def write_file(group):
if args.verbose:
print "Processing '%s' group..." % group
out = open(os.path.join(args.outputdir, '%s-5col.txt' % group), 'wt')
if group == 'train': reals, attacks = db.get_train_data()
elif group == 'devel': reals, attacks = db.get_devel_data()
elif group == 'test': reals, attacks = db.get_test_data()
else:
raise RuntimeError, "group parameter has to be train, devel or test"
total = len(reals) + len(attacks)
counter = 0
for obj in reals:
counter += 1
if args.verbose:
print "Processing file %s [%d/%d]..." % (obj.make_path(),
counter, total)
arr = obj.load(args.inputdir, '.hdf5')
arr = arr[~numpy.isnan(arr)] #remove NaN entries => invalid
avg = numpy.mean(arr[:args.average])
# This is a tremendous disencapsulation, but can't do it better for now
client_id = obj._File__f.client.id
out.write('%d %d %d %s %.5e\n' %
(client_id, client_id, client_id, obj.make_path(), avg))
for obj in attacks:
counter += 1
if args.verbose:
print "Processing file %s [%d/%d]..." % (obj.make_path(),
counter, total)
arr = obj.load(args.inputdir, '.hdf5')
arr = arr[~numpy.isnan(arr)] #remove NaN entries => invalid
avg = numpy.mean(arr[:args.average])
# This is a tremendous disencapsulation, but can't do it better for now
client_id = obj._File__f.client.id
out.write('%d %d attack %s %.5e\n' %
(client_id, client_id, obj.make_path(), avg))
out.close()
def write_file(group):
if args.verbose:
print("Processing '%s' group..." % group)
out = open(os.path.join(args.outputdir, '%s-5col.txt' % group), 'wt')
if group == 'train': reals, attacks = db.get_train_data()
elif group == 'devel': reals, attacks = db.get_devel_data()
elif group == 'test': reals, attacks = db.get_test_data()
else: raise RuntimeError("group parameter has to be train, devel or test")
total = len(reals) + len(attacks)
counter = 0
for obj in reals:
counter += 1
if args.verbose:
print("Processing file %s [%d/%d]..." % (obj.make_path(), counter, total))
arr = obj.load(args.inputdir, '.hdf5')
arr = arr[~numpy.isnan(arr)] #remove NaN entries => invalid
avg = numpy.mean(arr[beginScoresInterval:endScoresInterval])
# This is a tremendous disencapsulation, but can't do it better for now
client_id = obj.get_client_id()
out.write('%d %d %d %s %.5e\n' % (client_id, client_id, client_id,
obj.make_path(), avg))
for obj in attacks:
counter += 1
if args.verbose:
print("Processing file %s [%d/%d]..." % (obj.make_path(), counter, total))
arr = obj.load(args.inputdir, '.hdf5')
arr = arr[~numpy.isnan(arr)] #remove NaN entries => invalid
avg = numpy.mean(arr[beginScoresInterval:endScoresInterval])
# This is a tremendous disencapsulation, but can't do it better for now
client_id = obj.get_client_id()
out.write('%d %d attack %s %.5e\n' % (client_id, client_id,
obj.make_path(), avg))
out.close()
def main():
"""Main method"""
basedir = os.path.dirname(os.path.dirname(os.path.realpath(sys.argv[0])))
INPUTDIR = os.path.join(basedir, 'quantities')
OUTPUTDIR = os.path.join(basedir, 'mlp')
parser = argparse.ArgumentParser(description=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter)
parser.add_argument('inputdir', metavar='DIR', type=str, default=INPUTDIR, nargs='?', help='Base directory containing the 5-quantities to be loaded. Final MLP input will be generated from the input directory and concatenated column-wise to form the final training matrix (defaults to "%(default)s").')
parser.add_argument('outputdir', metavar='DIR', type=str, default=OUTPUTDIR, nargs='?', help='Base directory that will be used to save the results. The given value will be interpolated with time.strftime and then, os.environ (in this order), so you can include %%()s strings (e.g. %%(SGE_TASK_ID)s) to make up the final output directory path (defaults to "%(default)s").')
