sr = netG(lr)
batch_mse = ((sr - hr)**2).data.mean()
valing_results['mse'] += batch_mse * batch_size
batch_ssim = pytorch_ssim.ssim(sr, hr).data[0]
valing_results['ssims'] += batch_ssim * batch_size
valing_results['psnr'] = 10 * log10(
1 / (valing_results['mse'] / valing_results['batch_sizes']))
valing_results[
'ssim'] = valing_results['ssims'] / valing_results['batch_sizes']
val_bar.set_description(
desc='[converting LR images to SR images] PSNR: %.4f dB SSIM: %.4f'
% (valing_results['psnr'], valing_results['ssim']))
val_images.extend([
display_transform()(val_hr_restore.squeeze(0)),
display_transform()(hr.data.cpu().squeeze(0)),
display_transform()(sr.data.cpu().squeeze(0))
])
val_images = torch.stack(val_images)
val_images = torch.chunk(val_images, val_images.size(0) // 15)
val_save_bar = tqdm(val_images, desc='[saving training results]')
index = 1
for image in val_save_bar:
image = utils.make_grid(image, nrow=3, padding=5)
utils.save_image(image,
out_path + 'epoch_%d_index_%d.png' % (epoch, index),
padding=5)
index += 1
# save model parameters
for image_name, lr_image, hr_restore_img, hr_image in test_bar:
image_name = image_name[0]
lr_image = Variable(lr_image, volatile=True)
hr_image = Variable(hr_image, volatile=True)
# if torch.cuda.is_available():
# lr_image = lr_image.cuda()
# hr_image = hr_image.cuda()
sr_image = model(lr_image)
mse = ((hr_image - sr_image)**2).data.mean()
psnr = 10 * log10(1 / mse)
ssim = pytorch_ssim.ssim(sr_image, hr_image).data[0]
test_images = torch.stack([
display_transform()(hr_restore_img.squeeze(0)),
display_transform()(hr_image.data.cpu().squeeze(0)),
display_transform()(sr_image.data.cpu().squeeze(0))
])
image = utils.make_grid(test_images, nrow=3, padding=5)
utils.save_image(image, OUT_PATH + image_name, padding=5)
# save psnr\ssim
results['psnr'].append(psnr)
results['ssim'].append(ssim)
out_path = 'logs/statistics/'
saved_results = {'psnr': [], 'ssim': []}
for item in results.values():
psnr = np.array(item['psnr'])
ssim = np.array(item['ssim'])
valing_results['mse'] += batch_mse * batch_size
# valing_results['mse'] = valing_results['mse'] + batch_mse * batch_size
valing_results['ssims'] += batch_ssim * batch_size
# batch_ssim = pytorch_ssim.ssim(sr, hr).item()
valing_results['ssims'] = valing_results['ssims'] + batch_ssim * batch_size
valing_results['psnr'] = 10 * log10((hr.max()**2) / (valing_results['mse'] / valing_results['batch_sizes']))
valing_results['ssim'] = valing_results['ssims'] / valing_results['batch_sizes']
val_bar.set_description(
desc='[converting LR images to SR images] PSNR: %.4f dB SSIM: %.4f' % (
valing_results['psnr'], valing_results['ssim']))
# val_images.extend(
# [display_transform()(val_hr_restore.squeeze(0)), display_transform()(hr.data.cpu().squeeze(0)),
# display_transform()(sr.data.cpu().squeeze(0))])
val_images.extend(
[display_transform()(val_hr_restore.squeeze(0)), display_transform()(hr.detach().cpu().squeeze(0)),
display_transform()(sr.detach().cpu().squeeze(0))])
val_images = torch.stack(val_images)
val_images = torch.chunk(val_images, val_images.size(0) // 15)
val_save_bar = tqdm(val_images, desc='[saving training results]')
# val_image = torch.stack(val_images)
# val_images = torch.chunk(val_image, val_image.size(0) // 15)
# val_save_bar = tqdm(val_images, desc='[saving training results]')
index = 1
for image in val_save_bar:
image = utils.make_grid(image, nrow=3, padding=5)
utils.save_image(image, out_path + 'epoch_%d_index_%d.png' % (epoch, index), padding=5)
# img = utils.make_grid(image, nrow=3, padding=5)
# utils.save_image(img, out_path + 'epoch_%d_index_%d.png' % (epoch, index), padding=5)
index += 1
