def train(self, epoch):
"""
Train for one epoch.
:param epoch: current epoch
:type epoch: int
"""
self.encoder.train()
log('[Training] %d set encoder to train' % epoch)
self.decoder.train()
log('[Training] %d set decoder to train' % epoch)
self.classifier.train()
log('[Training] %d set classifier to train' % epoch)
num_batches = int(
math.ceil(self.train_images.shape[0] / self.args.batch_size))
assert self.encoder.training is True
permutation = numpy.random.permutation(self.train_images.shape[0])
permutation = numpy.concatenate(
(permutation, permutation[:self.args.batch_size]), axis=0)
for b in range(num_batches):
self.encoder_scheduler.update(epoch, float(b) / num_batches)
self.decoder_scheduler.update(epoch, float(b) / num_batches)
self.classifier_scheduler.update(epoch, float(b) / num_batches)
perm = permutation[b * self.args.batch_size:(b + 1) *
self.args.batch_size]
batch_images = common.torch.as_variable(self.train_images[perm],
self.args.use_gpu, True)
batch_images = batch_images.permute(0, 3, 1, 2)
output_mu, output_logvar = self.encoder(batch_images)
output_codes = self.reparameterize(output_mu, output_logvar)
output_images = self.decoder(output_codes)
output_real_classes = self.classifier(batch_images)
output_reconstructed_classes = self.classifier(output_images)
latent_loss = self.latent_loss(output_mu, output_logvar)
reconstruction_loss = self.reconstruction_loss(
batch_images, output_images)
decoder_loss = self.decoder_loss(output_reconstructed_classes)
discriminator_loss = self.discriminator_loss(
output_real_classes, output_reconstructed_classes)
self.encoder_scheduler.optimizer.zero_grad()
loss = latent_loss + self.args.beta * reconstruction_loss + self.args.gamma * decoder_loss + self.args.eta * torch.sum(
torch.abs(output_logvar))
loss.backward(retain_graph=True)
self.encoder_scheduler.optimizer.step()
self.decoder_scheduler.optimizer.zero_grad()
loss = self.args.beta * reconstruction_loss + self.args.gamma * decoder_loss
loss.backward(retain_graph=True)
self.decoder_scheduler.optimizer.step()
self.classifier_scheduler.optimizer.zero_grad()
loss = self.args.gamma * discriminator_loss
loss.backward()
self.classifier_scheduler.optimizer.step()
reconstruction_error = self.reconstruction_error(
batch_images, output_images)
iteration = epoch * num_batches + b + 1
self.train_statistics = numpy.vstack(
(self.train_statistics,
numpy.array([
iteration, iteration * self.args.batch_size,
min(num_batches, iteration),
min(num_batches, iteration) * self.args.batch_size,
reconstruction_loss.data, reconstruction_error.data,
latent_loss.data,
torch.mean(output_mu).item(),
torch.var(output_mu).item(),
torch.mean(output_logvar).item(),
decoder_loss.item(),
discriminator_loss.item(),
torch.mean(
torch.abs(list(
self.encoder.parameters())[0].grad)).item(),
torch.mean(
torch.abs(list(
self.decoder.parameters())[0].grad)).item(),
torch.mean(
torch.abs(list(
self.classifier.parameters())[0].grad)).item()
])))
skip = 10
if b % skip == skip // 2:
log('[Training] %d | %d: %g (%g) %g (%g, %g, %g)' % (
epoch,
b,
numpy.mean(self.train_statistics[max(0, iteration -
skip):iteration, 4]),
numpy.mean(self.train_statistics[max(0, iteration -
skip):iteration, 5]),
numpy.mean(self.train_statistics[max(0, iteration -
skip):iteration, 6]),
numpy.mean(self.train_statistics[max(0, iteration -
skip):iteration, 7]),
numpy.mean(self.train_statistics[max(0, iteration -
skip):iteration, 8]),
numpy.mean(self.train_statistics[max(0, iteration -
skip):iteration, 9]),
))
log('[Training] %d | %d: %g %g (%g, %g, %g)' % (
epoch,
b,
numpy.mean(self.train_statistics[max(0, iteration -
skip):iteration, 10]),
numpy.mean(self.train_statistics[max(0, iteration -
skip):iteration, 11]),
numpy.mean(self.train_statistics[max(0, iteration -
skip):iteration, 12]),
numpy.mean(self.train_statistics[max(0, iteration -
skip):iteration, 13]),
numpy.mean(self.train_statistics[max(0, iteration -
skip):iteration, 14]),
))
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 train(self):
"""
Train adversarially.