parser.add_argument('-b', '--batch-size', metavar='INT', type=int,
dest='batch', default=200, help='The number of samples per training iteration. Good values are greater than 100. Defaults to %(default)s')
parser.add_argument('-e', '--epoch', metavar='INT', type=int,
dest='epoch', default=1, help='This is the number of training steps that need to be executed before we attempt to measure the error on the development set. Defaults to %(default)s')
parser.add_argument('-n', '--hidden-neurons', metavar='INT', type=int,
dest='nhidden', default=5, help='The number hidden neurons in the (single) hidden layer of the MLP. Defaults to %(default)s')
parser.add_argument('-m', '--maximum-iterations', metavar='INT', type=int,
dest='maxiter', default=0, help='The maximum number of iterations to train for. A value of zero trains until a valley is detected on the development set. Defaults to %(default)s')
parser.add_argument('-i', '--no-improvements', metavar='INT', type=int,
dest='noimprov', default=0, help='The maximum number of iterations to wait for in case no improvements happen in the development set average RMSE. If that number of iterations is reached, the training is stopped. Values in the order of 10-20%% of the maximum number of iterations should be a reasonable default. If set to zero, do not consider this stop criteria. Defaults to %(default)s')
parser.add_argument('-f', '--overwrite', action='store_true',
dest='overwrite', default=False, help='If set and the destination directory exists, overwrite the results contained there')
parser.add_argument('-V', '--verbose', action='store_true', dest='verbose',
default=False, help='Increases this script verbosity')
# Adds database support using the common infrastructure
# N.B.: Only databases with 'video' support
import antispoofing.utils.db
antispoofing.utils.db.Database.create_parser(parser, 'video')
args = parser.parse_args()
start_time = time.time()
paramfile = ConfigParser.SafeConfigParser()
paramfile.add_section('time')
paramfile.set('time', 'start', time.asctime())
if args.verbose: print "Start time is", time.asctime()
if not os.path.exists(args.inputdir):
parser.error("input directory `%s' does not exist" % args.inputdir)
use_outputdir = time.strftime(args.outputdir) #interpolate time
use_outputdir = use_outputdir % os.environ #interpolate environment
if os.path.exists(use_outputdir):
if not args.overwrite:
parser.error("output directory '%s' exists and the overwrite flag was not set" % use_outputdir)
else:
bob.db.utils.makedirs_safe(use_outputdir)
if args.verbose: print "Output directory set to \"%s\"" % use_outputdir
use_inputdir = []
abspath = os.path.abspath(args.inputdir)
use_inputdir.append(abspath)
if args.verbose: print "Loading non-NaN entries from input files at '%s' database..." % args.name
db = args.cls(args)
data = {
'train': dict(zip(('real', 'attack'), db.get_train_data())),
'devel': dict(zip(('real', 'attack'), db.get_devel_data())),
'test' : dict(zip(('real', 'attack'), db.get_test_data())),
}
def merge_data(flist):
d = bob.io.load([k.make_path(use_inputdir[0], '.hdf5') for k in flist])
return d[~numpy.isnan(d.sum(axis=1)),:]
for key in data.keys():
for cls in data[key].keys():
if args.verbose: print "Loading %-5s/%-6s:" % (key, cls),
data[key][cls] = merge_data(data[key][cls])
if args.verbose: print len(data[key][cls])
if args.verbose: print "Training MLP..."
mlp, evolution = ml.rprop.make_mlp((data['train']['real'],
data['train']['attack']), (data['devel']['real'], data['devel']['attack']),
args.batch, args.nhidden, args.epoch, args.maxiter, args.noimprov,
args.verbose)
if args.verbose: print "Saving session information..."
def get_version(package):
__import__('pkg_resources').require(package)[0].version
paramfile.add_section('software')
for package in __import__('pkg_resources').require('antispoofing.motion'):
paramfile.set('software', package.key, package.version)
paramfile.add_section('environment')
cmdline = [os.path.realpath(sys.argv[0])] + sys.argv[1:]
paramfile.set('environment', 'command-line', ' '.join(cmdline))
paramfile.add_section('data')
datapath = [os.path.realpath(k) for k in use_inputdir]
paramfile.set('data', 'database', args.name)
paramfile.set('data', 'input', '\n'.join(datapath))
paramfile.set('data', 'train-real', str(len(data['train']['real'])))
paramfile.set('data', 'train-attack', str(len(data['train']['attack'])))
paramfile.set('data', 'devel-real', str(len(data['devel']['real'])))
paramfile.set('data', 'devel-attack', str(len(data['devel']['attack'])))
paramfile.set('data', 'test-real', str(len(data['test']['real'])))
paramfile.set('data', 'test-attack', str(len(data['test']['attack'])))
paramfile.add_section('mlp')
paramfile.set('mlp', 'shape', '-'.join([str(k) for k in mlp.shape]))
paramfile.set('mlp', 'batch-size', str(args.batch))
paramfile.set('mlp', 'epoch-size', str(args.epoch))
paramfile.set('mlp', 'maximum-iterations', str(args.maxiter))
if args.verbose: print "Saving MLP..."
