repo_name
stringlengths
6
97
path
stringlengths
3
341
text
stringlengths
8
1.02M
dutitello/parAnsys
examples/AngTang_MCAI_67.py
<gh_stars>1-10 """ Running example 6.7 from Ang & Tang 1984 """ import paransys import numpy as np # Call ParAnsys mc = paransys.MonteCarlo() # Console log On mc.Info(True) # Create random variables mc.CreateVar('y', 'gauss', 40, cv=0.125) mc.CreateVar('z', 'gauss', 50, cv=0.050) mc.CreateVar('m', 'gauss', 1000, std=0.200*1000) # First limit state (0) mc.CreateLimState('y*z-m') # Sampling for first limit state (0) # It create failures before =) k = 2 # For GAUSS and GUMBEL is better to use STD and for LOGN CV ;) mc.SetRandomVarSampl('y', 0, 'gauss', 40*(1-k*0.125), std=0.125*40) mc.SetRandomVarSampl('z', 0, 'gauss', 50*(1-k*0.050), std=0.050*50) mc.SetRandomVarSampl('m', 0, 'gauss', 1000*(1+k*0.200), std=0.200*1000) # Running values = mc.Run(100, 1000, 0.05, 0.005) # Export mc.ExportDataCSV('AngTang-MCAI-67', 'Comentarios?') # Figures mc.Graph(['N_Pf', 'N_Beta', 'N_CVPf'], show=True)
dutitello/parAnsys
paransys/ansys.py
<filename>paransys/ansys.py<gh_stars>1-10 # -*- coding: UTF-8 -*- """ This module calls ANSYS by Probabilistic Design System (PDS) as an FEM tool for running a couple of parametric analysis and get some variables back to Python. Built using ANSYS documentation from https://www.sharcnet.ca/Software/Ansys/ (not available anymore) This module makes no claim to own any rights to ANSYS, it's just an interface to call the program owned by ANSYS. Docs are available in https://dutitello.github.io/parAnsys/ """ import os import numpy as np import time class ANSYS(object): """ This class creates a couple of script files with some parameters defined here by the user, copy an APDL script file created by the user with the model to be analysed, run everything on ANSYS and get the results from some defined parameters back to Python. exec_loc : str, obligatory Location of ANSYS executable file. run_location : str, optional ANSYS working directory. Must be a separated directory. Defaults to ansys_anl inside current directory. jobname : str, optional ANSYS jobname. Defaults to \'file\'. nproc : int, optional Number of processors. Defaults to 2. override : bool, optional Attempts to delete the .lock file at working directory. It's useful when ANSYS was interrupted. Defaults to False cleardir : bool, optional Delete all the files from ANSYS working directory when call the Run command. Defaults to False add_flags : str, optional Additional flags to be called with ANSYS. If it's an academic version use add_flags='-aa_r' Do not use '-b -i -o'. Flags can be found at https://www.sharcnet.ca/Software/Ansys/16.2.3/en-us/help/ans_ope/Hlp_G_OPE3_1.html | | **Class methods:** """ def __init__(self, exec_loc=None, run_location=os.getcwd()+'\\ansys_anl\\', jobname='file', nproc=2, override=False, cleardir=False, add_flags=''): """ """ # Verify if the exec_loc is defined and real if exec_loc == None: exception = Exception('Undefined ANSYS executable location. \n You must define ANSYS executable location.') raise exception else: # Verify if the file exists if not os.path.isfile(exec_loc): exception = Exception('Invalid ANSYS executable file as \"%s\".' % exec_loc) raise exception # Save the options to the main object self.ANSYSprops = {} self.ANSYSprops['exec_loc'] = exec_loc self.ANSYSprops['run_location'] = run_location self.ANSYSprops['jobname'] = jobname self.ANSYSprops['nproc'] = nproc self.ANSYSprops['override'] = override self.ANSYSprops['cleardir'] = cleardir self.ANSYSprops['add_flags'] = add_flags # Create lists and set initial values # Variables names self.varInNames = [] self.varOutNames = [] # Variables values self.varInValues = {} self.varOutValues = {} # Analysis length self.length = 0 # Model properties self.Model = {} # Defalts to print always self.PrintR = True self._PrintR('ANSYS properties defined as:') self._PrintR(' Executable file: \"%s\".' % self.ANSYSprops['exec_loc']) self._PrintR(' Working directory: \"%s\".' % self.ANSYSprops['run_location']) self._PrintR(' Jobname: \"%s\".' % self.ANSYSprops['jobname']) self._PrintR(' Number of processors used: \"%s\".' % self.ANSYSprops['nproc']) self._PrintR(' Override lock file: \"%s\".' % self.ANSYSprops['override']) self._PrintR(' Clear working directory: \"%s\".' % self.ANSYSprops['cleardir']) self._PrintR(' Additional flags: \"%s\".' % self.ANSYSprops['add_flags']) def Info(self, act=False): """ Turn on/off the return of the commands to Python. Parameters ---------- act : bool, obligatory True turn On and False turn Off the return of the commands to Python. """ self.PrintR = act self._PrintR('Now the commands will send a return to Python (like this).') def _PrintR(self, value): """ Internal function to print or not the return of commands based on command Info() """ if self.PrintR: print(value, flush=True) else: pass def _writePDSfile(self): """ Internal function to create/write the APDL file with the PDS analysis. (pdsrun.inp) """ # Verify if self.length >0 if self.length <= 0: exception = Exception('Before run you must define the length of the analysis with SetLength()'+ 'and set the all the variables values.') raise exception # Run all the varInValues[var] looking for an empty for each in self.varInNames: if self.varInValues[each] == None: exception = Exception('Input variable \"%s\" has no defined values.' % each) raise exception # Try to open the file for writing try: runfile = '%s\\pdsrun.inp' % self.ANSYSprops['run_location'] f = open(runfile, 'wt') except: exception = Exception('Unable to open the pdsrun.inp file for writting.') raise exception # Now try to write try: self._PrintR('Writing the pdsrun.inp file.') # some ANSYS initial commands f.write('FINISH\n') f.write('/GOPR\n') f.write('KEYW,PR_SET,1\n') f.write('KEYW,PR_STRUC,1\n') # Open PDS module f.write('/PDS\n') # Clear the PDS database f.write('PDCLR,ALL\n') # Here we put the from SetModel(..) that is named here as "current.inp" f.write('PDANL,current,inp\n') # Input variables for each in self.varInNames: # Variables are declared with uniform distribution betwen 1.5*maxval and 0.5*minval cmd = 'PDVAR,%s,UNIF,%f,%f\n' % (each, min(0.5*min(self.varInValues[each]),-1), max(1,1.5*max(self.varInValues[each]))) f.write(cmd) # Output/control variables for each in self.varOutNames: # Just the varname and distrib=RESP cmd = 'PDVAR,%s,RESP\n' % (each) f.write(cmd) # We will use Monte Carlo simulation with user sampling, that way we can # simulate the points declared in each variable easily and without problems. f.write('PDMETH,MCS,USER\n') # The sample points file is named here as "current.samp" f.write('PDUSER,current,samp\n') # Runs the Analysis - The analysis is named as current so the results file # will be $jobname%_current.pdrs f.write('PDEXE,current\n') # Create a delay of 2s on ANSYS just to make sure that all were done f.write('/WAIT,2\n') # Close the file self._PrintR('Saving the pdsrun.inp file.') f.close() except: exception = Exception('Error while writting the pdsrun.inp file.') raise exception else: pass def _writeSAMPfile(self): """ Internal function to create/write the sample points file to run the PDS. (current.samp) The sample file format has the PDEXE name in the first line, and the second line is: ITER CYCL LOOP $INPUT VAR NAMES$ With a new column for each variable. The values for ITER and CYCL are always 1, LOOP is the number of solution range(0 to length) """ # self._PrintR('Writing the current.samp file.') # Create initial the header string header = 'current\nITER CYCL LOOP' # Create the initial format string format = '%d %d %d' # Create initial values array with ones samparray = np.zeros([self.Length, 3+len(self.varInNames)]) + 1 # Fill the loop column samparray[:, 2] = range(1, self.Length+1) # Each variable will add his name at header and a %15.8E in the format # A counter from 2 (3rd column) i = 2 for each in self.varInNames: # Add strings header = '%s %s' % (header, each) format = '%s %s' % (format, '%15.8E') # Add one in counter i += 1 # Place values try: samparray[:, i] = self.varInValues[each] except: exception = Exception('Error while passing the values of \"%s\" to the current.samp file.' % each) raise exception # Save the file try: sampfile = '%s\\current.samp' % self.ANSYSprops['run_location'] np.savetxt(sampfile, samparray, delimiter=' ', newline='\n', header=header, comments='', fmt=format) except: exception = Exception('Error while passing the values of \"%s\" to the current.samp file.' % each) raise exception pass def SetModel(self, inputname, extrafiles=[], directory=os.getcwd()): """ Set the input script file to be used and extra files that should be copied together. All this files must be in the same directory set in parameter directory. Parameters ---------- inputname : str, obligatory Name with extension of the script that will be executed in the analysis. The script must be done in function of the INPUT variables defined here, (as parameters of ANSYS), and must define/calculate ANSYS parameters with the results using the names defined here. extrafiles : list of strings, optional A list of strings containing extra files (with extension) that are necessary to run the script analys, could be an MODEL with the MESH already generated, for example. An example of extrafiles list is: extrafiles = ['temps.txt', 'model1.ans', 'file.db'] directory : str, optional If the script is not in the current running Python directory you should place the entire location, if it's in a subdirectory of current directory you can use '/dirname/filename.ext'. Defaults to current running Python directory. """ # Verify if the input script file exists if not os.path.isfile(str(directory+'\\'+inputname)): exception = Exception('Current input script file (\"%s\") does not exists in (\"%s\") to be copied.' % (inputname, directory)) raise exception # Copy the script file to workdirectory with the name 'current.inp' errcopy = os.system('copy /Y %s %s' % (str(directory+'\\'+inputname), str(self.ANSYSprops['run_location']+'\\current.inp'))) if errcopy != 0: exception = Exception('It was not possible to copy the input script file. (\"%s\").' % str(directory+'\\'+inputname)) raise exception # Copy the extra files for each in extrafiles: errcopy = os.system('copy /Y %s %s' % (str(directory+'\\'+each), str(self.ANSYSprops['run_location']+'\\'+each))) if errcopy != 0: exception = Exception('It was not possible to copy an extra file (\"%s\").' % each) raise exception # Clear last Model properties and set new self.Model = {} self.Model['inputname'] = inputname self.Model['extrafiles'] = extrafiles self.Model['directory'] = directory self._PrintR('Input script file and extra files copied to working directory.') self._PrintR(' Main APDL script: \"%s\".' % self.Model['inputname']) self._PrintR(' Extra model files: \"%s\".' % self.Model['extrafiles']) self._PrintR(' Input directory: \"%s\".' % self.Model['directory']) def _ClearForRun(self): """ Internal function to clear the entire directory or just the lock file before running """ # Verify the clear condition if self.ANSYSprops['cleardir']: # Try to clear the working directory self._PrintR('Cleaning the files from ANSYS working directory (\"%s\").' % self.ANSYSprops['run_location']) delhand = os.system('del /q %s\\*' % self.ANSYSprops['run_location']) if delhand != 0: exception = Exception('Unable to clear the ANSYS working directory.') raise exception # Put the model back in the working directory self.SetModel(self.Model['inputname'], self.Model['extrafiles'], self.Model['directory']) else: # Verify the override condition lockfile = '%s\\%s.lock' % (self.ANSYSprops['run_location'], self.ANSYSprops['jobname']) if self.ANSYSprops['override'] and os.path.isfile(lockfile): self._PrintR('Deleting lock file.') delhand = os.system('del /q %s' % lockfile) if delhand != 0: exception = Exception('Unable to delete lock file (\"%s\").' % lockfile) raise exception # Before execution ALWAYS erase the $jobname$.err file self._PrintR('Deleting old error log file.') errfile = '%s\\%s.err' % (self.ANSYSprops['run_location'], self.ANSYSprops['jobname']) delhand = os.system('del /q %s' % errfile) pass def _Run(self): # ANSYS Run Parameters https://www.sharcnet.ca/Software/Ansys/16.2.3/en-us/help/ans_ope/Hlp_G_OPE3_1.html # Ansys PDS commands = pdsrun.inp and out=pdsout.out # # DO NOT REMOVE THE SPACES BEFORE -np ! # cmd = 'start "ANSYS" /d "%s" /min /wait /b "%s" " -smp -np %d -j %s -b -i pdsrun.inp -o pdsout.out %s" ' % (self.ANSYSprops['run_location'], self.ANSYSprops['exec_loc'], self.ANSYSprops['nproc'], self.ANSYSprops['jobname'], self.ANSYSprops['add_flags']) self._PrintR('Running ANSYS.') # Get time before run ANSYS timei = time.time() ansyshand = os.system(cmd) if ansyshand != 0: exception = Exception('ANSYS exited with error id=%d. Please verify the output file (\"%s\").\n\n' %(ansyshand, self.ANSYSprops['run_location']+'\\pdsout.out')) raise exception # verify if it has an error on $jobname$.err #errfile = '%s\\%s.err' % (self.ANSYSprops['run_location'], self.ANSYSprops['jobname']) #f = open(errfile).read() #if '*** ERROR ***' in f: # exception = Exception('ANSYS exited with an error. Please verify the output file (%s).\n\n' % (self.ANSYSprops['run_location']+'\\pdsout.out')) # raise exception # Time after ANSYS timef = time.time() self._PrintR('Solution is done. It took %f minutes.' % ((timef-timei)/60)) def Run(self): """ Execute the analysis on ANSYS. """ # Verify if model was set if self.Model == {}: exception = Exception('Before running ANSYS you must define the model that will be analysed with SetModel().') raise exception # Verify if there are output variables if self.varOutNames == []: exception = Exception('Before running ANSYS you must define output variables.') raise exception # Performn the conditional clear self._ClearForRun() # Write the pdsrun.inp file self._writePDSfile() # Write the sample file self._writeSAMPfile() # Run self._Run() @property def Length(self): """ Return the number of analysis to be executed and the length of INPUT arrays. """ return self.length def SetLength(self, length): """ Define the number of analysis to be executed and the length of INPUT arrays. Must be set before the variables. Parameters ---------- length: int, obligatory Number of analysis. """ # control variable valuesset = False # if at least one InValue is set can't change de length for each in self.varInNames: if self.varInValues[each] != None: valuesset = True exception = Exception('At least one value were already set to INPUT variables. \n'+ 'To change the length you have to clear the variables values with ClearValues().') raise exception # if there is no value setted and length > 0 can change if valuesset == False: if length <= 0: exception = Exception('The analysis length must be greater than 0.') raise exception else: self.length = length self._PrintR('Analysis lenght set to %d.' % self.length) def CreateVarIn(self, name): """ Create an INPUT variable. Parameters ---------- name : str, obligatory Variable name. Do not use spaces in the name! The same name cannot be used with INPUT and OUTPUT variables. It's not case sensitivy, in fact it will be saved in uppercase because of ANSYS. """ if name in ['', None, ' ']: exception = Exception('You must define an unique name to each variable.') raise exception else: name = name.upper() if name in self.varInNames or name in self.varOutNames: exception = Exception('The same name cannot be used with INPUT and OUTPUT variables.') raise exception else: self.varInNames.append(name) self.varInValues[name] = None self._PrintR('ANSYS input variable \"%s\" created.' % name) def CreateVarOut(self, name): """ Create an OUTPUT variable. Parameters ---------- name : str, obligatory Variable name. Do not use spaces in the name! The same name cannot be used with INPUT and OUTPUT variables. It's not case sensitivy, in fact it will be saved in uppercase because of ANSYS. """ if name in ['', None, ' ']: exception = Exception('You must define an unique name to each variable.') raise exception else: name = name.upper() if name in self.varInNames or name in self.varOutNames: exception = Exception('The same name cannot be used with INPUT and OUTPUT variables.') raise exception else: self.varOutNames.append(name) self.varOutValues[name] = None self._PrintR('Variable \"%s\" declared as ANSYS output variable.' % name) def SetVarInValues(self, name, values): """ Set the values of an INPUT variable Parameters ---------- name : str, obligatory Input variable name that will receive the values. values : 1D np.array of floats, obligatory A 1D numpy.array() with the length of this class and the values to be analysed. If the array is not 1D just the first column (0) will be used. """ # Verify the set length if self.length <= 0: exception = Exception('Before set the values you must define the length of the analysis with SetLength().') raise exception # Verify the name if name in ['', None, ' ']: exception = Exception('You must define the name of the variable that will receive the values.') raise exception else: name = name.upper() if name not in self.varInNames: exception = Exception('This variable name is not declared.\n'+ 'Please use CreateVarIn("name") to declare it.') raise exception # Verify if it's a list, in this case transf to an np.array if type(values) == list: values = np.array(values) # Verify if it has more than one column try: values.shape[1] except: pass else: self._PrintR('The values that are trying to be set to \"%s\" has more than one column, just the first (0) will be used.' % (name)) values = np.copy(values[:,0]) # Verify if length is GT or LT the expected value if values.shape[0] < self.length: exception = Exception('The length of values that are trying to be set to \"%s\" is less than the defined length (%d).' % (name, self.length)) raise exception elif values.shape[0] > self.length: self._PrintR('The length of values that are trying to be set to \"%s\" is greater than the defined length (%d).\ \nJust the first %dth values will be used.' % (name, self.length, self.length)) values = values[:self.length] # Apply try: self.varInValues[name] = values except: exception = Exception('Error setting the values of \"%s\".' % name) raise exception else: self._PrintR('Values of \"%s\" were set.' % name) def GetVarOutValues(self): """ Return the values of the Results file from ANSYS for all the OUTPUT variables """ # Verify the existence of results file "$jobname$_current.pdrs" resultsfile = '%s\\%s_current.pdrs' % (self.ANSYSprops['run_location'], self.ANSYSprops['jobname']) if os.path.isfile(resultsfile): # Import the results file self._PrintR('Importing results from PDS results file (\"%s\").' % resultsfile) resultsALL = np.genfromtxt(resultsfile, names=True, skip_header=1) # Verify the existence of ERRORS errorcount = resultsALL['ERR'].sum() if errorcount > 0: self._PrintR('\n\n\nATTENTION:\nThe 4th column in the results file called ERR warns about simulations that results should not be used.\n'+ 'Current file has %d errors, you can acces this in the column ERR returned with the results.\n\n\n' % errorcount) _ = input('The execution is paused, press ENTER to continue.\n') # Get all the columns in varOutNames and varInNames the specified length results = {} #results['ERR'] = np.copy(resultsALL['ERR'][:self.length]) results['ERR'] = np.copy(resultsALL['ERR']) #for each in self.varInNames: # results[each] = np.copy(resultsALL[each]) for each in self.varOutNames: # This prevent bugs with 1 line results results[each] = np.atleast_1d(resultsALL[each]) for each in results: if results[each].size > self.length: results[each] = results[each][:self.length] return results else: # There is no results exception = Exception('There is no results file in current ANSYS working directory. \n'+ 'Please verify if the analysis was run.') raise exception def ClearAll(self): """ Clear all the properties (not from ANSYS object) """ # Clear all self.Model = {} self.varInNames = [] self.varOutNames = [] self.varInValues = {} self.varOutValues = {} self.length = 0 #self.PrintR = False self._PrintR('All the properties were cleared (not from ANSYS object).') def ClearValues(self): """ Clear the values of all variables. """ # Clear each input variable value for each in self.varInNames: self.varInValues[each] = None # Clear each output variable value for each in self.varOutNames: self.varOutValues[each] = None self._PrintR('The values of all variables were cleared. Now you can change the length parameter.')
dutitello/parAnsys
examples/AngTang_FORM_68.py
""" Running example 6.8 from Ang & Tang 1984 """ import paransys # Call ParAnsys form = paransys.FORM() # Console log On form.Info(True) # Create random variables form.CreateVar('y', 'logn', 40, cv=0.125) form.CreateVar('z', 'logn', 50, cv=0.050) form.CreateVar('m', 'gumbel', 1000, cv=0.200) # Create limit state form.SetLimState('y*z-m') # Run values = form.Run(dh=0.01, meth='iHLRF') form.ExportDataCSV('AngTang-FORM-68', 'Comentarios?')
dutitello/parAnsys
paransys/form.py
<gh_stars>1-10 ## -*- coding: UTF-8 -*- """ This module performns Reliability Analysis with First Order Reliability Method using Python. Please read the class docstring for more. Docs are available at https://dutitello.github.io/parAnsys/ """ import os import time from paransys.ansys import ANSYS import numpy as np import scipy.stats from math import * # It's necessry to evaluate limit states import math from inspect import isfunction class FORM(object): """ This class applies the First Order Reliability Method (FORM) inside Python using ParAnsys as a connection with ANSYS for evaluate FEM models. It is possible to run simulations without using ANSYS, just defining the limit state equation and all variables. This code was made following the ideia of ANSYS being a tool for getting the ultimate load of the structure. This works applying a displacement in the loaded node, and then getting the biggest reaction force on that node, following this way the limit state defined here is 'R-S', where R are the values get from ANSYS and S the values generated in Python. It's also possible to work applying the true load on ANSYS, it's just necessary to formulate a valid limit state equation. | | **Class methods:** """ #--------------------------------------------------------------------------- def __init__(self): """ """ # Before self._ANSYS = False self.PrintR = True self.limstate = None self._userf = None self.variableDistrib = {} self.variableConst = {} self.variableStartPt = {} self.controls = {} self.corlist = {} self._last_ck = 0 # Options self._options = {} self._options['iHLRF_forced_lambdk'] = 'auto' self._options['iHLRF_prod_ck'] = 2.00 self._options['iHLRF_add_ck'] = 0.00 self._options['iHLRF_par_a'] = 0.10 self._options['iHLRF_par_b'] = 0.50 self._options['iHLRF_step_lambdk_test'] = 4 self._options['rHLRF_relax'] = 0.50 self._options['APDLdebug'] = False # After self.variableDesPt = {} self.results = {} self._stnumb = 99 def ANSYS(self, exec_loc=None, run_location=os.getcwd()+'\\ansys_anl\\', jobname='file', nproc=2, override=False, cleardir=False, add_flags=''): """ If ANSYS will be used it defines ANSYS properties, for initialize the paransys.ANSYS class. Parameters ---------- exec_loc : str, obligatory Location of ANSYS executable file. run_location : str, optional ANSYS working directory. Recomended to be a separated directory. Defaults to ansys_anl on current directory. jobname : str, optional ANSYS jobname. Defaults to 'file'. nproc : int, optional Number of processors. Defaults to 2. override : bool, optional Attempts to delete the .lock file at working directory. It's useful when ANSYS was interrupted. Defaults to False cleardir : bool, optional Delete all the files from ANSYS working directory when call the Run command. Defaults to False add_flags : str, optional Additional flags to be called with ANSYS. If it's an academic version use add_flags='-aa_r' Do not use '-b -i -o' Flags can be found at https://www.sharcnet.ca/Software/Ansys/16.2.3/en-us/help/ans_ope/Hlp_G_OPE3_1.html """ self.ansys = ANSYS(exec_loc=exec_loc, run_location=run_location, jobname=jobname, nproc=nproc, override=override, cleardir=cleardir, add_flags=add_flags) self._ANSYS = True def Info(self, act=False): """ Turn on/off the return of the commands to Python. Parameters ---------- act : bool, obligatory True turn On and False turn Off the return of the commands to Python. """ self.PrintR = act if self._ANSYS: self.ansys.Info(act) #return self._PrintR('Now the commands will send a return to Python (like this).') def _PrintR(self, value): """ Internal function to print or not the return of commands based on command Info() """ if self.PrintR: return print(value, flush=True) else: pass def SetANSYSModel(self, inputname, extrafiles=[], directory=os.getcwd()): """ Set the input script file to be used on ANSYS and extra files that should be copied together. All this files must be in the same directory set in parameter directory. Parameters ---------- inputname : str, obligatory Name with extension of the script that will be executed in the analysis. The script must be done in function of the INPUT variables defined here, (as parameters of ANSYS), and must define/calculate ANSYS parameters with the results using the names defined here. extrafiles : list of strings, optional A list of strings containing extra files (with extension) that are necessary to run the script analys, could be an MODEL with the MESH already generated, for example. An example of extrafiles list is: ``extrafiles = ['temps.txt', 'model1.ans', 'file.db']`` directory : str, optional If the script is not in the current running Python directory you should place the entire location, if it's in a subdirectory of current directory you can use '/dirname/filename.ext'. Defaults to current running Python directory. """ if self._ANSYS: self.ansys.SetModel(inputname, extrafiles, directory) else: exception = Exception('ANSYS not declared yet. Before set ANSYS '+ 'model you must define ANSYS properties with ANSYS(...).') raise exception def SetANSYSOutVar(self, name): """ Defines a parameter/variable from ANSYS APDL script as an variable to return values for Python. Parameters ---------- name : str, obligatory Variable/Parameter name, as defined in APDL script. """ if self._ANSYS: self.ansys.CreateVarOut(name) else: exception = Exception('ANSYS not declared yet. Before set ANSYS '+ 'variables you must define ANSYS properties with ANSYS(...).') raise exception # Setting distribution of variables def CreateVar(self, name, distrib, mean, std=0, cv=None, par1=None, par2=None): """ Create a Variable, random or not. If it's used on ANSYS it need to be told, so after this use: >>> form.SetANSYSVar(name) Parameters ---------- name : str, obligatory Name of variable. distrib : str, obligatory Probabilistic variable distribution type. For all distributions Mean and Std are related to Normal distribution (the code determines the parameters for the desired distribution). Available types are: * gaussian (or gauss, normal); * lognormal (or log, logn, ln, lognorm); * gumbel (or gumb, type1); * constant (or const) - Constant value (doesn't need std). mean : float, obligatory Standard mean of variable values. std : float, optional Standard deviation of variable. You must define it or cv for variables that aren't constant, if both (cv and std) declared std will be used. For LogNormal variables it's recommend to use CV! cv : float, optional Coeficient of Variation of variable. You must define it or std for variables that aren't constant, if both (cv and std) declared std will be used. For LogNormal variables it's recommend to use CV! par1 and par2 : float, optional Parameters for future implementations. """ # Verify the variable name if name in ['', None, ' ']: exception = Exception('You must define the name of the variable that will receive the values.') raise exception # Set distribution name as lower case name = name.lower() distrib = distrib.lower() # CV or STD? if distrib not in ['constant', 'const', 'cons', 'c']: if cv == None: cv = std/mean else: std = cv*mean # Verify the distribution and then store the parameters # Gaussian variable if distrib in ['gauss', 'gaus', 'gaussian', 'normal', 'norm']: # Store values self.variableDistrib[name] = ['gauss', mean, std, cv] # Set the start point as the mean self.variableStartPt[name] = mean return self._PrintR('Variable \"%s\" defined as Gaussian with mean=%f, std. dev.=%f, CV=%f.' % (name, mean, std, cv)) # Lognormal distribution elif distrib in ['lognormal', 'logn', 'ln', 'log', 'lognorm']: # Store values self.variableDistrib[name] = ['logn', mean, std, cv] # Set the start point as the mean self.variableStartPt[name] = mean return self._PrintR('Variable \"%s\" defined as LogNormal with mean=%f, std. dev.=%f, CV=%f.' % (name, mean, std, cv)) # Gumbel distribution elif distrib in ['gumbel', 'gumb', 'type1']: # Store values self.variableDistrib[name] = ['gumbel', mean, std, cv] # Set the start point as the mean self.variableStartPt[name] = mean return self._PrintR('Variable \"%s\" defined as Gumbel with mean=%f, std. dev.=%f, CV=%f.' % (name, mean, std, cv)) # Constant value elif distrib in ['constant', 'const', 'cons', 'c']: # Store value self.variableConst[name] = mean return self._PrintR('Variable \"%s\" defined as Constant with value=%f' % (name, mean)) # or what? else: exception = Exception('Distribution \"%s\" set on variable \"%s\" is not recognized.' % (distrib.upper(), name)) raise exception def _SetControls(self, maxIter, tolRel, tolLS, dh, diff, meth): """ Internal from Run() Set the controls of process. Parameters ---------- maxIter : integer Maximum of iterations that can be performed. After this the process will stop with error. tolRel : float tolLS : float or string dh : float delta_h step when applying derivatives, value applied in reduced space. diff : str meth : str FORM method used. Available methods are: * HLRF: Hasofer Lind Rackwitz and Fiessler method. * iHLRF: improved Hasofer Lind Rackwitz and Fiessler method. """ if diff not in ['center', 'forward', 'backward']: exception = Exception('Invalid derivative method.') raise exception if min(maxIter, tolRel, dh) >= 0 and meth in ['HLRF', 'iHLRF', 'rHLRF'] and \ diff in ['center', 'forward', 'backward']: # Save controls variable self.controls['maxIter'] = maxIter self.controls['tolRel'] = tolRel self.controls['tolLS'] = tolLS self.controls['dh'] = dh self.controls['diff'] = diff self.controls['meth'] = meth self._PrintR('Process controls set as:') self._PrintR(' Limit of iterations: %d.' % maxIter) self._PrintR(' Relative error tolerance: %2.3E.' % tolRel) if isinstance(tolLS, str): self._PrintR(' Absolute LS error tolerance: auto.') else: self._PrintR(' Absolute LS error tolerance: %2.3E.' % tolLS) self._PrintR(' deltah (for derivatives): %2.3E.' % dh) self._PrintR(' Finite difference method: %s.' % diff) self._PrintR(' FORM Method: %s.' % meth) else: exception = Exception('Error while setting simulation controls. Please verify the set values.') raise exception def SetLimState(self, equat, userf=None): """ Set the limit state equation. Parameters ---------- equat : obligatory 1) It could be a string with the equation of the limit state. It must be write as a function of defined variables (In and Out). 2) It could be a Python function created by the user, just passing the function name in the place of the string. userf : function, optional An user defined function that could be used inside the limit state equation (string), called inside equat as ``userf()``. It's similar to use a function instead of a string in the equat parameter. For example, you can create a complex Python function with loops, ifs and whatever for evaluate the R part of your limit state function for a concrete beam. An example is showed after. First example: if ANSYS returns the maximum load on a truss as variable FxMAX, and applied loads to be tested are ``(g+q)*sin(theta)``, where ``g``, ``q``, theta are defined random variables created with ``CreateVar()``. .. code-block:: python form.SetLimState(equat='FxMAX-(g+q)*sin(theta)') Note that you can use math expressions as ``sin()``, ``cos()``, ``tan()``, ``sqrt()`` from Python math module inside the equation. | Second example: you have a steel bar in tension that hasn't hardening. It's stress is a function of ``(eps, fy, E)``, where ``eps`` is current deformation, ``fy`` is yield stress and ``E`` the elastic moduli, you can create inside your code an function like: .. code-block:: python def stress(eps, fy, E): if eps > fy/E: return fy else: return eps*E And now defining ``userf=stress`` we can: .. code-block:: python form.SetLimState(equat='userf(eps,fy,E)-q', userf=stress) where ``eps``, ``fy``, ``E`` and ``q`` are random variables. Note that the function inside the limit state equation should be called as ``userf()`` with the parameters from ``stress``. Or we can do the same using the functions instead of the string: .. code-block:: python form.SetLimState(equat=stress) """ # Store it self.limstate = equat if(type(self.limstate) == str): # String equation # Change equation to lowcase self.limstate = self.limstate.lower() self._userf = userf return self._PrintR('Limit state defined as "{}".'.format(equat)) else: # LS is a Python Function return self._PrintR('Limit state defined as "{}" function.'.format(self.limstate.__name__)) def SetANSYSVar(self, name): """ Set a variable as ANSYS variable. Parameters ---------- name : str, obligatory Name of variable. """ # Verify ANSYS object if self._ANSYS == False: exception = Exception('ANSYS not declared yet. Before set ANSYS '+ 'variables you must define ANSYS properties with ANSYS(...).') raise exception # Verify the variable name if name in ['', None, ' ']: exception = Exception('You must define the name of the variable.') raise exception else: name = name.lower() if name not in self.variableDistrib and name not in self.variableConst: exception = Exception('This variable name is not declared yet. '+ 'Only Random variables can be set as ANSYS variables.\n'+ 'Please use CreateVar() to declare it.') raise exception # Declare it on ansys object self.ansys.CreateVarIn(name) def SetStartPoint(self, name, value): """ Set the point, for each variable, that process will start. If it's not declared it will start with the mean value. Parameters ---------- name : str, obligatory Variable name. value : float, obligatory Starting point for this variable. """ # Set lower name = name.lower() # Verify if it's already created if name not in self.variableDistrib: exception = Exception('This variable name is not declared yet. '+ 'Before set the start value you must create the random '+ 'variable with CreateVar().') raise exception # If no error: store the value self.variableStartPt[name] = value # self._PrintR('Variable \"%s\" will start the process at point %f.' % (name, value)) def Options(self, option, value=None): """ Set extra options values. Parameters ---------- option : str, obligatory Name of option, listed next. value : optional Value to be set, type varies with option. If not defined it will return current value. ** Valid options:** For iHLRF method: * iHLRF_forced_lambdk : float Forced value when line search doesnt found a valid ``lambdak``. Being ``lambdak`` the step size. It could be set as ``'auto'``, when it is the complement of ``|y*.gradG|/(|y*|.|gradG|)``. Defaults to 'auto'. * iHLRF_prod_ck : float Scalar value that will be multiplied by calculated ``ck`` value. For a fix ``ck`` value turn it to 0 and then use 'iHLRF_add_ck'. * iHLRF_add_ck : float Scalar value that will be added to ``ck`` value. * iHLRF_par_a : float Value presented as ``a`` in line search equation for iHLRF. * iHLRF_par_b : float Value presented as ``b`` in line search equation for iHLRF, ``lambdak`` value is ``b**nk``. * iHLRF_step_lambdk_test : float Size of ``lambdak`` test block, after each block convergence is checked. For analyses using ANSYS: * APDLdebug: bool If it's true it will be print the dict with results imported from ANSYS at each call. Great use for APDL debug. ** If an invalid option or value is set the process could stop (or not).** """ if value == None: # Return current value return self._options[option] else: self._options[option] = value self._PrintR('Option \"%s\" set as \"%s\".' % (option, str(value))) def SetCorrel(self, var1, var2, correl): """ Set the correlation betwen two variables. The values will be transformed by the Nataf process before running. Parameters ---------- var1 : str, obligatory First variable name. var2 : str, obligatory Second variable name. correl : float, obligatory Correlation betwen var1 and var2. """ # Set lower var1 = var1.lower() var2 = var2.lower() # Verify if it's already created if var1 not in self.variableDistrib: exception = Exception('Variable "%s" is not declared yet. ' % var1 + 'Before set the correlation you must create the random '+ 'variable with CreateVar().') raise exception if var2 not in self.variableDistrib: exception = Exception('Variable "%s" is not declared yet. ' % var2 + 'Before set the correlation you must create the random '+ 'variable with CreateVar().') raise exception if var1 == var2: exception = Exception('You cannot change the correlation from a variable with itself.') raise exception if correl < -1 or correl > 1: exception = Exception('Correlation must be a value betwen -1 and +1.') raise exception # Store correlations on correlation list self.corlist self.corlist[var1, var2] = correl self.corlist[var2, var1] = correl self._PrintR('Correlation betwen \"%s\" and \"%s\" set as %f.' %(var1, var2, correl)) def _EquivNormal(self, name, pt): """ Internal function that determines the equivalent normal distribution parameters Parameters ---------- name : str, obligatory Name of variable that is being transformed. pt : float, obligatory Point that is being transformed. Returns ------- [mean, std] where: * mean : float Equivalent normal mean * std : float Equivalent normal standard deviation. """ # Copy the data to var name = name.lower() var = self.variableDistrib[name] # Gaussian if var[0] == 'gauss': # Doesn't need to transform mean = var[1] std = var[2] # Lognormal elif var[0] == 'logn': # LogNormal parameters qsi = math.sqrt(math.log(1 + (var[3])**2)) lmbd = math.log(var[1]) - 0.5*qsi**2 # Equiv parametes: analitycal since lognormal is derivated from normal std = pt*qsi mean = pt*(1-math.log(pt)+lmbd) # Gumbel elif var[0] == 'gumbel': # Gumbel parameters scl = math.sqrt(6)*var[2]/math.pi loc = var[1] - 0.57721*scl # Equiv with PDF/CDF functions # Gumbel pdf: scipy.stats.gumbel_r.pdf(x, loc, scl) # Gumbel cdf: scipy.stats.gumbel_r.cdf(x, loc, scl) # Normal pdf: scipy.stats.norm.pdf(x, mu, std) # Normal cdf: scipy.stats.norm.cdf(x, mu, std) # Normal icdf: scipy.stats.norm.ppf(q, mu, std) # invCum = icdf(gumbel_cdf()) invCum = scipy.stats.gumbel_r.cdf(pt, loc, scl) invCum = scipy.stats.norm.ppf(invCum, 0, 1) std = scipy.stats.norm.pdf(invCum, 0, 1) / scipy.stats.gumbel_r.pdf(pt, loc, scl) mean = pt - invCum*std # What?? else: exception = Exception('This couldnt happen!') raise exception # Return return [mean, std] def Run(self, maxIter=50, tolRel=0.01, tolLS='auto', dh=0.05, diff='forward', meth='iHLRF'): """ Run the FORM process. Parameters ---------- maxIter : integer, optional Maximum of iterations that can be performed. After this the process will stop with error. Defaults to 50. tolRel : float, optional Maximum **relative** error tolerance, for example on search for X point ``|X_k - X_(k-1)|/|X_(k-1)|<=tolRel``. Defaults to 0.005. tolLS : float, optional Maximum **absolute** error tolerance for limit state function, ``|G(X)|~=tolLS``. It should be calibrated based on the magnitude of limit state function. It's possible to automatically determine it using tolLS='auto', it will be set as (tolRel)*(first cycle limit state value). Defaults to 'auto'. dh : float, optional delta_h step when applying derivatives, value applied over means, as h=mean(x)*dh, so, ``f'(x) = (f(x+mean(x)*dh)-f(x)) / (mean(x)*dh)``. Defaults to 0.05. diff : str, optional Numeric derivative calcultation method. The possible mehtods are: * center: for finite difference method with central difference, ``f'(x) = (f(x+h)-f(x-h)) / (2h)``, it needs ``1 + 2*Nvars`` evaluations of the limit state function. * forward: for finite difference method with forward difference, ``f'(x) = (f(x+h)-f(x)) / h``, it needs ``1 + Nvars`` evaluations of the limit state function. * backward: for finite difference method with backward difference, ``f'(x) = (f(x)-f(x-h)) / h``, it needs ``1 + Nvars`` evaluations of the limit state function. Defaults to forward. meth : str, optional FORM method used. Available methods are: * HLRF: Hasofer Lind Rackwitz and Fiessler method. * iHLRF: improved Hasofer Lind Rackwitz and Fiessler method. Defaults to iHLRF. **Returns a dictionary with:** * status : integer Status of solution, values can be found after this list. * Pf : float Probability of failure. * Beta : float Reliability index. * {DesignPoint} : dictionary of values Dictionary with the design points for each variable. * {gradG} : dictionary of values Dictionary with the final gradient for each variable. * {alpha} : dictionary of values Dictionary with the final director cossines for each variable. * cycles : int Number of iterations performed to obtain the solution. **Status values:** * 0: no problem; * 1: warning, maximum of cycles reached with no convergence of CVPf; * 99: undefined error! """ #----------------------------------------------------------------------- # Set controls self._SetControls(maxIter=maxIter, tolRel=tolRel, tolLS=tolLS, dh=dh, diff=diff, meth=meth) #----------------------------------------------------------------------- #----------------------------------------------------------------------- # Verify the limit state equation if self.limstate == None: exception = Exception('Before Run you must define the limit state'+ 'equation with SetLimState().') raise exception #----------------------------------------------------------------------- #----------------------------------------------------------------------- # Verify if ANSYS is being used # if self._ANSYS: # Verify if the model is set if self.ansys.Model == {}: exception = Exception('Before running ANSYS you must define the model that will be analysed with SetANSYSModel().') raise exception #----------------------------------------------------------------------- #----------------------------------------------------------------------- # Copy the start point to self.variableDesPt() # self.variableDesPt = self.variableStartPt.copy() #----------------------------------------------------------------------- #----------------------------------------------------------------------- # Create the correlation Matrix, VarId list and Jyz/Jzy matrices # NInRandVars = len(self.variableDistrib) NInConstVars = len(self.variableConst) # NOutVars is ANSYS out vars NOutVars = 0 # correlMat start as an eye matrix, just with RandomVars! self.correlMat = np.eye(NInRandVars) # Create var id list varId = {} idx = 0 # For random variables for eachVar in self.variableDistrib: varId[eachVar] = idx idx += 1 # For constant variables for eachVar in self.variableConst: varId[eachVar] = idx idx += 1 # and Now ANSYS output variables, if using ANSYS if self._ANSYS: NOutVars = len(self.ansys.varOutNames) for eachVar in self.ansys.varOutNames: eachVar = eachVar.lower() varId[eachVar] = idx idx += 1 # Save varId in the object self.varId = varId # Run all the declareted correls for each in self.corlist: i = varId[each[0]] j = varId[each[1]] cor = self.corlist[each] # Apply Nataf to transform the correlation var1props = self.variableDistrib[each[0]] var2props = self.variableDistrib[each[1]] # Both are gauss if (var1props[0] == 'gauss' and var2props[0] == 'gauss'): cor = cor # Both are LN elif var1props[0] == 'logn' and var2props[0] == 'logn': cv1 = var1props[3] cv2 = var2props[3] cor = cor*(math.log(1+cor*cv1*cv2)/(cor*math.sqrt(math.log(1+cv1**2)*math.log(1+cv2**2)))) # Both are Gumbel elif var1props[0] == 'gumbel' and var2props[0] == 'gumbel': cor = cor*(1.064 - 0.069*cor + 0.005*cor**2) # One is gauss and other is logn elif (var1props[0] == 'gauss' and var2props[0] == 'logn') \ or (var2props[0] == 'gauss' and var1props[0] == 'logn'): # who is logn? if var1props[0] == 'logn': cv = var1props[3] else: cv = var2props[3] # cor is cor = cor*cv/math.sqrt(math.log(1+cv**2)) # One is gauss and other is gumbel elif (var1props[0] == 'gauss' and var2props[0] == 'gumbel') \ or (var2props[0] == 'gauss' and var1props[0] == 'gumbel'): cor = 1.031*cor # One is logn and other is gumbel elif (var1props[0] == 'logn' and var2props[0] == 'gumbel') \ or (var2props[0] == 'logn' and var1props[0] == 'gumbel'): # who is logn? if var1props[0] == 'logn': cv = var1props[3] else: cv = var2props[3] # cor is cor = cor*(1.029 + 0.001*cor + 0.014*cv + 0.004*cor**2 + 0.233*cv**2 - 0.197*cor*cv) # Forbiden zone else: exception = Exception('When applying NATAF on variables \"%s\" and \"%s\" the variables '+ 'conditions wasn\'t possible.') raise exception # Save it! self.correlMat[i, j] = cor self.correlMat[j, i] = cor # Obtain the transformation matrices Jyz/Jzy # Jzy is Lower matrix obtained by Cholesky at correlMat Jzy = np.linalg.cholesky(self.correlMat) # and Jyz is the inverse matrix Jyz = np.linalg.inv(Jzy) #----------------------------------------------------------------------- #----------------------------------------------------------------------- # CYCLEs # self._PrintR('\nStarting FORM process.') timei = time.time() # Start values lastcycle = False curVecRedPts = np.zeros(NInRandVars) valG = 0 curBeta = 0 self.results['Beta'] = [] self.results['Beta'].append(0) # Cycle (or iteration) start at 1, so +1 for cycle in range(1, 1+self.controls['maxIter']): #self._PrintR('----------------------------------------------------------------------------\n') self._PrintR('____________________________________________________________________________') self._PrintR('Iteration cycle %d.' % cycle) #------------------------------------------------------------------- # Mount mean vector, stddev matrix and transformations (Jxy/Jyx) matrices # matStd = np.zeros([NInRandVars, NInRandVars]) vecMean = np.zeros(NInRandVars) # Vector with current design point (and constant values) vecPts = np.zeros(NInRandVars+NInConstVars) for eachVar in self.variableDistrib: # Variable id id = varId[eachVar] # Current X point pt = self.variableDesPt[eachVar] # Transform [vecMean[id], matStd[id, id]] = self._EquivNormal(eachVar, pt) # Put point in vecPts vecPts[id] = pt for eachVar in self.variableConst: # Get variable id id = varId[eachVar] # Put point in vecPts vecPts[id] = self.variableConst[eachVar] # In case of correlations we need to apply that over Jxy/Jyx Jxy = matStd.dot(Jzy) Jyx = Jyz.dot(np.linalg.inv(matStd)) #------------------------------------------------------------------- #------------------------------------------------------------------- # Evaluate G and grad(G) # # Get dh dh = self.controls['dh'] if diff == 'center': # size is 1+2*NInRandVars matEvalPts = np.zeros([(1+2*NInRandVars+NInConstVars), (NInRandVars+NInConstVars+NOutVars)]) # All lines starts with design point/constant values, after random variables replace it matEvalPts[:, 0:(NInRandVars+NInConstVars)] = vecPts # Run all lines from [1,end] with step=2 # curId is current varId curId = 0 for eachLine in range(1, 1+2*NInRandVars, 2): # vecdh is not zero on current var item vecdh = np.zeros(NInRandVars) vecdh[curId] = dh # Odd line is +h matEvalPts[eachLine, 0:NInRandVars] = vecPts[0:NInRandVars] + vecMean*vecdh # and now -h matEvalPts[eachLine+1, 0:NInRandVars] = vecPts[0:NInRandVars] - vecMean*vecdh curId += 1 elif diff == 'forward': # size is 1+NInRandVars matEvalPts = np.zeros([(1+NInRandVars+NInConstVars), (NInRandVars+NInConstVars+NOutVars)]) # All lines starts with design point/constant values, after random variables replace it matEvalPts[:, 0:(NInRandVars+NInConstVars)] = vecPts curId = 0 for eachLine in range(1, 1+NInRandVars): # vecdh is not zero on current var item vecdh = np.zeros(NInRandVars) vecdh[curId] = dh matEvalPts[eachLine, 0:NInRandVars] = vecPts[0:NInRandVars] + vecMean*vecdh curId += 1 elif diff == 'backward': # size is 1+NInRandVars matEvalPts = np.zeros([(1+NInRandVars+NInConstVars), (NInRandVars+NInConstVars+NOutVars)]) # All lines starts with design point/constant values, after random variables replace it matEvalPts[:, 0:(NInRandVars+NInConstVars)] = vecPts curId = 0 for eachLine in range(1, 1+NInRandVars): # vecdh is not zero on current var item vecdh = np.zeros(NInRandVars) vecdh[curId] = dh matEvalPts[eachLine, 0:NInRandVars] = vecPts[0:NInRandVars] - vecMean*vecdh curId += 1 #------------------------------------------------------------------- # If using ANSYS send it's variables dependents to simulation. # if self._ANSYS: # Mount a list of matEvalPts lines that should be simunlated # Line 0 with X value is used for all not calculated lines + G(x) ansysSendingList = [0] if diff == 'center': for eachVar in self.ansys.varInNames: # Current variable == random or constant? if eachVar.lower() in self.variableDistrib: eachVar = eachVar.lower() ansysSendingList.append(2*varId[eachVar]+1) ansysSendingList.append(2*varId[eachVar]+2) elif diff == 'forward' or diff == 'backward': for eachVar in self.ansys.varInNames: # Current variable is random or constant? if eachVar.lower() in self.variableDistrib: eachVar = eachVar.lower() ansysSendingList.append(varId[eachVar]+1) # Set length as Ns Ns = len(ansysSendingList) self.ansys.ClearValues() self.ansys.SetLength(Ns) # Send from the list to ANSYS varInValues for eachVar in self.ansys.varInNames: eachVar = eachVar.lower() self.ansys.SetVarInValues(eachVar, matEvalPts[ansysSendingList, varId[eachVar]]) # Run ANSYS self.ansys.Run() # Get results from ANSYS resANSYS = self.ansys.GetVarOutValues() # If control APDL debug is true! if self._options['APDLdebug'] == True: self._PrintR('---\nPrinting \'resANSYS\' dict:') self._PrintR(resANSYS) self._PrintR('---\n') # Add ANSYS results to matEvalPts for eachVar in self.ansys.varOutNames: # First all lines has value from X matEvalPts[:, varId[eachVar.lower()]] = resANSYS[eachVar][0] # Now values are stored as in ansysSendingList matEvalPts[ansysSendingList, varId[eachVar.lower()]] = resANSYS[eachVar] #------------------------------------------------------------------- #------------------------------------------------------------------- # Evaluate Limit State (valG) and Gradient (in x) (gradGx) # self._PrintR('Evaluating limit state.') # dic of values in each evaluation varVal = {} # Eval Limit State: for eachVar in varId: varVal[eachVar] = matEvalPts[0, varId[eachVar]] # Test if limstate is a string or a function if(type(self.limstate) == str): varVal['userf'] = self._userf valG = eval(self.limstate, globals(), varVal) else: valG = self.limstate(**varVal) # tolLS 'auto' is tolRel*(initial valG) if cycle == 1 and tolLS == 'auto': tolLS = self.controls['tolRel']*valG self.controls['tolLS'] = tolLS self._PrintR('Limit state tolerance set to %f.' % tolLS) # Print limit state value: self._PrintR('Limit state value = %f (tolerance = %f).' % (valG, self.controls['tolLS'])) # Eval Gradient self._PrintR('Evaluating gradient.') curId = 0 gradGx = 0 gradGx = np.zeros(NInRandVars) #Derivatives only for random variables/input var if diff == 'center': for eachLine in range(1, (1+2*NInRandVars), 2): # G(X+dh) = val1 for eachVar in varId: varVal[eachVar] = matEvalPts[eachLine, varId[eachVar]] # Test if limstate is a string or a function if(type(self.limstate) == str): val1 = eval(self.limstate, globals(), varVal) else: val1 = self.limstate(**varVal) # G(X-dh) = val2 for eachVar in varId: varVal[eachVar] = matEvalPts[eachLine+1, varId[eachVar]] # Test if limstate is a string or a function if(type(self.limstate) == str): val2 = eval(self.limstate, globals(), varVal) else: val2 = self.limstate(**varVal) gradGx[curId] = (val1-val2)/(2*dh*vecMean[curId]) curId += 1 elif diff == 'forward': for eachLine in range(1, (1+NInRandVars)): # G(X+dh) = val1 for eachVar in varId: varVal[eachVar] = matEvalPts[eachLine, varId[eachVar]] # Test if limstate is a string or a function if(type(self.limstate) == str): val1 = eval(self.limstate, globals(), varVal) else: val1 = self.limstate(**varVal) gradGx[curId] = (val1-valG)/(dh*vecMean[curId]) curId += 1 elif diff == 'backward': for eachLine in range(1, (1+NInRandVars)): # G(X-dh) = val1 for eachVar in varId: varVal[eachVar] = matEvalPts[eachLine, varId[eachVar]] # Test if limstate is a string or a function if(type(self.limstate) == str): val1 = eval(self.limstate, globals(), varVal) else: val1 = self.limstate(**varVal) gradGx[curId] = (valG-val1)/(dh*vecMean[curId]) curId += 1 #--------------------------------------------------------------- #--------------------------------------------------------------- # Get gradG (of y) by gradGx and Jxy # gradG = (Jxy.T).dot(gradGx) #--------------------------------------------------------------- #--------------------------------------------------------------- # Print gradient and alpha # self._PrintR('Gradient and direction cosines (alpha):') self._PrintR(' VarName | grad(g(X_i)) | alpha(i)') idx = 0 absgrad = math.sqrt(gradG.dot(gradG)) for eachVar in self.variableDesPt: self._PrintR(' %15s | %+12.5E | %+9.5E' % (eachVar, gradG[idx], gradG[idx]/absgrad)) idx += 1 self._PrintR(' ') #--------------------------------------------------------------- #--------------------------------------------------------------- # the min valu of grad must be greater than 0, if isn't print a warning if min(abs(gradG)) == 0: self._PrintR('\n WARNING: One or more terms from gradG are equal to zero. \n Please pay attention to that!