def training_procedure(FLAGS):
"""
model definition
"""
encoder = Encoder(style_dim=FLAGS.style_dim, class_dim=FLAGS.class_dim)
encoder.apply(weights_init)
decoder = Decoder(style_dim=FLAGS.style_dim, class_dim=FLAGS.class_dim)
decoder.apply(weights_init)
# load saved models if load_saved flag is true
if FLAGS.load_saved:
encoder.load_state_dict(
torch.load(os.path.join('checkpoints', FLAGS.encoder_save)))
decoder.load_state_dict(
torch.load(os.path.join('checkpoints', FLAGS.decoder_save)))
"""
variable definition
"""
X_1 = torch.FloatTensor(FLAGS.batch_size, FLAGS.num_channels,
FLAGS.image_size, FLAGS.image_size)
X_2 = torch.FloatTensor(FLAGS.batch_size, FLAGS.num_channels,
FLAGS.image_size, FLAGS.image_size)
X_3 = torch.FloatTensor(FLAGS.batch_size, FLAGS.num_channels,
FLAGS.image_size, FLAGS.image_size)
style_latent_space = torch.FloatTensor(FLAGS.batch_size, FLAGS.style_dim)
"""
loss definitions
"""
cross_entropy_loss = nn.CrossEntropyLoss()
'''
add option to run on GPU
'''
if FLAGS.cuda:
encoder.cuda()
decoder.cuda()
cross_entropy_loss.cuda()
X_1 = X_1.cuda()
X_2 = X_2.cuda()
X_3 = X_3.cuda()
style_latent_space = style_latent_space.cuda()
"""
optimizer and scheduler definition
"""
auto_encoder_optimizer = optim.Adam(list(encoder.parameters()) +
list(decoder.parameters()),
lr=FLAGS.initial_learning_rate,
betas=(FLAGS.beta_1, FLAGS.beta_2))
reverse_cycle_optimizer = optim.Adam(list(encoder.parameters()),
lr=FLAGS.initial_learning_rate,
betas=(FLAGS.beta_1, FLAGS.beta_2))
# divide the learning rate by a factor of 10 after 80 epochs
auto_encoder_scheduler = optim.lr_scheduler.StepLR(auto_encoder_optimizer,
step_size=80,
gamma=0.1)
reverse_cycle_scheduler = optim.lr_scheduler.StepLR(
reverse_cycle_optimizer, step_size=80, gamma=0.1)
"""
training
"""
if torch.cuda.is_available() and not FLAGS.cuda:
print(
"WARNING: You have a CUDA device, so you should probably run with --cuda"
)
if not os.path.exists('checkpoints'):
os.makedirs('checkpoints')
if not os.path.exists('reconstructed_images'):
os.makedirs('reconstructed_images')
# load_saved is false when training is started from 0th iteration
if not FLAGS.load_saved:
with open(FLAGS.log_file, 'w') as log:
log.write(
'Epoch\tIteration\tReconstruction_loss\tKL_divergence_loss\tReverse_cycle_loss\n'
)
# load data set and create data loader instance
print('Loading MNIST paired dataset...')
paired_mnist = MNIST_Paired(root='mnist',
download=True,
train=True,
transform=transform_config)
loader = cycle(
DataLoader(paired_mnist,
batch_size=FLAGS.batch_size,
shuffle=True,
num_workers=0,
drop_last=True))
# initialize summary writer
writer = SummaryWriter()
for epoch in range(FLAGS.start_epoch, FLAGS.end_epoch):
print('')
print(
'Epoch #' + str(epoch) +
'..........................................................................'
)
# update the learning rate scheduler
auto_encoder_scheduler.step()
reverse_cycle_scheduler.step()
for iteration in range(int(len(paired_mnist) / FLAGS.batch_size)):
# A. run the auto-encoder reconstruction
image_batch_1, image_batch_2, _ = next(loader)
auto_encoder_optimizer.zero_grad()
X_1.copy_(image_batch_1)
X_2.copy_(image_batch_2)
style_mu_1, style_logvar_1, class_latent_space_1 = encoder(
Variable(X_1))
style_latent_space_1 = reparameterize(training=True,
mu=style_mu_1,
logvar=style_logvar_1)
kl_divergence_loss_1 = FLAGS.kl_divergence_coef * (
-0.5 * torch.sum(1 + style_logvar_1 - style_mu_1.pow(2) -
style_logvar_1.exp()))
kl_divergence_loss_1 /= (FLAGS.batch_size * FLAGS.num_channels *
FLAGS.image_size * FLAGS.image_size)
kl_divergence_loss_1.backward(retain_graph=True)
style_mu_2, style_logvar_2, class_latent_space_2 = encoder(
Variable(X_2))
style_latent_space_2 = reparameterize(training=True,
mu=style_mu_2,
logvar=style_logvar_2)
kl_divergence_loss_2 = FLAGS.kl_divergence_coef * (
-0.5 * torch.sum(1 + style_logvar_2 - style_mu_2.pow(2) -
style_logvar_2.exp()))
kl_divergence_loss_2 /= (FLAGS.batch_size * FLAGS.num_channels *
FLAGS.image_size * FLAGS.image_size)
kl_divergence_loss_2.backward(retain_graph=True)
reconstructed_X_1 = decoder(style_latent_space_1,
class_latent_space_2)
reconstructed_X_2 = decoder(style_latent_space_2,
class_latent_space_1)
reconstruction_error_1 = FLAGS.reconstruction_coef * mse_loss(
reconstructed_X_1, Variable(X_1))
reconstruction_error_1.backward(retain_graph=True)
reconstruction_error_2 = FLAGS.reconstruction_coef * mse_loss(
reconstructed_X_2, Variable(X_2))
reconstruction_error_2.backward()
reconstruction_error = (
reconstruction_error_1 +
reconstruction_error_2) / FLAGS.reconstruction_coef
kl_divergence_error = (kl_divergence_loss_1 + kl_divergence_loss_2
) / FLAGS.kl_divergence_coef
auto_encoder_optimizer.step()
# B. reverse cycle
image_batch_1, _, __ = next(loader)
image_batch_2, _, __ = next(loader)
reverse_cycle_optimizer.zero_grad()
X_1.copy_(image_batch_1)
X_2.copy_(image_batch_2)
style_latent_space.normal_(0., 1.)
