def get_parser(self):
"""
Get parser.
:return: parser
:rtype: argparse.ArgumentParser
"""
parser = argparse.ArgumentParser(description='Attack classifier.')
parser.add_argument('-test_images_file', default=paths.test_images_file(), help='HDF5 file containing images.', type=str)
parser.add_argument('-test_codes_file', default=paths.test_codes_file(), help='HDF5 file containing codes.', type=str)
parser.add_argument('-label_index', default=2, help='Label index.', type=int)
parser.add_argument('-classifier_file', default=paths.state_file('classifier'), help='Snapshot state file of classifier.', type=str)
parser.add_argument('-perturbations_file', default=paths.results_file('classifier/perturbations'), help='HDF5 file containing perturbations.', type=str)
parser.add_argument('-original_success_file', default=paths.results_file('classifier/success'), help='HDF5 file containing success.', type=str)
parser.add_argument('-transfer_success_file', default=paths.results_file('classifier/transfer_success', help='HDF5 file containing transfer success.'), type=str)
parser.add_argument('-original_accuracy_file', default=paths.results_file('classifier/accuracy'), help='HDF5 file containing accuracy.', type=str)
parser.add_argument('-transfer_accuracy_file', default=paths.results_file('classifier/transfer_accuracy', help='HDF5 file containing transfer accuracy.'), type=str)
parser.add_argument('-log_file', default=paths.log_file('classifier/attacks'), help='Log file.', type=str)
parser.add_argument('-batch_size', default=128, help='Batch size of attack.', type=int)
parser.add_argument('-no_gpu', dest='use_gpu', action='store_false')
# Some network parameters.
parser.add_argument('-network_architecture', default='standard', help='Classifier architecture to use.', type=str)
parser.add_argument('-network_activation', default='relu', help='Activation function to use.', type=str)
parser.add_argument('-network_no_batch_normalization', default=False, help='Do not use batch normalization.', action='store_true')
parser.add_argument('-network_channels', default=16, help='Channels of first convolutional layer, afterwards channels are doubled.', type=int)
parser.add_argument('-network_dropout', default=False, action='store_true', help='Whether to use dropout.')
parser.add_argument('-network_units', default='1024,1024,1024,1024', help='Units for MLP.')
return parser
def get_parser(self):
"""
Get parser.
:return: parser
:rtype: argparse.ArgumentParser
"""
parser = argparse.ArgumentParser(description='Test attacks on classifier.')
parser.add_argument('-test_images_file', default=paths.test_images_file(), help='HDF5 file containing dataset.', type=str)
parser.add_argument('-train_images_file', default=paths.train_images_file(), help='HDF5 file containing dataset.', type=str)
parser.add_argument('-test_codes_file', default=paths.test_codes_file(), help='HDF5 file containing labels.', type=str)
parser.add_argument('-label_index', default=2, help='Column index in label file.', type=int)
parser.add_argument('-accuracy_file', default=paths.results_file('learned_decoder/accuracy'), help='Correctly classified test samples of classifier.', type=str)
parser.add_argument('-perturbations_file', default=paths.results_file('learned_decoder/perturbations'), help='HDF5 file containing perturbations.', type=str)
parser.add_argument('-success_file', default=paths.results_file('learned_decoder/success'), help='HDF5 file indicating attack success.', type=str)
parser.add_argument('-plot_directory', default=paths.experiment_dir('learned_decoder'), help='Path to PNG plot file for success rate.', type=str)
parser.add_argument('-results_file', default='', help='Path to pickled results file.', type=str)
parser.add_argument('-batch_size', default=128, help='Batch size of attack.', type=int)
parser.add_argument('-plot_manifolds', default=False, action='store_true', help='Whether to plot manifolds.')
parser.add_argument('-no_gpu', dest='use_gpu', action='store_false')
# Network.
parser.add_argument('-N_theta', default=6, help='Numer of transformations.', type=int)
return parser
def get_parser(self):
"""
Get parser.
:return: parser
:rtype: argparse.ArgumentParser
"""
parser = argparse.ArgumentParser(
description='Attack decoder and classifier.')
parser.add_argument('-test_images_file',
default=paths.test_images_file(),
help='HDF5 file with test images.',
type=str)
parser.add_argument('-test_codes_file',
default=paths.test_codes_file(),
help='HDF5 file containing test codes.',
type=str)
parser.add_argument('-label_index',
default=2,
help='Label index.',
type=int)
parser.add_argument(
'-perturbations_file',
default=paths.results_file('decoder/perturbations'),
help='HDF5 file containing perturbations.',
type=str)
parser.add_argument(
'-perturbation_images_file',
default=paths.results_file('decoder/perturbation_images'),
help='HDF5 file for perturbation images.',
type=str)
parser.add_argument('-log_file',
default=paths.log_file('decoder/attacks'),
help='Log file.',
type=str)
parser.add_argument('-batch_size',
default=128,
help='Batch size of attack.',
type=int)
parser.add_argument('-no_gpu', dest='use_gpu', action='store_false')
# Some decoder parameters.
parser.add_argument('-N_theta',
default=6,
help='Numer of transformations.',
type=int)
return parser
def get_parser(self):
"""
Get parser.
:return: parser
:rtype: argparse.ArgumentParser
"""
parser = argparse.ArgumentParser(description='Train auto encoder.')
parser.add_argument('-train_images_file', default=paths.train_images_file(), help='HDF5 file containing dataset.', type=str)
parser.add_argument('-test_images_file', default=paths.test_images_file(), help='HDF5 file containing dataset.', type=str)
parser.add_argument('-train_codes_file', default=paths.train_codes_file(), help='HDF5 file containing dataset.', type=str)
parser.add_argument('-test_codes_file', default=paths.test_codes_file(), help='HDF5 file containing dataset.', type=str)
parser.add_argument('-label_index', default=2, help='Label index.', type=int)
parser.add_argument('-label', default=-1, help='Label to constrain to.', type=int)
parser.add_argument('-encoder_file', default=paths.state_file('encoder'), help='Snapshot state file.', type=str)
parser.add_argument('-decoder_file', default=paths.state_file('decoder'), help='Snapshot state file.', type=str)
parser.add_argument('-reconstruction_file', default=paths.results_file('reconstructions'), help='Reconstructions file.', type=str)
parser.add_argument('-interpolation_file', default=paths.results_file('interpolations'), help='Interpolation file.', type=str)
parser.add_argument('-random_file', default=paths.results_file('random'), help='Reconstructions file.', type=str)
parser.add_argument('-log_file', default=paths.log_file('auto_encoder'), help='Log file.', type=str)
parser.add_argument('-batch_size', default=64, help='Batch size.', type=int)
parser.add_argument('-latent_space_size', default=10, help='Size of latent space.', type=int)
parser.add_argument('-epochs', default=20, help='Number of epochs.', type=int)
parser.add_argument('-no_gpu', dest='use_gpu', action='store_false')
parser.add_argument('-base_lr', default=0.01, type=float, help='Base learning rate.')
parser.add_argument('-base_lr_decay', default=0.9, type=float, help='Base learning rate.')
parser.add_argument('-results_file', default='', help='Results file for evaluation.', type=str)
parser.add_argument('-training_file', default=paths.results_file('auto_encoder_training'), help='Training statistics file.', type=str)
parser.add_argument('-testing_file', default=paths.results_file('auto_encoder_testing'), help='Testing statistics file.', type=str)
parser.add_argument('-error_file', default=paths.image_file('auto_encoder_error'), help='Error plot file.', type=str)
parser.add_argument('-beta', default=1, help='Weight of KLD.', type=float)
parser.add_argument('-weight_decay', default=0.0001, help='Weight decay importance.', type=float)
parser.add_argument('-absolute_error', default=False, action='store_true', help='Use absolute loss.')