\n') #------------------------------------------------------------------- #------------------------------------------------------------------- # Get reduced uncorrelated points vector # # reduced uncorrelated points curVecRedPts = Jyx.dot(vecPts[:NInRandVars]-vecMean) # Current beta if cycle == 1: self.results['Beta'].append(math.sqrt(curVecRedPts.dot(curVecRedPts))) curBeta = self.results['Beta'][cycle] #------------------------------------------------------------------- #------------------------------------------------------------------- # FORM Method # if meth in ['HLRF', 'iHLRF', 'rHLRF']: #------------------------------------------------------------------- # HLRF - Rackwitz and Fiessler recursive method - Not Improved # # New reduced design point: newVecRedPts = gradG*(gradG.dot(curVecRedPts)-valG)*1/gradG.dot(gradG) #------------------------------------------------------------------- # Verify the convergence # if(abs(gradG.dot(curVecRedPts)) > 0.0): schwarzYgradG = abs(gradG.dot(curVecRedPts))/math.sqrt(gradG.dot(gradG)*curVecRedPts.dot(curVecRedPts)) else: schwarzYgradG = 0.0 self._PrintR('Schwarz inequality betwen y* and gradG = %f (it must be next to 1).' % schwarzYgradG) if lastcycle == True: if abs(valG) < self.controls['tolLS'] and (1-schwarzYgradG) < self.controls['tolRel']: #self._PrintR('\nFinal design point found on cycle %d.' % cycle) #self._PrintR('Performing a last cycle with final values.') #lastcycle = True self._PrintR('Convergence criterias checked.') self._PrintR('That\'s all folks.') self._stnumb = 0 break else: self._PrintR('Convergence criterias not checked.') self._PrintR('Sorry, it wasn\'t the last cycle...') lastcycle = False #------------------------------------------------------------------- # If not converged yet and meth is rHLRF or iHLRF # if meth == 'rHLRF': dk = newVecRedPts - curVecRedPts newVecRedPts = curVecRedPts + self._options['rHLRF_relax']*dk if meth == 'iHLRF': #------------------------------------------------------------------- # iHLRF - improved Rackwitz and Fiessler recursive method # # parameters par_a = self._options['iHLRF_par_a'] par_b = self._options['iHLRF_par_b'] # direction dk = newVecRedPts - curVecRedPts # Line search self._PrintR('Starting line search for iHLRF step.') # find ck # ck by Beck 2019 val1 = math.sqrt(curVecRedPts.dot(curVecRedPts)/gradG.dot(gradG)) val2 = 0.00 if abs(valG) >= self.controls['tolLS']: # yk+dk = newVecRedPts val2 = 1/2*newVecRedPts.dot(newVecRedPts)/abs(valG) ck = max(val1, val2)*self._options['iHLRF_prod_ck'] + self._options['iHLRF_add_ck'] # As ck must be greater than ||y*||/||gradG(y*)|| while ck < val1: ck = 2*ck # Ck must always grow ck = max(ck, self._last_ck) self._last_ck = ck self._PrintR('ck value: %f.' % ck) #----------------------------------------------------------- # Find lambdk: # # Initial calcs # m(y) = 1/2*y.y^T + c*|g(y)| # gradMy = grad(m(y)) gradMy = np.zeros(NInRandVars) gradMy = curVecRedPts + ck*valG/abs(valG)*gradG #-- # Now we search for lambdk # # myk = m(y) myk = 1/2*curVecRedPts.dot(curVecRedPts) + ck*abs(valG) # Find maxnk maxdk = max(abs(dk)) # If b**n*max(dk) is less than tolRel, y ~= y+b**n*dk maxnk = math.ceil(math.log(self.controls['tolRel']/maxdk)/math.log(par_b)) # I'm not a good guy and so we will do less! maxnk += 0 # But if limit state value doesn't change anymore it will stop! stepnk = self._options['iHLRF_step_lambdk_test'] # We need max(b**n), since 0<b<1 and n start as 0, where # the first value that the condition is true is the value, # so we can use stepnk to avoid evaluate like 100 ANSYS # simulations and use the first... nk = 0 valG_nk = 99999999.0 done = False forcenk = False self._PrintR('iHLRF step range from %f to %f (%.0f steps), being test step size %d.' % (par_b**nk, par_b**maxnk, maxnk, stepnk)) for step in range(math.ceil(maxnk/stepnk)): # current lenght curlen = min(stepnk, maxnk-(step)*stepnk) # points #matEvalPts = np.zeros([(1+2*NInRandVars+NInConstVars), (NInRandVars+NInConstVars+NOutVars)]) matEvalPts = np.zeros([curlen, (NInRandVars+NInConstVars+NOutVars)]) matEvalPts[:, 0:(NInRandVars+NInConstVars)] = vecPts lambdks = np.zeros(curlen) for eachnk in range(curlen): lambdks[eachnk] = par_b**nk matEvalPts[eachnk, 0:NInRandVars] = vecMean + Jxy.dot(curVecRedPts + lambdks[eachnk]*dk) nk += 1 # pass ANSYSvars to ANSYS if self._ANSYS: self.ansys.ClearValues() self.ansys.SetLength(curlen) # Send from the list to ANSYS varInValues for eachVar in self.ansys.varInNames: eachVar = eachVar.lower() self.ansys.SetVarInValues(eachVar, matEvalPts[:, varId[eachVar]]) # Run ANSYS self.ansys.Run() # Get results from ANSYS resANSYS = self.ansys.GetVarOutValues() # If control APDL debug is true! if self._options['APDLdebug'] == True: self._PrintR('---\nPrinting \'resANSYS\' dict:') self._PrintR(resANSYS) self._PrintR('---\n') # Add ANSYS results to matEvalPts for eachVar in self.ansys.varOutNames: matEvalPts[:, varId[eachVar.lower()]] = resANSYS[eachVar] #print(matEvalPts) #print(matEvalPts.shape) # evaluate the limit state function for eachnk for eachnk in range(curlen): # dic of values in each evaluation varVal = {} # Save last valG_nk to compare it valG_nk_old = valG_nk # Eval Limit State for eachVar in varId: varVal[eachVar] = matEvalPts[eachnk, varId[eachVar]] #valG_nk = eval(self.limstate, globals(), varVal) # Test if limstate is a string or a function if(type(self.limstate) == str): varVal['userf'] = self._userf valG_nk = eval(self.limstate, globals(), varVal) else: valG_nk = self.limstate(**varVal) # Verify the condition lambdk = lambdks[eachnk] curYk = curVecRedPts + lambdk*dk mynk = 1/2*curYk.dot(curYk)+ck*abs(valG_nk) # Correct way - REAL ARMIJO RULE target = -par_a*lambdk*gradMy.dot(dk) # Some people talk about this: gradMy.dot(gradMy), but ?? #target = -par_a*lambdk*gradMy.dot(gradMy) if (mynk - myk) <= target: self._PrintR('iHLRF step size is %f.' % lambdk) done = True break if valG_nk_old/valG_nk == 1: #if valG_nk_old/valG_nk >= (1-self.controls['tolRel']): forcenk = True break if done == True or forcenk == True: break if done == False or forcenk == True: if self._options['iHLRF_forced_lambdk'] == 'auto': lambdk = 1 - schwarzYgradG else: lambdk = self._options['iHLRF_forced_lambdk'] self._PrintR('iHLRF step not found, forcing to %f.' % lambdk) # Save new reduced point newVecRedPts = curVecRedPts + lambdk*dk else: exception = Exception('FORM method \"%s\" is not implemented.' % meth) raise exception #------------------------------------------------------------------- # Transform reduced points back to real space # newBeta = math.sqrt(newVecRedPts.dot(newVecRedPts)) self.results['Beta'].append(newBeta) NewVecPts = vecMean + Jxy.dot(newVecRedPts) # Save it to self.variableDesPt for eachVar in self.variableDesPt: self.variableDesPt[eachVar] = NewVecPts[varId[eachVar]] #------------------------------------------------------------------- #------------------------------------------------------------------- # Tell user how is it going # self._PrintR(' ') self._PrintR(' Current Beta = %2.3f.' % curBeta) self._PrintR(' ') self._PrintR(' VarName | Next Design Point') idx = 0 for eachVar in self.variableDesPt: self._PrintR(' %15s | %8.5E' % (eachVar, self.variableDesPt[eachVar])) idx += 1 self._PrintR(' ') #------------------------------------------------------------------- #------------------------------------------------------------------- # Verify the convergence # # Here we have a problem: I don't know from where Beck get this schwarzYgradG verificarion! # I've used that, but now I changed to the old fashion way. # To use that again you need to change the next verification and remove the # conditional lastcycle below too. # #### Sometimes it didn't converge because if this crap if np.linalg.norm(newVecRedPts) > 0: relErrorPoint = max(abs((curVecRedPts-newVecRedPts)/newVecRedPts)) else: relErrorPoint = 1.0 absErrorBeta = abs(curBeta-newBeta) self._PrintR('Maximum relative error on design point = %1.4f.' % relErrorPoint) self._PrintR('Absolute error betwen current and next beta = %1.4f.' % absErrorBeta) #### if abs(valG) < self.controls['tolLS'] and (1-schwarzYgradG) < self.controls['tolRel']: #### if abs(valG) < self.controls['tolLS'] and relErrorPoint < self.controls['tolRel']: if abs(valG) < self.controls['tolLS']: # limstate value is ok if absErrorBeta < self.controls['tolRel']: # Converged by absolute difference betwen betas self._PrintR('\nFinal design point found on cycle %d by absolute difference betwen two betas.' % cycle) self._stnumb = 0 lastcycle = True break elif (1-schwarzYgradG) < self.controls['tolRel']: lastcycle = True self._PrintR('\nFinal design point was probably found on cycle %d by Schwarz inequality betwen y* and gradG.' % cycle) self._PrintR('A new cycle will be started to confirm the limit state value and it\'s gradient.') else: lastcycle = False else: lastcycle = False #------------------------------------------------------------------- #----------------------------------------------------------------------- #----------------------------------------------------------------------- # After CYCLEs # timef = time.time() self.results['ElapsedTime'] = ((timef-timei)/60) # Save last cycles self._cycles = cycle # Get results #self.results['Beta'] = math.sqrt(newVecRedPts.dot(newVecRedPts)) self.results['Pf'] = scipy.stats.norm.cdf(-self.results['Beta'][-1]) # Put grad and alpha in self.results self.results['grad'] = {} self.results['alpha'] = {} absgrad = math.sqrt(gradG.dot(gradG)) idx = 0 for eachVar in self.variableDistrib: self.results['grad'][eachVar] = gradG[idx] self.results['alpha'][eachVar] = gradG[idx]/absgrad idx += 1 self._PrintR('\n\n============================================================================\n') # Verify if stopped without convergence if cycle == self.controls['maxIter']: self._stnumb = 1 self._PrintR(' WARNING:') self._PrintR(' The process was finished after reach the limit of iterations without') self._PrintR(' reach the error tolerance.\n') self._PrintR(' Total of iterations: %3.3E' % (self._cycles)) self._PrintR(' Probability of failure (Pf): %2.4E' % (self.results['Pf'])) self._PrintR(' Reliability index (Beta): %2.3f' % (self.results['Beta'][-1])) self._PrintR(' Elapsed time: %4.2f minutes.' % (self.results['ElapsedTime'])) self._PrintR(' Final values:') self._PrintR(' VarName | D. Point | grad(g(X_i)) | alpha(i) ') for eachVar in self.variableDesPt: self._PrintR(' %15s | %8.5E | %+12.5E | %+9.5E' % (eachVar, \ self.variableDesPt[eachVar], self.results['grad'][eachVar], self.results['alpha'][eachVar])) #self._PrintR(' %s = %3.5E' % (eachVar, self.variableDesPt[eachVar])) self._PrintR('\n============================================================================\n\n') #------------------------------------------------------------------- #----------------------------------------------------------------------- # Send the return # # put all in a dic finret = {} finret['status'] = self._stnumb finret['Pf'] = self.results['Pf'] finret['Beta'] = self.results['Beta'][-1] finret['DesignPoint'] = self.variableDesPt finret['gradG'] = self.results['grad'] finret['alpha'] = self.results['alpha'] finret['cycles'] = self._cycles return finret #----------------------------------------------------------------------- def ExportDataCSV(self, filename, description=None): """ Exports process data to a CSV file. Parameters ---------- filename : str, obligatory Name of file that will receive the values, doesn't need the extension ".csv", it will be placed automatically. description : str, optional A string that will be write in the beggining of the file. """ # Open file filename = filename+'.csv' try: f = open(filename, 'wt') except: exception = Exception('Unable to open the \"%s\" file for write data.' % filename) raise exception else: # Starts with sep=, for Microsoft Excel f.write('sep=,\n') # Description if description != None: f.write('%s\n\n' % description) f.write('Input data:\n') # Simulation Controllers: f.write('Process Controllers:\n') f.write(',Limit of iterations:,%d\n' % self.controls['maxIter']) f.write(',Relative error tolerance:,%2.3E\n' % self.controls['tolRel']) f.write(',Absolute LS error tolerance:,%2.3E\n' % self.controls['tolLS']) f.write(',deltah (for derivatives):,%2.3E\n' % self.controls['dh']) f.write(',Finite difference method:,%s\n' % self.controls['diff']) f.write(',FORM Method:,%s\n' % self.controls['meth']) # ANSYS Properties if self._ANSYS: f.write('\n') f.write('ANSYS Properties:\n') f.write(',ANSYS Model:\n') f.write(',,Model:,%s\n' % self.ansys.Model['inputname']) f.write(',,Extra files:,%s\n' % self.ansys.Model['extrafiles']) f.write(',,Input dir.:,%s\n' % self.ansys.Model['directory']) f.write(',ANSYS Input variables:\n') for eachVar in self.ansys.varInNames: f.write(',,%s\n' % eachVar) f.write(',ANSYS Output variables:\n') for eachVar in self.ansys.varOutNames: f.write(',,%s\n' % eachVar) f.write('\n') # Random variables f.write('Random variables:\n') f.write(',Name,Distribution,Mean,Standard Deviation,CV,Par1,Par2\n') for eachVar in self.variableDistrib: values = self.variableDistrib[eachVar] cmd = ',%s,%s' % (eachVar, values[0]) for eachVal in values[1:]: cmd = '%s,%8.5E' % (cmd, eachVal) f.write('%s\n' % cmd) f.write('\n') # Constant variables f.write('Constant variables:\n') f.write(',Name,Value\n') for eachVar in self.variableConst: cmd = ',%s,%8.5E' % (eachVar, self.variableConst[eachVar]) f.write('%s\n' % cmd) f.write('\n') # Correlation Matrix f.write('Correlation matrix:\n') # First line with Varnames: cmd = ',' idx = 0 for eachVar in self.varId: # Just random variables! if idx >= len(self.variableDistrib): break cmd = '%s,%s' % (cmd, eachVar) idx += 1 cmd = '%s\n' % cmd f.write(cmd) # Matrix lines with first column as Varname idx = 0 for eachLine in self.correlMat: cmd = ',%s' % list(self.varId.keys())[idx] # WTF DID I DO? But it works very well... xD # GO HORSE, GO GO GO GO GO! for eachVal in eachLine: cmd = '%s,%f' % (cmd, eachVal) cmd = '%s\n' % cmd f.write(cmd) idx += 1 f.write('\n') # Limit state f.write('Limit State:,"%s"\n' % (self.limstate)) f.write('\n') # Initial point f.write('Initial design point:\n') for eachVar in self.variableStartPt: f.write(',%s,%8.5E\n' % (eachVar, self.variableStartPt[eachVar])) f.write('\n') # Results part f.write('\nResults:\n') # Final results f.write('FORM results:\n') f.write(',Exit status:,%d,\n' % self._stnumb) f.write(',Total of iterations:,%d\n' % (self._cycles)) f.write(',Probability of failure (Pf):,%2.4E\n' % self.results['Pf']) f.write(',Reliability Index (Beta):,%2.3f\n' % self.results['Beta'][-1]) f.write(',Elapsed time (minutes):,%4.3f\n' % self.results['ElapsedTime']) f.write('\n') # Final Sampling Point f.write('Final values:\n') f.write(',Variable,D. Point,grad(g(X_i)),alpha(i)\n') for eachVar in self.variableDesPt: f.write(',%s,%8.5E,%11.5E,%9.5E\n' % (eachVar, \ self.variableDesPt[eachVar], self.results['grad'][eachVar], self.results['alpha'][eachVar])) f.write('\n') # Beta values f.write('Beta indexes of cycles:\n') f.write(',Cycle,Beta\n') idx = 1 for eachBeta in self.results['Beta'][1:]: f.write(',%d,%2.3f\n' % (idx, eachBeta)) idx += 1 # End f.close() self._PrintR('Simulation data exported to "%s".' % filename)
djivey/kinetic
_modules/generate.py
## generate various items on-demand. For use in sls files when no appropriate module exists. import random import string from cryptography.fernet import Fernet __virtualname__ = 'generate' def __virtual__(): return __virtualname__ def mac(prefix='52:54:00'): return '{0}:{1:02X}:{2:02X}:{3:02X}'.format(prefix, random.randint(0, 0xff), random.randint(0, 0xff), random.randint(0, 0xff)) def erlang_cookie(length = 20): return ''.join(random.choice(string.ascii_uppercase) for i in range(length)) def fernet_key(): return Fernet.generate_key().decode('utf-8')
djivey/kinetic
_modules/netcalc.py
<filename>_modules/netcalc.py import salt.utils.network as network __virtualname__ = 'netcalc' def __virtual__(): return __virtualname__ def gethosts(cidr): return network._network_hosts(cidr)
djivey/kinetic
_modules/minionmanage.py
<reponame>djivey/kinetic<filename>_modules/minionmanage.py import salt.utils.network as network import salt.modules.file as file __virtualname__ = 'minionmanage' def __virtual__(): return __virtualname__ def populate_cache(path): pending = file.readdir(path) pending.remove('.') pending.remove('..') return pending def populate_controller(path): pending = file.readdir(path) pending.remove('.') pending.remove('..') return pending def populate_controllerv2(path): pending = file.readdir(path) pending.remove('.') pending.remove('..') return pending def populate_storage(path): pending = file.readdir(path) pending.remove('.') pending.remove('..') return pending def populate_storagev2(path): pending = file.readdir(path) pending.remove('.') pending.remove('..') return pending def populate_compute(path): pending = file.readdir(path) pending.remove('.') pending.remove('..') return pending def populate_computev2(path): pending = file.readdir(path) pending.remove('.') pending.remove('..') return pending def populate_container(path): pending = file.readdir(path) pending.remove('.') pending.remove('..') return pending def populate_containerv2(path): pending = file.readdir(path) pending.remove('.') pending.remove('..') return pending
djivey/kinetic
_modules/cephx.py
<reponame>djivey/kinetic ## inspiration for simple salt-key module taken from ## https://github.com/ceph/ceph-ansible/blob/master/library/ceph_key.py#L26 import os, struct, time, base64 __virtualname__ = 'cephx' def __virtual__(): return __virtualname__ def make_key(): key = os.urandom(16) header = struct.pack('<hiih', 1, int(time.time()), 0, len(key)) secret = base64.b64encode(header + key) return secret.decode('utf-8')
dendisuhubdy/pyannote-audio
hubconf.py
<gh_stars>0 #!/usr/bin/env python # encoding: utf-8 # The MIT License (MIT) # Copyright (c) 2019-2020 CNRS # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE # SOFTWARE. # AUTHORS # <NAME> - http://herve.niderb.fr import os import yaml import pathlib import typing import functools import shutil import zipfile import torch from pyannote.audio.features import Pretrained as _Pretrained from pyannote.pipeline import Pipeline as _Pipeline dependencies = ['pyannote.audio', 'torch'] # shasum -a 256 models/sad_ami.zip _MODELS = { 'sad_dihard': 'ee924bd1751e6960e4e4322425dcdfdc77abec33a5f3ac1b74759229c176ff70', 'scd_dihard': '809aabd69c8116783b1d5cc1ba2f043d2b2d2fe693b34520c948ad5b2940e07c', 'ovl_dihard': '0ae57e5fc099b498db19aabc1d4f29e21cad44751227137909bd500048830dbd', 'emb_voxceleb': None, 'sad_ami': 'cb77c5ddfeec41288f428ee3edfe70aae908240e724a44c6592c8074462c6707', 'scd_ami': 'd2f59569c485ba3674130d441e9519993f26b7a1d3ad7d106739da0fc1dccea2', 'ovl_ami': 'debcb45c94d36b9f24550faba35c234b87cdaf367ac25729e4d8a140ac44fe64', 'emb_ami': '93c40c6fac98017f2655066a869766c536b10c8228e6a149a33e9d9a2ae80fd8', } # shasum -a 256 pipelines/dia_ami.zip _PIPELINES = { 'dia_dihard': None, 'dia_ami': '81bb175bbcdbcfe7989e09dd9afbbd853649d075a6ed63477cd8c288a179e77b', } _GITHUB = 'https://github.com/pyannote/pyannote-audio-models' _URL = f'{_GITHUB}/raw/master/{{kind}}s/{{name}}.zip' def _generic(name: str, duration: float = None, step: float = 0.25, batch_size: int = 32, device: typing.Optional[typing.Union[typing.Text, torch.device]] = None, pipeline: typing.Optional[bool] = None, force_reload: bool = False) -> typing.Union[_Pretrained, _Pipeline]: """Load pretrained model or pipeline Parameters ---------- name : str Name of pretrained model or pipeline duration : float, optional Override audio chunks duration. Defaults to the one used during training. step : float, optional Ratio of audio chunk duration used for the internal sliding window. Defaults to 0.25 (i.e. 75% overlap between two consecutive windows). Reducing this value might lead to better results (at the expense of slower processing). batch_size : int, optional Batch size used for inference. Defaults to 32. device : torch.device, optional Device used for inference. pipeline : bool, optional Wrap pretrained model in a (not fully optimized) pipeline. force_reload : bool Whether to discard the existing cache and force a fresh download. Defaults to use existing cache. Returns ------- pretrained: `Pretrained` or `Pipeline` Usage ----- >>> sad_pipeline = torch.hub.load('pyannote/pyannote-audio', 'sad_ami') >>> scores = model({'audio': '/path/to/audio.wav'}) """ model_exists = name in _MODELS pipeline_exists = name in _PIPELINES if model_exists and pipeline_exists: if pipeline is None: msg = ( f'Both a pretrained model and a pretrained pipeline called ' f'"{name}" are available. Use option "pipeline=True" to ' f'load the pipeline, and "pipeline=False" to load the model.') raise ValueError(msg) if pipeline: kind = 'pipeline' zip_url = _URL.format(kind=kind, name=name) sha256 = _PIPELINES[name] return_pipeline = True else: kind = 'model' zip_url = _URL.format(kind=kind, name=name) sha256 = _MODELS[name] return_pipeline = False elif pipeline_exists: if pipeline is None: pipeline = True if not pipeline: msg = ( f'Could not find any pretrained "{name}" model. ' f'A pretrained "{name}" pipeline does exist. ' f'Did you mean "pipeline=True"?' ) raise ValueError(msg) kind = 'pipeline' zip_url = _URL.format(kind=kind, name=name) sha256 = _PIPELINES[name] return_pipeline = True elif model_exists: if pipeline is None: pipeline = False kind = 'model' zip_url = _URL.format(kind=kind, name=name) sha256 = _MODELS[name] return_pipeline = pipeline if name.startswith('emb_') and return_pipeline: msg = ( f'Pretrained model "{name}" has no associated pipeline. Use ' f'"pipeline=False" or remove "pipeline" option altogether.' ) raise ValueError(msg) else: msg = ( f'Could not find any pretrained model nor pipeline called "{name}".' ) raise ValueError(msg) if sha256 is None: msg = ( f'Pretrained {kind} "{name}" is not available yet but will be ' f'released shortly. Stay tuned...' ) raise NotImplementedError(msg) # path where pre-trained models and pipelines are downloaded and cached hub_dir = pathlib.Path(os.environ.get("PYANNOTE_AUDIO_HUB", "~/.pyannote/hub")).expanduser().resolve() pretrained_dir = hub_dir / f'{kind}s' pretrained_subdir = pretrained_dir / f'{name}' pretrained_zip = pretrained_dir / f'{name}.zip' if not pretrained_subdir.exists() or force_reload: if pretrained_subdir.exists(): shutil.rmtree(pretrained_subdir) from pyannote.audio.utils.path import mkdir_p mkdir_p(pretrained_zip.parent) try: msg = ( f'Downloading pretrained {kind} "{name}" to "{pretrained_zip}".' ) print(msg) torch.hub.download_url_to_file(zip_url, pretrained_zip, hash_prefix=sha256, progress=True) except RuntimeError as e: shutil.rmtree(pretrained_subdir) msg = ( f'Failed to download pretrained {kind} "{name}".' f'Please try again.') raise RuntimeError(msg) # unzip downloaded file with zipfile.ZipFile(pretrained_zip) as z: z.extractall(path=pretrained_dir) if kind == 'model': params_yml, = pretrained_subdir.glob('*/*/*/*/params.yml') pretrained = _Pretrained(validate_dir=params_yml.parent, duration=duration, step=step, batch_size=batch_size, device=device) if return_pipeline: if name.startswith('sad_'): from pyannote.audio.pipeline.speech_activity_detection import SpeechActivityDetection pipeline = SpeechActivityDetection(scores=pretrained) elif name.startswith('scd_'): from pyannote.audio.pipeline.speaker_change_detection import SpeakerChangeDetection pipeline = SpeakerChangeDetection(scores=pretrained) elif name.startswith('ovl_'): from pyannote.audio.pipeline.overlap_detection import OverlapDetection pipeline = OverlapDetection(scores=pretrained) else: # this should never happen msg = ( f'Pretrained model "{name}" has no associated pipeline. Use ' f'"pipeline=False" or remove "pipeline" option altogether.' ) raise ValueError(msg) return pipeline.load_params(params_yml) return pretrained elif kind == 'pipeline': from pyannote.audio.pipeline.utils import load_pretrained_pipeline params_yml, *_ = pretrained_subdir.glob('*/*/params.yml') return load_pretrained_pipeline(params_yml.parent) sad_dihard = functools.partial(_generic, 'sad_dihard') scd_dihard = functools.partial(_generic, 'scd_dihard') ovl_dihard = functools.partial(_generic, 'ovl_dihard') dia_dihard = functools.partial(_generic, 'dia_dihard') sad_ami = functools.partial(_generic, 'sad_ami') scd_ami = functools.partial(_generic, 'scd_ami') ovl_ami = functools.partial(_generic, 'ovl_ami') emb_ami = functools.partial(_generic, 'emb_ami') dia_ami = functools.partial(_generic, 'dia_ami') emb_voxceleb = functools.partial(_generic, 'emb_voxceleb') sad = sad_dihard scd = scd_dihard ovl = ovl_dihard emb = emb_voxceleb dia = dia_dihard if __name__ == '__main__': DOCOPT = """Create torch.hub zip file from validation directory Usage: hubconf.py <validate_dir> hubconf.py (-h | --help) hubconf.py --version Options: -h --help Show this screen. --version Show version. """ from docopt import docopt from pyannote.audio.applications.base import create_zip arguments = docopt(DOCOPT, version='hubconf') validate_dir = pathlib.Path(arguments['<validate_dir>']) hub_zip = create_zip(validate_dir) print(f'Created file "{hub_zip.name}" in directory "{validate_dir}".')
dendisuhubdy/pyannote-audio
pyannote/audio/utils/path.py
<gh_stars>0 #!/usr/bin/env python # encoding: utf-8 # The MIT License (MIT) # Copyright (c) 2016-2020 CNRS # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE # SOFTWARE. # AUTHORS # <NAME> - http://herve.niderb.fr import os import errno from pathlib import Path from typing import Text from typing import Union from pyannote.database.protocol.protocol import ProtocolFile from pyannote.core import Segment from pyannote.core import SlidingWindowFeature import numpy as np def mkdir_p(path): """Create directory and all its parents if they do not exist This is the equivalent of Unix 'mkdir -p path' Parameter --------- path : str Path to new directory. Reference --------- http://stackoverflow.com/questions/600268/mkdir-p-functionality-in-python """ try: os.makedirs(path) except OSError as exc: # Python >2.5 if exc.errno == errno.EEXIST and os.path.isdir(path): pass else: raise exc class Pre___ed: """Wrapper around Precomputed, Pretrained, or torch.hub model This class allows user-facing APIs to support (torch.hub or locally) pretrained models or precomputed output interchangeably. * Pre___ed('sad_ami') is equivalent to torch.hub.load('pyannote/pyannote-audio', 'sad_ami') * Pre___ed('/path/to/xp/train/.../validate/...') is equivalent to Pretrained('/path/to/xp/train/.../validate/...') * Pre___ed('/path/to/xp/train/.../validate/.../apply/...') is equivalent to Precomputed('/path/to/xp/train/.../validate/.../apply/...') Bonus: Pre___ed('@scores') is equivalent to lambda f: f['scores'] Parameter --------- placeholder : Text, Path, Precomputed, or Pretrained """ def __init__(self, placeholder: Union[Text, Path, 'Precomputed', 'Pretrained']): super().__init__() from pyannote.audio.features import Pretrained from pyannote.audio.features import Precomputed if isinstance(placeholder, (Pretrained, Precomputed)): scorer = placeholder # if the path to a directory is provided elif Path(placeholder).is_dir(): directory = Path(placeholder) # if this succeeds, it means that 'placeholder' was indeed a path # to the output of "pyannote-audio ... apply" try: scorer = Precomputed(root_dir=directory) except Exception as e: scorer = None if scorer is None: # if this succeeds, it means that 'placeholder' was indeed a # path to the output of "pyannote-audio ... validate" try: scorer = Pretrained(validate_dir=directory) except Exception as e: scorer = None if scorer is None: msg = ( f'"{placeholder}" directory does not seem to be the path ' f'to precomputed features nor the path to a model ' f'validation step.' ) # otherwise it should be a string elif isinstance(placeholder, Text): # @key means that one should read the "key" key of protocol files if placeholder.startswith('@'): key = placeholder[1:] scorer = lambda current_file: current_file[key] # if string does not start with "@", it means that 'placeholder' # is the name of a torch.hub model else: try: import torch scorer = torch.hub.load('pyannote/pyannote-audio', placeholder) except Exception as e: msg = ( f'Could not load {placeholder} model from torch.hub. ' f'The following exception was raised:\n{e}') scorer = None # warn the user the something went wrong if scorer is None: raise ValueError(msg) self.scorer_ = scorer def crop(self, current_file: ProtocolFile, segment: Segment, mode: Text = 'center', fixed: float = None) -> np.ndarray: """Extract frames from a specific region Parameters ---------- current_file : ProtocolFile Protocol file segment : Segment Region of the file to process. mode : {'loose', 'strict', 'center'}, optional In 'strict' mode, only frames fully included in 'segment' support are returned. In 'loose' mode, any intersecting frames are returned. In 'center' mode, first and last frames are chosen to be the ones whose centers are the closest to 'segment' start and end times. Defaults to 'center'. fixed : float, optional Overrides 'segment' duration and ensures that the number of returned frames is fixed (which might otherwise not be the case because of rounding errors). Returns ------- frames : np.ndarray Frames. """ from pyannote.audio.features import Precomputed from pyannote.audio.features import Pretrained if isinstance(self.scorer_, (Precomputed, Pretrained)): return self.scorer_.crop(current_file, segment, mode=mode, fixed=fixed) return self.scorer_(current_file).crop(segment, mode=mode, fixed=fixed, return_data=True) def __call__(self, current_file) -> SlidingWindowFeature: """Extract frames from the whole file Parameters ---------- current_file : ProtocolFile Protocol file Returns ------- frames : np.ndarray Frames. """ return self.scorer_(current_file) # used to "inherit" most scorer_ attributes def __getattr__(self, name): return getattr(self.scorer_, name)
dendisuhubdy/pyannote-audio
pyannote/audio/labeling/tasks/base.py
<gh_stars>0 #!/usr/bin/env python # encoding: utf-8 # The MIT License (MIT) # Copyright (c) 2018-2020 CNRS # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE # SOFTWARE. # AUTHORS # <NAME> - http://herve.niderb.fr import torch import torch.nn.functional as F import numpy as np import scipy.signal from pyannote.core import Segment from pyannote.core import SlidingWindow from pyannote.core import Timeline from pyannote.core import Annotation from pyannote.core import SlidingWindowFeature from pyannote.database import get_unique_identifier from pyannote.database import get_annotated from pyannote.database.protocol.protocol import Protocol from pyannote.core.utils.numpy import one_hot_encoding from pyannote.audio.features import FeatureExtraction from pyannote.audio.features import RawAudio from pyannote.core.utils.random import random_segment from pyannote.core.utils.random import random_subsegment from pyannote.audio.train.trainer import Trainer from pyannote.audio.train.generator import BatchGenerator from pyannote.audio.train.task import Task, TaskType, TaskOutput from pyannote.audio.train.model import Resolution from pyannote.audio.train.model import RESOLUTION_CHUNK from pyannote.audio.train.model import RESOLUTION_FRAME class LabelingTaskGenerator(BatchGenerator): """Base batch generator for various labeling tasks This class should be inherited from: it should not be used directy Parameters ---------- feature_extraction : `pyannote.audio.features.FeatureExtraction` Feature extraction protocol : `pyannote.database.Protocol` subset : {'train', 'development', 'test'}, optional Protocol and subset. resolution : `pyannote.core.SlidingWindow`, optional Override `feature_extraction.sliding_window`. This is useful for models that include the feature extraction step (e.g. SincNet) and therefore output a lower sample rate than that of the input. Defaults to `feature_extraction.sliding_window` alignment : {'center', 'loose', 'strict'}, optional Which mode to use when cropping labels. This is useful for models that include the feature extraction step (e.g. SincNet) and therefore use a different cropping mode. Defaults to 'center'. duration : float, optional Duration of sub-sequences. Defaults to 3.2s. step : `float`, optional Sub-sequences step. Defaults to `duration`. Only used when `exhaustive` is True. batch_size : int, optional Batch size. Defaults to 32. per_epoch : float, optional Force total audio duration per epoch, in days. Defaults to total duration of protocol subset. exhaustive : bool, optional Ensure training files are covered exhaustively (useful in case of non-uniform label distribution). shuffle : bool, optional Shuffle exhaustive samples. Defaults to False. mask_dimension : `int`, optional When set, batches will have a "mask" key that provides a mask that has the same length as "y". This "mask" will be passed to the loss function has a way to weigh samples according to their "mask" value. The actual value of `mask_dimension` is used to select which dimension to use. This option assumes that `current_file["mask"]` contains a `SlidingWindowFeature` that can be used as masking. Defaults to not use masking. mask_logscale : `bool`, optional Set to True to indicate that mask values are log scaled. Will apply exponential. Defaults to False. Has not effect when `mask_dimension` is not set. """ def __init__(self, feature_extraction: FeatureExtraction, protocol: Protocol, subset='train', resolution: Resolution = None, alignment=None, duration=3.2, step=None, batch_size: int = 32, per_epoch: float = None, exhaustive=False, shuffle=False, mask_dimension=None, mask_logscale=False): self.feature_extraction = feature_extraction self.duration = duration # TODO: update 'step' semantics for consistency # TODO: should mean "ratio of duration" if step is None: step = duration self.step = step # TODO: create a reusable guess_resolution(self) function # TODO: in pyannote.audio.train.model if resolution in [None, RESOLUTION_FRAME]: resolution = self.feature_extraction.sliding_window elif resolution == RESOLUTION_CHUNK: resolution = SlidingWindow(duration=self.duration, step=self.step) self.resolution = resolution if alignment is None: alignment = 'center' self.alignment = alignment self.batch_size = batch_size self.exhaustive = exhaustive self.shuffle = shuffle self.mask_dimension = mask_dimension self.mask_logscale = mask_logscale total_duration = self._load_metadata(protocol, subset=subset) if per_epoch is None: per_epoch = total_duration / (24 * 60 * 60) self.per_epoch = per_epoch def postprocess_y(self, Y): """This function does nothing but return its input. It should be overriden by subclasses. Parameters ---------- Y : Returns ------- postprocessed : """ return Y def initialize_y(self, current_file): """Precompute y for the whole file Parameters ---------- current_file : `dict` File as provided by a pyannote.database protocol. Returns ------- y : `SlidingWindowFeature` Precomputed y for the whole file """ y, _ = one_hot_encoding(current_file['annotation'], get_annotated(current_file), self.resolution, labels=self.segment_labels_, mode='center') return SlidingWindowFeature(self.postprocess_y(y.data), y.sliding_window) def crop_y(self, y, segment): """Extract y for specified segment Parameters ---------- y : `pyannote.core.SlidingWindowFeature` Output of `initialize_y` above. segment : `pyannote.core.Segment` Segment for which to obtain y. Returns ------- cropped_y : (n_samples, dim) `np.ndarray` y for specified `segment` """ return y.crop(segment, mode=self.alignment, fixed=self.duration) def _load_metadata(self, protocol, subset='train') -> float: """Load training set metadata This function is called once at instantiation time, returns the total training set duration, and populates the following attributes: Attributes ---------- data_ : dict {'segments': <list of annotated segments>, 'duration': <total duration of annotated segments>, 'current_file': <protocol dictionary>, 'y': <labels as numpy array>} segment_labels_ : list Sorted list of (unique) labels in protocol. file_labels_ : dict of list Sorted lists of (unique) file labels in protocol Returns ------- duration : float Total duration of annotated segments, in seconds. """ self.data_ = {} segment_labels, file_labels = set(), dict() # loop once on all files for current_file in getattr(protocol, subset)(): # ensure annotation/annotated are cropped to actual file duration support = Segment(start=0, end=current_file['duration']) current_file['annotated'] = get_annotated(current_file).crop( support, mode='intersection') current_file['annotation'] = current_file['annotation'].crop( support, mode='intersection') # keep track of unique segment labels segment_labels.update(current_file['annotation'].labels()) # keep track of unique file labels for key, value in current_file.items(): if isinstance(value, (Annotation, Timeline, SlidingWindowFeature)): continue if key not in file_labels: file_labels[key] = set() file_labels[key].add(value) segments = [s for s in current_file['annotated'] if s.duration > self.duration] # corner case where no segment is long enough # and we removed them all... if not segments: continue # total duration of label in current_file (after removal of # short segments). duration = sum(s.duration for s in segments) # store all these in data_ dictionary datum = {'segments': segments, 'duration': duration, 'current_file': current_file} uri = get_unique_identifier(current_file) self.data_[uri] = datum self.file_labels_ = {k: sorted(file_labels[k]) for k in file_labels} self.segment_labels_ = sorted(segment_labels) for current_file in getattr(protocol, subset)(): uri = get_unique_identifier(current_file) if uri in self.data_: self.data_[uri]['y'] = self.initialize_y(current_file) return sum(datum['duration'] for datum in self.data_.values()) @property def specifications(self): """Task & sample specifications Returns ------- specs : `dict` ['task'] (`pyannote.audio.train.Task`) : task ['X']['dimension'] (`int`) : features dimension ['y']['classes'] (`list`) : list of classes """ specs = { 'task': Task(type=TaskType.MULTI_CLASS_CLASSIFICATION, output=TaskOutput.SEQUENCE), 'X': {'dimension': self.feature_extraction.dimension}, 'y': {'classes': self.segment_labels_}, } return specs def samples(self): if self.exhaustive: return self._sliding_samples() else: return self._random_samples() def _random_samples(self): """Random samples Returns ------- samples : generator Generator that yields {'X': ..., 'y': ...} samples indefinitely. """ uris = list(self.data_) durations = np.array([self.data_[uri]['duration'] for uri in uris]) probabilities = durations / np.sum(durations) while True: # choose file at random with probability # proportional to its (annotated) duration uri = uris[np.random.choice(len(uris), p=probabilities)] datum = self.data_[uri] current_file = datum['current_file'] # choose one segment at random with probability # proportional to its duration segment = next(random_segment(datum['segments'], weighted=True)) # choose fixed-duration subsegment at random subsegment = next(random_subsegment(segment, self.duration)) X = self.feature_extraction.crop(current_file, subsegment, mode='center', fixed=self.duration) y = self.crop_y(datum['y'], subsegment) sample = {'X': X, 'y': y} if self.mask_dimension is not None: # extract mask for current sub-segment mask = current_file['mask'].crop(subsegment, mode='center', fixed=self.duration) # use requested dimension (e.g. non-overlap scores) mask = mask[:, self.mask_dimension] if self.mask_logscale: mask = np.exp(mask) # it might happen that "mask" and "y" use different sliding # windows. therefore, we simply resample "mask" to match "y" if len(mask) != len(y): mask = scipy.signal.resample(mask, len(y), axis=0) sample['mask'] = mask for key, classes in self.file_labels_.items(): sample[key] = classes.index(current_file[key]) yield sample def _sliding_samples(self): uris = list(self.data_) durations = np.array([self.data_[uri]['duration'] for uri in uris]) probabilities = durations / np.sum(durations) sliding_segments = SlidingWindow(duration=self.duration, step=self.step) while True: np.random.shuffle(uris) # loop on all files for uri in uris: datum = self.data_[uri] # make a copy of current file current_file = dict(datum['current_file']) # compute features for the whole file features = self.feature_extraction(current_file) # randomly shift 'annotated' segments start time so that # we avoid generating exactly the same subsequence twice annotated = Timeline() for segment in get_annotated(current_file): shifted_segment = Segment( segment.start + np.random.random() * self.duration, segment.end) if shifted_segment: annotated.add(shifted_segment) if self.shuffle: samples = [] for sequence in sliding_segments(annotated): X = features.crop(sequence, mode='center', fixed=self.duration) y = self.crop_y(datum['y'], sequence) sample = {'X': X, 'y': y} if self.mask_dimension is not None: # extract mask for current sub-segment mask = current_file['mask'].crop(sequence, mode='center', fixed=self.duration) # use requested dimension (e.g. non-overlap scores) mask = mask[:, self.mask_dimension] if self.mask_logscale: mask = np.exp(mask) # it might happen that "mask" and "y" use different # sliding windows. therefore, we simply resample "mask" # to match "y" if len(mask) != len(y): mask = scipy.signal.resample(mask, len(y), axis=0) sample['mask'] = mask for key, classes in self.file_labels_.items(): sample[key] = classes.index(current_file[key]) if self.shuffle: samples.append(sample) else: yield sample if self.shuffle: np.random.shuffle(samples) for sample in samples: yield sample @property def batches_per_epoch(self): """Number of batches needed to complete an epoch""" duration_per_epoch = self.per_epoch * 24 * 60 * 60 duration_per_batch = self.duration * self.batch_size return int(np.ceil(duration_per_epoch / duration_per_batch)) class LabelingTask(Trainer): """Base class for various labeling tasks This class should be inherited from: it should not be used directy Parameters ---------- duration : float, optional Duration of sub-sequences. Defaults to 3.2s. batch_size : int, optional Batch size. Defaults to 32. per_epoch : float, optional Total audio duration per epoch, in days. Defaults to one day (1). """ def __init__(self, duration=3.2, batch_size=32, per_epoch=1): super(LabelingTask, self).__init__() self.duration = duration self.batch_size = batch_size self.per_epoch = per_epoch def get_batch_generator(self, feature_extraction, protocol, subset='train', resolution=None, alignment=None): """This method should be overriden by subclass Parameters ---------- feature_extraction : `pyannote.audio.features.FeatureExtraction` protocol : `pyannote.database.Protocol` subset : {'train', 'development'}, optional Defaults to 'train'. resolution : `pyannote.core.SlidingWindow`, optional Override `feature_extraction.sliding_window`. This is useful for models that include the feature extraction step (e.g. SincNet) and therefore output a lower sample rate than that of the input. alignment : {'center', 'loose', 'strict'}, optional Which mode to use when cropping labels. This is useful for models that include the feature extraction step (e.g. SincNet) and therefore use a different cropping mode. Defaults to 'center'. Returns ------- batch_generator : `LabelingTaskGenerator` """ return LabelingTaskGenerator( feature_extraction, protocol, subset=subset, resolution=resolution, alignment=alignment, duration=self.duration, step=self.step, per_epoch=self.per_epoch, batch_size=self.batch_size) @property def weight(self): """Class/task weights Returns ------- weight : None or `torch.Tensor` """ return None def on_train_start(self): """Set loss function (with support for class weights) loss_func_ = Function f(input, target, weight=None) -> loss value """ self.task_ = self.model_.task if self.task_.is_multiclass_classification: self.n_classes_ = len(self.model_.classes) def loss_func(input, target, weight=None, mask=None): if mask is None: return F.nll_loss(input, target, weight=weight, reduction='mean') else: return torch.mean( mask * F.nll_loss(input, target, weight=weight, reduction='none')) if self.task_.is_multilabel_classification: def loss_func(input, target, weight=None, mask=None): if mask is None: return F.binary_cross_entropy(input, target, weight=weight, reduction='mean') else: return torch.mean( mask * F.binary_cross_entropy(input, target, weight=weight, reduction='none')) if self.task_.is_regression: def loss_func(input, target, weight=None, mask=None): if mask is None: return F.mse_loss(input, target, reduction='mean') else: return torch.mean( mask * F.mse_loss(input, target, reduction='none')) self.loss_func_ = loss_func def batch_loss(self, batch): """Compute loss for current `batch` Parameters ---------- batch : `dict` ['X'] (`numpy.ndarray`) ['y'] (`numpy.ndarray`) ['mask'] (`numpy.ndarray`, optional) Returns ------- batch_loss : `dict` ['loss'] (`torch.Tensor`) : Loss """ # forward pass X = torch.tensor(batch['X'], dtype=torch.float32, device=self.device_) fX = self.model_(X) mask = None if self.task_.is_multiclass_classification: fX = fX.view((-1, self.n_classes_)) target = torch.tensor( batch['y'], dtype=torch.int64, device=self.device_).contiguous().view((-1, )) if 'mask' in batch: mask = torch.tensor( batch['mask'], dtype=torch.float32, device=self.device_).contiguous().view((-1, )) elif self.task_.is_multilabel_classification or \ self.task_.is_regression: target = torch.tensor( batch['y'], dtype=torch.float32, device=self.device_) if 'mask' in batch: mask = torch.tensor( batch['mask'], dtype=torch.float32, device=self.device_) weight = self.weight if weight is not None: weight = weight.to(device=self.device_) return { 'loss': self.loss_func_(fX, target, weight=weight, mask=mask), }
dendisuhubdy/pyannote-audio
pyannote/audio/pipeline/utils.py
<gh_stars>0 #!/usr/bin/env python # encoding: utf-8 # The MIT License (MIT) # Copyright (c) 2017-2020 CNRS # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE # SOFTWARE. # AUTHORS # <NAME> - http://herve.niderb.fr import yaml from pathlib import Path from pyannote.core import Annotation from pyannote.pipeline import Pipeline from pyannote.core.utils.helper import get_class_by_name def assert_string_labels(annotation: Annotation, name: str): """Check that annotation only contains string labels Parameters ---------- annotation : `pyannote.core.Annotation` Annotation. name : `str` Name of the annotation (used for user feedback in case of failure) """ if any(not isinstance(label, str) for label in annotation.labels()): msg = f'{name} must contain `str` labels only.' raise ValueError(msg) def assert_int_labels(annotation: Annotation, name: str): """Check that annotation only contains integer labels Parameters ---------- annotation : `pyannote.core.Annotation` Annotation. name : `str` Name of the annotation (used for user feedback in case of failure) """ if any(not isinstance(label, int) for label in annotation.labels()): msg = f'{name} must contain `int` labels only.' raise ValueError(msg) def load_pretrained_pipeline(train_dir: Path) -> Pipeline: """Load pretrained pipeline Parameters ---------- train_dir : Path Path to training directory (i.e. the one that contains `params.yml` created by calling `pyannote-pipeline train ...`) Returns ------- pipeline : Pipeline Pretrained pipeline """ config_yml = train_dir.parents[1] / 'config.yml' with open(config_yml, 'r') as fp: config = yaml.load(fp, Loader=yaml.SafeLoader) pipeline_name = config['pipeline']['name'] Klass = get_class_by_name( pipeline_name, default_module_name='pyannote.audio.pipeline') pipeline = Klass(**config['pipeline'].get('params', {})) return pipeline.load_params(train_dir / 'params.yml')
rumblefishrobotics/rf_common_ws
src/rf_display_ssd1306/nodes/rf_display_ssd1306_tester_node.py
<filename>src/rf_display_ssd1306/nodes/rf_display_ssd1306_tester_node.py #!/usr/bin/env python3 """ Node generating test messages for the display node. Could just use rostopic pub to rf_display/rf_display_topic RFDisplayData <payload>, but payload is a fair bit to keep typing / editing. This node is purely a testing tool. """ import rospy #from std_msgs.msg import String from rf_display.msg import RFDisplayData def publish_rf_display_messages(): # topic, message type, queue size pub = rospy.Publisher('rf_display_ssd1306_topic', RFDisplayData, queue_size=10) rospy.init_node('rf_display_ssd1306_tester_node', anonymous=True) rate = rospy.Rate(0.2) # 0.2 hz # create custom message display_msg = RFDisplayData() count = 0 while not rospy.is_shutdown(): count = count + 1 # Count just adds some changing text, on each line, with each message display_msg.screen_line1 = "line 1 (%s)" % count display_msg.screen_line2 = "line 2 (%s)" % count display_msg.screen_line3 = "line 3 (%s)" % count display_msg.screen_line4 = "line 4 (%s)" % count rospy.loginfo(display_msg) pub.publish(display_msg) rate.sleep() if __name__ == '__main__': try: publish_rf_display_messages() except rospy.ROSInterruptException: pass
rumblefishrobotics/rf_common_ws
src/rf_display_ssd1306/nodes/rf_display_ssd1306_node.py