mlpfile = bob.io.HDF5File(os.path.join(use_outputdir, 'mlp.hdf5'),'w')
mlp.save(mlpfile)
del mlpfile
if args.verbose: print "Saving result evolution..."
evofile = bob.io.HDF5File(os.path.join(use_outputdir,
'training-evolution.hdf5'),'w')
evolution.save(evofile)
del evofile
if args.verbose: print "Running analysis..."
evolution.report(mlp, (data['test']['real'], data['test']['attack']),
os.path.join(use_outputdir, 'plots.pdf'), paramfile)
paramfile.set('time', 'end', time.asctime())
total_time = int(time.time() - start_time)
diff = datetime.timedelta(seconds=total_time)
paramfile.set('time', 'duration', str(diff))
paramfile.set('time', 'host', socket.getfqdn())
if args.verbose: print "End time is", time.asctime()
if args.verbose: print "Total training time:", str(diff)
paramfile.write(open(os.path.join(use_outputdir, 'session.txt'), 'wb'))
if args.verbose: print "All done, bye!"
def main():
"""Main method"""
basedir = os.path.dirname(os.path.dirname(os.path.realpath(sys.argv[0])))
INPUTDIR = os.path.join(basedir, 'quantities')
OUTPUTDIR = os.path.join(basedir, 'lda')
parser = argparse.ArgumentParser(description=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter)
parser.add_argument('inputdir', metavar='DIR', type=str, default=INPUTDIR, nargs='?', help='Base directory containing the 5-quantities to be loaded. Final MLP input will be generated from the input directory and concatenated column-wise to form the final training matrix (defaults to "%(default)s").')
parser.add_argument('outputdir', metavar='DIR', type=str, default=OUTPUTDIR, nargs='?', help='Base directory that will be used to save the results. The given value will be interpolated with time.strftime and then, os.environ (in this order), so you can include %%()s strings (e.g. %%(SGE_TASK_ID)s) to make up the final output directory path (defaults to "%(default)s").')
parser.add_argument('-f', '--overwrite', action='store_true',
dest='overwrite', default=False, help='If set and the destination directory exists, overwrite the results contained there')
parser.add_argument('-V', '--verbose', action='store_true', dest='verbose',
default=False, help='Increases this script verbosity')
# Adds database support using the common infrastructure
# N.B.: Only databases with 'video' support
import antispoofing.utils.db
antispoofing.utils.db.Database.create_parser(parser, 'video')
args = parser.parse_args()
start_time = time.time()
paramfile = ConfigParser.SafeConfigParser()
paramfile.add_section('time')
paramfile.set('time', 'start', time.asctime())
if args.verbose: print "Start time is", time.asctime()
if not os.path.exists(args.inputdir):
parser.error("input directory `%s' does not exist" % args.inputdir)
use_outputdir = time.strftime(args.outputdir) #interpolate time
use_outputdir = use_outputdir % os.environ #interpolate environment
if os.path.exists(use_outputdir):
if not args.overwrite:
parser.error("output directory '%s' exists and the overwrite flag was not set" % use_outputdir)
else:
bob.db.utils.makedirs_safe(use_outputdir)
if args.verbose: print "Output directory set to \"%s\"" % use_outputdir
use_inputdir = []
abspath = os.path.abspath(args.inputdir)
use_inputdir.append(abspath)
if args.verbose: print "Loading non-NaN entries from input files at '%s' database..." % args.name
db = args.cls(args)
real, attack = db.get_train_data()
def merge_data(flist):
d = bob.io.load([k.make_path(use_inputdir[0], '.hdf5') for k in flist])
return d[~numpy.isnan(d.sum(axis=1)),:]
real = merge_data(real)
attack = merge_data(attack)
if args.verbose: print "Evaluating mean and standard deviation..."
from antispoofing.utils.ml.norm import calc_mean_std
mean, std = calc_mean_std(real, attack, nonStdZero=True)
if args.verbose: print "Training LDA..."