# Some network parameters.
parser.add_argument('-network_architecture', default='standard', help='Classifier architecture to use.', type=str)
parser.add_argument('-network_activation', default='relu', help='Activation function to use.', type=str)
parser.add_argument('-network_no_batch_normalization', default=False, help='Do not use batch normalization.', action='store_true')
parser.add_argument('-network_channels', default=16, help='Channels of first convolutional layer, afterwards channels are doubled.', type=int)
parser.add_argument('-network_dropout', default=False, action='store_true', help='Whether to use dropout.')
parser.add_argument('-network_units', default='1024,1024,1024,1024', help='Units for MLP.')
return parser
def get_parser(self):
"""
Get parser.
:return: parser
:rtype: argparse.ArgumentParser
"""
parser = argparse.ArgumentParser(
description='Attack decoder and classifier.')
parser.add_argument('-database_file',
default=paths.database_file(),
help='HDF5 file containing font prototype images.',
type=str)
parser.add_argument('-test_theta_file',
default=paths.test_theta_file(),
help='HDF5 file for thetas.',
type=str)
parser.add_argument('-test_codes_file',
default=paths.test_codes_file(),
help='HDF5 file containing codes.',
type=str)
parser.add_argument(
'-perturbations_file',
default=paths.results_file('decoder/perturbations'),
help='HDF5 file containing perturbations.',
type=str)
parser.add_argument(
'-perturbation_images_file',
default=paths.results_file('decoder/perturbation_images'),
help='HDF5 file for perturbation images.',
type=str)
parser.add_argument('-log_file',
default=paths.log_file('decoder/attacks'),
help='Log file.',
type=str)
parser.add_argument('-batch_size',
default=128,
help='Batch size of attack.',
type=int)
parser.add_argument('-no_gpu', dest='use_gpu', action='store_false')
return parser
def get_parser(self):
"""
Get parser.
:return: parser
:rtype: argparse.ArgumentParser
"""
parser = argparse.ArgumentParser(description='Attack decoder and classifier.')
parser.add_argument('-database_file', default=paths.database_file(), help='HDF5 file containing font prototype images.', type=str)
parser.add_argument('-test_codes_file', default=paths.test_codes_file(), help='HDF5 file containing codes.', type=str)
parser.add_argument('-test_theta_file', default=paths.test_theta_file(), help='HDF5 file containing transformations.', type=str)
parser.add_argument('-classifier_file', default=paths.state_file('classifier'), help='Snapshot state file of classifier.', type=str)
parser.add_argument('-accuracy_file', default=paths.results_file('decoder/accuracy'), help='Correctly classified test samples of classifier.', type=str)
parser.add_argument('-perturbations_file', default=paths.results_file('decoder/perturbations'), help='HDF5 file containing perturbations.', type=str)
parser.add_argument('-success_file', default=paths.results_file('decoder/success'), help='HDF5 file containing perturbations.', type=str)
parser.add_argument('-log_file', default=paths.log_file('decoder/attacks'), help='Log file.', type=str)
parser.add_argument('-attack', default='UntargetedBatchL2ClippedGradientDescent', help='Attack to try.', type=str)
parser.add_argument('-objective', default='UntargetedF6', help='Objective to use.', type=str)
parser.add_argument('-max_attempts', default=1, help='Maximum number of attempts per attack.', type=int)
parser.add_argument('-max_samples', default=20*128, help='How many samples from the test set to attack.', type=int)
parser.add_argument('-batch_size', default=128, help='Batch size of attack.', type=int)
parser.add_argument('-epsilon', default=0.1, help='Epsilon allowed for attacks.', type=float)
parser.add_argument('-c_0', default=0., help='Weight of norm.', type=float)
parser.add_argument('-c_1', default=0.1, help='Weight of bound, if not enforced through clipping or reparameterization.', type=float)
parser.add_argument('-c_2', default=0.5, help='Weight of objective.', type=float)
parser.add_argument('-max_iterations', default=100, help='Number of iterations for attack.', type=int)
parser.add_argument('-max_projections', default=5, help='Number of projections for alternating projection.', type=int)
parser.add_argument('-base_lr', default=0.005, help='Learning rate for attack.', type=float)
parser.add_argument('-no_gpu', dest='use_gpu', action='store_false')
parser.add_argument('-no_label_leaking', default=False, dest='no_label_leaking', action='store_true')
parser.add_argument('-on_manifold', default=False, dest='on_manifold', action='store_true')
parser.add_argument('-initialize_zero', default=False, action='store_true', help='Initialize attack at zero.')
# Some network parameters.
parser.add_argument('-network_architecture', default='standard', help='Classifier architecture to use.', type=str)
parser.add_argument('-network_activation', default='relu', help='Activation function to use.', type=str)
parser.add_argument('-network_no_batch_normalization', default=False, help='Do not use batch normalization.', action='store_true')
parser.add_argument('-network_channels', default=16, help='Channels of first convolutional layer, afterwards channels are doubled.', type=int)
parser.add_argument('-network_dropout', default=False, action='store_true', help='Whether to use dropout.')
parser.add_argument('-network_units', default='1024,1024,1024,1024', help='Units for MLP.')
return parser
def get_parser(self):
"""
Get parser.
:return: parser
:rtype: argparse.ArgumentParser
"""
parser = argparse.ArgumentParser(description='Detect attacks on classifier.')
parser.add_argument('-mode', default='svd', help='Mode.', type=str)
parser.add_argument('-database_file', default=paths.database_file(), help='HDF5 file containing font prototype images.', type=str)
parser.add_argument('-test_codes_file', default=paths.test_codes_file(), help='HDF5 file codes dataset.', type=str)
parser.add_argument('-test_theta_file', default=paths.test_theta_file(), help='HDF5 file containing transformations.', type=str)
parser.add_argument('-train_images_file', default=paths.train_images_file(), help='HDF5 file containing dataset.', type=str)
parser.add_argument('-test_images_file', default=paths.test_images_file(), help='HDF5 file containing dataset.', type=str)
parser.add_argument('-perturbations_file', default=paths.results_file('classifier/perturbations'), help='HDF5 file containing perturbations.', type=str)
parser.add_argument('-success_file', default=paths.results_file('classifier/success'), help='HDF5 file containing success indicators.', type=str)
parser.add_argument('-accuracy_file', default=paths.results_file('classifier/accuracy'), help='HDF5 file containing accuracy indicators.', type=str)
parser.add_argument('-batch_size', default=64, help='Batch size.', type=int)
parser.add_argument('-no_gpu', dest='use_gpu', action='store_false')
parser.add_argument('-pre_pca', default=20, help='PCA dimensionality reduction ebfore NN.', type=int)
parser.add_argument('-n_nearest_neighbors', default=50, help='Number of NNs to consider.', type=int)
parser.add_argument('-n_pca', default=10, help='Number of NNs to consider.', type=int)
parser.add_argument('-n_fit', default=100000, help='Training images to fit.', type=int)
parser.add_argument('-plot_directory', default=paths.experiment_dir('classifier/detection'), help='Plot directory.', type=str)
parser.add_argument('-max_samples', default=1000, help='Number of samples.', type=int)
# Some decoder parameters.
parser.add_argument('-decoder_files', default=paths.state_file('decoder'), help='Decoder files.', type=str)
parser.add_argument('-latent_space_size', default=10, help='Size of latent space.', type=int)
parser.add_argument('-decoder_architecture', default='standard', help='Architecture to use.', type=str)
parser.add_argument('-decoder_activation', default='relu', help='Activation function to use.', type=str)
parser.add_argument('-decoder_no_batch_normalization', default=False, help='Do not use batch normalization.', action='store_true')
parser.add_argument('-decoder_channels', default=16, help='Channels of first convolutional layer, afterwards channels are doubled.', type=int)
parser.add_argument('-decoder_dropout', default=False, action='store_true', help='Whether to use dropout.')
parser.add_argument('-decoder_units', default='1024,1024,1024,1024', help='Units for MLP.')
return parser
def get_parser(self):
"""
Get parser.