<gh_stars>0 #!/usr/bin/env python3 """ display_ssd1306 provides a ROS wrapper for the SSD1306 LED, using the Adafruit drivers for their display. It provides four lines of text, stacked one above the other. Each line has up to 32 characters. """ import time import subprocess import rospy from PIL import Image, ImageDraw, ImageFont from rf_display.msg import RFDisplayData from board import SCL, SDA import busio import adafruit_ssd1306 #------------------------------------------------------------------------------- # Constants #------------------------------------------------------------------------------- RF_DISPLAY_WIDTH = 128 RF_DISPLAY_HEIGHT = 32 #------------------------------------------------------------------------------- # Functions #------------------------------------------------------------------------------- def shutdown_hook(): """Clear the display on shutdown.""" global g_draw # Clear the display g_display.fill(0) g_display.show() def init_display(): """Startup initilization of display.""" global g_i2c global g_display global g_draw global g_image global g_font g_i2c = busio.I2C(SCL, SDA) g_display = adafruit_ssd1306.SSD1306_I2C(RF_DISPLAY_WIDTH, RF_DISPLAY_HEIGHT, g_i2c) # Clear the display g_display.fill(0) g_display.show() # Create a blank image for drawing width = g_display.width height = g_display.height g_image = Image.new("1", (width, height)) # Draw a filled box to clear the image g_draw = ImageDraw.Draw(g_image) g_draw.rectangle((0, 0, width, height), outline=0, fill=0) g_font = ImageFont.load_default() def display_callback(display_data): """Pushes incomming data to display onto display""" rospy.loginfo("(%s) heard: (%s),(%s),(%s),(%s)" % (rospy.get_name(), display_data.screen_line1, display_data.screen_line2, display_data.screen_line3, display_data.screen_line4)) g_draw.rectangle((0, 0, RF_DISPLAY_WIDTH, RF_DISPLAY_HEIGHT), outline=0, fill=0) top = -2 g_draw.text((0, top + 0), display_data.screen_line1, font=g_font, fill=255) g_draw.text((0, top + 8), display_data.screen_line2, font=g_font, fill=255) g_draw.text((0, top + 16), display_data.screen_line3, font=g_font, fill=255) g_draw.text((0, top + 24), display_data.screen_line4, font=g_font, fill=255) g_display.image(g_image) g_display.show() def listen_for_messages(): """Creates ROS display node and starts listening for data to display.""" init_display() rospy.init_node("rf_display_ssd1306_node") #(topic),(custom message name),(name of callback function) rospy.Subscriber("rf_display_ssd1306_topic", RFDisplayData, display_callback) rospy.spin() # Register the ROS shutdown hook # rospy.on_shutdown(shutdown_hook) #------------------------------------------------------------------------------- # Startup source singleton #------------------------------------------------------------------------------- if __name__ == '__main__': try: listen_for_messages() except rospy.ROSInterruptException: pass
backyio/go-timezone
tools/gen-var-timezones.py
<filename>tools/gen-var-timezones.py import json import sys from jinja2 import Template def var_timezones(): data = {} abbr_timezones_json = sys.argv[1] with open(abbr_timezones_json) as f: data = json.load(f) tpl_text = '''var timezones = map[string][]string{ {%- for abbr in timezones %} "{{ abbr }}": []string{ {%- for tz in timezones[abbr] %} "{{ tz }}", {%- endfor %} }, {%- endfor %} }''' tpl = Template(tpl_text) print(tpl.render({"timezones": data})) if __name__ == "__main__": var_timezones()
backyio/go-timezone
tools/gen-var-tzabbrinfos.py
import json import sys from jinja2 import Template def var_tzabbrinfos(): data = {} data2 = {} abbrs_json = sys.argv[1] military_abbrs_json = sys.argv[2] with open(abbrs_json) as f: data = json.load(f) with open(military_abbrs_json) as f: data2 = json.load(f) tpl_text = '''var tzAbbrInfos = map[string][]*TzAbbreviationInfo{ {%- for abbr in tzinfos %} "{{ abbr }}": []*TzAbbreviationInfo{ {%- for tz in tzinfos[abbr] %} { countryCode: "{{ tz["country_code"] }}", isDST: {{ tz["is_dst"] | lower }}, name: "{{ tz["name"] }}", offset: {{ tz["offset"] }}, offsetHHMM: "{{ tz["offset_hhmm"] }}", }, {%- endfor %} }, {%- endfor %} // military timezones {%- for abbr in military_tzinfos %} "{{ abbr }}": []*TzAbbreviationInfo{ { name: "{{ military_tzinfos[abbr]["name"] }}", offset: {{ military_tzinfos[abbr]["offset"] }}, offsetHHMM: "{{ military_tzinfos[abbr]["offset_hhmm"] }}", }, }, {%- endfor %} }''' tpl = Template(tpl_text) print(tpl.render({"tzinfos": data, "military_tzinfos": data2})) if __name__ == "__main__": var_tzabbrinfos()
backyio/go-timezone
tools/gen-var-tzinfos.py
import json import sys from jinja2 import Template def var_tzinfos(): data = {} timezones_json = sys.argv[1] with open(timezones_json) as f: data = json.load(f) tpl_text = '''var tzInfos = map[string]*TzInfo{ {%- for tz in tzinfos %} "{{ tz }}": &TzInfo{ longGeneric: "{{ tzinfos[tz]["long"]["generic"] }}", longStandard: "{{ tzinfos[tz]["long"]["standard"] }}", longDaylight: "{{ tzinfos[tz]["long"]["daylight"] }}", shortGeneric: "{{ tzinfos[tz]["short"]["generic"] }}", shortStandard: "{{ tzinfos[tz]["short"]["standard"] }}", shortDaylight: "{{ tzinfos[tz]["short"]["daylight"] }}", standardOffset: {{ tzinfos[tz]["standard_offset"] }}, daylightOffset: {{ tzinfos[tz]["daylight_offset"] }}, standardOffsetHHMM: "{{ tzinfos[tz]["standard_offset_hhmm"] }}", daylightOffsetHHMM: "{{ tzinfos[tz]["daylight_offset_hhmm"] }}", countryCode: "{{ tzinfos[tz]["country_code"] }}", isDeprecated: {{ tzinfos[tz]["is_deprecated"] | lower }}, linkTo: "{{ tzinfos[tz]["link_to"] }}", lastDST: {{ tzinfos[tz]["last_dst"] }}, }, {%- endfor %} }''' tpl = Template(tpl_text) print(tpl.render({"tzinfos": data})) if __name__ == "__main__": var_tzinfos()
kristoffer-paulsson/angelostools
src/angelostools/pyxscanner.py
<filename>src/angelostools/pyxscanner.py """ The .pyx scanner helps you to scan a whole package hierarchy for Cython pyx files and encloses them in Extensions from setuptools. """ import glob from pathlib import Path from setuptools import Extension class PyxScanner: """Scans hierarchically for .pyx files in given package.""" def __init__(self, base_path: str, glob: list = None, extra: dict = None, basic: dict = None): self.__base_path = str(Path(base_path)) self.__globlist = glob if glob else ["**.pyx"] self.__pkgdata = extra if extra else dict() self.__data = basic if basic else dict() def scan(self) -> list: """Build list of Extensions to be cythonized.""" glob_result = list() for pattern in self.__globlist: glob_path = str(Path(self.__base_path, pattern)) glob_result += glob.glob(glob_path, recursive=True) extensions = list() for module in glob_result: package = ".".join(Path(module[len(self.__base_path) + 1:-4]).parts) data = self.__pkgdata[package] if package in self.__pkgdata else {} core = {"name": package, "sources": [module]} kwargs = {**self.__data, **data, **core} extensions.append(Extension(**kwargs)) return extensions def list(self) -> list: """Build list of modules found.""" glob_result = list() for pattern in self.__globlist: glob_path = str(Path(self.__base_path, pattern)) glob_result += glob.glob(glob_path, recursive=True) modules = list() for module in glob_result: package = ".".join(Path(module[len(self.__base_path) + 1:-4]).parts) modules.append(package) return modules
kristoffer-paulsson/angelostools
src/angelostools/setup/script.py
<filename>src/angelostools/setup/script.py # # Copyright (c) 2018-2020 by <NAME> <<EMAIL>>. # # This software is available under the terms of the MIT license. Parts are licensed under # different terms if stated. The legal terms are attached to the LICENSE file and are # made available on: # # https://opensource.org/licenses/MIT # # SPDX-License-Identifier: MIT # # Contributors: # <NAME> - initial implementation # import subprocess from pathlib import Path from setuptools import Command class Script(Command): """Compile the executable""" user_options = [ ("name=", "n", "Name of the script."), ("prefix=", "p", "Possible prefix where to link against.") ] def initialize_options(self): """Initialize options""" self.name = None self.prefix = None def finalize_options(self): """Finalize options""" pass def run(self): dist = str(Path(self.prefix, "bin").resolve()) if self.prefix else str(Path("./bin").absolute()) home = str(Path("./").absolute()) subprocess.check_call( "cp {0} {2}/{1}".format( Path(home, "scripts", self.name + "_entry_point.sh"), self.name, dist), cwd=home, shell=True)
kristoffer-paulsson/angelostools
src/angelostools/nsscanner.py
<filename>src/angelostools/nsscanner.py """Scanners that comply with namespace packages.""" import os from pathlib import Path from typing import Union class NamespacePackageScanner: """A scanner for namespace packages.""" def __init__(self, namespace: str, root: Path = None): self.__namespace = namespace self.__root = root.resolve() if root else Path(os.curdir).resolve() self.__root_parts_cnt = len(self.__root.parts) def pkg_iter(self) -> None: """Iterator over all namespace packages""" for pkg_path in self.__root.glob(self.__namespace + "-*/"): yield pkg_path def pkg_name(self, pkg_path: Path) -> str: """Convert package path into its name.""" return pkg_path.parts[-1] @property def packages(self) -> list: """Property over all namespace packages.""" return [self.pkg_name(pkg_path) for pkg_path in self.pkg_iter()] def _dir_iter(self, pkg_path: Path, rel_path: Union[str, list], pattern: str) -> None: """Internal iterator for directories and extensions in a namespace package.""" for file_path in pkg_path.joinpath(rel_path).rglob(pattern): yield file_path def mod_iter(self, pkg_path: Path) -> None: """Iterate over all modules in named namespace package.""" for mod_path in self._dir_iter(pkg_path, "src", "*.pyx"): yield mod_path def tests_iter(self, pkg_path: Path) -> None: """Iterate over all tests in named namespace package.""" for mod_path in self._dir_iter(pkg_path, "tests", "test_*.py"): yield mod_path def mod_imp_path(self, mod_path: Path) -> str: """Converts module path to full package name.""" return ".".join(mod_path.parts[self.__root_parts_cnt+2:-1] + (mod_path.stem,))
kristoffer-paulsson/angelostools
src/angelostools/setup/vendor.py
<reponame>kristoffer-paulsson/angelostools # # Copyright (c) 2018-2020 by <NAME> <<EMAIL>>. # # This software is available under the terms of the MIT license. Parts are licensed under # different terms if stated. The legal terms are attached to the LICENSE file and are # made available on: # # https://opensource.org/licenses/MIT # # SPDX-License-Identifier: MIT # # Contributors: # <NAME> - initial implementation # """Vendor installer, downloads, compiles and install libraries from source.""" import logging import os import platform import shutil import subprocess import tarfile import urllib import zipfile from pathlib import Path from tempfile import TemporaryDirectory from abc import ABC, abstractmethod from setuptools import Command class VendorLibrary(ABC): """Base class for vendors.""" @abstractmethod def check(self) -> bool: """Check if target is satisfied.""" pass @abstractmethod def download(self): """Download source tarball.""" pass @abstractmethod def extract(self): """Extract source file.""" pass def uncompress(self, archive, target): """Uncompress Zip or Tar files.""" if zipfile.is_zipfile(archive): zar = zipfile.ZipFile(archive) zar.extractall(target) elif tarfile.is_tarfile(archive): tar = tarfile.open(archive) tar.extractall(target) tar.close() else: raise OSError("Unkown zip/archive format") @abstractmethod def build(self): """Build sources.""" pass @abstractmethod def install(self): """Install binaries.""" pass @abstractmethod def close(self): """Clean up temporary files.""" pass class VendorCompile(VendorLibrary): def __init__( self, base_dir: str, name: str, download: str, local: str, internal: str, check: str, prefix: str = None ): """ Example: name = "libsodium" download = "https://download.libsodium.org/libsodium/releases/libsodium-1.0.18-stable.tar.gz" local = "libsodium-1.0.18.tar.gz" internal = "libsodium-stable" target = "./usr/local/lib/libsodium.a" """ self._base = base_dir self._prefix = str(Path(prefix).resolve()) if isinstance(prefix, str) else "" self._name = name self._download = download self._local = local self._internal = internal self._check = check self._tarball = Path(self._base, "tarball", self._local) self._temp = TemporaryDirectory() self._archive = str(Path(self._temp.name, self._local)) self._target = str(Path(self._temp.name, self._name)) self._work = str(Path(self._temp.name, self._name, self._internal)) def check(self) -> bool: """Check if target is reached""" return Path(self._base, self._check).exists() def download(self): """Download sources tarball.""" if not self._tarball.exists(): urllib.request.urlretrieve(self._download, self._archive) shutil.copyfile(self._archive, str(self._tarball)) else: shutil.copyfile(str(self._tarball), self._archive) def extract(self): """Extract source file.""" self.uncompress(self._archive, self._target) def close(self): """Clean up temporary files.""" self._temp.cleanup() class VendorDownload(VendorLibrary): """Vendor installer for third party libraries i source code form.""" def __init__( self, base_dir: str, name: str, download: str, local: str, internal: str, check: str, prefix: str = None ): """ Example: name = "libsodium" download = "https://download.libsodium.org/libsodium/releases/libsodium-1.0.18-stable.tar.gz" local = "libsodium-1.0.18.tar.gz" internal = "libsodium-stable" target = "./usr/local/lib/libsodium.a" """ self._base = base_dir self._prefix = str(Path(prefix).resolve()) if isinstance(prefix, str) else "" self._name = name self._download = download self._local = local self._internal = internal self._check = check self._tarball = Path(self._base, "tarball", self._local) self._target = str(Path(self._base, "tarball", self._name)) def check(self) -> bool: """Check if target is reached""" return Path(self._base, self._check).exists() def download(self): """Download sources tarball.""" if not self._tarball.exists(): urllib.request.urlretrieve(self._download, self._tarball) def extract(self): """Extract source file.""" self.uncompress(str(self._tarball), self._target) def build(self): pass def install(self): pass def close(self): pass class Vendor(Command): """Install third party vendor libraries.""" user_options = [ ("base-dir=", "d", "Base directory."), ("compile=", "c", "Download, compile and install source tarball."), ("prefix=", "p", "Possible prefix where to install") ] def initialize_options(self): """Initialize options""" self.base_dir = None self.compile = None self.prefix = None def finalize_options(self): """Finalize options""" pass def do_compile(self): """Execute the compile command.""" if not self.compile: return for value in self.compile: logging.info(self.base_dir) klass = value["class"] del value["class"] library = klass(self.base_dir, **value, prefix=self.prefix) if not library.check(): library.download() library.extract() library.build() library.install() library.close() def run(self): """Install vendors.""" self.do_compile()
kristoffer-paulsson/angelostools
src/angelostools/setup/executable.py
<reponame>kristoffer-paulsson/angelostools # # Copyright (c) 2018-2020 by <NAME> <<EMAIL>>. # # This software is available under the terms of the MIT license. Parts are licensed under # different terms if stated. The legal terms are attached to the LICENSE file and are # made available on: # # https://opensource.org/licenses/MIT # # SPDX-License-Identifier: MIT # # Contributors: # <NAME> - initial implementation # import subprocess import sys import tempfile from pathlib import Path from setuptools import Command class Executable(Command): """Compile the executable""" user_options = [ ("name=", "n", "Entry name."), ("prefix=", "p", "Possible prefix where to link against.") ] def initialize_options(self): """Initialize options""" self.name = None self.prefix = None def finalize_options(self): """Finalize options""" pass def run(self): major, minor, _, _, _ = sys.version_info PY_VER = "{0}.{1}".format(major, minor) config = str( Path(self.prefix, "bin", "python{}-config".format(PY_VER)).resolve() ) if self.prefix else "python{}-config".format(PY_VER) dist = str(Path(self.prefix, "bin").resolve()) if self.prefix else str(Path("./bin").absolute()) temp = tempfile.TemporaryDirectory() temp_name = str(Path(temp.name, self.name).absolute()) home = str(Path("./").absolute()) cflags = subprocess.check_output( "{0} --cflags".format(config), stderr=subprocess.STDOUT, shell=True).decode() # Debian 10 specific cflags = cflags.replace("-specs=/usr/share/dpkg/no-pie-compile.specs", "") # https://docs.python.org/3.8/whatsnew/3.8.html#debug-build-uses-the-same-abi-as-release-build if major == 3 and minor >= 8: ldflags = subprocess.check_output( "{0} --ldflags --embed".format(config), stderr=subprocess.STDOUT, shell=True).decode() else: ldflags = subprocess.check_output( "{0} --ldflags".format(config), stderr=subprocess.STDOUT, shell=True).decode() subprocess.check_call( "cython --embed -3 -o {0}.c {1}".format( temp_name, Path("./scripts/", self.name+"_entry_point.pyx")), cwd=home, shell=True) subprocess.check_call( "gcc -o {0}.o -c {0}.c {1}".format( temp_name, cflags), cwd=temp.name, shell=True) subprocess.check_call( "gcc -rdynamic -o {0} {1}.o {2}".format( Path(dist, self.name), temp_name, ldflags), cwd=home, shell=True) # -rdynamic --> --export-dynamic temp.cleanup()
broadstripes/loggly-sidecar
run.py
import os, sys, re if('LOGGLY_AUTH_TOKEN' not in os.environ): print('Missing $LOGGLY_AUTH_TOKEN') sys.exit(1) if('LOGGLY_TAG' not in os.environ): print('Missing $LOGGLY_TAG') sys.exit(1) auth_token = os.environ['LOGGLY_AUTH_TOKEN'] tags = [] for tag in os.environ['LOGGLY_TAG'].split(","): tags.append('tag=\\"{}\\"'.format(tag)) with open('/etc/syslog-ng/syslog-ng.conf.tmpl') as conf_template_file: conf_template = conf_template_file.read() conf = re.sub('LOGGLY_AUTH_TOKEN', auth_token, conf_template) conf = re.sub('LOGGLY_TAG', ' '.join(tags), conf) with open('/etc/syslog-ng/syslog-ng.conf', 'w') as conf_file: conf_file.write(conf) os.execlp('syslog-ng', 'syslog-ng', '--foreground', '--stderr', '--verbose')
rkyoto/pipe_event
pipe_event2.py
<filename>pipe_event2.py ''' pipe_event2 =========== This module provides a Event class which behaves just like threading.Event but is based on two pipes created using os.pipe() functions. Before Python 3.3, monotonic time is not introduced so adjusting system clock may affect Event.wait() function if specific timeout is set. Following notes can be found in PEP 0418: "If a program uses the system time to schedule events or to implement a timeout, it may fail to run events at the right moment or stop the timeout too early or too late when the system time is changed manually or adjusted automatically by NTP." This module demonstrates an alternative Event implementation on Unix-like systems which is not affected by the above issue. ''' import sys import os import fcntl import select import threading def _clear_pipe(fd): try: while True: if not os.read(fd, 1024): break except OSError: pass class Event: def __init__(self): _r, _w = os.pipe() # create the pipe # set pipe to non-blocking fl = fcntl.fcntl(_r, fcntl.F_GETFL) fcntl.fcntl(_r, fcntl.F_SETFL, fl | os.O_NONBLOCK) self._r_fd = _r self._w_fd = _w self._lock = threading.Lock() def __del__(self): os.close(self._r_fd) os.close(self._w_fd) def is_set(self): return self.wait(0) # just poll the pipe def isSet(self): return self.is_set() def set(self): with self._lock: if not self.is_set(): os.write(self._w_fd, b'\n') def clear(self): with self._lock: _clear_pipe(self._r_fd) def wait(self, timeout=None): try: ret = select.select([self._r_fd], [], [], timeout)[0] if ret: return True except select.error as e: sys.stderr.write(str(e) + '\n') return False
rkyoto/pipe_event
pipe_event.py
''' pipe_event ========== This module provides a Event class which behaves just like threading.Event but is based on two pipes created using os.pipe() functions. Before Python 3.3, monotonic time is not introduced so adjusting system clock may affect Event.wait() function if specific timeout is set. Following notes can be found in PEP 0418: "If a program uses the system time to schedule events or to implement a timeout, it may fail to run events at the right moment or stop the timeout too early or too late when the system time is changed manually or adjusted automatically by NTP." This module demonstrates an alternative Event implementation on Unix-like systems which is not affected by the above issue. ''' import os import fcntl import select import threading class Event: def __init__(self): r_fd, w_fd = os.pipe() # create the pipes # set read() to non-blocking fl = fcntl.fcntl(r_fd, fcntl.F_GETFL) fcntl.fcntl(r_fd, fcntl.F_SETFL, fl | os.O_NONBLOCK) # create file objects self.r_pipe = os.fdopen(r_fd, 'rb', 0) self.w_pipe = os.fdopen(w_fd, 'wb', 0) self.lock = threading.Lock() # create a lock to guard the pipes def __del__(self): self.r_pipe.close() self.w_pipe.close() def is_set(self): return self.wait(0) # just poll the pipe def isSet(self): return self.is_set() def set(self): self.lock.acquire() try: if not self.is_set(): self.w_pipe.write(b'\n') except: self.lock.release() raise self.lock.release() def clear(self): self.lock.acquire() try: self.r_pipe.read() except: pass self.lock.release() def wait(self, timeout=None): ret = select.select([self.r_pipe], [], [], timeout)[0] return len(ret) > 0
chandru99/neuralnilm
neuralnilm/data/stridesource.py
from __future__ import print_function, division from copy import copy from datetime import timedelta import numpy as np import pandas as pd import nilmtk from nilmtk.timeframegroup import TimeFrameGroup from nilmtk.timeframe import TimeFrame from neuralnilm.data.source import Sequence from neuralnilm.utils import check_windows from neuralnilm.data.source import Source from neuralnilm.consts import DATA_FOLD_NAMES import logging logger = logging.getLogger(__name__) class StrideSource(Source): """ Attributes ---------- data : dict Structure example: {<train | unseen_appliances | unseen_activations_of_seen_appliances>: { <building_name>: pd.DataFrame of with 2 cols: mains, target }} _num_seqs : pd.Series with 2-level hierarchical index L0 : train, unseen_appliances, unseen_activations_of_seen_appliances L1 : building_names """ def __init__(self, target_appliance, seq_length, filename, windows, sample_period, stride=None, rng_seed=None): self.target_appliance = target_appliance self.seq_length = seq_length self.filename = filename check_windows(windows) self.windows = windows self.sample_period = sample_period self.stride = self.seq_length if stride is None else stride self._reset() super(StrideSource, self).__init__(rng_seed=rng_seed) # stop validation only when we've gone through all validation data self.num_batches_for_validation = None self._load_data_into_memory() self._compute_num_sequences_per_building() def _reset(self): self.data = {} self._num_seqs = pd.Series() def _load_data_into_memory(self): logger.info("Loading NILMTK data...") # Load dataset dataset = nilmtk.DataSet(self.filename) for fold, buildings_and_windows in self.windows.items(): for building_i, window in buildings_and_windows.items(): dataset.set_window(*window) elec = dataset.buildings[building_i].elec building_name = ( dataset.metadata['name'] + '_building_{}'.format(building_i)) # Mains logger.info( "Loading data for {}...".format(building_name)) mains_meter = elec.mains() mains_good_sections = mains_meter.good_sections() appliance_meter = elec[self.target_appliance] good_sections = appliance_meter.good_sections( sections=mains_good_sections) def load_data(meter): return meter.power_series_all_data( sample_period=self.sample_period, sections=good_sections).astype(np.float32).dropna() mains_data = load_data(mains_meter) appliance_data = load_data(appliance_meter) df = pd.DataFrame( {'mains': mains_data, 'target': appliance_data}, dtype=np.float32).dropna() del mains_data del appliance_data if not df.empty: self.data.setdefault(fold, {})[building_name] = df logger.info( "Loaded data from building {} for fold {}" " from {} to {}." .format(building_name, fold, df.index[0], df.index[-1])) dataset.store.close() logger.info("Done loading NILMTK mains data.") def _compute_num_sequences_per_building(self): index = [] all_num_seqs = [] for fold, buildings in self.data.items(): for building_name, df in buildings.items(): remainder = len(df) - self.seq_length num_seqs = np.ceil(remainder / self.stride) + 1 num_seqs = max(0 if df.empty else 1, int(num_seqs)) index.append((fold, building_name)) all_num_seqs.append(num_seqs) multi_index = pd.MultiIndex.from_tuples( index, names=["fold", "building_name"]) self._num_seqs = pd.Series(all_num_seqs, multi_index) def get_sequence(self, fold='train', enable_all_appliances=False): if enable_all_appliances: raise ValueError("`enable_all_appliances` is not implemented yet" " for StrideSource!") # select building building_divisions = self._num_seqs[fold].cumsum() total_seq_for_fold = self._num_seqs[fold].sum() building_row_i = 0 building_name = building_divisions.index[0] prev_division = 0 for seq_i in range(total_seq_for_fold): if seq_i == building_divisions.iloc[building_row_i]: prev_division = seq_i building_row_i += 1 building_name = building_divisions.index[building_row_i] seq_i_for_building = seq_i - prev_division start_i = seq_i_for_building * self.stride end_i = start_i + self.seq_length data_for_seq = self.data[fold][building_name].iloc[start_i:end_i] def get_data(col): data = data_for_seq[col].values n_zeros_to_pad = self.seq_length - len(data) data = np.pad( data, pad_width=(0, n_zeros_to_pad), mode='constant') return data[:, np.newaxis] seq = Sequence(self.seq_length) seq.input = get_data('mains') seq.target = get_data('target') assert len(seq.input) == self.seq_length assert len(seq.target) == self.seq_length # Set mask seq.weights = np.ones((self.seq_length, 1), dtype=np.float32) n_zeros_to_pad = self.seq_length - len(data_for_seq) if n_zeros_to_pad > 0: seq.weights[-n_zeros_to_pad:, 0] = 0 # Set metadata seq.metadata = { 'seq_i': seq_i, 'building_name': building_name, 'total_num_sequences': total_seq_for_fold, 'start_date': data_for_seq.index[0], 'end_date': data_for_seq.index[-1] } yield seq @classmethod def _attrs_to_remove_for_report(cls): return ['data', '_num_seqs', 'rng']
chandru99/neuralnilm
neuralnilm/data/datathread.py
from __future__ import print_function from threading import Thread, Event from queue import Queue, Empty class DataThread(Thread): def __init__(self, data_pipeline, max_queue_size=8, **get_batch_kwargs): super(DataThread, self).__init__(name='neuralnilm-data-process') self._stop = Event() self._queue = Queue(maxsize=max_queue_size) self.data_pipeline = data_pipeline self._get_batch_kwargs = get_batch_kwargs def run(self): while not self._stop.is_set(): batch = self.data_pipeline.get_batch(**self._get_batch_kwargs) self._queue.put(batch) def get_batch(self, timeout=30): if self.is_alive(): return self._queue.get(timeout=timeout) else: raise RuntimeError("Process is not running!") def stop(self): self._stop.set() try: self._queue.get(block=False) except Empty: pass self.join()
isuhao/java-cef
tools/git_util.py
# Copyright (c) 2014 The Chromium Embedded Framework Authors. All rights # reserved. Use of this source code is governed by a BSD-style license that # can be found in the LICENSE file from exec_util import exec_cmd import os def is_checkout(path): """ Returns true if the path represents a git checkout. """ return os.path.exists(os.path.join(path, '.git')) def get_hash(path = '.', branch = 'HEAD'): """ Returns the git hash for the specified branch/tag/hash. """ cmd = "git rev-parse %s" % branch result = exec_cmd(cmd, path) if result['out'] != '': return result['out'].strip() return 'Unknown' def get_url(path = '.'): """ Returns the origin url for the specified path. """ cmd = "git config --get remote.origin.url" result = exec_cmd(cmd, path) if result['out'] != '': return result['out'].strip() return 'Unknown' def get_commit_number(path = '.', branch = 'HEAD'): """ Returns the number of commits in the specified branch/tag/hash. """ cmd = "git rev-list --count %s" % branch result = exec_cmd(cmd, path) if result['out'] != '': return result['out'].strip() return '0' def get_changed_files(path = '.'): """ Retrieves the list of changed files. """ # not implemented return []
isuhao/java-cef
tools/make_readme.py
# Copyright (c) 2014 The Chromium Embedded Framework Authors. All rights # reserved. Use of this source code is governed by a BSD-style license that # can be found in the LICENSE file. from date_util import * from file_util import * from optparse import OptionParser import os import re import shlex import subprocess import git_util as git import sys import zipfile def get_readme_component(name): """ Loads a README file component. """ paths = [] # platform directory paths.append(os.path.join(script_dir, 'distrib', platform)) # shared directory paths.append(os.path.join(script_dir, 'distrib')) # load the file if it exists for path in paths: file = os.path.join(path, 'README.' +name + '.txt') if path_exists(file): return read_file(file) raise Exception('Readme component not found: ' + name) def create_readme(): """ Creates the README.TXT file. """ # gather the components header_data = get_readme_component('header') mode_data = get_readme_component('standard') redistrib_data = get_readme_component('redistrib') footer_data = get_readme_component('footer') # format the file data = header_data + '\n\n' + mode_data + '\n\n' + redistrib_data + '\n\n' + footer_data data = data.replace('$JCEF_URL$', jcef_url) data = data.replace('$JCEF_REV$', jcef_commit_hash) data = data.replace('$JCEF_VER$', jcef_ver) data = data.replace('$CEF_URL$', cef_url) data = data.replace('$CEF_VER$', cef_ver) data = data.replace('$CHROMIUM_URL$', chromium_url) data = data.replace('$CHROMIUM_VER$', chromium_ver) data = data.replace('$DATE$', date) if platform == 'win32': platform_str = 'Windows 32-bit' elif platform == 'win64': platform_str = 'Windows 64-bit' elif platform == 'macosx64': platform_str = 'Mac OS-X 64-bit' elif platform == 'linux32': platform_str = 'Linux 32-bit' elif platform == 'linux64': platform_str = 'Linux 64-bit' data = data.replace('$PLATFORM$', platform_str) write_file(os.path.join(output_dir, 'README.txt'), data) if not options.quiet: sys.stdout.write('Creating README.TXT file.\n') # cannot be loaded as a module if __name__ != "__main__": sys.stderr.write('This file cannot be loaded as a module!') sys.exit() # parse command-line options disc = """ This utility builds the JCEF README.txt for the distribution. """ parser = OptionParser(description=disc) parser.add_option('--output-dir', dest='outputdir', metavar='DIR', help='output directory [required]') parser.add_option('--platform', dest='platform', help='target platform for distribution [required]') parser.add_option('-q', '--quiet', action='store_true', dest='quiet', default=False, help='do not output detailed status information') (options, args) = parser.parse_args() # the outputdir option is required if options.outputdir is None or options.platform is None: parser.print_help(sys.stderr) sys.exit(1) output_dir = options.outputdir # Test the operating system. platform = options.platform; if (platform != 'linux32' and platform != 'linux64' and platform != 'macosx64' and platform != 'win32' and platform != 'win64'): print 'Unsupported target \"'+platform+'\"' sys.exit(1) # script directory script_dir = os.path.dirname(__file__) # JCEF root directory jcef_dir = os.path.abspath(os.path.join(script_dir, os.pardir)) # Read and parse the CEF version file. args = {} read_readme_file(os.path.join(jcef_dir, 'jcef_build', 'README.txt'), args) # retrieve url and revision information for CEF if not git.is_checkout(jcef_dir): raise Exception('Not a valid checkout: %s' % (jcef_dir)) jcef_commit_number = git.get_commit_number(jcef_dir) jcef_commit_hash = git.get_hash(jcef_dir) jcef_url = git.get_url(jcef_dir) jcef_ver = '%s.%s.%s.g%s' % (args['CEF_MAJOR'], args['CEF_BUILD'], jcef_commit_number, jcef_commit_hash[:7]) date = get_date() cef_ver = args['CEF_VER'] cef_url = args['CEF_URL'] chromium_ver = args['CHROMIUM_VER'] chromium_url = args['CHROMIUM_URL'] # create the README.TXT file create_readme()
tehtechguy/annt
src/setup.py
# setup.py # # Author : <NAME> # Contact : http://techtorials.me # Date Created : 03/30/15 # # Description : Installs the annt project # Python Version : 2.7.8 # # License : MIT License http://opensource.org/licenses/mit-license.php # Copyright : (c) 2015 <NAME> # Native imports from distutils.core import setup import shutil # Install the program setup( name='annt', version='0.5.0', description="Artificial Neural Network Toolbox", author='<NAME>', author_email='<EMAIL>', url='http://techtorials.me', packages=['annt', 'annt.examples', 'annt.experimental'], package_data={'annt.examples':['data/mnist.pkl']} ) # Remove the unnecessary build folder try: shutil.rmtree('build') except: pass
tehtechguy/annt
src/annt/util.py
<filename>src/annt/util.py # util.py # # Author : <NAME> # Contact : http://techtorials.me # Date Created : 04/02/15 # # Description : Utility module. # Python Version : 2.7.8 # # License : MIT License http://opensource.org/licenses/mit-license.php # Copyright : (c) 2015 <NAME> """ Utility module. This module handles any sort of accessory items. G{packagetree annt} """ __docformat__ = 'epytext' # Native imports import os, pkgutil, cPickle # Third party imports import numpy as np # Program imports from annt.exception_handler import BaseException, wrap_error ############################################################################### ########## Exception Handling ############################################################################### class InsufficientSamples(BaseException): """ Exception if too many samples were desired. """ def __init__(self, sample, navail, nsamples): """ Initialize this class. @param sample: The desired sample. @param navail: The number of available samples. @param nsamples: The desired number of samples to extract. """ self.msg = wrap_error('The sample, {0}, only has {1} instances. The ' \ 'requested number of samples of {2} is too large. Please reduce ' \ 'the desired number of samples and try again'.format(sample, navail, nsamples)) ############################################################################### ########## Primary Functions ############################################################################### def mnist_data(): """ Return the example U{MNIST<http://yann.lecun.com/exdb/mnist/>} data. This is merely a subset of the data. There are 80 samples per digit for the training set (800 total items) and 20 samples per digit for the testing set (200 total items). @return: A tuple of tuples of the following format: (train_data, train_labels), (test_data, test_labels) """ with open(os.path.join(pkgutil.get_loader('annt.examples').filename, 'data', 'mnist.pkl'), 'rb') as f: return cPickle.load(f) def one_hot(x, num_items): """ Convert an array into a one-hot encoding. The indices in x mark which bits should be set. The length of each sub-array will be determined by num_items. @param x: The array indexes to mark as valid. This should be a numpy array. @param num_items: The number of items each encoding should contain. This should be at least as large as the max value in x + 1. @return: An encoded array. """ y = np.repeat([np.zeros(num_items, dtype='uint8')], x.shape[0], 0) for i, ix in enumerate(x): y[i, ix] = 1 return y def threshold(x, thresh, min_value=-1, max_value=1): """ Threshold all of the data in a given matrix (2D). @param x: The array to threshold. @param thresh: The value to threshold at. @param min_value: The minimum value to set the data to. @param max_value: The maximum value to set the data to. @return: An encoded array. """ y = np.empty(x.shape); y.fill(min_value) max_idx = x >= thresh y[max_idx] = max_value return y def get_random_paired_indexes(y, nsamples, labels=None): """ Get a list of indexes corresponding to random selections of the data in y. A total of nsamples will be returned for each specified label. @param y: A numpy array consisting of labels. @param nsamples: The number of samples to obtain for each unique value in y. @param labels: This parameter should be a list of the labels to use. If it is None then all unqiue labels will be used. @return: A dictionary containing the indexes in y corresponding to each unique value in y. There will be a total of nsamples indexes. @raise InsufficientSamples: Raised if too many samples were desired to be selected. """ # Extract initial indexes if labels is None: keys = np.unique(y) else: keys = np.array(labels) idx = [np.where(key == y)[0] for key in keys] # Check to make sure it is possible for key, ix in zip(keys, idx): if ix.shape[0] < nsamples: raise InsufficientSamples(key, ix.shape[0], nsamples) # Shuffle the indexes for i in xrange(len(idx)): np.random.shuffle(idx[i]) # Build final result return {key:ix[:nsamples] for key, ix in zip(keys, idx)} def flip_random_bits(x, pct, states={-1:1, 1:-1}): """ Randomly flip bits in the input stream. @param x: The input data to work with. This must be a vector. @param pct: The percentage of bits to flip. @param states: A dictionary containing the state representations. Each valid state should be mapped to its own, unique, complement and vice-versa. For example, if the bit stream consists of the set (1, -1) and '1' is the inverse of '-1' and vice-versa, the states parameter should be set to {-1:1, 1:-1}. @return: The flipped array. """ y = np.copy(x) max_bits = int(x.shape[0] * pct) idx = np.arange(x.shape[0]) np.random.shuffle(idx) for ix in idx[:max_bits]: y[ix] = states[x[ix]] return y
tehtechguy/annt
src/annt/plot.py
# plot.py # # Author : <NAME> # Contact : http://techtorials.me # Date Created : 03/31/15 # # Description : Module for plotting. # Python Version : 2.7.8 # # License : MIT License http://opensource.org/licenses/mit-license.php # Copyright : (c) 2015 <NAME> """ Module for plotting. G{packagetree annt} """ __docformat__ = 'epytext' # Native imports import itertools # Third-Party imports import numpy as np import matplotlib.pyplot as plt from mpl_toolkits.mplot3d import Axes3D def plot_epoch(y_series, series_names=None, y_errs=None, y_label=None, title=None, semilog=False, legend_location='best', out_path=None, show=True): """ Basic plotter function for plotting various types of data against training epochs. Each item in the series should correspond to a single data point for that epoch. @param y_series: A tuple containing all of the desired series to plot. @param series_names: A tuple containing the names of the series. @param y_errs: The error in the y values. There should be one per series per datapoint. It is assumed this is the standard deviation, but any error will work. @param y_label: The label to use for the y-axis. @param title: The name of the plot. @param semilog: If True the y-axis will be plotted using a log(10) scale. @param legend_location: The location of where the legend should be placed. Refer to matplotlib's U{docs<http://matplotlib.org/api/pyplot_api.html# matplotlib.pyplot.legend>} for more details. @param out_path: The full path to where the image should be saved. The file extension of this path will be used as the format type. If this value is None then the plot will not be saved, but displayed only. @param show: If True the plot will be show upon creation. """ # Construct the basic plot fig, ax = plt.subplots() if title is not None : plt.title(title) if semilog : ax.set_yscale('log') if y_label is not None : ax.set_ylabel(y_label) ax.set_xlabel('Epoch') plt.xlim((1, max([x.shape[0] for x in y_series]))) colormap = plt.cm.brg colors = itertools.cycle([colormap(i) for i in np.linspace(0, 0.9, len(y_series))]) markers = itertools.cycle(['.', ',', 'o', 'v', '^', '<', '>', '1', '2', '3', '4', '8', 's', 'p', '*', 'p', 'h', 'H', '+', 'D', 'd', '|', '_', 'TICKLEFT', 'TICKRIGHT', 'TICKUP', 'TICKDOWN', 'CARETLEFT', 'CARETRIGHT', 'CARETUP', 'CARETDOWN']) # Add the data if y_errs is not None: for y, err in zip(y_series, y_errs): x = np.arange(1, x.shape[0] + 1) ax.errorbar(x, y, yerr=err, color=colors.next(), marker=markers.next()) else: for y in y_series: x = np.arange(1, x.shape[0] + 1) ax.scatter(x, y, color=colors.next(), marker=markers.next()) # Create the legend if series_names is not None: plt.legend(series_names, loc=legend_location) # Save the plot fig.set_size_inches(19.20, 10.80) if out_path is not None: plt.savefig(out_path, format=out_path.split('.')[-1], dpi = 100) # Show the plot and close it after the user is done if show: plt.show() plt.close() def plot_weights(weights, nrows, ncols, shape, title=None, cluster_titles=None, out_path=None, show=True): """ Plot the weight matrices for the network. @param weights: A numpy array containing a weight matrix. Each row in the array corresponds to a unique node. Each column corresponds to a weight value. @param nrows: The number of rows of plots to create. @param ncols: The number of columns of plots to create. @param shape: The shape of the weights. It is assumed that a 1D shape was used and is desired to be represented in 2D. Whatever shape is provided will be used to reshape the weights. For example, if you had a 28x28 image and each weight corresponded to one pixel, you would have a vector with a shape of (784, ). This vector would then need to be resized to your desired shape of (28, 28). @param title: The name of the plot. @param cluster_titles: The titles for each of the clusters. @param out_path: The full path to where the image should be saved. The file extension of this path will be used as the format type. If this value is None then the plot will not be saved, but displayed only. @param show: If True the plot will be show upon creation. """ # Construct the basic plot fig = plt.figure() if title is not None: fig.suptitle(title, fontsize=16) if cluster_titles is None: cluster_titles = ['Node {0}'.format(i) for i in xrange(len(weights))] # Add all of the figures to the grid for i, weight_set in enumerate(weights): ax = plt.subplot(nrows, ncols, i + 1) ax.set_title(cluster_titles[i]) ax.imshow(weight_set.reshape(shape), cmap=plt.cm.gray) ax.axes.get_xaxis().set_visible(False) ax.axes.get_yaxis().set_visible(False) # Save the plot fig.set_size_inches(19.20, 10.80) if out_path is not None: plt.savefig(out_path, format=out_path.split('.')[-1], dpi = 100) # Show the plot and close it after the user is done if show: plt.show() plt.close() def make_grid(data): """ Convert the properly spaced, but unorganized data into a proper 3D grid. @param data: A sequence containing of data of the form (x, y, z). x and y are independent variables and z is the dependent variable. @return: A tuple containing the new x, y, and z data. """ # Sort the data x, y, z = np.array(sorted(data, key=lambda x: (x[0], x[1]))).T xi = np.array(sorted(list(set(x)))) yi = np.array(sorted(list(set(y)))) xim, yim = np.meshgrid(xi, yi) zi = z.reshape(xim.shape) return (xim, yim, zi) def plot_surface(x, y, z, x_label=None, y_label=None, z_label=None, title=None, out_path=None, show=True): """ Basic plotter function for plotting surface plots @param x: A sequence containing the x-axis data. @param y: A sequence containing the y-axis data. @param z: A sequence containing the z-axis data. @param x_label: The label to use for the x-axis. @param y_label: The label to use for the y-axis. @param z_label: The label to use for the z-axis. @param title: The name of the plot. @param out_path: The full path to where the image should be saved. The file extension of this path will be used as the format type. If this value is None then the plot will not be saved, but displayed only. @param show: If True the plot will be show upon creation. """ # Construct the basic plot fig = plt.figure() ax = fig.add_subplot(111, projection='3d') surf = ax.plot_surface(x, y, z, cmap=plt.cm.jet, rstride=1, cstride=1, linewidth=0) fig.colorbar(surf, shrink=0.5, aspect=5) ax.view_init(azim=58, elev=28) # Add the labels if title is not None : plt.title(title) if x_label is not None : ax.set_xlabel(x_label) if y_label is not None : ax.set_ylabel(y_label) if z_label is not None : ax.set_zlabel(z_label) # Save the plot fig.set_size_inches(19.20, 10.80) if out_path is not None: plt.savefig(out_path, format=out_path.split('.')[-1], dpi = 100) # Show the plot and close it after the user is done if show: plt.show() plt.close()
tehtechguy/annt
src/annt/experimental/hopfield_network.py
# hopfield_network.py # # Author : <NAME> # Contact : http://techtorials.me # Date Created : 04/07/15 # # Description : Example showing how to create and use a hopfield network. # Python Version : 2.7.8 # # License : MIT License http://opensource.org/licenses/mit-license.php # Copyright : (c) 2015 <NAME> """ Example showing how to create and use a hopfield network. This example uses a reduced set of data from the U{MNIST<http://yann.lecun.com/exdb/mnist/>} dataset. G{packagetree annt} """ __docformat__ = 'epytext' # Native imports import os # Third party imports import numpy as np # Program imports from annt.util import mnist_data, threshold from annt.util import flip_random_bits from annt.experimental.hopfield_net import Hopfield from annt.plot import plot_weights def main(train_data, train_labels, nsamples=10, nstates=1, labels=[0], pct_noise=0.3, plot=True): """ Demonstrates a hopfield network using MNIST. @param train_data: The data to train with. This must be an iterable returning a numpy array. @param train_labels: The training labels. This must be an iterable with the same length as train_data. @param nsamples: The number of samples to use for generating the attractor states. @param nstates: The number of states to create. This will result in creating a number of attractor states equal to the amount of unique labels multiplied by this value. @param labels: This parameter should be a list of the labels to use. If it is None then all unique labels will be used. @param pct_noise: The percentage of noise to be added to the attractor state. @param plot: If True one or more plots are generated showing the attractor states. @return: A list of lists containing the attractor state, the attractor state with noise, and the found attractor state, respectively. """ # Create the network net = Hopfield() # Initialize the attractor states net.create_attractor_states(train_data, train_labels, nstates, nsamples, 255 / 2, labels=labels) # Check noise tolerance with all of the activation states all_states = [] for state in net.attractor_states.keys(): # Get the states states = [] states.append(np.array(state)) states.append(flip_random_bits(states[0], pct_noise)) states.append(net.step(states[1])) all_states.append(np.copy(states)) # Make the plot if plot: plot_weights(states, 1, 3, (28, 28), title='Hopfield Network - ' 'Noise Tolerance Example\nFlipped {0}% of the bits'.format( pct_noise * 100), cluster_titles=('Attractor State', 'Attractor State with Noise', 'Found Attractor')) return all_states def basic_sim(): """ Perform a basic simulation. """ # Get the data (train_data, train_labels), (test_data, test_labels) = mnist_data() # Run the network main(threshold(train_data, 255 / 2), train_labels) def vary_params(out_dir, show_plot=True, seed=None): """ Vary some parameters and generate some plots. @param out_dir: The directory to save the plots in. @param show_plot: If True the plot will be displayed upon creation. @param seed: The seed for the random number generator. This will force the network to work with the same data. """ # Get the data (train_data, train_labels), (test_data, test_labels) = mnist_data() train_d = threshold(train_data, 255 / 2) # Make the output directory try: os.makedirs(out_dir) except OSError: pass ########################################################################### ###### Vary number of attractors ########################################################################### cluster_titles = ['Attractor State', 'Input', 'Found Attractor'] for i in xrange(1, 4): np.random.seed(seed) states = np.array(main(train_d, train_labels, labels=np.arange(i), pct_noise=0, plot=False)) plot_weights(states.reshape(states.shape[1] * states.shape[0], states.shape[2]), i, 3, (28, 28), title='Hopfield Network - {0} ' 'Attractor State(s)'.format(i), cluster_titles=cluster_titles * i, out_path=os.path.join(out_dir, 'states_{0}.png'.format(i)), show=show_plot) ########################################################################### ###### Vary amount of noise on input ########################################################################### noise_ranges = (0.4, 0.6) cluster_titles = ['Attractor State', 'Input with {0}% Noise', 'Found Attractor'] * len(noise_ranges) for i, noise in enumerate(noise_ranges): cluster_titles[i * 3 + 1] = cluster_titles[i * 3 + 1].format(noise * 100) for i in xrange(1, 3): new_title = [y for x in [[cluster_titles[0], cluster_titles[3 * j + 1], cluster_titles[2]] * i for j in xrange(len(noise_ranges))] for y in x] states = [] for noise in noise_ranges: np.random.seed(seed) states.extend(main(train_d, train_labels, labels=np.arange(i), pct_noise=noise, plot=False)) states = np.array(states) plot_weights(states.reshape(states.shape[1] * states.shape[0], states.shape[2]), len(noise_ranges) * i, 3, (28, 28), title='Hopfield Network - Noise Tolerance'.format(noise * 100), cluster_titles=new_title, out_path=os.path.join(out_dir, 'noise_states_{0}.png'.format(i)), show=show_plot) if __name__ == '__main__': basic_sim() # vary_params(out_dir=r'D:\annt\test\Hopfield_Network', show_plot=False, # seed=123456789)
tehtechguy/annt
src/annt/activation.py