# index = index + 1
image_name = image_name[0]
lr_image = Variable(lr_image, volatile=True)
hr_image = Variable(hr_image, volatile=True)
if torch.cuda.is_available():
lr_image = lr_image.cuda()
hr_image = hr_image.cuda()
# print(lr_image.size())
sr_image = model(lr_image)
# print(sr_image.size())
# print(hr_image.size())
mse = ((hr_image - sr_image)**2).data.mean()
psnr = 10 * log10(1 / mse)
ssim = pytorch_ssim.ssim(sr_image, hr_image).data[0]
test_images = torch.stack([
display_transform()(hr_restore_img.squeeze(0)),
display_transform()(hr_image.data.cpu().squeeze(0)),
display_transform()(sr_image.data.cpu().squeeze(0))
])
image = utils.make_grid(test_images, nrow=3, padding=5)
utils.save_image(image,
out_path + image_name.split('.')[0] +
'_psnr_%.4f_ssim_%.4f.' % (psnr, ssim) +
image_name.split('.')[-1],
padding=5)
# save psnr\ssim
# print(image_name)
results['psnr'].append(psnr)
results['ssim'].append(ssim)
batch_mse = ((sr - hr)**2).data.mean()
valing_results['mse'] += batch_mse * batch_size
batch_ssim = pytorch_ssim.ssim(sr, hr).item()
valing_results['ssims'] += batch_ssim * batch_size
valing_results['psnr'] = 10 * log10(
1 /
(valing_results['mse'] / valing_results['batch_sizes']))
valing_results['ssim'] = valing_results[
'ssims'] / valing_results['batch_sizes']
val_bar.set_description(
desc=
'[converting LR images to SR images] PSNR: %.4f dB SSIM: %.4f'
% (valing_results['psnr'], valing_results['ssim']))
image1 = utils.make_grid(display_transform()(
val_hr_restore.squeeze(0)),
nrow=3,
padding=5)
# image1.，
# print(image1.size())
utils.save_image(image1,
out_path +
'epoch_%d_index_%d_hr_restore.png' %
(epoch, index),
padding=5)
image2 = utils.make_grid(display_transform()(
hr.data.cpu().squeeze(0)),
nrow=3,
padding=5)
# print(image2.size())
for image_name, lr_image, hr_restore_img, hr_image in test_bar:
image_name = image_name[0]
with torch.no_grad():
lr_image = Variable(lr_image)
hr_image = Variable(hr_image)
if torch.cuda.is_available():
lr_image = lr_image.cuda()
hr_image = hr_image.cuda()
with torch.no_grad():
sr_image = model(lr_image)
mse = ((hr_image - sr_image) ** 2).data.mean()
psnr = 10 * log10(1 / mse)
ssim = pytorch_ssim.ssim(sr_image, hr_image).data
test_images = torch.stack(
[display_transform()(lr_image.squeeze(0)), display_transform()(sr_image.data.cpu().squeeze(0)),
display_transform()(hr_image.data.cpu().squeeze(0))])
image = utils.make_grid(test_images, nrow=3, padding=5)
utils.save_image(image, out_path + image_name.split('.')[0] + '_psnr_%.4f_ssim_%.4f.' % (psnr, ssim) +
image_name.split('.')[-1], padding=5)
results[image_name.split('_')[0]]['psnr'].append(psnr)
results[image_name.split('_')[0]]['ssim'].append(ssim)
out_path = 'statistics/'
saved_results = {'psnr': [], 'ssim': []}
for item in results.values():
psnr = np.array(item['psnr'])
ssim = np.array([i.item() for i in item['ssim']])
if (len(psnr) == 0) or (len(ssim) == 0):
psnr = 'No data'
def main():
parser = ArgumentParser()
parser.add_argument("--augmentation", action='store_true')
parser.add_argument("--train-dataset-percentage", type=float, default=100)
parser.add_argument("--val-dataset-percentage", type=int, default=100)
parser.add_argument("--label-smoothing", type=float, default=0.9)
parser.add_argument("--validation-frequency", type=int, default=1)
args = parser.parse_args()
ENABLE_AUGMENTATION = args.augmentation
TRAIN_DATASET_PERCENTAGE = args.train_dataset_percentage
VAL_DATASET_PERCENTAGE = args.val_dataset_percentage
LABEL_SMOOTHING_FACTOR = args.label_smoothing