:return: parser
:rtype: argparse.ArgumentParser
"""
parser = argparse.ArgumentParser(description='Test attacks on classifier.')
parser.add_argument('-test_images_file', default=paths.test_images_file(), help='HDF5 file containing dataset.', type=str)
parser.add_argument('-train_images_file', default=paths.train_images_file(), help='HDF5 file containing dataset.', type=str)
parser.add_argument('-test_theta_file', default=paths.results_file('test_theta'), help='HDF5 file containing dataset.', type=str)
parser.add_argument('-train_theta_file', default=paths.results_file('train_theta'), help='HDF5 file containing dataset.', type=str)
parser.add_argument('-test_codes_file', default=paths.test_codes_file(), help='HDF5 file containing labels.', type=str)
parser.add_argument('-decoder_files', default=paths.state_file('decoder'), help='Decoder files.', type=str)
parser.add_argument('-label_index', default=2, help='Column index in label file.', type=int)
parser.add_argument('-accuracy_file', default=paths.results_file('learned_decoder/accuracy'), help='Correctly classified test samples of classifier.', type=str)
parser.add_argument('-perturbations_file', default=paths.results_file('learned_decoder/perturbations'), help='HDF5 file containing perturbations.', type=str)
parser.add_argument('-success_file', default=paths.results_file('learned_decoder/success'), help='HDF5 file indicating attack success.', type=str)
parser.add_argument('-plot_directory', default=paths.experiment_dir('learned_decoder'), help='Path to PNG plot file for success rate.', type=str)
parser.add_argument('-results_file', default='', help='Path to pickled results file.', type=str)
parser.add_argument('-batch_size', default=128, help='Batch size of attack.', type=int)
parser.add_argument('-plot_manifolds', default=False, action='store_true', help='Whether to plot manifolds.')
parser.add_argument('-no_gpu', dest='use_gpu', action='store_false')
parser.add_argument('-bound', default=2, help='Bound to consider for samples in latent space.', type=float)
# Some decoder parameters.
parser.add_argument('-latent_space_size', default=10, help='Size of latent space.', type=int)
parser.add_argument('-decoder_architecture', default='standard', help='Architecture to use.', type=str)
parser.add_argument('-decoder_activation', default='relu', help='Activation function to use.', type=str)
parser.add_argument('-decoder_no_batch_normalization', default=False, help='Do not use batch normalization.', action='store_true')
parser.add_argument('-decoder_channels', default=16, help='Channels of first convolutional layer, afterwards channels are doubled.', type=int)
parser.add_argument('-decoder_dropout', default=False, action='store_true', help='Whether to use dropout.')
parser.add_argument('-decoder_units', default='1024,1024,1024,1024', help='Units for MLP.')
return parser
def get_parser(self):
"""
Get parser.
:return: parser
:rtype: argparse.ArgumentParser
"""
parser = argparse.ArgumentParser(
description='Visualize images as mosaic.')
parser.add_argument('-images_file',
default=paths.results_file('reconstructions'),
help='HDF5 file containing images to visualize',
type=str)
parser.add_argument('-codes_file',
default='',
help='HDF5 label file.',
type=str)
parser.add_argument('-label_index',
default=2,
help='Label index.',
type=int)
parser.add_argument('-label', default=-1, help='Label.', type=int)
parser.add_argument('-cols',
default=8,
help='Number of columns.',
type=int)
parser.add_argument('-rows',
default=8,
help='Number of rows.',
type=int)
parser.add_argument('-vmin',
default=None,
help='Minimum value vor visualization.',
type=float)
parser.add_argument('-vmax',
default=None,
help='Maximum value for visualization.',
type=float)
parser.add_argument('-max_images',
default=10,
type=int,
help='Number of images to output.')
parser.add_argument('-output_directory',
default=paths.experiment_file('', ''),
help='Output directory.',
type=str)
return parser
def get_parser(self):
"""
Get parser.
:return: parser
:rtype: argparse.ArgumentParser
"""
parser = argparse.ArgumentParser(description='Train classifier.')
parser.add_argument('-train_images_file',
default=paths.train_images_file(),
help='HDF5 file containing dataset.',
type=str)
parser.add_argument('-train_codes_file',
default=paths.train_codes_file(),
help='HDF5 file containing codes.',
type=str)
parser.add_argument('-test_images_file',
default=paths.test_images_file(),
help='HDF5 file containing dataset.',
type=str)
parser.add_argument('-test_codes_file',
default=paths.test_codes_file(),
help='HDF5 file containing codes.',
type=str)
parser.add_argument('-state_file',
default=paths.state_file('classifier'),
help='Snapshot state file.',
type=str)
parser.add_argument('-log_file',
default=paths.log_file('classifier'),
help='Log file.',
type=str)
parser.add_argument('-training_file',
default=paths.results_file('training'),
help='Training statistics file.',
type=str)
parser.add_argument('-testing_file',
default=paths.results_file('testing'),
help='Testing statistics file.',
type=str)
parser.add_argument('-loss_file',
default=paths.image_file('loss'),
help='Loss plot file.',
type=str)
parser.add_argument('-error_file',
default=paths.image_file('error'),
help='Error plot file.',
type=str)
parser.add_argument('-gradient_file',
default=paths.image_file('gradient'),
help='Gradient plot file.',
type=str)
parser.add_argument('-label_index',
default=2,
help='Label index.',
type=int)
parser.add_argument('-training_samples',
default=-1,
help='Number of samples used for training.',
type=int)
parser.add_argument('-validation_samples',
default=0,
help='Number of samples for validation.',
type=int)
parser.add_argument('-test_samples',
default=-1,
help='Number of samples for validation.',
type=int)
parser.add_argument('-early_stopping',
default=False,
action='store_true',
help='Use early stopping.')
parser.add_argument(
'-random_samples',
default=False,
action='store_true',
help='Randomize the subsampling of the training set.')
parser.add_argument('-batch_size',
default=64,
help='Batch size.',
type=int)
parser.add_argument('-epochs',
default=10,
help='Number of epochs.',
type=int)
parser.add_argument('-weight_decay',
default=0.0001,
help='Weight decay importance.',
type=float)
parser.add_argument('-logit_decay',
default=0,
help='Logit decay importance.',
type=float)
parser.add_argument('-no_gpu', dest='use_gpu', action='store_false')
parser.add_argument('-skip',
default=5,
help='Verbosity in iterations.',
type=int)
parser.add_argument('-lr',
default=0.01,
type=float,
help='Base learning rate.')
parser.add_argument('-lr_decay',
default=0.9,
type=float,
help='Learning rate decay.')
parser.add_argument('-results_file',
default='',
help='Results file for evaluation.',
type=str)
parser.add_argument('-debug_directory',
default='',
help='Debug directory.',
type=str)
# Some network parameters.
parser.add_argument('-network_architecture',
default='standard',
help='Classifier architecture to use.',
type=str)
parser.add_argument('-network_activation',
default='relu',
help='Activation function to use.',
type=str)
parser.add_argument('-network_no_batch_normalization',
default=False,
help='Do not use batch normalization.',
action='store_true')
parser.add_argument(
'-network_channels',
default=16,
help=
'Channels of first convolutional layer, afterwards channels are doubled.',
type=int)
parser.add_argument('-network_dropout',
default=False,
action='store_true',
help='Whether to use dropout.')
parser.add_argument('-network_units',
default='1024,1024,1024,1024',
help='Units for MLP.')
return parser
def get_parser(self):
"""
Get parser.