<reponame>tehtechguy/annt # activation.py # # Author : <NAME> # Contact : http://techtorials.me # Date Created : 03/30/15 # # Description : Module for various activation functions. # Python Version : 2.7.8 # # License : MIT License http://opensource.org/licenses/mit-license.php # Copyright : (c) 2015 <NAME> """ Module for various activation functions. G{packagetree annt} """ __docformat__ = 'epytext' # Native imports from abc import ABCMeta, abstractmethod # Third party imports import numpy as np # Program imports from annt.exception_handler import BaseException, wrap_error ############################################################################### ########## Exception Handling ############################################################################### class UnsupportedActivationType(BaseException): """ Exception if the activation type is invalid. """ def __init__(self, type): """ Initialize this class. @param type: The type of activation function to use. """ self.msg = wrap_error('The type, {0}, is unsupported. The current ' 'types are {1}'.format(name, ', '.join(['linear', 'sigmoid']))) ############################################################################### ########## Functions ############################################################################### def get_activation(type): """ Returns a reference to an activation object. @param type: The type of activation function to use. @return: An activation object reference. @raise UnsupportedActivationType: Raised if type is invalid. """ if type == 'linear': return Linear elif type == 'sigmoid': return Sigmoid else: raise UnsupportedActivationType(type) def create_activation(type, **kargs): """ Creates an activation object instance. @param type: The type of activation function to use. @param kargs: Any keyword arguments. """ return get_activation(type)(**kargs) ############################################################################### ########## Class Template ############################################################################### class Activation(object): """ Base class description for an activation function. """ __metaclass__ = ABCMeta @abstractmethod def __init__(self): """ Initialize the class instance. """ @abstractmethod def compute(self, x): """ Compute the activation function. @param x: A numpy array representing the input data. This should be a vector. @return: A vector containing the element-wise result of applying the activation function to the input. """ ############################################################################### ########## Class Implementation ############################################################################### class Linear(Activation): """ Base class for a liner activation function. """ def __init__(self, m=1): """ Initializes this linear object. @param m: The slope of the line. Use the default value of "1" for the unity function. """ # Store the params self.m = m def compute(self, x): """ Compute the activation function. @param x: A numpy array representing the input data. This should be a vector. @return: A vector containing the element-wise result of applying the activation function to the input. """ return self.m * x def compute_derivative(self, x): """ Compute the activation function's derivative. @param x: A numpy array representing the input data. This should be a vector. @return: A vector containing the element-wise result of applying the activation function to the input. """ return np.repeat(self.m * len(x)) class Sigmoid(Activation): """ Base class for a sigmoid activation function. """ def __init__(self): """ Initializes this sigmoid object. """ pass def compute(self, x): """ Compute the activation function. @param x: A numpy array representing the input data. This should be a vector. @return: A vector containing the element-wise result of applying the activation function to the input. """ return 1 / (1 + np.exp(-x)) def compute_derivative(self, x): """ Compute the activation function's derivative. @param x: A numpy array representing the input data. This should be a vector. @return: A vector containing the element-wise result of applying the activation function to the input. """ y = self.compute(x) return y * (1 - y)
tehtechguy/annt
src/annt/timers.py
<reponame>tehtechguy/annt<filename>src/annt/timers.py # timers.py # # Author : <NAME> # Contact : http://techtorials.me # Date Created : 04/02/15 # # Description : Module used for timing. # Python Version : 2.7.8 # # License : MIT License http://opensource.org/licenses/mit-license.php # Copyright : (c) 2015 <NAME> """ Module used for timing. G{packagetree annt} """ __docformat__ = 'epytext' # Native imports import time, csv from itertools import izip ############################################################################### ########## Primary Functions ############################################################################### def pretty_time(elapsed_time): """ Get a string representing a formatted time in a pretty (human-readable) format. @param elapsed_time: The number of elapsed seconds. """ # The time labels labels = ('days', 'hours', 'minutes', 'seconds') formatted_times = [] # Compute the times days = elapsed_time / 86400. i_days = int(days) hours = (days - i_days) * 24 i_hours = int(hours) mins = (hours - i_hours) * 60 i_mins = int(mins) seconds = (mins - i_mins) * 60 times = (i_days, i_hours, i_mins, seconds) # Format all times but the last element for t, l in izip(times[:-1], labels[:-1]): if t != 0: formatted_times.append('{0} {1}'.format(t, l)) # Format the last element if times[-3] != 0: str = ', and {0:.3f} {1}' elif times[-2] != 0: str = ' and {0:.3f} {1}' else: str = '{0:.3f} {1}' # Return the formatted time return ', '.join(formatted_times) + str.format(times[-1], labels[-1]) ############################################################################### ########## Class Implementations ############################################################################### class SimpleTimer(object): """ Simple timer used to hold data just about a single timer. """ def __init__(self, name=None): """ Initialize this timer instance, starting the timer. """ # Init params self.name = name self.start() def start(self): """ Start the timer. """ self.start_time = time.time() def pause(self): """ Pauses the timer and compute the elapsed time. """ self.finish_time = time.time() self.elapsed_time += self.finish_time - self.start_time def stop(self): """ Stop the timer and compute the elapsed time. """ self.finish_time = time.time() self.elapsed_time = self.finish_time - self.start_time def get_elapsed_time(self, pretty=False): """ Return the elapsed time. @param pretty: If False the time is returned as the elapsed time in seconds. If True, the equivalent time breakdown is computed. @return: The elapsed time. """ # Determine how to format the time if not pretty: return self.elapsed_time else: return pretty_time(self.elapsed_time) class MultiTimer(object): """ Timer class used to work with multiple timers.. """ def __init__(self): """ Initialize this timer instance. """ self.timers = {} def add_timers(self, *timer_names): """ Adds a new timer to the class and starts the timer. @param timer_names: The name of the timer. """ for timer in timer_names: self.timers[timer] = SimpleTimer(timer) self.timers[timer].start() def stop_timers(self, *timer_names): """ Stop the specified timers. @param timer_names: One or more timer names to be stopped. """ for timer in timer_names: self.timers[timer].stop() def start_timers(self, *timer_names): """ Starts the specified timers. @param timer_names: One or more timer names to be started. """ for timer in timer_names: self.timers[timer].start() def pause_timers(self, *timer_names): """ Pauses the specified timers. @param timer_names: One or more timer names to be paused. """ for timer in timer_names: self.timers[timer].pause() def get_elapsed_time(self, timer_name, pretty=False): """ Retrieve the elapsed time for the specified timer. @param timer_name: The name of the timer to get the info for. @param pretty: If False the time is returned as the elapsed time in seconds. If True, the equivalent time breakdown is computed. @return: The elapsed time. """ return self.timers[timer_name].get_elapsed_time(pretty) def log_timers(self, out_path, header=True): """ Creates a log CSV file for all timers. Make sure to stop any timers before calling this. @param out_path: The full path to the CSV to write to. @param header: Flag denoting whether the header should be printed or not. """ with open(out_path, 'wb') as f: writer = csv.writer(f) if header: writer.writerow(self.timers.keys()) writer.writerow([timer.get_elapsed_time() for timer in self.timers.values()])
tehtechguy/annt
src/annt/examples/__init__.py
# __init__.py # # Author : <NAME> # Contact : http://techtorials.me # Date Created : 03/31/15 # # Description : Defines the examples package # Python Version : 2.7.8 # # License : MIT License http://opensource.org/licenses/mit-license.php # Copyright : (c) 2015 <NAME> """ This package contains some examples for working with the various types of networks. """ __docformat__ = 'epytext'
tehtechguy/annt
src/annt/experimental/hopfield_net.py
# hopfield_net.py # # Author : <NAME> # Contact : http://techtorials.me # Date Created : 04/07/15 # # Description : Module for a Hopfield network. # Python Version : 2.7.8 # # License : MIT License http://opensource.org/licenses/mit-license.php # Copyright : (c) 2015 <NAME> """ Module for a Hopfield network. G{packagetree annt} """ __docformat__ = 'epytext' # Third party imports import numpy as np # Program imports from annt.util import get_random_paired_indexes, threshold ############################################################################### ########## Class Implementation ############################################################################### class Hopfield(object): """ Base class for a Hopfield network. """ def __init__(self, attractor_states=None): """ Initializes this Hopfield network. @param attractor_states: The attractor states to use. This must be a dictionary containing a unique state as the key (this should be a tuple containing a vector of representing the current state). The value in the dictionary should be the corresponding label of the state. If None, the attractor states must be created using the "create_attractor_states" method. """ # Store the params self.attractor_states = attractor_states # Construct the weights (if possible) if self.attractor_states is not None: self.initialize_weights() def create_attractor_states(self, x, y, nstates, nsamples, thresh, labels=None): """ Create the attractor states based off the labeled data. Additionally, initialize the weights using the new attractor states. @param x: A numpy array consisting of the data to initialize with. @param y: A numpy array consisting of labels. @param nstates: The number of states to create. This will result in creating a number of attractor states equal to the amount of unique labels multiplied by this value. @param nsamples: The number of samples to use for each unique value in y. @param thresh: The value to threshold at. @param labels: This parameter should be a list of the labels to use. If it is None then all unique labels will be used. """ idxs = [get_random_paired_indexes(y, nsamples, labels) for _ in xrange(nstates)] self.attractor_states = {} for idx in idxs: for i, lbl in enumerate(idx): states = [x[ix] for ix in idx[lbl]] self.attractor_states[tuple(threshold(np.mean(states, 0), 0)) ] = lbl self.initialize_weights() def initialize_weights(self): """ Initialize the weights of the network. Initialization is done based off the attractor states. """ states = np.array(self.attractor_states.keys()).T self.weights = np.zeros((states.shape[0], states.shape[0])) for i in xrange(states.shape[0]): for j in xrange(states.shape[0]): self.weights[i][j] = np.dot(states[i], states[j]) / float( states.shape[0]) def step(self, x): """ Compute a single step of the network. @param x: The input data for this step. @return: The found attractor state. """ return threshold(np.inner(self.weights, x), 0)
tehtechguy/annt
src/annt/examples/multilayer_perceptron_network.py
# multilayer_perceptron_network.py # # Author : <NAME> # Contact : http://techtorials.me # Date Created : 03/31/15 # # Description : Example showing how to create and use a multilayer # perceptron network. # Python Version : 2.7.8 # # License : MIT License http://opensource.org/licenses/mit-license.php # Copyright : (c) 2015 <NAME> """ Example showing how to create and use a multilayer perceptron network. This example uses a reduced set of data from the U{MNIST<http://yann.lecun.com/exdb/mnist/>} dataset. G{packagetree annt} """ __docformat__ = 'epytext' # Native imports import os # Third party imports import numpy as np # Program imports from annt.util import one_hot, mnist_data from annt.net import MultilayerPerception from annt.plot import plot_epoch def main(train_data, train_labels, test_data, test_labels, nepochs=1, plot=True, verbose=True, hidden_layers=[100], bias=1, learning_rate=0.001, min_weight=-1, max_weight=1, hidden_activation_type='sigmoid'): """ Demonstrates a multilayer perceptron network using MNIST. @param train_data: The data to train with. This must be an iterable returning a numpy array. @param train_labels: The training labels. This must be an iterable with the same length as train_data. @param test_data: The data to test with. This must be an iterable returning a numpy array. @param test_labels: The testing labels. This must be an iterable with the same length as train_data. @param nepochs: The number of training epochs to perform. @param plot: If True, a plot will be created. @param verbose: If True, the network will print results after every iteration. @param hidden_layers: A sequence denoting the number of hidden layers and the number of nodes per hidden layer. For example, [100, 50] will result in the creation of two hidden layers with the first layer having 100 nodes and the second layer having 50 nodes. There is no limit to the number of layers and nodes. @param bias: The bias input. Set to "0" to disable. @param learning_rate: The learning rate to use. @param min_weight: The minimum weight value. @param max_weight: The maximum weight value. @param hidden_activation_type: The type activation function to use for the hidden neurons. This must be one of the classes implemented in L{annt.activation}. @return: A tuple containing the training and testing results, respectively. """ # Create the network shape = [train_data.shape[1]] + list(hidden_layers) + [10] net = MultilayerPerception( shape = shape, bias = bias, learning_rate = learning_rate, min_weight = min_weight, max_weight = max_weight, hidden_activation_type = hidden_activation_type ) # Simulate the network train_results, test_results = net.run(train_data, train_labels, test_data, test_labels, nepochs, verbose) # Plot the results if plot: plot_epoch(y_series=(train_results * 100, test_results * 100), series_names=('Train', 'Test'), y_label='Accuracy [%]', title='MLP - Example', legend_location='upper left') return train_results * 100, test_results * 100 def basic_sim(nepochs=100): """ Perform a basic simulation. @param nepochs: The number of training epochs to perform. """ # Get the data (train_data, train_labels), (test_data, test_labels) = mnist_data() # Scale pixel values to be between 0 and 1 # Convert labeled data to one-hot encoding # Run the network main(train_data/255., one_hot(train_labels, 10), test_data/255., one_hot(test_labels, 10), nepochs=nepochs) def bulk(niters, nepochs, verbose=True, plot=True, **kargs): """ Execute the main network across many networks. @param niters: The number of iterations to run for statistical purposes. @param nepochs: The number of training epochs to perform. @param verbose: If True, a simple iteration status will be printed. @param plot: If True, a plot will be generated. @param kargs: Any keyword arguments to pass to the main network simulation. @return: A tuple containing: (train_mean, train_std), (test_mean, test_std) """ # Simulate the network train_results = np.zeros((niters, nepochs)) test_results = np.zeros((niters, nepochs)) for i in xrange(niters): if verbose: print 'Executing iteration {0} of {1}'.format(i + 1, niters) train_results[i], test_results[i] = main(verbose=False, plot=False, nepochs=nepochs, **kargs) # Compute the mean costs train_mean = np.mean(train_results, 0) test_mean = np.mean(test_results, 0) # Compute the standard deviations train_std = np.std(train_results, 0) test_std = np.std(test_results, 0) if plot: plot_epoch(y_series=(train_mean, test_mean), legend_location='upper left', series_names=('Train', 'Test'), y_errs=(train_std, test_std), y_label='Accuracy [%]', title='MLP - Stats Example') return (train_mean, train_std), (test_mean, test_std) def bulk_sim(nepochs=100, niters=10): """ Perform a simulation across multiple iterations, for statistical purposes. @param nepochs: The number of training epochs to perform. @param niters: The number of iterations to run for statistical purposes. """ (train_data, train_labels), (test_data, test_labels) = mnist_data() bulk(train_data=train_data/255., train_labels=one_hot(train_labels, 10), test_data=test_data/255., test_labels=one_hot(test_labels, 10), nepochs=nepochs, niters=niters) def vary_params(out_dir, nepochs=100, niters=10, show_plot=True): """ Vary some parameters and generate some plots. @param out_dir: The directory to save the plots in. @param nepochs: The number of training epochs to perform. @param niters: The number of iterations to run for statistical purposes. @param show_plot: If True the plot will be displayed upon creation. """ # Get the data (train_data, train_labels), (test_data, test_labels) = mnist_data() train_d = train_data/255.; train_l = one_hot(train_labels, 10) test_d = test_data/255.; test_l = one_hot(test_labels, 10) # Make the output directory try: os.makedirs(out_dir) except OSError: pass ########################################################################### ###### Vary number of nodes for one layer ########################################################################### print 'Varying one layer' shapes = [[int(x)] for x in np.linspace(25, 250, 10)] train_results = np.zeros((len(shapes), nepochs)) train_stds = np.zeros((len(shapes), nepochs)) test_results = np.zeros((len(shapes), nepochs)) test_stds = np.zeros((len(shapes), nepochs)) series_names = ['Shape = {0}'.format(x) for x in shapes] for i, shape in enumerate(shapes): print 'Executing iteration {0} of {1}'.format(i + 1, len(shapes)) (train_results[i], train_stds[i]), (test_results[i], test_stds[i]) = \ bulk(train_data=train_d, train_labels=train_l, test_data=test_d, test_labels=test_l, plot=False, nepochs=nepochs, niters=niters, hidden_layers=shape, verbose=False) # Make training plot title = 'MLP - Training\n10 Iterations, Single Hidden Layer' out_path = os.path.join(out_dir, 'single-train.png') plot_epoch(y_series=train_results, series_names=series_names, title=title, y_errs=train_stds, y_label='Accuracy [%]', out_path=out_path, legend_location='upper left', show=show_plot) # Make testing plot title = 'MLP - Testing\n10 Iterations, Single Hidden Layer' out_path = os.path.join(out_dir, 'single-test.png') plot_epoch(y_series=test_results, series_names=series_names, title=title, y_errs=train_stds, y_label='Accuracy [%]', out_path=out_path, legend_location='upper left', show=show_plot) ########################################################################### ###### Vary number of nodes for two layers (only second layer varied) ########################################################################### print '\nVarying two layers' shapes = [[250, int(x)] for x in np.linspace(10, 100, 10)] train_results = np.zeros((len(shapes), nepochs)) train_stds = np.zeros((len(shapes), nepochs)) test_results = np.zeros((len(shapes), nepochs)) test_stds = np.zeros((len(shapes), nepochs)) series_names = ['Shape = {0}'.format(x) for x in shapes] for i, shape in enumerate(shapes): print 'Executing iteration {0} of {1}'.format(i + 1, len(shapes)) (train_results[i], train_stds[i]), (test_results[i], test_stds[i]) = \ bulk(train_data=train_d, train_labels=train_l, test_data=test_d, test_labels=test_l, plot=False, nepochs=nepochs, niters=niters, hidden_layers=shape, verbose=False) # Make training plot title = 'MLP - Training\n10 Iterations, Double Hidden Layer' out_path = os.path.join(out_dir, 'double-train.png') plot_epoch(y_series=train_results, series_names=series_names, title=title, y_errs=train_stds, y_label='Accuracy [%]', out_path=out_path, legend_location='upper left', show=show_plot) # Make testing plot title = 'MLP - Testing\n10 Iterations, Double Hidden Layer' out_path = os.path.join(out_dir, 'double-test.png') plot_epoch(y_series=test_results, series_names=series_names, title=title, y_errs=train_stds, y_label='Accuracy [%]', out_path=out_path, legend_location='upper left', show=show_plot) ########################################################################### ###### Vary number of nodes for three layers (only third layer varied) ########################################################################### print '\nVarying three layers' shapes = [[250, 100, int(x)] for x in np.linspace(5, 50, 10)] train_results = np.zeros((len(shapes), nepochs)) train_stds = np.zeros((len(shapes), nepochs)) test_results = np.zeros((len(shapes), nepochs)) test_stds = np.zeros((len(shapes), nepochs)) series_names = ['Shape = {0}'.format(x) for x in shapes] for i, shape in enumerate(shapes): print 'Executing iteration {0} of {1}'.format(i + 1, len(shapes)) (train_results[i], train_stds[i]), (test_results[i], test_stds[i]) = \ bulk(train_data=train_d, train_labels=train_l, test_data=test_d, test_labels=test_l, plot=False, nepochs=nepochs, niters=niters, hidden_layers=shape, verbose=False) # Make training plot title = 'MLP - Training\n10 Iterations, Triple Hidden Layer' out_path = os.path.join(out_dir, 'triple-train.png') plot_epoch(y_series=train_results, series_names=series_names, title=title, y_errs=train_stds, y_label='Accuracy [%]', out_path=out_path, legend_location='upper left', show=show_plot) # Make testing plot title = 'MLP - Testing\n10 Iterations, Triple Hidden Layer' out_path = os.path.join(out_dir, 'triple-test.png') plot_epoch(y_series=test_results, series_names=series_names, title=title, y_errs=train_stds, y_label='Accuracy [%]', out_path=out_path, legend_location='upper left', show=show_plot) if __name__ == '__main__': basic_sim() # bulk_sim() # vary_params(out_dir=r'D:\annt\test\MLP_Network', show_plot=False)
tehtechguy/annt
src/annt/experimental/__init__.py
<gh_stars>1-10 # __init__.py # # Author : <NAME> # Contact : http://techtorials.me # Date Created : 04/07/15 # # Description : Defines the experimental package # Python Version : 2.7.8 # # License : MIT License http://opensource.org/licenses/mit-license.php # Copyright : (c) 2015 <NAME> """ This package contains any experimental code. """ __docformat__ = 'epytext'
tehtechguy/annt
src/annt/examples/competitive_learning_network.py
<reponame>tehtechguy/annt<gh_stars>1-10 # competitive_learning_network.py # # Author : <NAME> # Contact : http://techtorials.me # Date Created : 04/02/15 # # Description : Example showing how to create and use a competitive learning # network. # Python Version : 2.7.8 # # License : MIT License http://opensource.org/licenses/mit-license.php # Copyright : (c) 2015 <NAME> """ Example showing how to create and use a competitive learning network. This example uses a reduced set of data from the U{MNIST<http://yann.lecun.com/exdb/mnist/>} dataset. G{packagetree annt} """ __docformat__ = 'epytext' # Native imports import os # Third party imports import numpy as np # Program imports from annt.util import mnist_data from annt.net import CompetitiveLearning from annt.plot import plot_epoch, plot_weights, plot_surface, make_grid def main(train_data, test_data, nepochs=1, plot=True, verbose=True, nclusters=10, learning_rate=0.001, boost_inc=0.1, boost_dec=0.01, duty_cycle=50, min_duty_cycle=5, min_weight=-1, max_weight=1, nrows=2, ncols=5, shape=(28, 28)): """ Demonstrates a competitive learning network using MNIST. @param train_data: The data to train with. This must be an iterable returning a numpy array. @param test_data: The data to test with. This must be an iterable returning a numpy array. @param nepochs: The number of training epochs to perform. @param plot: If True, a plot will be created. @param verbose: If True, the network will print results after every iteration. @param nclusters: The number of clusters. @param learning_rate: The learning rate to use. @param boost_inc: The amount to increment the boost by. @param boost_dec: The amount to decrement the boost by. @param duty_cycle: The history to retain for activations for each node. This is the period minimum activation is compared across. It is a rolling window. @param min_duty_cycle: The minimum duty cycle. If a node has not been active at least this many times, increment its boost value, else decrement it. @param min_weight: The minimum weight value. @param max_weight: The maximum weight value. @param nrows: The number of rows of plots to create for the clusters. @param ncols: The number of columns of plots to create for the clusters. @param shape: The shape of the weights. It is assumed that a 1D shape was used and is desired to be represented in 2D. Whatever shape is provided will be used to reshape the weights. For example, if you had a 28x28 image and each weight corresponded to one pixel, you would have a vector with a shape of (784, ). This vector would then need to be resized to your desired shape of (28, 28). @return: A tuple containing the training results, testing results, and weights, respectively. """ # Create the network net = CompetitiveLearning( ninputs = train_data.shape[1], nclusters = nclusters, learning_rate = learning_rate, boost_inc = boost_inc, boost_dec = boost_dec, duty_cycle = duty_cycle, min_duty_cycle = min_duty_cycle, min_weight = min_weight, max_weight = max_weight ) # Simulate the network train_results, test_results = net.run(train_data, test_data, nepochs, verbose) # Plot the results and clusters if plot: plot_epoch(y_series=(train_results, test_results), series_names=('Train', 'Test'), y_label='Cost', title='Clustering - Example', semilog=True) plot_weights(net.weights.T, nrows, ncols, shape, 'Clustering Weights - Example') return train_results, test_results, net.weights.T def basic_sim(nepochs=100): """ Perform a basic simulation. @param nepochs: The number of training epochs to perform. """ # Get the data (train_data, train_labels), (test_data, test_labels) = mnist_data() # Scale pixel values to be between 0 and 1 # Convert labeled data to one-hot encoding # Run the network main(train_data/255., test_data/255., nepochs=nepochs) def bulk(niters, nepochs, verbose=True, plot=True, **kargs): """ Execute the main network across many networks. @param niters: The number of iterations to run for statistical purposes. @param nepochs: The number of training epochs to perform. @param verbose: If True, a simple iteration status will be printed. @param plot: If True, a plot will be generated. @param kargs: Any keyword arguments to pass to the main network simulation. @return: A tuple containing: (train_mean, train_std), (test_mean, test_std) """ # Simulate the network train_results = np.zeros((niters, nepochs)) test_results = np.zeros((niters, nepochs)) for i in xrange(niters): if verbose: print 'Executing iteration {0} of {1}'.format(i + 1, niters) train_results[i], test_results[i], _ = main(verbose=False, plot=False, nepochs=nepochs, **kargs) # Compute the mean costs train_mean = np.mean(train_results, 0) test_mean = np.mean(test_results, 0) # Compute the standard deviations train_std = np.std(train_results, 0) test_std = np.std(test_results, 0) if plot: plot_epoch(y_series=(train_mean, test_mean), legend_location='upper left', series_names=('Train', 'Test'), y_errs=(train_std, test_std), y_label='Cost [%]', title='Clustering - Stats Example') return (train_mean, train_std), (test_mean, test_std) def bulk_sim(nepochs=100, niters=10): """ Perform a simulation across multiple iterations, for statistical purposes. @param nepochs: The number of training epochs to perform. @param niters: The number of iterations to run for statistical purposes. """ (train_data, train_labels), (test_data, test_labels) = mnist_data() bulk(train_data=train_data/255., test_data=test_data/255., nepochs=nepochs, niters=niters) def vary_params(out_dir, nepochs=100, niters=10, show_plot=True): """ Vary some parameters and generate some plots. @param out_dir: The directory to save the plots in. @param nepochs: The number of training epochs to perform. @param niters: The number of iterations to run for statistical purposes. @param show_plot: If True the plot will be displayed upon creation. """ # Get the data (train_data, train_labels), (test_data, test_labels) = mnist_data() train_d = train_data/255.; test_d = test_data/255. # Make the output directory try: os.makedirs(out_dir) except OSError: pass ########################################################################### ###### Vary number of clusters ########################################################################### print 'Varying number of clusters' nclusters = np.arange(5, 55, 5) weight_plot_params = [{'nrows':1, 'ncols':5}, {'nrows':2, 'ncols':5}, {'nrows':3, 'ncols':5}, {'nrows':4, 'ncols':5}, {'nrows':5, 'ncols':5}, {'nrows':3, 'ncols':10}, {'nrows':4, 'ncols':10}, {'nrows':4, 'ncols':10}, {'nrows':5, 'ncols':10}, {'nrows':5, 'ncols':10}] weight_results = [] train_results = np.zeros((len(nclusters), nepochs)) train_stds = np.zeros((len(nclusters), nepochs)) test_results = np.zeros((len(nclusters), nepochs)) test_stds = np.zeros((len(nclusters), nepochs)) series_names = ['Clusters = {0}'.format(x) for x in nclusters] for i, ncluster in enumerate(nclusters): print 'Executing iteration {0} of {1}'.format(i + 1, len(nclusters)) (train_results[i], train_stds[i]), (test_results[i], test_stds[i]) = \ bulk(train_data=train_d, test_data=test_d, plot=False, nepochs=nepochs, nclusters=ncluster, verbose=False, niters=niters) x, y, weights = main(train_data=train_d, test_data=test_d, plot=False, nepochs=nepochs, nclusters=ncluster, verbose=False) weight_results.append(weights) # Make training plot title = 'Competitive Learning Network - Training\n10 Iterations, ' \ 'Varying Number of Clusters' out_path = os.path.join(out_dir, 'clusters-train.png') plot_epoch(y_series=train_results, series_names=series_names, title=title, y_errs=train_stds, y_label='Cost', out_path=out_path, semilog=True, show=show_plot) # Make testing plot title = 'Competitive Learning Network - Testing\n10 Iterations, ' \ 'Varying Number of Clusters' out_path = os.path.join(out_dir, 'clusters-test.png') plot_epoch(y_series=test_results, series_names=series_names, title=title, y_errs=test_stds, y_label='Cost', out_path=out_path, semilog=True, show=show_plot) # Make weight plots title = 'Competitive Learning Network - Weights\n{0} Clusters' for weights, params, ncluster in zip(weight_results, weight_plot_params, nclusters): out_path = os.path.join(out_dir, 'weights-{0}.png'.format(ncluster)) plot_weights(weights=weights, title=title.format(ncluster), out_path=out_path, show=show_plot, shape=(28, 28), **params) ########################################################################### ###### Vary boost increment and decrement ########################################################################### print 'Varying boost increment and decrement amounts' space = np.linspace(0.001, .1, 100) boost_pairs = np.array([(x, y) for x in space for y in space]) train_results = np.zeros((len(boost_pairs), nepochs)) train_stds = np.zeros((len(boost_pairs), nepochs)) test_results = np.zeros((len(boost_pairs), nepochs)) test_stds = np.zeros((len(boost_pairs), nepochs)) for i, pair in enumerate(boost_pairs): print 'Executing iteration {0} of {1}'.format(i + 1, len(boost_pairs)) (train_results[i], train_stds[i]), (test_results[i], test_stds[i]) = \ bulk(train_data=train_d, test_data=test_d, plot=False, nepochs=nepochs, boost_inc=pair[0], boost_dec=pair[1], verbose=False, niters=niters) # Make training plot at last epoch title = 'Competitive Learning Network - Training\n10 Iterations, ' \ 'Epoch {0}, Varying Boost Increment and Decrement'.format(nepochs) out_path = os.path.join(out_dir, 'boost-train.png') plot_surface(*make_grid(np.array([boost_pairs.T[0], boost_pairs.T[1], train_results.T[-1]]).T), x_label='Boost Increment', out_path=out_path, y_label='Boost Decrement', title=title, show=show_plot, z_label='Cost') # Make testing plot at last epoch title = 'Competitive Learning Network - Testing\n10 Iterations, ' \ 'Epoch {0}, Varying Boost Increment and Decrement'.format(nepochs) out_path = os.path.join(out_dir, 'boost-test.png') plot_surface(*make_grid(np.array([boost_pairs.T[0], boost_pairs.T[1], test_results.T[-1]]).T), x_label='Boost Increment', out_path=out_path, y_label='Boost Decrement', title=title, show=show_plot, z_label='Cost') if __name__ == '__main__': basic_sim() # bulk_sim() # vary_params(out_dir=r'D:\annt\test\Clustering_Network', show_plot=False)
tehtechguy/annt
src/annt/net.py
<reponame>tehtechguy/annt<gh_stars>1-10 # net.py # # Author : <NAME> # Contact : http://techtorials.me # Date Created : 03/30/15 # # Description : Module for various artificial neural networks. # Python Version : 2.7.8 # # License : MIT License http://opensource.org/licenses/mit-license.php # Copyright : (c) 2015 <NAME> """ Module for various artificial neural networks. G{packagetree annt} """ __docformat__ = 'epytext' # Native imports from abc import ABCMeta, abstractmethod # Third party imports import numpy as np # Program imports from annt.activation import create_activation from annt.timers import MultiTimer, pretty_time ############################################################################### ########## Class Templates ############################################################################### class Net(object): """ Base class description for a network. """ __metaclass__ = ABCMeta @abstractmethod def __init__(self): """ Initialize the class instance. """ @abstractmethod def step(self, x, y=None): """ Compute a single step of the network. @param x: The input data to compute for this step. @param y: The expected output. """ @abstractmethod def run(self, train_data, train_labels, test_data, test_labels, nepochs=1): """ Simulate the entire network. @param train_data: The data to train with. This must be an iterable returning a numpy array. @param train_labels: The training labels. This must be an iterable with the same length as train_data. @param test_data: The data to test with. This must be an iterable returning a numpy array. @param test_labels: The testing labels. This must be an iterable with the same length as train_data. @param nepochs: The number of training epochs to perform. @return: A tuple containing the training and test costs/accuracies. """ @abstractmethod def initialize_weights(self, shape, min_weight=-1, max_weight=1): """ Initialize the weights of the network. Initialization is done randomly. @param shape: The number of nodes in the entire network, excluding any bias terms. This parameter must be a sequence. @param min_weight: The minimum weight value. @param max_weight: The maximum weight value. """ def _run(self, x, y=None): """ Execute the network for this batch of data @param x: The input data to compute for this step. @param y: The expected output. @return: The expected outputs for each input. """ if y is None: return np.array(map(self.step, x)) else: return np.array(map(self.step, x, y)) def enable_learning(self): """ Enables learning for the network. """ self.learning = True def disable_learning(self): """ Disables learning for the network. """ self.learning = False class SupervisedCostNet(Net): """ Base class description for network that computes a cost function using unsupervised training. """ __metaclass__ = ABCMeta @abstractmethod def cost(self, y, y_exp): """ Compute the cost function @param y: The true output. @param y_exp: The expected output. @return: The cost. """ def _print_train(self, epoch, nepochs, train_cost): """ Print the details about training. @param epoch: The current epoch number. @param nepochs: The total number of epochs. @param train_cost: A numpy array containing the training cost at each epoch. """ print '\nEpoch {0} of {1}:'.format(epoch + 1, nepochs) print ' Training Cost : {0}'.format(train_cost[epoch]) print ' Training Time : {0}'.format( self.timers.get_elapsed_time('train_epoch', True)) def _print_test(self, epoch, test_cost): """ Print the details about testing. @param epoch: The current epoch number. @param test_cost: A numpy array containing the testing cost at each epoch. """ print ' Testing Cost : {0}'.format(test_cost[epoch]) print ' Testing Time : {0}'.format( self.timers.get_elapsed_time('test_epoch', True)) def _print_final(self, nepochs, train_cost, test_cost): """ Print the details final details. @param nepochs: The total number of epochs. @param train_cost: A numpy array containing the training cost at each epoch. @param test_cost: A numpy array containing the testing cost at each epoch. """ print '\n' + '*' * 79 print '\nBest Training Cost : {0} at Epoch {1}'.format(np.min( train_cost), np.argmin(train_cost) + 1) print 'Best Testing Cost : {0} at Epoch {1}'.format(np.min( test_cost), np.argmin(test_cost) + 1) print '\nTotal Execution Time : {0}'.format( self.timers.get_elapsed_time('global', True)) print 'Total Training Time : {0}'.format( self.timers.get_elapsed_time('train', True)) print 'Average Training Epoch Time : {0}'.format( pretty_time(self.timers.get_elapsed_time('train') / nepochs)) print 'Total Testing Time : {0}'.format( self.timers.get_elapsed_time('test', True)) print 'Average Testing Epoch Time : {0}'.format( pretty_time(self.timers.get_elapsed_time('test') / nepochs)) def run(self, train_data, train_labels, test_data, test_labels, nepochs=1, verbose=True): """ Simulate the entire network. @param train_data: The data to train with. This must be an iterable returning a numpy array. @param train_labels: The training labels. This must be an iterable with the same length as train_data. @param test_data: The data to test with. This must be an iterable returning a numpy array. @param test_labels: The testing labels. This must be an iterable with the same length as train_data. @param nepochs: The number of training epochs to perform. @param verbose: If True, details will be printed after each epoch. @return: A tuple containing the training and test costs. """ # Make some timers self.timers = MultiTimer() self.timers.add_timers('global', 'train', 'train_epoch', 'test', 'test_epoch') self.timers.stop_timers('train', 'train_epoch', 'test', 'test_epoch') _run = self._run; cost = self.cost train_cost = np.zeros(nepochs); test_cost = np.zeros(nepochs) for i in xrange(nepochs): # Compute training cost self.timers.start_timers('train', 'train_epoch') self.enable_learning() train_cost[i] = cost(_run(train_data, train_labels), train_labels) self.timers.pause_timers('train') self.timers.stop_timers('train_epoch') if verbose: self._print_train(i, nepochs, train_cost) # Compute testing cost self.timers.start_timers('test', 'test_epoch') self.disable_learning() test_cost[i] = cost(_run(test_data), test_labels) self.timers.pause_timers('test') self.timers.stop_timers('test_epoch') if verbose: self._print_test(i, test_cost) self.timers.stop_timers('global') if verbose: self._print_final(nepochs, train_cost, test_cost) return (train_cost, test_cost) class UnsupervisedCostNet(SupervisedCostNet): """ Base class description for network that computes a cost function using unsupervised training. """ __metaclass__ = ABCMeta def run(self, train_data, test_data, nepochs=1, verbose=True): """ Simulate the entire network. @param train_data: The data to train with. This must be an iterable returning a numpy array. @param test_data: The data to test with. This must be an iterable returning a numpy array. @param nepochs: The number of training epochs to perform. @param verbose: If True, details will be printed after each epoch. @return: A tuple containing the training and test costs. """ # Make some timers self.timers = MultiTimer() self.timers.add_timers('global', 'train', 'train_epoch', 'test', 'test_epoch') self.timers.stop_timers('train', 'train_epoch', 'test', 'test_epoch') _run = self._run; cost = self.cost train_cost = np.zeros(nepochs); test_cost = np.zeros(nepochs) for i in xrange(nepochs): # Compute training cost self.timers.start_timers('train', 'train_epoch') self.enable_learning() train_cost[i] = cost(_run(train_data), train_data) self.timers.pause_timers('train') self.timers.stop_timers('train_epoch') if verbose: self._print_train(i, nepochs, train_cost) # Compute testing cost self.timers.start_timers('test', 'test_epoch') self.disable_learning() test_cost[i] = cost(_run(test_data), test_data) self.timers.pause_timers('test') self.timers.stop_timers('test_epoch') if verbose: self._print_test(i, test_cost) self.timers.stop_timers('global') if verbose: self._print_final(nepochs, train_cost, test_cost) return (train_cost, test_cost) class SupervisedAccuracyNet(Net): """ Base class description for network that computes an accuracy using supervised training. """ __metaclass__ = ABCMeta @abstractmethod def score(self, y, y_exp): """ Compute the accuracy. If there is competition between which node should be the winner, a random one is chosen. @param y: The true output. @param y_exp: The expected output. @return: The accuracy. """ def _print_train(self, epoch, nepochs, train_accuracy): """ Print the details about training. @param epoch: The current epoch number. @param nepochs: The total number of epochs. @param train_accuracy: A numpy array containing the training accuracy at each epoch. """ print '\nEpoch {0} of {1}:'.format(epoch + 1, nepochs) print ' Training Accuracy : {0}%'.format(train_accuracy[epoch] * 100) print ' Training Time : {0}'.format( self.timers.get_elapsed_time('train_epoch', True)) def _print_test(self, epoch, test_accuracy): """ Print the details about testing. @param epoch: The current epoch number. @param test_accuracy: A numpy array containing the testing accuracy at each epoch. """ print ' Testing Accuracy : {0}%'.format(test_accuracy[epoch] * 100) print ' Testing Time : {0}'.format( self.timers.get_elapsed_time('test_epoch', True)) def _print_final(self, nepochs, train_accuracy, test_accuracy): """ Print the details final details. @param nepochs: The total number of epochs. @param train_accuracy: A numpy array containing the training accuracy at each epoch. @param test_accuracy: A numpy array containing the testing accuracy at each epoch. """ print '\n' + '*' * 79 print '\nBest Training Accuracy : {0}% at Epoch {1}'.format(np.max( train_accuracy) * 100, np.argmax(train_accuracy) + 1) print 'Best Testing Accuracy : {0}% at Epoch {1}'.format(np.max( test_accuracy) * 100, np.argmax(test_accuracy) + 1) print '\nTotal Execution Time : {0}'.format( self.timers.get_elapsed_time('global', True)) print 'Total Training Time : {0}'.format( self.timers.get_elapsed_time('train', True)) print 'Average Training Epoch Time : {0}'.format( pretty_time(self.timers.get_elapsed_time('train') / nepochs)) print 'Total Testing Time : {0}'.format( self.timers.get_elapsed_time('test', True)) print 'Average Testing Epoch Time : {0}'.format( pretty_time(self.timers.get_elapsed_time('test') / nepochs)) def run(self, train_data, train_labels, test_data, test_labels, nepochs=1, verbose=True): """ Simulate the entire network. @param train_data: The data to train with. This must be an iterable returning a numpy array. @param train_labels: The training labels. This must be an iterable with the same length as train_data. @param test_data: The data to test with. This must be an iterable returning a numpy array. @param test_labels: The testing labels. This must be an iterable with the same length as train_data. @param nepochs: The number of training epochs to perform. @param verbose: If True, details will be printed after each epoch. @return: A tuple containing the training and test accuracies. """ # Make some timers self.timers = MultiTimer() self.timers.add_timers('global', 'train', 'train_epoch', 'test', 'test_epoch') self.timers.stop_timers('train', 'train_epoch', 'test', 'test_epoch') _run = self._run; score = self.score train_accuracy = np.zeros(nepochs); test_accuracy = np.zeros(nepochs) for i in xrange(nepochs): # Compute training accuracy self.timers.start_timers('train', 'train_epoch') self.enable_learning() train_accuracy[i] = score(_run(train_data, train_labels), train_labels) self.timers.pause_timers('train') self.timers.stop_timers('train_epoch') if verbose: self._print_train(i, nepochs, train_accuracy) # Compute testing accuracy self.timers.start_timers('test', 'test_epoch') self.disable_learning() test_accuracy[i] = score(_run(test_data), test_labels) self.timers.pause_timers('test') self.timers.stop_timers('test_epoch') if verbose: self._print_test(i, test_accuracy) self.timers.stop_timers('global') if verbose: self._print_final(nepochs, train_accuracy, test_accuracy) return (train_accuracy, test_accuracy) ############################################################################### ########## Class Implementation ############################################################################### class LinearRegressionNetwork(SupervisedCostNet): """ Base class for a liner regression network. """ def __init__(self, ninputs, bias=1, learning_rate=0.001, min_weight=-1, max_weight=1, activation_type='linear', activation_kargs={}, learning=True): """ Initializes this linear regression network. @param ninputs: The number of inputs (excluding the bias). @param bias: The bias input. Set to "0" to disable. @param learning_rate: The learning rate to use. @param min_weight: The minimum weight value. @param max_weight: The maximum weight value. @param activation_type: The type activation function to use. This must be one of the classes implemented in L{annt.activation}. @param activation_kargs: Any keyword arguments for the activation function. @param learning: Boolean denoting if the network is currently learning. """ # Store the params self.bias = np.array([bias]) self.learning_rate = learning_rate self.learning = learning # Initialize the activation function self.activation = create_activation(activation_type, **activation_kargs) # Construct the weights self.initialize_weights((ninputs + 1,), min_weight, max_weight) def initialize_weights(self, shape, min_weight=-1, max_weight=1): """ Initialize the weights of the network. Initialization is done randomly. @param shape: The number of nodes in the entire network, excluding any bias terms. This parameter must be a sequence. @param min_weight: The minimum weight value. @param max_weight: The maximum weight value. """ self.weights = np.random.uniform(min_weight, max_weight, shape[0]) def cost(self, y, y_exp): """ Compute the cost function @param y: The true output. @param y_exp: The expected output. @return: The cost. """ return np.sum(np.power((y_exp - y), 2)) / (2. * y_exp.shape[0]) def step(self, x, y=None): """ Compute a single step of the network. @param x: The input data to compute for this step. This must be a numpy array with a shape of (self.ninputs, ). @param y: The expected output. @return: The computed output. """ # Add bias full_x = np.concatenate((self.bias, x)) # Calculate the outputs y_est = self.activation.compute(np.dot(full_x, self.weights)) # Update the error using online learning if self.learning: self.weights += self.learning_rate * full_x * (y - y_est) return y_est class MultilayerPerception(SupervisedAccuracyNet): """ Base class for a multilayer perception. """ def __init__(self, shape, bias=1, learning_rate=0.001, min_weight=-1, max_weight=1, input_activation_type='linear', input_activation_kargs={'m':1}, hidden_activation_type='sigmoid', hidden_activation_kargs={}, learning=True): """ Initializes this multilayer perception network. @param shape: The number of layers and the number of nodes per layer (excluding the bias). @param bias: The bias input. This is applied to all input and hidden layers. Set to "0" to disable. @param learning_rate: The learning rate to use. @param min_weight: The minimum weight value. @param max_weight: The maximum weight value. @param input_activation_type: The type activation function to use for the input layer. This must be one of the classes implemented in L{annt.activation}. @param input_activation_kargs: Any keyword arguments for the input activation function. @param hidden_activation_type: The type activation function to use for the hidden layer. This must be one of the classes implemented in L{annt.activation}. @param hidden_activation_kargs: Any keyword arguments for the hidden activation function. @param learning: Boolean denoting if the network is currently learning. """ # Store the params self.bias = np.array([bias]) self.learning_rate = learning_rate self.learning = learning # Initialize the activation functions self.input_activation = create_activation(input_activation_type, **input_activation_kargs) self.hidden_activation = create_activation(hidden_activation_type, **hidden_activation_kargs) # Construct the weights self.initialize_weights(shape, min_weight, max_weight) # Construct the internal outputs and deltas new_shape = [1 + s for s in shape[:-1]] + [shape[-1]] self.outputs = np.array([np.zeros(s) for s in new_shape]) self.deltas = self.outputs.copy() def initialize_weights(self, shape, min_weight=-1, max_weight=1): """ Initialize the weights of the network. Initialization is done randomly. @param shape: The number of nodes in the entire network, excluding any bias terms. This parameter must be a sequence. @param min_weight: The minimum weight value. @param max_weight: The maximum weight value. """ # Input weights aren't trained, so they are ignored. All other weights # are set to be random. The Last dimension is incremented by 1 to allow # for the bias. self.weights = np.array([np.random.uniform(min_weight, max_weight, (c, p + 1)) for c, p in zip(shape[1:], shape[:-1])]) def score(self, y, y_exp): """ Compute the accuracy. If there is competition between which node should be the winner, a random one is chosen. @param y: The true output. @param y_exp: The expected output. @return: The accuracy. """ accuracy = 0. for predicted, expected in zip(y, y_exp): indexes = np.where(predicted == np.max(predicted))[0] np.random.shuffle(indexes) accuracy += 1 if expected[indexes[0]] == 1 else 0 return accuracy / y_exp.shape[0] def step(self, x, y=None): """ Compute a single step of the network. @param x: The input data to compute for this step. @param y: The expected output. @return: The computed outputs from the output layer. """ ####################################################################### ######## Calculate the outputs using forward propagation ####################################################################### # Calculate the outputs for the input layer self.outputs[0][0] = self.input_activation.compute(self.bias) self.outputs[0][1:] = self.input_activation.compute(x) # Calculate the outputs for the hidden layer(s) # - First hidden layer -> last hidden layer for layer, layer_weights in enumerate(self.weights[:-1], 1): self.outputs[layer][0] = self.hidden_activation.compute(self.bias) self.outputs[layer][1:] = self.hidden_activation.compute(np.inner( self.outputs[layer - 1], layer_weights)) # Calculate the outputs for the output layer self.outputs[-1] = self.hidden_activation.compute(np.inner( self.outputs[-2], self.weights[-1])) ####################################################################### ######## Train the network using backpropagation ####################################################################### if self.learning: ################################################################### ######## Calculate the deltas ################################################################### # Calculate output deltas self.deltas[-1] = self.outputs[-1] - y # Calculate hidden deltas # - Last hidden layer -> first hidden layer # Note that deltas are not computed for the bias for layer in xrange(-2, -self.deltas.shape[0], -1): self.deltas[layer] = self.hidden_activation.compute_derivative( self.outputs[layer][1:,]) * np.inner(self.deltas[layer + 1], self.weights[layer + 1].T[1:,]) ################################################################### ######## Update the weights ################################################################### # Update the weights # - Output -> first hidden layer # - Bias's weight -> last node's weight for layer in xrange(-1, -self.deltas.shape[0], -1): for i in xrange(self.weights[layer].shape[0]): self.weights[layer][i] += -self.learning_rate * \ self.deltas[layer][i] * self.outputs[layer - 1] # Return the outputs from the output layer return self.outputs[-1] class ExtremeLearningMachine(MultilayerPerception): """ Base class for an extreme learning machine. """ def step(self, x, y=None): """ Compute a single step of the network. @param x: The input data to compute for this step. @param y: The expected output. """ ####################################################################### ######## Calculate the outputs using forward propagation ####################################################################### # Calculate the outputs for the input layer self.outputs[0][0] = self.input_activation.compute(self.bias) self.outputs[0][1:] = self.input_activation.compute(x) # Calculate the outputs for the hidden layer(s) # - First hidden layer -> last hidden layer for layer, layer_weights in enumerate(self.weights[:-1], 1): self.outputs[layer][0] = self.hidden_activation.compute(self.bias) self.outputs[layer][1:] = self.hidden_activation.compute(np.inner( self.outputs[layer - 1], layer_weights)) # Calculate the outputs for the output layer self.outputs[-1] = self.hidden_activation.compute(np.inner( self.outputs[-2], self.weights[-1])) ####################################################################### ######## Train the network using backpropagation ####################################################################### if self.learning: ################################################################### ######## Calculate the deltas ################################################################### # Calculate output deltas self.deltas[-1] = self.outputs[-1] - y ################################################################### ######## Update the weights ################################################################### # Update the output weights for i in xrange(self.weights[-1].shape[0]): self.weights[-1][i] += -self.learning_rate * \ self.deltas[-1][i] * self.outputs[-2] # Return the outputs from the output layer return self.outputs[-1] class CompetitiveLearning(UnsupervisedCostNet): """ Base class for a competitive learning network (clustering). """ def __init__(self, ninputs, nclusters, learning_rate=0.001, boost_inc=0.1, boost_dec=0.01, duty_cycle=50, min_duty_cycle=5, min_weight=-1, max_weight=1, learning=True): """ Initializes this competitive learning network. @param ninputs: The number of inputs to the network. @param nclusters: The number of clusters. @param learning_rate: The learning rate to use. @param boost_inc: The amount to increment the boost by. @param boost_dec: The amount to decrement the boost by. @param duty_cycle: The history to retain for activations for each node. This is the period minimum activation is compared across. It is a rolling window. @param min_duty_cycle: The minimum duty cycle. If a node has not been active at least this many times, increment its boost value, else decrement it. @param min_weight: The minimum weight value. @param max_weight: The maximum weight value. @param learning: Boolean denoting if the network is currently learning. """ # Store the params self.learning_rate = learning_rate self.boost_inc = boost_inc self.boost_dec = boost_dec self.duty_cycle = duty_cycle self.min_duty_cycle = min_duty_cycle self.learning = learning # Construct the weights self.initialize_weights((ninputs, nclusters), min_weight, max_weight) # Construct the boost values self.boost = np.ones(nclusters) # Construct the outputs # - Each item represents a single cluster. # - Each cluster maintains a history of length two of its update. # - The first item refers to the current iteration. # - The second item refers to the previous iteration. # - The last item refers to the furthest iteration. self.outputs = np.ones((nclusters, duty_cycle)) def _update_boost(self): """ Update the boost values. """ for i, active in enumerate(self.outputs): if int(np.sum(active)) >= self.min_duty_cycle: self.boost[i] += self.boost_inc else: self.boost[i] = max(self.boost[i] - self.boost_dec, 0) def initialize_weights(self, shape, min_weight=-1, max_weight=1): """ Initialize the weights of the network. Initialization is done randomly. @param shape: The number of nodes in the entire network. This parameter must be a sequence. @param min_weight: The minimum weight value. @param max_weight: The maximum weight value. """ self.weights = np.random.uniform(min_weight, max_weight, shape) def cost(self, y, x): """ Compute the cost function @param y: The true output. @param x: The input. @return: The cost. """ cost = 0. for i, pattern in enumerate(x): for j, weights in enumerate(self.weights.T): cost += y[i][j] * np.sum((weights - pattern) ** 2) return cost / 2 def step(self, x): """ Compute a single step of the network. @param x: The input data to compute for this step. @return: The computed outputs. """ # Shift outputs self.outputs = np.roll(self.outputs, 1, 1) # Calculate the outputs for i, weights in enumerate(self.weights.T): self.outputs[i][0] = self.boost[i] * np.sum((weights - x) ** 2) min_ix = np.argmin(self.outputs.T[0]) self.outputs.T[0] = 0 self.outputs[min_ix][0] = 1 # Update the boosts self._update_boost() # Train the network if self.learning: for i, weights in enumerate(self.weights.T): self.weights.T[i] += self.learning_rate * self.outputs[i][0] \ * (x - weights) return self.outputs.T[0]