VALIDATION_FREQUENCY = args.validation_frequency
if ENABLE_AUGMENTATION:
augment_batch = AugmentPipe()
augment_batch.to(device)
else:
augment_batch = lambda x: x
augment_batch.p = 0
NUM_ADV_EPOCHS = round(NUM_ADV_BASELINE_EPOCHS /
(TRAIN_DATASET_PERCENTAGE / 100))
NUM_PRETRAIN_EPOCHS = round(NUM_BASELINE_PRETRAIN_EPOCHS /
(TRAIN_DATASET_PERCENTAGE / 100))
VALIDATION_FREQUENCY = round(VALIDATION_FREQUENCY /
(TRAIN_DATASET_PERCENTAGE / 100))
training_start = datetime.datetime.now().isoformat()
train_set = TrainDatasetFromFolder(train_dataset_dir,
patch_size=PATCH_SIZE,
upscale_factor=UPSCALE_FACTOR)
len_train_set = len(train_set)
train_set = Subset(
train_set,
list(
np.random.choice(
np.arange(len_train_set),
int(len_train_set * TRAIN_DATASET_PERCENTAGE / 100), False)))
val_set = ValDatasetFromFolder(val_dataset_dir,
upscale_factor=UPSCALE_FACTOR)
len_val_set = len(val_set)
val_set = Subset(
val_set,
list(
np.random.choice(np.arange(len_val_set),
int(len_val_set * VAL_DATASET_PERCENTAGE / 100),
False)))
train_loader = DataLoader(dataset=train_set,
num_workers=8,
batch_size=BATCH_SIZE,
shuffle=True,
pin_memory=True,
prefetch_factor=8)
val_loader = DataLoader(dataset=val_set,
num_workers=2,
batch_size=VAL_BATCH_SIZE,
shuffle=False,
pin_memory=True,
prefetch_factor=2)
epoch_validation_hr_dataset = HrValDatasetFromFolder(
val_dataset_dir) # Useful to compute FID metric
results_folder = Path(
f"results_{training_start}_CS:{PATCH_SIZE}_US:{UPSCALE_FACTOR}x_TRAIN:{TRAIN_DATASET_PERCENTAGE}%_AUGMENTATION:{ENABLE_AUGMENTATION}"
)
results_folder.mkdir(exist_ok=True)
writer = SummaryWriter(str(results_folder / "tensorboard_log"))
g_net = Generator(n_residual_blocks=NUM_RESIDUAL_BLOCKS,
upsample_factor=UPSCALE_FACTOR)
d_net = Discriminator(patch_size=PATCH_SIZE)
lpips_metric = lpips.LPIPS(net='alex')
g_net.to(device=device)
d_net.to(device=device)
lpips_metric.to(device=device)
g_optimizer = optim.Adam(g_net.parameters(), lr=1e-4)
d_optimizer = optim.Adam(d_net.parameters(), lr=1e-4)
bce_loss = BCELoss()
mse_loss = MSELoss()
bce_loss.to(device=device)
mse_loss.to(device=device)
results = {
'd_total_loss': [],
'g_total_loss': [],
'g_adv_loss': [],
'g_content_loss': [],
'd_real_mean': [],
'd_fake_mean': [],
'psnr': [],
'ssim': [],
'lpips': [],
'fid': [],
'rt': [],
'augment_probability': []
}
augment_probability = 0
num_images = len(train_set) * (NUM_PRETRAIN_EPOCHS + NUM_ADV_EPOCHS)
prediction_list = []
rt = 0
for epoch in range(1, NUM_PRETRAIN_EPOCHS + NUM_ADV_EPOCHS + 1):
train_bar = tqdm(train_loader, ncols=200)
running_results = {
'batch_sizes': 0,
'd_epoch_total_loss': 0,
'g_epoch_total_loss': 0,
'g_epoch_adv_loss': 0,
'g_epoch_content_loss': 0,
'd_epoch_real_mean': 0,
'd_epoch_fake_mean': 0,
'rt': 0,
'augment_probability': 0
}
image_percentage = epoch / (NUM_PRETRAIN_EPOCHS + NUM_ADV_EPOCHS) * 100
g_net.train()
d_net.train()
for data, target in train_bar:
augment_batch.p = torch.tensor([augment_probability],
device=device)
batch_size = data.size(0)
running_results["batch_sizes"] += batch_size
target = target.to(device)
data = data.to(device)
real_labels = torch.ones(batch_size, device=device)
fake_labels = torch.zeros(batch_size, device=device)
if epoch > NUM_PRETRAIN_EPOCHS:
# Discriminator training
d_optimizer.zero_grad(set_to_none=True)
d_real_output = d_net(augment_batch(target))
d_real_output_loss = bce_loss(
d_real_output, real_labels * LABEL_SMOOTHING_FACTOR)
fake_img = g_net(data)
d_fake_output = d_net(augment_batch(fake_img))
d_fake_output_loss = bce_loss(d_fake_output, fake_labels)
d_total_loss = d_real_output_loss + d_fake_output_loss