_, __, class_latent_space_1 = encoder(Variable(X_1))
_, __, class_latent_space_2 = encoder(Variable(X_2))
reconstructed_X_1 = decoder(Variable(style_latent_space),
class_latent_space_1.detach())
reconstructed_X_2 = decoder(Variable(style_latent_space),
class_latent_space_2.detach())
style_mu_1, style_logvar_1, _ = encoder(reconstructed_X_1)
style_latent_space_1 = reparameterize(training=False,
mu=style_mu_1,
logvar=style_logvar_1)
style_mu_2, style_logvar_2, _ = encoder(reconstructed_X_2)
style_latent_space_2 = reparameterize(training=False,
mu=style_mu_2,
logvar=style_logvar_2)
reverse_cycle_loss = FLAGS.reverse_cycle_coef * l1_loss(
style_latent_space_1, style_latent_space_2)
reverse_cycle_loss.backward()
reverse_cycle_loss /= FLAGS.reverse_cycle_coef
reverse_cycle_optimizer.step()
if (iteration + 1) % 10 == 0:
print('')
print('Epoch #' + str(epoch))
print('Iteration #' + str(iteration))
print('')
print('Reconstruction loss: ' +
str(reconstruction_error.data.storage().tolist()[0]))
print('KL-Divergence loss: ' +
str(kl_divergence_error.data.storage().tolist()[0]))
print('Reverse cycle loss: ' +
str(reverse_cycle_loss.data.storage().tolist()[0]))
# write to log
with open(FLAGS.log_file, 'a') as log:
log.write('{0}\t{1}\t{2}\t{3}\t{4}\n'.format(
epoch, iteration,
reconstruction_error.data.storage().tolist()[0],
kl_divergence_error.data.storage().tolist()[0],
reverse_cycle_loss.data.storage().tolist()[0]))
# write to tensorboard
writer.add_scalar(
'Reconstruction loss',
reconstruction_error.data.storage().tolist()[0],
epoch * (int(len(paired_mnist) / FLAGS.batch_size) + 1) +
iteration)
writer.add_scalar(
'KL-Divergence loss',
kl_divergence_error.data.storage().tolist()[0],
epoch * (int(len(paired_mnist) / FLAGS.batch_size) + 1) +
iteration)
writer.add_scalar(
'Reverse cycle loss',
reverse_cycle_loss.data.storage().tolist()[0],
epoch * (int(len(paired_mnist) / FLAGS.batch_size) + 1) +
iteration)
# save model after every 5 epochs
if (epoch + 1) % 5 == 0 or (epoch + 1) == FLAGS.end_epoch:
torch.save(encoder.state_dict(),
os.path.join('checkpoints', FLAGS.encoder_save))
torch.save(decoder.state_dict(),
os.path.join('checkpoints', FLAGS.decoder_save))
"""
save reconstructed images and style swapped image generations to check progress
"""
image_batch_1, image_batch_2, _ = next(loader)
image_batch_3, _, __ = next(loader)
X_1.copy_(image_batch_1)
X_2.copy_(image_batch_2)
X_3.copy_(image_batch_3)
style_mu_1, style_logvar_1, _ = encoder(Variable(X_1))
_, __, class_latent_space_2 = encoder(Variable(X_2))
style_mu_3, style_logvar_3, _ = encoder(Variable(X_3))
style_latent_space_1 = reparameterize(training=False,
mu=style_mu_1,
logvar=style_logvar_1)
style_latent_space_3 = reparameterize(training=False,
mu=style_mu_3,
logvar=style_logvar_3)
reconstructed_X_1_2 = decoder(style_latent_space_1,
class_latent_space_2)
reconstructed_X_3_2 = decoder(style_latent_space_3,
class_latent_space_2)
# save input image batch
image_batch = np.transpose(X_1.cpu().numpy(), (0, 2, 3, 1))
image_batch = np.concatenate(
(image_batch, image_batch, image_batch), axis=3)
imshow_grid(image_batch, name=str(epoch) + '_original', save=True)
# save reconstructed batch
reconstructed_x = np.transpose(
reconstructed_X_1_2.cpu().data.numpy(), (0, 2, 3, 1))
reconstructed_x = np.concatenate(
(reconstructed_x, reconstructed_x, reconstructed_x), axis=3)
imshow_grid(reconstructed_x,
name=str(epoch) + '_target',
save=True)
style_batch = np.transpose(X_3.cpu().numpy(), (0, 2, 3, 1))
style_batch = np.concatenate(
(style_batch, style_batch, style_batch), axis=3)
imshow_grid(style_batch, name=str(epoch) + '_style', save=True)
# save style swapped reconstructed batch
reconstructed_style = np.transpose(
reconstructed_X_3_2.cpu().data.numpy(), (0, 2, 3, 1))
reconstructed_style = np.concatenate(
(reconstructed_style, reconstructed_style,
reconstructed_style),
axis=3)
imshow_grid(reconstructed_style,
name=str(epoch) + '_style_target',
save=True)
def training_procedure(FLAGS):
"""
model definition
"""
encoder = Encoder(style_dim=FLAGS.style_dim, class_dim=FLAGS.class_dim)
encoder.apply(weights_init)
decoder = Decoder(style_dim=FLAGS.style_dim, class_dim=FLAGS.class_dim)
decoder.apply(weights_init)
# load saved models if load_saved flag is true
if FLAGS.load_saved:
encoder.load_state_dict(
torch.load(os.path.join('checkpoints', FLAGS.encoder_save)))
decoder.load_state_dict(
torch.load(os.path.join('checkpoints', FLAGS.decoder_save)))
"""
variable definition
"""
X = torch.FloatTensor(FLAGS.batch_size, 1, FLAGS.image_size,
FLAGS.image_size)
'''
add option to run on GPU
'''
if FLAGS.cuda:
encoder.cuda()
decoder.cuda()
X = X.cuda()
"""
optimizer definition
"""
auto_encoder_optimizer = optim.Adam(list(encoder.parameters()) +
list(decoder.parameters()),
lr=FLAGS.initial_learning_rate,
betas=(FLAGS.beta_1, FLAGS.beta_2))
"""
training
"""
if torch.cuda.is_available() and not FLAGS.cuda:
print(
"WARNING: You have a CUDA device, so you should probably run with --cuda"
)
if not os.path.exists('checkpoints'):
os.makedirs('checkpoints')
# load_saved is false when training is started from 0th iteration
if not FLAGS.load_saved:
with open(FLAGS.log_file, 'w') as log:
log.write(
'Epoch\tIteration\tReconstruction_loss\tStyle_KL_divergence_loss\tClass_KL_divergence_loss\n'
)
# load data set and create data loader instance
print('Loading MNIST dataset...')
mnist = datasets.MNIST(root='mnist',
download=True,
train=True,
transform=transform_config)
loader = cycle(
DataLoader(mnist,
batch_size=FLAGS.batch_size,
shuffle=True,
num_workers=0,
drop_last=True))
# initialize summary writer
writer = SummaryWriter()
for epoch in range(FLAGS.start_epoch, FLAGS.end_epoch):
print('')
print(
'Epoch #' + str(epoch) +
'..........................................................................'
)
for iteration in range(int(len(mnist) / FLAGS.batch_size)):
# load a mini-batch
image_batch, labels_batch = next(loader)
# set zero_grad for the optimizer
auto_encoder_optimizer.zero_grad()
X.copy_(image_batch)
style_mu, style_logvar, class_mu, class_logvar = encoder(
Variable(X))
grouped_mu, grouped_logvar = accumulate_group_evidence(
class_mu.data, class_logvar.data, labels_batch, FLAGS.cuda)
# kl-divergence error for style latent space
style_kl_divergence_loss = FLAGS.kl_divergence_coef * (
-0.5 * torch.sum(1 + style_logvar - style_mu.pow(2) -
style_logvar.exp()))
style_kl_divergence_loss /= (FLAGS.batch_size *
FLAGS.num_channels *
