random indicies to be sampled. Parameters ---------- n_samples Returns ------- base_ids_dec Indicies to be queried in decimal representation """ # n_samples = np.minimum(n_samples, self.N) base_inds_dec = [floor(random.uniform(0, 1) * self.N) for _ in range(n_samples)] return base_inds_dec def get_MDU(self, ret_num_subsample, ret_num_repeat, b, trans_times=False): """ Allows the QSFT Class to get the effective Ms, Ds and Us (subsampled transforms). Parameters ---------- ret_num_subsample ret_num_repeat b trans_times Returns ------- Ms_ret Ds_ret Us_ret """ Ms_ret = [] Ds_ret = [] Us_ret = [] Ts_ret = [] if ret_num_subsample <= self.num_subsample and ret_num_repeat <= self.num_repeat and b <= self.b: subsample_idx = np.random.choice(self.num_subsample, ret_num_subsample, replace=False) delay_idx = np.random.choice(self.num_repeat, ret_num_repeat, replace=False) for i in subsample_idx: Ms_ret.append(self.Ms[i][:, :b]) Ds_ret.append([]) Us_ret.append([]) Ts_ret.append([]) for j in delay_idx: Ds_ret[-1].append(self.Ds[i][j]) Us_ret[-1].append(self.Us[i][j][b]) Ts_ret[-1].append(self.transformTimes[i][j][b]) if trans_times: return Ms_ret, Ds_ret, Us_ret, Ts_ret else: return Ms_ret, Ds_ret, Us_ret else: raise ValueError("There are not enough Ms or Ds.") def _compute_subtransform(self, samples, b): transform = [gwht(row[::(self.q ** (self.b - b))], self.q, b) for row in samples] return transform def get_source_parity(self): return self.Ds[0][0].shape[0]
11,744
42.5
122
py
qsft
qsft-master/qsft/qsft.py
''' Class for computing the q-ary fourier transform of a function/signal ''' import time import numpy as np from qsft.reconstruct import singleton_detection from qsft.input_signal_subsampled import SubsampledSignal from qsft.utils import bin_to_dec, qary_vec_to_dec, sort_qary_vecs, calc_hamming_weight, dec_to_qary_vec from synt_exp.synt_src.synthetic_signal import SyntheticSubsampledSignal class QSFT: ''' Class to encapsulate the configuration of our fourier algorithm. Attributes --------- reconstruct_method_source : str method of reconstruction for source coding: "identity" - default setting, should be used unless you know that all indicies have low hamming weight "coded" - Currently only supports prime q, if you know the max hamming weight of less than t this option should be used and will greatly reduce complexity. Note a source_decoder object must also be passed reconstruct_method_channel : str Method of reconstruction for channel coding: "mle" - exact MLE computation. Fine for small problems but not recommended it is exponential in n "nso" - symbol-wise recovery suitable when a repetition type code is used "identity" - no channel coding, only use when there is no noise num_subsamples : int The number of different subsampling groups M used num_repeat : int When a repetition code is used for channel coding, (NSO) this is the number of repetitions b : int Size of the sub-sampling signal. In general, we need q^b = O(K) where K is the number of nonzero terms in the transform. In practice, any q^b > K typically works well. noise_sd : scalar A noise parameter. Roughly, the standard deviation of the noise if it was an additive gaussian. source_decoder : function A function that takes in a source coded index, and returns decoded value of that index. Only needed when reconstruct_method_source = "coded" ''' def __init__(self, **kwargs): self.reconstruct_method_source = kwargs.get("reconstruct_method_source") self.reconstruct_method_channel = kwargs.get("reconstruct_method_channel") self.num_subsample = kwargs.get("num_subsample") self.num_repeat = kwargs.get("num_repeat") self.b = kwargs.get("b") self.source_decoder = kwargs.get("source_decoder", None) def transform(self, signal, verbosity=0, report=False, timing_verbose=False, **kwargs): """ Computes the q-ary fourier transform of a signal object Arguments --------- signal : Signal Signal object to be transformed. verbosity : int Larger numbers lead to increased number of printouts timing_verbose : Boolean If set to True, outputs detailed information about the amount of time each transform step takes. report : Boolean If set to True this function returns optional outputs "runtime": transform_time + peeling_time, "n_samples": total number of samples,"locations": locations of nonzero indicies,"avg_hamming_weight" average hamming weight of non-zero indicies and "max_hamming_weight": the maximum hamming weight of a nonzero index Returns ------- gwht : dict Fourier transform (WHT) of the input signal runtime : scalar transform time + peeling time. n_samples : int number of samples used in computing the transform. locations : list List of nonzero indicies in the transform. avg_hamming_weight : scalar Average hamming wieght of non-zero indicies. max_hamming_weight : int Max hamming weight among the non-zero indicies. """ q = signal.q n = signal.n b = self.b omega = np.exp(2j * np.pi / q) result = [] gwht = {} gwht_counts = {} peeling_max = q ** n peeled = set([]) if isinstance(signal, SubsampledSignal): Ms, Ds, Us, Ts = signal.get_MDU(self.num_subsample, self.num_repeat, self.b, trans_times=True) else: raise NotImplementedError("QSFT currently only supports signals that inherit from SubsampledSignal") for i in range(len(Ds)): Us[i] = np.vstack(Us[i]) Ds[i] = np.vstack(Ds[i]) transform_time = np.sum(Ts) if timing_verbose: print(f"Transform Time:{transform_time}", flush=True) Us = np.array(Us) # print(Us) gamma = 0.5 cutoff = 1e-9 + (1 + gamma) * (signal.noise_sd ** 2) / (q ** b) # noise threshold cutoff = kwargs.get("cutoff", cutoff) if verbosity >= 2: print("cutoff = ", cutoff, flush=True) # begin peeling # index convention for peeling: 'i' goes over all M/U/S values # i.e. it refers to the index of the subsampling group (zero-indexed - off by one from the paper). # 'j' goes over all columns of the WHT subsample matrix, going from 0 to 2 ** b - 1. # e.g. (i, j) = (0, 2) refers to subsampling group 0, and aliased bin 2 (10 in binary) # which in the example of section 3.2 is the multiton X[0110] + X[1010] + W1[10] # a multiton will just store the (i, j)s in a list # a singleton will map from the (i, j)s to the true (binary) values k. # e.g. the singleton (0, 0), which in the example of section 3.2 is X[0100] + W1[00] # would be stored as the dictionary entry (0, 0): array([0, 1, 0, 0]). max_iter = 15 iter_step = 0 cont_peeling = True num_peeling = 0 peeling_start = time.time() if timing_verbose: start_time = time.time() while cont_peeling and num_peeling < peeling_max and iter_step < max_iter: iter_step += 1 if verbosity >= 2: print('-----') print("iter ", iter_step, flush=True) # print('the measurement matrix') # for U in Us: # print(U) # first step: find all the singletons and multitons. singletons = {} # dictionary from (i, j) values to the true index of the singleton, k. multitons = [] # list of (i, j) values indicating where multitons are. for i, (U, M, D) in enumerate(zip(Us, Ms, Ds)): for j, col in enumerate(U.T): if np.linalg.norm(col) ** 2 > cutoff * len(col): k = singleton_detection( col, method_channel=self.reconstruct_method_channel, method_source=self.reconstruct_method_source, q=q, source_parity=signal.get_source_parity(), nso_subtype="nso1", source_decoder=self.source_decoder ) signature = omega ** (D @ k) rho = np.dot(np.conjugate(signature), col) / D.shape[0] residual = col - rho * signature j_qary = dec_to_qary_vec([j], q, b).T[0] bin_matching = np.all((M.T @ k) % q == j_qary) if verbosity >= 5: print((i, j), np.linalg.norm(residual) ** 2, cutoff * len(col)) if (not bin_matching) or np.linalg.norm(residual) ** 2 > cutoff * len(col): multitons.append((i, j)) if verbosity >= 6: print("We have a Multiton") else: # declare as singleton singletons[(i, j)] = (k, rho) if verbosity >= 3: print("We have a Singleton at " + str(k)) else: if verbosity >= 6: print("We have a Zeroton") # all singletons and multitons are discovered if verbosity >= 5: print('singletons:') for ston in singletons.items(): print("\t{0} {1}\n".format(ston, bin_to_dec(ston[1][0]))) print("Multitons : {0}\n".format(multitons)) # if there were no multi-tons or single-tons, decrease cutoff if len(multitons) == 0 or len(singletons) == 0: cont_peeling = False # balls to peel balls_to_peel = set() ball_values = {} for (i, j) in singletons: k, rho = singletons[(i, j)] ball = tuple(k) # Must be a hashable type #qary_vec_to_dec(k, q) balls_to_peel.add(ball) ball_values[ball] = rho result.append((k, ball_values[ball])) if verbosity >= 5: print('these balls will be peeled') print(balls_to_peel) # peel for ball in balls_to_peel: num_peeling += 1 k = np.array(ball)[..., np.newaxis] potential_peels = [(l, qary_vec_to_dec(M.T.dot(k) % q, q)[0]) for l, M in enumerate(Ms)] if verbosity >= 6: k_dec = qary_vec_to_dec(k, q) peeled.add(int(k_dec)) print("Processing Singleton {0}".format(k_dec)) print(k) for (l, j) in potential_peels: print("The singleton appears in M({0}), U({1})".format(l, j)) for peel in potential_peels: signature_in_stage = omega ** (Ds[peel[0]] @ k) to_subtract = ball_values[ball] * signature_in_stage.reshape(-1, 1) # print(np.linalg.norm(Us[peel[0]][:, peel[1]]), np.linalg.norm(to_subtract)) if verbosity >= 6: print("Peeled ball {0} off bin {1}".format(qary_vec_to_dec(k, q), peel)) Us[peel[0]][:, peel[1]] -= np.array(to_subtract)[:, 0] if verbosity >= 5: print("Iteration Complete: The peeled indicies are:") print(np.sort(list(peeled))) loc = set() for k, value in result: # iterating over (i, j)s loc.add(tuple(k)) if tuple(k) in gwht_counts: gwht[tuple(k)] = (gwht[tuple(k)] * gwht_counts[tuple(k)] + value) / (gwht_counts[tuple(k)] + 1) gwht_counts[tuple(k)] = gwht_counts[tuple(k)] + 1 else: gwht[tuple(k)] = value gwht_counts[tuple(k)] = 1 if timing_verbose: print(f"Peeling Time:{time.time() - start_time}", flush=True) peeling_time = time.time() - peeling_start if not report: return gwht else: n_samples = np.prod(np.shape(np.array(Us))) if len(loc) > 0: loc = list(loc) if kwargs.get("sort", False): loc = sort_qary_vecs(loc) avg_hamming_weight = np.mean(calc_hamming_weight(loc)) max_hamming_weight = np.max(calc_hamming_weight(loc)) else: loc, avg_hamming_weight, max_hamming_weight = [], 0, 0 result = { "gwht": gwht, "runtime": transform_time + peeling_time, "n_samples": n_samples, "locations": loc, "avg_hamming_weight": avg_hamming_weight, "max_hamming_weight": max_hamming_weight } return result
12,144
42.067376
122
py
qsft
qsft-master/qsft/ReedSolomon.py
import galois from galois._codes._reed_solomon import decode_jit import numpy as np import math class ReedSolomon(galois.ReedSolomon): """ Class that extends galois.ReedSolomon. Mainly it is needed to implement syndrome decoding. Attributes --------- prime_field : GF.field The galois feild of base q s : """ def __init__(self, n: int, t: int, q: int): self.prime_field = galois.GF(q) self.s = math.ceil(math.log(n) / math.log(q)) nt = (q ** self.s) - 1 if n <= (q ** self.s) - 1 else (q ** (self.s + 1)) - 1 self.s = self.s if n <= (q ** self.s) - 1 else self.s + 1 self.ns = n super().__init__(n=nt, k=nt-2*t) def syndrome_decode(self, syndrome): """ Syndrome Decoding for the RS decoder Parameters ---------- syndrome : self.prime_field.array() The syndrome as represented in the prime field. (i.e., the raw output of D*k) Returns ------- err : prime_field.array The decoded value of k n_errors : the number of decoding errors """ # Invoke the JIT compiled function q = self.field.characteristic codeword = self.field.Zeros(self.n) syndrome_qs = self.field.Vector([syndrome[self.s * i:self.s * (i + 1)] for i in range(2 * self.t)]) dec_codeword, n_errors = decode_jit(self.field)(codeword[:, np.newaxis].T, syndrome_qs[:, np.newaxis].T, self.c, self.t, int(self.field.primitive_element)) return -dec_codeword[:, -self.ns:], n_errors def get_delay_matrix(self): """ Generates teh parity check matrix of the constructed ReedSolomon code in the prime field Returns ------- D : GF.array Coded delay matrix """ Hvec = self.H[:, -self.ns:].vector() p = self.get_parity_length() D = self.prime_field.Zeros((p+1, self.ns)) for i in range(self.ns): for j in range(2 * self.t): D[(self.s * j + 1):(self.s * (j + 1) + 1), i] = Hvec[j, i, :] return D def get_parity_length(self): """ Returns the number of parity symbols in the code Returns ------- P """ return 2*self.t*self.s
2,356
30.426667
120
py
qsft
qsft-master/qsft/lasso.py
import numpy as np from group_lasso import GroupLasso from sklearn.linear_model import Ridge import time from group_lasso._fista import ConvergenceWarning from sklearn.utils._testing import ignore_warnings from qsft.utils import calc_hamming_weight, dec_to_qary_vec, qary_ints @ignore_warnings(category=ConvergenceWarning) def lasso_decode(signal, n_samples, noise_sd = 0, refine=True, verbose=False, report=True): """ Implements Complex LASSO via Fast Iterative Soft Thresholding (FISTA) with optional Ridge Regression refinement Parameters --------- signal : Signal Signal object to be transformed. n_samples : int number of samples used in computing the transform. verbosity : bool If True printouts are increased. noise_sd : scalar Noise standard deviation. refine : bool If True Ridge Regression refinement is used. Returns ------- gwht : dict Fourier transform (WHT) of the input signal runtime : scalar transform time + peeling time. locations : list List of nonzero indicies in the transform. avg_hamming_weight : scalar Average hamming wieght of non-zero indicies. max_hamming_weight : int Max hamming weight among the non-zero indicies. """ q = signal.q n = signal.n N = q ** n dtype = int if (q ** 2)*n > 255 else np.uint8 start_time = time.time() if verbose: print("Setting up LASSO problem") (sample_idx_dec, y) = list(signal.signal_t.keys()), list(signal.signal_t.values()) sample_idx_dec = sample_idx_dec[:n_samples] y = y[:n_samples] # WARNING: ADD NOISE ONLY FOR SYNTHETIC SIGNALS if signal.is_synt: y += np.random.normal(0, noise_sd / np.sqrt(2), size=(len(y), 2)).view(np.complex).reshape(len(y)) sample_idx = dec_to_qary_vec(sample_idx_dec, q, n, dtype=dtype) y = np.concatenate((np.real(y), np.imag(y))) freqs = np.array(sample_idx).T @ qary_ints(n, q, dtype=dtype) X = np.exp(2j*np.pi*freqs/q).astype(np.csingle) X_ext = np.concatenate((np.concatenate((np.real(X), -np.imag(X)), axis=1), np.concatenate((np.imag(X), np.real(X)), axis=1))) groups = [i % N for i in range(2*N)] lasso_start = time.time() if verbose: print(f"Setup Time:{time.time() - start_time}sec") print("Running Iterations...") start_time = time.time() lasso = GroupLasso(groups=groups, group_reg=0.1, l1_reg=0, tol=1e-8, n_iter=25, supress_warning=True, fit_intercept=False) lasso.fit(X_ext, y) if verbose: print(f"LASSO fit time:{time.time() - start_time}sec") w = lasso.coef_ non_zero = np.nonzero(w[:, 0])[0] if len(non_zero) > 0 and refine: ridge = Ridge(alpha=1e-2, tol=1e-8) ridge.fit(X_ext[:, non_zero], y) w[non_zero] = ridge.coef_[:, np.newaxis] gwht = w[0:N] + 1j*w[N:(2*N)] gwht = np.reshape(gwht, [q] * n) gwht_dict = {} non_zero_pos = np.array(np.nonzero(gwht)).T for p in non_zero_pos: gwht_dict[tuple(p)] = gwht[tuple(p)] runtime = time.time() - lasso_start if not report: return gwht_dict else: if len(non_zero_pos) > 0: loc = list(non_zero_pos) avg_hamming_weight = np.mean(calc_hamming_weight(loc)) max_hamming_weight = np.max(calc_hamming_weight(loc)) else: loc, avg_hamming_weight, max_hamming_weight = [], 0, 0 result = { "gwht": gwht_dict, "n_samples": n_samples, "locations": loc, "runtime": runtime, "avg_hamming_weight": avg_hamming_weight, "max_hamming_weight": max_hamming_weight } return result