tehtechguy/annt
src/annt/examples/linear_regression_network.py
# linear_regression_network.py # # Author : <NAME> # Contact : http://techtorials.me # Date Created : 03/31/15 # # Description : Example showing how to create and use a linear regression # network. # Python Version : 2.7.8 # # License : MIT License http://opensource.org/licenses/mit-license.php # Copyright : (c) 2015 <NAME> """ Example showing how to create and use a linear regression network. This example uses a reduced set of data from the U{MNIST<http://yann.lecun.com/exdb/mnist/>} dataset. G{packagetree annt} """ __docformat__ = 'epytext' # Native imports import os # Third party imports import numpy as np # Program imports from annt.util import mnist_data from annt.net import LinearRegressionNetwork from annt.plot import plot_epoch def main(train_data, train_labels, test_data, test_labels, nepochs=1, plot=True, verbose=True, bias=1, learning_rate=0.001, min_weight=-1, max_weight=1, activation_type='linear', activation_kargs={'m':1}): """ Demonstrates a linear regression network using MNIST. @param train_data: The data to train with. This must be an iterable returning a numpy array. @param train_labels: The training labels. This must be an iterable with the same length as train_data. @param test_data: The data to test with. This must be an iterable returning a numpy array. @param test_labels: The testing labels. This must be an iterable with the same length as train_data. @param nepochs: The number of training epochs to perform. @param plot: If True, a plot will be created. @param verbose: If True, the network will print results after every iteration. @param bias: The bias input. Set to "0" to disable. @param learning_rate: The learning rate to use. @param min_weight: The minimum weight value. @param max_weight: The maximum weight value. @param activation_type: The type activation function to use. This must be one of the classes implemented in L{annt.activation}. @param activation_kargs: Any keyword arguments for the activation function. @return: A tuple containing the training and testing results, respectively. """ # Create the network net = LinearRegressionNetwork( ninputs = train_data.shape[1], bias = bias, learning_rate = learning_rate, min_weight = min_weight, max_weight = max_weight, activation_type = activation_type, activation_kargs = activation_kargs ) # Simulate the network train_results, test_results = net.run(train_data, train_labels, test_data, test_labels, nepochs, verbose) # Plot the results if plot: plot_epoch(y_series=(train_results, test_results), series_names=('Train', 'Test'), y_label='Cost', title='Linear Regression Network - Example', semilog=True) return train_results, test_results def basic_sim(nepochs=100): """ Perform a basic simulation. @param nepochs: The number of training epochs to perform. """ # Get the data (train_data, train_labels), (test_data, test_labels) = mnist_data() # Scale pixel values to be between 0 and 1 # Scale label values to be between 0 and 1 # Run the network main(train_data/255., train_labels/9., test_data/255., test_labels/9., nepochs=nepochs) def bulk(niters, nepochs, verbose=True, plot=True, **kargs): """ Execute the main network across many networks. @param niters: The number of iterations to run for statistical purposes. @param nepochs: The number of training epochs to perform. @param verbose: If True, a simple iteration status will be printed. @param plot: If True, a plot will be generated. @param kargs: Any keyword arguments to pass to the main network simulation. @return: A tuple containing: (train_mean, train_std), (test_mean, test_std) """ # Simulate the network train_results = np.zeros((niters, nepochs)) test_results = np.zeros((niters, nepochs)) for i in xrange(niters): if verbose: print 'Executing iteration {0} of {1}'.format(i + 1, niters) train_results[i], test_results[i] = main(verbose=False, plot=False, nepochs=nepochs, **kargs) # Compute the mean costs train_mean = np.mean(train_results, 0) test_mean = np.mean(test_results, 0) # Compute the standard deviations train_std = np.std(train_results, 0) test_std = np.std(test_results, 0) if plot: plot_epoch(y_series=(train_mean, test_mean), semilog=True, series_names=('Train', 'Test'), y_errs=(train_std, test_std), y_label='Cost', title='Linear Regression Network - Stats Example') return (train_mean, train_std), (test_mean, test_std) def bulk_sim(nepochs=100, niters=10): """ Perform a simulation across multiple iterations, for statistical purposes. @param nepochs: The number of training epochs to perform. @param niters: The number of iterations to run for statistical purposes. """ (train_data, train_labels), (test_data, test_labels) = mnist_data() bulk(train_data=train_data/255., train_labels=train_labels/9., test_data=test_data/255., test_labels=test_labels/9., nepochs=nepochs, niters=niters) def vary_params(out_dir, nepochs=100, niters=10): """ Vary some parameters and generate some plots. @param out_dir: The directory to save the plots in. @param nepochs: The number of training epochs to perform. @param niters: The number of iterations to run for statistical purposes. """ # Get the data (train_data, train_labels), (test_data, test_labels) = mnist_data() train_d = train_data/255.; train_l = train_labels/9. test_d = test_data/255.; test_l = test_labels/9. # Make the output directory try: os.makedirs(out_dir) except OSError: pass ########################################################################### ###### Vary learning rate ########################################################################### print 'Varying the learning rate' learning_rates = np.linspace(0.001, 0.01, 10) train_results = np.zeros((learning_rates.shape[0], nepochs)) train_stds = np.zeros((learning_rates.shape[0], nepochs)) test_results = np.zeros((learning_rates.shape[0], nepochs)) test_stds = np.zeros((learning_rates.shape[0], nepochs)) series_names = ['Learning Rate = {0}'.format(x) for x in learning_rates] for i, learning_rate in enumerate(learning_rates): print 'Executing iteration {0} of {1}'.format(i + 1, learning_rates.shape[0]) (train_results[i], train_stds[i]), (test_results[i], test_stds[i]) = \ bulk(train_data=train_d, train_labels=train_l, test_data=test_d, test_labels=test_l, plot=False, nepochs=nepochs, niters=niters, learning_rate=learning_rate, verbose=False) # Make training plot title = 'Linear Regression Network - Training\n10 Iterations, ' \ 'Varying Learning Rate' out_path = os.path.join(out_dir, 'learning_rate-train.png') plot_epoch(y_series=train_results, semilog=True, series_names=series_names, y_errs=train_stds, y_label='Cost', title=title, out_path=out_path) # Make testing plot title = 'Linear Regression Network - Testing\n10 Iterations, ' \ 'Varying Learning Rate' out_path = os.path.join(out_dir, 'learning_rate-test.png') plot_epoch(y_series=test_results, semilog=True, series_names=series_names, y_errs=train_stds, y_label='Cost', title=title, out_path=out_path) ########################################################################### ###### Vary slope ########################################################################### print '\nVarying the slope of the linear function' slopes = np.linspace(1, 10, 10) train_results = np.zeros((slopes.shape[0], nepochs)) train_stds = np.zeros((slopes.shape[0], nepochs)) test_results = np.zeros((slopes.shape[0], nepochs)) test_stds = np.zeros((slopes.shape[0], nepochs)) series_names = ['Slope = {0}'.format(x) for x in slopes] for i, slope in enumerate(slopes): print 'Executing iteration {0} of {1}'.format(i + 1, slopes.shape[0]) (train_results[i], train_stds[i]), (test_results[i], test_stds[i]) = \ bulk(train_data=train_d, train_labels=train_l, test_data=test_d, test_labels=test_l, plot=False, nepochs=nepochs, niters=niters, activation_kargs={'m':slope}, verbose=False) # Make training plot title = 'Linear Regression Network - Training\n10 Iterations, ' \ "Varying Activation Function's Slope" out_path = os.path.join(out_dir, 'slope-train.png') plot_epoch(y_series=train_results, semilog=True, series_names=series_names, y_errs=train_stds, y_label='Cost', title=title, out_path=out_path) # Make testing plot title = 'Linear Regression Network - Testing\n10 Iterations, ' \ "Varying Activation Function's Slope" out_path = os.path.join(out_dir, 'slope-test.png') plot_epoch(y_series=test_results, semilog=True, series_names=series_names, y_errs=train_stds, y_label='Cost', title=title, out_path=out_path) if __name__ == '__main__': basic_sim() # bulk_sim() # vary_params(out_dir=r'D:\annt\test\Linear_Regression_Network')
tehtechguy/annt
src/annt/__init__.py
# __init__.py # # Author : <NAME> # Contact : http://techtorials.me # Date Created : 03/30/15 # # Description : Defines the annt package # Python Version : 2.7.8 # # License : MIT License http://opensource.org/licenses/mit-license.php # Copyright : (c) 2015 <NAME> """ This is a collection of various components for creating artificial neural networks in Python. Legal ===== This code is licensed under the U{MIT license<http://opensource.org/ licenses/mit-license.php>}. Any included datasets may be licensed differently. Refer to the individual dataset for more details. Prerequisites ============= - U{Python 2.7.X<https://www.python.org/downloads/release/python-279/>} - U{Numpy<http://www.numpy.org/>} - U{matplotlib<http://matplotlib.org/>} Installation ============ 1. Install all prerequisites Assuming you have U{pip<https://pip.pypa.io/en/latest/installing.html>} installed, located in your X{Python27/Scripts} directory: X{pip install numpy matplotlib} 2. Install this package: X{python setup.py install}. The setup file is located in the "src" folder. Getting Started =============== - Click U{here<http://techtorials.me/mldata/index.html>} to access the API. - Check out the L{examples<annt.examples>}. Package Organization ==================== The annt package contains the following subpackages and modules: G{packagetree annt} Connectivity ============ The following image shows how everything is connected: G{importgraph} Developer Notes =============== The following notes are for developers only. Installation ------------ 1. Download and install U{graphviz<http://www.graphviz.org/Download.. php>} 2. Edit line 111 in X{dev/epydoc_config.txt} to point to the directory containing "dot.exe". This is part of the graphviz installation. 3. Download this repo and execute X{python setup.py install}. 4. Download and install U{Epydoc<http://sourceforge.net/projects/ epydoc/files>} Generating the API ------------------ From the root level, execute X{python epydoc --config=epydoc_config.txt annt} @group Examples: examples @author: U{<NAME><http://techtorials.me>} @requires: Python 2.7.X @version: 0.4.0 @license: U{The MIT License<http://opensource.org/licenses/mit-license.php>} @copyright: S{copy} 2015 <NAME> """ __docformat__ = 'epytext'
sqiangcao99/hgr_v2t
t2vretrieval/models/criterion.py
<filename>t2vretrieval/models/criterion.py<gh_stars>100-1000 import torch import torch.nn as nn import framework.configbase import framework.ops def cosine_sim(im, s): '''cosine similarity between all the image and sentence pairs ''' inner_prod = im.mm(s.t()) im_norm = torch.sqrt((im**2).sum(1).view(-1, 1) + 1e-18) s_norm = torch.sqrt((s**2).sum(1).view(1, -1) + 1e-18) sim = inner_prod / (im_norm * s_norm) return sim class ContrastiveLoss(nn.Module): '''compute contrastive loss ''' def __init__(self, margin=0, max_violation=False, direction='bi', topk=1): '''Args: direction: i2t for negative sentence, t2i for negative image, bi for both ''' super(ContrastiveLoss, self).__init__() self.margin = margin self.max_violation = max_violation self.direction = direction self.topk = topk def forward(self, scores, margin=None, average_batch=True): ''' Args: scores: image-sentence score matrix, (batch, batch) the same row of im and s are positive pairs, different rows are negative pairs ''' if margin is None: margin = self.margin batch_size = scores.size(0) diagonal = scores.diag().view(batch_size, 1) # positive pairs # mask to clear diagonals which are positive pairs pos_masks = torch.eye(batch_size).bool().to(scores.device) batch_topk = min(batch_size, self.topk) if self.direction == 'i2t' or self.direction == 'bi': d1 = diagonal.expand_as(scores) # same collumn for im2s (negative sentence) # compare every diagonal score to scores in its collumn # caption retrieval cost_s = (margin + scores - d1).clamp(min=0) cost_s = cost_s.masked_fill(pos_masks, 0) if self.max_violation: cost_s, _ = torch.topk(cost_s, batch_topk, dim=1) cost_s = cost_s / batch_topk if average_batch: cost_s = cost_s / batch_size else: if average_batch: cost_s = cost_s / (batch_size * (batch_size - 1)) cost_s = torch.sum(cost_s) if self.direction == 't2i' or self.direction == 'bi': d2 = diagonal.t().expand_as(scores) # same row for s2im (negative image) # compare every diagonal score to scores in its row cost_im = (margin + scores - d2).clamp(min=0) cost_im = cost_im.masked_fill(pos_masks, 0) if self.max_violation: cost_im, _ = torch.topk(cost_im, batch_topk, dim=0) cost_im = cost_im / batch_topk if average_batch: cost_im = cost_im / batch_size else: if average_batch: cost_im = cost_im / (batch_size * (batch_size - 1)) cost_im = torch.sum(cost_im) if self.direction == 'i2t': return cost_s elif self.direction == 't2i': return cost_im else: return cost_s + cost_im
sqiangcao99/hgr_v2t
framework/run_utils.py
import os import json import datetime import numpy as np import glob import framework.configbase def gen_common_pathcfg(path_cfg_file, is_train=False): path_cfg = framework.configbase.PathCfg() path_cfg.load(json.load(open(path_cfg_file))) output_dir = path_cfg.output_dir path_cfg.log_dir = os.path.join(output_dir, 'log') path_cfg.model_dir = os.path.join(output_dir, 'model') path_cfg.pred_dir = os.path.join(output_dir, 'pred') if not os.path.exists(path_cfg.log_dir): os.makedirs(path_cfg.log_dir) if not os.path.exists(path_cfg.model_dir): os.makedirs(path_cfg.model_dir) if not os.path.exists(path_cfg.pred_dir): os.makedirs(path_cfg.pred_dir) if is_train: timestamp = datetime.datetime.now().strftime('%Y-%m-%d-%H-%M-%S') path_cfg.log_file = os.path.join(path_cfg.log_dir, 'log-' + timestamp) else: path_cfg.log_file = None return path_cfg def find_best_val_models(log_dir, model_dir): step_jsons = glob.glob(os.path.join(log_dir, 'val.step.*.json')) epoch_jsons = glob.glob(os.path.join(log_dir, 'val.epoch.*.json')) val_names, val_scores = [], [] for i, json_file in enumerate(step_jsons + epoch_jsons): json_name = os.path.basename(json_file) scores = json.load(open(json_file)) val_names.append(json_name) val_scores.append(scores) measure_names = list(val_scores[0].keys()) model_files = {} for measure_name in measure_names: # for metrics: the lower the better if 'loss' in measure_name or 'medr' in measure_name or 'meanr' in measure_name: idx = np.argmin([scores[measure_name] for scores in val_scores]) # for metrics: the higher the better else: idx = np.argmax([scores[measure_name] for scores in val_scores]) json_name = val_names[idx] model_file = os.path.join(model_dir, 'epoch.%s.th'%(json_name.split('.')[2]) if 'epoch' in json_name \ else 'step.%s.th'%(json_name.split('.')[2])) model_files.setdefault(model_file, []) model_files[model_file].append(measure_name) name2file = {'-'.join(measure_name): model_file for model_file, measure_name in model_files.items()} return name2file
sqiangcao99/hgr_v2t
framework/modules/embeddings.py
<reponame>sqiangcao99/hgr_v2t """ Embeddings module """ import math import torch import torch.nn as nn class PositionalEncoding(nn.Module): """ Implements the sinusoidal positional encoding for non-recurrent neural networks. Implementation based on "Attention Is All You Need" Args: dim_embed (int): embedding size (even number) """ def __init__(self, dim_embed, max_len=100): super(PositionalEncoding, self).__init__() pe = torch.zeros(max_len, dim_embed) position = torch.arange(0, max_len).unsqueeze(1) div_term = torch.exp((torch.arange(0, dim_embed, 2, dtype=torch.float) * -(math.log(10000.0) / dim_embed))) pe[:, 0::2] = torch.sin(position.float() * div_term) pe[:, 1::2] = torch.cos(position.float() * div_term) self.pe = pe # size=(max_len, dim_embed) self.dim_embed = dim_embed def forward(self, emb, step=None): if emb.device != self.pe.device: self.pe = self.pe.to(emb.device) if step is None: # emb.size = (batch, seq_len, dim_embed) emb = emb + self.pe[:emb.size(1)] else: # emb.size = (batch, dim_embed) emb = emb + self.pe[step] return emb class Embedding(nn.Module): """Words embeddings for encoder/decoder. Args: word_vec_size (int): size of the dictionary of embeddings. word_vocab_size (int): size of dictionary of embeddings for words. position_encoding (bool): see :obj:`modules.PositionalEncoding` """ def __init__(self, word_vocab_size, word_vec_size, position_encoding=False, fix_word_embed=False, max_len=100): super(Embedding, self).__init__() self.word_vec_size = word_vec_size self.we = nn.Embedding(word_vocab_size, word_vec_size) if fix_word_embed: self.we.weight.requires_grad = False self.init_weight() self.position_encoding = position_encoding if self.position_encoding: self.pe = PositionalEncoding(word_vec_size, max_len=max_len) def init_weight(self): std = 1. / (self.word_vec_size**0.5) nn.init.uniform_(self.we.weight, -std, std) def forward(self, word_idxs, step=None): """Computes the embeddings for words. Args: word_idxs (`LongTensor`): index tensor size = (batch, seq_len) or (batch, ) Return: embeds: `FloatTensor`, size = (batch, seq_len, dim_embed) or (batch, dim_embed) """ embeds = self.we(word_idxs) if self.position_encoding: embeds = self.pe(embeds, step=step) return embeds
sqiangcao99/hgr_v2t
t2vretrieval/models/evaluation.py
import numpy as np def eval_q2m(scores, q2m_gts): ''' Image -> Text / Text -> Image Args: scores: (n_query, n_memory) matrix of similarity scores q2m_gts: list, each item is the positive memory ids of the query id Returns: scores: (recall@1, 5, 10, median rank, mean rank) gt_ranks: the best ranking of ground-truth memories ''' n_q, n_m = scores.shape gt_ranks = np.zeros((n_q, ), np.int32) for i in range(n_q): s = scores[i] sorted_idxs = np.argsort(-s) rank = n_m for k in q2m_gts[i]: tmp = np.where(sorted_idxs == k)[0][0] if tmp < rank: rank = tmp gt_ranks[i] = rank # compute metrics r1 = 100 * len(np.where(gt_ranks < 1)[0]) / n_q r5 = 100 * len(np.where(gt_ranks < 5)[0]) / n_q r10 = 100 * len(np.where(gt_ranks < 10)[0]) / n_q medr = np.median(gt_ranks) + 1 meanr = gt_ranks.mean() + 1 return (r1, r5, r10, medr, meanr)
sqiangcao99/hgr_v2t
t2vretrieval/readers/mpdata.py
<reponame>sqiangcao99/hgr_v2t<filename>t2vretrieval/readers/mpdata.py import os import json import numpy as np import torch.utils.data BOS, EOS, UNK = 0, 1, 2 class MPDataset(torch.utils.data.Dataset): def __init__(self, name_file, mp_ft_files, word2int_file, max_words_in_sent, ref_caption_file=None, is_train=False, _logger=None): if _logger is None: self.print_fn = print else: self.print_fn = _logger.info self.max_words_in_sent = max_words_in_sent self.is_train = is_train self.names = np.load(name_file) self.word2int = json.load(open(word2int_file)) self.mp_fts = [] for mp_ft_file in mp_ft_files: self.mp_fts.append(np.load(mp_ft_file)) self.mp_fts = np.concatenate(self.mp_fts, axis=-1) self.num_videos = len(self.mp_fts) self.print_fn('mp_fts size %s' % (str(self.mp_fts.shape))) if ref_caption_file is None: self.ref_captions = None else: self.ref_captions = json.load(open(ref_caption_file)) self.captions = set() self.pair_idxs = [] for i, name in enumerate(self.names): for j, sent in enumerate(self.ref_captions[name]): self.captions.add(sent) self.pair_idxs.append((i, j)) self.captions = list(self.captions) self.num_pairs = len(self.pair_idxs) self.print_fn('captions size %d' % self.num_pairs) def process_sent(self, sent, max_words): tokens = [self.word2int.get(w, UNK) for w in sent.split()] # # add BOS, EOS? # tokens = [BOS] + tokens + [EOS] tokens = tokens[:max_words] tokens_len = len(tokens) tokens = np.array(tokens + [EOS] * (max_words - tokens_len)) return tokens, tokens_len def __len__(self): if self.is_train: return self.num_pairs else: return self.num_videos def __getitem__(self, idx): out = {} if self.is_train: video_idx, cap_idx = self.pair_idxs[idx] name = self.names[video_idx] mp_ft = self.mp_fts[video_idx] sent = self.ref_captions[name][cap_idx] cap_ids, cap_len = self.process_sent(sent, self.max_words_in_sent) out['caption_ids'] = cap_ids out['caption_lens'] = cap_len else: name = self.names[idx] mp_ft = self.mp_fts[idx] out['names'] = name out['mp_fts'] = mp_ft return out def iterate_over_captions(self, batch_size): # the sentence order is the same as self.captions for s in range(0, len(self.captions), batch_size): e = s + batch_size cap_ids, cap_lens = [], [] for sent in self.captions[s: e]: cap_id, cap_len = self.process_sent(sent, self.max_words_in_sent) cap_ids.append(cap_id) cap_lens.append(cap_len) yield { 'caption_ids': np.array(cap_ids, np.int32), 'caption_lens': np.array(cap_lens, np.int32), } def collate_fn(data): outs = {} for key in ['names', 'mp_fts', 'caption_ids', 'caption_lens']: if key in data[0]: outs[key] = [x[key] for x in data] # reduce caption_ids lens if 'caption_lens' in outs: max_cap_len = np.max(outs['caption_lens']) outs['caption_ids'] = np.array(outs['caption_ids'])[:, :max_cap_len] return outs
sqiangcao99/hgr_v2t
t2vretrieval/encoders/video.py
import torch import torch.nn as nn import torch.nn.functional as F import framework.configbase class MPEncoderConfig(framework.configbase.ModuleConfig): def __init__(self): super().__init__() self.dim_fts = [2048] self.dim_embed = 1024 self.dropout = 0 class MPEncoder(nn.Module): def __init__(self, config): super().__init__() self.config = config input_size = sum(self.config.dim_fts) self.ft_embed = nn.Linear(input_size, self.config.dim_embed, bias=True) self.dropout = nn.Dropout(self.config.dropout) def forward(self, inputs): ''' Args: inputs: (batch, dim_fts) or (batch, max_seq_len, dim_fts) Return: embeds: (batch, dim_embed) or (batch, max_seq_len, dim_fts) ''' embeds = self.ft_embed(inputs) embeds = self.dropout(embeds) return embeds
sqiangcao99/hgr_v2t
t2vretrieval/encoders/graph.py
import math import torch import torch.nn as nn import torch.nn.functional as F class GCNLayer(nn.Module): def __init__(self, embed_size, dropout=0.0): super().__init__() self.embed_size = embed_size self.ctx_layer = nn.Linear(self.embed_size, self.embed_size, bias=False) self.layernorm = nn.LayerNorm(embed_size) self.dropout = nn.Dropout(dropout) def forward(self, node_fts, rel_edges): '''Args: node_fts: (batch_size, num_nodes, embed_size) rel_edges: (batch_size, num_nodes, num_nodes) ''' ctx_embeds = self.ctx_layer(torch.bmm(rel_edges, node_fts)) node_embeds = node_fts + self.dropout(ctx_embeds) node_embeds = self.layernorm(node_embeds) return node_embeds class AttnGCNLayer(GCNLayer): def __init__(self, embed_size, d_ff, dropout=0.0): super().__init__(embed_size, dropout=dropout) self.edge_attn_query = nn.Linear(embed_size, d_ff) self.edge_attn_key = nn.Linear(embed_size, d_ff) self.attn_denominator = math.sqrt(d_ff) def forward(self, node_fts, rel_edges): ''' Args: node_fts: (batch_size, num_nodes, embed_size) rel_edges: (batch_size, num_nodes, num_nodes) ''' # (batch_size, num_nodes, num_nodes) attn_scores = torch.einsum('bod,bid->boi', self.edge_attn_query(node_fts), self.edge_attn_key(node_fts)) / self.attn_denominator attn_scores = attn_scores.masked_fill(rel_edges == 0, -1e18) attn_scores = torch.softmax(attn_scores, dim=2) # some nodes do not connect with any edge attn_scores = attn_scores.masked_fill(rel_edges == 0, 0) ctx_embeds = self.ctx_layer(torch.bmm(attn_scores, node_fts)) node_embeds = node_fts + self.dropout(ctx_embeds) node_embeds = self.layernorm(node_embeds) return node_embeds class GCNEncoder(nn.Module): def __init__(self, dim_input, dim_hidden, num_hidden_layers, embed_first=False, dropout=0, attention=False): super().__init__() self.dim_input = dim_input self.dim_hidden = dim_hidden self.num_hidden_layers = num_hidden_layers self.embed_first = embed_first self.attention = attention if self.attention: gcn_fn = AttnGCNLayer else: gcn_fn = GCNLayer if self.embed_first: self.first_embedding = nn.Sequential( nn.Linear(self.dim_input, self.dim_hidden), nn.ReLU()) self.layers = nn.ModuleList() for k in range(num_hidden_layers): if self.attention: h2h = gcn_fn(self.dim_hidden, self.dim_hidden // 2, dropout=dropout) else: h2h = gcn_fn(self.dim_hidden, dropout=dropout) self.layers.append(h2h) def forward(self, node_fts, rel_edges): if self.embed_first: node_fts = self.first_embedding(node_fts) for k in range(self.num_hidden_layers): layer = self.layers[k] node_fts = layer(node_fts, rel_edges) # (batch_size, num_nodes, dim_hidden) return node_fts
sqiangcao99/hgr_v2t
t2vretrieval/miscs/semantic_role_labeling.py
import os import argparse import json from allennlp.predictors.predictor import Predictor def main(): parser = argparse.ArgumentParser() parser.add_argument('ref_caption_file') parser.add_argument('out_file') parser.add_argument('--cuda_device', default=-1, type=int) opts = parser.parse_args() predictor = Predictor.from_path("https://s3-us-west-2.amazonaws.com/allennlp/models/bert-base-srl-2019.06.17.tar.gz", cuda_device=opts.cuda_device) ref_caps = json.load(open(opts.ref_caption_file)) uniq_sents = set() for key, sents in ref_caps.items(): for sent in sents: uniq_sents.add(sent) uniq_sents = list(uniq_sents) print('unique sents', len(uniq_sents)) outs = {} if os.path.exists(opts.out_file): outs = json.load(open(opts.out_file)) for i, sent in enumerate(uniq_sents): if sent in outs: continue try: out = predictor.predict_tokenized(sent.split()) except KeyboardInterrupt: break except: continue outs[sent] = out if i % 1000 == 0: print('finish %d / %d = %.2f%%' % (i, len(uniq_sents), i / len(uniq_sents) * 100)) with open(opts.out_file, 'w') as f: json.dump(outs, f) if __name__ == '__main__': main()
sqiangcao99/hgr_v2t
t2vretrieval/readers/rolegraphs.py
import os import json import numpy as np import h5py import collections import torch import t2vretrieval.readers.mpdata ROLES = ['V', 'ARG1', 'ARG0', 'ARG2', 'ARG3', 'ARG4', 'ARGM-LOC', 'ARGM-MNR', 'ARGM-TMP', 'ARGM-DIR', 'ARGM-ADV', 'ARGM-PRP', 'ARGM-PRD', 'ARGM-COM', 'ARGM-MOD', 'NOUN'] class RoleGraphDataset(t2vretrieval.readers.mpdata.MPDataset): def __init__(self, name_file, attn_ft_files, word2int_file, max_words_in_sent, num_verbs, num_nouns, ref_caption_file, ref_graph_file, max_attn_len=20, load_video_first=False, is_train=False, _logger=None): if _logger is None: self.print_fn = print else: self.print_fn = _logger.info self.max_words_in_sent = max_words_in_sent self.is_train = is_train self.attn_ft_files = attn_ft_files self.max_attn_len = max_attn_len self.load_video_first = load_video_first self.names = np.load(name_file) self.word2int = json.load(open(word2int_file)) self.num_videos = len(self.names) self.print_fn('num_videos %d' % (self.num_videos)) if ref_caption_file is None: self.ref_captions = None else: self.ref_captions = json.load(open(ref_caption_file)) self.captions = set() self.pair_idxs = [] for i, name in enumerate(self.names): for j, sent in enumerate(self.ref_captions[name]): self.captions.add(sent) self.pair_idxs.append((i, j)) self.captions = list(self.captions) self.num_pairs = len(self.pair_idxs) self.print_fn('captions size %d' % self.num_pairs) if self.load_video_first: self.all_attn_fts, self.all_attn_lens = [], [] for name in self.names: attn_fts = self.load_attn_ft_by_name(name, self.attn_ft_files) attn_fts, attn_len = self.pad_or_trim_feature(attn_fts, self.max_attn_len, trim_type='select') self.all_attn_fts.append(attn_fts) self.all_attn_lens.append(attn_len) self.all_attn_fts = np.array(self.all_attn_fts) self.all_attn_lens = np.array(self.all_attn_lens) self.num_verbs = num_verbs self.num_nouns = num_nouns self.role2int = {} for i, role in enumerate(ROLES): self.role2int[role] = i self.role2int['C-%s'%role] = i self.role2int['R-%s'%role] = i self.ref_graphs = json.load(open(ref_graph_file)) def load_attn_ft_by_name(self, name, attn_ft_files): attn_fts = [] for i, attn_ft_file in enumerate(attn_ft_files): with h5py.File(attn_ft_file, 'r') as f: key = name.replace('/', '_') attn_ft = f[key][...] attn_fts.append(attn_ft) attn_fts = np.concatenate([attn_ft for attn_ft in attn_fts], axis=-1) return attn_fts def pad_or_trim_feature(self, attn_ft, max_attn_len, trim_type='top'): seq_len, dim_ft = attn_ft.shape attn_len = min(seq_len, max_attn_len) # pad if seq_len < max_attn_len: new_ft = np.zeros((max_attn_len, dim_ft), np.float32) new_ft[:seq_len] = attn_ft # trim else: if trim_type == 'top': new_ft = attn_ft[:max_attn_len] elif trim_type == 'select': idxs = np.round(np.linspace(0, seq_len-1, max_attn_len)).astype(np.int32) new_ft = attn_ft[idxs] return new_ft, attn_len def get_caption_outs(self, out, sent, graph): graph_nodes, graph_edges = graph #print(graph) verb_node2idxs, noun_node2idxs = {}, {} edges = [] out['node_roles'] = np.zeros((self.num_verbs + self.num_nouns, ), np.int32) # root node sent_ids, sent_len = self.process_sent(sent, self.max_words_in_sent) out['sent_ids'] = sent_ids out['sent_lens'] = sent_len # graph: add verb nodes node_idx = 1 out['verb_masks'] = np.zeros((self.num_verbs, self.max_words_in_sent), np.bool) for knode, vnode in graph_nodes.items(): k = node_idx - 1 if k >= self.num_verbs: break if vnode['role'] == 'V' and np.min(vnode['spans']) < self.max_words_in_sent: verb_node2idxs[knode] = node_idx for widx in vnode['spans']: if widx < self.max_words_in_sent: out['verb_masks'][k][widx] = True out['node_roles'][node_idx - 1] = self.role2int['V'] # add root to verb edge edges.append((0, node_idx)) node_idx += 1 # graph: add noun nodes node_idx = 1 + self.num_verbs out['noun_masks'] = np.zeros((self.num_nouns, self.max_words_in_sent), np.bool) for knode, vnode in graph_nodes.items(): k = node_idx - self.num_verbs - 1 if k >= self.num_nouns: break if vnode['role'] not in ['ROOT', 'V'] and np.min(vnode['spans']) < self.max_words_in_sent: noun_node2idxs[knode] = node_idx for widx in vnode['spans']: if widx < self.max_words_in_sent: out['noun_masks'][k][widx] = True out['node_roles'][node_idx - 1] = self.role2int.get(vnode['role'], self.role2int['NOUN']) node_idx += 1 # graph: add verb_node to noun_node edges for e in graph_edges: if e[0] in verb_node2idxs and e[1] in noun_node2idxs: edges.append((verb_node2idxs[e[0]], noun_node2idxs[e[1]])) edges.append((noun_node2idxs[e[1]], verb_node2idxs[e[0]])) num_nodes = 1 + self.num_verbs + self.num_nouns rel_matrix = np.zeros((num_nodes, num_nodes), dtype=np.float32) for src_nodeidx, tgt_nodeidx in edges: rel_matrix[tgt_nodeidx, src_nodeidx] = 1 # row norm for i in range(num_nodes): s = np.sum(rel_matrix[i]) if s > 0: rel_matrix[i] /= s out['rel_edges'] = rel_matrix return out def __getitem__(self, idx): out = {} if self.is_train: video_idx, cap_idx = self.pair_idxs[idx] name = self.names[video_idx] sent = self.ref_captions[name][cap_idx] out = self.get_caption_outs(out, sent, self.ref_graphs[sent]) else: video_idx = idx name = self.names[idx] if self.load_video_first: attn_fts, attn_len = self.all_attn_fts[video_idx], self.all_attn_lens[video_idx] else: attn_fts = self.load_attn_ft_by_name(name, self.attn_ft_files) attn_fts, attn_len = self.pad_or_trim_feature(attn_fts, self.max_attn_len, trim_type='select') out['names'] = name out['attn_fts'] = attn_fts out['attn_lens'] = attn_len return out def iterate_over_captions(self, batch_size): # the sentence order is the same as self.captions for s in range(0, len(self.captions), batch_size): e = s + batch_size data = [] for sent in self.captions[s: e]: out = self.get_caption_outs({}, sent, self.ref_graphs[sent]) data.append(out) outs = collate_graph_fn(data) yield outs def collate_graph_fn(data): outs = {} for key in ['names', 'attn_fts', 'attn_lens', 'sent_ids', 'sent_lens', 'verb_masks', 'noun_masks', 'node_roles', 'rel_edges']: if key in data[0]: outs[key] = [x[key] for x in data] batch_size = len(data) # reduce attn_lens if 'attn_fts' in outs: max_len = np.max(outs['attn_lens']) outs['attn_fts'] = np.stack(outs['attn_fts'], 0)[:, :max_len] # reduce caption_ids lens if 'sent_lens' in outs: max_cap_len = np.max(outs['sent_lens']) outs['sent_ids'] = np.array(outs['sent_ids'])[:, :max_cap_len] outs['verb_masks'] = np.array(outs['verb_masks'])[:, :, :max_cap_len] outs['noun_masks'] = np.array(outs['noun_masks'])[:, :, :max_cap_len] return outs
sqiangcao99/hgr_v2t
t2vretrieval/encoders/sentence.py
import torch import torch.nn as nn import torch.nn.functional as F import framework.configbase from framework.modules.embeddings import Embedding import framework.ops class SentEncoderConfig(framework.configbase.ModuleConfig): def __init__(self): super().__init__() self.num_words = 0 self.dim_word = 300 self.fix_word_embed = False self.rnn_type = 'gru' # gru, lstm self.bidirectional = True self.rnn_hidden_size = 1024 self.num_layers = 1 self.dropout = 0.5 def _assert(self): assert self.rnn_type in ['gru', 'lstm'], 'invalid rnn_type' class SentEncoder(nn.Module): def __init__(self, config): super().__init__() self.config = config self.embedding = Embedding(self.config.num_words, self.config.dim_word, fix_word_embed=self.config.fix_word_embed) dim_word = self.config.dim_word self.rnn = framework.ops.rnn_factory(self.config.rnn_type, input_size=dim_word, hidden_size=self.config.rnn_hidden_size, num_layers=self.config.num_layers, dropout=self.config.dropout, bidirectional=self.config.bidirectional, bias=True, batch_first=True) self.dropout = nn.Dropout(self.config.dropout) self.init_weights() def init_weights(self): directions = [''] if self.config.bidirectional: directions.append('_reverse') for layer in range(self.config.num_layers): for direction in directions: for name in ['i', 'h']: weight = getattr(self.rnn, 'weight_%sh_l%d%s'%(name, layer, direction)) nn.init.orthogonal_(weight.data) bias = getattr(self.rnn, 'bias_%sh_l%d%s'%(name, layer, direction)) nn.init.constant_(bias, 0) if name == 'i' and self.config.rnn_type == 'lstm': bias.data.index_fill_(0, torch.arange( self.config.rnn_hidden_size, self.config.rnn_hidden_size*2).long(), 1) def forward_text_encoder(self, word_embeds, seq_lens, init_states): # outs.size = (batch, seq_len, num_directions * hidden_size) outs, states = framework.ops.calc_rnn_outs_with_sort( self.rnn, word_embeds, seq_lens, init_states) return outs def forward(self, cap_ids, cap_lens, init_states=None, return_dense=False): ''' Args: cap_ids: LongTensor, (batch, seq_len) cap_lens: FloatTensor, (batch, ) Returns: if return_dense: embeds: FloatTensor, (batch, seq_len, embed_size) else: embeds: FloatTensor, (batch, embed_size) ''' word_embeds = self.embedding(cap_ids) hiddens = self.forward_text_encoder( self.dropout(word_embeds), cap_lens, init_states) batch_size, max_seq_len, hidden_size = hiddens.size() if self.config.bidirectional: splited_hiddens = torch.split(hiddens, self.config.rnn_hidden_size, dim=2) hiddens = (splited_hiddens[0] + splited_hiddens[1]) / 2 if return_dense: return hiddens else: sent_masks = framework.ops.sequence_mask(cap_lens, max_seq_len, inverse=False).float() sent_embeds = torch.sum(hiddens * sent_masks.unsqueeze(2), 1) / cap_lens.unsqueeze(1).float() return sent_embeds class SentAttnEncoder(SentEncoder): def __init__(self, config): super().__init__(config) self.ft_attn = nn.Linear(self.config.rnn_hidden_size, 1) self.softmax = nn.Softmax(dim=1) def forward(self, cap_ids, cap_lens, init_states=None, return_dense=False): hiddens = super().forward(cap_ids, cap_lens, init_states=init_states, return_dense=True) attn_scores = self.ft_attn(hiddens).squeeze(2) cap_masks = framework.ops.sequence_mask(cap_lens, max_len=attn_scores.size(1), inverse=False) attn_scores = attn_scores.masked_fill(cap_masks == 0, -1e18) attn_scores = self.softmax(attn_scores) if return_dense: return hiddens, attn_scores else: return torch.sum(hiddens * attn_scores.unsqueeze(2), 1)
sqiangcao99/hgr_v2t
t2vretrieval/encoders/mlvideo.py
<filename>t2vretrieval/encoders/mlvideo.py import torch import torch.nn as nn import torch.nn.functional as F import framework.configbase class MultilevelEncoderConfig(framework.configbase.ModuleConfig): def __init__(self): super().__init__() self.dim_fts = [2048] self.dim_embed = 1024 self.dropout = 0 self.num_levels = 3 self.share_enc = False class MultilevelEncoder(nn.Module): def __init__(self, config): super().__init__() self.config = config input_size = sum(self.config.dim_fts) self.dropout = nn.Dropout(self.config.dropout) num_levels = 1 if self.config.share_enc else self.config.num_levels self.level_embeds = nn.ModuleList([ nn.Linear(input_size, self.config.dim_embed, bias=True) for k in range(num_levels) ]) self.ft_attn = nn.Linear(self.config.dim_embed, 1, bias=True) def forward(self, inputs, input_lens): ''' Args: inputs: (batch, max_seq_len, dim_fts) Return: sent_embeds: (batch, dim_embed) verb_embeds: (batch, max_seq_len, dim_embed) noun_embeds: (batch, max_seq_len, dim_embed) ''' embeds = [] for k in range(self.config.num_levels): if self.config.share_enc: k = 0 embeds.append(self.dropout(self.level_embeds[k](inputs))) attn_scores = self.ft_attn(embeds[0]).squeeze(2) input_pad_masks = framework.ops.sequence_mask(input_lens, max_len=attn_scores.size(1), inverse=True) attn_scores = attn_scores.masked_fill(input_pad_masks, -1e18) attn_scores = torch.softmax(attn_scores, dim=1) sent_embeds = torch.sum(embeds[0] * attn_scores.unsqueeze(2), 1) return sent_embeds, embeds[1], embeds[2]
sqiangcao99/hgr_v2t
framework/modelbase.py
import os import time import json import numpy as np import torch import torch.nn as nn from torch import optim import torch.nn.functional as F import framework.logbase class ModelBase(object): def __init__(self, config, _logger=None, gpu_id=0): '''initialize model (support single GPU, otherwise need to be customized) ''' self.device = torch.device("cuda:%d"%gpu_id if torch.cuda.is_available() else "cpu") self.config = config if _logger is None: self.print_fn = print else: self.print_fn = _logger.info self.submods = self.build_submods() for submod in self.submods.values(): submod.to(self.device) self.criterion = self.build_loss() self.params, self.optimizer, self.lr_scheduler = self.build_optimizer() num_params, num_weights = 0, 0 for key, submod in self.submods.items(): for varname, varvalue in submod.state_dict().items(): self.print_fn('%s: %s, shape=%s, num:%d' % ( key, varname, str(varvalue.size()), np.prod(varvalue.size()))) num_params += 1 num_weights += np.prod(varvalue.size()) self.print_fn('num params %d, num weights %d'%(num_params, num_weights)) self.print_fn('trainable: num params %d, num weights %d'%( len(self.params), sum([np.prod(param.size()) for param in self.params]))) def build_submods(self): raise NotImplementedError('implement build_submods function: return submods') def build_loss(self): raise NotImplementedError('implement build_loss function: return criterion') def forward_loss(self, batch_data, step=None): raise NotImplementedError('implement forward_loss function: return loss and additional outs') def validate(self, val_reader, step=None): self.eval_start() # raise NotImplementedError('implement validate function: return metrics') def test(self, tst_reader, tst_pred_file, tst_model_file=None): if tst_model_file is not None: self.load_checkpoint(tst_model_file) self.eval_start() # raise NotImplementedError('implement test function') ########################## boilerpipe functions ######################## def build_optimizer(self): trn_params = [] trn_param_ids = set() per_param_opts = [] for key, submod in self.submods.items(): if self.config.subcfgs[key].freeze: for param in submod.parameters(): param.requires_grad = False else: params = [] for param in submod.parameters(): # sometimes we share params in different submods if param.requires_grad and id(param) not in trn_param_ids: params.append(param) trn_param_ids.add(id(param)) per_param_opts.append({ 'params': params, 'lr': self.config.base_lr * self.config.subcfgs[key].lr_mult, 'weight_decay': self.config.subcfgs[key].weight_decay, }) trn_params.extend(params) if len(trn_params) > 0: optimizer = optim.Adam(per_param_opts, lr=self.config.base_lr) lr_scheduler = optim.lr_scheduler.MultiStepLR(optimizer, milestones=self.config.decay_boundarys, gamma=self.config.decay_rate) else: optimizer, lr_scheduler = None, None print('no traiable parameters') return trn_params, optimizer, lr_scheduler def train_start(self): for key, submod in self.submods.items(): submod.train() torch.set_grad_enabled(True) def eval_start(self): for key, submod in self.submods.items(): submod.eval() torch.set_grad_enabled(False) def save_checkpoint(self, ckpt_file, submods=None): if submods is None: submods = self.submods state_dicts = {} for key, submod in submods.items(): state_dicts[key] = {} for varname, varvalue in submod.state_dict().items(): state_dicts[key][varname] = varvalue.cpu() torch.save(state_dicts, ckpt_file) def load_checkpoint(self, ckpt_file, submods=None): if submods is None: submods = self.submods state_dicts = torch.load(ckpt_file, map_location=lambda storage, loc: storage) num_resumed_vars = 0 for key, state_dict in state_dicts.items(): if key in submods: own_state_dict = submods[key].state_dict() new_state_dict = {} for varname, varvalue in state_dict.items(): if varname in own_state_dict: new_state_dict[varname] = varvalue num_resumed_vars += 1 own_state_dict.update(new_state_dict) submods[key].load_state_dict(own_state_dict) self.print_fn('number of resumed variables: %d'%num_resumed_vars) def pretty_print_metrics(self, prefix, metrics): metric_str = [] for measure, score in metrics.items(): metric_str.append('%s %.4f'%(measure, score)) metric_str = ' '.join(metric_str) self.print_fn('%s: %s' % (prefix, metric_str)) def get_current_base_lr(self): return self.optimizer.param_groups[0]['lr'] def train_one_batch(self, batch_data, step): self.optimizer.zero_grad() loss = self.forward_loss(batch_data, step=step) loss.backward() self.optimizer.step() loss_value = loss.data.item() if step is not None and self.config.monitor_iter > 0 and step % self.config.monitor_iter == 0: self.print_fn('\ttrn step %d lr %.8f %s: %.4f' % (step, self.get_current_base_lr(), 'loss', loss_value)) return {'loss': loss_value} def train_one_epoch(self, step, trn_reader, val_reader, model_dir, log_dir): self.train_start() avg_loss, n_batches = {}, {} for batch_data in trn_reader: loss = self.train_one_batch(batch_data, step) for loss_key, loss_value in loss.items(): avg_loss.setdefault(loss_key, 0) n_batches.setdefault(loss_key, 0) avg_loss[loss_key] += loss_value n_batches[loss_key] += 1 step += 1 if self.config.save_iter > 0 and step % self.config.save_iter == 0: self.save_checkpoint(os.path.join(model_dir, 'step.%d.th'%step)) if (self.config.save_iter > 0 and step % self.config.save_iter == 0) \ or (self.config.val_iter > 0 and step % self.config.val_iter == 0): metrics = self.validate(val_reader, step=step) with open(os.path.join(log_dir, 'val.step.%d.json'%step), 'w') as f: json.dump(metrics, f, indent=2) self.pretty_print_metrics('\tval step %d'%step, metrics) self.train_start() for loss_key, loss_value in avg_loss.items(): avg_loss[loss_key] = loss_value / n_batches[loss_key] return avg_loss, step def epoch_postprocess(self, epoch): if self.lr_scheduler is not None: self.lr_scheduler.step() def train(self, trn_reader, val_reader, model_dir, log_dir, resume_file=None): assert self.optimizer is not None if resume_file is not None: self.load_checkpoint(resume_file) # first validate metrics = self.validate(val_reader) self.pretty_print_metrics('init val', metrics) # training step = 0 for epoch in range(self.config.num_epoch): avg_loss, step = self.train_one_epoch( step, trn_reader, val_reader, model_dir, log_dir) self.pretty_print_metrics('epoch (%d/%d) trn'%(epoch, self.config.num_epoch), avg_loss) self.epoch_postprocess(epoch) if self.config.save_per_epoch: self.save_checkpoint(os.path.join(model_dir, 'epoch.%d.th'%epoch)) if self.config.val_per_epoch: metrics = self.validate(val_reader, step=step) with open(os.path.join(log_dir, 'val.epoch.%d.step.%d.json'%(epoch, step)), 'w') as f: json.dump(metrics, f, indent=2) self.pretty_print_metrics('epoch (%d/%d) val' % (epoch, self.config.num_epoch), metrics)
sqiangcao99/hgr_v2t
t2vretrieval/models/globalmatch.py