FLAGS.image_size * FLAGS.image_size)
style_kl_divergence_loss.backward(retain_graph=True)
# kl-divergence error for class latent space
class_kl_divergence_loss = FLAGS.kl_divergence_coef * (
-0.5 * torch.sum(1 + grouped_logvar - grouped_mu.pow(2) -
grouped_logvar.exp()))
class_kl_divergence_loss /= (FLAGS.batch_size *
FLAGS.num_channels *
FLAGS.image_size * FLAGS.image_size)
class_kl_divergence_loss.backward(retain_graph=True)
# reconstruct samples
"""
sampling from group mu and logvar for each image in mini-batch differently makes
the decoder consider class latent embeddings as random noise and ignore them
"""
style_latent_embeddings = reparameterize(training=True,
mu=style_mu,
logvar=style_logvar)
class_latent_embeddings = group_wise_reparameterize(
training=True,
mu=grouped_mu,
logvar=grouped_logvar,
labels_batch=labels_batch,
cuda=FLAGS.cuda)
reconstructed_images = decoder(style_latent_embeddings,
class_latent_embeddings)
reconstruction_error = FLAGS.reconstruction_coef * mse_loss(
reconstructed_images, Variable(X))
reconstruction_error.backward()
auto_encoder_optimizer.step()
if (iteration + 1) % 50 == 0:
print('')
print('Epoch #' + str(epoch))
print('Iteration #' + str(iteration))
print('')
print('Reconstruction loss: ' +
str(reconstruction_error.data.storage().tolist()[0]))
print('Style KL-Divergence loss: ' +
str(style_kl_divergence_loss.data.storage().tolist()[0]))
print('Class KL-Divergence loss: ' +
str(class_kl_divergence_loss.data.storage().tolist()[0]))
# write to log
with open(FLAGS.log_file, 'a') as log:
log.write('{0}\t{1}\t{2}\t{3}\t{4}\n'.format(
epoch, iteration,
reconstruction_error.data.storage().tolist()[0],
style_kl_divergence_loss.data.storage().tolist()[0],
class_kl_divergence_loss.data.storage().tolist()[0]))
# write to tensorboard
writer.add_scalar(
'Reconstruction loss',
reconstruction_error.data.storage().tolist()[0],
epoch * (int(len(mnist) / FLAGS.batch_size) + 1) + iteration)
writer.add_scalar(
'Style KL-Divergence loss',
style_kl_divergence_loss.data.storage().tolist()[0],
epoch * (int(len(mnist) / FLAGS.batch_size) + 1) + iteration)
writer.add_scalar(
'Class KL-Divergence loss',
class_kl_divergence_loss.data.storage().tolist()[0],
epoch * (int(len(mnist) / FLAGS.batch_size) + 1) + iteration)
# save checkpoints after every 5 epochs
if (epoch + 1) % 5 == 0 or (epoch + 1) == FLAGS.end_epoch:
torch.save(encoder.state_dict(),
os.path.join('checkpoints', FLAGS.encoder_save))
torch.save(decoder.state_dict(),
os.path.join('checkpoints', FLAGS.decoder_save))
def training_procedure(FLAGS):
"""
model definition
"""
encoder = Encoder(style_dim=FLAGS.style_dim, class_dim=FLAGS.class_dim)
encoder.apply(weights_init)
decoder = Decoder(style_dim=FLAGS.style_dim, class_dim=FLAGS.class_dim)
decoder.apply(weights_init)
discriminator = Discriminator()
discriminator.apply(weights_init)
# load saved models if load_saved flag is true
if FLAGS.load_saved:
raise Exception('This is not implemented')
encoder.load_state_dict(torch.load(os.path.join('checkpoints', FLAGS.encoder_save)))
decoder.load_state_dict(torch.load(os.path.join('checkpoints', FLAGS.decoder_save)))
"""
variable definition
"""
X_1 = torch.FloatTensor(FLAGS.batch_size, FLAGS.num_channels, FLAGS.image_size, FLAGS.image_size)
X_2 = torch.FloatTensor(FLAGS.batch_size, FLAGS.num_channels, FLAGS.image_size, FLAGS.image_size)
X_3 = torch.FloatTensor(FLAGS.batch_size, FLAGS.num_channels, FLAGS.image_size, FLAGS.image_size)
style_latent_space = torch.FloatTensor(FLAGS.batch_size, FLAGS.style_dim)
"""
loss definitions
"""
cross_entropy_loss = nn.CrossEntropyLoss()
adversarial_loss = nn.BCELoss()
'''
add option to run on GPU
'''
if FLAGS.cuda:
encoder.cuda()
decoder.cuda()
discriminator.cuda()
cross_entropy_loss.cuda()
adversarial_loss.cuda()
X_1 = X_1.cuda()
X_2 = X_2.cuda()
X_3 = X_3.cuda()
style_latent_space = style_latent_space.cuda()
"""
optimizer and scheduler definition
"""
auto_encoder_optimizer = optim.Adam(
list(encoder.parameters()) + list(decoder.parameters()),
lr=FLAGS.initial_learning_rate,
betas=(FLAGS.beta_1, FLAGS.beta_2)
)
reverse_cycle_optimizer = optim.Adam(
list(encoder.parameters()),
lr=FLAGS.initial_learning_rate,
betas=(FLAGS.beta_1, FLAGS.beta_2)
)
generator_optimizer = optim.Adam(
list(decoder.parameters()),
lr=FLAGS.initial_learning_rate,
betas=(FLAGS.beta_1, FLAGS.beta_2)
)
discriminator_optimizer = optim.Adam(
list(discriminator.parameters()),
lr=FLAGS.initial_learning_rate,
betas=(FLAGS.beta_1, FLAGS.beta_2)
)
# divide the learning rate by a factor of 10 after 80 epochs
auto_encoder_scheduler = optim.lr_scheduler.StepLR(auto_encoder_optimizer, step_size=80, gamma=0.1)
reverse_cycle_scheduler = optim.lr_scheduler.StepLR(reverse_cycle_optimizer, step_size=80, gamma=0.1)
generator_scheduler = optim.lr_scheduler.StepLR(generator_optimizer, step_size=80, gamma=0.1)
discriminator_scheduler = optim.lr_scheduler.StepLR(discriminator_optimizer, step_size=80, gamma=0.1)
# Used later to define discriminator ground truths
Tensor = torch.cuda.FloatTensor if FLAGS.cuda else torch.FloatTensor
"""
training
"""
if torch.cuda.is_available() and not FLAGS.cuda:
print("WARNING: You have a CUDA device, so you should probably run with --cuda")
if not os.path.exists('checkpoints'):
os.makedirs('checkpoints')
if not os.path.exists('reconstructed_images'):
os.makedirs('reconstructed_images')
# load_saved is false when training is started from 0th iteration
if not FLAGS.load_saved:
with open(FLAGS.log_file, 'w') as log:
headers = ['Epoch', 'Iteration', 'Reconstruction_loss', 'KL_divergence_loss', 'Reverse_cycle_loss']
if FLAGS.forward_gan:
headers.extend(['Generator_forward_loss', 'Discriminator_forward_loss'])
if FLAGS.reverse_gan:
headers.extend(['Generator_reverse_loss', 'Discriminator_reverse_loss'])
log.write('\t'.join(headers) + '\n')
# load data set and create data loader instance
print('Loading CIFAR paired dataset...')
paired_cifar = CIFAR_Paired(root='cifar', download=True, train=True, transform=transform_config)
loader = cycle(DataLoader(paired_cifar, batch_size=FLAGS.batch_size, shuffle=True, num_workers=0, drop_last=True))
# Save a batch of images to use for visualization
image_sample_1, image_sample_2, _ = next(loader)
image_sample_3, _, _ = next(loader)
# initialize summary writer
writer = SummaryWriter()
for epoch in range(FLAGS.start_epoch, FLAGS.end_epoch):
print('')
print('Epoch #' + str(epoch) + '..........................................................................')