"""
num_batches = int(
math.ceil(self.train_images.shape[0] / self.args.batch_size))
permutation = numpy.random.permutation(self.train_images.shape[0])
perturbation_permutation = numpy.random.permutation(
self.train_images.shape[0])
if self.args.safe:
perturbation_permutation = perturbation_permutation[
self.train_valid == 1]
else:
perturbation_permuation = permutation
for b in range(num_batches):
self.scheduler.update(self.epoch, float(b) / num_batches)
self.model.eval()
assert self.model.training is False
objective = self.objective_class()
split = self.args.batch_size // 2
if self.args.full_variant:
perm = numpy.concatenate(
(numpy.take(permutation,
range(b * self.args.batch_size,
b * self.args.batch_size + split),
mode='wrap'),
numpy.take(perturbation_permutation,
range(b * self.args.batch_size + split,
(b + 1) * self.args.batch_size),
mode='wrap')),
axis=0)
batch_images = common.torch.as_variable(
self.train_images[perm], self.args.use_gpu)
batch_classes = common.torch.as_variable(
self.train_codes[perm], self.args.use_gpu)
batch_theta = common.torch.as_variable(self.train_theta[perm],
self.args.use_gpu)
batch_images = batch_images.permute(0, 3, 1, 2)
attack = self.setup_attack(self.model, batch_images[:split],
batch_classes[:split])
success, perturbations, _, _, _ = attack.run(
objective, self.args.verbose)
batch_perturbations1 = common.torch.as_variable(
perturbations.astype(numpy.float32), self.args.use_gpu)
batch_perturbed_images1 = batch_images[:split] + batch_perturbations1
if isinstance(self.decoder, models.SelectiveDecoder):
self.decoder.set_code(batch_classes[split:])
attack = self.setup_decoder_attack(self.decoder_classifier,
batch_theta[split:],
batch_classes[split:])
attack.set_bound(torch.from_numpy(self.min_bound),
torch.from_numpy(self.max_bound))
decoder_success, decoder_perturbations, probabilities, norm, _ = attack.run(
objective, self.args.verbose)
batch_perturbed_theta = batch_theta[
split:] + common.torch.as_variable(decoder_perturbations,
self.args.use_gpu)
batch_perturbed_images2 = self.decoder(batch_perturbed_theta)
batch_perturbations2 = batch_perturbed_images2 - batch_images[
split:]
batch_input_images = torch.cat(
(batch_perturbed_images1, batch_perturbed_images2), dim=0)
self.model.train()
assert self.model.training is True
output_classes = self.model(batch_input_images)
self.scheduler.optimizer.zero_grad()
perturbation_loss = self.loss(batch_classes[:split],
output_classes[:split])
decoder_perturbation_loss = self.loss(batch_classes[split:],
output_classes[split:])
loss = (perturbation_loss + decoder_perturbation_loss) / 2
loss.backward()
self.scheduler.optimizer.step()
loss = loss.item()
perturbation_loss = perturbation_loss.item()
decoder_perturbation_loss = decoder_perturbation_loss.item()
gradient = torch.mean(
torch.abs(list(self.model.parameters())[0].grad))
gradient = gradient.item()
perturbation_error = self.error(batch_classes[:split],
output_classes[:split])
perturbation_error = perturbation_error.item()
decoder_perturbation_error = self.error(
batch_classes[split:], output_classes[split:])
decoder_perturbation_error = decoder_perturbation_error.item()
error = (perturbation_error + decoder_perturbation_error) / 2
else:
perm = numpy.concatenate((
numpy.take(
perturbation_permutation,
range(b * self.args.batch_size + split + split // 2,
(b + 1) * self.args.batch_size),
mode='wrap'),
numpy.take(
permutation,
range(b * self.args.batch_size,
b * self.args.batch_size + split + split // 2),
mode='wrap'),
),
axis=0)
batch_images = common.torch.as_variable(
self.train_images[perm], self.args.use_gpu)
batch_classes = common.torch.as_variable(
self.train_codes[perm], self.args.use_gpu)
batch_theta = common.torch.as_variable(self.train_theta[perm],
self.args.use_gpu)
batch_images = batch_images.permute(0, 3, 1, 2)
attack = self.setup_attack(self.model,
batch_images[split // 2:split],
batch_classes[split // 2:split])
success, perturbations, _, _, _ = attack.run(
objective, self.args.verbose)
batch_perturbations1 = common.torch.as_variable(
perturbations.astype(numpy.float32), self.args.use_gpu)
batch_perturbed_images1 = batch_images[
split // 2:split] + batch_perturbations1
if isinstance(self.decoder, models.SelectiveDecoder):
self.decoder.set_code(batch_classes[:split // 2])
attack = self.setup_decoder_attack(self.decoder_classifier,
batch_theta[:split // 2],
batch_classes[:split // 2])
attack.set_bound(torch.from_numpy(self.min_bound),
torch.from_numpy(self.max_bound))
decoder_success, decoder_perturbations, probabilities, norm, _ = attack.run(
objective, self.args.verbose)
batch_perturbed_theta = batch_theta[:split //
2] + common.torch.as_variable(
decoder_perturbations,
self.args.use_gpu)
batch_perturbed_images2 = self.decoder(batch_perturbed_theta)
batch_perturbations2 = batch_perturbed_images2 - batch_images[:split
//
2]
batch_input_images = torch.cat(
(batch_perturbed_images2, batch_perturbed_images1,
batch_images[split:]),
dim=0)
self.model.train()
assert self.model.training is True
output_classes = self.model(batch_input_images)
self.scheduler.optimizer.zero_grad()
loss = self.loss(batch_classes[split:], output_classes[split:])
perturbation_loss = self.loss(batch_classes[split // 2:split],
output_classes[split // 2:split])
decoder_perturbation_loss = self.loss(
batch_classes[:split // 2], output_classes[:split // 2])
l = (loss + perturbation_loss + decoder_perturbation_loss) / 3
l.backward()
self.scheduler.optimizer.step()
