def load_data(self):
"""
Load data.
"""
self.test_images = utils.read_hdf5(self.args.test_images_file).astype(numpy.float32)
if len(self.test_images.shape) < 4:
self.test_images = numpy.expand_dims(self.test_images, axis=3)
log('[Testing] read %s' % self.args.test_images_file)
self.test_codes = utils.read_hdf5(self.args.test_codes_file).astype(numpy.int)
self.test_codes = self.test_codes[:, self.args.label_index]
log('[Testing] read %s' % self.args.test_codes_file)
self.perturbations = utils.read_hdf5(self.args.perturbations_file).astype(numpy.float32)
self.N_attempts = self.perturbations.shape[0]
self.N_samples = self.perturbations.shape[1]
self.perturbations = numpy.swapaxes(self.perturbations, 0, 1)
if len(self.perturbations.shape) <= 4:
self.perturbations = self.perturbations.reshape((self.perturbations.shape[0] * self.perturbations.shape[1], self.perturbations.shape[2], self.perturbations.shape[3], 1))
else:
self.perturbations = self.perturbations.reshape((self.perturbations.shape[0] * self.perturbations.shape[1], self.perturbations.shape[2], self.perturbations.shape[3], self.perturbations.shape[4]))
log('[Testing] read %s' % self.args.perturbations_file)
self.original_success = utils.read_hdf5(self.args.original_success_file)
self.original_success = numpy.swapaxes(self.original_success, 0, 1)
self.original_success = self.original_success.reshape((self.original_success.shape[0] * self.original_success.shape[1]))
log('[Testing] read %s' % self.args.original_success_file)
self.original_accuracy = utils.read_hdf5(self.args.original_accuracy_file)
log('[Testing] read %s' % self.args.original_accuracy_file)
self.perturbation_codes = numpy.repeat(self.test_codes[:self.N_samples], self.N_attempts, axis=0)
self.transfer_success = numpy.copy(self.original_success)
def load_data(self):
"""
Load data and model.
"""
self.test_images = utils.read_hdf5(self.args.test_images_file).astype(numpy.float32)
log('[Testing] read %s' % self.args.test_images_file)
# For handling both color and gray images.
if len(self.test_images.shape) < 4:
self.test_images = numpy.expand_dims(self.test_images, axis=3)
log('[Testing] no color images, adjusted size')
self.resolution = self.test_images.shape[2]
log('[Testing] resolution %d' % self.resolution)
self.train_images = utils.read_hdf5(self.args.train_images_file).astype(numpy.float32)
# !
self.train_images = self.train_images.reshape((self.train_images.shape[0], -1))
log('[Testing] read %s' % self.args.train_images_file)
self.test_codes = utils.read_hdf5(self.args.test_codes_file).astype(numpy.int)
self.test_codes = self.test_codes[:, self.args.label_index]
self.N_class = numpy.max(self.test_codes) + 1
log('[Testing] read %s' % self.args.test_codes_file)
self.accuracy = utils.read_hdf5(self.args.accuracy_file)
log('[Testing] read %s' % self.args.accuracy_file)
self.perturbations = utils.read_hdf5(self.args.perturbations_file).astype(numpy.float32)
self.N_attempts = self.perturbations.shape[0]
# First, repeat relevant data.
self.test_images = numpy.repeat(self.test_images[:self.perturbations.shape[1]], self.N_attempts, axis=0)
self.perturbation_codes = numpy.repeat(self.test_codes[:self.perturbations.shape[1]], self.N_attempts, axis=0)
self.perturbation_codes = numpy.squeeze(self.perturbation_codes)
self.accuracy = numpy.repeat(self.accuracy[:self.perturbations.shape[1]], self.N_attempts, axis=0)
# Then, reshape the perturbations!
self.perturbations = numpy.swapaxes(self.perturbations, 0, 1)
self.perturbations = self.perturbations.reshape((self.perturbations.shape[0] * self.perturbations.shape[1], -1))
assert self.perturbations.shape[1] == self.args.N_theta
log('[Testing] read %s' % self.args.perturbations_file)
assert not numpy.any(self.perturbations != self.perturbations), 'NaN in perturbations'
self.success = utils.read_hdf5(self.args.success_file)
self.success = numpy.swapaxes(self.success, 0, 1)
self.success = self.success.reshape((self.success.shape[0] * self.success.shape[1]))
log('[Testing] read %s' % self.args.success_file)
log('[Testing] using %d input channels' % self.test_images.shape[3])
assert self.args.N_theta > 0 and self.args.N_theta <= 9
decoder = models.STNDecoder(self.args.N_theta)
# decoder.eval()
log('[Testing] set up STN decoder')
self.model = decoder
def load_data(self):
"""
Load data.
"""
test_codes = utils.read_hdf5(self.args.test_codes_file).astype(
numpy.int)
self.test_fonts = test_codes[:, 1]
self.test_classes = test_codes[:, 2]
log('[Testing] read %s' % self.args.test_codes_file)
self.perturbations = utils.read_hdf5(
self.args.perturbations_file).astype(numpy.float32)
self.N_attempts = self.perturbations.shape[0]
self.N_samples = self.perturbations.shape[1]
self.perturbations = numpy.swapaxes(self.perturbations, 0, 1)
self.perturbations = self.perturbations.reshape(
(self.perturbations.shape[0] * self.perturbations.shape[1], -1))
log('[Testing] read %s' % self.args.perturbations_file)
self.test_fonts = numpy.repeat(self.test_fonts[:self.N_samples],
self.N_attempts,
axis=0)
self.test_classes = numpy.repeat(self.test_classes[:self.N_samples],
self.N_attempts,
axis=0)
def load_data(self):
"""
Load data.
