def test_interpolation(self):
"""
Test interpolation.
"""
interpolations = None
perm = numpy.random.permutation(
numpy.array(range(self.pred_codes.shape[0])))
for i in range(50):
first = self.pred_codes[i]
second = self.pred_codes[perm[i]]
linfit = scipy.interpolate.interp1d([0, 1],
numpy.vstack([first, second]),
axis=0)
interpolations = common.numpy.concatenate(
interpolations, linfit(numpy.linspace(0, 1, 10)))
pred_images = None
num_batches = int(
math.ceil(interpolations.shape[0] / self.args.batch_size))
interpolations = interpolations.astype(numpy.float32)
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_codes = common.torch.as_variable(
interpolations[b_start:b_end], self.args.use_gpu)
# To get the correct images!
output_images = self.decoder(batch_codes)
output_images = numpy.squeeze(
numpy.transpose(output_images.cpu().detach().numpy(),
(0, 2, 3, 1)))
pred_images = common.numpy.concatenate(pred_images, output_images)
if b % 100 == 50:
log('[Testing] %d' % b)
utils.write_hdf5(self.args.interpolation_file, pred_images)
log('[Testing] wrote %s' % self.args.interpolation_file)
def test(self, epoch):
"""
Test the model.
:param epoch: current epoch
:type epoch: int
"""
self.encoder.eval()
log('[Training] %d set encoder to eval' % epoch)
self.decoder.eval()
log('[Training] %d set decoder to eval' % epoch)
self.classifier.eval()
log('[Training] %d set classifier to eval' % epoch)
latent_loss = 0
reconstruction_loss = 0
reconstruction_error = 0
decoder_loss = 0
discriminator_loss = 0
mean = 0
var = 0
logvar = 0
pred_images = None
pred_codes = None
num_batches = int(
math.ceil(self.test_images.shape[0] / self.args.batch_size))
assert self.encoder.training is False
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_images = batch_images.permute(0, 3, 1, 2)
output_mu, output_logvar = self.encoder(batch_images)
output_images = self.decoder(output_mu)
output_real_classes = self.classifier(batch_images)
output_reconstructed_classes = self.classifier(output_images)
# Latent loss.
e = self.latent_loss(output_mu, output_logvar)
latent_loss += e.item()
# Reconstruction loss.
e = self.reconstruction_loss(batch_images, output_images)
reconstruction_loss += e.item()
# Reconstruction error.
e = self.reconstruction_error(batch_images, output_images)
reconstruction_error += e.item()
e = self.decoder_loss(output_reconstructed_classes)
decoder_loss += e.item()
# Adversarial loss.
e = self.discriminator_loss(output_real_classes,
output_reconstructed_classes)
discriminator_loss += e.item()
mean += torch.mean(output_mu).item()
var += torch.var(output_mu).item()
logvar += torch.mean(output_logvar).item()
output_images = numpy.squeeze(
numpy.transpose(output_images.cpu().detach().numpy(),
(0, 2, 3, 1)))
pred_images = common.numpy.concatenate(pred_images, output_images)
output_codes = output_mu.cpu().detach().numpy()
pred_codes = common.numpy.concatenate(pred_codes, output_codes)
utils.write_hdf5(self.args.reconstruction_file, pred_images)
log('[Training] %d: wrote %s' % (epoch, self.args.reconstruction_file))
if utils.display():
png_file = self.args.reconstruction_file + '.%d.png' % epoch
if epoch == 0:
vis.mosaic(png_file, self.test_images[:225], 15, 5, 'gray', 0,
1)
else:
vis.mosaic(png_file, pred_images[:225], 15, 5, 'gray', 0, 1)
log('[Training] %d: wrote %s' % (epoch, png_file))
latent_loss /= num_batches
reconstruction_loss /= num_batches
reconstruction_error /= num_batches
decoder_loss /= num_batches
discriminator_loss /= num_batches
mean /= num_batches
var /= num_batches
logvar /= num_batches
log('[Training] %d: test %g (%g) %g (%g, %g, %g)' %
(epoch, reconstruction_loss, reconstruction_error, latent_loss,
mean, var, logvar))
log('[Training] %d: test %g %g' %
(epoch, decoder_loss, discriminator_loss))
num_batches = int(
math.ceil(self.train_images.shape[0] / self.args.batch_size))
iteration = epoch * num_batches
self.test_statistics = numpy.vstack(
(self.test_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, mean,
var, logvar, decoder_loss, discriminator_loss
])))
pred_images = None
if self.random_codes is None:
self.random_codes = common.numpy.truncated_normal(
(1000, self.args.latent_space_size)).astype(numpy.float32)
num_batches = int(
math.ceil(self.random_codes.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])
if b_start >= b_end: break
batch_codes = common.torch.as_variable(
self.random_codes[b_start:b_end], self.args.use_gpu)
output_images = self.decoder(batch_codes)
output_images = numpy.squeeze(
numpy.transpose(output_images.cpu().detach().numpy(),
(0, 2, 3, 1)))
pred_images = common.numpy.concatenate(pred_images, output_images)
utils.write_hdf5(self.args.random_file, pred_images)
log('[Training] %d: wrote %s' % (epoch, self.args.random_file))
if utils.display() and epoch > 0:
png_file = self.args.random_file + '.%d.png' % epoch
vis.mosaic(png_file, pred_images[:225], 15, 5, 'gray', 0, 1)
log('[Training] %d: wrote %s' % (epoch, png_file))
interpolations = None
perm = numpy.random.permutation(numpy.array(range(
pred_codes.shape[0])))
for i in range(50):
first = pred_codes[i]
second = pred_codes[perm[i]]
linfit = scipy.interpolate.interp1d([0, 1],
numpy.vstack([first, second]),
axis=0)
interpolations = common.numpy.concatenate(
interpolations, linfit(numpy.linspace(0, 1, 10)))
pred_images = None
num_batches = int(
math.ceil(interpolations.shape[0] / self.args.batch_size))
interpolations = interpolations.astype(numpy.float32)
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])
if b_start >= b_end: break
batch_codes = common.torch.as_variable(
interpolations[b_start:b_end], self.args.use_gpu)
output_images = self.decoder(batch_codes)
output_images = numpy.squeeze(
numpy.transpose(output_images.cpu().detach().numpy(),
(0, 2, 3, 1)))
pred_images = common.numpy.concatenate(pred_images, output_images)
if b % 100 == 50:
log('[Testing] %d' % b)
utils.write_hdf5(self.args.interpolation_file, pred_images)
log('[Testing] wrote %s' % self.args.interpolation_file)
if utils.display() and epoch > 0:
png_file = self.args.interpolation_file + '.%d.png' % epoch
vis.mosaic(png_file, pred_images[:100], 10, 5, 'gray', 0, 1)
log('[Training] %d: wrote %s' % (epoch, png_file))
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 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):
"""
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 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 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 = perturbation_loss = perturbation_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_theta = common.torch.as_variable(self.test_theta[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()
if self.args.strong_variant:
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)
e = self.loss(batch_classes, output_classes)
perturbation_loss += e.item()
e = self.error(batch_classes, output_classes)
perturbation_error += e.item()
loss /= num_batches
error /= num_batches
perturbation_loss /= num_batches
perturbation_error /= num_batches
log('[Training] %d: test %g (%g) %g (%g)' % (self.epoch, loss, error, perturbation_loss, perturbation_error))
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
]])))
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
]])))
def attack(self):
"""
Test the model.
"""
assert self.model is not None
assert self.model.training is False
assert self.test_images.shape[0] == self.test_codes.shape[0], 'number of samples has to match'
concatenate_axis = -1
if os.path.exists(self.args.perturbations_file) and os.path.exists(self.args.success_file):
self.original_perturbations = utils.read_hdf5(self.args.perturbations_file)
if self.test_images.shape[3] > 1:
assert len(self.original_perturbations.shape) == 5
else:
assert len(self.original_perturbations.shape) == 4
log('[Attack] read %s' % self.args.perturbations_file)
self.original_success = utils.read_hdf5(self.args.success_file)
log('[Attack] read %s' % self.args.success_file)
assert self.original_perturbations.shape[0] == self.original_success.shape[0]
assert self.original_perturbations.shape[1] == self.original_success.shape[1]
assert self.original_perturbations.shape[2] == self.test_images.shape[1]
assert self.original_perturbations.shape[3] == self.test_images.shape[2]#
if self.original_perturbations.shape[1] >= self.args.max_samples and self.original_perturbations.shape[0] >= self.args.max_attempts:
log('[Attack] found %d attempts, %d samples, requested no more' % (self.original_perturbations.shape[0], self.original_perturbations.shape[1]))
return
elif self.original_perturbations.shape[0] == self.args.max_attempts or self.original_perturbations.shape[1] == self.args.max_samples:
if self.original_perturbations.shape[0] == self.args.max_attempts:
self.test_images = self.test_images[self.original_perturbations.shape[1]:]
self.test_codes = self.test_codes[self.original_perturbations.shape[1]:]
self.args.max_samples = self.args.max_samples - self.original_perturbations.shape[1]
concatenate_axis = 1
log('[Attack] found %d attempts with %d perturbations, computing %d more perturbations' % (self.original_perturbations.shape[0], self.original_perturbations.shape[1], self.args.max_samples))
elif self.original_perturbations.shape[1] == self.args.max_samples:
self.args.max_attempts = self.args.max_attempts - self.original_perturbations.shape[0]
concatenate_axis = 0
log('[Attack] found %d attempts with %d perturbations, computing %d more attempts' % (self.original_perturbations.shape[0], self.original_perturbations.shape[1], self.args.max_attempts))