d_total_loss.backward()
d_optimizer.step()
d_real_mean = d_real_output.mean()
d_fake_mean = d_fake_output.mean()
# Generator training
g_optimizer.zero_grad(set_to_none=True)
fake_img = g_net(data)
if epoch > NUM_PRETRAIN_EPOCHS:
adversarial_loss = bce_loss(d_net(augment_batch(fake_img)),
real_labels) * ADV_LOSS_BALANCER
content_loss = mse_loss(fake_img, target)
g_total_loss = content_loss + adversarial_loss
else:
adversarial_loss = mse_loss(torch.zeros(
1, device=device), torch.zeros(
1,
device=device)) # Logging purposes, it is always zero
content_loss = mse_loss(fake_img, target)
g_total_loss = content_loss
g_total_loss.backward()
g_optimizer.step()
if epoch > NUM_PRETRAIN_EPOCHS and ENABLE_AUGMENTATION:
prediction_list.append(
(torch.sign(d_real_output - 0.5)).tolist())
if len(prediction_list) == RT_BATCH_SMOOTHING_FACTOR:
rt_list = [
prediction for sublist in prediction_list
for prediction in sublist
]
rt = mean(rt_list)
if mean(rt_list) > AUGMENT_PROB_TARGET:
augment_probability = min(
0.85,
augment_probability + AUGMENT_PROBABABILITY_STEP)
else:
augment_probability = max(
0.,
augment_probability - AUGMENT_PROBABABILITY_STEP)
prediction_list.clear()
running_results['g_epoch_total_loss'] += g_total_loss.to(
'cpu', non_blocking=True).detach() * batch_size
running_results['g_epoch_adv_loss'] += adversarial_loss.to(
'cpu', non_blocking=True).detach() * batch_size
running_results['g_epoch_content_loss'] += content_loss.to(
'cpu', non_blocking=True).detach() * batch_size
if epoch > NUM_PRETRAIN_EPOCHS:
running_results['d_epoch_total_loss'] += d_total_loss.to(
'cpu', non_blocking=True).detach() * batch_size
running_results['d_epoch_real_mean'] += d_real_mean.to(
'cpu', non_blocking=True).detach() * batch_size
running_results['d_epoch_fake_mean'] += d_fake_mean.to(
'cpu', non_blocking=True).detach() * batch_size
running_results['rt'] += rt * batch_size
running_results[
'augment_probability'] += augment_probability * batch_size
train_bar.set_description(
desc=f'[{epoch}/{NUM_ADV_EPOCHS + NUM_PRETRAIN_EPOCHS}] '
f'Loss_D: {running_results["d_epoch_total_loss"] / running_results["batch_sizes"]:.4f} '
f'Loss_G: {running_results["g_epoch_total_loss"] / running_results["batch_sizes"]:.4f} '
f'Loss_G_adv: {running_results["g_epoch_adv_loss"] / running_results["batch_sizes"]:.4f} '
f'Loss_G_content: {running_results["g_epoch_content_loss"] / running_results["batch_sizes"]:.4f} '
f'D(x): {running_results["d_epoch_real_mean"] / running_results["batch_sizes"]:.4f} '
f'D(G(z)): {running_results["d_epoch_fake_mean"] / running_results["batch_sizes"]:.4f} '
f'rt: {running_results["rt"] / running_results["batch_sizes"]:.4f} '
f'augment_probability: {running_results["augment_probability"] / running_results["batch_sizes"]:.4f}'
)
if epoch == 1 or epoch == (
NUM_PRETRAIN_EPOCHS + NUM_ADV_EPOCHS
) or epoch % VALIDATION_FREQUENCY == 0 or VALIDATION_FREQUENCY == 1:
torch.cuda.empty_cache()
gc.collect()
g_net.eval()
# ...
images_path = results_folder / Path(f'training_images_results')
images_path.mkdir(exist_ok=True)
with torch.no_grad():
val_bar = tqdm(val_loader, ncols=160)
val_results = {
'epoch_mse': 0,
'epoch_ssim': 0,
'epoch_psnr': 0,
'epoch_avg_psnr': 0,
'epoch_avg_ssim': 0,
'epoch_lpips': 0,
'epoch_avg_lpips': 0,
'epoch_fid': 0,
'batch_sizes': 0
}
val_images = torch.empty((0, 0))
epoch_validation_sr_dataset = None
for lr, val_hr_restore, hr in val_bar:
batch_size = lr.size(0)
val_results['batch_sizes'] += batch_size
hr = hr.to(device=device)
lr = lr.to(device=device)
sr = g_net(lr)
sr = torch.clamp(sr, 0., 1.)