import os import numpy as np import collections import json import torch import framework.ops import framework.configbase import framework.modelbase import t2vretrieval.encoders.video import t2vretrieval.encoders.sentence import t2vretrieval.models.criterion import t2vretrieval.models.evaluation from t2vretrieval.models.criterion import cosine_sim VISENC = 'video_encoder' TXTENC = 'text_encoder' class GlobalMatchModelConfig(framework.configbase.ModelConfig): def __init__(self): super().__init__() self.max_frames_in_video = None self.max_words_in_sent = 30 self.margin = 0.2 self.max_violation = False self.hard_topk = 1 self.loss_direction = 'bi' self.subcfgs[VISENC] = t2vretrieval.encoders.video.MPEncoderConfig() self.subcfgs[TXTENC] = t2vretrieval.encoders.sentence.SentEncoderConfig() class GlobalMatchModel(framework.modelbase.ModelBase): def build_submods(self): submods = { VISENC: t2vretrieval.encoders.video.MPEncoder(self.config.subcfgs[VISENC]), TXTENC: t2vretrieval.encoders.sentence.SentEncoder(self.config.subcfgs[TXTENC]), } return submods def build_loss(self): criterion = t2vretrieval.models.criterion.ContrastiveLoss( margin=self.config.margin, max_violation=self.config.max_violation, topk=self.config.hard_topk, direction=self.config.loss_direction) return criterion def forward_video_embed(self, batch_data): vid_fts = torch.FloatTensor(batch_data['mp_fts']).to(self.device) vid_embeds = self.submods[VISENC](vid_fts) return {'vid_embeds': vid_embeds} def forward_text_embed(self, batch_data): cap_ids = torch.LongTensor(batch_data['caption_ids']).to(self.device) cap_lens = torch.LongTensor(batch_data['caption_lens']).to(self.device) cap_embeds = self.submods[TXTENC](cap_ids, cap_lens) return {'cap_embeds': cap_embeds} def generate_scores(self, **kwargs): # compute image-sentence similarity vid_embeds = kwargs['vid_embeds'] cap_embeds = kwargs['cap_embeds'] scores = cosine_sim(vid_embeds, cap_embeds) # s[i, j] i: im_idx, j: s_idx return scores def forward_loss(self, batch_data, step=None): vid_enc_outs = self.forward_video_embed(batch_data) cap_enc_outs = self.forward_text_embed(batch_data) cap_enc_outs.update(vid_enc_outs) scores = self.generate_scores(**cap_enc_outs) loss = self.criterion(scores) if step is not None and self.config.monitor_iter > 0 and step % self.config.monitor_iter == 0: neg_scores = scores.masked_fill(torch.eye(len(scores), dtype=torch.bool).to(self.device), -1e10) self.print_fn('\tstep %d: pos mean scores %.2f, hard neg mean scores i2t %.2f, t2i %.2f'%( step, torch.mean(torch.diag(scores)), torch.mean(torch.max(neg_scores, 1)[0]), torch.mean(torch.max(neg_scores, 0)[0]))) return loss def evaluate_scores(self, tst_reader): vid_names, all_scores = [], [] cap_names = tst_reader.dataset.captions for vid_data in tst_reader: vid_names.extend(vid_data['names']) vid_enc_outs = self.forward_video_embed(vid_data) all_scores.append([]) for cap_data in tst_reader.dataset.iterate_over_captions(self.config.tst_batch_size): cap_enc_outs = self.forward_text_embed(cap_data) cap_enc_outs.update(vid_enc_outs) scores = self.generate_scores(**cap_enc_outs) all_scores[-1].append(scores.data.cpu().numpy()) all_scores[-1] = np.concatenate(all_scores[-1], axis=1) all_scores = np.concatenate(all_scores, axis=0) # (n_img, n_cap) return vid_names, cap_names, all_scores def calculate_metrics(self, scores, i2t_gts, t2i_gts): # caption retrieval cr1, cr5, cr10, cmedr, cmeanr = t2vretrieval.models.evaluation.eval_q2m(scores, i2t_gts) # image retrieval ir1, ir5, ir10, imedr, imeanr = t2vretrieval.models.evaluation.eval_q2m(scores.T, t2i_gts) # sum of recalls to be used for early stopping rsum = cr1 + cr5 + cr10 + ir1 + ir5 + ir10 metrics = collections.OrderedDict() metrics['ir1'] = ir1 metrics['ir5'] = ir5 metrics['ir10'] = ir10 metrics['imedr'] = imedr metrics['imeanr'] = imeanr metrics['cr1'] = cr1 metrics['cr5'] = cr5 metrics['cr10'] = cr10 metrics['cmedr'] = cmedr metrics['cmeanr'] = cmeanr metrics['rsum'] = rsum return metrics def evaluate(self, tst_reader, return_outs=False): vid_names, cap_names, scores = self.evaluate_scores(tst_reader) i2t_gts = [] for vid_name in vid_names: i2t_gts.append([]) for i, cap_name in enumerate(cap_names): if cap_name in tst_reader.dataset.ref_captions[vid_name]: i2t_gts[-1].append(i) t2i_gts = {} for i, t_gts in enumerate(i2t_gts): for t_gt in t_gts: t2i_gts.setdefault(t_gt, []) t2i_gts[t_gt].append(i) metrics = self.calculate_metrics(scores, i2t_gts, t2i_gts) if return_outs: outs = { 'vid_names': vid_names, 'cap_names': cap_names, 'scores': scores, } return metrics, outs else: return metrics def validate(self, val_reader, step=None): self.eval_start() metrics = self.evaluate(val_reader) return metrics def test(self, tst_reader, tst_pred_file, tst_model_file=None): if tst_model_file is not None: self.load_checkpoint(tst_model_file) self.eval_start() if tst_reader.dataset.ref_captions is None: vid_names, cap_names, scores = self.evaluate_scores(tst_reader) outs = { 'vid_names': vid_names, 'cap_names': cap_names, 'scores': scores, } metrics = None else: metrics, outs = self.evaluate(tst_reader, return_outs=True) with open(tst_pred_file, 'wb') as f: np.save(f, outs) return metrics
sqiangcao99/hgr_v2t
t2vretrieval/driver/configs/prepare_globalmatch_configs.py
<gh_stars>100-1000 import os import sys import argparse import numpy as np import json import t2vretrieval.models.globalmatch from t2vretrieval.models.globalmatch import VISENC, TXTENC def prepare_mp_globalmatch_model(root_dir): anno_dir = os.path.join(root_dir, 'annotation', 'RET') mp_ft_dir = os.path.join(root_dir, 'ordered_feature', 'MP') split_dir = os.path.join(root_dir, 'public_split') res_dir = os.path.join(root_dir, 'results', 'RET.released') mp_ft_names = ['resnet152.pth'] dim_mp_fts = [np.load(os.path.join(mp_ft_dir, mp_ft_name, 'val_ft.npy')).shape[-1] \ for mp_ft_name in mp_ft_names] num_words = len(np.load(os.path.join(anno_dir, 'int2word.npy'))) model_cfg = t2vretrieval.models.globalmatch.GlobalMatchModelConfig() model_cfg.max_words_in_sent = 30 model_cfg.margin = 0.2 model_cfg.max_violation = True #False model_cfg.hard_topk = 1 model_cfg.loss_direction = 'bi' model_cfg.trn_batch_size = 128 model_cfg.tst_batch_size = 1000 model_cfg.monitor_iter = 1000 model_cfg.summary_iter = 1000 visenc_cfg = model_cfg.subcfgs[VISENC] visenc_cfg.dim_fts = dim_mp_fts visenc_cfg.dim_embed = 1024 visenc_cfg.dropout = 0.2 txtenc_cfg = model_cfg.subcfgs[TXTENC] txtenc_cfg.num_words = num_words txtenc_cfg.dim_word = 300 txtenc_cfg.fix_word_embed = False txtenc_cfg.rnn_hidden_size = 1024 txtenc_cfg.num_layers = 1 txtenc_cfg.rnn_type = 'gru' # lstm, gru txtenc_cfg.bidirectional = True txtenc_cfg.dropout = 0.2 txtenc_name = '%s%s%s'%('bi' if txtenc_cfg.bidirectional else '', txtenc_cfg.rnn_type, '.fix' if txtenc_cfg.fix_word_embed else '') output_dir = os.path.join(res_dir, 'globalmatch', 'mp.vis.%s.txt.%s.%d.loss.%s%s.glove.init'% ('-'.join(mp_ft_names), txtenc_name, visenc_cfg.dim_embed, model_cfg.loss_direction, '.max.%d'%model_cfg.hard_topk if model_cfg.max_violation else '') ) print(output_dir) if not os.path.exists(output_dir): os.makedirs(output_dir) model_cfg.save(os.path.join(output_dir, 'model.json')) path_cfg = { 'output_dir': output_dir, 'mp_ft_files': {}, 'name_file': {}, 'word2int_file': os.path.join(anno_dir, 'word2int.json'), 'int2word_file': os.path.join(anno_dir, 'int2word.npy'), 'ref_caption_file': {}, } for setname in ['trn', 'val', 'tst']: path_cfg['mp_ft_files'][setname] = [ os.path.join(mp_ft_dir, mp_ft_name, '%s_ft.npy'%setname) for mp_ft_name in mp_ft_names ] path_cfg['name_file'][setname] = os.path.join(split_dir, '%s_names.npy'%setname) path_cfg['ref_caption_file'][setname] = os.path.join(anno_dir, 'ref_captions.json') with open(os.path.join(output_dir, 'path.json'), 'w') as f: json.dump(path_cfg, f, indent=2) if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument('root_dir') opts = parser.parse_args() prepare_mp_globalmatch_model(opts.root_dir)
sqiangcao99/hgr_v2t
framework/modules/global_attention.py
<gh_stars>100-1000 """ Global attention modules (Luong / Bahdanau) """ import torch import torch.nn as nn import torch.nn.functional as F class GlobalAttention(nn.Module): ''' Global attention takes a matrix and a query vector. It then computes a parameterized convex combination of the matrix based on the input query. Constructs a unit mapping a query `q` of size `dim` and a source matrix `H` of size `n x dim`, to an output of size `dim`. All models compute the output as :math:`c = sum_{j=1}^{SeqLength} a_j H_j` where :math:`a_j` is the softmax of a score function. However they differ on how they compute the attention score. * Luong Attention (dot, general): * dot: :math:`score(H_j,q) = H_j^T q` * general: :math:`score(H_j, q) = H_j^T W_a q` * Bahdanau Attention (mlp): * :math:`score(H_j, q) = w_a^T tanh(W_a q + U_a h_j)` Args: attn_size (int): dimensionality of query and key attn_type (str): type of attention to use, options [dot,general,mlp] ''' def __init__(self, query_size, attn_size, attn_type='dot'): super(GlobalAttention, self).__init__() self.query_size = query_size self.attn_size = attn_size self.attn_type = attn_type if self.attn_type == 'general': self.linear_in = nn.Linear(query_size, attn_size, bias=False) elif self.attn_type == 'mlp': self.linear_query = nn.Linear(query_size, attn_size, bias=True) self.attn_w = nn.Linear(attn_size, 1, bias=False) elif self.attn_type == 'dot': assert self.query_size == self.attn_size def forward(self, query, memory_keys, memory_values, memory_masks): """ Args: query (`FloatTensor`): (batch, query_size) memory_keys (`FloatTensor`): (batch, seq_len, attn_size) memory_values (`FloatTensor`): (batch, seq_len, attn_size) memory_masks (`LongTensor`): (batch, seq_len) Returns: attn_score: attention distributions (batch, seq_len) attn_memory: computed context vector, (batch, attn_size) """ batch_size, seq_len, attn_size = memory_keys.size() if self.attn_type == 'mlp': query_hidden = self.linear_query(query.unsqueeze(1)).expand( batch_size, seq_len, attn_size) # attn_hidden: # (batch, seq_len, attn_size) attn_hidden = torch.tanh(query_hidden + memory_keys) # attn_score: (batch, seq_len, 1) attn_score = self.attn_w(attn_hidden) elif self.attn_type == 'dot': # attn_score: (batch, seq_len, 1) attn_score = torch.bmm(memory_keys, query.unsqueeze(2)) elif self.attn_type == 'general': query_hidden = self.linear_in(query) attn_score = torch.bmm(memory_keys, query_hidden.unsqueeze(2)) # attn_score: (batch, seq_len) attn_score = attn_score.squeeze(2) if memory_masks is not None: attn_score = attn_score * memory_masks # memory mask [0, 1] attn_score = attn_score.masked_fill(memory_masks == 0, -1e18) attn_score = F.softmax(attn_score, dim=1) # make sure no item is attended when all memory_masks are all zeros if memory_masks is not None: attn_score = attn_score.masked_fill(memory_masks == 0, 0) attn_memory = torch.sum(attn_score.unsqueeze(2) * memory_values, 1) return attn_score, attn_memory #TODO class AdaptiveAttention(nn.Module): def __init__(self, query_size, attn_size): super(AdaptiveAttention, self).__init__() self.query_size = query_size self.attn_size = attn_size self.query_attn_conv = nn.Conv1d(query_size, attn_size, kernel_size=1, stride=1, padding=0, bias=True) self.sentinel_attn_conv = nn.Conv1d(query_size, attn_size, kernel_size=1, stride=1, padding=0, bias=False) self.attn_w = nn.Conv1d(attn_size, 1, kernel_size=1, stride=1, padding=0, bias=False) def forward(self, query, memory_keys, memory_values, memory_masks, sentinel): batch_size, _, enc_seq_len = memory_keys.size() query_hidden = self.query_attn_conv(query.unsqueeze(2)) sentinel_hidden = self.sentinel_attn_conv(sentinel.unsqueeze(2)) memory_keys_sentinel = torch.cat([memory_keys, sentinel_hidden], dim=2) attn_score = self.attn_w(F.tanh(query_hidden + memory_keys_sentinel)).squeeze(1) attn_score = F.softmax(attn_score, dim=1) masks = torch.cat([memory_masks, torch.ones(batch_size, 1).to(memory_masks.device)], dim=1) attn_score = attn_score * masks attn_score = attn_score / (torch.sum(attn_score, 1, keepdim=True) + 1e-10) attn_memory = torch.sum(attn_score[:, :-1].unsqueeze(1) * memory_values, 2) attn_memory = attn_memory + attn_score[:, -1:] * sentinel return attn_score, attn_memory
soumilbaldota/local_full-duplex_chatspace_using_sockets
CLIENT.py
<filename>CLIENT.py from socket import * import threading import time import sys time.sleep(10) def recv(): while True: s=socket(AF_INET,SOCK_STREAM) s.connect((sys.argv[1] ,5747)) s1='' while True: msg=(s.recv(8).decode('utf-8')) s1+=msg if len(msg)<=0: s.close() print(s1) break def send(): s=socket(AF_INET,SOCK_STREAM) s.bind((gethostname(),5743)) s.listen(5) while True: s1=input('>>>') conn,addr=s.accept() conn.send(bytes(s1.encode('utf-8'))) conn.close() t1=threading.Thread(target=recv) t1.start() t2=threading.Thread(target=send) t2.start()
Aghassi/rules_spa
spa/private/build_routes.bzl
<reponame>Aghassi/rules_spa """A macro for creating a webpack federation route module""" load("@aspect_rules_swc//swc:swc.bzl", "swc") # Defines this as an importable module area for shared macros and configs def build_route(name, entry, srcs, data, webpack, federation_shared_config): """ Macro that allows easy composition of routes from a multi route spa Args: name: name of a route (route) entry: the entry file to the route srcs: source files to be transpiled and bundled data: any dependencies the route needs to build webpack: the webpack module to invoke. The users must provide their own load statement for webpack before this macro is called federation_shared_config: a nodejs module file that exposes a map of dependencies to their shared module spec https://webpack.js.org/plugins/module-federation-plugin/#sharing-hints. An example of this is located within this repository under the private/webpack folder. """ build_name = name + "_route" # list of all transpilation targets from SWC to be passed to webpack deps = [ ":transpile_" + files.replace("//", "").replace("/", "_").split(".")[0] for files in srcs ] + data # buildifier: disable=no-effect [ swc( name = "transpile_" + s.replace("//", "").replace("/", "_").split(".")[0], args = [ "-C jsc.parser.jsx=true", "-C jsc.parser.syntax=typescript", "-C jsc.transform.react.runtime=automatic", "-C jsc.transform.react.development=false", "-C module.type=commonjs", ], srcs = [s], ) for s in srcs ] route_config = Label("//spa/private/webpack:webpack.route.config.js") webpack( name = name, args = [ "--env name=" + build_name, "--env entry=./$(execpath :transpile_" + name + ")", "--env SHARED_CONFIG=$(location %s)" % federation_shared_config, "--env BAZEL_SRC_PATH=$(execpath :transpile_" + name + ")", "--output-path=$(@D)", "--config=$(rootpath %s)" % route_config, ], data = [ route_config, federation_shared_config, Label("//spa/private/webpack:webpack.common.config.js"), ] + deps, output_dir = True, visibility = ["//src/client/routes:__pkg__"], )
Aghassi/rules_spa
spa/private/routes/generate_route_manifest.bzl
<filename>spa/private/routes/generate_route_manifest.bzl """A macro that wraps a script to generate an object that maps routes to their entry data""" load("@build_bazel_rules_nodejs//:index.bzl", "nodejs_binary", "npm_package_bin") def generate_route_manifest(name, routes): """Generates a json file that is a representation of all routes. Args: name: name of the invocation routes: the sources from the //src/routes rule to be used in the script Returns: the generated manifest """ nodejs_binary( name = "bin", data = [ Label("//spa/private/routes:generate-route-manifest.js"), ], entry_point = Label("//spa/private/routes:generate-route-manifest.js"), ) npm_package_bin( name = "route_manifest", tool = ":bin", data = [routes], args = [ "$(execpath %s)" % routes, ], stdout = "route.manifest.json", visibility = ["//visibility:public"], )
Aghassi/rules_spa
internal_deps.bzl
<filename>internal_deps.bzl """Our "development" dependencies Users should *not* need to install these. If users see a load() statement from these, that's a bug in our distribution. """ load("@bazel_tools//tools/build_defs/repo:git.bzl", "git_repository") load("@bazel_tools//tools/build_defs/repo:http.bzl", "http_archive") load("@bazel_tools//tools/build_defs/repo:utils.bzl", "maybe") def rules_spa_internal_deps(): "Fetch deps needed for local development" maybe( http_archive, name = "build_bazel_integration_testing", urls = [ "https://github.com/bazelbuild/bazel-integration-testing/archive/165440b2dbda885f8d1ccb8d0f417e6cf8c54f17.zip", ], strip_prefix = "bazel-integration-testing-165440b2dbda885f8d1ccb8d0f417e6cf8c54f17", sha256 = "2401b1369ef44cc42f91dc94443ef491208dbd06da1e1e10b702d8c189f098e3", ) maybe( http_archive, name = "io_bazel_rules_go", sha256 = "2b1641428dff9018f9e85c0384f03ec6c10660d935b750e3fa1492a281a53b0f", urls = [ "https://mirror.bazel.build/github.com/bazelbuild/rules_go/releases/download/v0.29.0/rules_go-v0.29.0.zip", "https://github.com/bazelbuild/rules_go/releases/download/v0.29.0/rules_go-v0.29.0.zip", ], ) maybe( http_archive, name = "bazel_gazelle", sha256 = "de69a09dc70417580aabf20a28619bb3ef60d038470c7cf8442fafcf627c21cb", urls = [ "https://mirror.bazel.build/github.com/bazelbuild/bazel-gazelle/releases/download/v0.24.0/bazel-gazelle-v0.24.0.tar.gz", "https://github.com/bazelbuild/bazel-gazelle/releases/download/v0.24.0/bazel-gazelle-v0.24.0.tar.gz", ], ) # Override bazel_skylib distribution to fetch sources instead # so that the gazelle extension is included # see https://github.com/bazelbuild/bazel-skylib/issues/250 maybe( http_archive, name = "bazel_skylib", sha256 = "07b4117379dde7ab382345c3b0f5edfc6b7cff6c93756eac63da121e0bbcc5de", strip_prefix = "bazel-skylib-1.1.1", urls = [ "https://mirror.bazel.build/github.com/bazelbuild/bazel-skylib/archive/1.1.1.tar.gz", "https://github.com/bazelbuild/bazel-skylib/archive/1.1.1.tar.gz", ], ) maybe( http_archive, name = "io_bazel_stardoc", sha256 = "c9794dcc8026a30ff67cf7cf91ebe245ca294b20b071845d12c192afe243ad72", urls = [ "https://mirror.bazel.build/github.com/bazelbuild/stardoc/releases/download/0.5.0/stardoc-0.5.0.tar.gz", "https://github.com/bazelbuild/stardoc/releases/download/0.5.0/stardoc-0.5.0.tar.gz", ], ) maybe( http_archive, name = "aspect_bazel_lib", sha256 = "8c8cf0554376746e2451de85c4a7670cc8d7400c1f091574c1c1ed2a65021a4c", url = "https://github.com/aspect-build/bazel-lib/releases/download/v0.2.6/bazel_lib-0.2.6.tar.gz", ) # The minimal version of rules_nodejs we need maybe( http_archive, name = "build_bazel_rules_nodejs", sha256 = "ddb78717b802f8dd5d4c01c340ecdc007c8ced5c1df7db421d0df3d642ea0580", urls = ["https://github.com/bazelbuild/rules_nodejs/releases/download/4.6.0/rules_nodejs-4.6.0.tar.gz"], ) # The minimal version of rules_swc that we need maybe( http_archive, name = "aspect_rules_swc", sha256 = "67d6020374627f60c6c1e5d5e1690fcdc4fa39952de8a727d3aabe265ca843be", strip_prefix = "rules_swc-0.1.0", url = "https://github.com/aspect-build/rules_swc/archive/v0.1.0.tar.gz", ) maybe( http_archive, name = "rules_nodejs", sha256 = "005c59bf299d15d1d9551f12f880b1a8967fa883654c897907a667cdbb77c7a6", urls = ["https://github.com/bazelbuild/rules_nodejs/releases/download/4.6.0/rules_nodejs-core-4.6.0.tar.gz"], ) # rules_js is needed for rules_swc maybe( http_archive, name = "aspect_rules_js", sha256 = "dd78b4911b7c2e6c6e919b85cd31572cc15e5baa62b9e7354d8a1065c67136e3", strip_prefix = "rules_js-0.3.1", url = "https://github.com/aspect-build/rules_js/archive/v0.3.1.tar.gz", ) # Rules Docker requirements maybe( http_archive, name = "io_bazel_rules_docker", sha256 = "92779d3445e7bdc79b961030b996cb0c91820ade7ffa7edca69273f404b085d5", strip_prefix = "rules_docker-0.20.0", urls = ["https://github.com/bazelbuild/rules_docker/releases/download/v0.20.0/rules_docker-v0.20.0.tar.gz"], ) # Required to offer local dev with ibazel maybe( git_repository, name = "com_github_ash2k_bazel_tools", commit = "4daedde3ec61a03db841c8a9ca68288972e25a82", remote = "https://github.com/ash2k/bazel-tools.git", )
Aghassi/rules_spa
spa/defs.bzl
<reponame>Aghassi/rules_spa "Public API re-exports" load("//spa/private/routes:generate_route_manifest.bzl", _generate_route_manifest = "generate_route_manifest") load("//spa/private:build_routes.bzl", _build_route = "build_route") load("//spa/private:build_server.bzl", _build_server = "build_server") load("//spa/private:host.bzl", _build_host = "build_host") generate_route_manifest = _generate_route_manifest build_route = _build_route build_server = _build_server build_host = _build_host
Aghassi/rules_spa
spa/private/build_server.bzl
<reponame>Aghassi/rules_spa """A macro for building the server bundle that will end up in a nodejs docker image""" load("@build_bazel_rules_nodejs//:index.bzl", "js_library") load("@aspect_rules_swc//swc:swc.bzl", "swc") # Defines this as an importable module area for shared macros and configs def build_server(name, srcs, data, **kwargs): """ Macro that construct the http server for the project Args: name: name of the package srcs: source files from the package data: any dependencies needed to build **kwargs: any other inputs to js_library """ # list of all transpilation targets from SWC to be passed to webpack deps = [ ":transpile_" + files.replace("//", "").replace("/", "_").split(".")[0] for files in srcs ] # buildifier: disable=no-effect [ swc( name = "transpile_" + s.replace("//", "").replace("/", "_").split(".")[0], args = [ "-C jsc.parser.jsx=true", "-C jsc.parser.syntax=typescript", "-C jsc.target=es2015", "-C module.type=commonjs", ], srcs = [s], ) for s in srcs ] js_library( name = name, srcs = deps + data, **kwargs )
Aghassi/rules_spa
spa/private/host.bzl
<reponame>Aghassi/rules_spa """A macro for creating a webpack federation host module""" load("@aspect_rules_swc//swc:swc.bzl", "swc") load("@build_bazel_rules_nodejs//:index.bzl", "pkg_web") # Defines this as an importable module area for shared macros and configs def build_host(entry, data, srcs, webpack, federation_shared_config): """ Macro that allows easy building of the main host of a SPA In addition to the usage of this macro, you will be required to pass in all required node_modules for compilation as well as anything the shared config relies on (most likely your package.json) into data Args: entry: the entry file to the route data: any dependencies the route needs to build including npm modules srcs: srcs files to be transpiled and sent to webpack webpack: the webpack module to invoke. The users must provide their own load statement for webpack before this macro is called federation_shared_config: a nodejs module file that exposes a map of dependencies to their shared module spec https://webpack.js.org/plugins/module-federation-plugin/#sharing-hints. An example of this is located within this repository under the private/webpack folder. """ # list of all transpilation targets from SWC to be passed to webpack deps = [ ":transpile_" + files.replace("//", "").replace("/", "_").split(".")[0] for files in srcs ] + data # buildifier: disable=no-effect [ swc( name = "transpile_" + s.replace("/", "_").split(".")[0], args = [ "-C jsc.parser.jsx=true", "-C jsc.parser.syntax=typescript", ], srcs = [s], ) for s in srcs ] host_config = Label("//spa/private/webpack:webpack.host.config.js") webpack( name = "host_build", args = [ "--env name=host", "--env entry=./$(location :transpile_host)", "--env SHARED_CONFIG=$(location %s)" % federation_shared_config, "--env BAZEL_SRC_PATH=$(execpath :transpile_host)", "--output-path=$(@D)", "--config=$(rootpath %s)" % host_config, ], data = [ host_config, federation_shared_config, Label("//spa/private/webpack:webpack.common.config.js"), Label("//spa/private/webpack:webpack.module-federation.shared.js"), ] + deps, output_dir = True, ) pkg_web( name = "host", srcs = [ ":host_build", ], additional_root_paths = ["%s/host_build" % native.package_name()], visibility = ["//visibility:public"], )
jpardobl/pandavro
example/example.py
<reponame>jpardobl/pandavro<filename>example/example.py import pandavro as pdx def main(): weather = pdx.from_avro('weather.avro') print(weather) pdx.to_avro('weather_out.avro', weather) if __name__ == '__main__': main()
duhines/CC-M7
action.py
""" Author: <NAME> Course: Computational Creativity Fall 2018 Project: M7: Playing with Words Date: last modified 12/13 Description: Class that allows characters to act. Includes the following classes: Actions - used to store and initialize all the actions Travel - implements the possibilities and consequences of a traveling in the world Injure - implements the consequences of one character injuring another Heal - implements the consequences of one character injuring another Sleep - implements the consequences of one character sleeping Notes: Any class implementing an action is expected to have a do() method that does any book-keeping associated with the action being performed and returns a string representing the consequences of the action. Any action that involves one character impacting another will need to be initialized for all the possible recipients of the action in the get_actions function in Actions. The path_finding function for the Traveling class currently assumes that all the places are connected--this would need to be update with some kind of path finding algorithm if every location was not connected to every other location. Sometimes, the justification function will be called when the clauses in question are all false. No idea why this happens, but it shoudn't. """ import random from random import choice class Actions: """ Class used to store all the actions available to a character. Implements the following method: get_actions """ def __init__(self, character, locations, characters, nearby_characters, social_network): self.character = character self.locations = locations self.characters = characters self.nearby_characters = nearby_characters self.social_network = social_network self.actions = self.get_actions() def get_actions(self): """ Purpose: return the actions available for a character """ # get list of characters that the acting character can interact with other_characters = self.nearby_characters.copy() if self.character in other_characters: other_characters.remove(self.character) actions = [] actions.append(Travel(self.character, self.locations)) actions.append(Sleep(self.character)) # since we can heal/injure other characters, we need an action for each # possible character that can be a recipient of one of these actions for character2 in other_characters: actions.append(Injure(self.character, character2, self.social_network)) actions.append(Heal(self.character, character2, self.social_network)) return actions class Travel: """ Class that implements a traveling action. """ def __init__(self, character, locations): self.character = character self.locations = locations self.pre_condition = self.character.health > 30 def do(self): """ Purpose: either travel randomly or travel due to characters goal """ character = self.character locations = self.locations if self.character.goal != None: if self.character.goal.place != None: character.location = self.path_find(character.goal.place) return character.name + ' traveled to ' + character.location + \ 'in order to ' + character.goal.statement else: options = self.locations['names'].copy() options.remove(self.character.location) if len(options) == 0: return '{} wanders around {}.'.format(self.character.name, self.character.location) character.location = choice(options) return character.name + ' travels to ' + character.location def path_find(self, goal): """ Purpose: if all places were not connected, then we would need to figure out how to get to where we want to go. TODO: extend this to work when there is not an edge between every location! """ return goal class Injure: """ Class that implements the injure action. """ def __init__(self, character1, character2, social_network): self.character1 = character1 self.character2 = character2 self.social_network = social_network self.clause1 = character1.personality.lawful_chaotic < -.5 self.clause2 = character1.personality.good_evil < -.5 self.clause3 = social_network.get_connection(character1, character2).brotherly < -1 self.clause4 = social_network.get_connection(character1, character2).lovers < -1 self.clause5 = character2.health > 1 self.pre_condition = (self.clause1 or self.clause2 or self.clause3 or self.clause4) and self.clause5 def do(self): """ Purpose: one character injures another. """ # TODO change this so that more animosity results in greater injury damage = random.randint(10, 50) self.character2.health -= damage if damage < 30: self.social_network.get_connection(self.character2, self.character1).adjust_all(-1) return '{} injures {} because {}.'.format(self.character1.name, self.character2.name, self.get_justification()) else: self.social_network.get_connection(self.character2, self.character1).adjust_all(-2) return '{} significantly injures {} because {}.'.format( self.character1.name, self.character2.name, self.get_justification()) def get_justification(self): """ Purpose: return a string representing the clause that allowed this action to happen. """ if self.clause1: return '{} is chotic'.format(self.character1.name) elif self.clause2: return '{} is evil'.format(self.character1.name) elif self.clause3: return '{} feels brotherly hate towards {}'.format( self.character1.name, self.character2.name) elif self.clause4: return '{} feels lover\'s hate towards {}'.format( self.character1.name, self.character2.name) else: return '{} made a mistake'.format(self.character1.name) class Heal: """ Class for the heal action. """ def __init__(self, character1, character2, social_network): self.character1 = character1 self.character2 = character2 self.social_network = social_network self.clause1 = character1.personality.lawful_chaotic > .5 self.clause2 = character1.personality.good_evil > .5 self.clause3 = social_network.get_connection(character1, character2).brotherly >= 1 self.clause4 = social_network.get_connection(character1, character2).lovers >= 1 self.clause5 = character2.health < 100 and self.character2.health > 0 self.pre_condition = (self.clause1 or self.clause2 or self.clause3 or self.clause4) and self.clause5 def do(self): """ Purpose: one character heals another character """ heal = random.randint(10, 50) self.character2.health += heal if heal < 30: self.social_network.get_connection(self.character2, self.character1).adjust_all(1) return '{} heals {} because {}.'.format(self.character1.name, self.character2.name, self.get_justification()) else: self.social_network.get_connection(self.character2, self.character1).adjust_all(2) return '{} significantly heals {} because {}.'.format( self.character1.name, self.character2.name, self.get_justification()) def get_justification(self): """ Purpose: return a string representing the clause that allowed this action to happen. """ if self.clause1: return '{} is lawful'.format(self.character1.name) elif self.clause2: return '{} is good'.format(self.character1.name) elif self.clause3: return '{} feels brotherly love towards {}'.format( self.character1.name, self.character2.name) elif self.clause4: return '{} feels lover\'s love towards {}'.format( self.character1.name, self.character2.name) else: return '{} made a mistake'.format(self.character1.name) class Sleep: """ Class for the sleep action. """ def __init__(self, character): self.character = character self.pre_condition = self.character.health > 20 def do(self): """ Purpose: sleep to regain some health. """ self.character.health += 30 return '{} sleeps to regain some strength.'.format(self.character.name)
duhines/CC-M7
knowledge/define_actions.py
import action actions = [] """ Basic actions: move from one place to another fall in love with another character attack another character heal another character """
duhines/CC-M7
quests.py
<filename>quests.py # THIS WHOLE FILE IS A BIG TODO #""" # Author: <NAME> # Course: Computational Creativity Fall 2018 # Project: M7: Playing with Words # Date: last modified 12/13 # Description: # Notes: # """ # """ # Want characters to have objectives--when a characters wishes to accomplish # something in the long term, then they will gain a quest or the narrator # will give characters quests by introducing random events. # """ # class Quest: # def __init__(self, statement): # self.statement = statement
duhines/CC-M7
character.py
""" Author: <NAME> Course: Computational Creativity Fall 2018 Project: M7: Playing with Words Date: last modified 12/13 Description: Notes: * personality representation: * -1 to 1 rating of the Big five personality traints * and/or -1 to 1 rating of evil->good, chaotic->lawful scale. * perhaps emotions are a modifier for actions: * ____ washed the dishes, mod: ANGRILY/SADLY/... """ import random import action as action_stuff class Personality: """ Class for representing a character's personality. Personality is quantified similarly to DnD (or the Bowdoin outing club LT exercies) where there are two personality parameters, one measures a scale of good to evil, and the other measures a scale of lawful to chaotic. These values are represented on a scale from -1 to 1: full good would be a value of 1, full lawful would be a value of 1 full evil would be a value of -1, full chaotic would be a value of -1 Implements the following methods: rand_personality for_narrative """ def __init__(self): self.good_evil = self.rand_personality() self.lawful_chaotic = self.rand_personality() # TODO add the ability fo rpersonalities to change over time # self.theta = random.random() + .01 # self.firmness = random.random() def rand_personality(self): """ Purpose: return random value [-1, 1] """ absolute_value = random.random() value = None if random.random() < .5: value = absolute_value * -1 else: value = absolute_value return value def for_narrative(self): """ Purpose: return string representation of the personality. """ good = '' lawful = '' if self.good_evil < -.75: good = 'very evil' elif self.good_evil < -.5: good = 'evil' elif self.good_evil < -.25: good = 'sort of evil' elif self.good_evil < .25 and self.good_evil > -.25: good = 'neither good nor evil' elif self.good_evil < .5: good = 'sort of good' elif self.good_evil < .75: good = 'good' elif self.good_evil < 1: good = 'very good' lawful = '' if self.lawful_chaotic < -.75: lawful = 'very chaotic' elif self.lawful_chaotic < -.5: lawful = 'chaotic' elif self.lawful_chaotic < -.25: lawful = 'sort of chaotic' elif self.lawful_chaotic < .25 and self.lawful_chaotic > -.25: lawful = 'neither lawful nor chaotic' elif self.lawful_chaotic < .5: lawful = 'sort of lawful' elif self.lawful_chaotic < .75: lawful = 'lawful' elif self.lawful_chaotic < 1: lawful = 'very lawful' return '{} and {}'.format(good, lawful) # This would be a big TODO if there was more time # def adjust_personality(self, demo_good_evil, demo_lawful_chaotic): # """ # Purpose: If a character ends up making an action that contradicts # their personality, then adjust their personality values with # respect to this characters personality maleability (theta). Note, # characters make actions outside their # """ # self.good_evil = self.demo_good_evil * (1 - self.theta) + demo_good_evil * self.theta # self.lawful_chaotic = self.lawful_chaotic * (1 - self.theta) + demo_lawful_chaotic * self.theta class Connection: """ Purpose: detail the connection between one character and another. Currently using the 4 types model from MEXICA: 1. brotherly-love / hate 2. lover's-love / hate 3. gratefulness / ingratitude 4. admiration, respect / disdain, disapproval Includes the following methods: init_value adjust_all adjust_brotherly adjust_lovers adjust_gratefulness adjust_admiration normalize_value get_string get_values """ def __init__(self): self.brotherly = self.init_value() self.lovers = self.init_value() self.gratefulness = self.init_value() self.admiration = self.init_value() def init_value(self): """ Purpose: return an integer between -3 and 3 for now """ absolute_value = random.randint(0, 3) value = None if random.random() < .5: value = absolute_value * -1 else: value = absolute_value return value def adjust_all(self, amount): """ Purpose: adjust all the connection values. """ self.adjust_brotherly(amount) self.adjust_lovers(amount) self.adjust_gratefulness(amount) self.adjust_admiration(amount) def adjust_brotherly(self, amount): """ Purpose: adjust a single connection value. """ self.brotherly = self.normalize_value(self.brotherly + amount) def adjust_lovers(self, amount): """ Purpose: adjust a single connection value. """ self.lovers = self.normalize_value(self.lovers + amount) def adjust_gratefulness(self, amount): """ Purpose: adjust a single connection value. """ self.gratefulness = self.normalize_value(self.gratefulness + amount) def adjust_admiration(self, amount): """ Purpose: adjust a single connection value. """ self.admiration = self.normalize_value(self.admiration + amount) def normalize_value(self, value): """ Purpose: adjust a value so that it remains in the range[-3, 3] TODO: expand this so it doesnt use magic numbers """ # keep values in the bounds (sorry that these are magic numbers) # TODO: flesh out connection system more if value > 3: return 3 if value < -3: return -3 return value def get_string(self): """ Purpose: return a string representation of a connection. brotherly-love / hate 2. lover's-love / hate 3. gratefulness / ingratitude 4. admiration, respect / disdain, disapproval """ string = 'brotherly-love / hate: {}, lover\'s-love / hate: {},' +\ ' gratefulness/ ingratitude: {}, respect / disdain: {}. ' return string.format(self.brotherly, self.lovers, self.gratefulness, self.admiration) def get_values(self): """ Purpose: return a list of the 4 connection parameters. """ return [self.brotherly, self.lovers, self.gratefulness, self.admiration] class SocialNetwork: """ Purpose: capture characters knowledge of each other and feelings towards the other characters. Use emotional links inspired by the MEXICA system: network = [character1_name->character2_name: connection] Includes the following methods: generate_network get_connection for_narrative for_fitness """ def __init__(self, characters): self.characters = characters self.network = self.generate_network() def generate_network(self): """ Purpose: generate a connection between all the characters """ network = {} for character1 in self.characters: all_other_characters = self.characters.copy() all_other_characters.remove(character1) for character2 in all_other_characters: # instantiate with a random connection network[character1.name+character2.name] = Connection() return network def get_connection(self, character1, character2): """ Purpose: get the connection betweeen two characters. """ connection = self.network[character1.name+character2.name] return connection def for_narrative(self): """ Purpose: return a string summarizing the characters connections. """ connection_details = [] for character1 in self.characters: all_other_characters = self.characters.copy() all_other_characters.remove(character1) for character2 in all_other_characters: connection = self.get_connection(character1, character2) connection_details.append('Connection from {} to {} is {}'\ .format(character1.name, character2.name, connection.get_string())) return connection_details def for_fitness(self): """ Purpose: return the connection information in a way that's easy to use to evaluate narrative fitness. """ connection_details = [] for character1 in self.characters: all_other_characters = self.characters.copy() all_other_characters.remove(character1) for character2 in all_other_characters: connection = self.get_connection(character1, character2) connection_details.append(connection.get_values()) return connection_details class Character: """ Character class that includes the following methods: check_health can_act action_maker tell_personality """ def __init__(self, name, personality, goal, health, location): self.name = name self.personality = personality self.goal = goal self.health = health self.location = location self.alive = True # TODO: maybe add emotions: self.emotions = emotions def __str__(self): return self.name def __repr__(self): return 'Character({}, {}, {}, {}, {})'.format(self.name, self.personality, self.goal, self.health, self.location) def check_health(self): """ Purpose: if a character has less than 1 health, then they are dead. """ if self.health < 1: self.alive = False def can_act(self, locations, characters, nearby_characters, social_network): """ Purpose: check to see if this character can do anything. """ all_actions = action_stuff.Actions(self, locations, characters, nearby_characters, social_network).actions for action in all_actions: if not action.pre_condition: all_actions.remove(action) if len(all_actions) == 0: return False return True def action_maker(self, locations, characters, nearby_characters, social_network): """ Purpose: choose an action that works the character towards their goal and is consistent with personality + emotions. """ self.health -= 5 self.check_health() if not self.alive: return '{} has died.'.format(self.name) all_actions = action_stuff.Actions(self, locations, characters, nearby_characters, social_network).actions for action in all_actions: if not action.pre_condition: all_actions.remove(action) if len(all_actions) == 0: return '{} does nothing.'.format(self.name) act = random.choice(all_actions) return act.do() def tell_personality(self): """ Purpose: return a string epressing this characters personality. """ return '{} is {}.'.format(self.name, self.personality.for_narrative())
duhines/CC-M7
clean_locations.py
<filename>clean_locations.py """ Author: <NAME> Course: Computational Creativity Fall 2018 Project: M7: Playing with Words Date: last modified 12/13 Description: This script is used to clean the originally copied and pasted location data. Notes: The way that locations are processed, each word of a multi-word location becomes a unique location. This results in some odd location names, but simplifies the locations by ensuring that all of them are just a single word. """ file = open('knowledge/locations.txt') locations = [] locations_so_far = set() for line in file.readlines(): as_string = str(line).replace('*', '') words = as_string.split() for word in words: if word not in locations_so_far: locations.append(word + '\n') locations_so_far.add(word) write_to = open('knowledge/cleaned_locations.txt', 'w') for location in locations: write_to.write(location)
duhines/CC-M7
knowledge.py
<filename>knowledge.py """ Author: <NAME> Course: Computational Creativity Fall 2018 Project: M7: Playing with Words Date: last modified 12/13 Description: File to set up knowledge by reading in information from a names and locations file in the knowledge base. Notes: Assumes that the names and locations file are a each a series of items separated by a newline. """ LOCATIONS_FILE = 'cleaned_locations.txt' NAMES_FILES = 'cleaned_names.txt' ACTIONS_FILE = 'actions.json' EXTENTION = 'knowledge/' class Knowledge: def __init__(self): self.names = self.get_names() self.locations = self.get_locations() def get_names(self): """ Purpose: read in names from the cleaned_names files and save them in a list object. """ file = open(EXTENTION + NAMES_FILES) names = [] for line in file.readlines(): names.append(line.strip()) return names def get_locations(self): """ Purpose: read in the locations from the cleaned_locations file and load them into a list object. The relationship between locations will be established by the narrator. """ file = open(EXTENTION + LOCATIONS_FILE) locations = [] for line in file.readlines(): locations.append(line.strip()) return locations
duhines/CC-M7
narrator.py
<reponame>duhines/CC-M7 """ Author: <NAME> Course: Computational Creativity Fall 2018 Project: M7: Playing with Words Date: last modified 12/13 Description: This module implements 3 classes: 1. Narrator - the entity that maintains the story knowledge such as setting and characters and keeps track of other meta data as the narrative progresses 2. Event - used an an object to store the time and event for a narrative event. 3. Episode - Structure that uses the narrator to generate a narrative of some length. Notes: * output will be a text script of TV episodes (maybe build this up to seasons?) """ import random import knowledge as knows import character from random import choice import os NUM_CHARACTERS = 10 NUM_LOCATIONS = 5 RATE_LOCATION_CONNECTED = 1 NUM_ACTIONS = 25 NUM_SCRIPTS = 100 RESULTS_PATH = 'results' RESULTS_NAME = 'script_' class Narrator: """ Purpose: control the flow of the story / keep track of story information Includes the following methods: characters_at can_act change_location init_locations init_characters """ def __init__(self): self.knowledge = knows.Knowledge() self.locations = self.init_locations() self.characters = self.init_characters() self.current_location = choice(self.locations['names']) self.social_network = character.SocialNetwork(self.characters) self.story_over = False def characters_at(self, location): """ Purpose: return a list of characters in a location. """ characters = [] for character in self.characters: if character.location == location: characters.append(character) return characters def can_act(self, nearby_characters): """ Purpose: return a sublist of characters than can actually make an action. """ can_act = nearby_characters.copy() for character in nearby_characters: if not character.can_act(self.locations, self.characters, nearby_characters, self.social_network): can_act.remove(character) if not character.alive: can_act.remove(character) return can_act def change_location(self): """ Purpose: change the current_location to a new location. """ possibilities = self.locations['names'].copy() # don't want to switch to the current location possibilities.remove(self.current_location) # don't want to focus on a location without any characters with_characters = possibilities.copy() for possibility in possibilities: if self.characters_at(possibility) == []: with_characters.remove(possibility) if len(possibilities) == 0: return 'There are no characters left.' else: new_location = choice(with_characters) self.current_location = new_location return 'We now shift our attention to {}'.format(self.current_location) def init_locations(self, num_locations=NUM_LOCATIONS): """ Purpose: create a kind of world map where there are locations from the knowledge base connected by edges. """ chosen_locations = [] for i in range(0, num_locations): location = choice(self.knowledge.locations) self.knowledge.locations.remove(location) chosen_locations.append(location) # dictionary to express connections between locations: # connections['location'] = [True, True, False, False,...] # ->True if there is a connection between the two locations, false # otherwise connections = {} for location in chosen_locations: options = chosen_locations.copy() options.remove(location) connected = [] for option in options: connected.append(random.random() < RATE_LOCATION_CONNECTED) connections[location] = connected locations = { 'names': chosen_locations, 'connections': connections } return locations def init_characters(self, num_characters=NUM_CHARACTERS): """ Purpose: create some character objects """ characters = [] for i in range(0, num_characters): # def __init__(self, name, personality, goal, health, location): name = choice(self.knowledge.names) # don't want multiple characters to have the same name self.knowledge.names.remove(name) personality = character.Personality() goal = None health = random.randint(0, 100) location = choice(self.locations['names']) characters.append(character.Character(name, personality, goal, health, location)) return characters class Event: """ Keep track of the time that an event occured during. Using a class because dot notation is better than dictionaries. """ def __init__(self, time, action): self.time = time self.action = action # TODO: flesh things out so that the same narrator can create multiple episodes # of narrative (as long as there are still characters left!) # class Season: # """ # Purpose: preserve details of characters and setting accross episodes # """ # def __init__(self, characters, setting): # return class Episode: """ Purpose: generated scripts organized into episodes, multiple episodes can be told by the same narrator (i.e. same characters and setting). Includes the following methods: write_narrative write_script get_narrative_from_action evaluate output_script """ def __init__(self, narrator, num_actions=NUM_ACTIONS): self.narrator = narrator self.narrative = [] self.script = [] self.num_actions = num_actions self.time = 0 self.starting_connections = [] self.final_connections = [] self.score = -1 def write_narrative(self): """ Purpose: determine the sequences of actions that make up the script. """ narrator = self.narrator curr_acts = 0 events = [] opener = 'We begin our story in {}. '.format(narrator.current_location) for character in narrator.characters_at(narrator.current_location): opener += '{} is here. '.format(character) opener += '\n' # this would be used for a string representation of the social # structure, but we don't use this initial_social_structre = narrator.social_network.for_narrative() self.starting_connections = narrator.social_network.for_fitness() for character in narrator.characters: opener += character.tell_personality() + '\n' events.append(Event(0, opener)) while curr_acts < self.num_actions: # want to only be focused on locations where there are characters than # can do things possible_characters = narrator.characters_at(narrator.current_location) can_act = narrator.can_act(possible_characters) if len(can_act) == 0: change_narrative_location = narrator.change_location() events.append(Event(self.time, change_narrative_location)) if change_narrative_location == 'There are no characters left.': # the story is OVER narrator.story_over = True break else: next_actor = choice(can_act) act = next_actor.action_maker(narrator.locations, narrator.characters, possible_characters, narrator.social_network) events.append(Event(self.time, act)) curr_acts += 1 self.time += 1 self.final_connections = narrator.social_network.for_fitness() closer = 'And thus ends the episode. \n' # this would be used for a string representation of the social # structure, but we don't use this end_social_structre = narrator.social_network.for_narrative() events.append(Event(self.time, closer)) self.narrative = events def write_script(self): """ Purpose: from a list of actions, get the narrative elements from each event. """ for event in self.narrative: self.script.append(self.get_narrative_from_action(event)) def get_narrative_from_action(self, event): """ Purpose: translate a event into a string representing the event. """ # TODO: set up way of getting narrative form that's more interesting # (BIG TODO) return str(event.action.strip()) def evalutate(self): """ Purpose: evalutate an episode of the script by summing the change in the connection parameters between the characters. """ changes = [] total_theta = 0 for i in range(0, len(self.starting_connections)): for j in range(0, len(self.starting_connections[i])): total_theta += abs(self.starting_connections[i][j] - self.final_connections[i][j]) self.score = total_theta return total_theta def output_script(self): """ Purpose: write the script to the output folder. """ # list the number of things in results folder, use that number # for a unique file name num_results = len(os.listdir(RESULTS_PATH)) file_name = RESULTS_NAME + str(num_results) + '.txt' file = open(RESULTS_PATH + '/' + file_name, 'w') for event in self.script: file.write(event + '\n') def main(): narrator = Narrator() episodes = [] bsf = [] bsf_score = 0 # generate 100 scripts using this narrator and choose the one with the # highest fitness, which is currently a measure of how dramatically # character connection values changed from the start of the episode # to the end. for i in range(0, NUM_SCRIPTS): episode = Episode(narrator) episode.write_narrative() episode.write_script() episode.evalutate() if episode.score > bsf_score: bsf = episode bsf_score = episode.score bsf.output_script() if __name__ == '__main__': main()
duhines/CC-M7
clean_names.py
""" Author: <NAME> Course: Computational Creativity Fall 2018 Project: M7: Playing with Words Date: last modified 12/13 Description: This is a script to clean names. The names are from the top 100 boys and girls names from 2018 from bounty.com Notes: Assumes the format of the original names file has the name first on each line. """ file = open('knowledge/names.txt') cleaned = [] for line in file.readlines(): if line == '\n': continue cleaned_name = line.split()[0] cleaned.append(cleaned_name + '\n') file.close() cleaned_file = open('knowledge/cleaned_names.txt', 'w') for name in cleaned: cleaned_file.write(name) cleaned_file.close()
GoHelpFund/aden_hash
setup.py
from distutils.core import setup, Extension help_hash_module = Extension('help_hash', sources = ['helpmodule.c', 'help.c', 'sha3/blake.c', 'sha3/bmw.c', 'sha3/groestl.c', 'sha3/jh.c', 'sha3/keccak.c', 'sha3/skein.c', 'sha3/cubehash.c', 'sha3/echo.c', 'sha3/luffa.c', 'sha3/simd.c', 'sha3/shavite.c'], include_dirs=['.', './sha3']) setup (name = 'help_hash', version = '1.3.1', description = 'Binding for Help X11 proof of work hashing.', ext_modules = [help_hash_module])
GoHelpFund/aden_hash
test.py