loss = loss.item()
perturbation_loss = perturbation_loss.item()
decoder_perturbation_loss = decoder_perturbation_loss.item()
gradient = torch.mean(
torch.abs(list(self.model.parameters())[0].grad))
gradient = gradient.item()
error = self.error(batch_classes[split:],
output_classes[split:])
error = error.item()
perturbation_error = self.error(
batch_classes[split // 2:split],
output_classes[split // 2:split])
perturbation_error = perturbation_error.item()
decoder_perturbation_error = self.error(
batch_classes[:split // 2], output_classes[:split // 2])
decoder_perturbation_error = decoder_perturbation_error.item()
iterations = numpy.mean(
success[success >= 0]) if numpy.sum(success >= 0) > 0 else -1
norm = numpy.mean(
numpy.linalg.norm(perturbations.reshape(
perturbations.shape[0], -1),
axis=1,
ord=self.norm))
success = numpy.sum(success >= 0) / self.args.batch_size
decoder_iterations = numpy.mean(
decoder_success[decoder_success >= 0]) if numpy.sum(
decoder_success >= 0) > 0 else -1
decoder_norm = numpy.mean(
numpy.linalg.norm(decoder_perturbations, axis=1,
ord=self.norm))
decoder_success = numpy.sum(
decoder_success >= 0) / self.args.batch_size
iteration = self.epoch * num_batches + b + 1
self.train_statistics = numpy.vstack((
self.train_statistics,
numpy.array([[
iteration, # iterations
iteration * (1 + self.args.max_iterations) *
self.args.batch_size, # samples seen
min(num_batches, iteration) * self.args.batch_size +
iteration * self.args.max_iterations *
self.args.batch_size, # unique samples seen
loss,
error,
perturbation_loss,
perturbation_error,
decoder_perturbation_loss,
decoder_perturbation_error,
success,
iterations,
norm,
decoder_success,
decoder_iterations,
decoder_norm,
gradient
]])))
if b % self.args.skip == self.args.skip // 2:
log('[Training] %d | %d: %g (%g) %g (%g) %g (%g) [%g]' % (
self.epoch,
b,
numpy.mean(self.train_statistics[
max(0, iteration - self.args.skip):iteration, 3]),
numpy.mean(self.train_statistics[
max(0, iteration - self.args.skip):iteration, 4]),
numpy.mean(self.train_statistics[
max(0, iteration - self.args.skip):iteration, 5]),
numpy.mean(self.train_statistics[
max(0, iteration - self.args.skip):iteration, 6]),
numpy.mean(self.train_statistics[
max(0, iteration - self.args.skip):iteration, 7]),
numpy.mean(self.train_statistics[
max(0, iteration - self.args.skip):iteration, 8]),
numpy.mean(self.train_statistics[
max(0, iteration - self.args.skip):iteration, -1]),
))
log('[Training] %d | %d: %g (%g, %g) %g (%g, %g)' % (
self.epoch,
b,
numpy.mean(self.train_statistics[
max(0, iteration - self.args.skip):iteration, 9]),
numpy.mean(self.train_statistics[
max(0, iteration - self.args.skip):iteration, 10]),
numpy.mean(self.train_statistics[
max(0, iteration - self.args.skip):iteration, 11]),
numpy.mean(self.train_statistics[
max(0, iteration - self.args.skip):iteration, 12]),
numpy.mean(self.train_statistics[
max(0, iteration - self.args.skip):iteration, 13]),
numpy.mean(self.train_statistics[
max(0, iteration - self.args.skip):iteration, 14]),
))
self.debug('clean.%d.png' % self.epoch,
batch_images.permute(0, 2, 3, 1))
self.debug('perturbed.%d.png' % self.epoch,
batch_perturbed_images1.permute(0, 2, 3, 1))
self.debug('perturbed2.%d.png' % self.epoch,
batch_perturbed_images2.permute(0, 2, 3, 1))
self.debug('perturbation.%d.png' % self.epoch,
batch_perturbations1.permute(0, 2, 3, 1),
cmap='seismic')
self.debug('perturbation2.%d.png' % self.epoch,
batch_perturbations2.permute(0, 2, 3, 1),
cmap='seismic')
def test(self):
"""
Test the model.
"""
self.model.eval()
assert self.model.training is False
log('[Training] %d set classifier to eval' % self.epoch)
loss = error = 0
num_batches = int(
math.ceil(self.args.test_samples / self.args.batch_size))
for b in range(num_batches):
perm = numpy.take(range(self.args.test_samples),
range(b * self.args.batch_size,
(b + 1) * self.args.batch_size),
mode='clip')
batch_images = common.torch.as_variable(self.test_images[perm],
self.args.use_gpu)
batch_classes = common.torch.as_variable(self.test_codes[perm],
self.args.use_gpu)
batch_images = batch_images.permute(0, 3, 1, 2)
output_classes = self.model(batch_images)
e = self.loss(batch_classes, output_classes)
loss += e.item()
a = self.error(batch_classes, output_classes)
error += a.item()
perturbation_loss = perturbation_error = success = iterations = norm = 0
num_batches = int(
math.ceil(self.args.attack_samples / self.args.batch_size))
assert self.args.attack_samples > 0 and self.args.attack_samples <= self.test_images.shape[
0]
for b in range(num_batches):
perm = numpy.take(range(self.args.attack_samples),
range(b * self.args.batch_size,
(b + 1) * self.args.batch_size),
mode='clip')
batch_images = common.torch.as_variable(self.test_images[perm],
self.args.use_gpu)
batch_classes = common.torch.as_variable(self.test_codes[perm],
self.args.use_gpu)
batch_images = batch_images.permute(0, 3, 1, 2)
objective = self.objective_class()
attack = self.setup_attack(self.model, batch_images, batch_classes)
s, p, _, _, _ = attack.run(objective, False)
batch_images = batch_images + common.torch.as_variable(
p.astype(numpy.float32), self.args.use_gpu)
output_classes = self.model(batch_images)
e = self.loss(batch_classes, output_classes)
perturbation_loss += e.item()
e = self.error(batch_classes, output_classes)
perturbation_error += e.item()
iterations += numpy.mean(
s[s >= 0]) if numpy.sum(s >= 0) > 0 else -1
norm += numpy.mean(
numpy.linalg.norm(p.reshape(p.shape[0], -1),
