"TSNE of Sign of Adv Gradients, SNNLCrossEntropy Model, factor:" +
str(FLAGS.SNNL_factor),
fontsize=42)
imscatter(X_embedded, x_test[:batch_size], zoom=2, cmap="Purples")
plt.savefig(output_dir + 'adversarial_gradients_SNNL_factor_' +
str(SNNL_factor) + '.png')
def main(argv=None):
SNNL_example(nb_epochs=FLAGS.nb_epochs,
batch_size=FLAGS.batch_size,
learning_rate=FLAGS.learning_rate,
nb_filters=FLAGS.nb_filters,
SNNL_factor=FLAGS.SNNL_factor,
output_dir=FLAGS.output_dir)
if __name__ == '__main__':
flags.DEFINE_integer('nb_filters', NB_FILTERS, 'Model size multiplier')
flags.DEFINE_integer('nb_epochs', NB_EPOCHS,
'Number of epochs to train model')
flags.DEFINE_integer('batch_size', BATCH_SIZE, 'Size of training batches')
flags.DEFINE_float('SNNL_factor', SNNL_FACTOR,
'Multiplier for Soft Nearest Neighbor Loss')
flags.DEFINE_float('learning_rate', LEARNING_RATE,
'Learning rate for training')
flags.DEFINE_string('output_dir', OUTPUT_DIR,
'output directory for saving figures')
tf.app.run()
import matplotlib.pyplot as plt
plt.close(figure)
_ = grid_visual(grid_viz_data)
return report
def main(argv=None):
from src.FGSM.cleverhans.cleverhans_tutorials import check_installation
check_installation(__file__)
mnist_tutorial_jsma(viz_enabled=FLAGS.viz_enabled,
nb_epochs=FLAGS.nb_epochs,
batch_size=FLAGS.batch_size,
source_samples=FLAGS.source_samples,
learning_rate=FLAGS.learning_rate)
if __name__ == '__main__':
flags.DEFINE_boolean('viz_enabled', VIZ_ENABLED,
'Visualize adversarial ex.')
flags.DEFINE_integer('nb_epochs', NB_EPOCHS,
'Number of epochs to train model')
flags.DEFINE_integer('batch_size', BATCH_SIZE, 'Size of training batches')
flags.DEFINE_integer('source_samples', SOURCE_SAMPLES,
'Nb of test inputs to attack')
flags.DEFINE_float('learning_rate', LEARNING_RATE,
'Learning rate for training')
tf.app.run()
trainer = TrainerSingleGPU(hparams)
trainer.model_train()
trainer.eval(inc_epoch=False)
return trainer.finish()
def main(argv=None):
f = {x: flags.FLAGS[x].value for x in dir(flags.FLAGS)}
HParams = namedtuple('HParams', f.keys())
hparams = HParams(**f)
run_trainer(hparams)
if __name__ == '__main__':
flags.DEFINE_integer('train_start', 0,
'Index of first training set example.')
flags.DEFINE_integer('train_end', 60000,
'Index of last training set example.')
flags.DEFINE_integer('test_start', 0, 'Index of first test set example.')
flags.DEFINE_integer('test_end', 10000, 'Index of last test set example.')
flags.DEFINE_integer('nb_epochs', 6, 'Number of epochs to train model.')
flags.DEFINE_integer('batch_size', 128, 'Size of training batches.')
flags.DEFINE_boolean('adv_train', False,
'Whether to do adversarial training.')
flags.DEFINE_boolean('save', True, 'Whether to save from a checkpoint.')
flags.DEFINE_string('save_dir', 'runs/X', 'Location to store logs/model.')
flags.DEFINE_string('model_type', 'madry',
'Model type: basic|madry|resnet_tf.')
