#!/usr/bin/env python #$ -cwd #$ -S /usr/bin/python ########################### usual importation of modules ##################### # import of standard module import sys import os from optparse import OptionParser # import the user configuration from ~/.pythonrc.py import user ######################### importation of pyVerif modules ##################### # import necessary classes from pyVerif from pyVerif.Executor import Executor from pyVerifseb.MyExperiments import GenerativeExperiment ##################### the "main" program ################## if __name__ == '__main__': print """ pyVerif program to perform a face verification experiment based on GMM/MAP adaptation. Authors: Sebastien Marcel, Johnny Mariethoz and Olivier Bornet Version: 1.0 Date: August 2006 """ # program usage usage = "usage: %prog [options]" parser = OptionParser(usage) # set options parser.set_description ("pyVerif program to perform a face verification experiment") parser.add_option ("-t", "--test", dest="test", help="Perform also test set", action="store_true", default=False) parser.add_option ("-d", "--features-dir", dest="features_dir", help="Directory containing the features files", type="string", default="none") parser.add_option ("-o", "--output-dir", dest="output_dir", help="Directory containing the results files", type="string", default="./noname") parser.add_option ("-p", "--protocol-dir", dest="protocol_dir", help="Directory containing the protocol files", type="string", default="lp") parser.add_option ("-b", "--base-dir", dest="base_dir", help="Directory containing protocols and features " + "(default xm2vts)", type="string", default="xm2vts") parser.add_option ("-v", "--variant", dest="variant", help="Variant of the protocol (default lp1)", type="string", default="lp1") parser.add_option ("-l", "--log-on-screen", dest="log_on_screen", help="Output log on screen", action="store_true", default=False) parser.add_option ("-n", "--number-of-gaussians", dest="number_of_gaussians", help="Number of Gaussians for each model " + "(default 512)", type="int", default=512) parser.add_option ("-m", "--map-factor", dest="map_factor", help="MAP adaptation factor (default 0.5)", type="float", default=0.5) parser.add_option ("-s", "--seed", dest="seed", help="seed (default -1)", type="int", default=-1) parser.add_option ("", "--prior-train", dest="prior_train", help="prior for world model training (default 0.1)", type="float", default=0.1) parser.add_option ("", "--prior-adapt", dest="prior_adapt", help="prior for model adaptation (default 0.1)", type="float", default=0.1) parser.add_option ("", "--variance-flooring", dest="variance_floor", help="variance flooring for training (default 0.0005)", type="float", default=0.0005) parser.add_option ("", "--accuracy", dest="accuracy", help="end of accuracy (default 0.0005)", type="float", default=0.0005) parser.add_option ("", "--kmean-iter", dest="kmean_iter", help="maximum number of K-means iterations for both training (default 500)", type="int", default=500) parser.add_option ("", "--em-iter", dest="em_iter", help="maximum number of EM iterations for both training (default 500)", type="int", default=500) parser.add_option ("", "--map-iter", dest="map_iter", help="maximum number of EM iterations for adaptation (default 10)", type="int", default=10) parser.add_option ("-w", "--worldmodel", dest="worldmodel", help="Prior world model to use", type="string", default="") parser.add_option ("", "--perf", dest="compute_perf", help="compute the performance", action="store_true", default=False) # parse options (options,args) = parser.parse_args () # check the number of options if len(args) != 0: parser.error("Error: incorrect number of arguments, try --help") if options.features_dir == "none": parser.error("Error: no feature dir provided") # # prepare the directories # # base dir where is this script (for instance /home/vision/marcel/work/experiments/faceverification/pyverif/verification) base_dir = os.path.join ('/home', 'vision', 'marcel', 'work', 'experiments', 'faceverification', 'pyverif', options.base_dir) # directory containing the features features_dir = os.path.join (base_dir, 'features', options.features_dir) # directory containing the binary programs bin_dir = os.path.join (base_dir, 'verification','bin') # directory for the output files out_dir = os.path.join (options.output_dir, options.variant) # base dir containing the protocol files base_dir_protocols = os.path.join (base_dir, 'protocols', options.protocol_dir) # directory containing the results (created if it does'nt exist) results_dir = os.path.join (out_dir, 'results') if not os.path.isdir (results_dir): os.makedirs (results_dir) # directory containing the temporary files (created if it does'nt exist) tmp_dir = os.path.join (out_dir, 'tmp') if not os.path.isdir (tmp_dir): os.makedirs (tmp_dir) # # prepare the experiment # # create an instance of the Experiment class experiment = GenerativeExperiment ( out_dir = out_dir, features_dir = features_dir, features_ext = '.bindata', bin_dir = bin_dir, tmp_dir = tmp_dir, log_on_screen = options.log_on_screen, log_file = os.path.join (out_dir, 'gmm-' + options.variant + '.log'), norm_file = 'train_world_' + options.variant + '.lst') # log the used options experiment.log_options (options.