axis=1,
ord=self.norm))
success += numpy.sum(s >= 0) / self.args.batch_size
decoder_perturbation_loss = decoder_perturbation_error = decoder_success = decoder_iterations = decoder_norm = 0
num_batches = int(
math.ceil(self.args.attack_samples / self.args.batch_size))
assert self.args.attack_samples > 0 and self.args.attack_samples <= self.test_images.shape[
0]
for b in range(num_batches):
perm = numpy.take(range(self.args.attack_samples),
range(b * self.args.batch_size,
(b + 1) * self.args.batch_size),
mode='clip')
batch_theta = common.torch.as_variable(self.test_theta[perm],
self.args.use_gpu)
batch_classes = common.torch.as_variable(self.test_codes[perm],
self.args.use_gpu)
objective = self.objective_class()
if isinstance(self.decoder, models.SelectiveDecoder):
self.decoder.set_code(batch_classes)
attack = self.setup_decoder_attack(self.decoder_classifier,
batch_theta, batch_classes)
attack.set_bound(torch.from_numpy(self.min_bound),
torch.from_numpy(self.max_bound))
s, p, _, _, _ = attack.run(objective, False)
perturbations = common.torch.as_variable(p, self.args.use_gpu)
batch_perturbed_theta = batch_theta + perturbations
batch_perturbed_images = self.decoder(batch_perturbed_theta)
output_classes = self.model(batch_perturbed_images)
e = self.loss(batch_classes, output_classes)
perturbation_loss += e.item()
a = self.error(batch_classes, output_classes)
perturbation_error += a.item()
decoder_iterations += numpy.mean(
s[s >= 0]) if numpy.sum(s >= 0) > 0 else -1
decoder_norm += numpy.mean(
numpy.linalg.norm(p.reshape(p.shape[0], -1),
axis=1,
ord=self.norm))
decoder_success += numpy.sum(s >= 0) / self.args.batch_size
loss /= num_batches
error /= num_batches
perturbation_loss /= num_batches
perturbation_error /= num_batches
success /= num_batches
iterations /= num_batches
norm /= num_batches
decoder_perturbation_loss /= num_batches
decoder_perturbation_error /= num_batches
decoder_success /= num_batches
decoder_iterations /= num_batches
decoder_norm /= num_batches
log('[Training] %d: test %g (%g) %g (%g) %g (%g)' %
(self.epoch, loss, error, perturbation_loss, perturbation_error,
decoder_perturbation_loss, decoder_perturbation_error))
log('[Training] %d: test %g (%g, %g) %g (%g, %g)' %
(self.epoch, success, iterations, norm, decoder_success,
decoder_iterations, decoder_norm))
num_batches = int(
math.ceil(self.train_images.shape[0] / self.args.batch_size))
iteration = self.epoch * num_batches
self.test_statistics = numpy.vstack((
self.test_statistics,
numpy.array([[
iteration, # iterations
iteration * (1 + self.args.max_iterations) *
self.args.batch_size, # samples seen
min(num_batches, iteration) * self.args.batch_size +
iteration * self.args.max_iterations *
self.args.batch_size, # unique samples seen
loss,
error,
perturbation_loss,
perturbation_error,
decoder_perturbation_loss,
decoder_perturbation_error,
success,
iterations,
norm,
decoder_success,
decoder_iterations,
decoder_norm,
]])))
def train(self, epoch):
"""
Train for one epoch.
:param epoch: current epoch
:type epoch: int
"""
assert self.encoder is not None and self.decoder is not None
assert self.scheduler is not None
self.auto_encoder.train()
log('[Training] %d set auto encoder to train' % epoch)
self.encoder.train()
log('[Training] %d set encoder to train' % epoch)
self.decoder.train()
log('[Training] %d set decoder to train' % epoch)
num_batches = int(math.ceil(self.train_images.shape[0]/self.args.batch_size))
assert self.encoder.training is True
permutation = numpy.random.permutation(self.train_images.shape[0])
permutation = numpy.concatenate((permutation, permutation[:self.args.batch_size]), axis=0)
for b in range(num_batches):
self.scheduler.update(epoch, float(b)/num_batches)
perm = permutation[b * self.args.batch_size: (b + 1) * self.args.batch_size]
batch_images = common.torch.as_variable(self.train_images[perm], self.args.use_gpu)
batch_images = batch_images.permute(0, 3, 1, 2)
output_images, output_mu, output_logvar = self.auto_encoder(batch_images)
reconstruction_loss = self.reconstruction_loss(batch_images, output_images)
self.scheduler.optimizer.zero_grad()
latent_loss = self.latent_loss(output_mu, output_logvar)
loss = self.args.beta*reconstruction_loss + latent_loss
loss.backward()
self.scheduler.optimizer.step()
reconstruction_loss = reconstruction_loss.item()
latent_loss = latent_loss.item()
reconstruction_error = self.reconstruction_error(batch_images, output_images)
reconstruction_error = reconstruction_error.item()
iteration = epoch*num_batches + b + 1
self.train_statistics = numpy.vstack((self.train_statistics, numpy.array([
iteration,
iteration * self.args.batch_size,
min(num_batches, iteration),
min(num_batches, iteration) * self.args.batch_size,
reconstruction_loss,
reconstruction_error,
latent_loss,
torch.mean(output_mu).item(),
torch.var(output_mu).item(),
torch.mean(output_logvar).item(),
])))
skip = 10
if b%skip == skip//2:
log('[Training] %d | %d: %g (%g) %g %g %g %g' % (
epoch,
b,
numpy.mean(self.train_statistics[max(0, iteration-skip):iteration, 4]),
numpy.mean(self.train_statistics[max(0, iteration-skip):iteration, 5]),
numpy.mean(self.train_statistics[max(0, iteration-skip):iteration, 6]),
numpy.mean(self.train_statistics[max(0, iteration-skip):iteration, 7]),
numpy.mean(self.train_statistics[max(0, iteration-skip):iteration, 8]),
numpy.mean(self.train_statistics[max(0, iteration-skip):iteration, 9]),
))
def train(self):
"""
Train with fair data augmentation.