3,870
28.549618
129
py
qsft
qsft-master/qsft/utils.py
''' Utility functions. ''' import numpy as np import scipy.fft as fft from group_lasso import GroupLasso from sklearn.linear_model import Ridge import itertools import math import random import time from scipy.spatial import ConvexHull import zlib import pickle import json import matplotlib.pyplot as plt def fwht(x): """Recursive implementation of the 1D Cooley-Tukey FFT""" # x = np.asarray(x, dtype=float) N = x.shape[0] if N == 1: return x else: X_even = fwht(x[0:(N//2)]) X_odd = fwht(x[(N//2):]) return np.concatenate([(X_even + X_odd), (X_even - X_odd)]) def gwht(x,q,n): """Computes the GWHT of an input signal with forward scaling""" x_tensor = np.reshape(x, [q] * n) x_tf = fft.fftn(x_tensor) / (q ** n) x_tf = np.reshape(x_tf, [q ** n]) return x_tf def gwht_tensored(x,q,n): """Computes the GWHT of an input signal with forward scaling""" x_tf = fft.fftn(x) / (q ** n) return x_tf def igwht(x,q,n): """Computes the IGWHT of an input signal with forward scaling""" x_tensor = np.reshape(x, [q] * n) x_tf = fft.ifftn(x_tensor) * (q ** n) x_tf = np.reshape(x_tf, [q ** n]) return x_tf def igwht_tensored(x,q,n): """Computes the IGWHT of an input signal with forward scaling""" x_tf = fft.ifftn(x) * (q ** n) return x_tf def bin_to_dec(x): n = len(x) c = 2**(np.arange(n)[::-1]) return c.dot(x).astype(np.int) def nth_roots_unity(n): return np.exp(-2j * np.pi / n * np.arange(n)) def near_nth_roots(ratios, q, eps): in_set = np.zeros(ratios.shape, dtype=bool) omega = nth_roots_unity(q) for i in range(q): in_set = in_set | (np.square(np.abs(ratios - omega[i])) < eps) is_singleton = in_set.all() return is_singleton def qary_vec_to_dec(x, q): n = x.shape[0] return np.array([q ** (n - (i + 1)) for i in range(n)], dtype=object) @ np.array(x, dtype=object) def dec_to_qary_vec(x, q, n): qary_vec = [] for i in range(n): qary_vec.append(np.array([a // (q ** (n - (i + 1))) for a in x], dtype=object)) x = x - (q ** (n-(i + 1))) * qary_vec[i] return np.array(qary_vec, dtype=int) def dec_to_bin(x, num_bits): assert x < 2**num_bits, "number of bits are not enough" u = bin(x)[2:].zfill(num_bits) u = list(u) u = [int(i) for i in u] return np.array(u) def binary_ints(m): ''' Returns a matrix where row 'i' is dec_to_bin(i, m), for i from 0 to 2 ** m - 1. From https://stackoverflow.com/questions/28111051/create-a-matrix-of-binary-representation-of-numbers-in-python. ''' a = np.arange(2 ** m, dtype=int)[np.newaxis,:] b = np.arange(m, dtype=int)[::-1,np.newaxis] return np.array(a & 2**b > 0, dtype=int) def angle_q(x,q): return (((np.angle(x) % (2*np.pi) // (np.pi/q)) + 1) // 2) % q # Can be made much faster def qary_ints(m, q, dtype=int): return np.array(list(itertools.product(np.arange(q), repeat=m)), dtype=dtype).T def comb(n, k): return math.factorial(n) // math.factorial(k) // math.factorial(n - k) def qary_ints_low_order(m, q, order): num_of_ks = np.sum([comb(m, o) * ((q-1) ** o) for o in range(order + 1)]) K = np.zeros((num_of_ks, m)) counter = 0 for o in range(order + 1): positions = itertools.combinations(np.arange(m), o) for pos in positions: K[counter:counter+((q-1) ** o), pos] = np.array(list(itertools.product(1 + np.arange(q-1), repeat=o))) counter += ((q-1) ** o) return K.T def base_ints(q, m): ''' Returns a matrix where row 'i' is the base-q representation of i, for i from 0 to q ** m - 1. Covers the functionality of binary_ints when n = 2, but binary_ints is faster for that case. ''' get_row = lambda i: np.array([int(j) for j in np.base_repr(i, base=q).zfill(m)]) return np.vstack((get_row(i) for i in range(q ** m))) def polymod(p1, p2, q, m): ''' Computes p1 modulo p2, and takes the coefficients modulo q. ''' p1 = np.trim_zeros(p1, trim='f') p2 = np.trim_zeros(p2, trim='f') while len(p1) >= len(p2) and len(p1) > 0: p1 -= p1[0] // p2[0] * np.pad(p2, (0, len(p1) - len(p2))) p1 = np.trim_zeros(p1, trim='f') return np.pad(np.mod(p1, q), (m + 1 - len(p1), 0)) def rref(A, b, q): ''' Row reduction, to easily solve finite field systems. ''' raise NotImplementedError() def sign(x): ''' Replacement for np.sign that matches the convention (footnote 2 on page 11). ''' return (1 - np.sign(x)) // 2 def flip(x): ''' Flip all bits in the binary array x. ''' return np.bitwise_xor(x, 1) def random_signal_strength_model(sparsity, a, b): magnitude = np.random.uniform(a, b, sparsity) phase = np.random.uniform(0, 2*np.pi, sparsity) return magnitude * np.exp(1j*phase) def best_convex_underestimator(points): hull = ConvexHull(points) vertices = points[hull.vertices] first_point_idx = np.argmin(vertices[:, 0]) last_point_idx = np.argmax(vertices[:, 0]) if last_point_idx == vertices.shape[0]: return vertices[first_point_idx:] if first_point_idx < last_point_idx: return vertices[first_point_idx:last_point_idx+1] else: return np.concatenate((vertices[first_point_idx:], vertices[:last_point_idx+1])) def sort_qary_vecs(qary_vecs): qary_vecs = np.array(qary_vecs) idx = np.lexsort(qary_vecs.T[::-1, :]) return qary_vecs[idx] def calc_hamming_weight(qary_vecs): qary_vecs = np.array(qary_vecs) return np.sum(qary_vecs != 0, axis = 1) def save_data(data, filename): with open(filename, 'wb') as f: f.write(zlib.compress(pickle.dumps(data, pickle.HIGHEST_PROTOCOL), 9)) def load_data(filename): start = time.time() with open(filename, 'rb') as f: data = pickle.loads(zlib.decompress(f.read())) return data class NpEncoder(json.JSONEncoder): def default(self, obj): if isinstance(obj, np.integer): return int(obj) if isinstance(obj, np.floating): return float(obj) if isinstance(obj, np.ndarray): return obj.tolist() return super(NpEncoder, self).default(obj)
6,283
29.357488
116
py
qsft
qsft-master/qsft/test_helper.py
import numpy as np from qsft.lasso import lasso_decode from qsft.qsft import QSFT from qsft.utils import gwht, dec_to_qary_vec, NpEncoder import json from qsft.query import get_reed_solomon_dec class TestHelper: def __init__(self, signal_args, methods, subsampling_args, test_args, exp_dir, subsampling=True): self.n = signal_args["n"] self.q = signal_args["q"] self.exp_dir = exp_dir self.subsampling = subsampling config_path = self.exp_dir / "config.json" config_exists = config_path.is_file() if not config_exists: config_dict = {"query_args": subsampling_args} with open(config_path, "w") as f: json.dump(config_dict, f, cls=NpEncoder) self.signal_args = signal_args self.subsampling_args = subsampling_args self.test_args = test_args if self.subsampling: if len(set(methods).intersection(["qsft"])) > 0: self.train_signal = self.load_train_data() # print("Quaternary Training data loaded.", flush=True) if len(set(methods).intersection(["qsft_binary"])) > 0: self.train_signal_binary = self.load_train_data_binary() # print("Binary Training data loaded.", flush=True) if len(set(methods).intersection(["lasso"])) > 0: self.train_signal_uniform = self.load_train_data_uniform() # print("Uniform Training data loaded.", flush=True) if len(set(methods).intersection(["qsft_coded"])) > 0: self.train_signal_coded = self.load_train_data_coded() # print("Uniform Training data loaded.", flush=True) self.test_signal = self.load_test_data() # print("Test data loaded.", flush=True) else: self.train_signal = self.load_full_data() self.test_signal = self.train_signal if any([m.startswith("binary") for m in methods]): raise NotImplementedError # TODO: implement the conversion # print("Full data loaded.", flush=True) def generate_signal(self, signal_args): raise NotImplementedError def load_train_data(self): signal_args = self.signal_args.copy() query_args = self.subsampling_args.copy() query_args.update({ "subsampling_method": "qsft", "query_method": "complex", "delays_method_source": "identity", "delays_method_channel": "nso" }) signal_args["folder"] = self.exp_dir / "train" signal_args["query_args"] = query_args return self.generate_signal(signal_args) def load_train_data_coded(self): signal_args = self.signal_args.copy() query_args = self.subsampling_args.copy() query_args.update({ "subsampling_method": "qsft", "query_method": "complex", "delays_method_source": "coded", "delays_method_channel": "nso", "t": signal_args["t"] }) signal_args["folder"] = self.exp_dir / "train_coded" signal_args["query_args"] = query_args return self.generate_signal(signal_args) def load_train_data_binary(self): return None # signal_args = self.signal_args.copy() # query_args = signal_args["query_args"] # signal_args["n_orig"] = signal_args["n"] # signal_args["q_orig"] = signal_args["q"] # factor = round(np.log(signal_args["q"]) / np.log(2)) # signal_args["n"] = factor * signal_args["n"] # signal_args["q"] = 2 # signal_args["query_args"]["subsampling_method"] = "qsft" # signal_args["query_args"]["b"] = factor * query_args["b"] # signal_args["query_args"]["all_bs"] = [factor * b for b in query_args["all_bs"]] # signal_args["query_args"]["num_repeat"] = max(1, query_args["num_repeat"] // factor) # signal_args["folder"] = self.exp_dir / "train_binary" # return self.generate_signal(signal_args) def load_train_data_uniform(self): signal_args = self.signal_args.copy() query_args = self.subsampling_args.copy() n_samples = query_args["num_subsample"] * (signal_args["q"] ** query_args["b"]) *\ query_args["num_repeat"] * (signal_args["n"] + 1) query_args = {"subsampling_method": "uniform", "n_samples": n_samples} signal_args["folder"] = self.exp_dir / "train_uniform" signal_args["query_args"] = query_args return self.generate_signal(signal_args) def load_test_data(self): signal_args = self.signal_args.copy() (self.exp_dir / "test").mkdir(exist_ok=True) signal_args["query_args"] = {"subsampling_method": "uniform", "n_samples": self.test_args.get("n_samples")} signal_args["folder"] = self.exp_dir / "test" signal_args["noise_sd"] = 0 return self.generate_signal(signal_args) def load_full_data(self): # TODO: implement return None def compute_model(self, method, model_kwargs, report=False, verbosity=0): if method == "gwht": return self._calculate_gwht(model_kwargs, report, verbosity) elif method == "qsft": return self._calculate_qsft(model_kwargs, report, verbosity) elif method == "qsft_binary": return self._calculate_qsft_binary(model_kwargs, report, verbosity) elif method == "qsft_coded": return self._calculate_qsft_coded(model_kwargs, report, verbosity) elif method == "lasso": return self._calculate_lasso(model_kwargs, report, verbosity) else: raise NotImplementedError() def test_model(self, method, **kwargs): if method == "qsft" or method == "qsft_coded" or method == "lasso": return self._test_qary(**kwargs) elif method == "qsft_binary": return self._test_binary(**kwargs) else: raise NotImplementedError() def _calculate_gwht(self, model_kwargs, report=False, verbosity=0): """ Calculates GWHT coefficients of the RNA fitness function. This will try to load them from the results folder, but will otherwise calculate from scratch. If save=True, then coefficients will be saved to the results folder. """ if verbosity >= 1: print("Finding all GWHT coefficients") beta = gwht(self.train_signal, q=4, n=self.n) print("Found GWHT coefficients") return beta def _calculate_qsft(self, model_kwargs, report=False, verbosity=0): """ Calculates GWHT coefficients of the RNA fitness function using QSFT. """ if verbosity >= 1: print("Estimating GWHT coefficients with QSFT") qsft = QSFT( reconstruct_method_source="identity", reconstruct_method_channel="nso", num_subsample=model_kwargs["num_subsample"], num_repeat=model_kwargs["num_repeat"], b=model_kwargs["b"] ) self.train_signal.noise_sd = model_kwargs["noise_sd"] out = qsft.transform(self.train_signal, verbosity=verbosity, timing_verbose=(verbosity >= 1), report=report) if verbosity >= 1: print("Found GWHT coefficients") return out def _calculate_qsft_coded(self, model_kwargs, report=False, verbosity=0): """ Calculates GWHT coefficients of the RNA fitness function using QSFT. """ if verbosity >= 1: print("Estimating GWHT coefficients with QSFT") decoder = get_reed_solomon_dec(self.signal_args["n"], self.signal_args["t"], self.signal_args["q"]) qsft = QSFT( reconstruct_method_source="coded", reconstruct_method_channel="nso", num_subsample=model_kwargs["num_subsample"], num_repeat=model_kwargs["num_repeat"], b=model_kwargs["b"], source_decoder=decoder ) self.train_signal_coded.noise_sd = model_kwargs["noise_sd"] out = qsft.transform(self.train_signal_coded, verbosity=verbosity, timing_verbose=(verbosity >= 1), report=report) if verbosity >= 1: print("Found GWHT coefficients") return out def _calculate_qsft_binary(self, model_kwargs, report=False, verbosity=0): """ Calculates GWHT coefficients of the RNA fitness function using QSFT. """ factor = round(np.log(self.q) / np.log(2)) if verbosity >= 1: print("Estimating GWHT coefficients with QSFT") qsft = QSFT( reconstruct_method_source="identity", reconstruct_method_channel="nso", num_subsample=model_kwargs["num_subsample"], num_repeat=max(1, model_kwargs["num_repeat"] // factor), b=factor * model_kwargs["b"], ) self.train_signal_binary.noise_sd = model_kwargs["noise_sd"] / factor out = qsft.transform(self.train_signal_binary, verbosity=verbosity, timing_verbose=(verbosity >= 1), report=report) if verbosity >= 1: print("Found GWHT coefficients") return out def _calculate_lasso(self, model_kwargs, report=False, verbosity=0): """ Calculates GWHT coefficients of the RNA fitness function using LASSO. This will try to load them from the results folder, but will otherwise calculate from scratch. If save=True, then coefficients will be saved to the results folder. """ if verbosity > 0: print("Finding Fourier coefficients with LASSO") self.train_signal_uniform.noise_sd = model_kwargs["noise_sd"] out = lasso_decode(self.train_signal_uniform, model_kwargs["n_samples"], noise_sd=model_kwargs["noise_sd"]) if verbosity > 0: print("Found Fourier coefficients") return out def _test_qary(self, beta): """ :param beta: :return: """ if len(beta.keys()) > 0: test_signal = self.test_signal.signal_t (sample_idx_dec, samples) = list(test_signal.keys()), list(test_signal.values()) batch_size = 10000 beta_keys = list(beta.keys()) beta_values = list(beta.values()) y_hat = [] for i in range(0, len(sample_idx_dec), batch_size): sample_idx_dec_batch = sample_idx_dec[i:i + batch_size] sample_idx_batch = dec_to_qary_vec(sample_idx_dec_batch, self.q, self.n) freqs = np.array(sample_idx_batch).T @ np.array(beta_keys).T H = np.exp(2j * np.pi * freqs / self.q) y_hat.append(H @ np.array(beta_values)) y_hat = np.concatenate(y_hat) return np.linalg.norm(y_hat - samples) ** 2 / np.linalg.norm(samples) ** 2 else: return 1 def _test_binary(self, beta): """ :param beta: :return: """ if len(beta.keys()) > 0: test_signal = self.test_signal.signal_t (sample_idx_dec, samples) = list(test_signal.keys()), list(test_signal.values()) batch_size = 10000 beta_keys = list(beta.keys()) beta_values = list(beta.values()) y_hat = [] for i in range(0, len(sample_idx_dec), batch_size): sample_idx_dec_batch = sample_idx_dec[i:i + batch_size] sample_idx_batch = dec_to_qary_vec(sample_idx_dec_batch, 2, 2 * self.n) freqs = np.array(sample_idx_batch).T @ np.array(beta_keys).T H = np.exp(2j * np.pi * freqs / 2) y_hat.append(H @ np.array(beta_values)) # TODO: Write with an if clause y_hat = np.abs(np.concatenate(y_hat)) return np.linalg.norm(y_hat - samples) ** 2 / np.linalg.norm(samples) ** 2 else: return 1