flags.DEFINE_string(
'attack_type_train', 'MadryEtAl_y_multigpu',
'Attack type for adversarial training:\
assert model_filepath.endswith('.joblib')
report_path = model_filepath[:-len('.joblib')] + suffix
goal = MaxConfidence()
bundle_examples_with_goal(sess,
model,
adv_x_list,
y,
goal,
report_path,
batch_size=FLAGS.batch_size)
if __name__ == '__main__':
flags.DEFINE_string('report_path', None, 'Report path')
flags.DEFINE_integer(
'train_start', TRAIN_START, 'Starting point (inclusive)'
'of range of train examples to use')
flags.DEFINE_integer(
'train_end', TRAIN_END, 'Ending point (non-inclusive) '
'of range of train examples to use')
flags.DEFINE_integer(
'test_start', TEST_START, 'Starting point '
'(inclusive) of range of test examples to use')
flags.DEFINE_integer(
'test_end', TEST_END, 'End point (non-inclusive) of '
'range of test examples to use')
flags.DEFINE_string('which_set', WHICH_SET, '"train" or "test"')
flags.DEFINE_integer('batch_size', BATCH_SIZE, 'batch size')
tf.app.run()
args=eval_params)
report.train_adv_train_adv_eval = accuracy
return report
def main(argv=None):
from src.FGSM.cleverhans.cleverhans_tutorials import check_installation
check_installation(__file__)
mnist_tutorial(nb_epochs=FLAGS.nb_epochs,
batch_size=FLAGS.batch_size,
learning_rate=FLAGS.learning_rate,
train_dir=FLAGS.train_dir,
filename=FLAGS.filename,
load_model=FLAGS.load_model)
if __name__ == '__main__':
flags.DEFINE_integer('nb_epochs', NB_EPOCHS,
'Number of epochs to train model')
flags.DEFINE_integer('batch_size', BATCH_SIZE, 'Size of training batches')
flags.DEFINE_float('learning_rate', LEARNING_RATE,
'Learning rate for training')
flags.DEFINE_string('train_dir', TRAIN_DIR,
'Directory where to save model.')
flags.DEFINE_string('filename', FILENAME, 'Checkpoint filename.')
flags.DEFINE_boolean('load_model', LOAD_MODEL,
'Load saved model or train.')
tf.app.run()
def main(argv=None):
"""
Print accuracies
"""
try:
_name_of_script, filepath = argv
except ValueError:
raise ValueError(argv)
print_accuracies(filepath=filepath, test_start=FLAGS.test_start,
test_end=FLAGS.test_end, which_set=FLAGS.which_set,
nb_iter=FLAGS.nb_iter, base_eps_iter=FLAGS.base_eps_iter,
batch_size=FLAGS.batch_size)
if __name__ == '__main__':
flags.DEFINE_integer('train_start', TRAIN_START, 'Starting point (inclusive)'
'of range of train examples to use')
flags.DEFINE_integer('train_end', TRAIN_END, 'Ending point (non-inclusive) '
'of range of train examples to use')
flags.DEFINE_integer('test_start', TEST_START, 'Starting point (inclusive) '
'of range of test examples to use')
flags.DEFINE_integer('test_end', TEST_END, 'End point (non-inclusive) of '
'range of test examples to use')
flags.DEFINE_integer('nb_iter', NB_ITER, 'Number of iterations of PGD')
flags.DEFINE_string('which_set', WHICH_SET, '"train" or "test"')
flags.DEFINE_integer('batch_size', BATCH_SIZE,
'Batch size for most jobs')
flags.DEFINE_float('base_eps_iter', BASE_EPS_ITER,
'epsilon per iteration, if data were in [0, 1]')
tf.app.run()
"""
try:
_name_of_script, filepath = argv
except ValueError:
raise ValueError(argv)
make_confidence_report_spsa(filepath=filepath, test_start=FLAGS.test_start,
test_end=FLAGS.test_end,
which_set=FLAGS.which_set,
report_path=FLAGS.report_path,
nb_iter=FLAGS.nb_iter,
batch_size=FLAGS.batch_size,
spsa_samples=FLAGS.spsa_samples,
spsa_iters=FLAGS.spsa_iters)
if __name__ == '__main__':
flags.DEFINE_integer('spsa_samples', SPSA_SAMPLES, 'Number samples for SPSA')
flags.DEFINE_integer('spsa_iters', SPSA.DEFAULT_SPSA_ITERS,
'Passed to SPSA.generate')
flags.DEFINE_integer('train_start', TRAIN_START, 'Starting point (inclusive)'
'of range of train examples to use')
flags.DEFINE_integer('train_end', TRAIN_END, 'Ending point (non-inclusive) '
'of range of train examples to use')
flags.DEFINE_integer('test_start', TEST_START,
'Starting point (inclusive) of range'
' of test examples to use')
flags.DEFINE_integer('test_end', TEST_END,
'End point (non-inclusive) of range'
' of test examples to use')
flags.DEFINE_integer('nb_iter', NB_ITER_SPSA, 'Number of iterations of SPSA')
flags.DEFINE_string('which_set', WHICH_SET, '"train" or "test"')
flags.DEFINE_string('report_path', REPORT_PATH, 'Path to save to')
check_installation(__file__)
mnist_tutorial_cw(viz_enabled=FLAGS.viz_enabled,
nb_epochs=FLAGS.nb_epochs,
batch_size=FLAGS.batch_size,
source_samples=FLAGS.source_samples,
learning_rate=FLAGS.learning_rate,
attack_iterations=FLAGS.attack_iterations,
model_path=FLAGS.model_path,
targeted=FLAGS.targeted)
if __name__ == '__main__':
flags.DEFINE_boolean('viz_enabled', VIZ_ENABLED,
'Visualize adversarial ex.')