"""
self.model.train()
assert self.model.training is True
assert self.decoder.training is False
split = self.args.batch_size // 2
num_batches = int(math.ceil(self.train_images.shape[0] / self.args.batch_size))
permutation = numpy.random.permutation(self.train_images.shape[0])
for b in range(num_batches):
self.scheduler.update(self.epoch, float(b) / num_batches)
perm = numpy.take(permutation, range(b*self.args.batch_size, (b+1)*self.args.batch_size), mode='wrap')
batch_images = common.torch.as_variable(self.train_images[perm], self.args.use_gpu)
batch_theta = common.torch.as_variable(self.train_theta[perm], self.args.use_gpu)
batch_classes = common.torch.as_variable(self.train_codes[perm], self.args.use_gpu)
batch_images = batch_images.permute(0, 3, 1, 2)
loss = error = gradient = 0
if self.args.full_variant:
for t in range(self.args.max_iterations):
if self.args.strong_variant:
# Here we want to sample form a truncated Gaussian
random = common.numpy.truncated_normal(batch_theta.size(), lower=-self.args.bound, upper=self.args.bound)
batch_perturbed_theta = common.torch.as_variable(random.astype(numpy.float32), self.args.use_gpu)
if isinstance(self.decoder, models.SelectiveDecoder):
self.decoder.set_code(batch_classes)
batch_perturbed_images = self.decoder(batch_perturbed_theta)
else:
random = common.numpy.uniform_ball(batch_theta.size(0), batch_theta.size(1), epsilon=self.args.epsilon, ord=self.norm)
batch_perturbed_theta = batch_theta + common.torch.as_variable(random.astype(numpy.float32), self.args.use_gpu)
batch_perturbed_theta = torch.min(common.torch.as_variable(self.max_bound, self.args.use_gpu), batch_perturbed_theta)
batch_perturbed_theta = torch.max(common.torch.as_variable(self.min_bound, self.args.use_gpu), batch_perturbed_theta)
if isinstance(self.decoder, models.SelectiveDecoder):
self.decoder.set_code(batch_classes)
batch_perturbed_images = self.decoder(batch_perturbed_theta)
output_classes = self.model(batch_perturbed_images)
self.scheduler.optimizer.zero_grad()
l = self.loss(batch_classes, output_classes)
l.backward()
self.scheduler.optimizer.step()
loss += l.item()
g = torch.mean(torch.abs(list(self.model.parameters())[0].grad))
gradient += g.item()
e = self.error(batch_classes, output_classes)
error += e.item()
batch_perturbations = batch_perturbed_images - batch_images
gradient /= self.args.max_iterations
loss /= self.args.max_iterations
error /= self.args.max_iterations
perturbation_loss = loss
perturbation_error = error
else:
output_classes = self.model(batch_images[:split])
self.scheduler.optimizer.zero_grad()
l = self.loss(batch_classes[:split], output_classes)
l.backward()
self.scheduler.optimizer.step()
loss = l.item()
gradient = torch.mean(torch.abs(list(self.model.parameters())[0].grad))
gradient = gradient.item()
e = self.error(batch_classes[:split], output_classes)
error = e.item()
perturbation_loss = perturbation_error = 0
for t in range(self.args.max_iterations):
if self.args.strong_variant:
# Here we want to sample form a truncated Gaussian
random = common.numpy.truncated_normal([split, batch_theta.size(1)], lower=-self.args.bound, upper=self.args.bound)
batch_perturbed_theta = common.torch.as_variable(random.astype(numpy.float32), self.args.use_gpu)
if isinstance(self.decoder, models.SelectiveDecoder):
self.decoder.set_code(batch_classes[split:])
batch_perturbed_images = self.decoder(batch_perturbed_theta)
else:
random = common.numpy.uniform_ball(split, batch_theta.size(1), epsilon=self.args.epsilon, ord=self.norm)
batch_perturbed_theta = batch_theta[split:] + common.torch.as_variable(random.astype(numpy.float32), self.args.use_gpu)
batch_perturbed_theta = torch.min(common.torch.as_variable(self.max_bound, self.args.use_gpu), batch_perturbed_theta)
batch_perturbed_theta = torch.max(common.torch.as_variable(self.min_bound, self.args.use_gpu), batch_perturbed_theta)
if isinstance(self.decoder, models.SelectiveDecoder):
self.decoder.set_code(batch_classes[split:])
batch_perturbed_images = self.decoder(batch_perturbed_theta)
output_classes = self.model(batch_perturbed_images)
self.scheduler.optimizer.zero_grad()
l = self.loss(batch_classes[split:], output_classes)
l.backward()
self.scheduler.optimizer.step()
perturbation_loss += l.item()
g = torch.mean(torch.abs(list(self.model.parameters())[0].grad))
gradient += g.item()
e = self.error(batch_classes[split:], output_classes)
perturbation_error += e.item()
batch_perturbations = batch_perturbed_images - batch_images[split:]
gradient /= self.args.max_iterations + 1
perturbation_loss /= self.args.max_iterations
perturbation_error /= self.args.max_iterations
iteration = self.epoch * num_batches + b + 1
self.train_statistics = numpy.vstack((self.train_statistics, numpy.array([[
iteration, # iterations
iteration * (1 + self.args.max_iterations) * self.args.batch_size, # samples seen
min(num_batches, iteration) * self.args.batch_size + iteration * self.args.max_iterations * self.args.batch_size, # unique samples seen
loss,
error,
perturbation_loss,
perturbation_error,
gradient
]])))
if b % self.args.skip == self.args.skip // 2:
log('[Training] %d | %d: %g (%g) %g (%g) [%g]' % (
self.epoch,
b,
numpy.mean(self.train_statistics[max(0, iteration - self.args.skip):iteration, 3]),
numpy.mean(self.train_statistics[max(0, iteration - self.args.skip):iteration, 4]),
numpy.mean(self.train_statistics[max(0, iteration - self.args.skip):iteration, 5]),
numpy.mean(self.train_statistics[max(0, iteration - self.args.skip):iteration, 6]),
numpy.mean(self.train_statistics[max(0, iteration - self.args.skip):iteration, -1]),
))
self.debug('clean.%d.png' % self.epoch, batch_images.permute(0, 2, 3, 1))
self.debug('perturbed.%d.png' % self.epoch, batch_perturbed_images.permute(0, 2, 3, 1))
self.debug('perturbation.%d.png' % self.epoch, batch_perturbations.permute(0, 2, 3, 1), cmap='seismic')
def train(self):
"""
Train adversarially.