12,024
40.608997
123
py
qsft
qsft-master/qsft/parallel_tests.py
import itertools from multiprocessing import Pool from tqdm import tqdm from functools import partial import numpy as np import pandas as pd from qsft.test_helper import TestHelper tqdm = partial(tqdm, position=0, leave=True) def _test(i): """ Runs a single instance of a test Parameters ---------- i index of the test Returns ------- result : dict Result of a test, containing a variety of important quantities """ df_row = test_df.iloc[i] num_subsample, num_repeat, b, noise_sd = int(df_row["num_subsample"]), int(df_row["num_repeat"]), int(df_row["b"]), df_row["noise_sd"] # set model arguments model_kwargs = {} model_kwargs["num_subsample"] = num_subsample model_kwargs["num_repeat"] = num_repeat model_kwargs["b"] = b model_kwargs["noise_sd"] = noise_sd model_kwargs["n_samples"] = num_subsample * (helper_obj.q ** b) * num_repeat * (helper_obj.n + 1) model_result = helper_obj.compute_model(method=method, model_kwargs=model_kwargs, report=True, verbosity=0) test_kwargs = {} test_kwargs["beta"] = model_result.get("gwht") del model_result['gwht'] del model_result['locations'] nmse = helper_obj.test_model(method=method, **test_kwargs) result = {} result["n"] = helper_obj.n result["q"] = helper_obj.q result["runtime"] = model_result.get("runtime") result["found_sparsity"] = len(test_kwargs["beta"]) result["n_samples"] = model_result.get("n_samples") result["ratio_samples"] = model_result.get("n_samples") / (helper_obj.q ** helper_obj.n) result["max_hamming_weight"] = model_result.get("max_hamming_weight") result["nmse"] = nmse result["method"] = method return result def run_tests(test_method, helper: TestHelper, iters, num_subsample_list, num_repeat_list, b_list, noise_sd_list, parallel=True): """ Run an experiment based on the given parameters Parameters ---------- test_method helper iters num_subsample_list num_repeat_list b_list noise_sd_list parallel Returns ------- """ global test_df global helper_obj global method helper_obj = helper method = test_method test_params_list = list(itertools.product(num_subsample_list, num_repeat_list, b_list, noise_sd_list, range(iters))) test_df = pd.DataFrame(data=test_params_list, columns=["num_subsample", "num_repeat", "b", "noise_sd", "iter"]) exp_count = len(test_df) pred = [] if parallel: with Pool() as pool: # run the tests in parallel pbar = tqdm(pool.imap(_test, range(exp_count)), total=exp_count) best_result = np.inf for result in pbar: pred.append(result) best_result = min(best_result, result['nmse']) pbar.set_postfix({"min NMSE": best_result}) else: pbar = tqdm(range(exp_count)) best_result = np.inf for i in pbar: result = _test(i) pred.append(result) if result['nmse'] < best_result: best_result = result['nmse'] df_row = test_df.iloc[i] num_subsample, num_repeat, b = int(df_row["num_subsample"]), int(df_row["num_repeat"]), int(df_row["b"]) pbar.set_postfix({"min NMSE": best_result, "b": b, "C": num_subsample, "P1": num_repeat}) results_df = pd.DataFrame(data=pred) results_df = pd.concat([test_df, results_df], axis=1, join="inner") return results_df
3,555
29.921739
138
py
qsft
qsft-master/qsft/__init__.py
0
0
0
py
qsft
qsft-master/qsft/query.py
''' Methods for the query generator: specifically, to 1. generate sparsity coefficients b and subsampling matrices M 2. get the indices of a signal subsample 3. compute a subsampled and delayed Walsh-Hadamard transform. ''' import time import numpy as np from qsft.utils import fwht, gwht, bin_to_dec, binary_ints, qary_ints from qsft.ReedSolomon import ReedSolomon def get_Ms_simple(n, b, q, num_to_get=None): ''' Sets Ms[0] = [I 0 ...], Ms[1] = [0 I ...], Ms[2] = [0 0 I 0 ...] and so forth. See get_Ms for full signature. ''' Ms = [] for i in range(num_to_get - 1, -1, -1): M = np.zeros((n, b), dtype=np.int32) M[(b * i) : (b * (i + 1)), :] = np.eye(b) Ms.append(M) return Ms def get_Ms_complex(n, b, q, num_to_get=None): """ Generate M uniformly at random. """ Ms = [] # TODO Prevent duplicate M (Not a problem for large n, m ) for i in range(num_to_get): M = np.random.randint(q, size=(n, b)) Ms.append(M) return Ms def get_Ms(n, b, q, num_to_get=None, method="simple"): ''' Gets subsampling matrices for different sparsity levels. Arguments --------- n : int log_q of the signal length (number of inputs to function). b : int subsampling dimension. num_to_get : int The number of M matrices to return. method : str The method to use. All methods referenced must use this signature (minus "method".) Returns ------- Ms : list of numpy.ndarrays, shape (n, b) The list of subsampling matrices. ''' if num_to_get is None: num_to_get = max(n // b, 3) if method == "simple" and num_to_get > n // b: raise ValueError("When query_method is 'simple', the number of M matrices to return cannot be larger than n // b") return { "simple": get_Ms_simple, "complex": get_Ms_complex }.get(method)(n, b, q, num_to_get) def get_D_identity(n, **kwargs): int_delays = np.zeros(n, ) int_delays = np.vstack((int_delays, np.eye(n))) return int_delays.astype(int) def get_D_random(n, **kwargs): ''' Gets a random delays matrix of dimension (num_delays, n). See get_D for full signature. ''' q=kwargs.get("q") num_delays = kwargs.get("num_delays") return np.random.choice(q, (num_delays, n)) def get_D_source_coded(n, **kwargs): q = kwargs.get("q") t_max = kwargs.get("t") D = ReedSolomon(n, t_max, q).get_delay_matrix() return np.array(D, dtype=int) def get_D_nso(n, D_source, **kwargs): ''' Get a repetition code based (NSO-SPRIGHT) delays matrix. See get_D for full signature. ''' num_repeat = kwargs.get("num_repeat") q = kwargs.get("q") random_offsets = get_D_random(n, q=q, num_delays=num_repeat) D = [] for row in random_offsets: modulated_offsets = (row - D_source) % q D.append(modulated_offsets) return D def get_D_channel_coded(n, D, **kwargs): raise NotImplementedError("One day this might be implemented") def get_D_channel_identity(n, D, **kwargs): q = kwargs.get("q") return [D % q] def get_D(n, **kwargs): ''' Delay generator: gets a delays matrix. Arguments --------- n : int number of bits: log2 of the signal length. Returns ------- D : numpy.ndarray of binary ints, dimension (num_delays, n). The delays matrix; if num_delays is not specified in kwargs, see the relevant sub-function for a default. ''' delays_method_source = kwargs.get("delays_method_source", "random") D = { "random": get_D_random, "identity": get_D_identity, "coded": get_D_source_coded }.get(delays_method_source)(n, **kwargs) delays_method_channel = kwargs.get("delays_method_channel", "identity") D = { "nso": get_D_nso, "coded": get_D_channel_coded, "identity": get_D_channel_identity }.get(delays_method_channel)(n, D, **kwargs) return D def subsample_indices(M, d): ''' Query generator: creates indices for signal subsamples. Arguments --------- M : numpy.ndarray, shape (n, b) The subsampling matrix; takes on binary values. d : numpy.ndarray, shape (n,) The subsampling offset; takes on binary values. Returns ------- indices : numpy.ndarray, shape (B,) The (decimal) subsample indices. Mostly for debugging purposes. ''' L = binary_ints(M.shape[1]) inds_binary = np.mod(np.dot(M, L).T + d, 2).T return bin_to_dec(inds_binary) def compute_delayed_gwht(signal, M, D, q): """ Computes the Fourier transform of the delayed signal for some M and for each row in the delay matrix D """ b = M.shape[1] L = np.array(qary_ints(b, q)) # List of all length b qary vectors base_inds = [(M @ L + np.outer(d, np.ones(q ** b, dtype=int))) % q for d in D] used_inds = np.swapaxes(np.array(base_inds), 0, 1) used_inds = np.reshape(used_inds, (used_inds.shape[0], -1)) samples_to_transform = signal.get_time_domain(base_inds) transform = np.array([gwht(row, q, b) for row in samples_to_transform]) return transform, used_inds def get_Ms_and_Ds(n, q, **kwargs): """ Based on the parameters provided in kwargs, generates Ms and Ds. """ timing_verbose = kwargs.get("timing_verbose", False) if timing_verbose: start_time = time.time() query_method = kwargs.get("query_method") b = kwargs.get("b") num_subsample = kwargs.get("num_subsample") Ms = get_Ms(n, b, q, method=query_method, num_to_get=num_subsample) if timing_verbose: print(f"M Generation:{time.time() - start_time}") Ds = [] if timing_verbose: start_time = time.time() D = get_D(n, q=q, **kwargs) if timing_verbose: print(f"D Generation:{time.time() - start_time}") for M in Ms: Ds.append(D) return Ms, Ds def compute_delayed_wht(signal, M, D): ''' Creates random delays, subsamples according to M and the random delays, and returns the subsample WHT along with the delays. Arguments --------- signal : Signal object The signal to subsample, delay, and compute the WHT of. M : numpy.ndarray, shape (n, b) The subsampling matrix; takes on binary values. num_delays : int The number of delays to apply; or, the number of rows in the delays matrix. force_identity_like : boolean Whether to make D = [0; I] like in the noiseless case; for debugging. ''' inds = np.array([subsample_indices(M, d) for d in D]) used_inds = set(np.unique(inds)) samples_to_transform = signal.signal_t[np.array([subsample_indices(M, d) for d in D])] # subsample to allow small WHTs return np.array([fwht(row) for row in samples_to_transform]), used_inds # compute the small WHTs def get_reed_solomon_dec(n, t_max, q): """ Gets a suitable reed solomon decoder Returns ------- A Reed Solomon syndrome decoder for a t_max error correcting code. """ primeset = [2, 3, 5, 7, 11, 13, 15, 17, 19, 23, 29] if q in primeset: return ReedSolomon(n, t_max, q).syndrome_decode else: raise NotImplementedError("q is not a prime number under 30!")
7,267
28.786885
122
py
qsft
qsft-master/qsft/reconstruct.py
''' Methods for the reconstruction engine; specifically, to: 1. carry out singleton detection 2. get the cardinalities of all bins in a subsampling group (debugging only). ''' import numpy as np from qsft.utils import angle_q def singleton_detection_noiseless(U_slice, **kwargs): ''' Finds the true index of a singleton, or the best-approximation singleton of a multiton. Assumes P = n + 1 and D = [0; I]. Arguments --------- U_slice : numpy.ndarray, (P,). The WHT component of a subsampled bin, with element i corresponding to delay i. Returns ------- k : numpy.ndarray Index of the corresponding right node, in binary form. ''' q = kwargs.get('q') angles = np.angle(U_slice) angles = q*(angles[1:] - angles[0])/(2*np.pi) angles = angles.round().astype(int) % q return angles def singleton_detection_coded(k, **kwargs): ''' Finds the true index of a singleton, or the best-approximation singleton of a multiton. Assumes the Delays matrix is generated by a code, and the syndrome decoder is passed to it. Arguments --------- U_slice : numpy.ndarray, (P,). The WHT component of a subsampled bin, with element i corresponding to delay i. Returns ------- k : numpy.ndarray Index of the corresponding right node, in binary form. ''' decoder = kwargs.get('source_decoder') dec = decoder(list(k)) return np.array(dec[0][0, :], dtype=np.int32) def singleton_detection_mle(U_slice, **kwargs): ''' Finds the true index of a singleton, or the best-approximation singleton of a multiton, in the presence of noise. Uses MLE: looks at the residuals created by peeling off each possible singleton. Arguments --------- U_slice : numpy.ndarray, (P,). The WHT component of a subsampled bin, with element i corresponding to delay i. selection : numpy.ndarray. The decimal preimage of the bin index, i.e. the list of potential singletons whose signature under M could be the j of the bin. S_slice : numpy.ndarray The set of signatures under the delays matrix D associated with each of the elements of 'selection'. n : int The signal's number of bits. Returns ------- k : numpy.ndarray, (n,) The index of the singleton. ''' selection, S_slice, q, n = kwargs.get("selection"), kwargs.get("S_slice"), kwargs.get("q"), kwargs.get("source_parity") P = S_slice.shape[0] alphas = 1/P * np.dot(np.conjugate(S_slice).T, U_slice) residuals = np.linalg.norm(U_slice - (alphas * S_slice).T, ord=2, axis=1) k_sel = np.argmin(residuals) return selection[k_sel], S_slice[:, k_sel] def singleton_detection_nso(U_slice, **kwargs): """ Singleton Detection Via NSO Algorithm nso1 - Multiplying by conjugate "Soft Decoding" nso2 - Quantized angle "Hard Decoding" """ nso_type = kwargs.get("nso_subtype", "nso1") if nso_type == "nso1": return singleton_detection_nso1(U_slice, **kwargs) elif nso_type == "nso2": return singleton_detection_nso2(U_slice, **kwargs) def singleton_detection_nso1(U_slice, **kwargs): """ Soft Decoding NSO algorithm """ q, p1 = kwargs.get("q"), kwargs.get("source_parity") q_roots = 2 * np.pi / q * np.arange(q + 1) U_slice_zero = U_slice[0::p1] k_sel_qary = np.zeros((p1-1, ), dtype=int) for i in range(1, p1): U_slice_i = U_slice[i::p1] angle = np.angle(np.mean(U_slice_zero * np.conjugate(U_slice_i))) % (2 * np.pi) idx = (np.abs(q_roots - angle)).argmin() % q k_sel_qary[i-1] = idx return k_sel_qary def singleton_detection_nso2(U_slice, **kwargs): """ Hard Decoding NSO Algorithm """ q, p1 = kwargs.get("q"), kwargs.get("source_parity") U_slice_zero = U_slice[0::p1] angle_0 = angle_q(U_slice_zero, q) k_sel_qary = np.zeros((p1-1, ), dtype=int) for i in range(1, p1): U_slice_i = U_slice[i::p1] angle = angle_q(U_slice_i, q) idx = np.round(np.mean((angle_0 - angle) % q)) % q k_sel_qary[i-1] = idx return k_sel_qary def singleton_detection(U_slice, method_source="identity", method_channel="identity", **kwargs): """ Recovers the index value k of a singleton. Parameters ---------- U_slice : np.array The relevant subsampled fourier transform to be considered method_source method of reconstruction for source coding: "identity" - default setting, should be used unless you know that all indicies have low hamming weight "coded" - Currently only supports prime q, if you know the max hamming weight of less than t this option should be used and will greatly reduce complexity. Note a source_decoder object must also be passed method_channel Method of reconstruction for channel coding: "mle" - exact MLE computation. Fine for small problems but not recommended it is exponential in n "nso" - symbol-wise recovery suitable when a repetition type code is used "identity" - no channel coding, only use when there is no noise Returns ------- Value of the computed singleton index k """ # Split detection into two phases, channel and source decoding k = { "mle": singleton_detection_mle, "nso": singleton_detection_nso, "identity": singleton_detection_noiseless, }.get(method_channel)(U_slice, **kwargs) if method_source != "identity": k = { "coded": singleton_detection_coded }.get(method_source)(k, **kwargs) return k
5,932
34.315476
131
py
qsft
qsft-master/qsft/input_signal.py