<filename>test.py import help_hash from binascii import unhexlify, hexlify import unittest # help block #1 # moo@b1:~/.help$ helpd getblockhash 1 # 000007d91d1254d60e2dd1ae580383070a4ddffa4c64c2eeb4a2f9ecc0414343 # moo@b1:~/.help$ helpd getblock 000007d91d1254d60e2dd1ae580383070a4ddffa4c64c2eeb4a2f9ecc0414343 # { # "hash" : "000007d91d1254d60e2dd1ae580383070a4ddffa4c64c2eeb4a2f9ecc0414343", # "confirmations" : 169888, # "size" : 186, # "height" : 1, # "version" : 2, # "merkleroot" : "ef3ee42b51e2a19c4820ef182844a36db1201c61eb0dec5b42f84be4ad1a1ca7", # "tx" : [ # "ef3ee42b51e2a19c4820ef182844a36db1201c61eb0dec5b42f84be4ad1a1ca7" # ], # "time" : 1390103681, # "nonce" : 128987, # "bits" : "1e0ffff0", # "difficulty" : 0.00024414, # "previousblockhash" : "00000ffd590b1485b3caadc19b22e6379c733355108f107a430458cdf3407ab6", # "nextblockhash" : "00000bafcc571ece7c5c436f887547ef41b574e10ef7cc6937873a74ef1efeae" # } header_hex = ("02000000" + "b67a40f3cd5804437a108f105533739c37e6229bc1adcab385140b59fd0f0000" + "a71c1aade44bf8425bec0deb611c20b16da3442818ef20489ca1e2512be43eef" "814cdb52" + "f0ff0f1e" + "dbf70100") best_hash = '434341c0ecf9a2b4eec2644cfadf4d0a07830358aed12d0ed654121dd9070000' class TestSequenceFunctions(unittest.TestCase): def setUp(self): self.block_header = unhexlify(header_hex) self.best_hash = best_hash def test_help_hash(self): self.pow_hash = hexlify(help_hash.getPoWHash(self.block_header)) self.assertEqual(self.pow_hash.decode(), self.best_hash) if __name__ == '__main__': unittest.main()
LoYungSum/gaze_correction_loyungsum
gaze_correction_system/flx.py
import tensorflow as tf import numpy as np import tf_utils import transformation img_crop = 3 def gen_agl_map(inputs, height, width,feature_dims): with tf.name_scope("gen_agl_map"): batch_size = tf.shape(inputs)[0] ret = tf.reshape(tf.tile(inputs,tf.constant([1,height*width])), [batch_size,height,width,feature_dims]) return ret def encoder(inputs, height, width, tar_dim): with tf.variable_scope('encoder'): dnn_blk_0 = tf_utils.dnn_blk(inputs, 16, name = 'dnn_blk_0') dnn_blk_1 = tf_utils.dnn_blk(dnn_blk_0, 16, name = 'dnn_blk_1') dnn_blk_2 = tf_utils.dnn_blk(dnn_blk_1, tar_dim, name = 'dnn_blk_2') agl_map = gen_agl_map(dnn_blk_2, height, width, tar_dim) return agl_map def apply_lcm(batch_img, light_weight): with tf.name_scope('apply_lcm'): img_wgts, pal_wgts = tf.split(light_weight, [1,1], 3) img_wgts = tf.tile(img_wgts, [1,1,1,3]) pal_wgts = tf.tile(pal_wgts, [1,1,1,3]) palette = tf.ones(tf.shape(batch_img), dtype = tf.float32) ret = tf.add(tf.multiply(batch_img, img_wgts), tf.multiply(palette, pal_wgts)) return ret def trans_module(inputs, structures, phase_train, name="trans_module"): with tf.variable_scope(name) as scope: cnn_blk_0 = tf_utils.cnn_blk(inputs, structures['depth'][0], structures['filter_size'][0], phase_train, name = 'cnn_blk_0') cnn_blk_1 = tf_utils.cnn_blk(cnn_blk_0, structures['depth'][1], structures['filter_size'][1], phase_train, name = 'cnn_blk_1') cnn_blk_2 = tf_utils.cnn_blk(tf.concat([cnn_blk_0,cnn_blk_1], axis=3), structures['depth'][2], structures['filter_size'][2], phase_train, name = 'cnn_blk_2') cnn_blk_3 = tf_utils.cnn_blk(tf.concat([cnn_blk_0,cnn_blk_1,cnn_blk_2], axis=3), structures['depth'][3], structures['filter_size'][3], phase_train, name = 'cnn_blk_3') cnn_4 = tf.layers.conv2d(inputs=cnn_blk_3, filters=structures['depth'][4], kernel_size=structures['filter_size'][4], padding="same", activation=None, use_bias=False, name="cnn_4") return cnn_4 def lcm_module(inputs, structures, phase_train, name="lcm_module"): with tf.variable_scope(name) as scope: cnn_blk_0 = tf_utils.cnn_blk(inputs, structures['depth'][0], structures['filter_size'][0], phase_train, name = 'cnn_blk_0') cnn_blk_1 = tf_utils.cnn_blk(cnn_blk_0, structures['depth'][1], structures['filter_size'][1], phase_train, name = 'cnn_blk_1') cnn_2 = tf.layers.conv2d(inputs=cnn_blk_1, filters=structures['depth'][2], kernel_size=structures['filter_size'][2], padding="same", activation=None, use_bias=False, name='cnn_2') lcm_map = tf.nn.softmax(cnn_2) return lcm_map def inference(input_img, input_fp, input_agl, phase_train, conf): """Build the Deepwarp model. Args: images, anchors_map of eye, angle Returns: lcm images """ corse_layer = {'depth':(32,64,64,32,16), 'filter_size':([5,5],[3,3],[3,3],[3,3],[1,1])} fine_layer = {'depth':(32,64,32,16,4), 'filter_size':([5,5],[3,3],[3,3],[3,3],[1,1])} lcm_layer = {'depth':(8,8,2), 'filter_size':([3,3],[3,3],[1,1])} with tf.variable_scope('warping_model'): agl_map = encoder(input_agl, conf.height, conf.width, conf.encoded_agl_dim) igt_inputs = tf.concat([input_img, input_fp, agl_map],axis=3) with tf.variable_scope('warping_module'): '''coarse module''' resized_igt_inputs = tf.layers.average_pooling2d(inputs=igt_inputs, pool_size=[2,2], strides=2, padding='same') cours_raw = trans_module(resized_igt_inputs, corse_layer, phase_train, name='coarse_level') cours_act = tf.nn.tanh(cours_raw) coarse_resize = tf.image.resize_images(cours_act, (conf.height, conf.width), method=tf.image.ResizeMethod.NEAREST_NEIGHBOR) coarse_out = tf.layers.average_pooling2d(inputs=coarse_resize, pool_size=[2, 2], strides=1, padding='same') '''fine module''' fine_input = tf.concat([igt_inputs, coarse_out],axis=3, name='fine_input') fine_out = trans_module(fine_input, fine_layer, phase_train, name='fine_level') flow_raw, lcm_input = tf.split(fine_out, [2,2], 3) flow = tf.nn.tanh(flow_raw) cfw_img = transformation.apply_transformation(flows = flow, img = input_img, num_channels=3) '''lcm module''' lcm_map = lcm_module(lcm_input, lcm_layer, phase_train, name="lcm_module") img_pred = apply_lcm(batch_img=cfw_img, light_weight=lcm_map) return img_pred, flow_raw, lcm_map def dist_loss(y_pred, y_, method="MAE"): with tf.variable_scope('img_dist_loss'): loss = 0 if(method == "L2"): loss = tf.sqrt(tf.reduce_sum(tf.square(y_pred - y_), axis=3, keep_dims = True)) elif (method == "MAE"): loss = tf.abs(y_pred - y_) loss = loss[:,img_crop:(-1)*img_crop,img_crop:(-1)*img_crop,:] loss = tf.reduce_sum(loss, axis = [1,2,3]) return tf.reduce_mean(loss, axis=0) def TVloss(inputs): with tf.variable_scope('TVloss'): dinputs_dx = inputs[:, :-1, :, :] - inputs[:, 1:, :, :] dinputs_dy = inputs[:, :, :-1, :] - inputs[:, :, 1:, :] dinputs_dx = tf.pad(dinputs_dx, [[0,0],[0, 1], [0, 0],[0,0]], "CONSTANT") dinputs_dy = tf.pad(dinputs_dy, [[0,0],[0, 0], [0, 1],[0,0]], "CONSTANT") tot_var = tf.add(tf.abs(dinputs_dx), tf.abs(dinputs_dy)) tot_var = tf.reduce_sum(tot_var, axis =3, keep_dims=True) return tot_var def TVlosses(eye_mask, ori_img, flow, lcm_map): with tf.variable_scope('TVlosses'): # eyeball_TVloss # calculate TV (dFlow(p)/dx + dFlow(p)/dy) TV_flow = TVloss(flow) # calculate the (1-D(p)) img_gray = tf.reduce_mean(ori_img, axis = 3, keep_dims=True) ones = tf.ones(shape = tf.shape(img_gray)) bright = ones - img_gray # calculate the F_e(p) eye_mask = tf.expand_dims(eye_mask, axis = 3) weights = tf.multiply(bright,eye_mask) TV_eye = tf.multiply(weights,TV_flow) # eyelid_TVloss lid_mask = ones - eye_mask TV_lid = tf.multiply(lid_mask,TV_flow) TV_eye = tf.reduce_sum(TV_eye, axis = [1,2,3]) TV_lid = tf.reduce_sum(TV_lid, axis = [1,2,3]) # lcm_map loss dist2cent = center_weight(tf.shape(lcm_map), base=0.005, boundary_penalty=3.0) TV_lcm = dist2cent*TVloss(lcm_map) TV_lcm = tf.reduce_sum(TV_lcm, axis = [1,2,3]) return tf.reduce_mean(TV_eye, axis=0), tf.reduce_mean(TV_lid, axis=0), tf.reduce_mean(TV_lcm, axis=0) def center_weight(shape, base=0.005, boundary_penalty=3.0): with tf.variable_scope('center_weight'): temp = boundary_penalty - base x = tf.pow(tf.abs(tf.lin_space(-1.0, 1.0,shape[1])),8) y = tf.pow(tf.abs(tf.lin_space(-1.0, 1.0,shape[2])),8) X, Y = tf.meshgrid(y, x) X = tf.expand_dims(X, axis=2) Y = tf.expand_dims(Y, axis=2) dist2cent = temp*tf.sqrt(tf.reduce_sum(tf.square(tf.concat([X,Y], axis=2)), axis=2)) + base dist2cent = tf.expand_dims(tf.tile(tf.expand_dims(dist2cent, axis=0), [shape[0],1,1]), axis=3) return dist2cent def lcm_adj(lcm_wgt): dist2cent = center_weight(tf.shape(lcm_wgt), base=0.005, boundary_penalty=3.0) with tf.variable_scope('lcm_adj'): _, loss = tf.split(lcm_wgt, [1,1], 3) loss = tf.reduce_sum(tf.abs(loss)*dist2cent, axis = [1,2,3]) return tf.reduce_mean(loss, axis=0) def loss(img_pred, img_, eye_mask, input_img, flow, lcm_wgt): with tf.variable_scope('losses'): loss_img = dist_loss(img_pred, img_, method = "L2") loss_eyeball, loss_eyelid, loss_lcm= TVlosses(eye_mask, input_img, flow, lcm_wgt) loss_lcm_adj = lcm_adj(lcm_wgt) losses = loss_img + loss_eyeball + loss_eyelid + loss_lcm_adj + loss_lcm tf.add_to_collection('losses', losses) return tf.add_n(tf.get_collection('losses'), name='total_loss'), loss_img
LoYungSum/gaze_correction_loyungsum
training/code_tf/model_train/transformation.py
import tensorflow as tf def repeat(x, num_repeats): with tf.name_scope("repeat"): ones = tf.ones((1, num_repeats), dtype='int32') x = tf.reshape(x, shape=(-1,1)) x = tf.matmul(x, ones) return tf.reshape(x, [-1]) def interpolate(image, x, y, output_size): with tf.name_scope("interpolate"): batch_size = tf.shape(image)[0] height = tf.shape(image)[1] width = tf.shape(image)[2] num_channels = tf.shape(image)[3] x = tf.cast(x , dtype='float32') y = tf.cast(y , dtype='float32') height_float = tf.cast(height, dtype='float32') width_float = tf.cast(width, dtype='float32') output_height = output_size[0] output_width = output_size[1] x = .5*(x + 1.0)*(width_float) y = .5*(y + 1.0)*(height_float) x0 = tf.cast(tf.floor(x), 'int32') x1 = x0 + 1 y0 = tf.cast(tf.floor(y), 'int32') y1 = y0 + 1 max_y = tf.cast(height - 1, dtype='int32') max_x = tf.cast(width - 1, dtype='int32') zero = tf.zeros([], dtype='int32') x0 = tf.clip_by_value(x0, zero, max_x) x1 = tf.clip_by_value(x1, zero, max_x) y0 = tf.clip_by_value(y0, zero, max_y) y1 = tf.clip_by_value(y1, zero, max_y) flat_image_dimensions = height*width pixels_batch = tf.range(batch_size)*flat_image_dimensions flat_output_dimensions = output_height*output_width base = repeat(pixels_batch, flat_output_dimensions) base_y0 = base + y0*width base_y1 = base + y1*width indices_a = base_y0 + x0 indices_b = base_y1 + x0 indices_c = base_y0 + x1 indices_d = base_y1 + x1 flat_image = tf.reshape(image, shape=(-1, num_channels)) flat_image = tf.cast(flat_image, dtype='float32') pixel_values_a = tf.gather(flat_image, indices_a) pixel_values_b = tf.gather(flat_image, indices_b) pixel_values_c = tf.gather(flat_image, indices_c) pixel_values_d = tf.gather(flat_image, indices_d) x0 = tf.cast(x0, 'float32') x1 = tf.cast(x1, 'float32') y0 = tf.cast(y0, 'float32') y1 = tf.cast(y1, 'float32') area_a = tf.expand_dims(((x1 - x) * (y1 - y)), 1) area_b = tf.expand_dims(((x1 - x) * (y - y0)), 1) area_c = tf.expand_dims(((x - x0) * (y1 - y)), 1) area_d = tf.expand_dims(((x - x0) * (y - y0)), 1) output = tf.add_n([area_a*pixel_values_a, area_b*pixel_values_b, area_c*pixel_values_c, area_d*pixel_values_d]) return output def meshgrid(height, width): with tf.name_scope("meshgrid"): y_linspace = tf.linspace(-1., 1., height) x_linspace = tf.linspace(-1., 1., width) x_coordinates, y_coordinates = tf.meshgrid(x_linspace, y_linspace) y_coordinates = tf.expand_dims(tf.reshape(y_coordinates, [-1]),0) x_coordinates = tf.expand_dims(tf.reshape(x_coordinates, [-1]),0) indices_grid = tf.concat([x_coordinates, y_coordinates], 0) return indices_grid def apply_transformation(flows, img, num_channels): with tf.name_scope("apply_transformation"): batch_size = tf.shape(img)[0] height = tf.shape(img)[1] width = tf.shape(img)[2] # num_channels = tf.shape(img)[3] output_size = (height, width) flow_channels = tf.shape(flows)[3] flows = tf.reshape(tf.transpose(flows, [0, 3, 1, 2]), [batch_size, flow_channels, height*width]) indices_grid = meshgrid(height, width) transformed_grid = tf.add(flows, indices_grid) x_s = tf.slice(transformed_grid, [0, 0, 0], [-1, 1, -1]) y_s = tf.slice(transformed_grid, [0, 1, 0], [-1, 1, -1]) x_s_flatten = tf.reshape(x_s, [-1]) y_s_flatten = tf.reshape(y_s, [-1]) transformed_image = interpolate(img, x_s_flatten, y_s_flatten, (height, width)) transformed_image = tf.reshape(transformed_image, [batch_size, height, width, num_channels]) return transformed_image
LoYungSum/gaze_correction_loyungsum
training/code_tf/model_train/tf_utils.py
import tensorflow as tf def batch_norm(x, train_phase, name='bn_layer'): #with tf.variable_scope(name) as scope: batch_norm = tf.layers.batch_normalization( inputs=x, momentum=0.9, epsilon=1e-5, center=True, scale=True, training = train_phase, name=name ) return batch_norm def cnn_blk(inputs, filters, kernel_size, phase_train, name = 'cnn_blk'): with tf.variable_scope(name) as scope: cnn = tf.layers.conv2d(inputs=inputs, filters=filters, kernel_size=kernel_size, padding="same", activation=None, use_bias=False, name="cnn") act = tf.nn.relu(cnn, name= "act") ret = batch_norm(act, phase_train) return ret def dnn_blk(inputs, nodes, name = 'dnn_blk'): with tf.variable_scope(name) as scope: dnn = tf.layers.dense(inputs=inputs, units=nodes, activation=None, name="dnn") ret = tf.nn.relu(dnn, name= "act") return ret
LoYungSum/gaze_correction_loyungsum
training/code_tf/model_train/config.py
#-*- coding: utf-8 -*- import argparse model_config = argparse.ArgumentParser() # model model_config.add_argument('--height', type=eval, default=48, help='') model_config.add_argument('--width', type=eval, default=64, help='') model_config.add_argument('--channel', type=eval, default=3, help='') model_config.add_argument('--agl_dim', type=eval, default=2, help='') model_config.add_argument('--encoded_agl_dim', type=eval, default=16, help='') model_config.add_argument('--early_stop', type=eval, default=16, help='') # hyper parameter model_config.add_argument('--lr', type=eval, default=0.001, help='') model_config.add_argument('--epochs', type=eval, default=500, help='') model_config.add_argument('--batch_size', type=eval, default=256, help='') model_config.add_argument('--dataset', type=str, default='dirl_48x64_example', help='') # training parameter model_config.add_argument('--tar_model', type=str, default='flx', help='') # model_config.add_argument('--tar_model', type=str, default='deepwarp', help='') model_config.add_argument('--loss_combination', type=str, default='l2sc', help='') model_config.add_argument('--ef_dim', type=eval, default=12, help='') model_config.add_argument('--eye', type=str, default="L", help='') #load trained weight model_config.add_argument('--load_weights', type=bool, default=False, help='') model_config.add_argument('--easy_mode', type=bool, default=True, help='') # model_config.add_argument('--load_weights', type=bool, default=True, help='') # model_config.add_argument('--easy_mode', type=bool, default=False, help='') # folders' path model_config.add_argument('--tb_dir', type=str, default='TFboard/', help='') model_config.add_argument('--data_dir', type=str, default='../../dataset/', help='') model_config.add_argument('--train_dir', type=str, default='training_inputs/', help='') model_config.add_argument('--valid_dir', type=str, default='valid_inputs/', help='') model_config.add_argument('--weight_dir', type=str, default='pt_ckpt/', help='') def get_config(): config, unparsed = model_config.parse_known_args() print(config) return config, unparsed
LoYungSum/gaze_correction_loyungsum
gaze_correction_system/focal_length_calibration.py
<filename>gaze_correction_system/focal_length_calibration.py #!/usr/bin/env python # coding: utf-8 # # Parameter settings # In[6]: import dlib # install dlib by "pip install cmake dlib" import cv2 import numpy as np # In[7]: # Please place your head in front of the camera about 50 cm d = 60 # cm # Please set your interpupillary distance (the distance between two eyes) in the code # or you can just set it to the average distance 6.3 cm P_IPD = 6.3 # cm # default image resolution video_res = [1280,720] # define the face detector from Dlib package detector = dlib.get_frontal_face_detector() predictor = dlib.shape_predictor("./lm_feat/shape_predictor_68_face_landmarks.dat") # detect face size with smaller resolustion for detection efficiency face_detect_size = [320,240] # In[8]: def get_eye_pos(shape, pos = "L"): if(pos == "R"): lc = 36 # idx for the left corner of the right eye rc = 39 # idx for the right corner of the right eye FP_seq = [36,37,38,39,40,41] # landmarkds for right eyes elif(pos == "L"): lc = 42 # idx for the left corner of the right eye rc = 45 # idx for the right corner of the right eye FP_seq = [45,44,43,42,47,46] # landmarkds for right eyes else: print("Error: Wrong pos parameter") eye_cx = (shape.part(rc).x+shape.part(lc).x)*0.5 eye_cy = (shape.part(rc).y+shape.part(lc).y)*0.5 eye_center = [eye_cx, eye_cy] eye_len = np.absolute(shape.part(rc).x - shape.part(lc).x) bx_d5w = eye_len*3/4 bx_h = 1.5*bx_d5w # Slightly moveing up the center of the bounding box # because the upper lids are more dynamic than the lower lids sft_up = bx_h*7/12 sft_low = bx_h*5/12 E_TL = (int(eye_cx-bx_d5w),int(eye_cy-sft_up)) E_RB = (int(eye_cx+bx_d5w),int(eye_cy+sft_low)) return eye_center, E_TL, E_RB # # Starting to capture your face, push "q" to leave the program # In[9]: vs = cv2.VideoCapture(1) while True: ret, recv_frame = vs.read() if ret == True: gray = cv2.cvtColor(recv_frame, cv2.COLOR_BGR2GRAY) face_detect_gray = cv2.resize(gray, (face_detect_size[0], face_detect_size[1])) # Detect the facial landmarks detections = detector(face_detect_gray, 0) x_ratio = video_res[0]/face_detect_size[0] y_ratio = video_res[1]/face_detect_size[1] LE_ach_maps=[] RE_ach_maps=[] for k,bx in enumerate(detections): target_bx = dlib.rectangle(left=int(bx.left()*x_ratio), right =int(bx.right()*x_ratio), top =int(bx.top()*y_ratio), bottom=int(bx.bottom()*y_ratio)) shape = predictor(gray, target_bx) # get the left and right eyes LE_center, L_E_TL, L_E_RB = get_eye_pos(shape, pos="L") RE_center, R_E_TL, R_E_RB = get_eye_pos(shape, pos="R") f = int(np.sqrt((LE_center[0]-RE_center[0])**2 + (LE_center[1]-RE_center[1])**2)*d/P_IPD) cv2.rectangle(recv_frame, (video_res[0]-150,0),(video_res[0],40), (255,255,255),-1 ) cv2.putText(recv_frame, 'f:'+str(f), (video_res[0]-140,15), cv2.FONT_HERSHEY_SIMPLEX, 0.4,(0,0,255),1,cv2.LINE_AA) # draw the regions of two eyes with blue cv2.rectangle(recv_frame, (L_E_TL[0],L_E_TL[1]),(L_E_RB[0],L_E_RB[1]), (255,0,0),1) cv2.rectangle(recv_frame, (R_E_TL[0],R_E_TL[1]),(R_E_RB[0],R_E_RB[1]), (255,0,0),1) # highlight the midlle point of the eye corners with green cv2.circle(recv_frame,(int(LE_center[0]),int(LE_center[1])), 2, (0,255,0), -1) cv2.circle(recv_frame,(int(RE_center[0]),int(RE_center[1])), 2, (0,255,0), -1) # draw facial landmarks with red for i in range(68): cv2.circle(recv_frame,(shape.part(i).x,shape.part(i).y), 2, (0,0,255), -1) cv2.imshow("Calibration", recv_frame) k = cv2.waitKey(10) if k == ord('q'): vs.release() cv2.destroyAllWindows() break else: pass # # remember to set the f value to the "config.py"# # In[10]: print("The focal length of your camera is", f, ",please set the value of f (--f) in the config.py")
LoYungSum/gaze_correction_loyungsum
gaze_correction_system/config.py
#-*- coding: utf-8 -*- import argparse model_config = argparse.ArgumentParser() # model parameters model_config.add_argument('--height', type=eval, default=48, help='') model_config.add_argument('--width', type=eval, default=64, help='') model_config.add_argument('--channel', type=eval, default=3, help='') model_config.add_argument('--ef_dim', type=eval, default=12, help='') model_config.add_argument('--agl_dim', type=eval, default=2, help='') model_config.add_argument('--encoded_agl_dim', type=eval, default=16, help='') #demo model_config.add_argument('--mod', type=str, default="flx", help='') model_config.add_argument('--weight_set', type=str, default="weights", help='') model_config.add_argument('--record_time', type=bool, default=False, help='') model_config.add_argument('--tar_ip', type=str, default='localhost', help='') model_config.add_argument('--sender_port', type=int, default=5005, help='') model_config.add_argument('--recver_port', type=int, default=5005, help='') model_config.add_argument('--uid', type=str, default='local', help='') model_config.add_argument('--P_IDP', type=eval, default=6.3, help='') model_config.add_argument('--f', type=eval, default=1000, help='') model_config.add_argument('--P_c_x', type=eval, default=0, help='') model_config.add_argument('--P_c_y', type=eval, default=-9.5, help='') model_config.add_argument('--P_c_z', type=eval, default=-1, help='') model_config.add_argument('--S_W', type=eval, default=28.5, help='') model_config.add_argument('--S_H', type=eval, default=18, help='') def get_config(): config, unparsed = model_config.parse_known_args() print(config) return config, unparsed
LoYungSum/gaze_correction_loyungsum
training/generate_training_inputs_from_raw_dataset/load_dataset2.py
<reponame>LoYungSum/gaze_correction_loyungsum import threading import numpy as np import tensorflow as tf import pickle import cv2 import os sur_agl_limit = 16 dif_agl_limit_v = 21 dif_agl_limit_h = 26 def seperate_eye_and_lid(imgs, anchor_maps): # start_time = time.time() imgs_eye = [] imgs_lid = [] imgs_eye_mask = [] imgs_lid_mask = [] for user_idx in range(len(imgs)): user_eye = [] user_lid = [] user_eye_mask = [] user_lid_mask = [] for img_idx in range(anchor_maps[user_idx].shape[0]): # get eye anchors anchors = [] for i in range(6): anchor = [int(np.where(anchor_maps[user_idx][img_idx,0,:,2*i+0] ==0)[0]),int(np.where(anchor_maps[user_idx][img_idx,:,0,2*i+1] ==0)[0])] anchors.append(anchor) anchors = np.array(anchors, np.int32) # create mask mask_eye = np.zeros((imgs[0].shape[1],imgs[0].shape[2]),np.uint8) cv2.fillPoly(mask_eye, [anchors], 1) mask_lid = np.ones((imgs[0].shape[1],imgs[0].shape[2]),np.uint8) - mask_eye # crop imaage #temp = cv2.cvtColor(mask_lid,cv2.COLOR_GRAY2RGB) img_eye = cv2.bitwise_and(imgs[user_idx][img_idx,...],imgs[user_idx][img_idx,...],mask = mask_eye)# + temp img_lid = cv2.bitwise_and(imgs[user_idx][img_idx,...],imgs[user_idx][img_idx,...],mask = mask_lid) user_eye.append(img_eye) user_lid.append(img_lid) user_eye_mask.append(mask_eye) user_lid_mask.append(mask_lid) user_eye = np.array(user_eye) user_lid = np.array(user_lid) user_eye_mask = np.array(user_eye_mask) user_lid_mask = np.array(user_lid_mask) imgs_eye.append(user_eye) imgs_lid.append(user_lid) imgs_eye_mask.append(user_eye_mask) imgs_lid_mask.append(user_lid_mask) # print("sep time %.4f" % (time.time()-start_time)) return imgs_eye, imgs_lid, imgs_eye_mask, imgs_lid_mask ### define your input schedual ##### def read_training_data(file_path): f = open(file_path, 'rb') data = pickle.load(f) f.close() return data def load(data_dir, dirs, eye): MyDict = {} imgs = [] agls = [] ps = [] anchor_maps = [] for d in dirs: print(os.path.join(data_dir, d)) data = read_training_data(os.path.join(data_dir, d) + '/' + d + str('_') + eye) imgs.append(np.asarray(data['img'], dtype= np.float32)/255.0) agls.append(np.concatenate([np.expand_dims(np.asarray(data['v'], dtype= np.float32), axis=1), np.expand_dims(np.asarray(data['h'], dtype= np.float32), axis=1)], axis = 1)) ps.append(np.asarray(data['p'], dtype= np.float32)) anchor_maps.append(np.asarray(data['anchor_map'], dtype= np.float32)) # sep image to eye and lid imgs_eye, imgs_lid, msk_eye, msk_lid = seperate_eye_and_lid(imgs, anchor_maps) MyDict['imgs_ori'] = imgs MyDict['agls'] = agls MyDict['ps'] = ps MyDict['anchor_maps'] = anchor_maps MyDict['imgs_eye'] = imgs_eye MyDict['imgs_lid'] = imgs_lid MyDict['msk_eye'] = msk_eye MyDict['msk_lid'] = msk_lid return MyDict def img_pair_list(agls, pose): ''' 10 dims: pose, uid, src_img_idx,tar_img_idx, src_v,src_h, tar_v,tar_h, agl_dif_v,agl_dif_h ''' for uid in range(len(agls)): n_agl = np.arange(len(agls[uid])) sur, tar = np.meshgrid(n_agl, n_agl) uid_pair = np.concatenate((np.expand_dims(np.repeat(pose, len(agls[uid])*len(agls[uid])), axis=1), np.expand_dims(np.repeat(uid, len(agls[uid])*len(agls[uid])), axis=1), np.expand_dims(np.reshape(sur,-1), axis=1), np.expand_dims(np.reshape(tar,-1), axis=1)), axis=1) if uid == 0: pairs = uid_pair src_agls = agls[uid][uid_pair[:,2],:] tar_agls = agls[uid][uid_pair[:,3],:] dif_agls = agls[uid][uid_pair[:,3],:] - agls[uid][uid_pair[:,2],:] else: pairs = np.concatenate((pairs, uid_pair), axis=0) # image index src_agls = np.concatenate((src_agls, agls[uid][uid_pair[:,2],:]), axis=0) # sourse angle tar_agls = np.concatenate((tar_agls, agls[uid][uid_pair[:,3],:]), axis=0) dif_agls = np.concatenate((dif_agls, agls[uid][uid_pair[:,3],:] - agls[uid][uid_pair[:,2],:]), axis=0) pairs = np.concatenate((pairs,src_agls,tar_agls,dif_agls), axis=1) return pairs.astype(np.int32) def data_iterator(input_dict, pairs, batch_size, shuffle = True): # print(input_dict.keys()) t_batch = int(len(pairs)/batch_size) while True: idxs = np.arange(0, len(pairs)) if(shuffle): np.random.shuffle(idxs) for batch_idx in range(t_batch-1): cur_idxs = idxs[(batch_idx*batch_size):((batch_idx+1)*batch_size)] pairs_batch = pairs[cur_idxs] out_dict = {} b_pose =[] b_uID =[] b_img_ori = [] b_sur_agl = [] b_tar_agl = [] b_fp = [] b_img__ori = [] b_msk_eye = [] for pair_idx in range(len(pairs_batch)): pose = str(pairs_batch[pair_idx,0]) uID = pairs_batch[pair_idx,1] surID = pairs_batch[pair_idx,2] tarID = pairs_batch[pair_idx,3] b_pose.append(pose) b_uID.append(uID) b_img_ori.append(input_dict[pose]['imgs_ori'][uID][surID]) b_sur_agl.append(input_dict[pose]['agls'][uID][surID]) b_tar_agl.append(input_dict[pose]['agls'][uID][tarID]) b_fp.append(input_dict[pose]['anchor_maps'][uID][surID]) b_img__ori.append(input_dict[pose]['imgs_ori'][uID][tarID]) b_msk_eye.append(input_dict[pose]['msk_eye'][uID][surID]) out_dict['pose'] = np.asarray(b_pose) out_dict['uID'] = np.asarray(b_uID) out_dict['imgs_ori'] = np.asarray(b_img_ori) out_dict['fp'] = np.asarray(b_fp) out_dict['sur_agl'] = np.asarray(b_sur_agl) out_dict['tar_agl'] = np.asarray(b_tar_agl) out_dict['imgs__ori'] = np.asarray(b_img__ori) out_dict['msk_eye'] = np.asarray(b_msk_eye) yield out_dict def shuffle_data_batch(data_batch): idxs = np.arange(0, data_batch['imgs_ori'].shape[0]) np.random.shuffle(idxs) for i in data_batch.keys(): # print(data_batch.keys()) data_batch[i] = data_batch[i][idxs,...] return data_batch def get_dict(conf, tar_path): tar_dir = os.path.join(conf.data_dir, conf.dataset, tar_path) pose_dirs = np.asarray([d for d in os.listdir(tar_dir) if os.path.isdir(os.path.join(tar_dir, d))]) print("Pose dirs", pose_dirs) tar_dicts = {} for p in pose_dirs: print('pose', p) tar_dirs = np.asarray([d for d in os.listdir(os.path.join(tar_dir, p)) if os.path.isdir(os.path.join(tar_dir, p, d))]) print("Dirs", tar_dirs) # load training inputs tar_dict = load(data_dir=os.path.join(tar_dir, p), dirs = tar_dirs, eye = conf.eye) tar_dict['pairs'] = img_pair_list(tar_dict['agls'], int(p)) tar_dicts[p] = tar_dict return tar_dicts def get_easy_hard_iter(input_dicts, batch_size): sur_agl_limit = 16 dif_agl_limit_v = 21 dif_agl_limit_h = 26 pairs = [] for pose in list(input_dicts.keys()): if pose == list(input_dicts.keys())[0]: pairs = input_dicts[pose]['pairs'] else: pairs = np.concatenate((pairs, input_dicts[pose]['pairs']), axis = 0) # image index all_idx = np.arange(len(pairs)) easy_idx = np.where((np.abs(pairs[:,4]) < sur_agl_limit) & (np.abs(pairs[:,5]) < sur_agl_limit) & (np.abs(pairs[:,8]) < dif_agl_limit_v) & (np.abs(pairs[:,9]) < dif_agl_limit_h))[0] hard_idx = np.setdiff1d(all_idx, easy_idx) if (len(all_idx) != (len(hard_idx) + len(easy_idx))): sys.exit("[T] Easy and Hard sets separation error") print("E {}; H {}; ALL {}".format(len(easy_idx),len(hard_idx),len(all_idx))) easy_iter_ = data_iterator(input_dicts, pairs[easy_idx,:], batch_size) hard_iter_ = data_iterator(input_dicts, pairs[hard_idx,:], batch_size) return easy_iter_, hard_iter_, len(easy_idx),len(hard_idx) def merge_batches(cur_batch, tar_batch): for i in cur_batch.keys(): cur_batch[i] = np.concatenate((cur_batch[i], tar_batch[i]), axis=0) return cur_batch
sh4nth/maps
scratch/scripts.py
<reponame>sh4nth/maps import pandas as pd import urllib, json from concurrent.futures import ThreadPoolExecutor allDistricts = {} with open('List_of_Districts.txt') as f: for dist in f.read().splitlines(): allDistricts[dist.split(',')[0]] = dist.replace('_', ' ') allDistricts['Andamans'] = None allDistricts['Nicobars'] = None allDistricts['Lakshadweep'] = None allDistricts['Lahul_and_Spiti'] = 'Lahul, Himachal Pradesh, India' allDistricts['Leh_Ladakh'] = 'Leh 194101' allDistricts['East_Godavari'] = 'Kakinada, Andhra Pradesh, India' allDistricts['South_Garo_Hills'] = 'Chokpot, Meghalaya, India' allDistricts['Mayurbhanj'] = 'Baripada, Odisha, India' allDistricts['Kurung_Kumey'] = 'Aalo 791001' allDistricts['Upper_Dibang_Valley'] = 'Anini 792101' allDistricts['Upper_Siang'] = 'Yingkiong 791002' allDistricts['Upper_Subansiri'] = 'Gengi 791125' allDistricts['Karbi_Anglong'] = 'Diphu, Assam, India' def getDirections(address, mode='transit', origin='New Delhi', log=False): apiKey = 'your apiKey from https://developers.google.com/maps/documentation/directions/ (Click get a Key)' url = 'https://maps.googleapis.com/maps/api/directions/json?origin='+origin+'&key='+apiKey+'&destination=' + address + '&mode=' + mode ntries = 5 response = None for _ in range(ntries): try: response = urllib.urlopen(url) break # success except IOError as err: pass # noOp else: # all ntries failed raise err # re-raise the last timeout error data = json.loads(response.read()) if log: print data time = data['routes'] if len(time) != 0: time = time[0]['legs'][0]['duration'] return time else: return None mode = 'driving' #,'walking','transit' executor = ThreadPoolExecutor(max_workers=10) origins_done = ['Kolkata', 'Delhi', 'Greater_Bombay', 'Bangalore', 'Chennai','Srinagar', 'Kohima', 'Nagpur', 'Gandhinagar'] origins = [] for origin_id in origins: times[origin_id] = {} origin = allDistricts[origin_id] for dest_id in allDistricts: dest = allDistricts[dest_id] if dest is None or origin is None: times[origin_id][dest_id] = None else: times[origin_id][dest_id] = executor.submit(getDirections, dest, mode, origin=origin) for dest_id in allDistricts: if times[origin_id][dest_id] is not None: times[origin_id][dest_id] = times[origin_id][dest_id].result() if times[origin_id][dest_id] is not None: times[origin_id][dest_id] = times[origin_id][dest_id]['value'] print '.', else: print 'o',origin_id,'->',dest_id, '(',allDistricts[origin_id],allDistricts[dest_id],')' with open("/tmp/data.csv","w") as f: f.write('origin') for dest_id in sorted(allDistricts): f.write(',' + dest_id) f.write('\n') for origin_id in origins_done: f.write(origin_id) for dest_id in sorted(allDistricts): f.write(','+str(times[origin_id][dest_id])) f.write('\n')
Swipe650/Py-Timer
pytimer.py
<reponame>Swipe650/Py-Timer<filename>pytimer.py #!/usr/bin/env python3 # encoding: utf-8 """ Count down seconds from a given minute value using the Tkinter GUI toolkit that comes with Python. Basic Tk version by vegaseat (I extended this): https://www.daniweb.com/programming/software-development/threads/464062/countdown-clock-with-python <NAME>, alias <NAME> https://github.com/jabbalaci/Pomodoro-Timer Modified by: Swipe650 https://github.com/Swipe650 """ from tkinter import Tk, PhotoImage import os import re import shlex import socket import sys import time import tkinter from tkinter import * from collections import OrderedDict from subprocess import PIPE, STDOUT, Popen from subprocess import call import tkinter as tk # Package dependencies: """ wmctrl, xdotool, sox and tkinter / tk packages: Arh based distros: sudo pacman -S tk Ubuntu: sudo apt-get install python3-tk """ go_on = None # will be set later VERSION = '0.1' MINUTES = 0 # 0 minutes is the default WINDOW_TITLE = 'pytimer' DEBUG = True PROJECT_DIR = os.path.dirname(os.path.abspath(__file__)) SOUND_FILE = '{d}/timer_done.wav'.format(d=PROJECT_DIR) SOUND_VOLUME = '0.75' hostname = socket.gethostname() if hostname == 'laptop': SOUND_VOLUME = 0.75 required_commands = [ '/usr/bin/wmctrl', # in package wmctrl '/usr/bin/xdotool', # in package xdotool '/usr/bin/play', # in package sox ] def check_required_commands(): """ Verify if the external binaries are available. """ for cmd in required_commands: if not os.path.isfile(cmd): print("Error: the command '{0}' is not available! Abort.".format(cmd)) sys.exit(1) check_required_commands() ################################### ## window and process management ## ################################### def get_simple_cmd_output(cmd, stderr=STDOUT): """ Execute a simple external command and get its output. The command contains no pipes. Error messages are redirected to the standard output by default. """ args = shlex.split(cmd) return Popen(args, stdout=PIPE, stderr=stderr).communicate()[0].decode("utf8") def get_wmctrl_output(): """ Parses the output of wmctrl and returns a list of ordered dicts. """ cmd = "wmctrl -lGpx" lines = [line for line in get_simple_cmd_output(cmd) .encode('ascii', 'ignore') .decode('ascii').split("\n") if line] res = [] for line in lines: pieces = line.split() d = OrderedDict() d['wid'] = pieces[0] d['desktop'] = int(pieces[1]) d['pid'] = int(pieces[2]) d['geometry'] = [int(x) for x in pieces[3:7]] d['window_class'] = pieces[7] d['client_machine_name'] = pieces[8] d['window_title'] = ' '.join(pieces[9:]) res.append(d) # return res def get_wid_by_title(title_regexp): """ Having the window title (as a regexp), return its wid. If not found, return None. """ for d in get_wmctrl_output(): m = re.search(title_regexp, d['window_title']) if m: return d['wid'] # return None def activate_window_by_id(wid): """ Put the focus on and activate the the window with the given ID. """ os.system('xdotool windowactivate {wid}'.format(wid=wid)) def switch_to_window(title_regexp): """ Put the focus on the window with the specified title. """ wid = get_wid_by_title(title_regexp) if wid: if DEBUG: print('# window id:', wid) wid = int(wid, 16) if DEBUG: print('# switching to the other window') activate_window_by_id(wid) else: if DEBUG: print('# not found') ######### ## GUI ## ######### def formatter(sec): # format as 2 digit integers, fills with zero to the left # divmod() gives minutes, seconds #return "{:02d}:{:02d}".format(*divmod(sec, 60)) hours, remainder = divmod(sec, 3600) mins, sec = divmod(remainder, 60) #if MINUTES > 60: return "{:1d}:{:02d}:{:02d}".format(hours, mins, sec) #else: #return "{:02d}:{:02d}".format(mins, sec) #return "{:2d}:{:02d}:{:02d}".format(hours, mins, sec) def play_sound(): os.system("play -q -v {vol} {fname} &".format( vol=SOUND_VOLUME, fname=SOUND_FILE )) def reset(event=None): entry.delete(len(entry.get())-1) entry.delete(len(entry.get())-1) entry.delete(len(entry.get())-1) global go_on go_on = False time_str.set(formatter(MINUTES * 60)) #clear_vars() root.update() def mute(event=None): call(["kill", "-9", "play"]) def close(event=None): call(["kill", "-9", "play"]) root.destroy() # SET button code: def onset(event=None): global SET SET = int(entry.get()) MINUTES = SET global onset onset = True time_str.set(formatter(MINUTES * 60)) root.update() #def count_down(event=None): global go_on go_on = True if onset == True: MINUTES = SET for t in range(MINUTES * 60 - 1, -1, -1): if t == 0: play_sound() switch_to_window(WINDOW_TITLE) time_str.set(formatter(t)) root.update() # delay one second for _ in range(2): time.sleep(0.5) # if minimized then maximized, root.update() # it's more responsive this way if not go_on: return reset() def center(win): """ centers a tkinter window :param win: the root or Toplevel window to center from http://stackoverflow.com/a/10018670/232485 """ win.update_idletasks() width = win.winfo_width() frm_width = win.winfo_rootx() - win.winfo_x() win_width = width + 2 * frm_width height = win.winfo_height() titlebar_height = win.winfo_rooty() - win.winfo_y() win_height = height + titlebar_height + frm_width x = win.winfo_screenwidth() // 2 - win_width // 2 y = win.winfo_screenheight() // 2 - win_height // 2 win.geometry('{}x{}+{}+{}'.format(width, height, x, y)) win.deiconify() def print_usage(): print(""" Swipe650's Pytimer v{ver} Usage: {fname} [parameter] Parameters: -h, --help this help -play play the sound and quit (for testing the volume) """.strip().format(ver=VERSION, fname=sys.argv[0])) if len(sys.argv) > 1: param = sys.argv[1] if param in ['-h', '--help']: print_usage() sys.exit(0) elif param == '-play': # for testing the volume play_sound() sys.exit(0) elif re.search(r'\d+', param): try: MINUTES = int(param) except ValueError: print("Error: unknown option.") sys.exit(1) else: print("Error: unknown option.") sys.exit(1) root = tk.Tk() root.wm_title(WINDOW_TITLE) #root.wm_geometry ("-30-80") root.wm_geometry ("-2030-0") root.resizable(width=False, height=False) #root.geometry('{}x{}'.format(195, 200)) img = PhotoImage(file='/home/swipe/bin/pytimer/pytimer_icon.png') root.tk.call('wm', 'iconphoto', root._w, img) mainframe = tk.Frame(root, padx="3",pady="3") mainframe.grid(column=0, row=0, sticky=(N, W, E, S)) mainframe.columnconfigure(0, weight=1) mainframe.rowconfigure(0, weight=1) time_str = tk.StringVar() # create the time display label, give it a large font # label auto-adjusts to the font label_font = ('helvetica', 34) tk.Label(root, textvariable=time_str, font=label_font, bg='white', fg='red', relief='raised', bd=3).grid(padx=5, pady=5) time_str.set(formatter(MINUTES * 60)) root.update() # Input box entry = Entry(root, width=3) entry.grid(column=0, row=2, pady=3, sticky=(N)) entry.focus() entry.bind('<Return>', onset) # create buttons and activates them with enter key #startbtn = tk.Button(root, text='Start', command=count_down) #startbtn.grid(column=0, row=2, sticky=(E)) #startbtn.bind('<Return>', count_down) #closebtn = tk.Button(root, text='Close', command=root.destroy) closebtn = tk.Button(root, text='Close', command=close) closebtn.grid(column=0, row=2, sticky=(W)) closebtn.bind('<Return>', close) resetbtn = tk.Button(root, text='Reset', command=reset) resetbtn.grid(column=0, row=2, sticky=(N)) resetbtn.bind('<Return>', reset) mutebtn = tk.Button(root, text='Mute', command=mute) mutebtn.grid(column=0, row=2, sticky=(E)) mutebtn.bind('<Return>', mute) #setbtn = tk.Button(root, text=' Set ', command=onset) #setbtn.grid(column=0, row=2, sticky=(W)) #setbtn.bind('<Return>', onset) # start the GUI event loop #root.wm_attributes("-topmost", 1) # always on top root.wm_attributes("-topmost", 0) # never on top #center(root) root.mainloop()
all4dich/ldap_tools
src/main/test_python/call_hello.py
import re import os import logging from ldap_tools.common import LDAPClient logger = logging.getLogger("add-lge-users-to-nexus") if __name__ == "__main__": host = "192.168.3.11" username = "lge\\allessunjoo.park" password = input("Password: ") a = LDAPClient(host, username, password) a.search_root = 'ou=LGE Users,dc=lge,dc=net' try: lge_department_name = "TV DevOps개발" except: lge_department_name = input("LGE department name: ") r = a.get_members(lge_department_name) for member in r: user_cn = str(member.cn) user_mail = str(member.mail) print(user_mail)
all4dich/ldap_tools
src/main/test_python/find-departments.py
<reponame>all4dich/ldap_tools import re import os import logging from ldap_tools.common import LDAPClient import argparse arg_parser = argparse.ArgumentParser() arg_parser.add_argument("-u", dest="username", required=True) arg_parser.add_argument("-p", dest="password", required=True) args = arg_parser.parse_args() if __name__ == "__main__": host = "172.16.17.32" username = f"lge\\{args.username}" password = <PASSWORD> a = LDAPClient(host, username, password) a.search_root = 'ou=LGE Users,dc=lge,dc=net' try: lge_department_name = "TV DevOps개발" except: lge_department_name = input("LGE department name: ") departments = a.find_department(lge_department_name) for each_dept in departments: print(each_dept.name.value)
all4dich/ldap_tools
src/main/test_python/test-call-ldap-server.py
<filename>src/main/test_python/test-call-ldap-server.py from ldap_tools.common import LDAPClient import subprocess import re if __name__ == "__main__": host = "192.168.127.12" username = "lge\\addhost" password = input(f"Password for {username}") a = LDAPClient(host,username,password) a.search_root = 'ou=LGE Users,dc=lge,dc=net' # b = a.get_child(oc="posixGroup") # print(b) # members = list(map(lambda x: x.entry_dn, b)) # print("\n".join(members)) # c = a.get_objects("ou=TV SW Engineering") # print(c) # user = "jaewooki.lee" # d = a.get_objects(f"sAMAccountName=*{user}*") # for e in d: # print(e) team = "TV SW CI팀" r = a.get_members(team) for member in r: print(member.cn)
all4dich/ldap_tools
src/main/python/ldap_tools/common.py
from ldap3 import Server, Connection, ALL, NTLM import logging logger = logging.getLogger() logger.setLevel(logging.WARN) formatter = logging.Formatter('%(levelname)7s:%(filename)s:%(lineno)d:%(funcName)10s: %(message)s') ch = logging.StreamHandler() ch.setLevel(logging.WARN) ch.setFormatter(formatter) logger.addHandler(ch) class LDAPClient: def __init__(self, host, username, password, authentication=NTLM, search_root=None): self.host = host self._username = username self._password = password self.authentication = authentication self._search_root = search_root self._search_attributes = ['name', 'mail', 'mobile', 'cn', 'department', 'description', 'displayNamePrintable', 'displayName', 'objectClass'] @property def username(self): return self._username @username.setter def username(self, val): self._username = val @property def password(self): return self._password @password.setter def password(self, val): self._password = val @property def search_root(self): return self._search_root @search_root.setter def search_root(self, value): self._search_root = value @property def search_attributes(self): return self._search_attributes @search_attributes.setter def search_attributes(self, value): self._search_attributes = value def get_connection(self): """ Return a connection to LDAP server :return: conn """ server = Server(self.host) conn = Connection(server, user=self._username, password=self._password, authentication=self.authentication) conn.bind() return conn def get_objects(self, obj_name): """ Return a list of objects with a query 'obj_name' 'obj_name' is a query statement written by a caller Example. cn=<NAME> :param obj_name: :return: """ conn = self.get_connection() if self._search_root is None: logger.warning("Set 'search_root' and try again") return None conn.search(self._search_root, f"({obj_name})", attributes=self._search_attributes) return conn.entries def get_departments(self, dept_name): return self.get_objects(f"cn={dept_name}") def get_members(self, dept_name): """ Get a list of members who are under 'dept_name' :param dept_name: Department name :return: """ members = [] conn = self.get_connection() if self._search_root is None: logger.warning("Set 'search_root' and try again") return None # Get a dapartment object depts = self.get_objects(f"&(objectClass=organizationalUnit)(ou=*{dept_name}*)") for each_dept in depts: member_search_base = each_dept.entry_dn conn.search(member_search_base, '(objectClass=person)', attributes=self._search_attributes) for each_entry in conn.entries: members.append(each_entry) return members def get_child(self, oc="person"): """ Return all child elements under 'self._search_root' :param oc: :return: """ if self._search_root is None: logger.warning("Set 'search_root' and try again") return None conn = self.get_connection() conn.search(self._search_root, f"(objectClass={oc})", attributes=self._search_attributes) return conn.entries def find_department(self, dept_str): """ Find deparments that each has its deparment name with substring 'dept_str' :param dept_str: Deparment name's subsring :return: List of departments. Each department's objectClass is 'group' """ conn = self.get_connection() if self._search_root is None: logger.warning("Set 'search_root' and try again") return None # Get a dapartment object dept_str_converted = dept_str.replace("(", "*").replace(")", "*") depts = self.get_objects(f"&(objectClass=group)(cn=*{dept_str_converted}*)") return depts
gauchm/mlstream
mlstream/utils.py
import sys import pickle import json from pathlib import Path from typing import Dict, List from datetime import datetime import h5py import pandas as pd import numpy as np import scipy as sp from tqdm import tqdm from .datasets import LumpedBasin from .datautils import store_static_attributes def create_h5_files(data_root: Path, out_file: Path, basins: List, dates: List, forcing_vars: List, seq_length: int, allow_negative_target: bool): """Creates H5 training set. Parameters ---------- data_root : Path Path to the main directory of the data set out_file : Path Path of the location where the hdf5 file should be stored basins : List List containing the gauge ids dates : List List of start and end date of the discharge period to use, when combining the data. forcing_vars : List Names of forcing variables seq_length : int Length of the requested input sequences allow_negative_target : bool, optional If False, will remove samples with negative target value from the dataset. Raises ------ FileExistsError If file at this location already exists. """ if out_file.is_file(): raise FileExistsError(f"File already exists at {out_file}") with h5py.File(out_file, 'w') as out_f: input_data = out_f.create_dataset('input_data', shape=(0, seq_length, len(forcing_vars)), maxshape=(None, seq_length, len(forcing_vars)), chunks=True, dtype=np.float32, compression='gzip') target_data = out_f.create_dataset('target_data', shape=(0, 1), maxshape=(None, 1), chunks=True, dtype=np.float32, compression='gzip') q_stds = out_f.create_dataset('q_stds', shape=(0, 1), maxshape=(None, 1), dtype=np.float32, compression='gzip', chunks=True) sample_2_basin = out_f.create_dataset('sample_2_basin', shape=(0, ), maxshape=(None, ), dtype="S10", compression='gzip', chunks=True) scalers = None for basin in tqdm(basins, file=sys.stdout): dataset = LumpedBasin(data_root=data_root, basin=basin, forcing_vars=forcing_vars, is_train=True, train_basins=basins, seq_length=seq_length, dates=dates, scalers=scalers, allow_negative_target=allow_negative_target, with_attributes=False) if len(dataset) == 0: print (f"No data for basin {basin}. Skipping it.") continue # Reuse scalers across datasets to save computation time if scalers is None: scalers = dataset.input_scalers, dataset.output_scalers, dataset.static_scalers num_samples = len(dataset) total_samples = input_data.shape[0] + num_samples # store input and output samples input_data.resize((total_samples, seq_length, len(forcing_vars))) target_data.resize((total_samples, 1)) input_data[-num_samples:, :, :] = dataset.x target_data[-num_samples:, :] = dataset.y # additionally store std of discharge of this basin for each sample q_stds.resize((total_samples, 1)) q_std_array = np.array([dataset.q_std] * num_samples, dtype=np.float32).reshape(-1, 1) q_stds[-num_samples:, :] = q_std_array sample_2_basin.resize((total_samples, )) str_arr = np.array([basin.encode("ascii", "ignore")] * num_samples) sample_2_basin[-num_samples:] = str_arr out_f.flush() def store_results(user_cfg: Dict, run_cfg: Dict, results: Dict): """Stores prediction results in a pickle file. Parameters ---------- user_cfg : Dict Dictionary containing the user entered evaluation config run_cfg : Dict Dictionary containing the run config loaded from the cfg.json file results : Dict DataFrame containing the observed and predicted discharge. """ if run_cfg["no_static"]: file_name = user_cfg["run_dir"] / f"results_no_static_seed{run_cfg['seed']}.p" else: if run_cfg["concat_static"]: file_name = user_cfg["run_dir"] / f"results_concat_static_seed{run_cfg['seed']}.p" else: file_name = user_cfg["run_dir"] / f"results_seed{run_cfg['seed']}.p" with (file_name).open('wb') as fp: pickle.dump(results, fp) print(f"Successfully stored results at {file_name}") def prepare_data(cfg: Dict, basins: List) -> Dict: """Pre-processes training data. Parameters ---------- cfg : Dict Dictionary containing the run config basins : List List containing the gauge ids Returns ------- Dict Dictionary containing the updated run config. """ # create database file containing the static basin attributes cfg["db_path"] = cfg["run_dir"] / "static_attributes.db" store_static_attributes(cfg["data_root"], db_path=cfg["db_path"], attribute_names=cfg["static_attributes"]) # create .h5 files for train and validation data cfg["train_file"] = cfg["train_dir"] / 'train_data.h5' create_h5_files(data_root=cfg["data_root"], out_file=cfg["train_file"], basins=basins, dates=[cfg["start_date"], cfg["end_date"]], forcing_vars=cfg["forcing_attributes"], seq_length=cfg["seq_length"], allow_negative_target=cfg["allow_negative_target"]) return cfg def setup_run(cfg: Dict) -> Dict: """Creates the folder structure for the experiment. Parameters ---------- cfg : Dict Dictionary containing the run config Returns ------- Dict Dictionary containing the updated run config """ cfg["start_time"] = str(datetime.now()) if not cfg["run_dir"].is_dir(): cfg["train_dir"] = cfg["run_dir"] / 'data' / 'train' cfg["train_dir"].mkdir(parents=True) cfg["val_dir"] = cfg["run_dir"] / 'data' / 'val' cfg["val_dir"].mkdir(parents=True) else: raise RuntimeError('There is already a folder at {}'.format(cfg["run_dir"])) # dump a copy of cfg to run directory with (cfg["run_dir"] / 'cfg.json').open('w') as fp: temp_cfg = {} for key, val in cfg.items(): if isinstance(val, Path): temp_cfg[key] = str(val) elif isinstance(val, pd.Timestamp): temp_cfg[key] = val.strftime(format="%d%m%Y") elif isinstance(val, np.ndarray): temp_cfg[key] = val.tolist() # np.ndarrays are not serializable elif 'param_dist' in key: temp_dict = {} for k, v in val.items(): if isinstance(v, sp.stats._distn_infrastructure.rv_frozen): temp_dict[k] = f"{v.dist.name}{v.args}, *kwds={v.kwds}" else: temp_dict[k] = str(v) temp_cfg[key] = str(temp_dict) else: temp_cfg[key] = val json.dump(temp_cfg, fp, sort_keys=True, indent=4) return cfg def nse(qsim: np.ndarray, qobs: np.ndarray) -> float: """Calculates NSE, ignoring NANs in ``qobs``. .. math:: \\text{NSE} = 1 - \\frac{\\sum_{t=1}^T{(q_s^t - q_o^t)^2}}{\\sum_{t=1}^T{(q_o^t - \\bar{q}_o)^2}} Parameters ---------- qsim : np.ndarray Predicted streamflow qobs : np.ndarray Ground truth streamflow Returns ------- nse : float The prediction's NSE Raises ------ ValueError If lenghts of qsim and qobs are not equal. """ if len(qsim) != len(qobs): raise ValueError(f"Lenghts of qsim {len(qsim)} and qobs {len(qobs)} mismatch.") qsim = qsim[~np.isnan(qobs)] qobs = qobs[~np.isnan(qobs)] return 1 - (np.sum(np.square(qsim - qobs)) / np.sum(np.square(qobs - np.mean(qobs))))
gauchm/mlstream
mlstream/__init__.py
# # Machine Learning for Streamflow Prediction #
gauchm/mlstream
mlstream/models/xgboost.py
