if batches_done % opt.sample_interval == 0:
# Keep a record of losses for plotting.
epochs.append(epoch + i / len(dataloader_train))
g_losses.append(g_loss.item())
d_losses.append(d_loss.item())
# Generate 5 images from the validation set.
batch_A, batch_B = next(iter(dataloader_val))
imgs_val_A = Variable(batch_A.type(Tensor))
imgs_val_B = Variable(batch_B.type(Tensor))
imgs_fake_B = generator(imgs_val_A).detach().data
# For visualization purposes.
imgs_val_A = imgs_val_A.add_(1).div_(2)
imgs_fake_B = imgs_fake_B.add_(1).div_(2)
fake_val = torch.cat((imgs_val_A, imgs_fake_B), dim=2)
# Display results on visdom page.
vis.line(X=torch.stack([Tensor(epochs)] * len(loss_legend),
dim=1),
Y=torch.stack((Tensor(d_losses), Tensor(g_losses)),
dim=1),
opts={
'title': 'loss over time',
'legend': loss_legend,
'xlabel': 'epoch',
'ylabel': 'loss',
'width': 512,
'height': 512
imgs_fake_hr = generator(imgs_val_lr).detach().data
# For visualization purposes.
imgs_val_lr = torch.nn.functional.upsample(
imgs_val_lr,
size=(imgs_fake_hr.size(2), imgs_fake_hr.size(3)),
mode='bilinear')
imgs_val_lr = imgs_val_lr.mul_(
Tensor([0.229, 0.224, 0.225]).view(1, 3, 1, 1)).add_(
Tensor([0.485, 0.456, 0.406]).view(1, 3, 1, 1))
imgs_val_hr = imgs_val_hr.mul_(
Tensor([0.229, 0.224, 0.225]).view(1, 3, 1, 1)).add_(
Tensor([0.485, 0.456, 0.406]).view(1, 3, 1, 1))
imgs_fake_hr = imgs_fake_hr.add_(
torch.abs(torch.min(imgs_fake_hr))).div_(
torch.max(imgs_fake_hr) - torch.min(imgs_fake_hr))
fake_val = torch.cat((imgs_val_lr, imgs_val_hr, imgs_fake_hr),
dim=2)
# Display results on visdom page.
vis.line(X=torch.stack([Tensor(epochs)] * len(loss_legend),
dim=1),
Y=torch.stack((Tensor(d_losses), Tensor(g_losses)),
dim=1),
opts={
'title': 'loss over time',
'legend': loss_legend,
'xlabel': 'epoch',
'ylabel': 'loss',
'width': 512,