:return: parser
:rtype: argparse.ArgumentParser
"""
parser = argparse.ArgumentParser(description='Test classifier.')
parser.add_argument('-test_images_file',
default=paths.test_images_file(),
help='HDF5 file containing dataset.',
type=str)
parser.add_argument('-test_codes_file',
default=paths.test_codes_file(),
help='HDF5 file containing codes.',
type=str)
parser.add_argument('-state_file',
default=paths.state_file('classifier'),
help='Snapshot state file.',
type=str)
parser.add_argument(
'-accuracy_file',
default=paths.results_file('classifier/accuracy'),
help='Correctly classified test samples of classifier.',
type=str)
parser.add_argument('-label_index',
default=2,
help='Label index.',
type=int)
parser.add_argument('-results_file',
default='',
help='Results file for evaluation.',
type=str)
parser.add_argument('-batch_size',
default=64,
help='Batch size.',
type=int)
parser.add_argument('-no_gpu', dest='use_gpu', action='store_false')
parser.add_argument('-log_file',
default=paths.log_file('test_classifier'),
help='Log file.',
type=str)
# Some network parameters.
parser.add_argument('-network_architecture',
default='standard',
help='Classifier architecture to use.',
type=str)
parser.add_argument('-network_activation',
default='relu',
help='Activation function to use.',
type=str)
parser.add_argument('-network_no_batch_normalization',
default=False,
help='Do not use batch normalization.',
action='store_true')
parser.add_argument(
'-network_channels',
default=16,
help=
'Channels of first convolutional layer, afterwards channels are doubled.',
type=int)
parser.add_argument('-network_dropout',
default=False,
action='store_true',
help='Whether to use dropout.')
parser.add_argument('-network_units',
default='1024,1024,1024,1024',
help='Units for MLP.')
return parser
def get_parser(self):
"""
Get parser.
:return: parser
:rtype: argparse.ArgumentParser
"""
parser = argparse.ArgumentParser(description='Test auto encoder.')
parser.add_argument('-train_images_file',
default=paths.train_images_file(),
help='HDF5 file containing dataset.',
type=str)
parser.add_argument('-test_images_file',
default=paths.test_images_file(),
help='HDF5 file containing dataset.',
type=str)
parser.add_argument('-train_codes_file',
default=paths.train_codes_file(),
help='HDF5 file containing dataset.',
type=str)
parser.add_argument('-test_codes_file',
default=paths.test_codes_file(),
help='HDF5 file containing dataset.',
type=str)
parser.add_argument('-train_theta_file',
default=paths.results_file('train_theta'),
help='HDF5 file for codes.',
type=str)
parser.add_argument('-test_theta_file',
default=paths.results_file('test_theta'),
help='HDF5 file for codes.',
type=str)
parser.add_argument('-label_index',
default=2,
help='Label index.',
type=int)
parser.add_argument('-label',
default=-1,
help='Label to constrain to.',
type=int)
parser.add_argument('-encoder_file',
default=paths.state_file('encoder'),
help='Snapshot state file.',
type=str)
parser.add_argument('-decoder_file',
default=paths.state_file('decoder'),
help='Snapshot state file.',
type=str)
parser.add_argument('-reconstruction_file',
default=paths.results_file('reconstructions'),
help='Reconstructions file.',
type=str)
parser.add_argument('-train_reconstruction_file',
default='',
help='Reconstructions file.',
type=str)
parser.add_argument('-random_file',
default=paths.results_file('random'),
help='Reconstructions file.',
type=str)
parser.add_argument('-interpolation_file',
default=paths.results_file('interpolation'),
help='Interpolations file.',
type=str)
parser.add_argument('-batch_size',
default=64,
help='Batch size.',
type=int)
parser.add_argument('-latent_space_size',
default=10,
help='Size of latent space.',
type=int)
parser.add_argument('-no_gpu', dest='use_gpu', action='store_false')
parser.add_argument('-results_file',
default='',
help='Results file for evaluation.',
type=str)
parser.add_argument('-output_directory',
default='',
help='Output directory for plots.',
type=str)
parser.add_argument('-log_file',
default=paths.log_file('test_auto_encoder'),
help='Log file.',
type=str)
# Some network parameters.
parser.add_argument('-network_architecture',
default='standard',
help='Architecture type.')
parser.add_argument('-network_activation',
default='relu',
help='Activation function to use.',
type=str)
parser.add_argument('-network_no_batch_normalization',
default=False,
help='Do not use batch normalization.',
action='store_true')
parser.add_argument(
'-network_channels',
default=16,
help=
'Channels of first convolutional layer, afterwards channels are doubled.',
type=int)
parser.add_argument('-network_dropout',
default=False,
action='store_true',
help='Whether to use dropout.')
parser.add_argument('-network_units',
default='1024,1024,1024,1024',
help='Units for MLP.')
return parser
def train(self, base_directory, method, label):
"""
Train VAE-GAN.