"""
self.test_images = utils.read_hdf5(self.args.test_images_file)
log('[Testing] read %s' % self.args.test_images_file)
if len(self.test_images.shape) < 4:
self.test_images = numpy.expand_dims(self.test_images, axis=3)
self.resolution = self.test_images.shape[1]
self.image_channels = self.test_images.shape[3]
self.test_codes = utils.read_hdf5(self.args.test_codes_file)
self.test_codes = self.test_codes[:, self.args.label_index]
log('[Testing] read %s' % self.args.test_codes_file)
self.perturbations = utils.read_hdf5(
self.args.perturbations_file).astype(numpy.float32)
self.N_attempts = self.perturbations.shape[0]
self.N_samples = self.perturbations.shape[1]
self.perturbations = numpy.swapaxes(self.perturbations, 0, 1)
self.perturbations = self.perturbations.reshape(
(self.perturbations.shape[0] * self.perturbations.shape[1], -1))
log('[Testing] read %s' % self.args.perturbations_file)
self.test_codes = numpy.repeat(self.test_codes[:self.N_samples],
self.N_attempts,
axis=0)
self.test_images = numpy.repeat(self.test_images[:self.N_samples],
self.N_attempts,
axis=0)
def test(self):
"""
Test classifier to identify valid samples to attack.
"""
num_batches = int(
math.ceil(self.perturbations.shape[0] / self.args.batch_size))
for b in range(num_batches):
b_start = b * self.args.batch_size
b_end = min((b + 1) * self.args.batch_size,
self.perturbations.shape[0])
batch_fonts = self.test_fonts[b_start:b_end]
batch_classes = self.test_classes[b_start:b_end]
batch_code = numpy.concatenate(
(common.numpy.one_hot(batch_fonts, self.N_font),
common.numpy.one_hot(batch_classes, self.N_class)),
axis=1).astype(numpy.float32)
batch_inputs = common.torch.as_variable(
self.perturbations[b_start:b_end], self.args.use_gpu)
batch_code = common.torch.as_variable(batch_code,
self.args.use_gpu)
# This basically allows to only optimize over theta, keeping the font/class code fixed.
self.model.set_code(batch_code)
output_images = self.model(batch_inputs)
output_images = numpy.squeeze(
numpy.transpose(output_images.cpu().detach().numpy(),
(0, 2, 3, 1)))
self.perturbation_images = common.numpy.concatenate(
self.perturbation_images, output_images)
if b % 100 == 0:
log('[Testing] computing perturbation images %d' % b)
utils.makedir(os.path.dirname(self.args.perturbation_images_file))
if len(self.perturbation_images.shape) > 3:
self.perturbation_images = self.perturbation_images.reshape(
self.N_samples, self.N_attempts,
self.perturbation_images.shape[1],
self.perturbation_images.shape[2],
self.perturbation_images.shape[3])
else:
self.perturbation_images = self.perturbation_images.reshape(
self.N_samples, self.N_attempts,
self.perturbation_images.shape[1],
self.perturbation_images.shape[2])
self.perturbation_images = numpy.swapaxes(self.perturbation_images, 0,
1)
utils.write_hdf5(self.args.perturbation_images_file,
self.perturbation_images)
log('[Testing] wrote %s' % self.args.perturbation_images_file)
def test(self):
"""
Test classifier to identify valid samples to attack.
"""
num_batches = int(
math.ceil(self.perturbations.shape[0] / self.args.batch_size))
for b in range(num_batches):
b_start = b * self.args.batch_size
b_end = min((b + 1) * self.args.batch_size,
self.perturbations.shape[0])
batch_images = common.torch.as_variable(
self.test_images[b_start:b_end], self.args.use_gpu)
batch_inputs = common.torch.as_variable(
self.perturbations[b_start:b_end], self.args.use_gpu)
self.model.set_image(batch_images)
output_images = self.model(batch_inputs)
output_images = numpy.squeeze(
numpy.transpose(output_images.cpu().detach().numpy(),
(0, 2, 3, 1)))
self.perturbation_images = common.numpy.concatenate(
self.perturbation_images, output_images)
if b % 100 == 0:
log('[Testing] computing perturbation images %d' % b)
utils.makedir(os.path.dirname(self.args.perturbation_images_file))
if len(self.perturbation_images.shape) > 3:
self.perturbation_images = self.perturbation_images.reshape(
self.N_samples, self.N_attempts,
self.perturbation_images.shape[1],
self.perturbation_images.shape[2],
self.perturbation_images.shape[3])
else:
self.perturbation_images = self.perturbation_images.reshape(
self.N_samples, self.N_attempts,
self.perturbation_images.shape[1],
self.perturbation_images.shape[2])
self.perturbation_images = numpy.swapaxes(self.perturbation_images, 0,
1)
utils.write_hdf5(self.args.perturbation_images_file,
self.perturbation_images)
log('[Testing] wrote %s' % self.args.perturbation_images_file)
def compute_statistics(self):
"""
Compute statistics based on distances.