if not epoch_validation_sr_dataset:
epoch_validation_sr_dataset = SingleTensorDataset(
(sr.cpu() * 255).to(torch.uint8))
else:
epoch_validation_sr_dataset = ConcatDataset(
(epoch_validation_sr_dataset,
SingleTensorDataset(
(sr.cpu() * 255).to(torch.uint8))))
batch_mse = ((sr - hr)**2).data.mean() # Pixel-wise MSE
val_results['epoch_mse'] += batch_mse * batch_size
batch_ssim = pytorch_ssim.ssim(sr, hr).item()
val_results['epoch_ssim'] += batch_ssim * batch_size
val_results['epoch_avg_ssim'] = val_results[
'epoch_ssim'] / val_results['batch_sizes']
val_results['epoch_psnr'] += 20 * log10(
hr.max() / (batch_mse / batch_size)) * batch_size
val_results['epoch_avg_psnr'] = val_results[
'epoch_psnr'] / val_results['batch_sizes']
val_results['epoch_lpips'] += torch.mean(
lpips_metric(hr * 2 - 1, sr * 2 - 1)).to(
'cpu', non_blocking=True).detach() * batch_size
val_results['epoch_avg_lpips'] = val_results[
'epoch_lpips'] / val_results['batch_sizes']
val_bar.set_description(
desc=
f"[converting LR images to SR images] PSNR: {val_results['epoch_avg_psnr']:4f} dB "
f"SSIM: {val_results['epoch_avg_ssim']:4f} "
f"LPIPS: {val_results['epoch_avg_lpips']:.4f} ")
if val_images.size(0) * val_images.size(
1) < NUM_LOGGED_VALIDATION_IMAGES * 3:
if val_images.size(0) == 0:
val_images = torch.hstack(
(display_transform(CENTER_CROP_SIZE)
(val_hr_restore).unsqueeze(0).transpose(0, 1),
display_transform(CENTER_CROP_SIZE)(
hr.data.cpu()).unsqueeze(0).transpose(
0, 1),
display_transform(CENTER_CROP_SIZE)(
sr.data.cpu()).unsqueeze(0).transpose(
0, 1)))
else:
val_images = torch.cat((
val_images,
torch.hstack(
(display_transform(CENTER_CROP_SIZE)(
val_hr_restore).unsqueeze(0).transpose(
0, 1),
display_transform(CENTER_CROP_SIZE)(
hr.data.cpu()).unsqueeze(0).transpose(
0, 1),
display_transform(CENTER_CROP_SIZE)(
sr.data.cpu()).unsqueeze(0).transpose(
0, 1)))))
val_results['epoch_fid'] = calculate_metrics(
epoch_validation_sr_dataset,
epoch_validation_hr_dataset,
cuda=True,
fid=True,
verbose=True
)['frechet_inception_distance'] # Set batch_size=1 if you get memory error (inside calculate metric function)
val_images = val_images.view(
(NUM_LOGGED_VALIDATION_IMAGES // 4, -1, 3,
CENTER_CROP_SIZE, CENTER_CROP_SIZE))
val_save_bar = tqdm(val_images,
desc='[saving validation results]',
ncols=160)
for index, image_batch in enumerate(val_save_bar, start=1):
image_grid = utils.make_grid(image_batch,
nrow=3,
padding=5)
writer.add_image(
f'progress{image_percentage:.1f}_index_{index}.png',
image_grid)
# save loss / scores / psnr /ssim
results['d_total_loss'].append(running_results['d_epoch_total_loss'] /
running_results['batch_sizes'])
results['g_total_loss'].append(running_results['g_epoch_total_loss'] /
running_results['batch_sizes'])
results['g_adv_loss'].append(running_results['g_epoch_adv_loss'] /
running_results['batch_sizes'])
results['g_content_loss'].append(
running_results['g_epoch_content_loss'] /
running_results['batch_sizes'])
results['d_real_mean'].append(running_results['d_epoch_real_mean'] /
running_results['batch_sizes'])
results['d_fake_mean'].append(running_results['d_epoch_fake_mean'] /
running_results['batch_sizes'])
results['rt'].append(running_results['rt'] /
running_results['batch_sizes'])
results['augment_probability'].append(
running_results['augment_probability'] /
running_results['batch_sizes'])
if epoch == 1 or epoch == (
NUM_PRETRAIN_EPOCHS + NUM_ADV_EPOCHS
) or epoch % VALIDATION_FREQUENCY == 0 or VALIDATION_FREQUENCY == 1:
results['psnr'].append(val_results['epoch_avg_psnr'])
results['ssim'].append(val_results['epoch_avg_ssim'])
results['lpips'].append(val_results['epoch_avg_lpips'])
results['fid'].append(val_results['epoch_fid'])
for metric, metric_values in results.items():
if epoch == 1 or epoch == (
NUM_PRETRAIN_EPOCHS + NUM_ADV_EPOCHS) or epoch % VALIDATION_FREQUENCY == 0 or VALIDATION_FREQUENCY == 1 or \
metric not in ["psnr", "ssim", "lpips", "fid"]:
writer.add_scalar(metric, metric_values[-1],
int(image_percentage * num_images * 0.01))
if epoch == 1 or epoch == (
NUM_PRETRAIN_EPOCHS + NUM_ADV_EPOCHS
) or epoch % VALIDATION_FREQUENCY == 0 or VALIDATION_FREQUENCY == 1:
# save model parameters
models_path = results_folder / "saved_models"
models_path.mkdir(exist_ok=True)
torch.save(
{
'progress': image_percentage,