"""
encoder_file = paths.state_file('%s/encoder' % base_directory)
decoder_file = paths.state_file('%s/decoder' % base_directory)
classifier_file = paths.state_file('%s/classifier' % base_directory)
reconstruction_file = paths.results_file('%s/reconstructions' %
base_directory)
interpolation_file = paths.results_file('%s/interpolation' %
base_directory)
random_file = paths.results_file('%s/random' % base_directory)
log_file = paths.log_file('%s/vaegan' % base_directory)
results_file = paths.pickle_file('%s/results' % base_directory)
training_file = paths.results_file('%s/training' % base_directory)
testing_file = paths.results_file('%s/testing' % base_directory)
error_file = paths.image_file('%s/error' % base_directory)
gradient_file = paths.image_file('%s/gradient' % base_directory)
test_results_file = paths.pickle_file('%s/test_results' %
base_directory)
if not os.path.exists(encoder_file) or not os.path.exists(
decoder_file):
if method.find('vaegan') >= 0:
arguments = [
'-train_images_file=%s' % self.train_images_file,
'-train_codes_file=%s' % self.train_codes_file,
'-test_images_file=%s' % self.test_images_file,
'-test_codes_file=%s' % self.test_codes_file,
'-label_index=%d' % self.label_index,
'-label=%d' % label,
'-encoder_file=%s' % encoder_file,
'-decoder_file=%s' % decoder_file,
'-classifier_file=%s' % classifier_file,
'-reconstruction_file=%s' % reconstruction_file,
'-interpolation_file=%s' % interpolation_file,
'-random_file=%s' % random_file,
'-log_file=%s' % log_file,
'-results_file=%s' % results_file,
'-training_file=%s' % training_file,
'-testing_file=%s' % testing_file,
'-error_file=%s' % error_file,
'-gradient_file=%s' % gradient_file,
'-latent_space_size=%d' % self.args.latent_space_size,
'-beta=%g' % self.args.beta,
'-gamma=%g' % self.args.gamma,
'-epochs=%d' % self.args.epochs,
'-eta=%g' % self.args.eta,
] + self.training_parameters + self.network_parameters + self.classifier_parameters
log(arguments)
if method == 'vaegan2':
train = TrainVAEGAN2(arguments)
else:
raise NotImplementedError()
train.main()
elif method == 'vae':
arguments = [
'-train_images_file=%s' % self.train_images_file,
'-train_codes_file=%s' % self.train_codes_file,
'-test_images_file=%s' % self.test_images_file,
'-test_codes_file=%s' % self.test_codes_file,
'-label_index=%d' % self.label_index,
'-label=%d' % label,
'-encoder_file=%s' % encoder_file,
'-decoder_file=%s' % decoder_file,
'-reconstruction_file=%s' % reconstruction_file,
'-interpolation_file=%s' % interpolation_file,
'-random_file=%s' % random_file,
'-log_file=%s' % log_file,
'-results_file=%s' % results_file,
'-training_file=%s' % training_file,
'-testing_file=%s' % testing_file,
'-error_file=%s' % error_file,
'-latent_space_size=%d' % self.args.latent_space_size,
'-beta=%g' % self.args.beta,
'-epochs=%d' % self.args.epochs,
] + self.training_parameters + self.network_parameters
train = TrainVariationalAutoEncoder(arguments)
train.main()
else:
raise NotImplementedError()
train_theta_file = paths.results_file('%s/train_theta' %
base_directory)
test_theta_file = paths.results_file('%s/test_theta' % base_directory)
if self.args.reevaluate or not os.path.exists(test_results_file):
test = TestVariationalAutoEncoder([
'-train_images_file=%s' % self.train_images_file,
'-test_images_file=%s' % self.test_images_file,
'-train_codes_file=%s' % self.train_codes_file,
'-test_codes_file=%s' % self.test_codes_file,
'-train_theta_file=%s' % train_theta_file,
'-test_theta_file=%s' % test_theta_file,
'-label_index=%d' % self.label_index,
'-label=%d' % label,
'-encoder_file=%s' % encoder_file,
'-decoder_file=%s' % decoder_file,
'-batch_size=256',
'-results_file=%s' % test_results_file,
'-reconstruction_file=',
'-random_file=',
'-interpolation_file=',
'-output_directory=',
'-latent_space_size=%d' % self.args.latent_space_size,
] + self.network_parameters)
test.main()
results = test.results
else:
results = utils.read_pickle(test_results_file)
self.results[base_directory] = (results['reconstruction_error'],
results['code_mean'],
results['code_var'])
if utils.display():
random_directory = paths.experiment_dir(
'%s/random/' % base_directory, experiment=self.experiment())
if self.args.reevaluate or not os.path.exists(random_directory):
visualize_mosaic = VisualizeMosaic([
'-images_file=%s' % random_file,
'-output_directory=%s' % random_directory,
])
visualize_mosaic.main()
random_directory = paths.experiment_dir(
'%s/random_/' % base_directory, experiment=self.experiment())
# if self.args.reevaluate or not os.path.exists(random_directory):
visualize_individual = VisualizeIndividual([
'-images_file=%s' % random_file,
'-output_directory=%s' % random_directory,
])
visualize_individual.main()
if utils.display():
reconstruction_directory = paths.experiment_dir(
'%s/reconstruction/' % base_directory,
experiment=self.experiment())
if self.args.reevaluate or not os.path.exists(
reconstruction_directory):
visualize_mosaic = VisualizeMosaic([
'-images_file=%s' % reconstruction_file,
'-output_directory=%s' % reconstruction_directory,
])
visualize_mosaic.main()
reconstruction_directory = paths.experiment_dir(
'%s/reconstruction_/' % base_directory,
experiment=self.experiment())
# if self.args.reevaluate or not os.path.exists(reconstruction_directory):
visualize_individual = VisualizeIndividual([
'-images_file=%s' % reconstruction_file,
'-output_directory=%s' % reconstruction_directory,
])
visualize_individual.main()
if utils.display():
interpolation_directory = paths.experiment_dir(
'%s/interpolation/' % base_directory,
experiment=self.experiment())
if self.args.reevaluate or not os.path.exists(
interpolation_directory):
visualize_mosaic = VisualizeMosaic([
'-images_file=%s' % interpolation_file,
'-output_directory=%s' % interpolation_directory,
])
visualize_mosaic.main()
interpolation_directory = paths.experiment_dir(
'%s/interpolation_/' % base_directory,
experiment=self.experiment())
# if self.args.reevaluate or not os.path.exists(reconstruction_directory):
visualize_individual = VisualizeIndividual([
'-images_file=%s' % interpolation_file,
'-output_directory=%s' % interpolation_directory,
])
visualize_individual.main()
if utils.display():
original_directory = paths.experiment_dir(
'%s/original/' % base_directory, experiment=self.experiment())
if self.args.reevaluate or not os.path.exists(original_directory):
visualize_mosaic = VisualizeMosaic([
'-images_file=%s' % self.test_images_file,
'-codes_file=%s' % self.test_codes_file,
'-label_index=%d' % self.label_index,
'-label=%d' % label,
'-output_directory=%s' % original_directory,
])
visualize_mosaic.main()
original_directory = paths.experiment_dir(
'%s/original_/' % base_directory, experiment=self.experiment())
# if self.args.reevaluate or not os.path.exists(reconstruction_directory):
visualize_individual = VisualizeIndividual([
'-images_file=%s' % self.test_images_file,
'-codes_file=%s' % self.test_codes_file,
'-label_index=%d' % self.label_index,
'-label=%d' % label,
'-output_directory=%s' % original_directory,
])
visualize_individual.main()
def get_parser(self):
"""
Get parser.
:return: parser
:rtype: argparse.ArgumentParser
"""
parser = argparse.ArgumentParser(description='Train classifier.')