# update the learning rate scheduler
auto_encoder_scheduler.step()
reverse_cycle_scheduler.step()
generator_scheduler.step()
discriminator_scheduler.step()
for iteration in range(int(len(paired_cifar) / FLAGS.batch_size)):
# Adversarial ground truths
valid = Variable(Tensor(FLAGS.batch_size, 1).fill_(1.0), requires_grad=False)
fake = Variable(Tensor(FLAGS.batch_size, 1).fill_(0.0), requires_grad=False)
# A. run the auto-encoder reconstruction
image_batch_1, image_batch_2, _ = next(loader)
auto_encoder_optimizer.zero_grad()
X_1.copy_(image_batch_1)
X_2.copy_(image_batch_2)
style_mu_1, style_logvar_1, class_latent_space_1 = encoder(Variable(X_1))
style_latent_space_1 = reparameterize(training=True, mu=style_mu_1, logvar=style_logvar_1)
kl_divergence_loss_1 = FLAGS.kl_divergence_coef * (
- 0.5 * torch.sum(1 + style_logvar_1 - style_mu_1.pow(2) - style_logvar_1.exp())
)
kl_divergence_loss_1 /= (FLAGS.batch_size * FLAGS.num_channels * FLAGS.image_size * FLAGS.image_size)
kl_divergence_loss_1.backward(retain_graph=True)
style_mu_2, style_logvar_2, class_latent_space_2 = encoder(Variable(X_2))
style_latent_space_2 = reparameterize(training=True, mu=style_mu_2, logvar=style_logvar_2)
kl_divergence_loss_2 = FLAGS.kl_divergence_coef * (
- 0.5 * torch.sum(1 + style_logvar_2 - style_mu_2.pow(2) - style_logvar_2.exp())
)
kl_divergence_loss_2 /= (FLAGS.batch_size * FLAGS.num_channels * FLAGS.image_size * FLAGS.image_size)
kl_divergence_loss_2.backward(retain_graph=True)
reconstructed_X_1 = decoder(style_latent_space_1, class_latent_space_2)
reconstructed_X_2 = decoder(style_latent_space_2, class_latent_space_1)
reconstruction_error_1 = FLAGS.reconstruction_coef * mse_loss(reconstructed_X_1, Variable(X_1))
reconstruction_error_1.backward(retain_graph=True)
reconstruction_error_2 = FLAGS.reconstruction_coef * mse_loss(reconstructed_X_2, Variable(X_2))
reconstruction_error_2.backward()
reconstruction_error = (reconstruction_error_1 + reconstruction_error_2) / FLAGS.reconstruction_coef
kl_divergence_error = (kl_divergence_loss_1 + kl_divergence_loss_2) / FLAGS.kl_divergence_coef
auto_encoder_optimizer.step()
# A-1. Discriminator training during forward cycle
if FLAGS.forward_gan:
# Training generator
generator_optimizer.zero_grad()
g_loss_1 = adversarial_loss(discriminator(Variable(reconstructed_X_1)), valid)
g_loss_2 = adversarial_loss(discriminator(Variable(reconstructed_X_2)), valid)
gen_f_loss = (g_loss_1 + g_loss_2) / 2.0
gen_f_loss.backward()
generator_optimizer.step()
# Training discriminator
discriminator_optimizer.zero_grad()
real_loss_1 = adversarial_loss(discriminator(Variable(X_1)), valid)
real_loss_2 = adversarial_loss(discriminator(Variable(X_2)), valid)
fake_loss_1 = adversarial_loss(discriminator(Variable(reconstructed_X_1)), fake)
fake_loss_2 = adversarial_loss(discriminator(Variable(reconstructed_X_2)), fake)
dis_f_loss = (real_loss_1 + real_loss_2 + fake_loss_1 + fake_loss_2) / 4.0
dis_f_loss.backward()
discriminator_optimizer.step()
# B. reverse cycle
image_batch_1, _, __ = next(loader)
image_batch_2, _, __ = next(loader)
reverse_cycle_optimizer.zero_grad()
X_1.copy_(image_batch_1)
X_2.copy_(image_batch_2)
style_latent_space.normal_(0., 1.)
_, __, class_latent_space_1 = encoder(Variable(X_1))
_, __, class_latent_space_2 = encoder(Variable(X_2))
reconstructed_X_1 = decoder(Variable(style_latent_space), class_latent_space_1.detach())
reconstructed_X_2 = decoder(Variable(style_latent_space), class_latent_space_2.detach())
style_mu_1, style_logvar_1, _ = encoder(reconstructed_X_1)
style_latent_space_1 = reparameterize(training=False, mu=style_mu_1, logvar=style_logvar_1)
style_mu_2, style_logvar_2, _ = encoder(reconstructed_X_2)
style_latent_space_2 = reparameterize(training=False, mu=style_mu_2, logvar=style_logvar_2)
reverse_cycle_loss = FLAGS.reverse_cycle_coef * l1_loss(style_latent_space_1, style_latent_space_2)
reverse_cycle_loss.backward()
reverse_cycle_loss /= FLAGS.reverse_cycle_coef
reverse_cycle_optimizer.step()
# B-1. Discriminator training during reverse cycle
if FLAGS.reverse_gan:
# Training generator
generator_optimizer.zero_grad()
g_loss_1 = adversarial_loss(discriminator(Variable(reconstructed_X_1)), valid)
g_loss_2 = adversarial_loss(discriminator(Variable(reconstructed_X_2)), valid)
gen_r_loss = (g_loss_1 + g_loss_2) / 2.0
gen_r_loss.backward()
generator_optimizer.step()
# Training discriminator
discriminator_optimizer.zero_grad()
real_loss_1 = adversarial_loss(discriminator(Variable(X_1)), valid)
real_loss_2 = adversarial_loss(discriminator(Variable(X_2)), valid)
fake_loss_1 = adversarial_loss(discriminator(Variable(reconstructed_X_1)), fake)
fake_loss_2 = adversarial_loss(discriminator(Variable(reconstructed_X_2)), fake)
dis_r_loss = (real_loss_1 + real_loss_2 + fake_loss_1 + fake_loss_2) / 4.0
dis_r_loss.backward()
discriminator_optimizer.step()
if (iteration + 1) % 10 == 0:
print('')
print('Epoch #' + str(epoch))
print('Iteration #' + str(iteration))
print('')
print('Reconstruction loss: ' + str(reconstruction_error.data.storage().tolist()[0]))
print('KL-Divergence loss: ' + str(kl_divergence_error.data.storage().tolist()[0]))
print('Reverse cycle loss: ' + str(reverse_cycle_loss.data.storage().tolist()[0]))
if FLAGS.forward_gan:
print('Generator F loss: ' + str(gen_f_loss.data.storage().tolist()[0]))
print('Discriminator F loss: ' + str(dis_f_loss.data.storage().tolist()[0]))
if FLAGS.reverse_gan:
print('Generator R loss: ' + str(gen_r_loss.data.storage().tolist()[0]))
print('Discriminator R loss: ' + str(dis_r_loss.data.storage().tolist()[0]))
# write to log
with open(FLAGS.log_file, 'a') as log:
row = []
row.append(epoch)
row.append(iteration)
row.append(reconstruction_error.data.storage().tolist()[0])
row.append(kl_divergence_error.data.storage().tolist()[0])
row.append(reverse_cycle_loss.data.storage().tolist()[0])
if FLAGS.forward_gan:
row.append(gen_f_loss.data.storage().tolist()[0])
row.append(dis_f_loss.data.storage().tolist()[0])
if FLAGS.reverse_gan:
row.append(gen_r_loss.data.storage().tolist()[0])
row.append(dis_r_loss.data.storage().tolist()[0])
row = [str(x) for x in row]
log.write('\t'.join(row) + '\n')
# write to tensorboard
writer.add_scalar('Reconstruction loss', reconstruction_error.data.storage().tolist()[0],
epoch * (int(len(paired_cifar) / FLAGS.batch_size) + 1) + iteration)
writer.add_scalar('KL-Divergence loss', kl_divergence_error.data.storage().tolist()[0],
epoch * (int(len(paired_cifar) / FLAGS.batch_size) + 1) + iteration)
writer.add_scalar('Reverse cycle loss', reverse_cycle_loss.data.storage().tolist()[0],
epoch * (int(len(paired_cifar) / FLAGS.batch_size) + 1) + iteration)
if FLAGS.forward_gan:
writer.add_scalar('Generator F loss', gen_f_loss.data.storage().tolist()[0],
epoch * (int(len(paired_cifar) / FLAGS.batch_size) + 1) + iteration)
writer.add_scalar('Discriminator F loss', dis_f_loss.data.storage().tolist()[0],
epoch * (int(len(paired_cifar) / FLAGS.batch_size) + 1) + iteration)
if FLAGS.reverse_gan:
writer.add_scalar('Generator R loss', gen_r_loss.data.storage().tolist()[0],
epoch * (int(len(paired_cifar) / FLAGS.batch_size) + 1) + iteration)
writer.add_scalar('Discriminator R loss', dis_r_loss.data.storage().tolist()[0],
epoch * (int(len(paired_cifar) / FLAGS.batch_size) + 1) + iteration)
# save model after every 5 epochs
if (epoch + 1) % 5 == 0 or (epoch + 1) == FLAGS.end_epoch:
torch.save(encoder.state_dict(), os.path.join('checkpoints', FLAGS.encoder_save))
torch.save(decoder.state_dict(), os.path.join('checkpoints', FLAGS.decoder_save))
"""
save reconstructed images and style swapped image generations to check progress
"""
X_1.copy_(image_sample_1)
X_2.copy_(image_sample_2)
X_3.copy_(image_sample_3)
style_mu_1, style_logvar_1, _ = encoder(Variable(X_1))
_, __, class_latent_space_2 = encoder(Variable(X_2))
style_mu_3, style_logvar_3, _ = encoder(Variable(X_3))
style_latent_space_1 = reparameterize(training=False, mu=style_mu_1, logvar=style_logvar_1)
style_latent_space_3 = reparameterize(training=False, mu=style_mu_3, logvar=style_logvar_3)
reconstructed_X_1_2 = decoder(style_latent_space_1, class_latent_space_2)
reconstructed_X_3_2 = decoder(style_latent_space_3, class_latent_space_2)
# save input image batch
image_batch = np.transpose(X_1.cpu().numpy(), (0, 2, 3, 1))
if FLAGS.num_channels == 1:
image_batch = np.concatenate((image_batch, image_batch, image_batch), axis=3)
imshow_grid(image_batch, name=str(epoch) + '_original', save=True)
# save reconstructed batch
reconstructed_x = np.transpose(reconstructed_X_1_2.cpu().data.numpy(), (0, 2, 3, 1))
if FLAGS.num_channels == 1:
reconstructed_x = np.concatenate((reconstructed_x, reconstructed_x, reconstructed_x), axis=3)
imshow_grid(reconstructed_x, name=str(epoch) + '_target', save=True)
style_batch = np.transpose(X_3.cpu().numpy(), (0, 2, 3, 1))
if FLAGS.num_channels == 1:
style_batch = np.concatenate((style_batch, style_batch, style_batch), axis=3)
imshow_grid(style_batch, name=str(epoch) + '_style', save=True)
# save style swapped reconstructed batch
reconstructed_style = np.transpose(reconstructed_X_3_2.cpu().data.numpy(), (0, 2, 3, 1))
if FLAGS.num_channels == 1:
reconstructed_style = np.concatenate((reconstructed_style, reconstructed_style, reconstructed_style), axis=3)
imshow_grid(reconstructed_style, name=str(epoch) + '_style_target', save=True)
decoder = Decoder()
decoder.apply(weights_init)
# load saved models if load_saved flag is true
if LOAD_SAVED:
encoder.load_state_dict(
torch.load(os.path.join('checkpoints', ENCODER_SAVE)))
decoder.load_state_dict(
torch.load(os.path.join('checkpoints', DECODER_SAVE)))
# loss definition
mse_loss = nn.MSELoss()
# add option to run on gpu
if (CUDA):
encoder.cuda()
decoder.cuda()
mse_loss.cuda()
# optimizer
optimizer = torch.optim.Adam(list(encoder.parameters()) +
list(decoder.parameters()),
lr=LR,
betas=(BETA1, BETA2))
if torch.cuda.is_available() and not CUDA:
print(
"WARNING: You have a CUDA device, so you should probably run with --cuda"
)
# creating directories
FLAGS = parser.parse_args()
if __name__ == '__main__':
"""
model definitions
"""
encoder = Encoder(style_dim=FLAGS.style_dim, class_dim=FLAGS.class_dim)
decoder = Decoder(style_dim=FLAGS.style_dim, class_dim=FLAGS.class_dim)
encoder.load_state_dict(
torch.load(os.path.join('checkpoints', FLAGS.encoder_save), map_location=lambda storage, loc: storage))
decoder.load_state_dict(
torch.load(os.path.join('checkpoints', FLAGS.decoder_save), map_location=lambda storage, loc: storage))
encoder.cuda()
decoder.cuda()
if not os.path.exists('reconstructed_images'):
os.makedirs('reconstructed_images')
# load data set and create data loader instance
'''
print('Loading MNIST paired dataset...')
paired_mnist = MNIST_Paired(root='mnist', download=True, train=False, transform=transform_config)
loader = cycle(DataLoader(paired_mnist, batch_size=FLAGS.batch_size, shuffle=True, num_workers=0, drop_last=True))
image_array = []
for i in range(0, 11):
image_array.append([])
# add a blank image in the top row
def training_procedure(FLAGS):
"""
model definition
"""
encoder = Encoder(style_dim=FLAGS.style_dim, class_dim=FLAGS.class_dim)
encoder.apply(weights_init)
decoder = Decoder(style_dim=FLAGS.style_dim, class_dim=FLAGS.class_dim)
decoder.apply(weights_init)
# load saved models if load_saved flag is true
if FLAGS.load_saved:
encoder.load_state_dict(
torch.load(os.path.join(savedir, FLAGS.encoder_save)))
decoder.load_state_dict(
torch.load(os.path.join(savedir, FLAGS.decoder_save)))
'''
add option to run on GPU
'''
if FLAGS.cuda:
encoder.cuda()
decoder.cuda()
"""
optimizer definition
"""
auto_encoder_optimizer = optim.Adam(list(encoder.parameters()) +
list(decoder.parameters()),
lr=FLAGS.initial_learning_rate,
betas=(FLAGS.beta_1, FLAGS.beta_2))
"""
training
"""
if torch.cuda.is_available() and not FLAGS.cuda:
print(
"WARNING: You have a CUDA device, so you should probably run with --cuda"
)
savedir = 'checkpoints_%d' % (FLAGS.batch_size)
if not os.path.exists(savedir):
os.makedirs(savedir)
# load_saved is false when training is started from 0th iteration
if not FLAGS.load_saved:
with open(FLAGS.log_file, 'w') as log:
log.write(
'Epoch\tIteration\tReconstruction_loss\tStyle_KL_divergence_loss\tClass_KL_divergence_loss\n'
)
# load data set and create data loader instance
print('Loading MNIST dataset...')
mnist = datasets.MNIST(root='mnist',
download=True,
train=True,
transform=transform_config)
# Creating data indices for training and validation splits:
dataset_size = len(mnist)
indices = list(range(dataset_size))
split = 10000
np.random.seed(0)
np.random.shuffle(indices)
train_indices, val_indices = indices[split:], indices[:split]
train_mnist, val_mnist = torch.utils.data.random_split(
mnist, [dataset_size - split, split])
# Creating PT data samplers and loaders:
weights_train = torch.ones(len(mnist))
weights_test = torch.ones(len(mnist))
weights_train[val_mnist.indices] = 0
weights_test[train_mnist.indices] = 0
counts = torch.zeros(10)
for i in range(10):
idx_label = mnist.targets[train_mnist.indices].eq(i)
counts[i] = idx_label.sum()
max = float(counts.max())
sum_counts = float(counts.sum())
for i in range(10):
idx_label = mnist.targets[train_mnist.indices].eq(
i).nonzero().squeeze()
weights_train[train_mnist.indices[idx_label]] = (sum_counts /
counts[i])
train_sampler = SubsetRandomSampler(train_mnist.indices)
valid_sampler = SubsetRandomSampler(val_mnist.indices)
kwargs = {'num_workers': 1, 'pin_memory': True} if FLAGS.cuda else {}
loader = DataLoader(mnist,
batch_size=FLAGS.batch_size,
sampler=train_sampler,
**kwargs)
valid_loader = DataLoader(mnist,
batch_size=FLAGS.batch_size,
sampler=valid_sampler,
**kwargs)
monitor = torch.zeros(FLAGS.end_epoch - FLAGS.start_epoch, 4)
# initialize summary writer
writer = SummaryWriter()
for epoch in range(FLAGS.start_epoch, FLAGS.end_epoch):
print('')
print(
'Epoch #' + str(epoch) +
'..........................................................................'
