class Model():
def __init__(self, args):
self.args = args
self.pretrained = False
self.epoch = 0
self.G = Generator()
self.D = Discriminator()
self.g_optimizer = optim.Adam(self.G.parameters(), lr=1E-4)
self.d_optimizer = optim.Adam(self.D.parameters(), lr=1E-4)
self.g_scheduler = optim.lr_scheduler.StepLR(self.g_optimizer,
step_size=40)
self.d_scheduler = optim.lr_scheduler.StepLR(self.d_optimizer,
step_size=40)
self.train_losses = []
self.val_losses = []
if args.load_model:
self._load_state(args.load_model)
# extract all layers prior to the last softmax of VGG-19
vgg19_layers = list(models.vgg19(pretrained=True).features)[:36]
self.vgg19 = nn.Sequential(*vgg19_layers).eval()
for param in self.vgg19.parameters():
param.requires_grad = False
self.mse_loss = torch.nn.MSELoss()
self.bce_loss = torch.nn.BCELoss()
def train(self, train_dataloader, val_dataloader=None):
self.D.to(device)
self.G.to(device)
self.vgg19.to(device)
""" Pretrain Generator """
if not self.pretrained:
log_message("Starting pretraining")
self._pretrain(train_dataloader)
self._save_state()
if val_dataloader:
val_g_loss, _ = self.evaluate(val_dataloader)
log_message("Pretrain G loss: {:.4f}".format(val_g_loss))
""" Real Training """
log_message("Starting training")
while self.epoch < self.args.epochs:
# Train one epoch
self.D.train()
self.G.train()
g_loss, d_loss = self._run_epoch(train_dataloader, train=True)
self.train_losses.append([g_loss, d_loss])
self.g_scheduler.step()
self.d_scheduler.step()
self.epoch += 1
log_message("Epoch: {}/{}".format(self.epoch, self.args.epochs))
# Print evaluation
train_string = "Train G loss: {:.4f} | Train D loss: {:.4f}".format(
g_loss, d_loss)
if self.epoch % self.args.eval_epochs == 0:
if val_dataloader:
val_g_loss, val_d_loss = self.evaluate(val_dataloader)
self.val_losses.append([val_g_loss, val_d_loss])
train_string += " | Val G loss: {:.4f} | Val D loss: {:.4f}".format(
val_g_loss, val_d_loss)
log_message(train_string)
# Save the model
if self.epoch % self.args.save_epochs == 0:
self._save_state()
log_message("Finished training")
self._save_state()
def evaluate(self, dataloader):
self.D.eval()
self.G.eval()
with torch.no_grad():
return self._run_epoch(dataloader, train=False)
def generate(self, dataloader):
def to_image(tensor):
array = tensor.data.cpu().numpy()
array = array.transpose((1, 2, 0))
array = np.clip(255.0 * (array + 1) / 2, 0, 255)
array = np.uint8(array)
return Image.fromarray(array)
self.D.eval()
self.G.eval()
if not os.path.exists(self.args.generate_dir):
os.mkdir(self.args.generate_dir)
with torch.no_grad():
for batch in dataloader:
low_res = batch['low_res'].to(device)
hi_res = batch['high_res']
generated = self.G(low_res)
for i in range(len(generated)):
naive = np.clip(
255.0 * low_res[i].data.cpu().numpy().transpose(
(1, 2, 0)), 0, 255)
naive = Image.fromarray(np.uint8(naive))
naive = naive.resize((96, 96), Image.BICUBIC)
fake_im = to_image(generated[i])
real_im = to_image(hi_res[i])
naive.save(
os.path.join(self.args.generate_dir,
"{}_naive.png".format(i)))
fake_im.save(
os.path.join(self.args.generate_dir,
"{}_fake.png".format(i)))
real_im.save(
os.path.join(self.args.generate_dir,
"{}_real.png".format(i)))
if i > 10:
return
def _load_state(self, fname):
if torch.cuda.is_available():
map_location = lambda storage, loc: storage.cuda()
else:
map_location = 'cpu'
state = torch.load(fname, map_location=map_location)
self.pretrained = state["pretrained"]
self.epoch = state["epoch"]
self.train_losses = state["train_losses"]
self.val_losses = state["val_losses"]
self.G.load_state_dict(state["G"])
self.D.load_state_dict(state["D"])
self.g_optimizer.load_state_dict(state["g_optimizer"])
self.d_optimizer.load_state_dict(state["d_optimizer"])
self.g_scheduler.load_state_dict(state["g_scheduler"])
self.d_scheduler.load_state_dict(state["d_scheduler"])
for state in self.d_optimizer.state.values():
for k, v in state.items():
if torch.is_tensor(v):
state[k] = v.to(device)
for state in self.g_optimizer.state.values():
for k, v in state.items():
if torch.is_tensor(v):
state[k] = v.to(device)
def _save_state(self):
if not os.path.exists(self.args.save_dir):
os.mkdir(self.args.save_dir)
fname = "%s/save_%d.pkl" % (self.args.save_dir, self.epoch)