parser.add_argument('-train_images_file',
default=paths.train_images_file(),
help='HDF5 file containing dataset.',
type=str)
parser.add_argument('-train_codes_file',
default=paths.train_codes_file(),
help='HDF5 file containing codes.',
type=str)
parser.add_argument('-train_theta_file',
default=paths.results_file('train_theta'),
help='HDF5 file containing transformations.',
type=str)
parser.add_argument('-test_theta_file',
default=paths.results_file('test_theta'),
help='HDF5 file containing transformations.',
type=str)
parser.add_argument('-test_images_file',
default=paths.test_images_file(),
help='HDF5 file containing dataset.',
type=str)
parser.add_argument('-test_codes_file',
default=paths.test_codes_file(),
help='HDF5 file containing codes.',
type=str)
parser.add_argument('-decoder_files',
default=paths.state_file('decoder'),
help='Decoder files.',
type=str)
parser.add_argument('-latent_space_size',
default=10,
help='Size of latent space.',
type=int)
parser.add_argument('-label_index',
default=2,
help='Column index in label file.',
type=int)
parser.add_argument(
'-state_file',
default=paths.state_file('robust_manifold_classifier'),
help='Snapshot state file.',
type=str)
parser.add_argument(
'-log_file',
default=paths.log_file('robust_manifold_classifier'),
help='Log file.',
type=str)
parser.add_argument(
'-training_file',
default=paths.results_file('robust_manifold_training'),
help='Training statistics file.',
type=str)
parser.add_argument(
'-testing_file',
default=paths.results_file('robust_manifold_testing'),
help='Testing statistics file.',
type=str)
parser.add_argument('-loss_file',
default=paths.image_file('loss'),
help='Loss plot file.',
type=str)
parser.add_argument('-error_file',
default=paths.image_file('error'),
help='Error plot file.',
type=str)
parser.add_argument('-success_file',
default=paths.image_file('robust_success'),
help='Success rate plot file.',
type=str)
parser.add_argument('-gradient_file',
default='',
help='Gradient plot file.',
type=str)
parser.add_argument(
'-random_samples',
default=False,
action='store_true',
help='Randomize the subsampling of the training set.')
parser.add_argument('-training_samples',
default=-1,
help='Number of samples used for training.',
type=int)
parser.add_argument('-test_samples',
default=-1,
help='Number of samples for validation.',
type=int)
parser.add_argument('-validation_samples',
default=0,
help='Number of samples for validation.',
type=int)
parser.add_argument('-early_stopping',
default=False,
action='store_true',
help='Use early stopping.')
parser.add_argument('-attack_samples',
default=1000,
help='Samples to attack.',
type=int)
parser.add_argument('-batch_size',
default=64,
help='Batch size.',
type=int)
parser.add_argument('-epochs',
default=10,
help='Number of epochs.',
type=int)
parser.add_argument('-weight_decay',
default=0.0001,
help='Weight decay importance.',
type=float)
parser.add_argument('-logit_decay',
default=0,
help='Logit decay importance.',
type=float)
parser.add_argument('-no_gpu', dest='use_gpu', action='store_false')
parser.add_argument('-skip',
default=5,
help='Verbosity in iterations.',
type=int)
parser.add_argument('-lr',
default=0.005,
type=float,
help='Base learning rate.')
parser.add_argument('-lr_decay',
default=0.9,
type=float,
help='Learning rate decay.')
parser.add_argument('-results_file',
default='',
help='Results file for evaluation.',
type=str)
parser.add_argument(
'-bound',
default=2,
help=
'Bound used to define "safe" latent codes to compute adversarial examples on.',
type=float)
parser.add_argument('-debug_directory',
default='',
help='Debug directory.',
type=str)
# Some network parameters.
parser.add_argument('-network_architecture',
default='standard',
help='Classifier architecture to use.',
type=str)
parser.add_argument('-network_activation',
default='relu',
help='Activation function to use.',
type=str)
parser.add_argument('-network_no_batch_normalization',
default=False,
help='Do not use batch normalization.',
action='store_true')
parser.add_argument(
'-network_channels',
default=16,
help=
'Channels of first convolutional layer, afterwards channels are doubled.',
type=int)
parser.add_argument('-network_dropout',
default=False,
action='store_true',
help='Whether to use dropout.')
parser.add_argument('-network_units',
default='1024,1024,1024,1024',
help='Units for MLP.')
# Decoder parameters.
parser.add_argument('-decoder_architecture',
default='standard',
help='Architecture to use.',
type=str)
parser.add_argument('-decoder_activation',
default='relu',
help='Activation function to use.',
type=str)
parser.add_argument('-decoder_no_batch_normalization',
default=False,
help='Do not use batch normalization.',
action='store_true')
parser.add_argument(
'-decoder_channels',
default=16,
help=
'Channels of first convolutional layer, afterwards channels are doubled.',
type=int)
parser.add_argument('-decoder_dropout',
default=False,
action='store_true',
help='Whether to use dropout.')
parser.add_argument('-decoder_units',
default='1024,1024,1024,1024',
help='Units for MLP.')
# Attack parameters.
parser.add_argument('-attack',
default='UntargetedBatchL2ClippedGradientDescent',
help='Attack to try.',
type=str)
parser.add_argument('-objective',
default='UntargetedF6',
help='Objective to use.',
type=str)
parser.add_argument('-epsilon',
default=1,
help='Epsilon allowed for attacks.',
type=float)
parser.add_argument('-c_0',
default=0.,
help='Weight of norm.',
type=float)
parser.add_argument(
'-c_1',
default=0.1,
help=
'Weight of bound, if not enforced through clipping or reparameterization.',
type=float)
parser.add_argument('-c_2',
default=0.5,
help='Weight of objective.',
type=float)
parser.add_argument('-max_iterations',
default=10,
help='Number of iterations for attack.',
type=int)
parser.add_argument(
'-max_projections',
default=5,
help='Number of projections for alternating projection.',
type=int)
parser.add_argument('-base_lr',
default=0.005,
help='Learning rate for attack.',
type=float)
parser.add_argument('-decoder_epsilon',
default=1,
help='Epsilon allowed for attacks.',
type=float)
parser.add_argument('-decoder_c_0',
default=0.,
help='Weight of norm.',
type=float)
parser.add_argument(
'-decoder_c_1',
default=0.1,
help=
'Weight of bound, if not enforced through clipping or reparameterization.',
type=float)
parser.add_argument('-decoder_c_2',
default=0.5,
help='Weight of objective.',
type=float)
parser.add_argument('-decoder_max_iterations',
default=10,
help='Number of iterations for attack.',
type=int)
parser.add_argument(
'-decoder_max_projections',
default=5,
help='Number of projections for alternating projection.',
type=int)
parser.add_argument('-decoder_base_lr',
default=0.005,
help='Learning rate for attack.',
type=float)
parser.add_argument('-verbose',
action='store_true',
default=False,
help='Verbose attacks.')
parser.add_argument('-anneal_epochs',
default=0,
help='Anneal iterations in the first epochs.',
type=int)
# Variants.
parser.add_argument('-full_variant',
default=False,
action='store_true',
help='100% variant.')
parser.add_argument('-safe',
default=False,
action='store_true',
help='Save variant.')
parser.add_argument('-training_mode',
default=False,
action='store_true',
help='Training mode variant for attack.')
return parser
def get_parser(self):
"""
Get parser.
:return: parser
:rtype: argparse.ArgumentParser
"""
parser = argparse.ArgumentParser(
description='Attack decoder and classifier.')
parser.add_argument('-test_images_file',
default=paths.test_images_file(),
help='HDF5 file with test images.',
type=str)
parser.add_argument('-test_codes_file',
default=paths.test_codes_file(),
help='HDF5 file containing test codes.',
type=str)
parser.add_argument('-label_index',
default=2,
help='Label index.',
type=int)
parser.add_argument('-decoder_files',
default=paths.state_file('decoder'),
help='Decoder state file.',
type=str)
parser.add_argument(
'-perturbations_file',
default=paths.results_file('decoder/perturbations'),
help='HDF5 file containing perturbations.',
type=str)
parser.add_argument(
'-perturbation_images_file',
default=paths.results_file('decoder/perturbation_images'),
help='HDF5 file for perturbation images.',
type=str)
parser.add_argument('-log_file',
default=paths.log_file('decoder/attacks'),
help='Log file.',
type=str)
parser.add_argument('-batch_size',
default=128,
help='Batch size of attack.',
type=int)
parser.add_argument('-no_gpu', dest='use_gpu', action='store_false')
# Some decoder parameters.
parser.add_argument('-latent_space_size',
default=10,
help='Size of latent space.',
type=int)
parser.add_argument('-decoder_architecture',
default='standard',
help='Architecture to use.',
type=str)
parser.add_argument('-decoder_activation',
default='relu',
help='Activation function to use.',
type=str)
parser.add_argument('-decoder_no_batch_normalization',
default=False,
help='Do not use batch normalization.',
action='store_true')
parser.add_argument(
'-decoder_channels',
default=16,
help=
'Channels of first convolutional layer, afterwards channels are doubled.',
type=int)
parser.add_argument('-decoder_units',
default='1024,1024,1024,1024',
help='Units for MLP.')
return parser
def get_parser(self):
"""
Get parser.