from antispoofing.utils.ml.norm import zeromean_unitvar_norm
real = zeromean_unitvar_norm(real, mean, std)
attack = zeromean_unitvar_norm(attack, mean, std)
from antispoofing.utils.ml.lda import make_lda
machine = make_lda([real, attack])
# adjust some details of the final machine to be saved
machine.resize(machine.shape[0], 1)
# so we get real and attacks on the "right" side of the axis
machine.weights = -1 * machine.weights
machine.input_subtract = mean
machine.input_divide = std
if args.verbose: print "Performance evaluation:"
real, attack = db.get_devel_data()
real = merge_data(real)
attack = merge_data(attack)
pos = machine(real)[:,0]
neg = machine(attack)[:,0]
thres = bob.measure.eer_threshold(neg, pos)
far, frr = bob.measure.farfrr(neg, pos, thres)
good_neg = bob.measure.correctly_classified_negatives(neg, thres).sum()
good_pos = bob.measure.correctly_classified_positives(pos, thres).sum()
print " -> EER @ devel set threshold: %.5e" % thres
print " -> Devel set results:"
print " * FAR : %.3f%% (%d/%d)" % (100*far, len(neg)-good_neg, len(neg))
print " * FRR : %.3f%% (%d/%d)" % (100*frr, len(pos)-good_pos, len(pos))
print " * HTER: %.3f%%" % (50*(far+frr))
real, attack = db.get_test_data()
real = merge_data(real)
attack = merge_data(attack)
pos = machine(real)[:,0]
neg = machine(attack)[:,0]
far, frr = bob.measure.farfrr(neg, pos, thres)
good_neg = bob.measure.correctly_classified_negatives(neg, thres).sum()
good_pos = bob.measure.correctly_classified_positives(pos, thres).sum()
print " -> Test set results:"
print " * FAR: %.3f%% (%d/%d)" % (100*far, len(neg)-good_neg, len(neg))
print " * FRR: %.3f%% (%d/%d)" % (100*frr, len(pos)-good_pos, len(pos))
print " * HTER: %.3f%%" % (50*(far+frr))
if args.verbose: print "Saving session information..."
def get_version(package):
__import__('pkg_resources').require(package)[0].version
paramfile.add_section('software')
for package in __import__('pkg_resources').require('antispoofing.motion'):
paramfile.set('software', package.key, package.version)
paramfile.add_section('environment')
cmdline = [os.path.realpath(sys.argv[0])] + sys.argv[1:]
paramfile.set('environment', 'command-line', ' '.join(cmdline))
paramfile.add_section('data')
datapath = [os.path.realpath(k) for k in use_inputdir]
paramfile.set('data', 'database', args.name)
paramfile.set('data', 'input', '\n'.join(datapath))
paramfile.set('data', 'train-real', str(len(real)))
paramfile.set('data', 'train-attack', str(len(attack)))
paramfile.add_section('lda')
paramfile.set('lda', 'shape', '-'.join([str(k) for k in machine.shape]))
if args.verbose: print "Saving LDA machine..."
machfile = bob.io.HDF5File(os.path.join(use_outputdir, 'lda.hdf5'),'w')
machine.save(machfile)
del machfile
paramfile.set('time', 'end', time.asctime())
total_time = int(time.time() - start_time)
diff = datetime.timedelta(seconds=total_time)
paramfile.set('time', 'duration', str(diff))
paramfile.set('time', 'host', socket.getfqdn())
if args.verbose: print "End time is", time.asctime()
if args.verbose: print "Total training time:", str(diff)
paramfile.write(open(os.path.join(use_outputdir, 'session.txt'), 'wb'))
if args.verbose: print "All done, bye!"
def main():
"""Main method"""
basedir = os.path.dirname(os.path.dirname(os.path.realpath(sys.argv[0])))
INPUTDIR = os.path.join(basedir, 'quantities')
OUTPUTDIR = os.path.join(basedir, 'mlp')
parser = argparse.ArgumentParser(
description=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter)
parser.add_argument(
'inputdir',
metavar='DIR',
type=str,
default=INPUTDIR,
nargs='?',
help=
'Base directory containing the 5-quantities to be loaded. Final MLP input will be generated from the input directory and concatenated column-wise to form the final training matrix (defaults to "%(default)s").'
)
parser.add_argument(
'outputdir',
metavar='DIR',
type=str,
default=OUTPUTDIR,
nargs='?',
help=
'Base directory that will be used to save the results. The given value will be interpolated with time.strftime and then, os.environ (in this order), so you can include %%()s strings (e.g. %%(SGE_TASK_ID)s) to make up the final output directory path (defaults to "%(default)s").'