# can't squeeze here!
if self.test_images.shape[3] > 1:
self.perturbations = numpy.zeros((self.args.max_attempts, self.args.max_samples, self.test_images.shape[1], self.test_images.shape[2], self.test_images.shape[3]))
else:
self.perturbations = numpy.zeros((self.args.max_attempts, self.args.max_samples, self.test_images.shape[1], self.test_images.shape[2]))
self.success = numpy.ones((self.args.max_attempts, self.args.max_samples), dtype=int) * -1
if self.args.attack.find('Batch') >= 0:
batch_size = min(self.args.batch_size, self.args.max_samples)
else:
batch_size = 1
objective = self.objective_class()
num_batches = int(math.ceil(self.args.max_samples/batch_size))
for i in range(num_batches): # self.test_images.shape[0]
if i*batch_size == self.args.max_samples:
break
i_start = i*batch_size
i_end = min((i+1)*batch_size, self.args.max_samples)
batch_images = common.torch.as_variable(self.test_images[i_start: i_end], self.args.use_gpu)
batch_classes = common.torch.as_variable(numpy.array(self.test_codes[i_start: i_end]), self.args.use_gpu)
batch_images = batch_images.permute(0, 3, 1, 2)
t = 0
while t < self.args.max_attempts:
attack = self.setup_attack(batch_images, batch_classes)
success, perturbations, probabilities, norm, _ = attack.run(objective)
assert not numpy.any(perturbations != perturbations), perturbations
# Note that we save the perturbed image, not only the perturbation!
self.perturbations[t][i_start: i_end] = numpy.squeeze(numpy.transpose(perturbations + batch_images.cpu().numpy(), (0, 2, 3, 1)))
self.success[t][i_start: i_end] = success
# IMPORTANT: The adversarial examples are not considering whether the classifier is
# actually correct to start with.
t += 1
log('[Attack] %d: completed' % i)
if concatenate_axis >= 0:
if self.perturbations.shape[0] == self.args.max_attempts:
self.perturbations = numpy.concatenate((self.original_perturbations, self.perturbations), axis=concatenate_axis)
self.success = numpy.concatenate((self.original_success, self.success), axis=concatenate_axis)
log('[Attack] concatenated')
utils.write_hdf5(self.args.perturbations_file, self.perturbations)
log('[Attack] wrote %s' % self.args.perturbations_file)
utils.write_hdf5(self.args.success_file, self.success)
log('[Attack] wrote %s' % self.args.success_file)
def load_data(self):
"""
Load data.
"""
assert self.args.batch_size % 4 == 0
self.train_images = utils.read_hdf5(
self.args.train_images_file).astype(numpy.float32)
log('[Training] read %s' % self.args.train_images_file)
self.test_images = utils.read_hdf5(self.args.test_images_file).astype(
numpy.float32)
log('[Training] read %s' % self.args.test_images_file)