state = {
"pretrained": self.pretrained,
"epoch": self.epoch,
"G": self.G.state_dict(),
"D": self.D.state_dict(),
"g_optimizer": self.g_optimizer.state_dict(),
"d_optimizer": self.d_optimizer.state_dict(),
"g_scheduler": self.g_scheduler.state_dict(),
"d_scheduler": self.d_scheduler.state_dict(),
"train_losses": self.train_losses,
"val_losses": self.val_losses
}
torch.save(state, fname)
def _pretrain(self, dataloader):
self.G.train()
for i in range(self.args.pretrain_epochs):
log_message("Pretrain Epoch: {}/{}".format(
i, self.args.pretrain_epochs))
for batch in dataloader:
low_res = batch['low_res'].to(device)
high_res = batch['high_res'].to(device)
self.g_optimizer.zero_grad()
generated = self.G(low_res)
# Optimize pixel loss
g_loss = self.mse_loss(generated, high_res)
g_loss.backward()
self.g_optimizer.step()
self.pretrained = True
def _run_epoch(self, dataloader, train):
g_losses, d_losses = [], []
for batch in dataloader:
low_res = batch['low_res'].to(device)
high_res = batch['high_res'].to(device)
batch_size = high_res.size(0)
real = torch.ones((batch_size, 1), requires_grad=False).to(device)
fake = torch.zeros((batch_size, 1), requires_grad=False).to(device)
""" Discriminator """
generated = self.G(low_res)
self.d_optimizer.zero_grad()
real_loss = self.bce_loss(self.D(high_res), real)
fake_loss = self.bce_loss(self.D(generated), fake)
d_loss = real_loss + fake_loss
d_losses.append(d_loss.item())
if train:
d_loss.backward()
self.d_optimizer.step()
""" Generator """
generated = self.G(low_res)
self.g_optimizer.zero_grad()
# take a [B, C, W, H] batch of [-1, 1] images, normalize, then run through vgg19
def vgg_features(image):
mean = torch.tensor(
[0.485, 0.456,
0.406]).unsqueeze(0).unsqueeze(2).unsqueeze(3).to(device)
std = torch.tensor(
[0.229, 0.224,
0.225]).unsqueeze(0).unsqueeze(2).unsqueeze(3).to(device)
image = (image + 1) / 2
image = (image - mean) / std
return self.vgg19(image)
pixel_loss = self.mse_loss(high_res, generated)
content_loss = self.mse_loss(vgg_features(high_res),
vgg_features(generated))
adversarial_loss = self.bce_loss(self.D(generated), real)
g_loss = pixel_loss + 0.006 * content_loss + 1E-3 * adversarial_loss
g_losses.append(g_loss.item())
if train:
g_loss.backward()
self.g_optimizer.step()
return np.mean(g_losses), np.mean(d_losses)
class Solver(object):
def __init__(self, config, data_loader):
self.generator = None
self.discriminator = None
self.g_optimizer = None
self.d_optimizer = None
self.g_conv_dim = config.g_conv_dim
self.d_conv_dim = config.d_conv_dim
self.z_dim = config.z_dim
self.beta1 = config.beta1
self.beta2 = config.beta2
self.image_size = config.image_size
self.data_loader = data_loader
self.num_epochs = config.num_epochs
self.batch_size = config.batch_size
self.sample_size = config.sample_size
self.lr = config.lr
self.log_step = config.log_step
self.sample_step = config.sample_step
self.sample_path = config.sample_path
self.model_path = config.model_path
self.epoch = config.epoch
self.build_model()
self.plotter = Plotter()
def build_model(self):
"""Build generator and discriminator."""
self.generator = Generator(z_dim=self.z_dim)
print(count_parameters(self.generator))
self.discriminator = Discriminator()
print(count_parameters(self.discriminator))
self.g_optimizer = optim.Adam(self.generator.parameters(),
self.lr, (self.beta1, self.beta2))
self.d_optimizer = optim.Adam(self.discriminator.parameters(),
self.lr*1, (self.beta1, self.beta2))
if self.epoch:
g_path = os.path.join(self.model_path, 'generator-%d.pkl' % self.epoch)
d_path = os.path.join(self.model_path, 'discriminator-%d.pkl' % self.epoch)
g_optim_path = os.path.join(self.model_path, 'gen-optim-%d.pkl' % self.epoch)
d_optim_path = os.path.join(self.model_path, 'dis-optim-%d.pkl' % self.epoch)
self.generator.load_state_dict(torch.load(g_path))
self.discriminator.load_state_dict(torch.load(d_path))
self.g_optimizer.load_state_dict(torch.load(g_optim_path))
self.d_optimizer.load_state_dict(torch.load(d_optim_path))
if torch.cuda.is_available():
self.generator.cuda()
self.discriminator.cuda()
def to_variable(self, x):
"""Convert tensor to variable."""
if torch.cuda.is_available():
x = x.cuda()
return Variable(x)
def to_data(self, x):
"""Convert variable to tensor."""