)
parser.add_argument(
'-b',
'--batch-size',
metavar='INT',
type=int,
dest='batch',
default=200,
help=
'The number of samples per training iteration. Good values are greater than 100. Defaults to %(default)s'
)
parser.add_argument(
'-e',
'--epoch',
metavar='INT',
type=int,
dest='epoch',
default=1,
help=
'This is the number of training steps that need to be executed before we attempt to measure the error on the development set. Defaults to %(default)s'
)
parser.add_argument(
'-n',
'--hidden-neurons',
metavar='INT',
type=int,
dest='nhidden',
default=5,
help=
'The number hidden neurons in the (single) hidden layer of the MLP. Defaults to %(default)s'
)
parser.add_argument(
'-m',
'--maximum-iterations',
metavar='INT',
type=int,
dest='maxiter',
default=0,
help=
'The maximum number of iterations to train for. A value of zero trains until a valley is detected on the development set. Defaults to %(default)s'
)
parser.add_argument(
'-i',
'--no-improvements',
metavar='INT',
type=int,
dest='noimprov',
default=0,
help=
'The maximum number of iterations to wait for in case no improvements happen in the development set average RMSE. If that number of iterations is reached, the training is stopped. Values in the order of 10-20%% of the maximum number of iterations should be a reasonable default. If set to zero, do not consider this stop criteria. Defaults to %(default)s'
)
parser.add_argument(
'-f',
'--overwrite',
action='store_true',
dest='overwrite',
default=False,
help=
'If set and the destination directory exists, overwrite the results contained there'
)
parser.add_argument('-V',
'--verbose',
action='store_true',
dest='verbose',
default=False,
help='Increases this script verbosity')
# Adds database support using the common infrastructure
# N.B.: Only databases with 'video' support
import antispoofing.utils.db
antispoofing.utils.db.Database.create_parser(parser, 'video')
args = parser.parse_args()
start_time = time.time()
paramfile = ConfigParser.SafeConfigParser()
paramfile.add_section('time')
paramfile.set('time', 'start', time.asctime())
if args.verbose: print "Start time is", time.asctime()
if not os.path.exists(args.inputdir):
parser.error("input directory `%s' does not exist" % args.inputdir)
use_outputdir = time.strftime(args.outputdir) #interpolate time
use_outputdir = use_outputdir % os.environ #interpolate environment
if os.path.exists(use_outputdir):
if not args.overwrite:
parser.error(
"output directory '%s' exists and the overwrite flag was not set"
% use_outputdir)
else:
bob.db.utils.makedirs_safe(use_outputdir)
if args.verbose: print "Output directory set to \"%s\"" % use_outputdir
use_inputdir = []
abspath = os.path.abspath(args.inputdir)
use_inputdir.append(abspath)
if args.verbose:
print "Loading non-NaN entries from input files at '%s' database..." % args.name
db = args.cls(args)
data = {
'train': dict(zip(('real', 'attack'), db.get_train_data())),
'devel': dict(zip(('real', 'attack'), db.get_devel_data())),
'test': dict(zip(('real', 'attack'), db.get_test_data())),
}
def merge_data(flist):
d = bob.io.load([k.make_path(use_inputdir[0], '.hdf5') for k in flist])
return d[~numpy.isnan(d.sum(axis=1)), :]
for key in data.keys():
for cls in data[key].keys():
if args.verbose: print "Loading %-5s/%-6s:" % (key, cls),
data[key][cls] = merge_data(data[key][cls])
if args.verbose: print len(data[key][cls])
if args.verbose: print "Training MLP..."
mlp, evolution = ml.rprop.make_mlp(
(data['train']['real'], data['train']['attack']),
(data['devel']['real'], data['devel']['attack']), args.batch,
args.nhidden, args.epoch, args.maxiter, args.noimprov, args.verbose)
if args.verbose: print "Saving session information..."