"""
num_attempts = self.perturbations.shape[0]
perturbations = numpy.swapaxes(self.perturbations, 0, 1)
perturbations = perturbations.reshape(
(perturbations.shape[0] * perturbations.shape[1],
perturbations.shape[2]))
success = numpy.swapaxes(self.success, 0, 1)
success = success.reshape((success.shape[0] * success.shape[1]))
probabilities = numpy.swapaxes(self.probabilities, 0, 1)
probabilities = probabilities.reshape(
(probabilities.shape[0] * probabilities.shape[1], -1))
confidences = numpy.max(probabilities, 1)
perturbation_probabilities = self.test_probabilities[:self.success.
shape[1]]
perturbation_probabilities = numpy.repeat(perturbation_probabilities,
num_attempts,
axis=0)
perturbation_confidences = numpy.max(perturbation_probabilities, 1)
probability_ratios = confidences / perturbation_confidences
raw_overall_success = success >= 0
log('[Testing] %d valid attacks' % numpy.sum(raw_overall_success))
# For off-manifold attacks this should not happen, but save is save.
if not numpy.any(raw_overall_success):
for type in [
'raw_success', 'raw_iteration', 'raw_roc',
'raw_confidence_weighted_success', 'raw_confidence',
'raw_ratios'
]:
self.results[type] = 0
if self.args.results_file:
utils.write_pickle(self.args.results_file, self.results)
log('[Testing] wrote %s' % self.args.results_file)
log('[Testing] no successful attacks found, no plots')
return
#
# We compute some simple statistics:
# - raw success rate: fraction of successful attack without considering epsilon
# - corrected success rate: fraction of successful attacks within epsilon-ball
# - raw average perturbation: average distance to original samples (for successful attacks)
# - corrected average perturbation: average distance to original samples for perturbations
# within epsilon-ball (for successful attacks).
# These statistics can also be computed per class.
# And these statistics are computed with respect to three norms.
if self.args.plot_directory and utils.display():
iterations = success[raw_overall_success]
x = numpy.arange(numpy.max(iterations) + 1)
y = numpy.bincount(iterations)
plot_file = os.path.join(self.args.plot_directory, 'iterations')
plot.bar(plot_file,
x,
y,
title='Distribution of Iterations of Successful Attacks',
xlabel='Number of Iterations',
ylabel='Count')
log('[Testing] wrote %s' % plot_file)
plot_file = os.path.join(self.args.plot_directory, 'probabilities')
plot.histogram(plot_file, confidences[raw_overall_success], 50)
log('[Testing] wrote %s' % plot_file)
plot_file = os.path.join(self.args.plot_directory,
'probability_ratios')
plot.histogram(plot_file, probability_ratios, 50)
log('[Testing] wrote %s' % plot_file)
plot_file = os.path.join(self.args.plot_directory,
'test_probabilities')
plot.histogram(
plot_file, self.test_probabilities[
numpy.arange(self.test_probabilities.shape[0]),
self.test_codes], 50)
log('[Testing] wrote %s' % plot_file)
y_true = numpy.concatenate(
(numpy.zeros(confidences.shape[0]),
numpy.ones(perturbation_confidences.shape[0])))
y_score = numpy.concatenate((confidences, perturbation_confidences))
roc_auc_score = sklearn.metrics.roc_auc_score(y_true, y_score)
self.results['raw_roc'] = roc_auc_score
self.results['raw_confidence_weighted_success'] = numpy.sum(
confidences[raw_overall_success]) / numpy.sum(
perturbation_confidences)
self.results['raw_confidence'] = numpy.mean(
probabilities[raw_overall_success])
self.results['raw_ratios'] = numpy.mean(
probability_ratios[raw_overall_success])
self.results['raw_success'] = numpy.sum(
raw_overall_success) / success.shape[0]
self.results['raw_iteration'] = numpy.average(
success[raw_overall_success])
if self.args.results_file:
utils.write_pickle(self.args.results_file, self.results)
log('[Testing] wrote %s' % self.args.results_file)
def test(self):
"""
Test classifier to identify valid samples to attack.
"""
self.model.eval()
assert self.model.training is False
assert self.perturbation_codes.shape[0] == self.perturbations.shape[0]
assert self.test_codes.shape[0] == self.test_images.shape[0]
assert len(self.perturbations.shape) == 4
assert len(self.test_images.shape) == 4
perturbations_accuracy = None
num_batches = int(math.ceil(self.perturbations.shape[0] / self.args.batch_size))
for b in range(num_batches):
b_start = b * self.args.batch_size
b_end = min((b + 1) * self.args.batch_size, self.perturbations.shape[0])
batch_perturbations = common.torch.as_variable(self.perturbations[b_start: b_end], self.args.use_gpu)
batch_classes = common.torch.as_variable(self.perturbation_codes[b_start: b_end], self.args.use_gpu)
batch_perturbations = batch_perturbations.permute(0, 3, 1, 2)
output_classes = self.model(batch_perturbations)
values, indices = torch.max(torch.nn.functional.softmax(output_classes, dim=1), dim=1)
errors = torch.abs(indices - batch_classes)
perturbations_accuracy = common.numpy.concatenate(perturbations_accuracy, errors.data.cpu().numpy())
for n in range(batch_perturbations.size(0)):
log('[Testing] %d: original success=%d, transfer accuracy=%d' % (n, self.original_success[b_start + n], errors[n].item()))
self.transfer_success[perturbations_accuracy == 0] = -1
self.transfer_success = self.transfer_success.reshape((self.N_samples, self.N_attempts))
self.transfer_success = numpy.swapaxes(self.transfer_success, 0, 1)
utils.makedir(os.path.dirname(self.args.transfer_success_file))
utils.write_hdf5(self.args.transfer_success_file, self.transfer_success)
log('[Testing] wrote %s' % self.args.transfer_success_file)
num_batches = int(math.ceil(self.test_images.shape[0] / self.args.batch_size))
for b in range(num_batches):
b_start = b * self.args.batch_size
b_end = min((b + 1) * self.args.batch_size, self.test_images.shape[0])
batch_images = common.torch.as_variable(self.test_images[b_start: b_end], self.args.use_gpu)
batch_classes = common.torch.as_variable(self.test_codes[b_start: b_end], self.args.use_gpu)
batch_images = batch_images.permute(0, 3, 1, 2)
output_classes = self.model(batch_images)
values, indices = torch.max(torch.nn.functional.softmax(output_classes, dim=1), dim=1)
errors = torch.abs(indices - batch_classes)
self.transfer_accuracy = common.numpy.concatenate(self.transfer_accuracy, errors.data.cpu().numpy())
if b % 100 == 0:
log('[Testing] computing accuracy %d' % b)
self.transfer_accuracy = self.transfer_accuracy == 0
log('[Testing] original accuracy=%g' % (numpy.sum(self.original_accuracy)/float(self.original_accuracy.shape[0])))
log('[Testing] transfer accuracy=%g' % (numpy.sum(self.transfer_accuracy)/float(self.transfer_accuracy.shape[0])))
log('[Testing] accuracy difference=%g' % (numpy.sum(self.transfer_accuracy != self.original_accuracy)/float(self.transfer_accuracy.shape[0])))
log('[Testing] accuracy difference on %d samples=%g' % (self.N_samples, numpy.sum(self.transfer_accuracy[:self.N_samples] != self.original_accuracy[:self.N_samples])/float(self.N_samples)))
self.transfer_accuracy = numpy.logical_and(self.transfer_accuracy, self.original_accuracy)
utils.makedir(os.path.dirname(self.args.transfer_accuracy_file))
utils.write_hdf5(self.args.transfer_accuracy_file, self.transfer_accuracy)
log('[Testing] wrote %s' % self.args.transfer_accuracy_file)
def load_data(self):
"""
Load data and model.