""" A shell Class for common interface to an input signal. This class should be extended when implemented """ import numpy as np from qsft.utils import gwht_tensored, igwht_tensored, save_data, load_data from pathlib import Path class Signal: """ Class to encapsulate a time domain signal and its q-ary Fourier transform. Attributes --------- n : int number of bits: number of function inputs. q : int Locations of true peaks in the W-H spectrum. Elements must be integers in [0, q ** n - 1]. noise_sd : scalar The standard deviation of the added noise. signal_t Time domain representation of the signal. signal_w Frequency domain representation of the signal, calc_w If True and signal_w is not included, it is computed based on signal_t. foldername If signal_t is not provided, the signal will be read from {foldername}/signal_t.pickle. If signal_t is provided, a copy of the signal is written to {foldername}/signal_t.pickle """ def __init__(self, **kwargs): self._set_params(**kwargs) self._init_signal() def _set_params(self, **kwargs): self.n = kwargs.get("n") self.q = kwargs.get("q") self.noise_sd = kwargs.get("noise_sd", 0) self.N = self.q ** self.n self.signal_t = kwargs.get("signal_t") self.signal_w = kwargs.get("signal_w") self.calc_w = kwargs.get("calc_w", False) self.foldername = kwargs.get("folder") self.is_synt = False def _init_signal(self): if self.signal_t is None: signal_path = Path(f"{self.foldername}/signal_t.pickle") if signal_path.is_file(): self.signal_t = load_data(Path(f"{self.foldername}/signal_t.pickle")) else: self.sample() Path(f"{self.foldername}").mkdir(exist_ok=True) save_data(self.signal_t, Path(f"{self.foldername}/signal_t.pickle")) if self.calc_w and self.signal_w is None: self.signal_w = gwht_tensored(self.signal_t, self.q, self.n) if np.linalg.norm(self.signal_t - igwht_tensored(self.signal_w, self.q, self.n)) / self.N < 1e-5: print("verified transform") def sample(self): raise NotImplementedError def shape(self): ''' shape: returns the shape of the time domain signal. Returns ------- shape of time domain signal ''' return tuple([self.q for i in range(self.n)]) def get_time_domain(self, inds): ''' Arguments --------- inds: tuple of 1d n-element arrays that represent the indicies to be queried Returns ------- indices : linear output of the queried indicies ''' inds = np.array(inds) if len(inds.shape) == 3: return [self.signal_t[tuple(inds)] for inds in inds] elif len(inds.shape) == 2: return self.signal_t[tuple(inds)]
3,034
30.28866
118
py
sort
sort-master/sort.py
""" SORT: A Simple, Online and Realtime Tracker Copyright (C) 2016-2020 Alex Bewley alex@bewley.ai This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version. This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details. You should have received a copy of the GNU General Public License along with this program. If not, see <http://www.gnu.org/licenses/>. """ from __future__ import print_function import os import numpy as np import matplotlib matplotlib.use('TkAgg') import matplotlib.pyplot as plt import matplotlib.patches as patches from skimage import io import glob import time import argparse from filterpy.kalman import KalmanFilter np.random.seed(0) def linear_assignment(cost_matrix): try: import lap _, x, y = lap.lapjv(cost_matrix, extend_cost=True) return np.array([[y[i],i] for i in x if i >= 0]) # except ImportError: from scipy.optimize import linear_sum_assignment x, y = linear_sum_assignment(cost_matrix) return np.array(list(zip(x, y))) def iou_batch(bb_test, bb_gt): """ From SORT: Computes IOU between two bboxes in the form [x1,y1,x2,y2] """ bb_gt = np.expand_dims(bb_gt, 0) bb_test = np.expand_dims(bb_test, 1) xx1 = np.maximum(bb_test[..., 0], bb_gt[..., 0]) yy1 = np.maximum(bb_test[..., 1], bb_gt[..., 1]) xx2 = np.minimum(bb_test[..., 2], bb_gt[..., 2]) yy2 = np.minimum(bb_test[..., 3], bb_gt[..., 3]) w = np.maximum(0., xx2 - xx1) h = np.maximum(0., yy2 - yy1) wh = w * h o = wh / ((bb_test[..., 2] - bb_test[..., 0]) * (bb_test[..., 3] - bb_test[..., 1]) + (bb_gt[..., 2] - bb_gt[..., 0]) * (bb_gt[..., 3] - bb_gt[..., 1]) - wh) return(o) def convert_bbox_to_z(bbox): """ Takes a bounding box in the form [x1,y1,x2,y2] and returns z in the form [x,y,s,r] where x,y is the centre of the box and s is the scale/area and r is the aspect ratio """ w = bbox[2] - bbox[0] h = bbox[3] - bbox[1] x = bbox[0] + w/2. y = bbox[1] + h/2. s = w * h #scale is just area r = w / float(h) return np.array([x, y, s, r]).reshape((4, 1)) def convert_x_to_bbox(x,score=None): """ Takes a bounding box in the centre form [x,y,s,r] and returns it in the form [x1,y1,x2,y2] where x1,y1 is the top left and x2,y2 is the bottom right """ w = np.sqrt(x[2] * x[3]) h = x[2] / w if(score==None): return np.array([x[0]-w/2.,x[1]-h/2.,x[0]+w/2.,x[1]+h/2.]).reshape((1,4)) else: return np.array([x[0]-w/2.,x[1]-h/2.,x[0]+w/2.,x[1]+h/2.,score]).reshape((1,5)) class KalmanBoxTracker(object): """ This class represents the internal state of individual tracked objects observed as bbox. """ count = 0 def __init__(self,bbox): """ Initialises a tracker using initial bounding box. """ #define constant velocity model self.kf = KalmanFilter(dim_x=7, dim_z=4) self.kf.F = np.array([[1,0,0,0,1,0,0],[0,1,0,0,0,1,0],[0,0,1,0,0,0,1],[0,0,0,1,0,0,0], [0,0,0,0,1,0,0],[0,0,0,0,0,1,0],[0,0,0,0,0,0,1]]) self.kf.H = np.array([[1,0,0,0,0,0,0],[0,1,0,0,0,0,0],[0,0,1,0,0,0,0],[0,0,0,1,0,0,0]]) self.kf.R[2:,2:] *= 10. self.kf.P[4:,4:] *= 1000. #give high uncertainty to the unobservable initial velocities self.kf.P *= 10. self.kf.Q[-1,-1] *= 0.01 self.kf.Q[4:,4:] *= 0.01 self.kf.x[:4] = convert_bbox_to_z(bbox) self.time_since_update = 0 self.id = KalmanBoxTracker.count KalmanBoxTracker.count += 1 self.history = [] self.hits = 0 self.hit_streak = 0 self.age = 0 def update(self,bbox): """ Updates the state vector with observed bbox. """ self.time_since_update = 0 self.history = [] self.hits += 1 self.hit_streak += 1 self.kf.update(convert_bbox_to_z(bbox)) def predict(self): """ Advances the state vector and returns the predicted bounding box estimate. """ if((self.kf.x[6]+self.kf.x[2])<=0): self.kf.x[6] *= 0.0 self.kf.predict() self.age += 1 if(self.time_since_update>0): self.hit_streak = 0 self.time_since_update += 1 self.history.append(convert_x_to_bbox(self.kf.x)) return self.history[-1] def get_state(self): """ Returns the current bounding box estimate. """ return convert_x_to_bbox(self.kf.x) def associate_detections_to_trackers(detections,trackers,iou_threshold = 0.3): """ Assigns detections to tracked object (both represented as bounding boxes) Returns 3 lists of matches, unmatched_detections and unmatched_trackers """ if(len(trackers)==0): return np.empty((0,2),dtype=int), np.arange(len(detections)), np.empty((0,5),dtype=int) iou_matrix = iou_batch(detections, trackers) if min(iou_matrix.shape) > 0: a = (iou_matrix > iou_threshold).astype(np.int32) if a.sum(1).max() == 1 and a.sum(0).max() == 1: matched_indices = np.stack(np.where(a), axis=1) else: matched_indices = linear_assignment(-iou_matrix) else: matched_indices = np.empty(shape=(0,2)) unmatched_detections = [] for d, det in enumerate(detections): if(d not in matched_indices[:,0]): unmatched_detections.append(d) unmatched_trackers = [] for t, trk in enumerate(trackers): if(t not in matched_indices[:,1]): unmatched_trackers.append(t) #filter out matched with low IOU matches = [] for m in matched_indices: if(iou_matrix[m[0], m[1]]<iou_threshold): unmatched_detections.append(m[0]) unmatched_trackers.append(m[1]) else: matches.append(m.reshape(1,2)) if(len(matches)==0): matches = np.empty((0,2),dtype=int) else: matches = np.concatenate(matches,axis=0) return matches, np.array(unmatched_detections), np.array(unmatched_trackers) class Sort(object): def __init__(self, max_age=1, min_hits=3, iou_threshold=0.3): """ Sets key parameters for SORT """ self.max_age = max_age self.min_hits = min_hits self.iou_threshold = iou_threshold self.trackers = [] self.frame_count = 0 def update(self, dets=np.empty((0, 5))): """ Params: dets - a numpy array of detections in the format [[x1,y1,x2,y2,score],[x1,y1,x2,y2,score],...] Requires: this method must be called once for each frame even with empty detections (use np.empty((0, 5)) for frames without detections). Returns the a similar array, where the last column is the object ID. NOTE: The number of objects returned may differ from the number of detections provided. """ self.frame_count += 1 # get predicted locations from existing trackers. trks = np.zeros((len(self.trackers), 5)) to_del = [] ret = [] for t, trk in enumerate(trks): pos = self.trackers[t].predict()[0] trk[:] = [pos[0], pos[1], pos[2], pos[3], 0] if np.any(np.isnan(pos)): to_del.append(t) trks = np.ma.compress_rows(np.ma.masked_invalid(trks)) for t in reversed(to_del): self.trackers.pop(t) matched, unmatched_dets, unmatched_trks = associate_detections_to_trackers(dets,trks, self.iou_threshold) # update matched trackers with assigned detections for m in matched: self.trackers[m[1]].update(dets[m[0], :]) # create and initialise new trackers for unmatched detections for i in unmatched_dets: trk = KalmanBoxTracker(dets[i,:]) self.trackers.append(trk) i = len(self.trackers) for trk in reversed(self.trackers): d = trk.get_state()[0] if (trk.time_since_update < 1) and (trk.hit_streak >= self.min_hits or self.frame_count <= self.min_hits): ret.append(np.concatenate((d,[trk.id+1])).reshape(1,-1)) # +1 as MOT benchmark requires positive i -= 1 # remove dead tracklet if(trk.time_since_update > self.max_age): self.trackers.pop(i) if(len(ret)>0): return np.concatenate(ret) return np.empty((0,5)) def parse_args(): """Parse input arguments.""" parser = argparse.ArgumentParser(description='SORT demo') parser.add_argument('--display', dest='display', help='Display online tracker output (slow) [False]',action='store_true') parser.add_argument("--seq_path", help="Path to detections.", type=str, default='data') parser.add_argument("--phase", help="Subdirectory in seq_path.", type=str, default='train') parser.add_argument("--max_age", help="Maximum number of frames to keep alive a track without associated detections.", type=int, default=1) parser.add_argument("--min_hits", help="Minimum number of associated detections before track is initialised.", type=int, default=3) parser.add_argument("--iou_threshold", help="Minimum IOU for match.", type=float, default=0.3) args = parser.parse_args() return args if __name__ == '__main__': # all train args = parse_args() display = args.display phase = args.phase total_time = 0.0 total_frames = 0 colours = np.random.rand(32, 3) #used only for display if(display): if not os.path.exists('mot_benchmark'): print('\n\tERROR: mot_benchmark link not found!\n\n Create a symbolic link to the MOT benchmark\n (https://motchallenge.net/data/2D_MOT_2015/#download). E.g.:\n\n $ ln -s /path/to/MOT2015_challenge/2DMOT2015 mot_benchmark\n\n') exit() plt.ion() fig = plt.figure() ax1 = fig.add_subplot(111, aspect='equal') if not os.path.exists('output'): os.makedirs('output') pattern = os.path.join(args.seq_path, phase, '*', 'det', 'det.txt') for seq_dets_fn in glob.glob(pattern): mot_tracker = Sort(max_age=args.max_age, min_hits=args.min_hits, iou_threshold=args.iou_threshold) #create instance of the SORT tracker seq_dets = np.loadtxt(seq_dets_fn, delimiter=',') seq = seq_dets_fn[pattern.find('*'):].split(os.path.sep)[0] with open(os.path.join('output', '%s.txt'%(seq)),'w') as out_file: print("Processing %s."%(seq)) for frame in range(int(seq_dets[:,0].max())): frame += 1 #detection and frame numbers begin at 1 dets = seq_dets[seq_dets[:, 0]==frame, 2:7] dets[:, 2:4] += dets[:, 0:2] #convert to [x1,y1,w,h] to [x1,y1,x2,y2] total_frames += 1 if(display): fn = os.path.join('mot_benchmark', phase, seq, 'img1', '%06d.jpg'%(frame)) im =io.imread(fn) ax1.imshow(im) plt.title(seq + ' Tracked Targets') start_time = time.time() trackers = mot_tracker.update(dets) cycle_time = time.time() - start_time total_time += cycle_time for d in trackers: print('%d,%d,%.2f,%.2f,%.2f,%.2f,1,-1,-1,-1'%(frame,d[4],d[0],d[1],d[2]-d[0],d[3]-d[1]),file=out_file) if(display): d = d.astype(np.int32) ax1.add_patch(patches.Rectangle((d[0],d[1]),d[2]-d[0],d[3]-d[1],fill=False,lw=3,ec=colours[d[4]%32,:])) if(display): fig.canvas.flush_events() plt.draw() ax1.cla() print("Total Tracking took: %.3f seconds for %d frames or %.1f FPS" % (total_time, total_frames, total_frames / total_time)) if(display): print("Note: to get real runtime results run without the option: --display")
11,739
34.468278
242
py
cogcn
cogcn-main/cogcn/utils.py
import pickle as pkl import os import networkx as nx import numpy as np import scipy.sparse as sp import torch import pandas as pd from sklearn.metrics import roc_auc_score, average_precision_score from matplotlib import pyplot as plt def load_data_cma(dataset): adj_file = os.path.join(dataset, "struct.csv") feat_file = os.path.join(dataset, "content.csv") # Load reate adjacency matrix adj = pd.read_csv(adj_file, header=None) adj = adj.values adj = nx.from_numpy_matrix(adj) adj = nx.adjacency_matrix(adj) print("Adjacency matrix shape:", adj.shape) # Load features feat = pd.read_csv(feat_file, header=None) feat = feat.values features = torch.FloatTensor(feat) print("Features shape:", features.shape) return adj, features def parse_index_file(filename): index = [] for line in open(filename): index.append(int(line.strip())) return index def sparse_to_tuple(sparse_mx): if not sp.isspmatrix_coo(sparse_mx): sparse_mx = sparse_mx.tocoo() coords = np.vstack((sparse_mx.row, sparse_mx.col)).transpose() values = sparse_mx.data shape = sparse_mx.shape return coords, values, shape def preprocess_graph(adj): adj = sp.coo_matrix(adj) adj_ = adj + sp.eye(adj.shape[0]) rowsum = np.array(adj_.sum(1)) degree_mat_inv_sqrt = sp.diags(np.power(rowsum, -0.5).flatten()) adj_normalized = adj_.dot(degree_mat_inv_sqrt).transpose().dot(degree_mat_inv_sqrt).tocoo() # return sparse_to_tuple(adj_normalized) return sparse_mx_to_torch_sparse_tensor(adj_normalized) def sparse_mx_to_torch_sparse_tensor(sparse_mx): """Convert a scipy sparse matrix to a torch sparse tensor.""" sparse_mx = sparse_mx.tocoo().astype(np.float32) indices = torch.from_numpy( np.vstack((sparse_mx.row, sparse_mx.col)).astype(np.int64)) values = torch.from_numpy(sparse_mx.data) shape = torch.Size(sparse_mx.shape) return torch.sparse.FloatTensor(indices, values, shape) def plot_losses(losses, epoch_mark): for i in range(4): plt.subplot(2,2,i+1) plt.plot(losses[:,i]) plt.axvline(epoch_mark, color='r') plt.axvline(epoch_mark*2, color='g') plt.show()
2,240
29.69863
95
py
cogcn
cogcn-main/cogcn/model.py
import torch import torch.nn as nn import torch.nn.functional as F from layers import GraphConvolution class GCNAE(nn.Module): def __init__(self, input_feat_dim, hidden_dim1, hidden_dim2, dropout): super(GCNAE, self).__init__() self.encgc1 = GraphConvolution(input_feat_dim, hidden_dim1, dropout, act=F.relu) self.encgc2 = GraphConvolution(hidden_dim1, hidden_dim2, dropout, act=lambda x: x) self.decgc1 = GraphConvolution(hidden_dim2, hidden_dim1, dropout, act=F.relu) self.decgc2 = GraphConvolution(hidden_dim1, input_feat_dim, dropout, act=lambda x: x) def encode(self, x, adj): hidden1 = self.encgc1(x, adj) hidden2 = self.encgc2(hidden1, adj) return hidden2 def decode(self, hidden, adj): hidden1 = self.decgc1(hidden, adj) recon = self.decgc2(hidden1, adj) return recon def forward(self, x, adj): enc = self.encode(x, adj) dec = self.decode(enc, adj) return dec, enc class InnerProductDecoder(nn.Module): """Decoder for using inner product for prediction.""" def __init__(self, dropout, act=torch.sigmoid): super(InnerProductDecoder, self).__init__() self.dropout = dropout self.act = act def forward(self, z): z = F.dropout(z, self.dropout, training=self.training) adj = self.act(torch.mm(z, z.t())) return adj