def get_version(package):
__import__('pkg_resources').require(package)[0].version
paramfile.add_section('software')
for package in __import__('pkg_resources').require('antispoofing.motion'):
paramfile.set('software', package.key, package.version)
paramfile.add_section('environment')
cmdline = [os.path.realpath(sys.argv[0])] + sys.argv[1:]
paramfile.set('environment', 'command-line', ' '.join(cmdline))
paramfile.add_section('data')
datapath = [os.path.realpath(k) for k in use_inputdir]
paramfile.set('data', 'database', args.name)
paramfile.set('data', 'input', '\n'.join(datapath))
paramfile.set('data', 'train-real', str(len(data['train']['real'])))
paramfile.set('data', 'train-attack', str(len(data['train']['attack'])))
paramfile.set('data', 'devel-real', str(len(data['devel']['real'])))
paramfile.set('data', 'devel-attack', str(len(data['devel']['attack'])))
paramfile.set('data', 'test-real', str(len(data['test']['real'])))
paramfile.set('data', 'test-attack', str(len(data['test']['attack'])))
paramfile.add_section('mlp')
paramfile.set('mlp', 'shape', '-'.join([str(k) for k in mlp.shape]))
paramfile.set('mlp', 'batch-size', str(args.batch))
paramfile.set('mlp', 'epoch-size', str(args.epoch))
paramfile.set('mlp', 'maximum-iterations', str(args.maxiter))
if args.verbose: print "Saving MLP..."
mlpfile = bob.io.HDF5File(os.path.join(use_outputdir, 'mlp.hdf5'), 'w')
mlp.save(mlpfile)
del mlpfile
if args.verbose: print "Saving result evolution..."
evofile = bob.io.HDF5File(
os.path.join(use_outputdir, 'training-evolution.hdf5'), 'w')
evolution.save(evofile)
del evofile
if args.verbose: print "Running analysis..."
evolution.report(mlp, (data['test']['real'], data['test']['attack']),
os.path.join(use_outputdir, 'plots.pdf'), paramfile)
paramfile.set('time', 'end', time.asctime())
total_time = int(time.time() - start_time)
diff = datetime.timedelta(seconds=total_time)
paramfile.set('time', 'duration', str(diff))
paramfile.set('time', 'host', socket.getfqdn())
if args.verbose: print "End time is", time.asctime()
if args.verbose: print "Total training time:", str(diff)
paramfile.write(open(os.path.join(use_outputdir, 'session.txt'), 'wb'))
if args.verbose: print "All done, bye!"
def main():
"""Main method"""
basedir = os.path.dirname(os.path.dirname(os.path.realpath(sys.argv[0])))
INPUTDIR = os.path.join(basedir, 'quantities')
OUTPUTDIR = os.path.join(basedir, 'lda')
parser = argparse.ArgumentParser(
description=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter)
parser.add_argument(
'inputdir',
metavar='DIR',
type=str,
default=INPUTDIR,
nargs='?',
help=
'Base directory containing the 5-quantities to be loaded. Final MLP input will be generated from the input directory and concatenated column-wise to form the final training matrix (defaults to "%(default)s").'
)
parser.add_argument(
'outputdir',
metavar='DIR',
type=str,
default=OUTPUTDIR,
nargs='?',
help=
'Base directory that will be used to save the results. The given value will be interpolated with time.strftime and then, os.environ (in this order), so you can include %%()s strings (e.g. %%(SGE_TASK_ID)s) to make up the final output directory path (defaults to "%(default)s").'
)
parser.add_argument(
'-f',
'--overwrite',
action='store_true',
dest='overwrite',
default=False,
help=
'If set and the destination directory exists, overwrite the results contained there'
)
parser.add_argument('-V',
'--verbose',
action='store_true',
dest='verbose',
default=False,
help='Increases this script verbosity')
# Adds database support using the common infrastructure
# N.B.: Only databases with 'video' support
import antispoofing.utils.db
antispoofing.utils.db.Database.create_parser(parser, 'video')
args = parser.parse_args()
start_time = time.time()
paramfile = ConfigParser.SafeConfigParser()
paramfile.add_section('time')
paramfile.set('time', 'start', time.asctime())
if args.verbose: print "Start time is", time.asctime()
if not os.path.exists(args.inputdir):
parser.error("input directory `%s' does not exist" % args.inputdir)
use_outputdir = time.strftime(args.outputdir) #interpolate time
use_outputdir = use_outputdir % os.environ #interpolate environment
if os.path.exists(use_outputdir):
if not args.overwrite:
parser.error(
"output directory '%s' exists and the overwrite flag was not set"
% use_outputdir)
else:
bob.db.utils.makedirs_safe(use_outputdir)
if args.verbose: print "Output directory set to \"%s\"" % use_outputdir
use_inputdir = []
abspath = os.path.abspath(args.inputdir)
use_inputdir.append(abspath)
if args.verbose:
print "Loading non-NaN entries from input files at '%s' database..." % args.name
db = args.cls(args)
real, attack = db.get_train_data()
def merge_data(flist):
d = bob.io.load([k.make_path(use_inputdir[0], '.hdf5') for k in flist])
return d[~numpy.isnan(d.sum(axis=1)), :]
real = merge_data(real)
attack = merge_data(attack)
if args.verbose: print "Evaluating mean and standard deviation..."