"""
with logw('[Detection] read %s' % self.args.train_images_file):
self.nearest_neighbor_images = utils.read_hdf5(
self.args.train_images_file)
assert len(self.nearest_neighbor_images.shape) == 3
with logw('[Detection] read %s' % self.args.test_images_file):
self.test_images = utils.read_hdf5(self.args.test_images_file)
if len(self.test_images.shape) < 4:
self.test_images = numpy.expand_dims(self.test_images, axis=3)
with logw('[Detection] read %s' % self.args.train_codes_file):
self.train_codes = utils.read_hdf5(self.args.train_codes_file)
with logw('[Detection] read %s' % self.args.test_codes_file):
self.test_codes = utils.read_hdf5(self.args.test_codes_file)
with logw('[Detection] read %s' % self.args.test_theta_file):
self.test_theta = utils.read_hdf5(self.args.test_theta_file)
with logw('[Detection] read %s' % self.args.perturbations_file):
self.perturbations = utils.read_hdf5(self.args.perturbations_file)
assert len(self.perturbations.shape) == 3
with logw('[Detection] read %s' % self.args.success_file):
self.success = utils.read_hdf5(self.args.success_file)
with logw('[Detection] read %s' % self.args.accuracy_file):
self.accuracy = utils.read_hdf5(self.args.accuracy_file)
self.perturbations = numpy.swapaxes(self.perturbations, 0, 1)
num_attempts = self.perturbations.shape[1]
self.test_images = self.test_images[:self.perturbations.shape[0]]
self.train_images = self.nearest_neighbor_images[:self.perturbations.
shape[0]]
self.test_codes = self.test_codes[:self.perturbations.shape[0]]
self.accuracy = self.accuracy[:self.perturbations.shape[0]]
self.test_theta = self.test_theta[:self.perturbations.shape[0]]
self.perturbations = self.perturbations.reshape(
(self.perturbations.shape[0] * self.perturbations.shape[1],
self.perturbations.shape[2]))
self.success = numpy.swapaxes(self.success, 0, 1)
self.success = self.success.reshape(
(self.success.shape[0] * self.success.shape[1]))
self.accuracy = numpy.repeat(self.accuracy, num_attempts, axis=0)
self.test_images = numpy.repeat(self.test_images, num_attempts, axis=0)
self.train_images = numpy.repeat(self.train_images,
num_attempts,
axis=0)
self.test_codes = numpy.repeat(self.test_codes, num_attempts, axis=0)
self.test_theta = numpy.repeat(self.test_theta, num_attempts, axis=0)
max_samples = self.args.max_samples
self.success = self.success[:max_samples]
self.accuracy = self.accuracy[:max_samples]
self.perturbations = self.perturbations[:max_samples]
self.test_images = self.test_images[:max_samples]
self.train_images = self.train_images[:max_samples]
self.test_codes = self.test_codes[:max_samples]
self.test_theta = self.test_theta[:max_samples]
with logw('[Testing] read %s' % self.args.database_file):
database = utils.read_hdf5(self.args.database_file)
self.N_font = database.shape[0]
self.N_class = database.shape[1]
self.N_theta = self.test_theta.shape[1]
database = database.reshape((database.shape[0] * database.shape[1],
database.shape[2], database.shape[3]))
database = torch.from_numpy(database)
if self.args.use_gpu:
database = database.cuda()
database = torch.autograd.Variable(database, False)
self.model = models.AlternativeOneHotDecoder(
database, self.N_font, self.N_class, self.N_theta)
self.model.eval()
self.compute_images()
def load_data_and_model(self):
"""
Load data and model.
