Spaces:
Running on Zero
Running on Zero
File size: 6,451 Bytes
d43892c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 | import random
import numpy as np
from PIL import Image
from PIL import ImageFilter
import torchvision.transforms.functional as TF
from torchvision import transforms
import torch
def to_tensor_and_norm(imgs, labels):
# to tensor
imgs = [TF.to_tensor(img) for img in imgs]
labels = [torch.from_numpy(np.array(img, np.uint8)).unsqueeze(dim=0)
for img in labels]
imgs = [TF.normalize(img, mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
for img in imgs]
return imgs, labels
class CDDataAugmentation:
def __init__(
self,
img_size,
with_random_hflip=False,
with_random_vflip=False,
with_random_rot=False,
with_random_crop=False,
with_scale_random_crop=False,
with_random_blur=False,
):
self.img_size = img_size
if self.img_size is None:
self.img_size_dynamic = True
else:
self.img_size_dynamic = False
self.with_random_hflip = with_random_hflip
self.with_random_vflip = with_random_vflip
self.with_random_rot = with_random_rot
self.with_random_crop = with_random_crop
self.with_scale_random_crop = with_scale_random_crop
self.with_random_blur = with_random_blur
def transform(self, imgs, labels, to_tensor=True):
"""
:param imgs: [ndarray,]
:param labels: [ndarray,]
:return: [ndarray,],[ndarray,]
"""
# resize image and covert to tensor
imgs = [TF.to_pil_image(img) for img in imgs]
if self.img_size is None:
self.img_size = None
if not self.img_size_dynamic:
if imgs[0].size != (self.img_size, self.img_size):
imgs = [TF.resize(img, [self.img_size, self.img_size], interpolation=3)
for img in imgs]
else:
self.img_size = imgs[0].size[0]
labels = [TF.to_pil_image(img) for img in labels]
if len(labels) != 0:
if labels[0].size != (self.img_size, self.img_size):
labels = [TF.resize(img, [self.img_size, self.img_size], interpolation=0)
for img in labels]
random_base = 0.5
if self.with_random_hflip and random.random() > 0.5:
imgs = [TF.hflip(img) for img in imgs]
labels = [TF.hflip(img) for img in labels]
if self.with_random_vflip and random.random() > 0.5:
imgs = [TF.vflip(img) for img in imgs]
labels = [TF.vflip(img) for img in labels]
if self.with_random_rot and random.random() > random_base:
angles = [90, 180, 270]
index = random.randint(0, 2)
angle = angles[index]
imgs = [TF.rotate(img, angle) for img in imgs]
labels = [TF.rotate(img, angle) for img in labels]
if self.with_random_crop and random.random() > 0:
i, j, h, w = transforms.RandomResizedCrop(size=self.img_size). \
get_params(img=imgs[0], scale=(0.8, 1.0), ratio=(1, 1))
imgs = [TF.resized_crop(img, i, j, h, w,
size=(self.img_size, self.img_size),
interpolation=Image.CUBIC)
for img in imgs]
labels = [TF.resized_crop(img, i, j, h, w,
size=(self.img_size, self.img_size),
interpolation=Image.NEAREST)
for img in labels]
if self.with_scale_random_crop:
# rescale
scale_range = [1, 1.2]
target_scale = scale_range[0] + random.random() * (scale_range[1] - scale_range[0])
imgs = [pil_rescale(img, target_scale, order=3) for img in imgs]
labels = [pil_rescale(img, target_scale, order=0) for img in labels]
# crop
imgsize = imgs[0].size # h, w
box = get_random_crop_box(imgsize=imgsize, cropsize=self.img_size)
imgs = [pil_crop(img, box, cropsize=self.img_size, default_value=0)
for img in imgs]
labels = [pil_crop(img, box, cropsize=self.img_size, default_value=255)
for img in labels]
if self.with_random_blur and random.random() > 0:
radius = random.random()
imgs = [img.filter(ImageFilter.GaussianBlur(radius=radius))
for img in imgs]
if to_tensor:
# to tensor
imgs = [TF.to_tensor(img) for img in imgs]
labels = [torch.from_numpy(np.array(img, np.uint8)).unsqueeze(dim=0)
for img in labels]
imgs = [TF.normalize(img, mean=[0.5, 0.5, 0.5],std=[0.5, 0.5, 0.5])
for img in imgs]
return imgs, labels
def pil_crop(image, box, cropsize, default_value):
assert isinstance(image, Image.Image)
img = np.array(image)
if len(img.shape) == 3:
cont = np.ones((cropsize, cropsize, img.shape[2]), img.dtype)*default_value
else:
cont = np.ones((cropsize, cropsize), img.dtype)*default_value
cont[box[0]:box[1], box[2]:box[3]] = img[box[4]:box[5], box[6]:box[7]]
return Image.fromarray(cont)
def get_random_crop_box(imgsize, cropsize):
h, w = imgsize
ch = min(cropsize, h)
cw = min(cropsize, w)
w_space = w - cropsize
h_space = h - cropsize
if w_space > 0:
cont_left = 0
img_left = random.randrange(w_space + 1)
else:
cont_left = random.randrange(-w_space + 1)
img_left = 0
if h_space > 0:
cont_top = 0
img_top = random.randrange(h_space + 1)
else:
cont_top = random.randrange(-h_space + 1)
img_top = 0
return cont_top, cont_top+ch, cont_left, cont_left+cw, img_top, img_top+ch, img_left, img_left+cw
def pil_rescale(img, scale, order):
assert isinstance(img, Image.Image)
height, width = img.size
target_size = (int(np.round(height*scale)), int(np.round(width*scale)))
return pil_resize(img, target_size, order)
def pil_resize(img, size, order):
assert isinstance(img, Image.Image)
if size[0] == img.size[0] and size[1] == img.size[1]:
return img
if order == 3:
resample = Image.BICUBIC
elif order == 0:
resample = Image.NEAREST
return img.resize(size[::-1], resample)
|