)
elbo_epoch = 0
term1_epoch = 0
term2_epoch = 0
term3_epoch = 0
for it, (image_batch, labels_batch) in enumerate(loader):
# set zero_grad for the optimizer
auto_encoder_optimizer.zero_grad()
X = image_batch.cuda().detach().clone()
elbo, reconstruction_proba, style_kl_divergence_loss, class_kl_divergence_loss = process(
FLAGS, X, labels_batch, encoder, decoder)
(-elbo).backward()
auto_encoder_optimizer.step()
elbo_epoch += elbo
term1_epoch += reconstruction_proba
term2_epoch += style_kl_divergence_loss
term3_epoch += class_kl_divergence_loss
print("Elbo epoch %.2f" % (elbo_epoch / (it + 1)))
print("Rec. Proba %.2f" % (term1_epoch / (it + 1)))
print("KL style %.2f" % (term2_epoch / (it + 1)))
print("KL content %.2f" % (term3_epoch / (it + 1)))
# save checkpoints after every 5 epochs
if (epoch + 1) % 5 == 0 or (epoch + 1) == FLAGS.end_epoch:
monitor[epoch, :] = eval(FLAGS, valid_loader, encoder, decoder)
torch.save(
encoder.state_dict(),
os.path.join(savedir, FLAGS.encoder_save + '_e%d' % epoch))
torch.save(
decoder.state_dict(),
os.path.join(savedir, FLAGS.decoder_save + '_e%d' % epoch))
print("VAL elbo %.2f" % (monitor[epoch, 0]))
print("VAL Rec. Proba %.2f" % (monitor[epoch, 1]))
print("VAL KL style %.2f" % (monitor[epoch, 2]))
print("VAL KL content %.2f" % (monitor[epoch, 3]))
torch.save(monitor, os.path.join(savedir, 'monitor_e%d' % epoch))
def training_procedure(FLAGS):
"""
model definition
"""
encoder = Encoder(style_dim=FLAGS.style_dim, class_dim=FLAGS.class_dim)
encoder.apply(weights_init)
decoder = Decoder(style_dim=FLAGS.style_dim, class_dim=FLAGS.class_dim)
decoder.apply(weights_init)
discriminator = Discriminator()
discriminator.apply(weights_init)
# load saved models if load_saved flag is true
if FLAGS.load_saved:
encoder.load_state_dict(torch.load(os.path.join('checkpoints', FLAGS.encoder_save)))
decoder.load_state_dict(torch.load(os.path.join('checkpoints', FLAGS.decoder_save)))
discriminator.load_state_dict(torch.load(os.path.join('checkpoints', FLAGS.discriminator_save)))
"""
variable definition
"""
real_domain_labels = 1
fake_domain_labels = 0
X_1 = torch.FloatTensor(FLAGS.batch_size, FLAGS.num_channels, FLAGS.image_size, FLAGS.image_size)
X_2 = torch.FloatTensor(FLAGS.batch_size, FLAGS.num_channels, FLAGS.image_size, FLAGS.image_size)
X_3 = torch.FloatTensor(FLAGS.batch_size, FLAGS.num_channels, FLAGS.image_size, FLAGS.image_size)
domain_labels = torch.LongTensor(FLAGS.batch_size)
style_latent_space = torch.FloatTensor(FLAGS.batch_size, FLAGS.style_dim)
"""
loss definitions
"""
cross_entropy_loss = nn.CrossEntropyLoss()
'''
add option to run on GPU
'''
if FLAGS.cuda:
encoder.cuda()
decoder.cuda()
discriminator.cuda()
cross_entropy_loss.cuda()
X_1 = X_1.cuda()
X_2 = X_2.cuda()
X_3 = X_3.cuda()
domain_labels = domain_labels.cuda()
style_latent_space = style_latent_space.cuda()
"""
optimizer definition
"""
auto_encoder_optimizer = optim.Adam(
list(encoder.parameters()) + list(decoder.parameters()),
lr=FLAGS.initial_learning_rate,
betas=(FLAGS.beta_1, FLAGS.beta_2)
)
discriminator_optimizer = optim.Adam(
list(discriminator.parameters()),
lr=FLAGS.initial_learning_rate,
betas=(FLAGS.beta_1, FLAGS.beta_2)
)
generator_optimizer = optim.Adam(
list(encoder.parameters()) + list(decoder.parameters()),
lr=FLAGS.initial_learning_rate,
betas=(FLAGS.beta_1, FLAGS.beta_2)
)
"""
training
"""
if torch.cuda.is_available() and not FLAGS.cuda:
print("WARNING: You have a CUDA device, so you should probably run with --cuda")
if not os.path.exists('checkpoints'):
os.makedirs('checkpoints')
# load_saved is false when training is started from 0th iteration
if not FLAGS.load_saved:
with open(FLAGS.log_file, 'w') as log:
log.write('Epoch\tIteration\tReconstruction_loss\tKL_divergence_loss\t')
log.write('Generator_loss\tDiscriminator_loss\tDiscriminator_accuracy\n')
# load data set and create data loader instance
print('Loading MNIST paired dataset...')
paired_mnist = MNIST_Paired(root='mnist', download=True, train=True, transform=transform_config)
loader = cycle(DataLoader(paired_mnist, batch_size=FLAGS.batch_size, shuffle=True, num_workers=0, drop_last=True))
# initialise variables
discriminator_accuracy = 0.
# initialize summary writer
writer = SummaryWriter()
for epoch in range(FLAGS.start_epoch, FLAGS.end_epoch):
print('')
print('Epoch #' + str(epoch) + '..........................................................................')
for iteration in range(int(len(paired_mnist) / FLAGS.batch_size)):
# A. run the auto-encoder reconstruction
image_batch_1, image_batch_2, _ = next(loader)
auto_encoder_optimizer.zero_grad()
X_1.copy_(image_batch_1)
X_2.copy_(image_batch_2)
style_mu_1, style_logvar_1, class_1 = encoder(Variable(X_1))
style_1 = reparameterize(training=True, mu=style_mu_1, logvar=style_logvar_1)
kl_divergence_loss_1 = - 0.5 * torch.sum(1 + style_logvar_1 - style_mu_1.pow(2) - style_logvar_1.exp())
kl_divergence_loss_1 /= (FLAGS.batch_size * FLAGS.num_channels * FLAGS.image_size * FLAGS.image_size)
kl_divergence_loss_1.backward(retain_graph=True)
_, __, class_2 = encoder(Variable(X_2))
reconstructed_X_1 = decoder(style_1, class_1)
reconstructed_X_2 = decoder(style_1, class_2)
reconstruction_error_1 = mse_loss(reconstructed_X_1, Variable(X_1))
reconstruction_error_1.backward(retain_graph=True)
reconstruction_error_2 = mse_loss(reconstructed_X_2, Variable(X_1))
reconstruction_error_2.backward()
reconstruction_error = reconstruction_error_1 + reconstruction_error_2
kl_divergence_error = kl_divergence_loss_1
auto_encoder_optimizer.step()
# B. run the generator
for i in range(FLAGS.generator_times):
generator_optimizer.zero_grad()
image_batch_1, _, __ = next(loader)
image_batch_3, _, __ = next(loader)
domain_labels.fill_(real_domain_labels)
X_1.copy_(image_batch_1)
X_3.copy_(image_batch_3)
style_mu_1, style_logvar_1, _ = encoder(Variable(X_1))
style_1 = reparameterize(training=True, mu=style_mu_1, logvar=style_logvar_1)
kl_divergence_loss_1 = - 0.5 * torch.sum(1 + style_logvar_1 - style_mu_1.pow(2) - style_logvar_1.exp())
kl_divergence_loss_1 /= (FLAGS.batch_size * FLAGS.num_channels * FLAGS.image_size * FLAGS.image_size)
kl_divergence_loss_1.backward(retain_graph=True)
_, __, class_3 = encoder(Variable(X_3))
reconstructed_X_1_3 = decoder(style_1, class_3)
output_1 = discriminator(Variable(X_3), reconstructed_X_1_3)
generator_error_1 = cross_entropy_loss(output_1, Variable(domain_labels))
generator_error_1.backward(retain_graph=True)
style_latent_space.normal_(0., 1.)