# For handling both color and gray images.
if len(self.train_images.shape) < 4:
self.train_images = numpy.expand_dims(self.train_images, axis=3)
self.test_images = numpy.expand_dims(self.test_images, axis=3)
log('[Training] no color images, adjusted size')
self.resolution = self.test_images.shape[2]
log('[Training] resolution %d' % self.resolution)
self.train_codes = utils.read_hdf5(self.args.train_codes_file).astype(
numpy.int)
assert self.train_codes.shape[1] >= self.args.label_index + 1
self.train_codes = self.train_codes[:, self.args.label_index]
log('[Training] read %s' % self.args.train_codes_file)
self.N_class = numpy.max(self.train_codes) + 1
self.test_codes = utils.read_hdf5(self.args.test_codes_file).astype(
numpy.int)
assert self.test_codes.shape[1] >= self.args.label_index + 1
self.test_codes = self.test_codes[:, self.args.label_index]
log('[Training] read %s' % self.args.test_codes_file)
self.train_theta = utils.read_hdf5(self.args.train_theta_file).astype(
numpy.float32)
log('[Training] read %s' % self.args.train_theta_file)
assert self.test_images.shape[0] == self.test_codes.shape[0]
self.min_bound = numpy.min(self.train_theta, axis=0)
self.max_bound = numpy.max(self.train_theta, axis=0)
log('[Training] min bound: %s' % ' '.join(
['%g' % self.min_bound[i]
for i in range(self.min_bound.shape[0])]))
log('[Training] max bound: %s' % ' '.join(
['%g' % self.max_bound[i]
for i in range(self.max_bound.shape[0])]))
self.test_theta = utils.read_hdf5(self.args.test_theta_file).astype(
numpy.float32)
log('[Training] read %s' % self.args.test_theta_file)
assert self.train_codes.shape[0] == self.train_images.shape[0]
assert self.test_codes.shape[0] == self.test_images.shape[0]
assert self.train_theta.shape[
0] == self.train_images.shape[0], '%s != %s' % ('x'.join(
list(map(str, self.train_theta.shape))), 'x'.join(
list(map(str, self.train_images.shape))))
assert self.test_theta.shape[0] == self.test_images.shape[0]
# Select subset of samples
if self.args.training_samples < 0:
self.args.training_samples = self.train_images.shape[0]
else:
self.args.training_samples = min(self.args.training_samples,
self.train_images.shape[0])
log('[Training] using %d training samples' %
self.args.training_samples)
if self.args.test_samples < 0:
self.args.test_samples = self.test_images.shape[0]
else:
self.args.test_samples = min(self.args.test_samples,
self.test_images.shape[0])
if self.args.early_stopping:
assert self.args.validation_samples > 0
assert self.args.training_samples + self.args.validation_samples <= self.train_images.shape[
0]
self.val_images = self.train_images[self.train_images.shape[0] -
self.args.validation_samples:]
self.val_codes = self.train_codes[self.train_codes.shape[0] -
self.args.validation_samples:]
self.train_images = self.train_images[:self.train_images.shape[0] -
self.args.validation_samples]
self.train_codes = self.train_codeſ[:self.train_codes.shape[0] -
self.args.validation_samples]
assert self.val_images.shape[
0] == self.args.validation_samples and self.val_codes.shape[
0] == self.args.validation_samples
if self.args.random_samples:
perm = numpy.random.permutation(self.train_images.shape[0] // 10)
perm = perm[:self.args.training_samples // 10]
perm = numpy.repeat(perm, self.N_class, axis=0) * 10 + numpy.tile(
numpy.array(range(self.N_class)), (perm.shape[0]))
self.train_images = self.train_images[perm]
self.train_codes = self.train_codes[perm]
self.train_theta = self.train_theta[perm]
else:
self.train_images = self.train_images[:self.args.training_samples]
self.train_codes = self.train_codes[:self.args.training_samples]
self.train_theta = self.train_theta[:self.args.training_samples]
self.train_valid = (numpy.max(numpy.abs(self.train_theta), axis=1) <=
self.args.bound).astype(int)
self.test_valid = (numpy.max(numpy.abs(self.test_theta), axis=1) <=
self.args.bound).astype(int)
# Check that the dataset is balanced.
number_samples = self.train_codes.shape[0] // self.N_class
for c in range(self.N_class):
number_samples_ = numpy.sum(self.train_codes == c)
if number_samples_ != number_samples:
log(
'[Training] dataset not balanced, class %d should have %d samples but has %d'
% (c, number_samples, number_samples_), LogLevel.WARNING)
def load_data(self):
"""
Load data.