"""
split = self.args.batch_size // 2
num_batches = int(
math.ceil(self.train_images.shape[0] / self.args.batch_size))
permutation = numpy.random.permutation(self.train_images.shape[0])
for b in range(num_batches):
self.scheduler.update(self.epoch, float(b) / num_batches)
perm = numpy.take(permutation,
range(b * self.args.batch_size,
(b + 1) * self.args.batch_size),
mode='wrap')
batch_images = common.torch.as_variable(self.train_images[perm],
self.args.use_gpu)
batch_theta = common.torch.as_variable(self.train_theta[perm],
self.args.use_gpu)
batch_images = batch_images.permute(0, 3, 1, 2)
batch_fonts = self.train_codes[perm, 1]
batch_classes = self.train_codes[perm, self.args.label_index]
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_code = common.torch.as_variable(batch_code,
self.args.use_gpu)
batch_classes = common.torch.as_variable(batch_classes,
self.args.use_gpu)
self.model.eval()
assert self.model.training is False
if self.args.full_variant:
objective = self.objective_class()
self.decoder.set_code(batch_code)
attack = self.setup_attack(self.decoder_classifier,
batch_theta, batch_classes)
attack.set_bound(torch.from_numpy(self.min_bound),
torch.from_numpy(self.max_bound))
success, perturbations, probabilities, norm, _ = attack.run(
objective, self.args.verbose)
batch_perturbed_theta = batch_theta + common.torch.as_variable(
perturbations, self.args.use_gpu)
batch_perturbed_images = self.decoder(batch_perturbed_theta)
batch_perturbations = batch_perturbed_images - batch_images
self.model.train()
assert self.model.training is True
output_classes = self.model(batch_perturbed_images)
self.scheduler.optimizer.zero_grad()
loss = self.loss(batch_classes, output_classes)
loss.backward()
self.scheduler.optimizer.step()
loss = perturbation_loss = loss.item()
gradient = torch.mean(
torch.abs(list(self.model.parameters())[0].grad))
gradient = gradient.item()
error = self.error(batch_classes, output_classes)
error = perturbation_error = error.item()
else:
objective = self.objective_class()
self.decoder.set_code(batch_code[split:])
attack = self.setup_attack(self.decoder_classifier,
batch_theta[split:],
batch_classes[split:])
attack.set_bound(torch.from_numpy(self.min_bound),
torch.from_numpy(self.max_bound))
success, perturbations, probabilities, norm, _ = attack.run(
objective, self.args.verbose)
batch_perturbed_theta = batch_theta[
split:] + common.torch.as_variable(perturbations,
self.args.use_gpu)
batch_perturbed_images = self.decoder(batch_perturbed_theta)
batch_perturbations = batch_perturbed_images - batch_images[
split:]
self.model.train()
assert self.model.training is True
batch_input_images = torch.cat(
(batch_images[:split], batch_perturbed_images), dim=0)
output_classes = self.model(batch_input_images)
self.scheduler.optimizer.zero_grad()
loss = self.loss(batch_classes[:split], output_classes[:split])
perturbation_loss = self.loss(batch_classes[split:],
output_classes[split:])
l = (loss + perturbation_loss) / 2
l.backward()
self.scheduler.optimizer.step()
loss = loss.item()
perturbation_loss = perturbation_loss.item()
gradient = torch.mean(
torch.abs(list(self.model.parameters())[0].grad))
gradient = gradient.item()
error = self.error(batch_classes[:split],
output_classes[:split])
error = error.item()
perturbation_error = self.error(batch_classes[split:],
output_classes[split:])
perturbation_error = perturbation_error.item()
iterations = numpy.mean(
success[success >= 0]) if numpy.sum(success >= 0) > 0 else -1
norm = numpy.mean(
numpy.linalg.norm(perturbations.reshape(
perturbations.shape[0], -1),
axis=1,
ord=self.norm))
success = numpy.sum(success >= 0) / (self.args.batch_size // 2)
iteration = self.epoch * num_batches + b + 1
self.train_statistics = numpy.vstack((
self.train_statistics,
numpy.array([[
iteration, # iterations
iteration * (1 + self.args.max_iterations) *
self.args.batch_size, # samples seen
min(num_batches, iteration) * self.args.batch_size +
iteration * self.args.max_iterations *
self.args.batch_size, # unique samples seen
loss,
error,
perturbation_loss,
perturbation_error,
success,
iterations,
norm,
gradient
]])))
if b % self.args.skip == self.args.skip // 2:
log('[Training] %d | %d: %g (%g) %g (%g) [%g]' % (
self.epoch,
b,
numpy.mean(self.train_statistics[
max(0, iteration - self.args.skip):iteration, 3]),
numpy.mean(self.train_statistics[
max(0, iteration - self.args.skip):iteration, 4]),
numpy.mean(self.train_statistics[
max(0, iteration - self.args.skip):iteration, 5]),
numpy.mean(self.train_statistics[
max(0, iteration - self.args.skip):iteration, 6]),
numpy.mean(self.train_statistics[
max(0, iteration - self.args.skip):iteration, -1]),
))
log('[Training] %d | %d: %g (%g, %g)' % (
self.epoch,
b,
numpy.mean(self.train_statistics[
max(0, iteration - self.args.skip):iteration, 7]),
numpy.mean(self.train_statistics[
max(0, iteration - self.args.skip):iteration, 8]),
numpy.mean(self.train_statistics[
max(0, iteration - self.args.skip):iteration, 9]),
))
self.debug('clean.%d.png' % self.epoch,
batch_images.permute(0, 2, 3, 1))
self.debug('perturbed.%d.png' % self.epoch,
batch_perturbed_images.permute(0, 2, 3, 1))
self.debug('perturbation.%d.png' % self.epoch,
batch_perturbations.permute(0, 2, 3, 1),
cmap='seismic')
def train(self):
"""
Train with fair data augmentation.