"""
database = utils.read_hdf5(self.args.database_file).astype(
numpy.float32)
log('[Visualization] read %s' % self.args.database_file)
N_font = database.shape[0]
N_class = database.shape[1]
resolution = database.shape[2]
database = database.reshape((database.shape[0] * database.shape[1],
database.shape[2], database.shape[3]))
database = torch.from_numpy(database)
if self.args.use_gpu:
database = database.cuda()
database = torch.autograd.Variable(database, False)
self.test_images = utils.read_hdf5(self.args.test_images_file).astype(
numpy.float32)
if len(self.test_images.shape) < 4:
self.test_images = numpy.expand_dims(self.test_images, axis=3)
self.perturbations = utils.read_hdf5(
self.args.perturbations_file).astype(numpy.float32)
self.perturbations = numpy.swapaxes(self.perturbations, 0, 1)
log('[Visualization] read %s' % self.args.perturbations_file)
self.success = utils.read_hdf5(self.args.success_file)
self.success = numpy.swapaxes(self.success, 0, 1)
log('[Visualization] read %s' % self.args.success_file)
self.accuracy = utils.read_hdf5(self.args.accuracy_file)
log('[Visualization] read %s' % self.args.success_file)
self.test_theta = utils.read_hdf5(self.args.test_theta_file).astype(
numpy.float32)
self.test_theta = self.test_theta[:self.perturbations.shape[0]]
N_theta = self.test_theta.shape[1]
log('[Visualization] using %d N_theta' % N_theta)
log('[Visualization] read %s' % self.args.test_theta_file)
self.test_codes = utils.read_hdf5(self.args.test_codes_file).astype(
numpy.int)
self.test_codes = self.test_codes[:self.perturbations.shape[0]]
self.test_codes = self.test_codes[:, 1:3]
self.test_codes = numpy.concatenate(
(common.numpy.one_hot(self.test_codes[:, 0], N_font),
common.numpy.one_hot(self.test_codes[:, 1], N_class)),
axis=1).astype(numpy.float32)
log('[Attack] read %s' % self.args.test_codes_file)
image_channels = 1 if N_theta <= 7 else 3
network_units = list(map(int, self.args.network_units.split(',')))
log('[Visualization] using %d input channels' % image_channels)
self.classifier = models.Classifier(
N_class,
resolution=(image_channels, resolution, resolution),
architecture=self.args.network_architecture,
activation=self.args.network_activation,
batch_normalization=not self.args.network_no_batch_normalization,
start_channels=self.args.network_channels,
dropout=self.args.network_dropout,
units=network_units)
self.decoder = models.AlternativeOneHotDecoder(database, N_font,
N_class, N_theta)
self.decoder.eval()
assert os.path.exists(
self.args.classifier_file
), 'state file %s not found' % self.args.classifier_file
state = State.load(self.args.classifier_file)
log('[Visualization] read %s' % self.args.classifier_file)
self.classifier.load_state_dict(state.model)
if self.args.use_gpu and not cuda.is_cuda(self.classifier):
log('[Visualization] classifier is not CUDA')
self.classifier = self.classifier.cuda()
log('[Visualization] loaded classifier')
self.classifier.eval()
log('[Visualization] set classifier to eval')
def load_data_and_model(self):
"""
Load data and model.
"""
self.test_images = utils.read_hdf5(self.args.test_images_file).astype(
numpy.float32)
if len(self.test_images.shape) < 4:
self.test_images = numpy.expand_dims(self.test_images, axis=3)
resolution = self.test_images.shape[2]
log('[Visualization] read %s' % self.args.test_images_file)
self.test_codes = utils.read_hdf5(self.args.test_codes_file).astype(
numpy.int)
self.test_codes = self.test_codes[:, self.args.label_index]
N_class = numpy.max(self.test_codes) + 1
log('[Visualization] read %s' % self.args.test_codes_file)
self.perturbations = utils.read_hdf5(
self.args.perturbations_file).astype(numpy.float32)
if len(self.perturbations.shape) < 5:
self.perturbations = numpy.expand_dims(self.perturbations, axis=4)
self.perturbations = numpy.swapaxes(self.perturbations, 0, 1)
self.test_images = self.test_images[:self.perturbations.shape[0]]
log('[Visualization] read %s' % self.args.perturbations_file)
self.success = utils.read_hdf5(self.args.success_file)
self.success = numpy.swapaxes(self.success, 0, 1)
self.success = self.success >= 0
log('[Visualization] read %s' % self.args.success_file)
if self.args.selection_file:
selection = utils.read_hdf5(self.args.selection_file)
log('[Visualization] read %s' % self.args.selection_file)
selection = numpy.swapaxes(selection, 0, 1)
selection = selection[:self.success.shape[0]]
selection = selection >= 0
assert len(selection.shape) == len(self.success.shape)
self.success = numpy.logical_and(self.success, selection)
log('[Visualization] updated selection')
self.accuracy = utils.read_hdf5(self.args.accuracy_file)
log('[Visualization] read %s' % self.args.success_file)
log('[Visualization] using %d input channels' %
self.test_images.shape[3])
network_units = list(map(int, self.args.network_units.split(',')))
self.model = models.Classifier(
N_class,
resolution=(self.test_images.shape[3], self.test_images.shape[1],
self.test_images.shape[2]),
architecture=self.args.network_architecture,
activation=self.args.network_activation,
batch_normalization=not self.args.network_no_batch_normalization,
start_channels=self.args.network_channels,
dropout=self.args.network_dropout,
units=network_units)
assert os.path.exists(
self.args.classifier_file
), 'state file %s not found' % self.args.classifier_file
state = State.load(self.args.classifier_file)
log('[Visualization] read %s' % self.args.classifier_file)
self.model.load_state_dict(state.model)
if self.args.use_gpu and not cuda.is_cuda(self.model):
log('[Visualization] classifier is not CUDA')
self.model = self.model.cuda()
log('[Visualization] loaded classifier')
self.model.eval()
log('[Visualization] set model to eval')
def load_data_and_model(self):
"""
Load data and model.