"""
assert self.args.batch_size % 4 == 0
self.database = utils.read_hdf5(self.args.database_file).astype(
numpy.float32)
log('[Training] read %s' % self.args.database_file)
self.N_font = self.database.shape[0]
self.N_class = self.database.shape[1]
self.database = self.database.reshape(
(self.database.shape[0] * self.database.shape[1],
self.database.shape[2], self.database.shape[3]))
self.database = torch.from_numpy(self.database)
if self.args.use_gpu:
self.database = self.database.cuda()
self.database = torch.autograd.Variable(self.database, False)
self.train_images = utils.read_hdf5(
self.args.train_images_file).astype(numpy.float32)
log('[Training] read %s' % self.args.train_images_file)
self.test_images = utils.read_hdf5(self.args.test_images_file).astype(
numpy.float32)
log('[Training] read %s' % self.args.test_images_file)
# For handling both color and gray images.
if len(self.train_images.shape) < 4:
self.train_images = numpy.expand_dims(self.train_images, axis=3)
self.test_images = numpy.expand_dims(self.test_images, axis=3)
log('[Training] no color images, adjusted size')
self.resolution = self.train_images.shape[2]
log('[Training] resolution %d' % self.resolution)
self.train_codes = utils.read_hdf5(self.args.train_codes_file).astype(
numpy.int)
assert self.train_codes.shape[1] == 3
log('[Training] read %s' % self.args.train_codes_file)
self.test_codes = utils.read_hdf5(self.args.test_codes_file).astype(
numpy.int)
assert self.test_codes.shape[1] == 3
log('[Training] read %s' % self.args.test_codes_file)
self.train_theta = utils.read_hdf5(self.args.train_theta_file).astype(
numpy.float32)
log('[Training] read %s' % self.args.train_theta_file)
self.min_bound = numpy.min(self.train_theta, axis=0)
self.max_bound = numpy.max(self.train_theta, axis=0)
self.test_theta = utils.read_hdf5(self.args.test_theta_file).astype(
numpy.float32)
log('[Training] read %s' % self.args.test_theta_file)
assert self.train_codes.shape[0] == self.train_images.shape[0]
assert self.test_codes.shape[0] == self.test_images.shape[0]
assert self.train_theta.shape[0] == self.train_images.shape[0]
assert self.test_theta.shape[0] == self.test_images.shape[0]
# Select subset of samples
if self.args.training_samples < 0:
self.args.training_samples = self.train_images.shape[0]
else:
self.args.training_samples = min(self.args.training_samples,
self.train_images.shape[0])
log('[Training] found %d classes' % self.N_class)
log('[Training] using %d training samples' %
self.args.training_samples)
if self.args.test_samples < 0:
self.args.test_samples = self.test_images.shape[0]
else:
self.args.test_samples = min(self.args.test_samples,
self.test_images.shape[0])
if self.args.early_stopping:
assert self.args.validation_samples > 0
assert self.args.training_samples + self.args.validation_samples <= self.train_images.shape[
0]
self.val_images = self.train_images[self.train_images.shape[0] -
self.args.validation_samples:]
self.val_codes = self.train_codes[self.train_codes.shape[0] -
self.args.validation_samples:,
self.args.label_index]
self.train_images = self.train_images[:self.train_images.shape[0] -
self.args.validation_samples]
self.train_codes = self.train_codes[:self.train_codes.shape[0] -
self.args.validation_samples]
assert self.val_images.shape[
0] == self.args.validation_samples and self.val_codes.shape[
0] == self.args.validation_samples
if self.args.random_samples:
perm = numpy.random.permutation(self.train_images.shape[0] // 10)
perm = perm[:self.args.training_samples // 10]
perm = numpy.repeat(perm, self.N_class, axis=0) * 10 + numpy.tile(
numpy.array(range(self.N_class)), (perm.shape[0]))
self.train_images = self.train_images[perm]
self.train_codes = self.train_codes[perm]
self.train_theta = self.train_theta[perm]
else:
self.train_images = self.train_images[:self.args.training_samples]
self.train_codes = self.train_codes[:self.args.training_samples]
self.train_theta = self.train_theta[:self.args.training_samples]
# Check that the dataset is balanced.
number_samples = self.train_codes.shape[0] // self.N_class
for c in range(self.N_class):
number_samples_ = numpy.sum(
self.train_codes[:, self.args.label_index] == c)
if number_samples_ != number_samples:
log(
'[Training] dataset not balanced, class %d should have %d samples but has %d'
% (c, number_samples, number_samples_), LogLevel.WARNING)
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 test(self):
"""
Test the model.