1,415
31.930233
93
py
cogcn
cogcn-main/cogcn/kmeans.py
import sys import torch import torch.nn as nn from sklearn.cluster import KMeans class Clustering(object): def __init__(self, K, n_init=5, max_iter=250): self.K = K self.n_init = n_init self.max_iter = max_iter self.u = None self.M = None def cluster(self, embed): embed_np = embed.detach().cpu().numpy() clustering = KMeans(n_clusters=self.K, n_init=self.n_init, max_iter=self.max_iter) clustering.fit(embed_np) self.M = clustering.labels_ self.u = self._compute_centers(self.M, embed_np) def get_loss(self, embed): loss = torch.Tensor([0.]) #TODO: This may be slightly inefficient, we can fix it later to use matrix multiplications for i, clusteridx in enumerate(self.M): x = embed[i] c = self.u[clusteridx] difference = x - c err = torch.sum(torch.mul(difference, difference)) loss += err return loss def get_membership(self): return self.M def _compute_centers(self,labels, embed): """ sklearn kmeans may not give accurate cluster centers in some cases (see doc), so we compute ourselves """ clusters = {} for i,lbl in enumerate(labels): if clusters.get(lbl) is None: clusters[lbl] = [] clusters[lbl].append(torch.FloatTensor(embed[i])) centers = {} for k in clusters: all_embed = torch.stack(clusters[k]) center = torch.mean(all_embed, 0) centers[k] = center return centers
1,646
28.410714
109
py
cogcn
cogcn-main/cogcn/layers.py
import torch import torch.nn.functional as F from torch.nn.modules.module import Module from torch.nn.parameter import Parameter class GraphConvolution(Module): """ Simple GCN layer, similar to https://arxiv.org/abs/1609.02907 """ def __init__(self, in_features, out_features, dropout=0., act=F.relu): super(GraphConvolution, self).__init__() self.in_features = in_features self.out_features = out_features self.dropout = dropout self.act = act self.weight = Parameter(torch.FloatTensor(in_features, out_features)) self.reset_parameters() def reset_parameters(self): torch.nn.init.xavier_uniform_(self.weight) def forward(self, input, adj): input = F.dropout(input, self.dropout, self.training) support = torch.mm(input, self.weight) output = torch.spmm(adj, support) output = self.act(output) return output def __repr__(self): return self.__class__.__name__ + ' (' \ + str(self.in_features) + ' -> ' \ + str(self.out_features) + ')'
1,110
30.742857
77
py
cogcn
cogcn-main/cogcn/train.py
from __future__ import division from __future__ import print_function import argparse import time import sys import os import pickle import numpy as np import scipy.sparse as sp import torch import torch.nn as nn from torch import optim from matplotlib import pyplot as plt from model import GCNAE from optimizer import compute_structure_loss, compute_attribute_loss, update_o1, update_o2 from utils import load_data_cma, preprocess_graph, plot_losses from kmeans import Clustering parser = argparse.ArgumentParser() parser.add_argument('--model', type=str, default='gcn_vae', help="models used") parser.add_argument('--seed', type=int, default=42, help='Random seed.') # 5000 epochs is where the pretrain loss starts hovering around 5.0 - 6.0 parser.add_argument('--k', type=int, default=6, help='Number of clusters.') parser.add_argument('--preepochs', type=int, default=350, help='Number of epochs to pre-train.') parser.add_argument('--clusepochs', type=int, default=1, help='Number of epochs to pre-train for clustering.') parser.add_argument('--epochs', type=int, default=300, help='Number of epochs to train.') parser.add_argument('--hidden1', type=int, default=64, help='Number of units in hidden layer 1.') parser.add_argument('--hidden2', type=int, default=32, help='Number of units in hidden layer 2.') parser.add_argument('--lr', type=float, default=0.01, help='Initial learning rate.') parser.add_argument('--lambda1', type=float, default=0.1, help='Structure loss weight.') parser.add_argument('--lambda2', type=float, default=0.1, help='Attribute loss weight.') parser.add_argument('--lambda3', type=float, default=0.8, help='Clustering loss weight.') parser.add_argument('--dropout', type=float, default=0.2, help='Dropout rate (1 - keep probability).') parser.add_argument('--dataset-str', type=str, default=None, help='type of dataset.') parser.add_argument('--outfile', type=str, default='embeddings', help='output embeddings file.') args = parser.parse_args() def gae_for(args): print("Using {} dataset".format(args.dataset_str)) adj, features = load_data_cma(args.dataset_str) n_nodes, feat_dim = features.shape # Some preprocessing adj_norm = preprocess_graph(adj) model = GCNAE(feat_dim, args.hidden1, args.hidden2, args.dropout) optimizer = optim.Adam(model.parameters(), lr=args.lr) gamma = 0.98 schedule_update_interval = 400 scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma=gamma) # initialize all outlier scores with the equal values summing to 1 init_value = [1./n_nodes] * n_nodes o_1 = torch.FloatTensor(init_value) # structural outlier o_2 = torch.FloatTensor(init_value) # attribute outlier lossfn = nn.MSELoss(reduction='none') optimizer = optim.Adam(model.parameters(), lr=args.lr) lambda1 = args.lambda1 / (args.lambda1 + args.lambda2) lambda2 = args.lambda2 / (args.lambda1 + args.lambda2) kmeans = Clustering(args.k) # PRETRAIN ON STRUCTURE AND ATTRIBUTE LOSSES, NO OUTLIER LOSS for epoch in range(args.preepochs): model.train() optimizer.zero_grad() recon, embed = model(features, adj_norm) structure_loss = compute_structure_loss(adj_norm, embed, o_1) attribute_loss = compute_attribute_loss(lossfn, features, recon, o_2) loss = lambda1 * structure_loss + lambda2 * attribute_loss # Update the functions F and G (embedding network) loss.backward() cur_loss = loss.item() optimizer.step() if (epoch+1) % 100 == 0: #print("Epoch:", '%04d' % (epoch + 1), "train_loss=", "{:.5f}".format(cur_loss), "time=", "{:.5f}".format(time.time() - t), "Pretrain:",pretrain) print("Epoch:", '%04d' % (epoch + 1), "train_loss=", "{:.5f}".format(cur_loss), "lr=", "{:.5f}".format(scheduler.get_last_lr()[0])) # Initialize clusters recon, embed = model(features, adj_norm) kmeans.cluster(embed) # TRAIN ON ALL THREE LOSES WITH OUTLIER UPDATES for epoch in range(args.epochs): # Update the values of O_i1 and O_i2 o_1 = update_o1(adj_norm, embed) o_2 = update_o2(features, recon) if (epoch+1) % schedule_update_interval == 0: scheduler.step() model.train() optimizer.zero_grad() recon, embed = model(features, adj_norm) kmeans.cluster(embed) structure_loss = compute_structure_loss(adj_norm, embed, o_1) attribute_loss = compute_attribute_loss(lossfn, features, recon, o_2) clustering_loss = kmeans.get_loss(embed) loss = (args.lambda1 * structure_loss) + (args.lambda2 * attribute_loss) + (args.lambda3 * clustering_loss) #loss = (args.lambda1 * structure_loss) + (args.lambda2 * attribute_loss) # Update the functions F and G (embedding network) loss.backward() cur_loss = loss.item() optimizer.step() if (epoch+1) % 100 == 0: #print("Epoch:", '%04d' % (epoch + 1), "train_loss=", "{:.5f}".format(cur_loss), "time=", "{:.5f}".format(time.time() - t), "Pretrain:",pretrain) print("Epoch:", '%04d' % (epoch + 1), "train_loss=", "{:.5f}".format(cur_loss), "lr=", "{:.5f}".format(scheduler.get_last_lr()[0])) # Extract embeddings adj_norm = preprocess_graph(adj) recon, embed = model(features, adj_norm) embed = embed.detach().cpu().numpy() embfile = os.path.join(args.dataset_str, args.outfile+".pkl") with open(embfile,"wb") as f: pickle.dump(embed, f) o_1 = o_1.detach().cpu().numpy() o_2 = o_2.detach().cpu().numpy() outlfile = os.path.join(args.dataset_str, args.outfile+"_outliers.pkl") with open(outlfile,"wb") as f: pickle.dump([o_1, o_2], f) membfile = os.path.join(args.dataset_str, args.outfile+"_membership.pkl") with open(membfile,"wb") as f: pickle.dump(kmeans.get_membership(), f) if __name__ == '__main__': gae_for(args)
5,980
39.412162
157
py
cogcn
cogcn-main/cogcn/optimizer.py
import sys import torch import torch.nn as nn import torch.nn.modules.loss import torch.nn.functional as F from sklearn.cluster import KMeans def compute_attribute_loss(lossfn, features, recon, outlier_wt): loss = lossfn(features, recon) loss = loss.sum(dim=1) outlier_wt = torch.log(1/outlier_wt) attr_loss = torch.sum(torch.mul(outlier_wt, loss)) return attr_loss def compute_structure_loss(adj, embed, outlier_wt): # to compute F(x_i).F(x_j) embeddot = torch.mm(embed, torch.transpose(embed, 0, 1)) adj_tensor = adj.to_dense() # compute A_ij - F(x_i)*F(x_j) difference = adj_tensor - embeddot # square difference and sum loss = torch.sum(torch.mul(difference, difference), dim=1) outlier_wt = torch.log(1/outlier_wt) struct_loss = torch.sum(torch.mul(outlier_wt, loss)) return struct_loss def update_o1(adj, embed): # to compute F(x_i).F(x_j) embed = embed.data embeddot = torch.mm(embed, torch.transpose(embed, 0, 1)) adj_tensor = adj.to_dense() # compute A_ij - F(x_i)*F(x_j) difference = adj_tensor - embeddot # square difference and sum error = torch.sum(torch.mul(difference, difference), dim=1) # compute the denominator normalization_factor = torch.sum(error) # normalize the errors o1 = error/normalization_factor return o1 def update_o2(features, recon): features = features.data recon = recon.data # error = x - F(G(x)) error = features - recon # error now = (x - F(G(x)))^2, summed across dim 1 error = torch.sum(torch.mul(error, error), dim=1) # compute the denominator normalization_factor = torch.sum(error) # normalize the errors o2 = error/normalization_factor return o2
1,768
23.915493
64
py
deepglo
deepglo-master/DeepGLO/DeepGLO.py
from __future__ import print_function import torch, h5py import numpy as np from scipy.io import loadmat from torch.nn.utils import weight_norm import torch.nn as nn import torch.optim as optim import numpy as np # import matplotlib from torch.autograd import Variable import sys import itertools import torch.nn.functional as F import copy import os import gc from DeepGLO.data_loader import * from sklearn.decomposition import NMF use_cuda = True #### Assuming you have a GPU ###### from DeepGLO.utilities import * from DeepGLO.LocalModel import * from DeepGLO.metrics import * import copy import random import pickle np.random.seed(111) torch.cuda.manual_seed(111) torch.manual_seed(111) random.seed(111) def get_model(A, y, lamb=0): """ Regularized least-squares """ n_col = A.shape[1] return np.linalg.lstsq( A.T.dot(A) + lamb * np.identity(n_col), A.T.dot(y), rcond=None ) class DeepGLO(object): def __init__( self, Ymat, vbsize=150, hbsize=256, num_channels_X=[32, 32, 32, 32, 1], num_channels_Y=[32, 32, 32, 32, 1], kernel_size=7, dropout=0.2, rank=64, kernel_size_Y=7, lr=0.0005, val_len=24, end_index=20000, normalize=False, start_date="2016-1-1", freq="H", covariates=None, use_time=True, dti=None, svd=False, period=None, forward_cov=False, ): self.start_date = start_date self.freq = freq self.covariates = covariates self.use_time = use_time self.dti = dti self.dropout = dropout self.period = period self.forward_cov = forward_cov self.Xseq = TemporalConvNet( num_inputs=1, num_channels=num_channels_X, kernel_size=kernel_size, dropout=dropout, init=True, ) if normalize: self.s = np.std(Ymat[:, 0:end_index], axis=1) # self.s[self.s == 0] = 1.0 self.s += 1.0 self.m = np.mean(Ymat[:, 0:end_index], axis=1) self.Ymat = (Ymat - self.m[:, None]) / self.s[:, None] self.mini = np.abs(np.min(self.Ymat)) self.Ymat = self.Ymat + self.mini else: self.Ymat = Ymat self.normalize = normalize n, T = self.Ymat.shape t0 = end_index + 1 if t0 > T: self.Ymat = np.hstack([self.Ymat, self.Ymat[:, -1].reshape(-1, 1)]) if svd: indices = np.random.choice(self.Ymat.shape[0], rank, replace=False) X = self.Ymat[indices, 0:t0] mX = np.std(X, axis=1) mX[mX == 0] = 1.0 X = X / mX[:, None] Ft = get_model(X.transpose(), self.Ymat[:, 0:t0].transpose(), lamb=0.1) F = Ft[0].transpose() self.X = torch.from_numpy(X).float() self.F = torch.from_numpy(F).float() else: R = torch.zeros(rank, t0).float() X = torch.normal(R, 0.1) C = torch.zeros(n, rank).float() F = torch.normal(C, 0.1) self.X = X.float() self.F = F.float() self.vbsize = vbsize self.hbsize = hbsize self.num_channels_X = num_channels_X self.num_channels_Y = num_channels_Y self.kernel_size_Y = kernel_size_Y self.rank = rank self.kernel_size = kernel_size self.lr = lr self.val_len = val_len self.end_index = end_index self.D = data_loader( Ymat=self.Ymat, vbsize=vbsize, hbsize=hbsize, end_index=end_index, val_len=val_len, shuffle=False, ) def tensor2d_to_temporal(self, T): T = T.view(1, T.size(0), T.size(1)) T = T.transpose(0, 1) return T def temporal_to_tensor2d(self, T): T = T.view(T.size(0), T.size(2)) return T def calculate_newX_loss_vanilla(self, Xn, Fn, Yn, Xf, alpha): Yout = torch.mm(Fn, Xn) cr1 = nn.L1Loss() cr2 = nn.MSELoss() l1 = cr2(Yout, Yn) / torch.mean(Yn ** 2) l2 = cr2(Xn, Xf) / torch.mean(Xf ** 2) return (1 - alpha) * l1 + alpha * l2 def recover_future_X( self, last_step, future, cpu=True, num_epochs=50, alpha=0.5, vanilla=True, tol=1e-7, ): rg = max( 1 + 2 * (self.kernel_size - 1) * 2 ** (len(self.num_channels_X) - 1), 1 + 2 * (self.kernel_size_Y - 1) * 2 ** (len(self.num_channels_Y) - 1), ) X = self.X[:, last_step - rg : last_step] X = self.tensor2d_to_temporal(X) outX = self.predict_future(model=self.Xseq, inp=X, future=future, cpu=cpu) outX = self.temporal_to_tensor2d(outX) Xf = outX[:, -future::] Yn = self.Ymat[:, last_step : last_step + future] Yn = torch.from_numpy(Yn).float() if cpu: self.Xseq = self.Xseq.cpu() else: Yn = Yn.cuda() Xf = Xf.cuda() Fn = self.F Xt = torch.zeros(self.rank, future).float() Xn = torch.normal(Xt, 0.1) if not cpu: Xn = Xn.cuda() lprev = 0 for i in range(num_epochs): Xn = Variable(Xn, requires_grad=True) optim_Xn = optim.Adam(params=[Xn], lr=self.lr) optim_Xn.zero_grad() loss = self.calculate_newX_loss_vanilla( Xn, Fn.detach(), Yn.detach(), Xf.detach(), alpha ) loss.backward() optim_Xn.step() # Xn = torch.clamp(Xn.detach(), min=0) if np.abs(lprev - loss.cpu().item()) <= tol: break if i % 1000 == 0: print("Recovery Loss: " + str(loss.cpu().item())) lprev = loss.cpu().item() self.Xseq = self.Xseq.cuda() return Xn.detach() def step_factX_loss(self, inp, out, last_vindex, last_hindex, reg=0.0): Xout = self.X[:, last_hindex + 1 : last_hindex + 1 + out.size(2)] Fout = self.F[self.D.I[last_vindex : last_vindex + out.size(0)], :] if use_cuda: Xout = Xout.cuda() Fout = Fout.cuda() Xout = Variable(Xout, requires_grad=True) out = self.temporal_to_tensor2d(out) optim_X = optim.Adam(params=[Xout], lr=self.lr) Hout = torch.matmul(Fout, Xout) optim_X.zero_grad() loss = torch.mean(torch.pow(Hout - out.detach(), 2)) l2 = torch.mean(torch.pow(Xout, 2)) r = loss.detach() / l2.detach() loss = loss + r * reg * l2 loss.backward() optim_X.step() # Xout = torch.clamp(Xout, min=0) self.X[:, last_hindex + 1 : last_hindex + 1 + inp.size(2)] = Xout.cpu().detach() return loss def step_factF_loss(self, inp, out, last_vindex, last_hindex, reg=0.0): Xout = self.X[:, last_hindex + 1 : last_hindex + 1 + out.size(2)] Fout = self.F[self.D.I[last_vindex : last_vindex + out.size(0)], :] if use_cuda: Xout = Xout.cuda() Fout = Fout.cuda() Fout = Variable(Fout, requires_grad=True) optim_F = optim.Adam(params=[Fout], lr=self.lr) out = self.temporal_to_tensor2d(out) Hout = torch.matmul(Fout, Xout) optim_F.zero_grad() loss = torch.mean(torch.pow(Hout - out.detach(), 2)) l2 = torch.mean(torch.pow(Fout, 2)) r = loss.detach() / l2.detach() loss = loss + r * reg * l2 loss.backward() optim_F.step() self.F[ self.D.I[last_vindex : last_vindex + inp.size(0)], : ] = Fout.cpu().detach() return loss def step_temporal_loss_X(self, inp, last_vindex, last_hindex): Xin = self.X[:, last_hindex : last_hindex + inp.size(2)] Xout = self.X[:, last_hindex + 1 : last_hindex + 1 + inp.size(2)] for p in self.Xseq.parameters(): p.requires_grad = False if use_cuda: Xin = Xin.cuda() Xout = Xout.cuda() Xin = Variable(Xin, requires_grad=True) Xout = Variable(Xout, requires_grad=True) optim_out = optim.Adam(params=[Xout], lr=self.lr) Xin = self.tensor2d_to_temporal(Xin) Xout = self.tensor2d_to_temporal(Xout) hatX = self.Xseq(Xin) optim_out.zero_grad() loss = torch.mean(torch.pow(Xout - hatX.detach(), 2)) loss.backward() optim_out.step() # Xout = torch.clamp(Xout, min=0) temp = self.temporal_to_tensor2d(Xout.detach()) self.X[:, last_hindex + 1 : last_hindex + 1 + inp.size(2)] = temp return loss def predict_future_batch(self, model, inp, future=10, cpu=True): if cpu: model = model.cpu() inp = inp.cpu() else: inp = inp.cuda() out = model(inp) output = out[:, :, out.size(2) - 1].view(out.size(0), out.size(1), 1) out = torch.cat((inp, output), dim=2) torch.cuda.empty_cache() for i in range(future - 1): inp = out out = model(inp) output = out[:, :, out.size(2) - 1].view(out.size(0), out.size(1), 1) out = torch.cat((inp, output), dim=2) torch.cuda.empty_cache() out = self.temporal_to_tensor2d(out) out = np.array(out.cpu().detach()) return out def predict_future(self, model, inp, future=10, cpu=True, bsize=90): n = inp.size(0) inp = inp.cpu() ids = np.arange(0, n, bsize) ids = list(ids) + [n] out = self.predict_future_batch(model, inp[ids[0] : ids[1], :, :], future, cpu) torch.cuda.empty_cache() for i in range(1, len(ids) - 1): temp = self.predict_future_batch( model, inp[ids[i] : ids[i + 1], :, :], future, cpu ) torch.cuda.empty_cache() out = np.vstack([out, temp]) out = torch.from_numpy(out).float() return self.tensor2d_to_temporal(out) def predict_global( self, ind, last_step=100, future=10, cpu=False, normalize=False, bsize=90 ): if ind is None: ind = np.arange(self.Ymat.shape[0]) if cpu: self.Xseq = self.Xseq.cpu() self.Xseq = self.Xseq.eval() rg = max( 1 + 2 * (self.kernel_size - 1) * 2 ** (len(self.num_channels_X) - 1), 1 + 2 * (self.kernel_size_Y - 1) * 2 ** (len(self.num_channels_Y) - 1), ) X = self.X[:, last_step - rg : last_step] n = X.size(0) T = X.size(1) X = self.tensor2d_to_temporal(X) outX = self.predict_future( model=self.Xseq, inp=X, future=future, cpu=cpu, bsize=bsize ) outX = self.temporal_to_tensor2d(outX) F = self.F Y = torch.matmul(F, outX) Y = np.array(Y[ind, :].cpu().detach()) self.Xseq = self.Xseq.cuda() del F torch.cuda.empty_cache() for p in self.Xseq.parameters(): p.requires_grad = True if normalize: Y = Y - self.mini Y = Y * self.s[ind, None] + self.m[ind, None] return Y else: return Y def train_Xseq(self, Ymat, num_epochs=20, early_stop=False, tenacity=3): seq = self.Xseq num_channels = self.num_channels_X kernel_size = self.kernel_size vbsize = min(self.vbsize, Ymat.shape[0] / 2) for p in seq.parameters(): p.requires_grad = True TC = LocalModel( Ymat=Ymat, num_inputs=1, num_channels=num_channels, kernel_size=kernel_size, vbsize=vbsize, hbsize=self.hbsize, normalize=False, end_index=self.end_index - self.val_len, val_len=self.val_len, lr=self.lr, num_epochs=num_epochs, ) TC.train_model(early_stop=early_stop, tenacity=tenacity) self.Xseq = TC.seq def train_factors( self, reg_X=0.0, reg_F=0.0, mod=5, early_stop=False, tenacity=3, ind=None, seed=False, ): self.D.epoch = 0 self.D.vindex = 0 self.D.hindex = 0 if use_cuda: self.Xseq = self.Xseq.cuda() for p in self.Xseq.parameters(): p.requires_grad = True l_F = [0.0] l_X = [0.0] l_X_temporal = [0.0] iter_count = 0 vae = float("inf") scount = 0 Xbest = self.X.clone() Fbest = self.F.clone() while self.D.epoch < self.num_epochs: last_epoch = self.D.epoch last_vindex = self.D.vindex last_hindex = self.D.hindex inp, out, vindex, hindex = self.D.next_batch(option=1) if use_cuda: inp = inp.float().cuda() out = out.float().cuda() if iter_count % mod >= 0: l1 = self.step_factF_loss(inp, out, last_vindex, last_hindex, reg=reg_F) l_F = l_F + [l1.cpu().item()] if iter_count % mod >= 0: l1 = self.step_factX_loss(inp, out, last_vindex, last_hindex, reg=reg_X) l_X = l_X + [l1.cpu().item()] if seed == False and iter_count % mod == 1: l2 = self.step_temporal_loss_X(inp, last_vindex, last_hindex) l_X_temporal = l_X_temporal + [l2.cpu().item()] iter_count = iter_count + 1 if self.D.epoch > last_epoch: print("Entering Epoch# ", self.D.epoch) print("Factorization Loss F: ", np.mean(l_F)) print("Factorization Loss X: ", np.mean(l_X)) print("Temporal Loss X: ", np.mean(l_X_temporal)) if ind is None: ind = np.arange(self.Ymat.shape[0]) else: ind = ind inp = self.predict_global( ind, last_step=self.end_index - self.val_len, future=self.val_len, cpu=False, ) R = self.Ymat[ind, self.end_index - self.val_len : self.end_index] S = inp[:, -self.val_len : :] ve = np.abs(R - S).mean() / np.abs(R).mean() print("Validation Loss (Global): ", ve) if ve <= vae: vae = ve scount = 0 Xbest = self.X.clone() Fbest = self.F.clone() # Xseqbest = TemporalConvNet( # num_inputs=1, # num_channels=self.num_channels_X, # kernel_size=self.kernel_size, # dropout=self.dropout, # ) # Xseqbest.load_state_dict(self.Xseq.state_dict()) Xseqbest = pickle.loads(pickle.dumps(self.Xseq)) else: scount += 1 if scount > tenacity and early_stop: print("Early Stopped") self.X = Xbest self.F = Fbest self.Xseq = Xseqbest if use_cuda: self.Xseq = self.Xseq.cuda() break def create_Ycov(self): t0 = self.end_index + 1 self.D.epoch = 0 self.D.vindex = 0 self.D.hindex = 0 Ycov = copy.deepcopy(self.Ymat[:, 0:t0]) Ymat_now = self.Ymat[:, 0:t0] if use_cuda: self.Xseq = self.Xseq.cuda() self.Xseq = self.Xseq.eval() while self.D.epoch < 1: last_epoch = self.D.epoch last_vindex = self.D.vindex last_hindex = self.D.hindex inp, out, vindex, hindex = self.D.next_batch(option=1) if use_cuda: inp = inp.cuda() Xin = self.tensor2d_to_temporal(self.X[:, last_hindex : last_hindex + inp.size(2)]).cuda() Xout = self.temporal_to_tensor2d(self.Xseq(Xin)).cpu() Fout = self.F[self.D.I[last_vindex : last_vindex + out.size(0)], :] output = np.array(torch.matmul(Fout, Xout).detach()) Ycov[ last_vindex : last_vindex + output.shape[0], last_hindex + 1 : last_hindex + 1 + output.shape[1], ] = output for p in self.Xseq.parameters(): p.requires_grad = True if self.period is None: Ycov_wc = np.zeros(shape=[Ycov.shape[0], 1, Ycov.shape[1]]) if self.forward_cov: Ycov_wc[:, 0, 0:-1] = Ycov[:, 1::] else: Ycov_wc[:, 0, :] = Ycov else: Ycov_wc = np.zeros(shape=[Ycov.shape[0], 2, Ycov.shape[1]]) if self.forward_cov: Ycov_wc[:, 0, 0:-1] = Ycov[:, 1::] else: Ycov_wc[:, 0, :] = Ycov Ycov_wc[:, 1, self.period - 1 : :] = Ymat_now[:, 0 : -(self.period - 1)] return Ycov_wc def train_Yseq(self, num_epochs=20, early_stop=False, tenacity=7): Ycov = self.create_Ycov() self.Yseq = LocalModel( self.Ymat, num_inputs=1, num_channels=self.num_channels_Y, kernel_size=self.kernel_size_Y, dropout=self.dropout, vbsize=self.vbsize, hbsize=self.hbsize, num_epochs=num_epochs, lr=self.lr, val_len=self.val_len, test=True, end_index=self.end_index - self.val_len, normalize=False, start_date=self.start_date, freq=self.freq, covariates=self.covariates, use_time=self.use_time, dti=self.dti, Ycov=Ycov, ) self.Yseq.train_model(early_stop=early_stop, tenacity=tenacity) def train_all_models( self, init_epochs=100, alt_iters=10, y_iters=200, tenacity=7, mod=5 ): print("Initializing Factors.....") self.num_epochs = init_epochs self.train_factors() if alt_iters % 2 == 1: alt_iters += 1 print("Starting Alternate Training.....") for i in range(1, alt_iters): if i % 2 == 0: print( "--------------------------------------------Training Factors. Iter#: " + str(i) + "-------------------------------------------------------" ) self.num_epochs = 300 self.train_factors( seed=False, early_stop=True, tenacity=tenacity, mod=mod ) else: print( "--------------------------------------------Training Local Model. Iter#: " + str(i) + "-------------------------------------------------------" ) self.num_epochs = 300 T = np.array(self.X.cpu().detach()) self.train_Xseq( Ymat=T, num_epochs=self.num_epochs, early_stop=True, tenacity=tenacity, ) self.num_epochs = y_iters self.train_Yseq(num_epochs=y_iters, early_stop=True, tenacity=tenacity) def predict( self, ind=None, last_step=100, future=10, cpu=False, normalize=False, bsize=90 ): if ind is None: ind = np.arange(self.Ymat.shape[0]) if cpu: self.Xseq = self.Xseq.cpu() self.Yseq.seq = self.Yseq.seq.eval() self.Xseq = self.Xseq.eval() rg = max( 1 + 2 * (self.kernel_size - 1) * 2 ** (len(self.num_channels_X) - 1), 1 + 2 * (self.kernel_size_Y - 1) * 2 ** (len(self.num_channels_Y) - 1), ) covs = self.Yseq.covariates[:, last_step - rg : last_step + future] # print(covs.shape) yc = self.predict_global( ind=ind, last_step=last_step, future=future, cpu=cpu, normalize=False, bsize=bsize, ) if self.period is None: ycovs = np.zeros(shape=[yc.shape[0], 1, yc.shape[1]]) if self.forward_cov: ycovs[:, 0, 0:-1] = yc[:, 1::] else: ycovs[:, 0, :] = yc else: ycovs = np.zeros(shape=[yc.shape[0], 2, yc.shape[1]]) if self.forward_cov: ycovs[:, 0, 0:-1] = yc[:, 1::] else: ycovs[:, 0, :] = yc period = self.period while last_step + future - (period - 1) > last_step + 1: period += self.period ycovs[:, 1, period - 1 : :] = self.Ymat[ :, last_step - rg : last_step + future - (period - 1) ] ### this seems like we are looking ahead, but it will not use the last coordinate, which is the only new point added # print(ycovs.shape) Y = self.Yseq.predict_future( data_in=self.Ymat[ind, last_step - rg : last_step], covariates=covs, ycovs=ycovs, future=future, cpu=cpu, bsize=bsize, normalize=False, ) if normalize: Y = Y - self.mini Y = Y * self.s[ind, None] + self.m[ind, None] return Y else: return Y def rolling_validation(self, Ymat, tau=24, n=7, bsize=90, cpu=False, alpha=0.3): prevX = self.X.clone() prev_index = self.end_index out = self.predict( last_step=self.end_index, future=tau, bsize=bsize, cpu=cpu, normalize=self.normalize, ) out_global = self.predict_global( np.arange(self.Ymat.shape[0]), last_step=self.end_index, future=tau, cpu=cpu, normalize=self.normalize, bsize=bsize, ) predicted_values = [] actual_values = [] predicted_values_global = [] S = out[:, -tau::] S_g = out_global[:, -tau::] predicted_values += [S] predicted_values_global += [S_g] R = Ymat[:, self.end_index : self.end_index + tau] actual_values += [R] print("Current window wape: " + str(wape(S, R))) self.Xseq = self.Xseq.eval() self.Yseq.seq = self.Yseq.seq.eval() for i in range(n - 1): Xn = self.recover_future_X( last_step=self.end_index + 1, future=tau, num_epochs=100000, alpha=alpha, vanilla=True, cpu=True, ) self.X = torch.cat([self.X, Xn], dim=1) self.end_index += tau out = self.predict( last_step=self.end_index, future=tau, bsize=bsize, cpu=cpu, normalize=self.normalize, ) out_global = self.predict_global( np.arange(self.Ymat.shape[0]), last_step=self.end_index, future=tau, cpu=cpu, normalize=self.normalize, bsize=bsize, ) S = out[:, -tau::] S_g = out_global[:, -tau::] predicted_values += [S] predicted_values_global += [S_g] R = Ymat[:, self.end_index : self.end_index + tau] actual_values += [R] print("Current window wape: " + str(wape(S, R))) predicted = np.hstack(predicted_values) predicted_global = np.hstack(predicted_values_global) actual = np.hstack(actual_values) dic = {} dic["wape"] = wape(predicted, actual) dic["mape"] = mape(predicted, actual) dic["smape"] = smape(predicted, actual) dic["mae"] = np.abs(predicted - actual).mean() dic["rmse"] = np.sqrt(((predicted - actual) ** 2).mean()) dic["nrmse"] = dic["rmse"] / np.sqrt(((actual) ** 2).mean()) dic["wape_global"] = wape(predicted_global, actual) dic["mape_global"] = mape(predicted_global, actual) dic["smape_global"] = smape(predicted_global, actual) dic["mae_global"] = np.abs(predicted_global - actual).mean() dic["rmse_global"] = np.sqrt(((predicted_global - actual) ** 2).mean()) dic["nrmse_global"] = dic["rmse"] / np.sqrt(((actual) ** 2).mean()) baseline = Ymat[:, Ymat.shape[1] - n * tau - tau : Ymat.shape[1] - tau] dic["baseline_wape"] = wape(baseline, actual) dic["baseline_mape"] = mape(baseline, actual) dic["baseline_smape"] = smape(baseline, actual) self.X = prevX self.end_index = prev_index return dic
25,258
32.235526
131
py
deepglo
deepglo-master/DeepGLO/data_loader.py