reconstructed_X_latent_3 = decoder(Variable(style_latent_space), class_3)
output_2 = discriminator(Variable(X_3), reconstructed_X_latent_3)
generator_error_2 = cross_entropy_loss(output_2, Variable(domain_labels))
generator_error_2.backward()
generator_error = generator_error_1 + generator_error_2
kl_divergence_error += kl_divergence_loss_1
generator_optimizer.step()
# C. run the discriminator
for i in range(FLAGS.discriminator_times):
discriminator_optimizer.zero_grad()
# train discriminator on real data
domain_labels.fill_(real_domain_labels)
image_batch_1, _, __ = next(loader)
image_batch_2, image_batch_3, _ = next(loader)
X_1.copy_(image_batch_1)
X_2.copy_(image_batch_2)
X_3.copy_(image_batch_3)
real_output = discriminator(Variable(X_2), Variable(X_3))
discriminator_real_error = cross_entropy_loss(real_output, Variable(domain_labels))
discriminator_real_error.backward()
# train discriminator on fake data
domain_labels.fill_(fake_domain_labels)
style_mu_1, style_logvar_1, _ = encoder(Variable(X_1))
style_1 = reparameterize(training=False, mu=style_mu_1, logvar=style_logvar_1)
_, __, class_3 = encoder(Variable(X_3))
reconstructed_X_1_3 = decoder(style_1, class_3)
fake_output = discriminator(Variable(X_3), reconstructed_X_1_3)
discriminator_fake_error = cross_entropy_loss(fake_output, Variable(domain_labels))
discriminator_fake_error.backward()
# total discriminator error
discriminator_error = discriminator_real_error + discriminator_fake_error
# calculate discriminator accuracy for this step
target_true_labels = torch.cat((torch.ones(FLAGS.batch_size), torch.zeros(FLAGS.batch_size)), dim=0)
if FLAGS.cuda:
target_true_labels = target_true_labels.cuda()
discriminator_predictions = torch.cat((real_output, fake_output), dim=0)
_, discriminator_predictions = torch.max(discriminator_predictions, 1)
discriminator_accuracy = (discriminator_predictions.data == target_true_labels.long()
).sum().item() / (FLAGS.batch_size * 2)
if discriminator_accuracy < FLAGS.discriminator_limiting_accuracy:
discriminator_optimizer.step()
if (iteration + 1) % 50 == 0:
print('')
print('Epoch #' + str(epoch))
print('Iteration #' + str(iteration))
print('')
print('Reconstruction loss: ' + str(reconstruction_error.data.storage().tolist()[0]))
print('KL-Divergence loss: ' + str(kl_divergence_error.data.storage().tolist()[0]))
print('')
print('Generator loss: ' + str(generator_error.data.storage().tolist()[0]))
print('Discriminator loss: ' + str(discriminator_error.data.storage().tolist()[0]))
print('Discriminator accuracy: ' + str(discriminator_accuracy))
print('..........')
# write to log
with open(FLAGS.log_file, 'a') as log:
log.write('{0}\t{1}\t{2}\t{3}\t{4}\t{5}\t{6}\n'.format(
epoch,
iteration,
reconstruction_error.data.storage().tolist()[0],
kl_divergence_error.data.storage().tolist()[0],
generator_error.data.storage().tolist()[0],
discriminator_error.data.storage().tolist()[0],
discriminator_accuracy
))
# write to tensorboard
writer.add_scalar('Reconstruction loss', reconstruction_error.data.storage().tolist()[0],
epoch * (int(len(paired_mnist) / FLAGS.batch_size) + 1) + iteration)
writer.add_scalar('KL-Divergence loss', kl_divergence_error.data.storage().tolist()[0],
epoch * (int(len(paired_mnist) / FLAGS.batch_size) + 1) + iteration)
writer.add_scalar('Generator loss', generator_error.data.storage().tolist()[0],
epoch * (int(len(paired_mnist) / FLAGS.batch_size) + 1) + iteration)
writer.add_scalar('Discriminator loss', discriminator_error.data.storage().tolist()[0],
epoch * (int(len(paired_mnist) / FLAGS.batch_size) + 1) + iteration)
writer.add_scalar('Discriminator accuracy', discriminator_accuracy * 100,
epoch * (int(len(paired_mnist) / FLAGS.batch_size) + 1) + iteration)
# save model after every 5 epochs
if (epoch + 1) % 5 == 0 or (epoch + 1) == FLAGS.end_epoch:
torch.save(encoder.state_dict(), os.path.join('checkpoints', FLAGS.encoder_save))
torch.save(decoder.state_dict(), os.path.join('checkpoints', FLAGS.decoder_save))
torch.save(discriminator.state_dict(), os.path.join('checkpoints', FLAGS.discriminator_save))
def training_procedure(FLAGS):
"""
model definition
"""
encoder = Encoder(nv_dim=FLAGS.nv_dim, nc_dim=FLAGS.nc_dim)
encoder.apply(weights_init)
decoder = Decoder(nv_dim=FLAGS.nv_dim, nc_dim=FLAGS.nc_dim)
decoder.apply(weights_init)
discriminator = Discriminator()
discriminator.apply(weights_init)
# load saved models if load_saved flag is true
if FLAGS.load_saved:
encoder.load_state_dict(
torch.load(os.path.join('checkpoints', FLAGS.encoder_save)))
decoder.load_state_dict(
torch.load(os.path.join('checkpoints', FLAGS.decoder_save)))
discriminator.load_state_dict(
torch.load(os.path.join('checkpoints', FLAGS.discriminator_save)))
"""
variable definition
"""
real_domain_labels = 1
fake_domain_labels = 0
X_1 = torch.FloatTensor(FLAGS.batch_size, FLAGS.num_channels,
FLAGS.image_size, FLAGS.image_size)
X_2 = torch.FloatTensor(FLAGS.batch_size, FLAGS.num_channels,
FLAGS.image_size, FLAGS.image_size)
X_3 = torch.FloatTensor(FLAGS.batch_size, FLAGS.num_channels,
FLAGS.image_size, FLAGS.image_size)
domain_labels = torch.LongTensor(FLAGS.batch_size)
"""
loss definitions
"""
cross_entropy_loss = nn.CrossEntropyLoss()
'''
add option to run on GPU
'''
if FLAGS.cuda:
encoder.cuda()
decoder.cuda()
discriminator.cuda()
cross_entropy_loss.cuda()
X_1 = X_1.cuda()
X_2 = X_2.cuda()
X_3 = X_3.cuda()
domain_labels = domain_labels.cuda()
"""
optimizer definition
"""
auto_encoder_optimizer = optim.Adam(list(encoder.parameters()) +
list(decoder.parameters()),
lr=FLAGS.initial_learning_rate,
betas=(FLAGS.beta_1, FLAGS.beta_2))
discriminator_optimizer = optim.Adam(list(discriminator.parameters()),
lr=FLAGS.initial_learning_rate,
betas=(FLAGS.beta_1, FLAGS.beta_2))
generator_optimizer = optim.Adam(list(encoder.parameters()) +
list(decoder.parameters()),
lr=FLAGS.initial_learning_rate,
betas=(FLAGS.beta_1, FLAGS.beta_2))
"""
training
"""
if torch.cuda.is_available() and not FLAGS.cuda:
print(
"WARNING: You have a CUDA device, so you should probably run with --cuda"
)
if not os.path.exists('checkpoints'):
os.makedirs('checkpoints')
if not os.path.exists('reconstructed_images'):
os.makedirs('reconstructed_images')
# load_saved is false when training is started from 0th iteration
if not FLAGS.load_saved:
with open(FLAGS.log_file, 'w') as log:
log.write('Epoch\tIteration\tReconstruction_loss\t')
log.write(
'Generator_loss\tDiscriminator_loss\tDiscriminator_accuracy\n')
# load data set and create data loader instance
print('Loading MNIST paired dataset...')
paired_mnist = MNIST_Paired(root='mnist',
download=True,
train=True,
transform=transform_config)
loader = cycle(
DataLoader(paired_mnist,
batch_size=FLAGS.batch_size,
shuffle=True,
num_workers=0,
drop_last=True))
# initialise variables
discriminator_accuracy = 0.