'g_net': g_net.state_dict(),
'd_net': g_net.state_dict(),
# 'g_optimizer': g_optimizer.state_dict(), Uncomment this if you want resume training
# 'd_optimizer': d_optimizer.state_dict(),
},
str(models_path / f'progress_{image_percentage:.1f}.tar'))
def validation_step(dataloader: DataLoader,
netG: nn.Module,
out_dir: str,
idx_epoch: int,
num_epochs: int,
num_print: int = 5) -> dict:
netG.eval()
num_samples = len(dataloader)
batch_sizes = 0
step_ = int(math.ceil(num_samples) / num_print)
t1 = time.time()
with torch.no_grad():
val_results = {
'mse': 0,
'ssims': 0,
'psnr': 0,
'ssim': 0,
'batch_sizes': 0
}
val_images = []
for idx_val, data_val in enumerate(dataloader):
val_lr, val_hr_restore, val_hr = data_val['lr'], data_val[
'lr_up'], data_val['hr']
batch_size = val_lr.size(0)
batch_sizes += batch_size
val_hr_restore = x_preprocess(val_hr_restore, to_device=None)
lr = x_preprocess(val_lr, to_device=to_device)
hr = x_preprocess(val_hr, to_device=to_device)
sr = netG(lr)
#
batch_mse = float(((sr - hr)**2).mean())
val_results['mse'] += batch_mse * batch_size
batch_ssim = float(pytorch_ssim.ssim(sr, hr)) # .data[0]
val_results['ssims'] += batch_ssim * batch_size
val_results['psnr'] = 10 * log10(
1 / (val_results['mse'] / batch_sizes))
val_results['ssim'] = val_results['ssims'] / batch_sizes
if (idx_val % step_) == 0:
print('\t\t(VAL) [{}/{}] <- MSE/SSIM = {:0.3f}/{:0.3f}'.format(
idx_val, num_samples, batch_mse, batch_ssim))
val_images.extend([
display_transform()(val_hr_restore.squeeze(0)),
display_transform()(hr.data.cpu().squeeze(0)),
display_transform()(sr.data.cpu().squeeze(0))
])
chunk_size = 3 * 5
size_round = chunk_size * int(math.floor(len(val_images) / chunk_size))
val_images = val_images[:size_round]
val_images = torch.stack(val_images)
val_images = torch.chunk(val_images, val_images.size(0) // chunk_size)
for idx_image, image in enumerate(val_images): # val_save_bar:
image = utils.make_grid(image, nrow=3, padding=5)
path_img_out = os.path.join(
out_dir,
'epoch_{:05d}_index_{:02d}.png'.format(idx_epoch, idx_image))
utils.save_image(image, path_img_out, padding=5)
dt = time.time() - t1
val_results = {k: v / batch_sizes for k, v in val_results.items()}
tmp_ = ', '.join(
['{}: {:0.2f}'.format(k, v) for k, v in val_results.items()])
print('\t\t(VALIDATION) ({}/{}) dt ~{:0.2f} (s), {}'.format(
idx_epoch, num_epochs, dt, tmp_))
return val_results
batch_mse = ((sr - hr)**2).data.mean()
valing_results['mse'] += batch_mse * batch_size
batch_ssim = pytorch_ssim.ssim(sr, hr).item()
valing_results['ssims'] += batch_ssim * batch_size
valing_results['psnr'] = 10 * log10(
(hr.max()**2) / (valing_results['mse'] /
valing_results['batch_sizes']))
valing_results['ssim'] = valing_results[
'ssims'] / valing_results['batch_sizes']
val_bar.set_description(
desc=
'[converting LR images to SR images] PSNR: %.4f dB SSIM: %.4f'
% (valing_results['psnr'], valing_results['ssim']))
val_images.extend([
display_transform()(lr.squeeze(0)),
display_transform()(sr.data.cpu().squeeze(0)),
display_transform()(hr.data.cpu().squeeze(0))
])
val_images = torch.stack(val_images)
val_images = torch.chunk(val_images,
val_images.size()[0] // 15)
val_save_bar = tqdm(val_images,
desc='[saving training results]')
index = 1
for image in val_save_bar:
image = utils.make_grid(image, nrow=3, padding=5)
utils.save_image(image,
out_path + 'epoch_%d_index_%d.png' %
(epoch, index),
padding=5)
def main(args):
if (not os.path.exists('data/dataset.pt')):
# Make sure the bit depth is 24, 8 = Gray scale
df = pd.read_pickle('data/dataset_files.gzip')
df = df[(df['width'] > 100) & (df['height'] > 100)]
train_df, val_df = train_test_split(df,
test_size=0.2,
random_state=42,
shuffle=True)
_, val_similar = dataframe_find_similar_images(
val_df, batch_size=args.batch_size)
# Create the train dataset
train_filenames = train_df['filename'].tolist()
train_set = TrainDatasetFromList(train_filenames,
crop_size=args.crop_size,
upscale_factor=args.upscale_factor)
val_sets = list()
for val_df in val_similar:
val_filenames = val_df['filename'].tolist()
val_set = ValDatasetFromList(val_filenames,
upscale_factor=args.upscale_factor)
val_sets.append(val_set)
train_sampler = torch.utils.data.RandomSampler(train_set)
val_sampler = torch.utils.data.SequentialSampler(val_set)
data_to_save = {