"""
self.test_images = utils.read_hdf5(self.args.test_images_file).astype(
numpy.float32)
if len(self.test_images.shape) < 4:
self.test_images = numpy.expand_dims(self.test_images, axis=3)
resolution = (self.test_images.shape[3], self.test_images.shape[1],
self.test_images.shape[2])
log('[Visualization] read %s' % self.args.test_images_file)
self.perturbations = utils.read_hdf5(
self.args.perturbations_file).astype(numpy.float32)
self.perturbations = numpy.swapaxes(self.perturbations, 0, 1)
log('[Visualization] read %s' % self.args.perturbations_file)
self.success = utils.read_hdf5(self.args.success_file)
self.success = numpy.swapaxes(self.success, 0, 1)
log('[Visualization] read %s' % self.args.success_file)
self.accuracy = utils.read_hdf5(self.args.accuracy_file)
log('[Visualization] read %s' % self.args.success_file)
self.test_theta = utils.read_hdf5(self.args.test_theta_file).astype(
numpy.float32)
self.test_theta = self.test_theta[:self.perturbations.shape[0]]
log('[Visualization] read %s' % self.args.test_theta_file)
self.test_codes = utils.read_hdf5(self.args.test_codes_file).astype(
numpy.int)
self.test_codes = self.test_codes[:, self.args.label_index]
self.N_class = numpy.max(self.test_codes) + 1
self.test_codes = self.test_codes[:self.perturbations.shape[0]]
log('[Visualization] read %s' % self.args.test_codes_file)
network_units = list(map(int, self.args.network_units.split(',')))
self.classifier = models.Classifier(
self.N_class,
resolution=resolution,
architecture=self.args.network_architecture,
activation=self.args.network_activation,
batch_normalization=not self.args.network_no_batch_normalization,
start_channels=self.args.network_channels,
dropout=self.args.network_dropout,
units=network_units)
assert os.path.exists(
self.args.classifier_file
), 'state file %s not found' % self.args.classifier_file
state = State.load(self.args.classifier_file)
log('[Visualization] read %s' % self.args.classifier_file)
self.classifier.load_state_dict(state.model)
if self.args.use_gpu and not cuda.is_cuda(self.classifier):
log('[Visualization] classifier is not CUDA')
self.classifier = self.classifier.cuda()
log('[Visualization] loaded classifier')
self.classifier.eval()
log('[Visualization] set classifier to eval')
assert self.args.decoder_files
decoder_files = self.args.decoder_files.split(',')
for decoder_file in decoder_files:
assert os.path.exists(
decoder_file), 'could not find %s' % decoder_file
log('[Visualization] using %d input channels' %
self.test_images.shape[3])
decoder_units = list(map(int, self.args.decoder_units.split(',')))
if len(decoder_files) > 1:
log('[Visualization] loading multiple decoders')
decoders = []
for i in range(len(decoder_files)):
decoder = models.LearnedDecoder(
self.args.latent_space_size,
resolution=resolution,
architecture=self.args.decoder_architecture,
start_channels=self.args.decoder_channels,
activation=self.args.decoder_activation,
batch_normalization=not self.args.
decoder_no_batch_normalization,
units=decoder_units)
state = State.load(decoder_files[i])
decoder.load_state_dict(state.model)
if self.args.use_gpu and not cuda.is_cuda(decoder):
decoder = decoder.cuda()
decoders.append(decoder)
decoder.eval()
log('[Visualization] loaded %s' % decoder_files[i])
self.decoder = models.SelectiveDecoder(decoders,
resolution=resolution)
else:
log('[Visualization] loading one decoder')
decoder = models.LearnedDecoder(
self.args.latent_space_size,
resolution=resolution,
architecture=self.args.decoder_architecture,
start_channels=self.args.decoder_channels,
activation=self.args.decoder_activation,
batch_normalization=not self.args.
decoder_no_batch_normalization,
units=decoder_units)
state = State.load(decoder_files[0])
decoder.load_state_dict(state.model)
if self.args.use_gpu and not cuda.is_cuda(decoder):
decoder = decoder.cuda()
decoder.eval()
log('[Visualization] read decoder')
self.decoder = decoder
def load_data(self):
"""
Load data and model.
"""
self.test_images = utils.read_hdf5(self.args.test_images_file).astype(
numpy.float32)
log('[Testing] read %s' % self.args.test_images_file)
if len(self.test_images.shape) <= 3:
self.test_images = numpy.expand_dims(self.test_images, axis=3)
log('[Testing] no color images, adjusted size')
self.train_images = utils.read_hdf5(
self.args.train_images_file).astype(numpy.float32)
# !
self.train_images = self.train_images.reshape(
(self.train_images.shape[0], -1))
log('[Testing] read %s' % self.args.train_images_file)
self.test_theta = utils.read_hdf5(self.args.test_theta_file).astype(
numpy.float32)
log('[Testing] read %s' % self.args.test_theta_file)
self.train_theta = utils.read_hdf5(self.args.train_theta_file).astype(
numpy.float32)
log('[Testing] read %s' % self.args.train_theta_file)
self.test_codes = utils.read_hdf5(
self.args.test_codes_file).astype(int)
log('[Testing] read %s' % self.args.test_codes_file)
self.accuracy = utils.read_hdf5(self.args.accuracy_file)
log('[Testing] read %s' % self.args.accuracy_file)
self.perturbations = utils.read_hdf5(
self.args.perturbations_file).astype(numpy.float32)
self.N_attempts = self.perturbations.shape[0]
# First, repeat relevant data.
self.perturbation_theta = numpy.repeat(
self.test_theta[:self.perturbations.shape[1]],
self.N_attempts,
axis=0)
self.perturbation_codes = numpy.repeat(
self.test_codes[:self.perturbations.shape[1]],
self.N_attempts,
axis=0)
self.accuracy = numpy.repeat(
self.accuracy[:self.perturbations.shape[1]],
self.N_attempts,
axis=0)
# Then, reshape the perturbations!
self.perturbations = numpy.swapaxes(self.perturbations, 0, 1)
self.perturbations = self.perturbations.reshape(
(self.perturbations.shape[0] * self.perturbations.shape[1], -1))
log('[Testing] read %s' % self.args.perturbations_file)
self.success = utils.read_hdf5(self.args.success_file)
self.success = numpy.swapaxes(self.success, 0, 1)
self.success = self.success.reshape(
(self.success.shape[0] * self.success.shape[1]))
log('[Testing] read %s' % self.args.success_file)
database = utils.read_hdf5(self.args.database_file)
log('[Testing] read %s' % self.args.database_file)
self.N_font = database.shape[0]
self.N_class = database.shape[1]
N_theta = self.test_theta.shape[1]
log('[Testing] using %d N_theta' % N_theta)
database = database.reshape((database.shape[0] * database.shape[1],
database.shape[2], database.shape[3]))
database = torch.from_numpy(database)
if self.args.use_gpu:
database = database.cuda()
database = torch.autograd.Variable(database, False)
self.model = models.AlternativeOneHotDecoder(database, self.N_font,
self.N_class, N_theta)
self.model.eval()
def load_data(self):
"""
Load data and model.