"""
self.model.eval()
log('[Training] %d set classifier to eval' % self.epoch)
loss = error = perturbation_loss = perturbation_error = 0
num_batches = int(
math.ceil(self.args.test_samples / self.args.batch_size))
assert self.model.training is False
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_theta = common.torch.as_variable(self.test_theta[perm],
self.args.use_gpu)
batch_images = batch_images.permute(0, 3, 1, 2)
batch_fonts = self.test_codes[perm, 1]
batch_classes = self.test_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)
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()
if self.args.strong_variant:
min_bound = numpy.repeat(self.min_bound.reshape(1, -1),
batch_theta.size(0),
axis=0)
max_bound = numpy.repeat(self.max_bound.reshape(1, -1),
batch_theta.size(0),
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)
l = self.loss(batch_classes, output_classes)
perturbation_loss += l.item()
e = self.error(batch_classes, output_classes)
perturbation_error += e.item()
loss /= num_batches
error /= num_batches
perturbation_loss /= num_batches
perturbation_error /= num_batches
log('[Training] %d: test %g (%g) %g (%g)' %
(self.epoch, loss, error, perturbation_loss, perturbation_error))
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
]])))
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')
up = np.zeros(len(instance.nodes) * 3)
bc_dofs = []
for bc in boundary_conditions:
set_nodes = []
if bc.type == 'surface':
base = results_odb.rootAssembly.surfaces[bc.set_name]
elif bc.type == 'node_set':
base = results_odb.rootAssembly.nodeSets[bc.set_name]
idx = base.instances.index(instance)
nodes = base.nodes[idx]
print(len(nodes), "in set", bc.set_name)
for n in nodes:
set_nodes.append(3 * (n.label - 1) + bc.component - 1)
bc_dofs.append(3 * (n.label - 1) + bc.component - 1)
bc_dofs = np.unique(np.array(bc_dofs))
with open('up.pkl') as pickle_handle:
up_red = pickle.load(pickle_handle)
bc_set = set(bc_dofs)
j = 0
for i in range(up.shape[0]):
if i not in bc_set:
up[i] = up_red[j]
j += 1
idx = np.argmin(up[1::3])
print(idx)
for i, n in enumerate(node_labels):
permanent_deformation[i, :] = up[3 * (n - 1):3 * n]
results_odb.close()
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 load_model(self):
"""
Load the decoder.
"""
assert self.args.N_theta > 0 and self.args.N_theta <= 9
min_translation_x, max_translation_x = map(
float, self.args.translation_x.split(','))
min_translation_y, max_translation_y = map(
float, self.args.translation_y.split(','))
min_shear_x, max_shear_x = map(float, self.args.shear_x.split(','))
min_shear_y, max_shear_y = map(float, self.args.shear_y.split(','))
min_scale, max_scale = map(float, self.args.scale.split(','))
min_rotation, max_rotation = map(float, self.args.rotation.split(','))
min_color, max_color = self.args.color, 1
self.min_bound = numpy.array([
min_translation_x,
min_translation_y,
min_shear_x,
min_shear_y,
min_scale,
min_rotation,
min_color,
min_color,
min_color,
])
self.max_bound = numpy.array([
max_translation_x, max_translation_y, max_shear_x, max_shear_y,
max_scale, max_rotation, max_color, max_color, max_color
])
self.min_bound = self.min_bound[:self.args.N_theta].astype(
numpy.float32)
self.max_bound = self.max_bound[:self.args.N_theta].astype(
numpy.float32)
decoder = models.STNDecoder(self.args.N_theta)
log('[Attack] set up STN decoder')
classifier = models.Classifier(
self.N_class,
resolution=(self.test_images.shape[3], self.test_images.shape[1],
self.test_images.shape[2]),
architecture='standard',
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)
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('[Attack] read %s' % self.args.classifier_file)
classifier.load_state_dict(state.model)
if self.args.use_gpu and not cuda.is_cuda(classifier):
log('[Attack] classifier is not CUDA')
classifier = classifier.cuda()
classifier.eval()
log('[Attack] loaded classifier')
self.model = models.DecoderClassifier(decoder, classifier)
log('[Training] set up decoder classifier')