from antispoofing.utils.ml.norm import calc_mean_std
mean, std = calc_mean_std(real, attack, nonStdZero=True)
if args.verbose: print "Training LDA..."
from antispoofing.utils.ml.norm import zeromean_unitvar_norm
real = zeromean_unitvar_norm(real, mean, std)
attack = zeromean_unitvar_norm(attack, mean, std)
from antispoofing.utils.ml.lda import make_lda
machine = make_lda([real, attack])
# adjust some details of the final machine to be saved
machine.resize(machine.shape[0], 1)
# so we get real and attacks on the "right" side of the axis
machine.weights = -1 * machine.weights
machine.input_subtract = mean
machine.input_divide = std
if args.verbose: print "Performance evaluation:"
real, attack = db.get_devel_data()
real = merge_data(real)
attack = merge_data(attack)
pos = machine(real)[:, 0]
neg = machine(attack)[:, 0]
thres = bob.measure.eer_threshold(neg, pos)
far, frr = bob.measure.farfrr(neg, pos, thres)
good_neg = bob.measure.correctly_classified_negatives(neg, thres).sum()
good_pos = bob.measure.correctly_classified_positives(pos, thres).sum()
print " -> EER @ devel set threshold: %.5e" % thres
print " -> Devel set results:"
print " * FAR : %.3f%% (%d/%d)" % (100 * far, len(neg) - good_neg,
len(neg))
print " * FRR : %.3f%% (%d/%d)" % (100 * frr, len(pos) - good_pos,
len(pos))
print " * HTER: %.3f%%" % (50 * (far + frr))
real, attack = db.get_test_data()
real = merge_data(real)
attack = merge_data(attack)
pos = machine(real)[:, 0]
neg = machine(attack)[:, 0]
far, frr = bob.measure.farfrr(neg, pos, thres)
good_neg = bob.measure.correctly_classified_negatives(neg, thres).sum()
good_pos = bob.measure.correctly_classified_positives(pos, thres).sum()
print " -> Test set results:"
print " * FAR: %.3f%% (%d/%d)" % (100 * far, len(neg) - good_neg,
len(neg))
print " * FRR: %.3f%% (%d/%d)" % (100 * frr, len(pos) - good_pos,
len(pos))
print " * HTER: %.3f%%" % (50 * (far + frr))
if args.verbose: print "Saving session information..."
def get_version(package):
__import__('pkg_resources').require(package)[0].version
paramfile.add_section('software')
for package in __import__('pkg_resources').require('antispoofing.motion'):
paramfile.set('software', package.key, package.version)
paramfile.add_section('environment')
cmdline = [os.path.realpath(sys.argv[0])] + sys.argv[1:]
paramfile.set('environment', 'command-line', ' '.join(cmdline))
paramfile.add_section('data')
datapath = [os.path.realpath(k) for k in use_inputdir]
paramfile.set('data', 'database', args.name)
paramfile.set('data', 'input', '\n'.join(datapath))
paramfile.set('data', 'train-real', str(len(real)))
paramfile.set('data', 'train-attack', str(len(attack)))
paramfile.add_section('lda')
paramfile.set('lda', 'shape', '-'.join([str(k) for k in machine.shape]))
if args.verbose: print "Saving LDA machine..."
machfile = bob.io.HDF5File(os.path.join(use_outputdir, 'lda.hdf5'), 'w')
machine.save(machfile)
del machfile
paramfile.set('time', 'end', time.asctime())
total_time = int(time.time() - start_time)
diff = datetime.timedelta(seconds=total_time)
paramfile.set('time', 'duration', str(diff))
paramfile.set('time', 'host', socket.getfqdn())
if args.verbose: print "End time is", time.asctime()
if args.verbose: print "Total training time:", str(diff)
paramfile.write(open(os.path.join(use_outputdir, 'session.txt'), 'wb'))
if args.verbose: print "All done, bye!"