"""
with logw('[Detection] read %s' % self.args.train_images_file):
self.nearest_neighbor_images = utils.read_hdf5(self.args.train_images_file)
assert len(self.nearest_neighbor_images.shape) == 3
with logw('[Detection] read %s' % self.args.test_images_file):
self.test_images = utils.read_hdf5(self.args.test_images_file)
if len(self.test_images.shape) < 4:
self.test_images = numpy.expand_dims(self.test_images, axis=3)
with logw('[Detection] read %s' % self.args.perturbations_file):
self.perturbations = utils.read_hdf5(self.args.perturbations_file)
assert len(self.perturbations.shape) == 4
with logw('[Detection] read %s' % self.args.success_file):
self.success = utils.read_hdf5(self.args.success_file)
with logw('[Detection] read %s' % self.args.accuracy_file):
self.accuracy = utils.read_hdf5(self.args.accuracy_file)
self.perturbations = numpy.swapaxes(self.perturbations, 0, 1)
num_attempts = self.perturbations.shape[1]
self.test_images = self.test_images[:self.perturbations.shape[0]]
self.train_images = self.nearest_neighbor_images[:self.perturbations.shape[0]]
self.accuracy = self.accuracy[:self.perturbations.shape[0]]
self.perturbations = self.perturbations.reshape((self.perturbations.shape[0]*self.perturbations.shape[1], self.perturbations.shape[2], self.perturbations.shape[3]))
self.success = numpy.swapaxes(self.success, 0, 1)
self.success = self.success.reshape((self.success.shape[0]*self.success.shape[1]))
self.accuracy = numpy.repeat(self.accuracy, num_attempts, axis=0)
self.test_images = numpy.repeat(self.test_images, num_attempts, axis=0)
self.train_images = numpy.repeat(self.train_images, num_attempts, axis=0)
max_samples = self.args.max_samples
self.success = self.success[:max_samples]
self.accuracy = self.accuracy[:max_samples]
self.perturbations = self.perturbations[:max_samples]
self.test_images = self.test_images[:max_samples]
self.train_images = self.train_images[:max_samples]
if self.args.mode == 'true':
assert self.args.database_file
assert self.args.test_codes_file
assert self.args.test_theta_file
self.test_codes = utils.read_hdf5(self.args.test_codes_file)
log('[Detection] read %s' % self.args.test_codes_file)
self.test_theta = utils.read_hdf5(self.args.test_theta_file)
log('[Detection] read %s' % self.args.test_theta_file)
self.test_codes = self.test_codes[:self.perturbations.shape[0]]
self.test_theta = self.test_theta[:self.perturbations.shape[0]]
self.test_codes = numpy.repeat(self.test_codes, num_attempts, axis=0)
self.test_theta = numpy.repeat(self.test_theta, num_attempts, axis=0)
self.test_codes = self.test_codes[:max_samples]
self.test_theta = self.test_theta[:max_samples]
database = utils.read_hdf5(self.args.database_file)
log('[Detection] read %s' % self.args.database_file)
self.N_font = database.shape[0]
self.N_class = database.shape[1]
self.N_theta = self.test_theta.shape[1]
database = database.reshape((database.shape[0]*database.shape[1], database.shape[2], database.shape[3]))
database = torch.from_numpy(database)
if self.args.use_gpu:
database = database.cuda()
database = torch.autograd.Variable(database, False)
self.model = models.AlternativeOneHotDecoder(database, self.N_font, self.N_class, self.N_theta)
self.model.eval()
log('[Detection] initialized decoder')
elif self.args.mode == 'appr':
assert self.args.decoder_files
assert self.args.test_codes_file
assert self.args.test_theta_file
self.test_codes = utils.read_hdf5(self.args.test_codes_file)
log('[Detection] read %s' % self.args.test_codes_file)
self.test_theta = utils.read_hdf5(self.args.test_theta_file)
log('[Detection] read %s' % self.args.test_theta_file)
self.test_codes = self.test_codes[:self.perturbations.shape[0]]
self.test_theta = self.test_theta[:self.perturbations.shape[0]]
self.test_codes = numpy.repeat(self.test_codes, num_attempts, axis=0)
self.test_theta = numpy.repeat(self.test_theta, num_attempts, axis=0)
self.test_codes = self.test_codes[:max_samples]
self.test_theta = self.test_theta[:max_samples]
assert self.args.decoder_files
decoder_files = self.args.decoder_files.split(',')
for decoder_file in decoder_files:
assert os.path.exists(decoder_file), 'could not find %s' % decoder_file
resolution = [1 if len(self.test_images.shape) <= 3 else self.test_images.shape[3], self.test_images.shape[1], self.test_images.shape[2]]
decoder_units = list(map(int, self.args.decoder_units.split(',')))
if len(decoder_files) > 1:
log('[Detection] loading multiple decoders')
decoders = []
for i in range(len(decoder_files)):
decoder = models.LearnedDecoder(self.args.latent_space_size,
resolution=resolution,
architecture=self.args.decoder_architecture,
start_channels=self.args.decoder_channels,
activation=self.args.decoder_activation,
batch_normalization=not self.args.decoder_no_batch_normalization,
units=decoder_units)
state = State.load(decoder_files[i])
decoder.load_state_dict(state.model)
if self.args.use_gpu and not cuda.is_cuda(decoder):
decoder = decoder.cuda()
decoders.append(decoder)
decoder.eval()
log('[Detection] loaded %s' % decoder_files[i])
self.model = models.SelectiveDecoder(decoders, resolution=resolution)
else:
log('[Detection] loading one decoder')
decoder = models.LearnedDecoder(self.args.latent_space_size,
resolution=resolution,
architecture=self.args.decoder_architecture,
start_channels=self.args.decoder_channels,
activation=self.args.decoder_activation,
batch_normalization=not self.args.decoder_no_batch_normalization,
units=decoder_units)
state = State.load(decoder_files[0])
decoder.load_state_dict(state.model)
if self.args.use_gpu and not cuda.is_cuda(decoder):
decoder = decoder.cuda()
decoder.eval()
log('[Detection] read decoder')
self.model = decoder
def load_data(self):
"""
Load data and model.