"""
self.model.train()
assert self.model.training is True
split = self.args.batch_size // 2
num_batches = int(
math.ceil(self.train_images.shape[0] / self.args.batch_size))
permutation = numpy.random.permutation(self.train_images.shape[0])
for b in range(num_batches):
self.scheduler.update(self.epoch, float(b) / num_batches)
perm = numpy.take(permutation,
range(b * self.args.batch_size,
(b + 1) * self.args.batch_size),
mode='wrap')
batch_images = common.torch.as_variable(self.train_images[perm],
self.args.use_gpu)
batch_theta = common.torch.as_variable(self.train_theta[perm],
self.args.use_gpu)
batch_images = batch_images.permute(0, 3, 1, 2)
batch_fonts = self.train_codes[perm, 1]
batch_classes = self.train_codes[perm, self.args.label_index]
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_code = common.torch.as_variable(batch_code,
self.args.use_gpu)
batch_classes = common.torch.as_variable(batch_classes,
self.args.use_gpu)
loss = error = gradient = 0
if self.args.full_variant:
for t in range(self.args.max_iterations):
if self.args.strong_variant:
# Here, we want to uniformly sample all allowed transformations, so that's OK.
min_bound = numpy.repeat(self.min_bound.reshape(1, -1),
self.args.batch_size,
axis=0)
max_bound = numpy.repeat(self.max_bound.reshape(1, -1),
self.args.batch_size,
axis=0)
random = numpy.random.uniform(
min_bound, max_bound,
(batch_theta.size(0), batch_theta.size(1)))
batch_perturbed_theta = common.torch.as_variable(
random.astype(numpy.float32), self.args.use_gpu)
self.decoder.set_code(batch_code)
batch_perturbed_images = self.decoder(
batch_perturbed_theta)
else:
random = common.numpy.uniform_ball(
batch_theta.size(0),
batch_theta.size(1),
epsilon=self.args.epsilon,
ord=self.norm)
batch_perturbed_theta = batch_theta + common.torch.as_variable(
random.astype(numpy.float32), self.args.use_gpu)
batch_perturbed_theta = torch.min(
common.torch.as_variable(self.max_bound,
self.args.use_gpu),
batch_perturbed_theta)
batch_perturbed_theta = torch.max(
common.torch.as_variable(self.min_bound,
self.args.use_gpu),
batch_perturbed_theta)
self.decoder.set_code(batch_code)
batch_perturbed_images = self.decoder(
batch_perturbed_theta)
output_classes = self.model(batch_perturbed_images)
self.scheduler.optimizer.zero_grad()
l = self.loss(batch_classes, output_classes)
l.backward()
self.scheduler.optimizer.step()
loss += l.item()
g = torch.mean(
torch.abs(list(self.model.parameters())[0].grad))
gradient += g.item()
e = self.error(batch_classes, output_classes)
error += e.item()
batch_perturbations = batch_perturbed_images - batch_images
gradient /= self.args.max_iterations
loss /= self.args.max_iterations
error /= self.args.max_iterations
perturbation_loss = loss
perturbation_error = error
else:
output_classes = self.model(batch_images[:split])
self.scheduler.optimizer.zero_grad()
l = self.loss(batch_classes[:split], output_classes)
l.backward()
self.scheduler.optimizer.step()
loss = l.item()
gradient = torch.mean(
torch.abs(list(self.model.parameters())[0].grad))
gradient = gradient.item()
e = self.error(batch_classes[:split], output_classes)
error = e.item()
perturbation_loss = perturbation_error = 0
for t in range(self.args.max_iterations):
if self.args.strong_variant:
# Again, sampling all possible transformations.
min_bound = numpy.repeat(self.min_bound.reshape(1, -1),
split,
axis=0)
max_bound = numpy.repeat(self.max_bound.reshape(1, -1),
split,
axis=0)
random = numpy.random.uniform(
min_bound, max_bound, (split, batch_theta.size(1)))
batch_perturbed_theta = common.torch.as_variable(
random.astype(numpy.float32), self.args.use_gpu)
self.decoder.set_code(batch_code[split:])
batch_perturbed_images = self.decoder(
batch_perturbed_theta)
else:
random = common.numpy.uniform_ball(
split,
batch_theta.size(1),
epsilon=self.args.epsilon,
ord=self.norm)
batch_perturbed_theta = batch_theta[
split:] + common.torch.as_variable(
random.astype(numpy.float32),
self.args.use_gpu)
batch_perturbed_theta = torch.min(
common.torch.as_variable(self.max_bound,
self.args.use_gpu),
batch_perturbed_theta)
batch_perturbed_theta = torch.max(
common.torch.as_variable(self.min_bound,
self.args.use_gpu),
batch_perturbed_theta)
self.decoder.set_code(batch_code[split:])
batch_perturbed_images = self.decoder(
batch_perturbed_theta)
output_classes = self.model(batch_perturbed_images)
self.scheduler.optimizer.zero_grad()
l = self.loss(batch_classes[split:], output_classes)
l.backward()
self.scheduler.optimizer.step()
perturbation_loss += l.item()
g = torch.mean(
torch.abs(list(self.model.parameters())[0].grad))
gradient += g.item()
e = self.error(batch_classes[split:], output_classes)
perturbation_error += e.item()
batch_perturbations = batch_perturbed_images - batch_images[
split:]
gradient /= self.args.max_iterations + 1
perturbation_loss /= self.args.max_iterations
perturbation_error /= self.args.max_iterations
iteration = self.epoch * num_batches + b + 1
self.train_statistics = numpy.vstack((
self.train_statistics,
numpy.array([[
iteration, # iterations
iteration * (1 + self.args.max_iterations) *
self.args.batch_size, # samples seen
min(num_batches, iteration) * self.args.batch_size +
iteration * self.args.max_iterations *
self.args.batch_size, # unique samples seen
loss,
error,
perturbation_loss,
perturbation_error,
gradient
]])))
if b % self.args.skip == self.args.skip // 2:
log('[Training] %d | %d: %g (%g) %g (%g) [%g]' % (
self.epoch,
b,
numpy.mean(self.train_statistics[
max(0, iteration - self.args.skip):iteration, 3]),
numpy.mean(self.train_statistics[
max(0, iteration - self.args.skip):iteration, 4]),
numpy.mean(self.train_statistics[
max(0, iteration - self.args.skip):iteration, 5]),
numpy.mean(self.train_statistics[
max(0, iteration - self.args.skip):iteration, 6]),
numpy.mean(self.train_statistics[
max(0, iteration - self.args.skip):iteration, -1]),
))
self.debug('clean.%d.png' % self.epoch,
batch_images.permute(0, 2, 3, 1))
self.debug('perturbed.%d.png' % self.epoch,
batch_perturbed_images.permute(0, 2, 3, 1))
self.debug('perturbation.%d.png' % self.epoch,
batch_perturbations.permute(0, 2, 3, 1),
cmap='seismic')
def test(self):
"""
Test the model.
"""
self.model.eval()
log('[Training] %d set classifier to eval' % self.epoch)
assert self.model.training is False
loss = error = perturbation_loss = perturbation_error = success = iterations = norm = 0
num_batches = int(math.ceil(self.args.test_samples/self.args.batch_size))
for b in range(num_batches):
perm = numpy.take(range(self.args.test_samples), range(b*self.args.batch_size, (b+1)*self.args.batch_size), mode='clip')
batch_images = common.torch.as_variable(self.test_images[perm], self.args.use_gpu)
batch_classes = common.torch.as_variable(self.test_codes[perm], self.args.use_gpu)
batch_images = batch_images.permute(0, 3, 1, 2)
output_classes = self.model(batch_images)
e = self.loss(batch_classes, output_classes)
loss += e.data # 0-dim tensor
a = self.error(batch_classes, output_classes)
error += a.data
loss /= num_batches
error /= num_batches
num_batches = int(math.ceil(self.args.attack_samples/self.args.batch_size))
assert self.args.attack_samples > 0 and self.args.attack_samples <= self.test_images.shape[0]
for b in range(num_batches):
perm = numpy.take(range(self.args.attack_samples), range(b*self.args.batch_size, (b+1)*self.args.batch_size), mode='clip')
batch_images = common.torch.as_variable(self.test_images[perm], self.args.use_gpu)
batch_classes = common.torch.as_variable(self.test_codes[perm], self.args.use_gpu)
batch_images = batch_images.permute(0, 3, 1, 2)
objective = self.objective_class()
attack = self.setup_attack(self.model, batch_images, batch_classes)
s, p, _, _, _ = attack.run(objective, False)
batch_images = batch_images + common.torch.as_variable(p.astype(numpy.float32), self.args.use_gpu)
output_classes = self.model(batch_images)
e = self.loss(batch_classes, output_classes)
perturbation_loss += e.item()
e = self.error(batch_classes, output_classes)
perturbation_error += e.item()
iterations += numpy.mean(s[s >= 0]) if numpy.sum(s >= 0) > 0 else -1
norm += numpy.mean(numpy.linalg.norm(p.reshape(p.shape[0], -1), axis=1, ord=self.norm))
success += numpy.sum(s >= 0)/self.args.batch_size
perturbation_error /= num_batches
perturbation_loss /= num_batches
success /= num_batches
iterations /= num_batches
norm /= num_batches
log('[Training] %d: test %g (%g) %g (%g)' % (self.epoch, loss, error, perturbation_loss, perturbation_error))
log('[Training] %d: test %g (%g, %g)' % (self.epoch, success, iterations, norm))
num_batches = int(math.ceil(self.train_images.shape[0]/self.args.batch_size))
iteration = self.epoch*num_batches
self.test_statistics = numpy.vstack((self.test_statistics, numpy.array([[
iteration,
iteration * self.args.batch_size,
min(num_batches, iteration) * self.args.batch_size,
loss,
error,
perturbation_loss,
perturbation_error,
success,
iterations,
norm
]])))