:return: parser
:rtype: argparse.ArgumentParser
"""
parser = argparse.ArgumentParser(
description='Test attacks on classifier.')
parser.add_argument('-classifier_file',
default=paths.state_file('classifier'),
help='Snapshot state file of classifier.',
type=str)
parser.add_argument('-test_images_file',
default=paths.test_images_file(),
help='HDF5 file containing dataset.',
type=str)
parser.add_argument('-test_codes_file',
default=paths.test_codes_file(),
help='HDF5 file containing dataset.',
type=str)
parser.add_argument('-label_index',
default=2,
help='Column index in label file.',
type=int)
parser.add_argument(
'-perturbations_file',
default=paths.results_file('classifier/perturbations'),
help='HDF5 file containing perturbations.',
type=str)
parser.add_argument('-success_file',
default=paths.results_file('classifier/success'),
help='HDF5 file indicating attack success.',
type=str)
parser.add_argument(
'-probabilities_file',
default=paths.results_file('classifier/probabilities'),
help='HDF5 file containing attack probabilities.')
parser.add_argument('-results_file',
default='',
help='Path to pickled results file.',
type=str)
parser.add_argument('-plot_directory',
default=paths.experiment_dir('classifier'),
help='Path to PNG plot file for success rate.',
type=str)
parser.add_argument('-plot_manifolds',
default=False,
action='store_true',
help='Whether to plot manifolds.')
parser.add_argument('-no_gpu', dest='use_gpu', action='store_false')
parser.add_argument('-batch_size',
default=128,
help='Batch size of attack.',
type=int)
# Some network parameters.
parser.add_argument('-network_architecture',
default='standard',
help='Classifier architecture to use.',
type=str)
parser.add_argument('-network_activation',
default='relu',
help='Activation function to use.',
type=str)
parser.add_argument('-network_no_batch_normalization',
default=False,
help='Do not use batch normalization.',
action='store_true')
parser.add_argument(
'-network_channels',
default=16,
help=
'Channels of first convolutional layer, afterwards channels are doubled.',
type=int)
parser.add_argument('-network_dropout',
default=False,
action='store_true',
help='Whether to use dropout.')
parser.add_argument('-network_units',
default='1024,1024,1024,1024',
help='Units for MLP.')
return parser
def get_parser(self):
"""
Get parser.
:return: parser
:rtype: argparse.ArgumentParser
"""
parser = argparse.ArgumentParser(
description='Test attacks on classifier.')
parser.add_argument('-test_images_file',
default=paths.test_images_file(),
help='HDF5 file containing dataset.',
type=str)
parser.add_argument('-train_images_file',
default=paths.train_images_file(),
help='HDF5 file containing dataset.',
type=str)
parser.add_argument('-test_theta_file',
default=paths.test_theta_file(),
help='HDF5 file containing dataset.',
type=str)
parser.add_argument('-train_theta_file',
default=paths.train_theta_file(),
help='HDF5 file containing dataset.',
type=str)
parser.add_argument('-test_codes_file',
default=paths.test_codes_file(),
help='HDF5 file containing labels.',
type=str)
parser.add_argument('-label_index',
default=2,
help='Column index in label file.',
type=int)
parser.add_argument(
'-accuracy_file',
default=paths.results_file('classifier/accuracy'),
help='Correctly classified test samples of classifier.',
type=str)
parser.add_argument(
'-perturbations_file',
default=paths.results_file('classifier/perturbations'),
help='HDF5 file containing perturbations.',
type=str)
parser.add_argument('-success_file',
default=paths.results_file('classifier/success'),
help='HDF5 file indicating attack success.',
type=str)
parser.add_argument('-results_file',
default='',
help='Path to pickled results file.',
type=str)
parser.add_argument('-plot_directory',
default=paths.experiment_dir('classifier'),
help='Path to PNG plot file for success rate.',
type=str)
parser.add_argument('-plot_manifolds',
default=False,
action='store_true',
help='Whether to plot manifolds.')
parser.add_argument('-latent',
default=False,
action='store_true',
help='Latent statistics.')
return parser
def get_parser(self):
"""
Get parser.
:return: parser
:rtype: argparse.ArgumentParser
"""
parser = argparse.ArgumentParser(
description='Visualize attacks on decoder and classifier.')
parser.add_argument('-test_images_file',
default=paths.test_images_file(),
help='HDF5 file containing images.',
type=str)
parser.add_argument('-test_theta_file',
default=paths.results_file('test_theta'),
help='HDF5 file containing transformations.',
type=str)
parser.add_argument('-test_codes_file',
default=paths.test_codes_file(),
help='HDF5 file containing codes.',
type=str)
parser.add_argument('-label_index',
default=2,
help='Label index.',
type=int)
parser.add_argument('-decoder_files',
default=paths.state_file('decoder'),
help='Decoder state file.',
type=str)
parser.add_argument('-classifier_file',
default=paths.state_file('classifier'),
help='Snapshot state file of classifier.',
type=str)
parser.add_argument(
'-perturbations_file',
default=paths.results_file('learned_decoder/perturbations'),
help='HDF5 file containing perturbations.',
type=str)
parser.add_argument(
'-success_file',
default=paths.results_file('learned_decoder/success'),
help='HDF5 file containing perturbations.',
type=str)
parser.add_argument(
'-accuracy_file',
default=paths.results_file('learned_decoder/accuracy'),
help='Correctly classified test samples of classifier.',
type=str)
parser.add_argument(
'-output_directory',
default=paths.experiment_dir('decoder/perturbations'),
help='Directory to store visualizations.',
type=str)
parser.add_argument('-batch_size',
default=128,
help='Batch size of attack.',
type=int)
parser.add_argument('-no_gpu', dest='use_gpu', action='store_false')
# Some network parameters.
parser.add_argument('-network_architecture',
default='standard',
help='Classifier architecture to use.',
type=str)
parser.add_argument('-network_activation',
default='relu',
help='Activation function to use.',
type=str)
parser.add_argument('-network_no_batch_normalization',
default=False,
help='Do not use batch normalization.',
action='store_true')
parser.add_argument(
'-network_channels',
default=16,
help=
'Channels of first convolutional layer, afterwards channels are doubled.',
type=int)
parser.add_argument('-network_dropout',
default=False,
action='store_true',
help='Whether to use dropout.')
parser.add_argument('-network_units',
default='1024,1024,1024,1024',
help='Units for MLP.')