"""
self.test_images = utils.read_hdf5(self.args.test_images_file).astype(numpy.float32)
log('[Testing] read %s' % self.args.test_images_file)
# For handling both color and gray images.
if len(self.test_images.shape) < 4:
self.test_images = numpy.expand_dims(self.test_images, axis=3)
log('[Testing] no color images, adjusted size')
self.resolution = self.test_images.shape[2]
log('[Testing] resolution %d' % self.resolution)
self.train_images = utils.read_hdf5(self.args.train_images_file).astype(numpy.float32)
# !
self.train_images = self.train_images.reshape((self.train_images.shape[0], -1))
log('[Testing] read %s' % self.args.train_images_file)
self.test_theta = utils.read_hdf5(self.args.test_theta_file).astype(numpy.float32)
log('[Testing] read %s' % self.args.test_theta_file)
self.train_theta = utils.read_hdf5(self.args.train_theta_file).astype(numpy.float32)
log('[Testing] read %s' % self.args.train_theta_file)
self.test_codes = utils.read_hdf5(self.args.test_codes_file).astype(numpy.int)
self.test_codes = self.test_codes[:, self.args.label_index]
self.N_class = numpy.max(self.test_codes) + 1
log('[Testing] read %s' % self.args.test_codes_file)
self.accuracy = utils.read_hdf5(self.args.accuracy_file)
log('[Testing] read %s' % self.args.accuracy_file)
self.perturbations = utils.read_hdf5(self.args.perturbations_file).astype(numpy.float32)
self.N_attempts = self.perturbations.shape[0]
assert not numpy.any(self.perturbations != self.perturbations), 'NaN in perturbations'
# First, repeat relevant data.
self.perturbation_theta = numpy.repeat(self.test_theta[:self.perturbations.shape[1]], self.N_attempts, axis=0)
self.perturbation_codes = numpy.repeat(self.test_codes[:self.perturbations.shape[1]], self.N_attempts, axis=0)
self.perturbation_codes = numpy.squeeze(self.perturbation_codes)
self.accuracy = numpy.repeat(self.accuracy[:self.perturbations.shape[1]], self.N_attempts, axis=0)
# Then, reshape the perturbations!
self.perturbations = numpy.swapaxes(self.perturbations, 0, 1)
self.perturbations = self.perturbations.reshape((self.perturbations.shape[0] * self.perturbations.shape[1], -1))
log('[Testing] read %s' % self.args.perturbations_file)
self.success = utils.read_hdf5(self.args.success_file)
self.success = numpy.swapaxes(self.success, 0, 1)
self.success = self.success.reshape((self.success.shape[0] * self.success.shape[1]))
log('[Testing] read %s' % self.args.success_file)
assert self.args.decoder_files
decoder_files = self.args.decoder_files.split(',')
for decoder_file in decoder_files:
assert os.path.exists(decoder_file), 'could not find %s' % decoder_file
log('[Testing] using %d input channels' % self.test_images.shape[3])
decoder_units = list(map(int, self.args.decoder_units.split(',')))
if len(decoder_files) > 1:
log('[Testing] loading multiple decoders')
decoders = []
for i in range(len(decoder_files)):
decoder = models.LearnedDecoder(self.args.latent_space_size, resolution=(self.test_images.shape[3], self.test_images.shape[1], self.test_images.shape[2]),
architecture=self.args.decoder_architecture,
start_channels=self.args.decoder_channels,
activation=self.args.decoder_activation,
batch_normalization=not self.args.decoder_no_batch_normalization,
units=decoder_units)
state = State.load(decoder_files[i])
decoder.load_state_dict(state.model)
if self.args.use_gpu and not cuda.is_cuda(decoder):
decoder = decoder.cuda()
decoders.append(decoder)
decoder.eval()
log('[Testing] loaded %s' % decoder_files[i])
self.model = models.SelectiveDecoder(decoders, resolution=(self.test_images.shape[3], self.test_images.shape[1], self.test_images.shape[2]))
else:
log('[Testing] loading one decoder')
decoder = models.LearnedDecoder(self.args.latent_space_size, resolution=(self.test_images.shape[3], self.test_images.shape[1], self.test_images.shape[2]),
architecture=self.args.decoder_architecture,
start_channels=self.args.decoder_channels,
activation=self.args.decoder_activation,
batch_normalization=not self.args.decoder_no_batch_normalization,
units=decoder_units)
state = State.load(decoder_files[0])
decoder.load_state_dict(state.model)
if self.args.use_gpu and not cuda.is_cuda(decoder):
decoder = decoder.cuda()
decoder.eval()
log('[Testing] read decoder')
self.model = decoder