import torch, h5py import numpy as np from scipy.io import loadmat import torch.nn as nn import torch.optim as optim import numpy as np # import matplotlib from torch.autograd import Variable import itertools from sklearn.preprocessing import normalize import datetime import json import os, sys import pandas as pd import pyarrow.parquet as pq from DeepGLO.Ftree import * class data_loader(object): """ Data Loader Class for DeepGLO """ def __init__( self, Ymat, covariates=None, Ycov=None, vbsize=200, hbsize=100, end_index=20000, val_len=30, shuffle=False, ): """ Argeuments: Ymat: time-series matrix n*T covariates: global covariates common for all time series r*T, where r is the number of covariates Ycov: per time-series covariates n*l*T, l such covariates per time-series All of the above arguments are numpy arrays vbsize: vertical batch size hbsize: horizontal batch size end_index: training and validation set is only from 0:end_index val_len: validation length. The last 'val_len' time-points for every time-series is the validation set shuffle: data is shuffles if True (this is deprecated and set to False) """ n, T = Ymat.shape self.vindex = 0 self.hindex = 0 self.epoch = 0 self.vbsize = vbsize self.hbsize = hbsize self.Ymat = Ymat self.val_len = val_len self.end_index = end_index self.val_index = np.random.randint(0, n - self.vbsize - 5) self.shuffle = shuffle self.I = np.array(range(n)) self.covariates = covariates self.Ycov = Ycov def next_batch(self, option=1): """ Arguments: option = 1 means data is returned as pytorch tensor of shape nd*cd*td where nd is vbsize, hb is hsize and cd is the number os channels (depends on covariates) option = 0 is deprecated Returns: inp: input batch out: one shifted output batch vindex: strating vertical index of input batch hindex: starting horizontal index of input batch """ n, T = self.Ymat.shape if self.hindex + self.hbsize + 1 >= self.end_index: pr_hindex = self.hindex self.hindex = 0 if self.vindex + self.vbsize >= n: pr_vindex = self.vindex self.vindex = 0 self.epoch = self.epoch + 1 if self.shuffle: I = np.random.choice(n, n, replace=False) self.I = I self.Ymat = self.Ymat[self.I, :] else: pr_vindex = self.vindex self.vindex = self.vindex + self.vbsize else: pr_hindex = self.hindex self.hindex = self.hindex + self.hbsize pr_vindex = self.vindex data = self.Ymat[ int(pr_vindex) : int(pr_vindex + self.vbsize), int(pr_hindex) : int(min(self.end_index, pr_hindex + self.hbsize)), ] out_data = self.Ymat[ int(pr_vindex) : int(pr_vindex + self.vbsize), int(pr_hindex + 1) : int(min(self.end_index, pr_hindex + self.hbsize) + 1), ] nd, Td = data.shape if self.covariates is not None: covs = self.covariates[ :, int(pr_hindex) : int(min(self.end_index, pr_hindex + self.hbsize)) ] rcovs = np.repeat( covs.reshape(1, covs.shape[0], covs.shape[1]), repeats=nd, axis=0 ) if self.Ycov is not None: ycovs = self.Ycov[ int(pr_vindex) : int(pr_vindex + self.vbsize), :, int(pr_hindex) : int(min(self.end_index, pr_hindex + self.hbsize)), ] if option == 1: inp = torch.from_numpy(data).view(1, nd, Td) out = torch.from_numpy(out_data).view(1, nd, Td) if self.covariates is not None: rcovs = torch.from_numpy(rcovs).float() if self.Ycov is not None: ycovs = torch.from_numpy(ycovs).float() inp = inp.transpose(0, 1).float() if self.covariates is not None: inp = torch.cat((inp, rcovs), 1) if self.Ycov is not None: inp = torch.cat((inp, ycovs), 1) out = out.transpose(0, 1).float() else: inp = torch.from_numpy(data).float() out = torch.from_numpy(out_data).float() inp[torch.isnan(inp)] = 0 out[torch.isnan(out)] = 0 return inp, out, self.vindex, self.hindex def supply_test(self, option=1): """ Supplies validation set in the same format as above """ n, T = self.Ymat.shape index = self.val_index in_data = self.Ymat[ int(index) : int(index + self.vbsize), int(self.end_index) : int(self.end_index + self.val_len), ] out_data = self.Ymat[ int(index) : int(index + self.vbsize), int(self.end_index + 1) : int(self.end_index + self.val_len + 1), ] nd, Td = in_data.shape if self.covariates is not None: covs = self.covariates[ :, int(self.end_index) : int(self.end_index + self.val_len) ] rcovs = np.repeat( covs.reshape(1, covs.shape[0], covs.shape[1]), repeats=nd, axis=0 ) if self.Ycov is not None: ycovs = self.Ycov[ int(index) : int(index + self.vbsize), :, int(self.end_index) : int(self.end_index + self.val_len), ] if option == 1: inp = torch.from_numpy(in_data).view(1, nd, Td) inp = inp.transpose(0, 1).float() if self.covariates is not None: rcovs = torch.from_numpy(rcovs).float() if self.Ycov is not None: ycovs = torch.from_numpy(ycovs).float() out = torch.from_numpy(out_data).view(1, nd, Td) if self.covariates is not None: inp = torch.cat((inp, rcovs), 1) if self.Ycov is not None: inp = torch.cat((inp, ycovs), 1) out = out.transpose(0, 1).float() else: inp = torch.from_numpy(in_data).float() out = torch.from_numpy(out_data).float() return inp, out, self.vindex, self.hindex
6,610
34.735135
167
py
deepglo
deepglo-master/DeepGLO/Ftree.py
import numpy as np import pandas as pd class FplusTreeSampling(object): """ F+ tree for sampling from a large population Construct in O(N) time Sample and update in O(log(N)) time """ def __init__(self, dimension, weights=None): self.dimension = dimension self.layers = int(np.ceil(np.log2(dimension))) self.F = [np.array([])] * self.layers self.initialize(weights) def initialize(self, weights=None): """ initialize F+ tree with uniform weights """ # initialzie last layer with weights if weights is None: weight = 1.0 / self.dimension self.F[-1] = np.ones((self.dimension,)) * weight else: self.F[-1] = weights # for l in range(self.layers-2, -1 , -1): # weight *= 2 # length = int(np.ceil(self.F[l+1].shape[0]/2.0)) # self.F[l] = np.ones((length,)) * weight # if len(self.F[l+1])%2 != 0 : # self.F[l][-1] = self.F[l+1][-1] # else: # self.F[l][-1] = self.F[l+1][-1] + self.F[l+1][-2] # assert(self.F[0][0] + self.F[0][1] == 1.0) for l in range(self.layers - 2, -1, -1): length = int(np.ceil(self.F[l + 1].shape[0] / 2.0)) self.F[l] = np.ones((length,)) if len(self.F[l + 1]) % 2 != 0: self.F[l][:-1] = self.F[l + 1][:-1].reshape((-1, 2)).sum(axis=1) self.F[l][-1] = self.F[l + 1][-1] else: self.F[l] = self.F[l + 1].reshape((-1, 2)).sum(axis=1) def print_graph(self): if self.dimension > 1000: print("Are you crazy?") return for fl in self.F: for prob in fl: print(prob, " ") print("||") def total_weight(self): """ return the total weight sum """ return self.F[0][0] + self.F[0][1] def get_weight(self, indices): """ return the weight of given indices """ return self.F[-1][indices] def sample_batch(self, batch_size): """ sample a batch without replacement """ indices = np.zeros((batch_size,), dtype=np.int) weights = np.zeros((batch_size,), dtype=np.float) for i in range(batch_size): indices[i] = self.__sample() weights[i] = self.F[-1][indices[i]] self.__update(indices[i], 0) # wighout replacement self.update_batch(indices, weights) # resume their original weights return indices def update_batch(self, indices, probs): """ update weights of a given batch """ for i, p in zip(indices, probs): self.__update(i, p) def __sample(self): """ sample a single node, in log(N) time """ u = np.random.sample() * self.F[0][0] i = 0 for fl in self.F[1:]: # i_left = 2*i # i_right = 2*i +1 if u > fl[2 * i] and fl.shape[0] >= 2 * (i + 1): # then chose i_right u -= fl[2 * i] i = 2 * i + 1 else: i = 2 * i return i def __update(self, idx, prob): """ update weight of a single node, in log(N) time """ delta = prob - self.F[-1][idx] for l in range(self.layers - 1, -1, -1): self.F[l][idx] += delta idx = idx // 2
3,508
29.513043
82
py
deepglo
deepglo-master/DeepGLO/utilities.py
import pandas as pd import numpy as np import datetime def last_days(num=60, date=datetime.datetime(2018, 6, 20)): y = [str(date.year) + "%02d" % date.month + "%02d" % date.day] for i in range(1, num): d = date - datetime.timedelta(days=i) y = y + [str(d.year) + "%02d" % d.month + "%02d" % d.day] return y def date_range(d1=datetime.datetime(2018, 3, 19), d2=datetime.datetime(2018, 6, 20)): td = d2 - d1 ndays = td.days y = [str(d1.year) + "%02d" % d1.month + "%02d" % d1.day] for i in range(1, ndays + 1): d = d1 + datetime.timedelta(days=i) y = y + [str(d.year) + "%02d" % d.month + "%02d" % d.day] return y
680
29.954545
85
py
deepglo
deepglo-master/DeepGLO/time.py
import pandas as pd import numpy as np class TimeCovariates(object): def __init__(self, start_date, num_ts=100, freq="H", normalized=True): self.start_date = start_date self.num_ts = num_ts self.freq = freq self.normalized = normalized self.dti = pd.date_range(self.start_date, periods=self.num_ts, freq=self.freq) def _minute_of_hour(self): minutes = np.array(self.dti.minute, dtype=np.float) if self.normalized: minutes = minutes / 59.0 - 0.5 return minutes def _hour_of_day(self): hours = np.array(self.dti.hour, dtype=np.float) if self.normalized: hours = hours / 23.0 - 0.5 return hours def _day_of_week(self): dayWeek = np.array(self.dti.dayofweek, dtype=np.float) if self.normalized: dayWeek = dayWeek / 6.0 - 0.5 return dayWeek def _day_of_month(self): dayMonth = np.array(self.dti.day, dtype=np.float) if self.normalized: dayMonth = dayMonth / 30.0 - 0.5 return dayMonth def _day_of_year(self): dayYear = np.array(self.dti.dayofyear, dtype=np.float) if self.normalized: dayYear = dayYear / 364.0 - 0.5 return dayYear def _month_of_year(self): monthYear = np.array(self.dti.month, dtype=np.float) if self.normalized: monthYear = monthYear / 11.0 - 0.5 return monthYear def _week_of_year(self): weekYear = np.array(self.dti.weekofyear, dtype=np.float) if self.normalized: weekYear = weekYear / 51.0 - 0.5 return weekYear def get_covariates(self): MOH = self._minute_of_hour().reshape(1, -1) HOD = self._hour_of_day().reshape(1, -1) DOM = self._day_of_month().reshape(1, -1) DOW = self._day_of_week().reshape(1, -1) DOY = self._day_of_year().reshape(1, -1) MOY = self._month_of_year().reshape(1, -1) WOY = self._week_of_year().reshape(1, -1) all_covs = [MOH, HOD, DOM, DOW, DOY, MOY, WOY] return np.vstack(all_covs)
2,134
30.865672
86
py
deepglo
deepglo-master/DeepGLO/metrics.py
import numpy as np def smape(P, A): nz = np.where(A > 0) Pz = P[nz] Az = A[nz] return np.mean(2 * np.abs(Az - Pz) / (np.abs(Az) + np.abs(Pz))) def mape(P, A): nz = np.where(A > 0) Pz = P[nz] Az = A[nz] return np.mean(np.abs(Az - Pz) / np.abs(Az)) def wape(P, A): return np.mean(np.abs(A - P)) / np.mean(np.abs(A)) def confidence_score(func, P1, A1, num): P = P1.flatten() A = A1.flatten() values = [] for i in range(num): I = np.random.choice(len(P), int(0.63 * len(P)), replace=False) p = P[I] a = A[I] values += [func(p, a)] return np.mean(values), np.std(values) def confidence_score_dim(func, P1, A1): composite = np.hstack([P1, A1]) n, m = P1.shape values = np.apply_along_axis(lambda x: func(x[0:m], x[m::]), axis=0, arr=composite) return np.mean(values), np.std(values)
898
18.543478
87
py
deepglo
deepglo-master/DeepGLO/__init__.py
# Implement your code here.
28
13.5
27
py
deepglo
deepglo-master/DeepGLO/LocalModel.py
import torch, h5py import numpy as np from scipy.io import loadmat from torch.nn.utils import weight_norm import torch.nn as nn import torch.optim as optim import numpy as np # import matplotlib from torch.autograd import Variable import itertools import torch.nn.functional as F from DeepGLO.data_loader import * use_cuda = True #### Assuming you have a GPU ###### from DeepGLO.utilities import * from DeepGLO.time import * from DeepGLO.metrics import * import random import pickle np.random.seed(111) torch.cuda.manual_seed(111) torch.manual_seed(111) random.seed(111) class Chomp1d(nn.Module): def __init__(self, chomp_size): super(Chomp1d, self).__init__() self.chomp_size = chomp_size def forward(self, x): return x[:, :, : -self.chomp_size].contiguous() class TemporalBlock(nn.Module): def __init__( self, n_inputs, n_outputs, kernel_size, stride, dilation, padding, dropout=0.1, init=True, ): super(TemporalBlock, self).__init__() self.kernel_size = kernel_size self.conv1 = weight_norm( nn.Conv1d( n_inputs, n_outputs, kernel_size, stride=stride, padding=padding, dilation=dilation, ) ) self.chomp1 = Chomp1d(padding) self.relu1 = nn.ReLU() self.dropout1 = nn.Dropout(dropout) self.conv2 = weight_norm( nn.Conv1d( n_outputs, n_outputs, kernel_size, stride=stride, padding=padding, dilation=dilation, ) ) self.chomp2 = Chomp1d(padding) self.relu2 = nn.ReLU() self.dropout2 = nn.Dropout(dropout) self.net = nn.Sequential( self.conv1, self.chomp1, self.relu1, self.dropout1, self.conv2, self.chomp2, self.relu2, self.dropout2, ) self.downsample = ( nn.Conv1d(n_inputs, n_outputs, 1) if n_inputs != n_outputs else None ) self.init = init self.relu = nn.ReLU() self.init_weights() def init_weights(self): if self.init: nn.init.normal_(self.conv1.weight, std=1e-3) nn.init.normal_(self.conv2.weight, std=1e-3) self.conv1.weight[:, 0, :] += ( 1.0 / self.kernel_size ) ###new initialization scheme self.conv2.weight += 1.0 / self.kernel_size ###new initialization scheme nn.init.normal_(self.conv1.bias, std=1e-6) nn.init.normal_(self.conv2.bias, std=1e-6) else: nn.init.xavier_uniform_(self.conv1.weight) nn.init.xavier_uniform_(self.conv2.weight) if self.downsample is not None: self.downsample.weight.data.normal_(0, 0.1) def forward(self, x): out = self.net(x) res = x if self.downsample is None else self.downsample(x) return self.relu(out + res) class TemporalBlock_last(nn.Module): def __init__( self, n_inputs, n_outputs, kernel_size, stride, dilation, padding, dropout=0.2, init=True, ): super(TemporalBlock_last, self).__init__() self.kernel_size = kernel_size self.conv1 = weight_norm( nn.Conv1d( n_inputs, n_outputs, kernel_size, stride=stride, padding=padding, dilation=dilation, ) ) self.chomp1 = Chomp1d(padding) self.relu1 = nn.ReLU() self.dropout1 = nn.Dropout(dropout) self.conv2 = weight_norm( nn.Conv1d( n_outputs, n_outputs, kernel_size, stride=stride, padding=padding, dilation=dilation, ) ) self.chomp2 = Chomp1d(padding) self.relu2 = nn.ReLU() self.dropout2 = nn.Dropout(dropout) self.net = nn.Sequential( self.conv1, self.chomp1, self.dropout1, self.conv2, self.chomp2, self.dropout2, ) self.downsample = ( nn.Conv1d(n_inputs, n_outputs, 1) if n_inputs != n_outputs else None ) self.init = init self.relu = nn.ReLU() self.init_weights() def init_weights(self): if self.init: nn.init.normal_(self.conv1.weight, std=1e-3) nn.init.normal_(self.conv2.weight, std=1e-3) self.conv1.weight[:, 0, :] += ( 1.0 / self.kernel_size ) ###new initialization scheme self.conv2.weight += 1.0 / self.kernel_size ###new initialization scheme nn.init.normal_(self.conv1.bias, std=1e-6) nn.init.normal_(self.conv2.bias, std=1e-6) else: nn.init.xavier_uniform_(self.conv1.weight) nn.init.xavier_uniform_(self.conv2.weight) if self.downsample is not None: self.downsample.weight.data.normal_(0, 0.1) def forward(self, x): out = self.net(x) res = x if self.downsample is None else self.downsample(x) return out + res class TemporalConvNet(nn.Module): def __init__(self, num_inputs, num_channels, kernel_size=2, dropout=0.1, init=True): super(TemporalConvNet, self).