# initialize summary writer
writer = SummaryWriter()
for epoch in range(FLAGS.start_epoch, FLAGS.end_epoch):
print('')
print(
'Epoch #' + str(epoch) +
'..........................................................................'
)
for iteration in range(int(len(paired_mnist) / FLAGS.batch_size)):
# A. run the auto-encoder reconstruction
image_batch_1, image_batch_2, labels_batch_1 = next(loader)
auto_encoder_optimizer.zero_grad()
X_1.copy_(image_batch_1)
X_2.copy_(image_batch_2)
nv_1, nc_1 = encoder(Variable(X_1))
nv_2, nc_2 = encoder(Variable(X_2))
reconstructed_X_1 = decoder(nv_1, nc_2)
reconstructed_X_2 = decoder(nv_2, nc_1)
reconstruction_error_1 = mse_loss(reconstructed_X_1, Variable(X_1))
reconstruction_error_1.backward(retain_graph=True)
reconstruction_error_2 = mse_loss(reconstructed_X_2, Variable(X_2))
reconstruction_error_2.backward()
reconstruction_error = reconstruction_error_1 + reconstruction_error_2
if FLAGS.train_auto_encoder:
auto_encoder_optimizer.step()
# B. run the adversarial part of the architecture
# B. a) run the discriminator
for i in range(FLAGS.discriminator_times):
discriminator_optimizer.zero_grad()
# train discriminator on real data
domain_labels.fill_(real_domain_labels)
image_batch_1, image_batch_2, labels_batch_1 = next(loader)
X_1.copy_(image_batch_1)
X_2.copy_(image_batch_2)
real_output = discriminator(Variable(X_1), Variable(X_2))
discriminator_real_error = FLAGS.disc_coef * cross_entropy_loss(
real_output, Variable(domain_labels))
discriminator_real_error.backward()
# train discriminator on fake data
domain_labels.fill_(fake_domain_labels)
image_batch_3, _, labels_batch_3 = next(loader)
X_3.copy_(image_batch_3)
nv_3, nc_3 = encoder(Variable(X_3))
# reconstruction is taking common factor from X_1 and varying factor from X_3
reconstructed_X_3_1 = decoder(nv_3, encoder(Variable(X_1))[1])
fake_output = discriminator(Variable(X_1), reconstructed_X_3_1)
discriminator_fake_error = FLAGS.disc_coef * cross_entropy_loss(
fake_output, Variable(domain_labels))
discriminator_fake_error.backward()
# total discriminator error
discriminator_error = discriminator_real_error + discriminator_fake_error
# calculate discriminator accuracy for this step
target_true_labels = torch.cat((torch.ones(
FLAGS.batch_size), torch.zeros(FLAGS.batch_size)),
dim=0)
if FLAGS.cuda:
target_true_labels = target_true_labels.cuda()
discriminator_predictions = torch.cat(
(real_output, fake_output), dim=0)
_, discriminator_predictions = torch.max(
discriminator_predictions, 1)
discriminator_accuracy = (discriminator_predictions.data
== target_true_labels.long()).sum(
).item() / (FLAGS.batch_size * 2)
if discriminator_accuracy < FLAGS.discriminator_limiting_accuracy and FLAGS.train_discriminator:
discriminator_optimizer.step()
# B. b) run the generator
for i in range(FLAGS.generator_times):
generator_optimizer.zero_grad()
image_batch_1, _, labels_batch_1 = next(loader)
image_batch_3, __, labels_batch_3 = next(loader)
domain_labels.fill_(real_domain_labels)
X_1.copy_(image_batch_1)
X_3.copy_(image_batch_3)
nv_3, nc_3 = encoder(Variable(X_3))
# reconstruction is taking common factor from X_1 and varying factor from X_3
reconstructed_X_3_1 = decoder(nv_3, encoder(Variable(X_1))[1])
output = discriminator(Variable(X_1), reconstructed_X_3_1)
generator_error = FLAGS.gen_coef * cross_entropy_loss(
output, Variable(domain_labels))
generator_error.backward()
if FLAGS.train_generator:
generator_optimizer.step()
# print progress after 10 iterations
if (iteration + 1) % 10 == 0:
print('')
print('Epoch #' + str(epoch))
print('Iteration #' + str(iteration))
print('')
print('Reconstruction loss: ' +
str(reconstruction_error.data.storage().tolist()[0]))
print('Generator loss: ' +
str(generator_error.data.storage().tolist()[0]))
print('')
print('Discriminator loss: ' +
str(discriminator_error.data.storage().tolist()[0]))
print('Discriminator accuracy: ' + str(discriminator_accuracy))
print('..........')
# write to log
with open(FLAGS.log_file, 'a') as log:
log.write('{0}\t{1}\t{2}\t{3}\t{4}\t{5}\n'.format(
epoch, iteration,
reconstruction_error.data.storage().tolist()[0],
generator_error.data.storage().tolist()[0],
discriminator_error.data.storage().tolist()[0],
discriminator_accuracy))
# write to tensorboard
writer.add_scalar(
'Reconstruction loss',
reconstruction_error.data.storage().tolist()[0],
epoch * (int(len(paired_mnist) / FLAGS.batch_size) + 1) +
iteration)
writer.add_scalar(
'Generator loss',
generator_error.data.storage().tolist()[0],
epoch * (int(len(paired_mnist) / FLAGS.batch_size) + 1) +
iteration)
writer.add_scalar(
'Discriminator loss',
discriminator_error.data.storage().tolist()[0],
epoch * (int(len(paired_mnist) / FLAGS.batch_size) + 1) +
iteration)
# save model after every 5 epochs
if (epoch + 1) % 5 == 0 or (epoch + 1) == FLAGS.end_epoch:
torch.save(encoder.state_dict(),
os.path.join('checkpoints', FLAGS.encoder_save))
torch.save(decoder.state_dict(),
os.path.join('checkpoints', FLAGS.decoder_save))
torch.save(discriminator.state_dict(),
os.path.join('checkpoints', FLAGS.discriminator_save))
"""
save reconstructed images and style swapped image generations to check progress
"""
image_batch_1, image_batch_2, labels_batch_1 = next(loader)
image_batch_3, _, __ = next(loader)
X_1.copy_(image_batch_1)
X_2.copy_(image_batch_2)
X_3.copy_(image_batch_3)
nv_1, nc_1 = encoder(Variable(X_1))
nv_2, nc_2 = encoder(Variable(X_2))
nv_3, nc_3 = encoder(Variable(X_3))
reconstructed_X_1 = decoder(nv_1, nc_2)
reconstructed_X_3_2 = decoder(nv_3, nc_2)
# save input image batch
image_batch = np.transpose(X_1.cpu().numpy(), (0, 2, 3, 1))
image_batch = np.concatenate(
(image_batch, image_batch, image_batch), axis=3)
imshow_grid(image_batch, name=str(epoch) + '_original', save=True)
# save reconstructed batch
reconstructed_x = np.transpose(
reconstructed_X_1.cpu().data.numpy(), (0, 2, 3, 1))
reconstructed_x = np.concatenate(
(reconstructed_x, reconstructed_x, reconstructed_x), axis=3)
imshow_grid(reconstructed_x,
name=str(epoch) + '_target',
save=True)
# save cross reconstructed batch
style_batch = np.transpose(X_3.cpu().numpy(), (0, 2, 3, 1))
style_batch = np.concatenate(
(style_batch, style_batch, style_batch), axis=3)
imshow_grid(style_batch, name=str(epoch) + '_style', save=True)
reconstructed_style = np.transpose(
reconstructed_X_3_2.cpu().data.numpy(), (0, 2, 3, 1))
reconstructed_style = np.concatenate(
(reconstructed_style, reconstructed_style,
reconstructed_style),
axis=3)
imshow_grid(reconstructed_style,
name=str(epoch) + '_style_target',
save=True)