flags.DEFINE_integer('nb_epochs', NB_EPOCHS,
'Number of epochs to train model')
flags.DEFINE_integer('batch_size', BATCH_SIZE, 'Size of training batches')
flags.DEFINE_integer('source_samples', SOURCE_SAMPLES,
'Number of test inputs to attack')
flags.DEFINE_float('learning_rate', LEARNING_RATE,
'Learning rate for training')
flags.DEFINE_string('model_path', MODEL_PATH,
'Path to save or load the model file')
flags.DEFINE_integer('attack_iterations', ATTACK_ITERATIONS,
'Number of iterations to run attack; 1000 is good')
flags.DEFINE_boolean('targeted', TARGETED,
'Run the tutorial in targeted mode?')
tf.app.run()
raise ValueError(argv)
make_confidence_report(filepath=filepath,
test_start=FLAGS.test_start,
test_end=FLAGS.test_end,
which_set=FLAGS.which_set,
report_path=FLAGS.report_path,
mc_batch_size=FLAGS.mc_batch_size,
nb_iter=FLAGS.nb_iter,
base_eps_iter=FLAGS.base_eps_iter,
batch_size=FLAGS.batch_size,
save_advx=FLAGS.save_advx)
if __name__ == '__main__':
flags.DEFINE_integer(
'train_start', TRAIN_START, 'Starting point (inclusive)'
'of range of train examples to use')
flags.DEFINE_integer(
'train_end', TRAIN_END, 'Ending point (non-inclusive) '
'of range of train examples to use')
flags.DEFINE_integer(
'test_start', TEST_START, 'Starting point (inclusive) '
'of range of test examples to use')
flags.DEFINE_integer(
'test_end', TEST_END, 'End point (non-inclusive) of '
'range of test examples to use')
flags.DEFINE_integer('nb_iter', NB_ITER, 'Number of iterations of PGD')
flags.DEFINE_string('which_set', WHICH_SET, '"train" or "test"')
flags.DEFINE_string('report_path', REPORT_PATH, 'Path to save to')
flags.DEFINE_integer('mc_batch_size', MC_BATCH_SIZE,
'Batch size for MaxConfidence')
mnist_blackbox(nb_classes=FLAGS.nb_classes,
batch_size=FLAGS.batch_size,
learning_rate=FLAGS.learning_rate,
nb_epochs=FLAGS.nb_epochs,
holdout=FLAGS.holdout,
data_aug=FLAGS.data_aug,
nb_epochs_s=FLAGS.nb_epochs_s,
lmbda=FLAGS.lmbda,
aug_batch_size=FLAGS.data_aug_batch_size)
if __name__ == '__main__':
# General flags
flags.DEFINE_integer('nb_classes', NB_CLASSES,
'Number of classes in problem')
flags.DEFINE_integer('batch_size', BATCH_SIZE, 'Size of training batches')
flags.DEFINE_float('learning_rate', LEARNING_RATE,
'Learning rate for training')
# Flags related to oracle
flags.DEFINE_integer('nb_epochs', NB_EPOCHS,
'Number of epochs to train model')
# Flags related to substitute
flags.DEFINE_integer('holdout', HOLDOUT, 'Test set holdout for adversary')
flags.DEFINE_integer('data_aug', DATA_AUG,
'Number of substitute data augmentations')
flags.DEFINE_integer('nb_epochs_s', NB_EPOCHS_S,
'Training epochs for substitute')
flags.DEFINE_float('lmbda', LMBDA, 'Lambda from arxiv.org/abs/1602.02697')