# Some decoder parameters.
parser.add_argument('-latent_space_size',
default=10,
help='Size of latent space.',
type=int)
parser.add_argument('-decoder_architecture',
default='standard',
help='Architecture to use.',
type=str)
parser.add_argument('-decoder_activation',
default='relu',
help='Activation function to use.',
type=str)
parser.add_argument('-decoder_no_batch_normalization',
default=False,
help='Do not use batch normalization.',
action='store_true')
parser.add_argument(
'-decoder_channels',
default=16,
help=
'Channels of first convolutional layer, afterwards channels are doubled.',
type=int)
parser.add_argument('-decoder_units',
default='1024,1024,1024,1024',
help='Units for MLP.')
return parser
def get_parser(self):
"""
Get parser.
:return: parser
:rtype: argparse.ArgumentParser
"""
parser = argparse.ArgumentParser(description='Train classifier.')
parser.add_argument('-train_images_file',
default=paths.train_images_file(),
help='HDF5 file containing dataset.',
type=str)
parser.add_argument('-train_codes_file',
default=paths.train_codes_file(),
help='HDF5 file containing codes.',
type=str)
parser.add_argument('-test_images_file',
default=paths.test_images_file(),
help='HDF5 file containing dataset.',
type=str)
parser.add_argument('-test_codes_file',
default=paths.test_codes_file(),
help='HDF5 file containing codes.',
type=str)
parser.add_argument('-state_file',
default=paths.state_file('stn_classifier'),
help='Snapshot state file.',
type=str)
parser.add_argument('-log_file',
default=paths.log_file('stn_classifier'),
help='Log file.',
type=str)
parser.add_argument('-training_file',
default=paths.results_file('stn_training'),
help='Training statistics file.',
type=str)
parser.add_argument('-testing_file',
default=paths.results_file('stn_testing'),
help='Testing statistics file.',
type=str)
parser.add_argument('-loss_file',
default=paths.image_file('loss'),
help='Loss plot file.',
type=str)
parser.add_argument('-error_file',
default=paths.image_file('error'),
help='Error plot file.',
type=str)
parser.add_argument('-gradient_file',
default='',
help='Gradient plot file.',
type=str)
parser.add_argument('-label_index',
default=2,
help='Label index.',
type=int)
parser.add_argument(
'-random_samples',
default=False,
action='store_true',
help='Randomize the subsampling of the training set.')
parser.add_argument('-training_samples',
default=-1,
help='Number of samples used for training.',
type=int)
parser.add_argument('-test_samples',
default=-1,
help='Number of samples for validation.',
type=int)
parser.add_argument('-validation_samples',
default=0,
help='Number of samples for validation.',
type=int)
parser.add_argument('-early_stopping',
default=False,
action='store_true',
help='Use early stopping.')
parser.add_argument('-batch_size',
default=64,
help='Batch size.',
type=int)
parser.add_argument('-epochs',
default=10,
help='Number of epochs.',
type=int)
parser.add_argument('-weight_decay',
default=0.0001,
help='Weight decay importance.',
type=float)
parser.add_argument('-logit_decay',
default=0,
help='Logit decay importance.',
type=float)
parser.add_argument('-no_gpu', dest='use_gpu', action='store_false')
parser.add_argument('-skip',
default=5,
help='Verbosity in iterations.',
type=int)
parser.add_argument('-lr',
default=0.005,
type=float,
help='Base learning rate.')
parser.add_argument('-lr_decay',
default=0.9,
type=float,
help='Learning rate decay.')
parser.add_argument('-results_file',
default='',
help='Results file for evaluation.',
type=str)
parser.add_argument('-debug_directory',
default='',
help='Debug directory.',
type=str)
# Some network parameters.
parser.add_argument('-network_architecture',
default='standard',
help='Classifier architecture to use.',
type=str)
parser.add_argument('-network_activation',
default='relu',
help='Activation function to use.',
type=str)
parser.add_argument('-network_no_batch_normalization',
default=False,
help='Do not use batch normalization.',
action='store_true')
parser.add_argument(
'-network_channels',
default=16,
help=
'Channels of first convolutional layer, afterwards channels are doubled.',
type=int)
parser.add_argument('-network_dropout',
default=False,
action='store_true',
help='Whether to use dropout.')
parser.add_argument('-network_units',
default='1024,1024,1024,1024',
help='Units for MLP.')
# Attack parameters.
parser.add_argument('-epsilon',
default=1,
help='Epsilon allowed for attacks.',
type=float)
parser.add_argument('-max_iterations',
default=10,
help='Number of iterations for attack.',
type=int)
parser.add_argument('-N_theta',
default=6,
help='Numer of transformations.',
type=int)
parser.add_argument('-translation_x',
default='-0.2,0.2',
type=str,
help='Minimum and maximum translation in x.')
parser.add_argument('-translation_y',
default='-0.2,0.2',
type=str,
help='Minimum and maximum translation in y')
parser.add_argument('-shear_x',
default='-0.5,0.5',
type=str,
help='Minimum and maximum shear in x.')
parser.add_argument('-shear_y',
default='-0.5,0.5',
type=str,
help='Minimum and maximum shear in y.')
parser.add_argument('-scale',
default='0.9,1.1',
type=str,
help='Minimum and maximum scale.')
parser.add_argument('-rotation',
default='%g,%g' % (-math.pi / 4, math.pi / 4),
type=str,
help='Minimum and maximum rotation.')
parser.add_argument('-color',
default=0.5,
help='Minimum color value, maximum is 1.',
type=float)
# Variants.
parser.add_argument('-norm',
default='inf',
help='Norm to use.',
type=float)
parser.add_argument('-full_variant',
default=False,
action='store_true',
help='100% variant.')
parser.add_argument('-strong_variant',
default=False,
action='store_true',
help='Strong data augmentation variant.')
return parser
def get_parser(self):
"""
Get parser.
:return: parser
:rtype: argparse.ArgumentParser
"""
parser = argparse.ArgumentParser(description='Sample auto encoder.')
parser.add_argument('-decoder_file',
default=paths.state_file('decoder'),
help='Snapshot state file.',
type=str)
parser.add_argument('-test_images_file',
default=paths.test_images_file(),
help='HDF5 file containing test images.',
type=str)
parser.add_argument('-images_file',
default=paths.results_file('sampled_images'),
help='HDF5 file for sampled test images.',
type=str)
parser.add_argument('-theta_file',
default=paths.results_file('sampled_theta'),
help='HDF5 file for sampled train theta.',
type=str)
parser.add_argument('-N_samples',
default=40000,
help='Number of samples.',
type=int)
parser.add_argument('-bound',
default=2,
help='Truncated normal bound.',
type=float)
parser.add_argument('-batch_size',
default=128,
help='Batch size.',
type=int)
parser.add_argument('-latent_space_size',
default=10,
help='Size of latent space.',
type=int)
parser.add_argument('-no_gpu', dest='use_gpu', action='store_false')
# Some network parameters.
parser.add_argument('-network_architecture',
default='standard',
help='Classifier architecture to use.',
type=str)
parser.add_argument('-network_activation',
default='relu',
help='Activation function to use.',
type=str)
parser.add_argument('-network_no_batch_normalization',
default=False,
help='Do not use batch normalization.',
action='store_true')
parser.add_argument(
'-network_channels',
default=16,
help=
'Channels of first convolutional layer, afterwards channels are doubled.',
type=int)
parser.add_argument('-network_units',
default='1024,1024,1024,1024',
help='Units for MLP.')
return parser