if torch.cuda.is_available():
x = x.cpu()
return x.data
def reset_grad(self):
"""Zero the gradient buffers."""
self.discriminator.zero_grad()
self.generator.zero_grad()
def denorm(self, x):
"""Convert range (-1, 1) to (0, 1)"""
out = (x + 1) / 2
return out.clamp(0, 1)
def train(self):
"""Train generator and discriminator."""
fixed_noise = self.to_variable(torch.randn(self.batch_size, self.z_dim))
total_step = len(self.data_loader)
for epoch in range(self.epoch, self.epoch + self.num_epochs) if self.epoch else range(self.num_epochs):
for i, images in enumerate(self.data_loader):
if len(images) != self.batch_size:
continue
# self.plotter.draw_kernels(self.discriminator)
for p in self.discriminator.parameters():
p.requires_grad = True
#===================== Train D =====================#
images = self.to_variable(images)
images.retain_grad()
batch_size = images.size(0)
noise = self.to_variable(torch.randn(batch_size, self.z_dim))
# Train D to recognize real images as real.
outputs = self.discriminator(images)
real_loss = torch.mean((outputs - 1) ** 2) # L2 loss instead of Binary cross entropy loss (this is optional for stable training)
# real_loss = torch.mean(outputs - 1)
# Train D to recognize fake images as fake.
fake_images = self.generator(noise)
fake_images.retain_grad()
outputs = self.discriminator(fake_images)
fake_loss = torch.mean(outputs ** 2)
# fake_loss = torch.mean(outputs)
# gradient penalty
gp_loss = calc_gradient_penalty(self.discriminator, images, fake_images)
# Backprop + optimize
d_loss = fake_loss + real_loss + gp_loss
self.reset_grad()
d_loss.backward()
self.d_optimizer.step()
if i % 10 == 0:
self.plotter.draw_activations(fake_images.grad[0], original=fake_images[0])
g_losses = []
for p in self.discriminator.parameters():
p.requires_grad = False
#===================== Train G =====================#
for g_batch in range(5):
noise = self.to_variable(torch.randn(batch_size, self.z_dim))
# Train G so that D recognizes G(z) as real.
fake_images = self.generator(noise)
outputs = self.discriminator(fake_images)
g_loss = torch.mean((outputs - 1) ** 2)
# g_loss = -torch.mean(outputs)
# Backprop + optimize
self.reset_grad()
g_loss.backward()
# if g_loss.item() < 0.5 * d_loss.item():
# break
self.g_optimizer.step()
g_losses.append("%.3f"%g_loss.clone().item())
# print the log info
if (i+1) % self.log_step == 0:
print('Epoch [%d/%d], Step[%d/%d], d_real_loss: %.4f, '
'd_fake_loss: %.4f, gp_loss: %s, g_loss: %s'
%(epoch+1, self.num_epochs, i+1, total_step,
real_loss.item(), fake_loss.item(), gp_loss.item(), ", ".join(g_losses)))
# save the sampled images
# print((i+1)%self.sample_step)
if (i) % self.sample_step == 0:
print("saving samples")
fake_images = self.generator(fixed_noise)
if not os.path.exists(self.sample_path):
os.makedirs(self.sample_path)
torchvision.utils.save_image(self.denorm(fake_images.data),
os.path.join(self.sample_path,
'fake_samples-%d-%d.png' %(epoch+1, i+1)))
# save the model parameters for each epoch
if epoch % 5 == 0:
if not os.path.exists(self.model_path):
os.mkdir(self.model_path)
g_path = os.path.join(self.model_path, 'generator-%d.pkl' %(epoch+1))
d_path = os.path.join(self.model_path, 'discriminator-%d.pkl' %(epoch+1))
g_optim_path = os.path.join(self.model_path, 'gen-optim-%d.pkl' % (epoch + 1))
d_optim_path = os.path.join(self.model_path, 'dis-optim-%d.pkl' % (epoch + 1))
torch.save(self.generator.state_dict(), g_path)
torch.save(self.discriminator.state_dict(), d_path)
torch.save(self.g_optimizer.state_dict(), g_optim_path)
torch.save(self.d_optimizer.state_dict(), d_optim_path)
def sample(self):
# Load trained parameters
g_path = os.path.join(self.model_path, 'generator-%d.pkl' % self.num_epochs)
d_path = os.path.join(self.model_path, 'discriminator-%d.pkl' % self.num_epochs)
self.generator.load_state_dict(torch.load(g_path))
self.discriminator.load_state_dict(torch.load(d_path))
self.generator.eval()
self.discriminator.eval()
# Sample the images
noise = self.to_variable(torch.randn(self.sample_size, self.z_dim))
fake_images = self.generator(noise)
sample_path = os.path.join(self.sample_path, 'fake_samples-final.png')
torchvision.utils.save_image(self.denorm(fake_images.data), sample_path, nrow=12)
print("Saved sampled images to '%s'" %sample_path)