__init__() layers = [] self.num_channels = num_channels self.num_inputs = num_inputs self.kernel_size = kernel_size self.dropout = dropout num_levels = len(num_channels) for i in range(num_levels): dilation_size = 2 ** i in_channels = num_inputs if i == 0 else num_channels[i - 1] out_channels = num_channels[i] if i == num_levels - 1: layers += [ TemporalBlock_last( in_channels, out_channels, kernel_size, stride=1, dilation=dilation_size, padding=(kernel_size - 1) * dilation_size, dropout=dropout, init=init, ) ] else: layers += [ TemporalBlock( in_channels, out_channels, kernel_size, stride=1, dilation=dilation_size, padding=(kernel_size - 1) * dilation_size, dropout=dropout, init=init, ) ] self.network = nn.Sequential(*layers) def forward(self, x): return self.network(x) class LocalModel(object): def __init__( self, Ymat, num_inputs=1, num_channels=[32, 32, 32, 32, 32, 1], kernel_size=7, dropout=0.2, vbsize=300, hbsize=128, num_epochs=100, lr=0.0005, val_len=10, test=True, end_index=120, normalize=False, start_date="2016-1-1", freq="H", covariates=None, use_time=False, dti=None, Ycov=None, ): """ Arguments: Ymat: input time-series n*T num_inputs: always set to 1 num_channels: list containing channel progression of temporal comvolution network kernel_size: kernel size of temporal convolution filters dropout: dropout rate for each layer vbsize: vertical batch size hbsize: horizontal batch size num_epochs: max. number of epochs lr: learning rate val_len: validation length test: always set to True end_index: no data is touched fro training or validation beyond end_index normalize: normalize dataset before training or not start_data: start data in YYYY-MM-DD format (give a random date if unknown) freq: "H" hourly, "D": daily and for rest see here: https://pandas.pydata.org/pandas-docs/stable/user_guide/timeseries.html#timeseries-offset-aliases covariates: global covariates common for all time series r*T, where r is the number of covariates Ycov: per time-series covariates n*l*T, l such covariates per time-series use_time: if false, default trime-covriates are not used dti: date time object can be explicitly supplied here, leave None if default options are to be used """ self.start_date = start_date if use_time: self.time = TimeCovariates( start_date=start_date, freq=freq, normalized=True, num_ts=Ymat.shape[1] ) if dti is not None: self.time.dti = dti time_covariates = self.time.get_covariates() if covariates is None: self.covariates = time_covariates else: self.covariates = np.vstack([time_covariates, covariates]) else: self.covariates = covariates self.Ycov = Ycov self.freq = freq self.vbsize = vbsize self.hbsize = hbsize self.num_inputs = num_inputs self.num_channels = num_channels self.num_epochs = num_epochs self.lr = lr self.val_len = val_len self.Ymat = Ymat self.test = test self.end_index = end_index self.normalize = normalize self.kernel_size = kernel_size if normalize: Y = Ymat m = np.mean(Y[:, 0 : self.end_index], axis=1) s = np.std(Y[:, 0 : self.end_index], axis=1) # s[s == 0] = 1.0 s += 1.0 Y = (Y - m[:, None]) / s[:, None] mini = np.abs(np.min(Y)) self.Ymat = Y + mini self.m = m self.s = s self.mini = mini if self.Ycov is not None: self.num_inputs += self.Ycov.shape[1] if self.covariates is not None: self.num_inputs += self.covariates.shape[0] self.seq = TemporalConvNet( num_inputs=self.num_inputs, num_channels=num_channels, kernel_size=kernel_size, dropout=dropout, init=True, ) self.seq = self.seq.float() self.D = data_loader( Ymat=self.Ymat, vbsize=vbsize, hbsize=hbsize, end_index=end_index, val_len=val_len, covariates=self.covariates, Ycov=self.Ycov, ) self.val_len = val_len if use_cuda: self.seq = self.seq.cuda() def __loss__(self, out, target, dic=None): criterion = nn.L1Loss() return criterion(out, target) / torch.abs(target.data).mean() def __prediction__(self, data): dic = None out = self.seq(data) return out, dic def train_model(self, early_stop=False, tenacity=10): """ early_stop: set true for using early stop tenacity: patience for early_stop """ print("Training Local Model(Tconv)") if use_cuda: self.seq = self.seq.cuda() optimizer = optim.Adam(params=self.seq.parameters(), lr=self.lr) iter_count = 0 loss_all = [] loss_test_all = [] vae = float("inf") scount = 0 while self.D.epoch < self.num_epochs: last_epoch = self.D.epoch inp, out_target, _, _ = self.D.next_batch() if self.test: inp_test, out_target_test, _, _ = self.D.supply_test() current_epoch = self.D.epoch if use_cuda: inp = inp.cuda() out_target = out_target.cuda() inp = Variable(inp) out_target = Variable(out_target) optimizer.zero_grad() out, dic = self.__prediction__(inp) loss = self.__loss__(out, out_target, dic) iter_count = iter_count + 1 for p in self.seq.parameters(): p.requires_grad = True loss.backward() for p in self.seq.parameters(): p.grad.data.clamp_(max=1e5, min=-1e5) optimizer.step() loss_all = loss_all + [loss.cpu().item()] if self.test: if use_cuda: inp_test = inp_test.cuda() out_target_test = out_target_test.cuda() inp_test = Variable(inp_test) out_target_test = Variable(out_target_test) out_test, dic = self.__prediction__(inp_test) losst = self.__loss__(out_test, out_target_test, dic) loss_test_all = loss_test_all + [losst.cpu().item()] if current_epoch > last_epoch: ve = loss_test_all[-1] print("Entering Epoch# ", current_epoch) print("Train Loss:", np.mean(loss_all)) print("Validation Loss:", ve) if ve <= vae: vae = ve scount = 0 # self.saved_seq = TemporalConvNet( # num_inputs=self.seq.num_inputs, # num_channels=self.seq.num_channels, # kernel_size=self.seq.kernel_size, # dropout=self.seq.dropout, # ) # self.saved_seq.load_state_dict(self.seq.state_dict()) self.saved_seq = pickle.loads(pickle.dumps(self.seq)) else: scount += 1 if scount > tenacity and early_stop: self.seq = self.saved_seq if use_cuda: self.seq = self.seq.cuda() break def convert_to_input(self, data, cuda=True): n, m = data.shape inp = torch.from_numpy(data).view(1, n, m) inp = inp.transpose(0, 1).float() if cuda: inp = inp.cuda() return inp def convert_covariates(self, data, covs, cuda=True): nd, td = data.shape rcovs = np.repeat( covs.reshape(1, covs.shape[0], covs.shape[1]), repeats=nd, axis=0 ) rcovs = torch.from_numpy(rcovs).float() if cuda: rcovs = rcovs.cuda() return rcovs def convert_ycovs(self, data, ycovs, cuda=True): nd, td = data.shape ycovs = torch.from_numpy(ycovs).float() if cuda: ycovs = ycovs.cuda() return ycovs def convert_from_output(self, T): out = T.view(T.size(0), T.size(2)) return np.array(out.cpu().detach()) def predict_future_batch( self, data, covariates=None, ycovs=None, future=10, cpu=False ): if cpu: self.seq = self.seq.cpu() else: self.seq = self.seq.cuda() inp = self.convert_to_input(data) if covariates is not None: cov = self.convert_covariates(data, covariates) inp = torch.cat((inp, cov[:, :, 0 : inp.size(2)]), 1) if ycovs is not None: ycovs = self.convert_ycovs(data, ycovs) inp = torch.cat((inp, ycovs[:, :, 0 : inp.size(2)]), 1) if cpu: inp = inp.cpu() cov = cov.cpu() ycovs = ycovs.cpu() out, dic = self.__prediction__(inp) ci = inp.size(2) output = out[:, :, out.size(2) - 1].view(out.size(0), out.size(1), 1) if covariates is not None: output = torch.cat( (output, cov[:, :, ci].view(cov.size(0), cov.size(1), 1)), 1 ) if ycovs is not None: output = torch.cat( (output, ycovs[:, :, ci].view(ycovs.size(0), ycovs.size(1), 1)), 1 ) out = torch.cat((inp, output), dim=2) for i in range(future - 1): inp = out out, dic = self.__prediction__(inp) output = out[:, :, out.size(2) - 1].view(out.size(0), out.size(1), 1) ci += 1 if covariates is not None: output = torch.cat( (output, cov[:, :, ci].view(cov.size(0), cov.size(1), 1)), 1 ) if ycovs is not None: output = torch.cat( (output, ycovs[:, :, ci].view(ycovs.size(0), ycovs.size(1), 1)), 1 ) out = torch.cat((inp, output), dim=2) out = out[:, 0, :].view(out.size(0), 1, out.size(2)) out = out.cuda() y = self.convert_from_output(out) self.seq = self.seq.cuda() return y def predict_future( self, data_in, covariates=None, ycovs=None, future=10, cpu=False, bsize=40, normalize=False, ): """ data_in: input past data in same format of Ymat covariates: input past covariates ycovs: input past individual covariates future: number of time-points to predict cpu: if true then gpu is not used bsize: batch size for processing (determine according to gopu memory limits) normalize: should be set according to the normalization used in the class initialization """ if normalize: data = (data_in - self.m[:, None]) / self.s[:, None] data += self.mini else: data = data_in n, T = data.shape I = list(np.arange(0, n, bsize)) I.append(n) bdata = data[range(I[0], I[1]), :] if ycovs is not None: out = self.predict_future_batch( bdata, covariates, ycovs[range(I[0], I[1]), :, :], future, cpu ) else: out = self.predict_future_batch(bdata, covariates, None, future, cpu) for i in range(1, len(I) - 1): bdata = data[range(I[i], I[i + 1]), :] self.seq = self.seq.cuda() if ycovs is not None: temp = self.predict_future_batch( bdata, covariates, ycovs[range(I[i], I[i + 1]), :, :], future, cpu ) else: temp = self.predict_future_batch(bdata, covariates, None, future, cpu) out = np.vstack([out, temp]) if normalize: temp = (out - self.mini) * self.s[:, None] + self.m[:, None] out = temp return out def rolling_validation(self, Ymat, tau=24, n=7, bsize=90, cpu=False, alpha=0.3): last_step = Ymat.shape[1] - tau * n rg = 1 + 2 * (self.kernel_size - 1) * 2 ** (len(self.num_channels) - 1) self.seq = self.seq.eval() if self.covariates is not None: covs = self.covariates[:, last_step - rg : last_step + tau] else: covs = None if self.Ycov is not None: ycovs = self.Ycov[:, :, last_step - rg : last_step + tau] else: ycovs = None data_in = Ymat[:, last_step - rg : last_step] out = self.predict_future( data_in, covariates=covs, ycovs=ycovs, future=tau, cpu=cpu, bsize=bsize, normalize=self.normalize, ) predicted_values = [] actual_values = [] S = out[:, -tau::] predicted_values += [S] R = Ymat[:, last_step : last_step + tau] actual_values += [R] print("Current window wape: " + str(wape(S, R))) for i in range(n - 1): last_step += tau rg = 1 + 2 * (self.kernel_size - 1) * 2 ** (len(self.num_channels) - 1) if self.covariates is not None: covs = self.covariates[:, last_step - rg : last_step + tau] else: covs = None if self.Ycov is not None: ycovs = self.Ycov[:, :, last_step - rg : last_step + tau] else: ycovs = None data_in = Ymat[:, last_step - rg : last_step] out = self.predict_future( data_in, covariates=covs, ycovs=ycovs, future=tau, cpu=cpu, bsize=bsize, normalize=self.normalize, ) S = out[:, -tau::] predicted_values += [S] R = Ymat[:, last_step : last_step + tau] actual_values += [R] print("Current window wape: " + str(wape(S, R))) predicted = np.hstack(predicted_values) actual = np.hstack(actual_values) dic = {} dic["wape"] = wape(predicted, actual) dic["mape"] = mape(predicted, actual) dic["smape"] = smape(predicted, actual) dic["mae"] = np.abs(predicted - actual).mean() dic["rmse"] = np.sqrt(((predicted - actual) ** 2).mean()) dic["nrmse"] = dic["rmse"] / np.sqrt(((actual) ** 2).mean()) baseline = Ymat[:, Ymat.shape[1] - n * tau - tau : Ymat.shape[1] - tau] dic["baseline_wape"] = wape(baseline, actual) dic["baseline_mape"] = mape(baseline, actual) dic["baseline_smape"] = smape(baseline, actual) return dic
21,683
31.804841
157
py
deepglo
deepglo-master/run_scripts/run_traffic.py
#### OS and commanline arguments import sys import multiprocessing as mp import gzip import subprocess from pathlib import Path import argparse import logging import os sys.path.append('./') #### DeepGLO model imports from DeepGLO.metrics import * from DeepGLO.DeepGLO import * from DeepGLO.LocalModel import * import pandas as pd import numpy as np import pickle import random np.random.seed(111) torch.cuda.manual_seed(111) torch.manual_seed(111) random.seed(111) import json def str2bool(v): if isinstance(v, bool): return v if v.lower() in ("yes", "true", "t", "y", "1"): return True elif v.lower() in ("no", "false", "f", "n", "0"): return False else: raise argparse.ArgumentTypeError("Boolean value expected.") def bool2str(b): if b: return "true" else: return "false" Ymat = np.load("./datasets/traffic.npy") vbsize = 128 ## vertical batch size hbsize = 256 ## horizontal batch size num_channels_X = [32, 32, 32, 32, 32, 1] ## number of channels for local model num_channels_Y = [32, 32, 32, 32, 32, 1] ## number of channels for hybrid model kernel_size = 7 ## kernel size for local models dropout = 0.2 ## dropout during training rank = 64 ## rank of global model kernel_size_Y = 7 ## kernel size of hybrid model lr = 0.0005 ## learning rate val_len = 24 ## validation length end_index = Ymat.shape[1] - 24 * 7 ## models will not look beyond this during training start_date = "2012-1-1" ## start date time for the time-series freq = "H" ## frequency of data covariates = None ## no covraites specified use_time = True ## us time covariates dti = None ## no spcified time covariates (using default) svd = False ## factor matrices are not initialized by NMF period = None ## periodicity of 24 is expected, leave it out if not known y_iters = 300 ## max. number of iterations while training Tconv models init_epochs = 100 ## max number of iterations while initialiozing factors forward_cov = False logging.basicConfig( stream=sys.stdout, format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S", level=logging.INFO, ) logger = logging.getLogger(__name__) def main(args): DG = DeepGLO( Ymat, vbsize=vbsize, hbsize=hbsize, num_channels_X=num_channels_X, num_channels_Y=num_channels_Y, kernel_size=kernel_size, dropout=dropout, rank=rank, kernel_size_Y=kernel_size_Y, lr=lr, val_len=val_len, end_index=end_index, normalize=normalize, start_date=start_date, freq=freq, covariates=covariates, use_time=use_time, dti=dti, svd=svd, period=period, forward_cov=forward_cov, ) DG.train_all_models(y_iters=y_iters, init_epochs=init_epochs) result_dic = DG.rolling_validation( Ymat=Ymat, tau=24, n=7, bsize=100, cpu=False, alpha=0.3 ) print(result_dic) out_path = Path( ".", "results", "result_dictionary_traffic_" + bool2str(normalize) + ".pkl", ) pickle.dump(result_dic, open(out_path, "wb")) if __name__ == "__main__": parser = argparse.ArgumentParser() parser.add_argument( "--normalize", type=str2bool, required=True, help="normalize for training or not", ) args = parser.parse_args() global normalize normalize = args.normalize main(args)
3,498
24.727941
87
py
deepglo
deepglo-master/run_scripts/run_wiki.py
#### OS and commanline arguments import sys import multiprocessing as mp import gzip import subprocess from pathlib import Path import argparse import logging import os sys.path.append('./') #### DeepGLO model imports from DeepGLO.metrics import * from DeepGLO.DeepGLO import * from DeepGLO.LocalModel import * import pandas as pd import numpy as np import pickle import random np.random.seed(111) torch.cuda.manual_seed(111) torch.manual_seed(111) random.seed(111) import json def str2bool(v): if isinstance(v, bool): return v if v.lower() in ("yes", "true", "t", "y", "1"): return True elif v.lower() in ("no", "false", "f", "n", "0"): return False else: raise argparse.ArgumentTypeError("Boolean value expected.") def bool2str(b): if b: return "true" else: return "false" Ymat = np.load("./datasets/wiki.npy") vbsize = 2048 ## vertical batch size hbsize = 256 ## horizontal batch size num_channels_X = [32, 32, 32, 32, 1] ## number of channels for local model num_channels_Y = [32, 32, 32, 32, 1] ## number of channels for hybrid model kernel_size = 7 ## kernel size for local models dropout = 0.2 ## dropout during training rank = 128 ## rank of global model kernel_size_Y = 7 ## kernel size of hybrid model lr = 0.0005 ## learning rate val_len = 14 ## validation length end_index = Ymat.shape[1] - 14 * 4 ## models will not look beyond this during training start_date = "2012-1-1" ## start date time for the time-series freq = "D" ## frequency of data covariates = None ## no covraites specified use_time = True ## us time covariates dti = None ## no specified time covariates (using default) svd = True ## factor matrices are initialized by NMF period = 7 ## periodicity of 7 is expected, leave it out if not known y_iters = 300 ## max. number of iterations while training Tconv models init_epochs = 100 ## max number of iterations while initializing factors forward_cov = False logging.basicConfig( stream=sys.stdout, format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S", level=logging.INFO, ) logger = logging.getLogger(__name__) def main(args): DG = DeepGLO( Ymat, vbsize=vbsize, hbsize=hbsize, num_channels_X=num_channels_X, num_channels_Y=num_channels_Y, kernel_size=kernel_size, dropout=dropout, rank=rank, kernel_size_Y=kernel_size_Y, lr=lr, val_len=val_len, end_index=end_index, normalize=normalize, start_date=start_date, freq=freq, covariates=covariates, use_time=use_time, dti=dti, svd=svd, period=period, forward_cov=forward_cov, ) DG.train_all_models(y_iters=y_iters, init_epochs=init_epochs) result_dic = DG.rolling_validation( Ymat=Ymat, tau=14, n=4, bsize=100, cpu=False, alpha=0.5 ) print(result_dic) out_path = Path( ".", "results", "result_dictionary_wiki_" + bool2str(normalize) + ".pkl", ) pickle.dump(result_dic, open(out_path, "wb")) if __name__ == "__main__": parser = argparse.ArgumentParser() parser.add_argument( "--normalize", type=str2bool, required=True, help="normalize for training or not", ) args = parser.parse_args() global normalize normalize = args.normalize main(args)
3,475
24.940299
87
py