__dict__) # load the specified protocol experiment.load_protocols (base_dir_protocols, options.variant) # specify the main program to train and test a GMM cmd = "gmm" # check if a world model is provided if not options.worldmodel: prior_model = os.path.join (experiment.world_models_dir, experiment.world_model_name) else: prior_model = options.worldmodel print "using prior model %s" % prior_model experiment.world_models_dir = prior_model[:prior_model.rfind('/')] experiment.world_model_name = prior_model[prior_model.rfind('/')+1:len(prior_model)] #print experiment.world_models_dir, experiment.world_model_name # options to train an GMM (world model training) train_opt = "-bin " train_opt = train_opt + "-prior %g " % options.prior_train train_opt = train_opt + "-iterk %d " % options.kmean_iter train_opt = train_opt + "-iterg %d " % options.em_iter train_opt = train_opt + "-threshold %g " % options.variance_floor train_opt = train_opt + "-e %g " % options.accuracy train_opt = train_opt + "-seed %d " % options.seed train_opt = train_opt + "-n_gaussians %s " % options.number_of_gaussians train_opt = train_opt + "-dir %TMPDIR -save %MODEL %FILES " train_opt = train_opt + " > %TMPDIR/train_model.stdout" # options to adapt a GMM (client model training) adapt_opt = "--adapt -bin " adapt_opt = adapt_opt + "-prior %g " % options.prior_adapt adapt_opt = adapt_opt + "-iterg %d " % options.map_iter adapt_opt = adapt_opt + "-e %g " % options.accuracy adapt_opt = adapt_opt + "-map %s " % options.map_factor adapt_opt = adapt_opt + "-dir %TMPDIR -save %MODEL " + prior_model + " %FILES " adapt_opt = adapt_opt + " > %TMPDIR/adapt_model.stdout" # options to compute the scores against a GMM (world/client model testing) test_opt = "--test -bin -dir %TMPDIR %MODEL %FILES %SCORES " test_opt = test_opt + " > %TMPDIR/test_model.stdout" # set main program and options experiment.set_cmds (cmd=cmd, world_options=train_opt, test_options=test_opt, client_options=adapt_opt) # # run the experiment # # train the world model if no world model provided if not options.worldmodel: experiment.train_world_model () # compute all the scores against the world model for the development set experiment.build_world_scores (for_dev = True) # train all the client models experiment.train_clients_models (for_dev = True) # compute all the scores against the clients models for the development set experiment.build_clients_scores (for_dev = True) # merge scores from the world model and from the client model for the development set experiment.merge_score (os.path.join (results_dir, 'scores-dev'), for_dev = True) if options.test: # compute all the scores against the world model for the test set experiment.build_world_scores (for_dev = False) # Warning: no client models to train in test with XM2VTS if options.base_dir != "xm2vts": # train all the client models experiment.train_clients_models (for_dev=False) # compute all the scores against the clients models for the test set experiment.build_clients_scores (for_dev = False) # merge scores from the world model and from the client model for the test set experiment.merge_score (os.path.join (results_dir, 'scores-test'), for_dev = False) if options.compute_perf: # compute the performance with pymeasure executor = Executor() executor.enable_log(on_screen=True,filename=experiment.executor.log_filename) cmd_perf = "pyerror" perf_opt = "-b --dev-set=" + os.path.join (results_dir, 'scores-dev') + " " + os.path.join (results_dir, 'scores-test') executor.run(cmd_perf,perf_opt) # generate DET/EPC curves cmd_perf = "pydet" perf_opt = "-b -s " + os.path.join (results_dir, 'det.eps') + " " + os.path.join (results_dir, 'scores-dev') + " " + os.path.join (results_dir, 'scores-test') executor.run(cmd_perf,perf_opt) cmd_perf = "pyepc" perf_opt = "-b -n 10 -s " + os.path.join (results_dir, 'epc.eps') + " " + os.path.join (results_dir, 'scores-dev') + " " + os.path.join (results_dir, 'scores-test') executor.run(cmd_perf,perf_opt)