<filename>mlstream/models/xgboost.py from typing import Dict, List from pathlib import Path import pickle import numpy as np import pandas as pd from sklearn.model_selection import RandomizedSearchCV from torch.utils.data import DataLoader try: import xgboost as xgb except ImportError: print("Importing the optional dependency xgboost failed, but required to train an XGBoost model.") raise from ..datasets import LumpedBasin, LumpedH5 from .base_models import LumpedModel from .nseloss import XGBNSEObjective class LumpedXGBoost(LumpedModel): """Wrapper for XGBoost model on lumped data. """ def __init__(self, no_static: bool = False, concat_static: bool = True, use_mse: bool = False, run_dir: Path = None, n_jobs: int = 1, seed: int = 0, n_estimators: int = 100, learning_rate: float = 0.01, early_stopping_rounds: int = None, n_cv: int = 5, param_dist: Dict = None, param_search_n_estimators: int = None, param_search_n_iter: int = None, param_search_early_stopping_rounds: int = None, reg_search_param_dist: Dict = None, reg_search_n_iter: int = None, model_path: Path = None): if not no_static and not concat_static: raise ValueError("XGBoost has to use concat_static.") self.model = None self.use_mse = use_mse self.n_estimators = n_estimators self.learning_rate = learning_rate self.early_stopping_rounds = early_stopping_rounds self.n_cv = n_cv self.param_dist = param_dist self.param_search_n_estimators = param_search_n_estimators self.param_search_n_iter = param_search_n_iter self.param_search_early_stopping_rounds = param_search_early_stopping_rounds self.reg_search_param_dist = reg_search_param_dist self.reg_search_n_iter = reg_search_n_iter self.run_dir = run_dir self.n_jobs = n_jobs self.seed = seed if model_path is not None: self.load(model_path) def load(self, model_path: Path): self.model = pickle.load(open(model_path, 'rb')) def train(self, ds: LumpedH5) -> None: # Create train/val sets loader = DataLoader(ds, batch_size=len(ds), num_workers=self.n_jobs) data = next(iter(loader)) # don't use static variables if len(data) == 3: x, y, q_stds = data # this shouldn't happen since we raise an exception if concat_static is False. else: raise ValueError("XGBoost has to use concat_static.") x = x.reshape(len(x), -1).numpy() y = y.reshape(-1).numpy() q_stds = q_stds.reshape(-1).numpy() # define loss function if not self.use_mse: # slight hack to enable NSE on XGBoost: replace the target with a unique id # so we can figure out the corresponding q_std during the loss calculation. y_actual = y.copy() y = np.arange(len(y)) loss = XGBNSEObjective(y, y_actual, q_stds) self.objective = loss.nse_objective_xgb_sklearn_api self.eval_metric = loss.nse_metric_xgb self.scoring = loss.neg_nse_metric_sklearn else: self.objective = 'reg:squarederror' self.eval_metric = 'rmse' self.scoring = 'neg_mean_squared_error' num_val_samples = int(len(x) * 0.2) val_indices = np.random.choice(range(len(x)), size=num_val_samples, replace=False) train_indices = np.setdiff1d(range(len(x)), val_indices) val = [(x[train_indices], y[train_indices]), (x[val_indices], y[val_indices])] if self.model is None: print("Performing parameter searches.") if self.param_dist is None or self.n_cv is None \ or self.param_search_n_iter is None \ or self.param_search_n_estimators is None \ or self.param_search_early_stopping_rounds is None \ or self.reg_search_param_dist is None \ or self.reg_search_n_iter is None: raise ValueError("Need to pass parameter search configuration or load model.") best_params = self._param_search(x[train_indices], y[train_indices], val, self.param_search_n_estimators, self.param_dist, self.param_search_n_iter).best_params_ print(f"Best parameters: {best_params}.") # Find regularization parameters in separate search for k, v in best_params.items(): self.reg_search_param_dist[k] = [v] model = self._param_search(x[train_indices], y[train_indices], val, self.param_search_n_estimators, self.reg_search_param_dist, self.reg_search_n_iter) print(f"Best regularization parameters: {model.best_params_}.") cv_results = pd.DataFrame(model.cv_results_).sort_values(by='mean_test_score', ascending=False) print(cv_results.filter(regex='param_|mean_test_score|mean_train_score', axis=1).head()) print(cv_results.loc[model.best_index_, ['mean_train_score', 'mean_test_score']]) xgb_params = model.best_params_ else: print('Using model parameters from provided XGBoost model.') xgb_params = self.model.get_xgb_params() self.model = xgb.XGBRegressor() self.model.set_params(**xgb_params) self.model.n_estimators = self.n_estimators self.model.learning_rate = self.learning_rate self.model.objective = self.objective self.model.random_state = self.seed self.model.n_jobs = self.n_jobs print(self.model.get_xgb_params()) print("Fitting model.") self.model.fit(x[train_indices], y[train_indices], eval_set=val, verbose=True, eval_metric=self.eval_metric, early_stopping_rounds=self.early_stopping_rounds) model_path = self.run_dir / 'model.pkl' with open(model_path, 'wb') as f: pickle.dump(self.model, f) print(f"Model saved as {model_path}.") def predict(self, ds: LumpedBasin) -> np.ndarray: loader = DataLoader(ds, batch_size=len(ds), shuffle=False, num_workers=4) data = next(iter(loader)) if len(data) == 2: x, y = data # this shouldn't happen since we didn't allow concat_static = False in training. else: raise ValueError("XGBoost has to use concat_static or no_static.") x = x.reshape(len(x), -1) y = y.reshape(-1) return self.model.predict(x), y def _param_search(self, x_train: np.ndarray, y_train: np.ndarray, eval_set: List, n_estimators: int, param_dist: Dict, n_iter: int) -> Dict: """Performs a cross-validated random parameter search. Parameters ---------- x_train : np.ndarray Training input y_train : np.ndarray Training ground truth eval_set : List List of evaluation sets to report metrics on n_estimators : int Number of trees to train param_dist : Dict Search space of parameter distributions n_iter : int Number of random parameter samples to test Returns ------- RandomizedSearchCV Fitted random search instance (with ``refit=False``) """ model = xgb.XGBRegressor(n_estimators=n_estimators, objective=self.objective, n_jobs=1, random_state=self.seed) model = RandomizedSearchCV(model, param_dist, n_iter=n_iter, cv=self.n_cv, return_train_score=True, scoring=self.scoring, n_jobs=self.n_jobs, random_state=self.seed, refit=False, verbose=5, error_score='raise') model.fit(x_train, y_train, eval_set=eval_set, eval_metric=self.eval_metric, early_stopping_rounds=self.param_search_early_stopping_rounds, verbose=False) return model
gauchm/mlstream
mlstream/datasets.py
from pathlib import Path from typing import List, Tuple, Dict import h5py import torch import pandas as pd import numpy as np from tqdm import tqdm from torch.utils.data import Dataset from .datautils import (load_discharge, load_forcings_lumped, load_static_attributes, reshape_data) from .scaling import InputScaler, OutputScaler, StaticAttributeScaler class LumpedBasin(Dataset): """PyTorch data set to work with the raw text files for lumped (daily basin-aggregated) forcings and streamflow. Parameters ---------- data_root : Path Path to the main directory of the data set basin : str Gauge-id of the basin forcing_vars : List Names of forcing variables to use dates : List Start and end date of the period. is_train : bool If True, discharge observations are normalized and invalid discharge samples are removed train_basins : List List of basins used in the training of the experiment this Dataset is part of. Needed to create the correct feature scalers (the ones that are calculated on these basins) seq_length : int Length of the input sequence with_attributes : bool, optional If True, loads and returns addtionaly attributes, by default False concat_static : bool, optional If true, adds catchment characteristics at each time step to the meteorological forcing input data, by default True db_path : str, optional Path to sqlite3 database file containing the catchment characteristics, by default None allow_negative_target : bool, optional If False, will remove samples with negative target value from the dataset. scalers : Tuple[InputScaler, OutputScaler, Dict[str, StaticAttributeScaler]], optional Scalers to normalize and resale input, output, and static variables. If not provided, the scalers will be initialized at runtime, which will result in poor performance if many datasets are created. Instead, it makes sense to re-use the scalers across datasets. """ def __init__(self, data_root: Path, basin: str, forcing_vars: List, dates: List, is_train: bool, train_basins: List, seq_length: int, with_attributes: bool = False, concat_static: bool = True, db_path: str = None, allow_negative_target: bool = False, scalers: Tuple[InputScaler, OutputScaler, Dict[str, StaticAttributeScaler]] = None): self.data_root = data_root self.basin = basin self.forcing_vars = forcing_vars self.seq_length = seq_length self.is_train = is_train self.train_basins = train_basins self.dates = dates self.with_attributes = with_attributes self.concat_static = concat_static self.db_path = db_path self.allow_negative_target = allow_negative_target if scalers is not None: self.input_scalers, self.output_scalers, self.static_scalers = scalers else: self.input_scalers, self.output_scalers, self.static_scalers = None, None, {} if self.input_scalers is None: self.input_scalers = InputScaler(self.data_root, self.train_basins, self.dates[0], self.dates[1], self.forcing_vars) if self.output_scalers is None: self.output_scalers = OutputScaler(self.data_root, self.train_basins, self.dates[0], self.dates[1]) # placeholder to store std of discharge, used for rescaling losses during training self.q_std = None # placeholder to store start and end date of entire period (incl warmup) self.period_start = None self.period_end = None self.attribute_names = None self.x, self.y = self._load_data() if self.with_attributes: self.attributes = self._load_attributes() self.num_samples = self.x.shape[0] def __len__(self): return self.num_samples def __getitem__(self, idx: int): if self.with_attributes: if self.concat_static: x = torch.cat([self.x[idx], self.attributes.repeat((self.seq_length, 1))], dim=-1) return x, self.y[idx] else: return self.x[idx], self.attributes, self.y[idx] else: return self.x[idx], self.y[idx] def _load_data(self) -> Tuple[torch.Tensor, torch.Tensor]: """Loads input and output data from text files. """ # we use (seq_len) time steps before start for warmup df = load_forcings_lumped(self.data_root, [self.basin])[self.basin] qobs = load_discharge(self.data_root, basins=[self.basin]).set_index('date')['qobs'] if not self.is_train and len(qobs) == 0: tqdm.write(f"Treating {self.basin} as validation basin (no streamflow data found).") qobs = pd.Series(np.nan, index=df.index, name='qobs') df = df.loc[self.dates[0]:self.dates[1]] qobs = qobs.loc[self.dates[0]:self.dates[1]] if len(qobs) != len(df): print(f"Length of forcings {len(df)} and observations {len(qobs)} \ doesn't match for basin {self.basin}") df['qobs'] = qobs # store first and last date of the selected period self.period_start = df.index[0] self.period_end = df.index[-1] # use all meteorological variables as inputs x = np.array([df[var].values for var in self.forcing_vars]).T y = np.array([df['qobs'].values]).T # normalize data, reshape for LSTM training and remove invalid samples x = self.input_scalers.normalize(x) x, y = reshape_data(x, y, self.seq_length) if self.is_train: # Delete all samples where discharge is NaN if np.sum(np.isnan(y)) > 0: tqdm.write(f"Deleted {np.sum(np.isnan(y))} NaNs in basin {self.basin}.") x = np.delete(x, np.argwhere(np.isnan(y)), axis=0) y = np.delete(y, np.argwhere(np.isnan(y)), axis=0) # Deletes all records with invalid discharge if not self.allow_negative_target and np.any(y < 0): tqdm.write(f"Deleted {np.sum(y < 0)} negative values in basin {self.basin}.") x = np.delete(x, np.argwhere(y < 0)[:, 0], axis=0) y = np.delete(y, np.argwhere(y < 0)[:, 0], axis=0) # store std of discharge before normalization self.q_std = np.std(y) y = self.output_scalers.normalize(y) # convert arrays to torch tensors x = torch.from_numpy(x.astype(np.float32)) y = torch.from_numpy(y.astype(np.float32)) return x, y def _load_attributes(self) -> torch.Tensor: df = load_static_attributes(self.db_path, [self.basin], drop_lat_lon=True) # normalize data for feature in [f for f in df.columns if f[:7] != 'onehot_']: if feature not in self.static_scalers or self.static_scalers[feature] is None: self.static_scalers[feature] = \ StaticAttributeScaler(self.db_path, self.train_basins, feature) df[feature] = self.static_scalers[feature].normalize(df[feature]) # store attribute names self.attribute_names = df.columns # store feature as PyTorch Tensor attributes = df.loc[df.index == self.basin].values return torch.from_numpy(attributes.astype(np.float32)) class LumpedH5(Dataset): """PyTorch data set to work with pre-packed hdf5 data base files. Should be used only in combination with the files processed from `create_h5_files` in the `utils` module. Parameters ---------- h5_file : Path Path to hdf5 file, containing the bundled data basins : List List containing the basin ids db_path : str Path to sqlite3 database file, containing the catchment characteristics concat_static : bool If true, adds catchment characteristics at each time step to the meteorological forcing input data, by default True cache : bool, optional If True, loads the entire data into memory, by default False no_static : bool, optional If True, no catchment attributes are added to the inputs, by default False """ def __init__(self, h5_file: Path, basins: List, db_path: str, concat_static: bool = True, cache: bool = False, no_static: bool = False): self.h5_file = h5_file self.basins = basins self.db_path = db_path self.concat_static = concat_static self.cache = cache self.no_static = no_static # Placeholder for catchment attributes stats self.df = None self.attribute_names = None # preload data if cached is true if self.cache: (self.x, self.y, self.sample_2_basin, self.q_stds) = self._preload_data() # load attributes into data frame self._load_attributes() # determine number of samples once if self.cache: self.num_samples = self.y.shape[0] else: with h5py.File(h5_file, 'r') as f: self.num_samples = f["target_data"].shape[0] def __len__(self): return self.num_samples def __getitem__(self, idx: int): if self.cache: x = self.x[idx] y = self.y[idx] basin = self.sample_2_basin[idx] q_std = self.q_stds[idx] else: with h5py.File(self.h5_file, 'r') as f: x = f["input_data"][idx] y = f["target_data"][idx] basin = f["sample_2_basin"][idx] basin = basin.decode("ascii") q_std = f["q_stds"][idx] if not self.no_static: # get attributes from data frame and create 2d array with copies attributes = self.df.loc[self.df.index == basin].values if self.concat_static: attributes = np.repeat(attributes, repeats=x.shape[0], axis=0) # combine meteorological obs with static attributes x = np.concatenate([x, attributes], axis=1).astype(np.float32) else: attributes = torch.from_numpy(attributes.astype(np.float32)) # convert to torch tensors x = torch.from_numpy(x.astype(np.float32)) y = torch.from_numpy(y.astype(np.float32)) q_std = torch.from_numpy(q_std) if self.no_static: return x, y, q_std else: if self.concat_static: return x, y, q_std else: return x, attributes, y, q_std def _preload_data(self): print("Preloading training data.") with h5py.File(self.h5_file, 'r') as f: x = f["input_data"][:] y = f["target_data"][:] str_arr = f["sample_2_basin"][:] str_arr = [x.decode("ascii") for x in str_arr] q_stds = f["q_stds"][:] return x, y, str_arr, q_stds def _get_basins(self): if self.cache: basins = list(set(self.sample_2_basin)) else: with h5py.File(self.h5_file, 'r') as f: str_arr = f["sample_2_basin"][:] str_arr = [x.decode("ascii") for x in str_arr] basins = list(set(str_arr)) return basins def _load_attributes(self): df = load_static_attributes(self.db_path, self.basins, drop_lat_lon=True) # normalize data self.attribute_scalers = {} for feature in [f for f in df.columns if f[:7] != 'onehot_']: self.attribute_scalers[feature] = \ StaticAttributeScaler(self.db_path, self.basins, feature) df[feature] = self.attribute_scalers[feature].normalize(df[feature]) # store attribute names self.attribute_names = df.columns self.df = df
gauchm/mlstream
mlstream/experiment.py
<filename>mlstream/experiment.py import json import random from pathlib import Path from typing import Dict, List, Tuple from tqdm import tqdm import pandas as pd import numpy as np import torch from .datasets import LumpedBasin, LumpedH5 from .scaling import InputScaler, OutputScaler, StaticAttributeScaler from .utils import setup_run, prepare_data, nse from .datautils import load_static_attributes class Experiment: """Main entrypoint for training and prediction. """ def __init__(self, data_root: Path, is_train: bool, run_dir: Path, start_date: str = None, end_date: str = None, basins: List = None, forcing_attributes: List = None, static_attributes: List = None, seq_length: int = 10, concat_static: bool = False, no_static: bool = False, cache_data: bool = False, n_jobs: int = 1, seed: int = 0, allow_negative_target: bool = False, run_metadata: Dict = {}): """Initializes the experiment. Parameters ---------- data_root : Path Path to the base directory of the data set. is_train : bool If True, will setup folder structure for training. run_dir: Path Path to store experiment results in. start_date : str, optional Start date (training start date if ``is_train``, else validation start date, ddmmyyyy) end_date : str, optional End date (training end date if ``is_train``, else validation end date, ddmmyyyy) basins : List, optional List of basins to use during training, or basins to predict during prediction. forcing_attributes : List, optional Names of forcing attributes to use. static_attributes : List, optional Names of static basin attributes to use. seq_length : int, optional Length of historical forcings to feed the model. Default 10 cache_data : bool, optional If True, will preload all data in memory for training. Default False concat_static : bool, optional If True, will concatenate static basin attributes with forcings for model input. no_static : bool, optional If True, will not use static basin attributes as model input. n_jobs : int, optional Number of workers to use for training. Default 1 seed : int, optional Seed to use for training. Default 0 allow_negative_target : bool, optional If False, will ignore training samples with negative target value from the dataset, and will clip predictions to values >= 0. This value should be False when predicting discharge, but True when predicting values that can be negative. run_metadata : dict, optional Optional dictionary of values to store in cfg.json for documentation purpose. """ self.model = None self.results = {} self.cfg = { "data_root": data_root, "run_dir": run_dir, "start_date": pd.to_datetime(start_date, format='%d%m%Y'), "end_date": pd.to_datetime(end_date, format='%d%m%Y'), "basins": basins, "forcing_attributes": forcing_attributes, "static_attributes": static_attributes, "seq_length": seq_length, "cache_data": cache_data, "concat_static": concat_static, "no_static": no_static, "seed": seed, "n_jobs": n_jobs, "allow_negative_target": allow_negative_target } self.cfg.update(run_metadata) if is_train: # create folder structure for this run self.cfg = setup_run(self.cfg) # prepare data for training self.cfg = prepare_data(cfg=self.cfg, basins=basins) def set_model(self, model) -> None: """Set the model to use in the experiment. """ self.model = model def train(self) -> None: """Train model. """ if self.model is None: raise AttributeError("Model is not set.") # fix random seeds random.seed(self.cfg["seed"]) np.random.seed(self.cfg["seed"]) torch.cuda.manual_seed(self.cfg["seed"]) torch.manual_seed(self.cfg["seed"]) # prepare PyTorch DataLoader ds = LumpedH5(h5_file=self.cfg["train_file"], basins=self.cfg["basins"], db_path=self.cfg["db_path"], concat_static=self.cfg["concat_static"], cache=self.cfg["cache_data"], no_static=self.cfg["no_static"]) self.model.train(ds) def predict(self) -> Dict: """Generates predictions with a trained model. Returns ------- results : Dict Dictionary containing the DataFrame of predictions and observations for each basin. """ with open(self.cfg["run_dir"] / 'cfg.json', 'r') as fp: run_cfg = json.load(fp) if self.model is None: raise AttributeError("Model is not set.") # self.cfg["start_date"] contains validation start date, # run_cfg["start_date"] the training start date run_cfg["start_date"] = pd.to_datetime(run_cfg["start_date"], format='%d%m%Y') run_cfg["end_date"] = pd.to_datetime(run_cfg["end_date"], format='%d%m%Y') self.cfg["start_date"] = pd.to_datetime(self.cfg["start_date"], format='%d%m%Y') self.cfg["end_date"] = pd.to_datetime(self.cfg["end_date"], format='%d%m%Y') # create scalers input_scalers = InputScaler(self.cfg["data_root"], run_cfg["basins"], run_cfg["start_date"], run_cfg["end_date"], run_cfg["forcing_attributes"]) output_scalers = OutputScaler(self.cfg["data_root"], run_cfg["basins"], run_cfg["start_date"], run_cfg["end_date"]) static_scalers = {} db_path = self.cfg["run_dir"] / "static_attributes.db" df = load_static_attributes(db_path, run_cfg["basins"], drop_lat_lon=True) for feature in [f for f in df.columns if 'onehot' not in f]: static_scalers[feature] = StaticAttributeScaler(db_path, run_cfg["basins"], feature) # self.cfg["basins"] contains the test basins, run_cfg["basins"] the train basins. for basin in tqdm(self.cfg["basins"]): ds_test = LumpedBasin(data_root=Path(self.cfg["data_root"]), basin=basin, forcing_vars=run_cfg["forcing_attributes"], dates=[self.cfg["start_date"], self.cfg["end_date"]], is_train=False, train_basins=run_cfg["basins"], seq_length=run_cfg["seq_length"], with_attributes=not run_cfg["no_static"], concat_static=run_cfg["concat_static"], db_path=db_path, allow_negative_target=run_cfg["allow_negative_target"], scalers=(input_scalers, output_scalers, static_scalers)) preds, obs = self.predict_basin(ds_test, allow_negative_target=run_cfg["allow_negative_target"]) date_range = pd.date_range(start=self.cfg["start_date"] + pd.DateOffset(days=run_cfg["seq_length"] - 1), end=self.cfg["end_date"]) df = pd.DataFrame(data={'qobs': obs.flatten(), 'qsim': preds.flatten()}, index=date_range) self.results[basin] = df return self.results def predict_basin(self, ds: LumpedBasin, allow_negative_target: bool = False) -> Tuple[np.ndarray, np.ndarray]: """Predicts a single basin. Parameters ---------- ds : LumpedBasin Dataset for the basin to predict allow_negative_target : bool, default False If False, will clip predictions to values >= 0 Returns ------- preds : np.ndarray Array containing the (rescaled) network prediction for the entire data period obs : np.ndarray Array containing the observed discharge for the entire data period """ preds, obs = self.model.predict(ds) preds = ds.output_scalers.rescale(preds) if not allow_negative_target: preds[preds < 0] = 0 return preds, obs def get_nses(self) -> Dict: """Calculates the experiment's NSE for each calibration basin. Validation basins are ignored since they don't provide ground truth. Returns ------- nses : Dict Dictionary mapping basin ids to their NSE Raises ------ AttributeError If called before predicting """ if len(self.results) == 0: raise AttributeError("No results to evaluate.") nses = {} for basin, df in self.results.items(): # ignore validation basins that have no ground truth if not all(pd.isna(df['qobs'])): nses[basin] = nse(df['qsim'].values, df['qobs'].values) return nses
gauchm/mlstream
mlstream/datautils.py
import re from typing import Dict, List, Tuple from pathlib import Path import sqlite3 import pandas as pd import numpy as np import netCDF4 as nc from numba import njit def get_basin_list(data_root: Path, basin_type: str) -> List: """Returns the list of basin names. If basin_type is 'C' or 'V', the gauge_info.csv needs to contain a column 'Cal_Val' that indicates the basin type. Parameters ---------- data_root : Path Path to base data directory, which contains a folder 'gauge_info' with the ``gauge_info.csv`` file basin_type : str 'C' to return calibration stations only, 'V' to return validation stations only, '*' to return all stations Returns ------- list List of basin name strings """ if basin_type not in ['*', 'C', 'V']: raise ValueError('Illegal basin type') gauge_info_file = data_root / 'gauge_info' / 'gauge_info.csv' gauge_info = pd.read_csv(gauge_info_file) if basin_type != '*': if 'Cal_Val' not in gauge_info.columns: raise RuntimeError('gauge_info.csv needs column "Cal_Val" to filter for basin types.') gauge_info = gauge_info[gauge_info['Cal_Val'] == basin_type] if 'Gauge_ID' not in gauge_info.columns: raise RuntimeError('gauge_info.csv has no column "Gauge_ID".') basins = gauge_info['Gauge_ID'].str.zfill(8).values return np.unique(basins).tolist() def load_discharge(data_root: Path, basins: List = None) -> pd.DataFrame: """Loads observed discharge for (calibration) gauging stations. Parameters ---------- data_root : Path Path to base data directory, which contains a directory 'discharge' with one or more nc-files. basins : List, optional List of basins for which to return data. If None (default), all basins are returned Returns ------- pd.DataFrame A DataFrame with columns [date, basin, qobs], where 'qobs' contains the streamflow. """ discharge_dir = data_root / 'discharge' files = [f for f in discharge_dir.glob('*') if f.is_file()] data_streamflow = pd.DataFrame(columns=['date', 'basin', 'qobs']) found_basins = [] for f in files: try: file_format = f.name.split(".")[-1] if file_format == "nc": q_nc = nc.Dataset(f, 'r') file_basins = np.array([f.zfill(8) for f in q_nc['station_id'][:]]) if basins is not None: # some basins might be in multiple NC-files. We only load them once. target_basins = [i for i, b in enumerate(file_basins) if b in basins and b not in found_basins] else: target_basins = [i for i, b in enumerate(file_basins) if b not in found_basins] if len(target_basins) > 0: time = nc.num2date(q_nc['time'][:], q_nc['time'].units, q_nc['time'].calendar) data = pd.DataFrame(q_nc['Q'][target_basins, :].T, index=time, columns=file_basins[target_basins]) data = data.unstack().reset_index().rename({'level_0': 'basin', 'level_1': 'date', 0: 'qobs'}, axis=1) found_basins += data['basin'].unique().tolist() data_streamflow = data_streamflow.append(data, ignore_index=True, sort=True) q_nc.close() else: raise NotImplementedError(f"Discharge format {file_format} not supported.") except Exception as e: print (f"Couldn't load discharge from {f.name}. Skipping ...") return data_streamflow def load_forcings_lumped(data_root: Path, basins: List = None) -> Dict: """Loads basin-lumped forcings. Parameters ---------- data_root : Path Path to base data directory, which contains the directory 'forcings/lumped/', which contains one .rvt/.csv/.txt -file per basin. basins : List, optional List of basins for which to return data. Default (None) returns data for all basins. Returns ------- dict Dictionary of forcings (pd.DataFrame) per basin """ lumped_dir = data_root / 'forcings' / 'lumped' basin_files = [f for f in lumped_dir.glob('*') if f.is_file()] basin_forcings = {} for f in basin_files: try: basin = f.name.split('_')[-1][:-4].zfill(8) file_format = f.name.split(".")[-1] if basins is not None and basin not in basins: continue if file_format == 'rvt': with open(f) as fp: next(fp) start_date = next(fp)[:10] columns = re.split(r',\s+', next(fp).replace('\n', ''))[1:] data = pd.read_csv(f, sep=r',\s*', skiprows=4, skipfooter=1, names=columns, dtype=float, header=None, usecols=range(len(columns)), engine='python') elif file_format in ['txt', 'csv']: with open(f) as fp: columns = re.split(',', next(fp).replace('\n', ''))[1:] start_date = next(fp)[:10] data = pd.read_csv(f, dtype=float, usecols=range(1, len(columns)+1)) else: raise NotImplementedError(f"Forcing format {file_format} not supported.") data.index = pd.date_range(start_date, periods=len(data), freq='D') basin_forcings[basin] = data except Exception as e: print (f"Couldn't load lumped forcings from {f.name}. Skipping ...") return basin_forcings def store_static_attributes(data_root: Path, db_path: Path = None, attribute_names: List = None): """Loads catchment characteristics from text file and stores them in a sqlite3 table Parameters ---------- data_root : Path Path to the main directory of the data set db_path : Path, optional Path to where the database file should be saved. If None, stores the database in data_root/static_attributes.db. Default: None attribute_names : List, optional List of attribute names to use. Default: use all attributes. Raises ------ RuntimeError If attributes folder can not be found. """ f = data_root / 'gauge_info' / 'gauge_info.csv' gauge_info = pd.read_csv(f).rename({'Gauge_ID': 'basin'}, axis=1) gauge_info["basin"] = gauge_info["basin"].str.zfill(8) gauge_info.set_index('basin', inplace=True) if attribute_names is not None: static_attributes = gauge_info.loc[:, attribute_names] else: static_attributes = gauge_info if db_path is None: db_path = data_root / 'static_attributes.db' with sqlite3.connect(str(db_path)) as conn: # insert into databse static_attributes.to_sql('basin_attributes', conn) print(f"Successfully stored basin attributes in {db_path}.") def load_static_attributes(db_path: Path, basins: List, drop_lat_lon: bool = True, keep_features: List = None) -> pd.DataFrame: """Loads attributes from database file into DataFrame and one-hot-encodes non-numerical features. Parameters ---------- db_path : Path Path to sqlite3 database file basins : List List containing the basin id drop_lat_lon : bool If True, drops latitude and longitude column from final data frame, by default True keep_features : List If a list is passed, a pd.DataFrame containing these features will be returned. By default, returns a pd.DataFrame containing the features used for training. Returns ------- pd.DataFrame Attributes in a pandas DataFrame. Index is basin id. """ with sqlite3.connect(str(db_path)) as conn: df = pd.read_sql("SELECT * FROM 'basin_attributes'", conn, index_col='basin') # drop lat/lon col if drop_lat_lon: df = df.drop(['Lat_outlet', 'Lon_outlet'], axis=1, errors='ignore') # drop invalid attributes if keep_features is not None: drop_names = [c for c in df.columns if c not in keep_features] df = df.drop(drop_names, axis=1) # one-hot-encoding non_numeric_features = [f for f in df.columns if df[f].dtype == object] df = pd.get_dummies(df, columns=non_numeric_features, prefix=[f"onehot_{f}" for f in non_numeric_features]) # drop rows of basins not contained in data set df = df.loc[basins] return df @njit def reshape_data(x: np.ndarray, y: np.ndarray, seq_length: int) -> Tuple[np.ndarray, np.ndarray]: """Reshape data into LSTM many-to-one input samples Parameters ---------- x : np.ndarray Input features of shape [num_samples, num_features] y : np.ndarray Output feature of shape [num_samples, 1] seq_length : int Length of the requested input sequences. Returns ------- x_new : np.ndarray Reshaped input features of shape [num_samples*, seq_length, num_features], where num_samples* is equal to num_samples - seq_length + 1, due to the need of a warm start at the beginning y_new : np.ndarray The target value for each sample in x_new """ num_samples, num_features = x.shape x_new = np.zeros((num_samples - seq_length + 1, seq_length, num_features)) y_new = np.zeros((num_samples - seq_length + 1, 1)) for i in range(0, x_new.shape[0]): x_new[i, :, :num_features] = x[i:i + seq_length, :] y_new[i, :] = y[i + seq_length - 1, 0] return x_new, y_new
gauchm/mlstream
mlstream/models/sklearn_models.py
<filename>mlstream/models/sklearn_models.py<gh_stars>1-10 from pathlib import Path import pickle import numpy as np from sklearn.base import BaseEstimator from torch.utils.data import DataLoader from ..datasets import LumpedBasin, LumpedH5 from .base_models import LumpedModel class LumpedSklearnRegression(LumpedModel): """Wrapper for scikit-learn regression models on lumped data. """ def __init__(self, model: BaseEstimator, no_static: bool = False, concat_static: bool = True, run_dir: Path = None, n_jobs: int = 1): if not no_static and not concat_static: raise ValueError("Sklearn regression has to use concat_static.") self.model = model self.run_dir = run_dir self.n_jobs = n_jobs def load(self, model_file: Path) -> None: self.model = pickle.load(open(model_file, 'rb')) def train(self, ds: LumpedH5) -> None: # Create train/val sets loader = DataLoader(ds, batch_size=len(ds), num_workers=self.n_jobs) data = next(iter(loader)) # don't use static variables if len(data) == 3: x, y, _ = data # ignore q_stds # this shouldn't happen since we raise an exception if concat_static is False. else: raise ValueError("Sklearn regression has to use concat_static.") x = x.reshape(len(x), -1) y = y.reshape(len(y)) print("Fitting model.") self.model.fit(x, y) model_path = self.run_dir / 'model.pkl' with open(model_path, 'wb') as f: pickle.dump(self.model, f) print(f"Model saved as {model_path}.") def predict(self, ds: LumpedBasin) -> np.ndarray: loader = DataLoader(ds, batch_size=len(ds), shuffle=False, num_workers=4) data = next(iter(loader)) if len(data) == 2: x, y = data # this shouldn't happen since we didn't allow concat_static = False in training. else: raise ValueError("sklearn regression has to use concat_static or no_static.") x = x.reshape(len(x), -1) y = y.reshape(len(y)) return self.model.predict(x), y
gauchm/mlstream
setup.py
from setuptools import setup import os def readme(): with open(os.path.dirname(os.path.realpath(__file__)) + '/README.md') as f: return f.read() requires = [ 'pandas', 'numpy', 'h5py', 'tqdm', 'netCDF4', 'numba', 'scipy', ] # to save build resources, we mock torch and xgboost while building the docs if not os.getenv('READTHEDOCS'): requires.append('torch') setup(name='mlstream', version='0.1.2', description='Machine learning for streamflow prediction', long_description=readme(), long_description_content_type='text/markdown', keywords='ml hydrology streamflow machine learning', url='http://github.com/gauchm/mlstream', author='<NAME>', author_email='<EMAIL>', license='Apache-2.0', packages=['mlstream', 'mlstream.models'], install_requires=requires, include_package_data=True, zip_safe=False)
gauchm/mlstream
mlstream/models/base_models.py
<gh_stars>1-10 from pathlib import Path import numpy as np from ..datasets import (LumpedBasin, LumpedH5) class LumpedModel: """Model that operates on lumped (daily, basin-averaged) inputs. """ def load(self, model_file: Path) -> None: """Loads a trained and pickled model. Parameters ---------- model_file : Path Path to the stored model. """ pass def train(self, ds: LumpedH5) -> None: """Trains the model. Parameters ---------- ds : LumpedH5 Training dataset """ pass def predict(self, ds: LumpedBasin) -> np.ndarray: """Generates predictions for a basin. Parameters ---------- ds : LumpedBasin Dataset of the basin to predict. Returns ------- np.ndarray Array of predictions. """ pass
gauchm/mlstream
mlstream/models/nseloss.py
<gh_stars>1-10 import torch import numpy as np class NSELoss(torch.nn.Module): """Calculates (batch-wise) NSE Loss. Each sample i is weighted by 1 / (std_i + eps)^2, where std_i is the standard deviation of the discharge of the basin to which the sample belongs. Parameters ---------- eps : float Constant, added to the weight for numerical stability and smoothing, default to 0.1 """ def __init__(self, eps: float = 0.1): super().__init__() self.eps = eps def forward(self, y_pred: torch.Tensor, y_true: torch.Tensor, q_stds: torch.Tensor): """Calculates the batch-wise NSE loss function. Parameters ---------- y_pred : torch.Tensor Tensor containing the network prediction. y_true : torch.Tensor Tensor containing the true discharge values q_stds : torch.Tensor Tensor containing the discharge std (calculated over training period) of each sample Returns ------- torch.Tenor The batch-wise NSE-Loss """ squared_error = (y_pred - y_true)**2 weights = 1 / (q_stds + self.eps)**2 scaled_loss = weights * squared_error return torch.mean(scaled_loss) class XGBNSEObjective: """Custom NSE XGBoost objective. This is a bit of a hack: We use a unique dummy target value for each sample, allowing us to look up the q_std that corresponds to the sample's station. When calculating the loss, we replace the dummy with the actual target so the model learns the right thing. """ def __init__(self, dummy_target, actual_target, q_stds, eps: float = 0.1): self.dummy_target = dummy_target.reshape(-1) self.actual_target = actual_target.reshape(-1) self.q_stds = q_stds.reshape(-1) self.eps = eps def nse_objective_xgb_sklearn_api(self, y_true, y_pred): """NSE objective for XGBoost (sklearn API). """ indices = np.searchsorted(self.dummy_target, y_true) normalization = ((self.q_stds[indices] + self.eps)**2) grad = 2 * (y_pred - self.actual_target[indices]) / normalization hess = 2.0 / normalization return grad, hess def nse_objective_xgb(self, y_pred, dtrain): """NSE objective for XGBoost (non-sklearn API). """ y_true = dtrain.get_label() indices = np.searchsorted(self.dummy_target, y_true) normalization = ((self.q_stds[indices] + self.eps)**2) grad = 2 * (y_pred - self.actual_target[indices]) / normalization hess = 2.0 / normalization return grad, hess def nse(self, y_pred, y_true, q_stds): squared_error = (y_pred - y_true)**2 weights = 1 / (q_stds + self.eps)**2 return np.mean(weights * squared_error) def nse_metric_xgb(self, y_pred, y_true): """NSE metric for XGBoost. """ indices = np.searchsorted(self.dummy_target, y_true.get_label()) nse = self.nse(y_pred, self.actual_target[indices], self.q_stds[indices]) return 'nse', nse def neg_nse_metric_sklearn(self, estimator, X, y_true): """Negative NSE metric for sklearn. """ y_pred = estimator.predict(X) indices = np.searchsorted(self.dummy_target, y_true) return -self.nse(y_pred, self.actual_target[indices], self.q_stds[indices])
gauchm/mlstream
mlstream/models/lstm.py
""" Large parts of this implementation are taken over from https://github.com/kratzert/ealstm_regional_modeling. """ import sys from pathlib import Path from typing import Tuple, Dict import numpy as np from tqdm import tqdm import torch import torch.nn as nn from torch.utils.data import DataLoader, SubsetRandomSampler from ..datasets import LumpedBasin, LumpedH5 from .base_models import LumpedModel from .nseloss import NSELoss DEVICE = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") torch.backends.cudnn.deterministic = True torch.backends.cudnn.benchmark = False class LumpedLSTM(LumpedModel): """(EA-)LSTM model for lumped data. """ def __init__(self, num_dynamic_vars: int, num_static_vars: int, use_mse: bool = True, no_static: bool = False, concat_static: bool = False, run_dir: Path = None, n_jobs: int = 1, hidden_size: int = 256, learning_rate: float = 1e-3, learning_rates: Dict = {}, epochs: int = 30, initial_forget_bias: int = 5, dropout: float = 0.0, batch_size: int = 256, clip_norm: bool = True, clip_value: float = 1.0): input_size_stat = 0 if no_static else num_static_vars input_size_dyn = num_dynamic_vars if (no_static or not concat_static) \ else num_static_vars + num_dynamic_vars self.no_static = no_static self.run_dir = run_dir self.epochs = epochs self.batch_size = batch_size self.learning_rates = learning_rates self.learning_rates[0] = learning_rate self.clip_norm = clip_norm self.clip_value = clip_value self.n_jobs = n_jobs self.use_mse = use_mse self.model = Model(input_size_dyn=input_size_dyn, input_size_stat=input_size_stat, concat_static=concat_static, no_static=no_static, hidden_size=hidden_size, initial_forget_bias=initial_forget_bias, dropout=dropout).to(DEVICE) self.loss_func = nn.MSELoss() if use_mse else NSELoss() def load(self, model_file: Path) -> None: self.model.load_state_dict(torch.load(model_file, map_location=DEVICE)) def train(self, ds: LumpedH5) -> None: val_indices = np.random.choice(len(ds), size=int(0.1 * len(ds)), replace=False) train_indices = [i for i in range(len(ds)) if i not in val_indices] train_sampler = SubsetRandomSampler(train_indices) val_sampler = SubsetRandomSampler(val_indices) self.train_loader = DataLoader(ds, self.batch_size, sampler=train_sampler, drop_last=False, num_workers=self.n_jobs) self.val_loader = DataLoader(ds, self.batch_size, sampler=val_sampler, drop_last=False, num_workers=self.n_jobs) self.optimizer = torch.optim.Adam(self.model.parameters(), lr=self.learning_rates[0]) for epoch in range(1, self.epochs + 1): # set new learning rate if epoch in self.learning_rates.keys(): for param_group in self.optimizer.param_groups: param_group["lr"] = self.learning_rates[epoch] self._train_epoch(epoch) val_loss = self._val_epoch() print(f"# Epoch {epoch}: validation loss: {val_loss:.7f}.") model_path = self.run_dir / f"model_epoch{epoch}.pt" torch.save(self.model.state_dict(), str(model_path)) print(f"Model saved as {model_path}.") def predict(self, ds: LumpedBasin) -> np.ndarray: self.model.eval() loader = DataLoader(ds, batch_size=1024, shuffle=False, num_workers=4) preds, obs = None, None with torch.no_grad(): for data in loader: if len(data) == 2: x, y = data x = x.to(DEVICE) p = self.model(x)[0] elif len(data) == 3: x_d, x_s, y = data x_d, x_s = x_d.to(DEVICE), x_s.to(DEVICE) p = self.model(x_d, x_s[:, 0, :])[0] if preds is None: preds = p.detach().cpu() obs = y else: preds = torch.cat((preds, p.detach().cpu()), 0) obs = torch.cat((obs, y), 0) return preds.numpy(), obs.numpy() def _train_epoch(self, epoch: int): """Trains model for a single epoch. Parameters ---------- epoch : int Current Number of epoch """ self.model.train() # process bar handle pbar = tqdm(self.train_loader, file=sys.stdout) pbar.set_description(f'# Epoch {epoch}') # Iterate in batches over training set running_loss = 0 for i, data in enumerate(pbar): # delete old gradients self.optimizer.zero_grad() # forward pass through LSTM if len(data) == 3: x, y, q_stds = data x, y, q_stds = x.to(DEVICE), y.to(DEVICE), q_stds.to(DEVICE) predictions = self.model(x)[0] # forward pass through EALSTM elif len(data) == 4: x_d, x_s, y, q_stds = data x_d, x_s, y = x_d.to(DEVICE), x_s.to(DEVICE), y.to(DEVICE) predictions = self.model(x_d, x_s[:, 0, :])[0] # MSELoss if self.use_mse: loss = self.loss_func(predictions, y) # NSELoss needs std of each basin for each sample else: q_stds = q_stds.to(DEVICE) loss = self.loss_func(predictions, y, q_stds) # calculate gradients loss.backward() if self.clip_norm: torch.nn.utils.clip_grad_norm_(self.model.parameters(), self.clip_value) # perform parameter update self.optimizer.step() running_loss += loss.item() pbar.set_postfix_str(f"Loss: {loss.item():.6f} / Mean: {running_loss / (i+1):.6f}") def _val_epoch(self) -> float: """Calculates loss on validation set during training. Returns ------- loss : float Mean validation loss """ self.model.eval() loss = 0.0 with torch.no_grad(): for data in self.val_loader: # forward pass through LSTM if len(data) == 3: x, y, q_stds = data x, y, q_stds = x.to(DEVICE), y.to(DEVICE), q_stds.to(DEVICE) predictions = self.model(x)[0] # forward pass through EALSTM elif len(data) == 4: x_d, x_s, y, q_stds = data x_d, x_s, y = x_d.to(DEVICE), x_s.to(DEVICE), y.to(DEVICE) predictions = self.model(x_d, x_s[:, 0, :])[0] # MSELoss if self.use_mse: loss += self.loss_func(predictions, y).item() # NSELoss needs std of each basin for each sample else: q_stds = q_stds.to(DEVICE) loss += self.loss_func(predictions, y, q_stds).item() return loss / len(self.val_loader) class Model(nn.Module): """Wrapper class that connects LSTM/EA-LSTM with fully connceted layer""" def __init__(self, input_size_dyn: int, input_size_stat: int, hidden_size: int, initial_forget_bias: int = 5, dropout: float = 0.0, concat_static: bool = False, no_static: bool = False): """Initializes the model. Parameters ---------- input_size_dyn : int Number of dynamic input features. input_size_stat : int Number of static input features (used in the EA-LSTM input gate). hidden_size : int Number of LSTM cells/hidden units. initial_forget_bias : int Value of the initial forget gate bias. (default: 5) dropout : float Dropout probability in range(0,1). (default: 0.0) concat_static : bool If True, uses standard LSTM otherwise uses EA-LSTM no_static : bool If True, runs standard LSTM """ super(Model, self).__init__() self.input_size_dyn = input_size_dyn self.input_size_stat = input_size_stat self.hidden_size = hidden_size self.initial_forget_bias = initial_forget_bias self.dropout_rate = dropout self.concat_static = concat_static self.no_static = no_static if self.concat_static or self.no_static: self.lstm = LSTM(input_size=input_size_dyn, hidden_size=hidden_size, initial_forget_bias=initial_forget_bias) else: self.lstm = EALSTM(input_size_dyn=input_size_dyn, input_size_stat=input_size_stat, hidden_size=hidden_size, initial_forget_bias=initial_forget_bias) self.dropout = nn.Dropout(p=dropout) self.fc = nn.Linear(hidden_size, 1) def forward(self, x_d: torch.Tensor, x_s: torch.Tensor = None) \ -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: """Run forward pass through the model. Parameters ---------- x_d : torch.Tensor Tensor containing the dynamic input features of shape [batch, seq_length, n_features] x_s : torch.Tensor, optional Tensor containing the static catchment characteristics, by default None Returns ------- out : torch.Tensor Tensor containing the network predictions h_n : torch.Tensor Tensor containing the hidden states of each time step c_n : torch,Tensor Tensor containing the cell states of each time step """ if self.concat_static or self.no_static: h_n, c_n = self.lstm(x_d) else: h_n, c_n = self.lstm(x_d, x_s) last_h = self.dropout(h_n[:, -1, :]) out = self.fc(last_h) return out, h_n, c_n class EALSTM(nn.Module): """Implementation of the Entity-Aware-LSTM (EA-LSTM) Model details: https://arxiv.org/abs/1907.08456 Parameters ---------- input_size_dyn : int Number of dynamic features, which are those, passed to the LSTM at each time step. input_size_stat : int Number of static features, which are those that are used to modulate the input gate. hidden_size : int Number of hidden/memory cells. batch_first : bool, optional If True, expects the batch inputs to be of shape [batch, seq, features] otherwise, the shape has to be [seq, batch, features], by default True. initial_forget_bias : int, optional Value of the initial forget gate bias, by default 0 """ def __init__(self, input_size_dyn: int, input_size_stat: int, hidden_size: int, batch_first: bool = True, initial_forget_bias: int = 0): super(EALSTM, self).__init__() self.input_size_dyn = input_size_dyn self.input_size_stat = input_size_stat self.hidden_size = hidden_size self.batch_first = batch_first self.initial_forget_bias = initial_forget_bias # create tensors of learnable parameters self.weight_ih = nn.Parameter(torch.FloatTensor(input_size_dyn, 3 * hidden_size)) self.weight_hh = nn.Parameter(torch.FloatTensor(hidden_size, 3 * hidden_size)) self.weight_sh = nn.Parameter(torch.FloatTensor(input_size_stat, hidden_size)) self.bias = nn.Parameter(torch.FloatTensor(3 * hidden_size)) self.bias_s = nn.Parameter(torch.FloatTensor(hidden_size)) # initialize parameters self.reset_parameters() def reset_parameters(self): """Initialize all learnable parameters of the LSTM""" nn.init.orthogonal_(self.weight_ih.data) nn.init.orthogonal_(self.weight_sh) weight_hh_data = torch.eye(self.hidden_size) weight_hh_data = weight_hh_data.repeat(1, 3) self.weight_hh.data = weight_hh_data nn.init.constant_(self.bias.data, val=0) nn.init.constant_(self.bias_s.data, val=0) if self.initial_forget_bias != 0: self.bias.data[:self.hidden_size] = self.initial_forget_bias def forward(self, x_d: torch.Tensor, x_s: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]: """Performs a forward pass on the model. Parameters ---------- x_d : torch.Tensor Tensor, containing a batch of sequences of the dynamic features. Shape has to match the format specified with batch_first. x_s : torch.Tensor Tensor, containing a batch of static features. Returns ------- h_n : torch.Tensor The hidden states of each time step of each sample in the batch. c_n : torch.Tensor The cell states of each time step of each sample in the batch. """ if self.batch_first: x_d = x_d.transpose(0, 1) seq_len, batch_size, _ = x_d.size() h_0 = x_d.data.new(batch_size, self.hidden_size).zero_() c_0 = x_d.data.new(batch_size, self.hidden_size).zero_() h_x = (h_0, c_0) # empty lists to temporally store all intermediate hidden/cell states h_n, c_n = [], [] # expand bias vectors to batch size bias_batch = (self.bias.unsqueeze(0).expand(batch_size, *self.bias.size())) # calculate input gate only once because inputs are static bias_s_batch = (self.bias_s.unsqueeze(0).expand(batch_size, *self.bias_s.size())) i = torch.sigmoid(torch.addmm(bias_s_batch, x_s, self.weight_sh)) # perform forward steps over input sequence for t in range(seq_len): h_0, c_0 = h_x # calculate gates gates = (torch.addmm(bias_batch, h_0, self.weight_hh) + torch.mm(x_d[t], self.weight_ih)) f, o, g = gates.chunk(3, 1) c_1 = torch.sigmoid(f) * c_0 + i * torch.tanh(g) h_1 = torch.sigmoid(o) * torch.tanh(c_1) # store intermediate hidden/cell state in list h_n.append(h_1) c_n.append(c_1) h_x = (h_1, c_1) h_n = torch.stack(h_n, 0) c_n = torch.stack(c_n, 0) if self.batch_first: h_n = h_n.transpose(0, 1) c_n = c_n.transpose(0, 1) return h_n, c_n class LSTM(nn.Module): """Implementation of the standard LSTM. Parameters ---------- input_size : int Number of input features hidden_size : int Number of hidden/memory cells. batch_first : bool, optional If True, expects the batch inputs to be of shape [batch, seq, features] otherwise, the shape has to be [seq, batch, features], by default True. initial_forget_bias : int, optional Value of the initial forget gate bias, by default 0 """ def __init__(self, input_size: int, hidden_size: int, batch_first: bool = True, initial_forget_bias: int = 0): super(LSTM, self).__init__() self.input_size = input_size self.hidden_size = hidden_size self.batch_first = batch_first self.initial_forget_bias = initial_forget_bias # create tensors of learnable parameters self.weight_ih = nn.Parameter(torch.FloatTensor(input_size, 4 * hidden_size)) self.weight_hh = nn.Parameter(torch.FloatTensor(hidden_size, 4 * hidden_size)) self.bias = nn.Parameter(torch.FloatTensor(4 * hidden_size)) # initialize parameters self.reset_parameters() def reset_parameters(self): """Initializes all learnable parameters of the LSTM. """ nn.init.orthogonal_(self.weight_ih.data) weight_hh_data = torch.eye(self.hidden_size) weight_hh_data = weight_hh_data.repeat(1, 4) self.weight_hh.data = weight_hh_data nn.init.constant_(self.bias.data, val=0) if self.initial_forget_bias != 0: self.bias.data[:self.hidden_size] = self.initial_forget_bias def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]: """Performs a forward pass on the model. Parameters ---------- x : torch.Tensor Tensor, containing a batch of input sequences. Format must match the specified format, defined by the batch_first agrument. Returns ------- h_n : torch.Tensor The hidden states of each time step of each sample in the batch. c_n : torch.Tensor The cell states of each time step of each sample in the batch. """ if self.batch_first: x = x.transpose(0, 1) seq_len, batch_size, _ = x.size() h_0 = x.data.new(batch_size, self.hidden_size).zero_() c_0 = x.data.new(batch_size, self.hidden_size).zero_() h_x = (h_0, c_0) # empty lists to temporally store all intermediate hidden/cell states h_n, c_n = [], [] # expand bias vectors to batch size bias_batch = (self.bias.unsqueeze(0).expand(batch_size, *self.bias.size())) # perform forward steps over input sequence for t in range(seq_len): h_0, c_0 = h_x # calculate gates gates = (torch.addmm(bias_batch, h_0, self.weight_hh) + torch.mm(x[t], self.weight_ih)) f, i, o, g = gates.chunk(4, 1) c_1 = torch.sigmoid(f) * c_0 + torch.sigmoid(i) * torch.tanh(g) h_1 = torch.sigmoid(o) * torch.tanh(c_1) # store intermediate hidden/cell state in list h_n.append(h_1) c_n.append(c_1) h_x = (h_1, c_1) h_n = torch.stack(h_n, 0) c_n = torch.stack(c_n, 0) if self.batch_first: h_n = h_n.transpose(0, 1) c_n = c_n.transpose(0, 1) return h_n, c_n