'train_dataset': train_set,
"val_datasets": val_sets,
'train_sampler': train_sampler,
'val_sampler': val_sampler
}
torch.save(data_to_save, 'data/dataset.pt')
else:
datasets = torch.load('data/dataset.pt')
train_set = datasets['train_dataset']
val_sets = datasets['val_datasets']
train_sampler = datasets['train_sampler']
val_sampler = datasets['val_sampler']
train_loader = DataLoader(dataset=train_set,
batch_size=args.batch_size,
num_workers=args.num_workers,
sampler=train_sampler)
val_loaders = list()
for val_set in val_sets:
val_loaders.append(
DataLoader(dataset=val_set,
batch_size=args.batch_size,
num_workers=args.num_workers,
shuffle=False))
netG = Generator(args.upscale_factor)
print('# generator parameters:',
sum(param.numel() for param in netG.parameters()))
netD = Discriminator()
print('# discriminator parameters:',
sum(param.numel() for param in netD.parameters()))
generator_criterion = GeneratorLoss()
if torch.cuda.is_available():
netG.cuda()
netD.cuda()
generator_criterion.cuda()
optimizerG = optim.Adam(netG.parameters())
optimizerD = optim.Adam(netD.parameters())
results = {
'd_loss': [],
'g_loss': [],
'd_score': [],
'g_score': [],
'psnr': [],
'ssim': []
}
start_epoch = 1
if args.resume:
import glob
netG_files = glob.glob(
os.path.join(args.output_dir,
'netG_epoch_%d_*.pth' % (args.upscale_factor)))
netD_files = glob.glob(
os.path.join(args.output_dir,
'netD_epoch_%d_*.pth' % (args.upscale_factor)))
if (len(netG_files) > 0):
netG_file = max(netG_files, key=os.path.getctime)
netD_file = max(netD_files, key=os.path.getctime)
netG.load_state_dict(torch.load(netG_file))
netD.load_state_dict(torch.load(netD_file))
start_epoch = len(netG_files)
for epoch in range(start_epoch, args.epochs + 1):
train_bar = tqdm(train_loader)
running_results = {
'batch_sizes': 0,
'd_loss': 0,
'g_loss': 0,
'd_score': 0,
'g_score': 0
}
dscaler = torch.cuda.amp.GradScaler(
) # Creates once at the beginning of training #* Discriminator
gscaler = torch.cuda.amp.GradScaler() #* Generator
netG.train()
netD.train()
for data, target in train_bar:
with torch.cuda.amp.autocast(): # Mix precision
batch_size = data.size(0)
running_results['batch_sizes'] += batch_size
############################
# (1) Update D network: maximize D(x)-1-D(G(z))
###########################
netD.zero_grad()
real_img = Variable(target, requires_grad=False)
if torch.cuda.is_available():
real_img = real_img.cuda()
z = Variable(data)
if torch.cuda.is_available():
z = z.cuda()
fake_img = netG(z)
real_out = netD(real_img).mean(
) # Discriminator Takes in the real image and predicts whether it's real
fake_out = netD(fake_img).mean(
) # Discriminator takes in the fake image and predicts if it's fake
d_loss = 1 - real_out + fake_out # Minimizing the loss would mean real_out=1 and fake out = 0. so it knows the real image it knows the fake image
# d_loss.backward(retain_graph=True)
# optimizerD.step()
############################
# (2) Update G network: minimize 1-D(G(z)) + Perception Loss + Image Loss + TV Loss
###########################
netG.zero_grad()
g_loss = generator_criterion(fake_out.detach(), fake_img,
real_img.detach())
dscaler.scale(d_loss).backward(retain_graph=True)
gscaler.scale(g_loss).backward()
dscaler.step(optimizerD)
dscaler.update()
gscaler.step(optimizerG)
gscaler.update()
fake_img = netG(z)
fake_out = netD(fake_img).mean()
# loss for current batch before optimization
running_results['g_loss'] += g_loss.item() * batch_size
running_results['d_loss'] += d_loss.item() * batch_size
running_results['d_score'] += real_out.item() * batch_size
running_results['g_score'] += fake_out.item() * batch_size
train_bar.set_description(
desc=
'[%d/%d] Loss_D: %.4f Loss_G: %.4f D(x): %.4f D(G(z)): %.4f' %
(epoch, args.epochs,
running_results['d_loss'] / running_results['batch_sizes'],
running_results['g_loss'] / running_results['batch_sizes'],
running_results['d_score'] / running_results['batch_sizes'],
running_results['g_score'] / running_results['batch_sizes']))
# save model parameters
torch.save(
netG.state_dict(),
os.path.join(args.output_dir, 'netG_epoch_%d_%d.pth' %
(args.upscale_factor, epoch)))
torch.save(
netD.state_dict(),
os.path.join(args.output_dir, 'netD_epoch_%d_%d.pth' %
(args.upscale_factor, epoch)))
if epoch % args.validation_epoch == 0 and epoch != 0:
netG.eval()
with torch.no_grad():
val_results = {
'mse': 0,
'ssims': 0,
'psnr': 0,
'ssim': 0,
'batch_sizes': 0
}
val_images = []
for i in trange(len(val_loaders), desc='Running validation'):
val_loader = val_loaders[i]
for val_lr, val_hr_restore, val_hr in val_loader:
batch_size = val_lr.size(0)
val_results['batch_sizes'] += batch_size
lr = val_lr
hr = val_hr
if torch.cuda.is_available():
lr = lr.cuda()
hr = hr.cuda()
sr = netG(lr)
batch_mse = ((sr - hr)**2).data.mean()
val_results['mse'] += batch_mse * batch_size
batch_ssim = pytorch_ssim.ssim(sr, hr).item()
val_results['ssims'] += batch_ssim * batch_size
val_results['psnr'] = 10 * log10(
(hr.max()**2) /
(val_results['mse'] / val_results['batch_sizes']))
val_results['ssim'] = val_results[
'ssims'] / val_results['batch_sizes']
# val_bar.set_description(
# desc='[converting LR images to SR images] PSNR: %.4f dB SSIM: %.4f' % (
# val_results['psnr'], val_results['ssim']))
# convert the validation images
val_hr_restore_squeeze = val_hr_restore.squeeze(0)
hr_squeeze = hr.data.cpu().squeeze(0)
sr_squeeze = sr.data.cpu().squeeze(0)
for b in range(batch_size):
val_hr = val_hr_restore_squeeze[b]
hr_temp = hr_squeeze[b]
sr_temp = sr_squeeze[b]
val_images.extend([
display_transform()(val_hr),
display_transform()(hr_temp),
display_transform()(sr_temp)
])
val_images = torch.stack(val_images)
val_images = torch.chunk(val_images, val_images.size(0) // 15)
val_save_bar = tqdm(val_images,
desc='[saving training results]')
index = 1
for image in val_save_bar:
image = utils.make_grid(image, nrow=3, padding=5)
utils.save_image(
image,
os.path.join(
args.output_dir,
'epoch_%d_upscale_%d_index_%d.png' %
(epoch, args.upscale_factor, index)))
index += 1
# save loss\scores\psnr\ssim
results['d_loss'].append(running_results['d_loss'] /
running_results['batch_sizes'])
results['g_loss'].append(running_results['g_loss'] /
running_results['batch_sizes'])
results['d_score'].append(running_results['d_score'] /
running_results['batch_sizes'])
results['g_score'].append(running_results['g_score'] /
running_results['batch_sizes'])
results['psnr'].append(val_results['psnr'])
results['ssim'].append(val_results['ssim'])
if epoch % 10 == 0 and epoch != 0:
out_path = 'statistics/'
data_frame = pd.DataFrame(data={
'Loss_D': results['d_loss'],
'Loss_G': results['g_loss'],
'Score_D': results['d_score'],
'Score_G': results['g_score'],
'PSNR': results['psnr'],
'SSIM': results['ssim']
},
index=range(1, epoch + 1))
data_frame.to_csv(out_path + 'srf_' + str(args.upscale_factor) +
'_train_results.csv',
index_label='Epoch')
for image_name, lr_image, hr_restore_img, hr_image in test_bar:
with torch.no_grad():
image_name = image_name[0]
lr_image = Variable(lr_image, volatile=True)
hr_image = Variable(hr_image, volatile=True)
if torch.cuda.is_available():
lr_image = lr_image.cuda()
hr_image = hr_image.cuda()
sr_image = model(lr_image)
mse = ((hr_image - sr_image)**2).data.mean()
psnr = 10 * log10(1 / mse)
ssim = float(pytorch_ssim.ssim(sr_image, hr_image))
test_out = display_transform()(hr_restore_img.squeeze(0))
test_gt = display_transform()(hr_image.data.cpu().squeeze(0))
#test_images = torch.stack(
# [display_transform()(hr_restore_img.squeeze(0)), display_transform()(hr_image.data.cpu().squeeze(0)),
# display_transform()(sr_image.data.cpu().squeeze(0))])
# image = utils.make_grid(test_images, nrow=3, padding=5)
# utils.save_image(test_out, out_path + image_name.split('.')[0] + '_psnr_%.4f_ssim_%.4f.' % (psnr, ssim) +
# image_name.split('.')[-1], padding=5)
utils.save_image(test_gt, out_path_2 + image_name)
utils.save_image(test_out, out_path + image_name, padding=5)
# utils.save_image(test_out,)
# save psnr\ssim
results[image_name.split('_')[0]]['psnr'].append(psnr)
results[image_name.split('_')[0]]['ssim'].append(ssim)
def main_train(path_trn: str, path_val: str,
crop_size: int, upscale_factor: int, num_epochs: int,
num_workers: int, to_device: str = 'cuda:0', batch_size: int = 64):
to_device = get_device(to_device)
train_set = TrainDatasetFromFolder(path_trn, crop_size=crop_size, upscale_factor=upscale_factor)
val_set = ValDatasetFromFolder(path_val, upscale_factor=upscale_factor)
# train_set = TrainDatasetFromFolder('data/VOC2012/train', crop_size=crop_size, upscale_factor=upscale_factor)
# val_set = ValDatasetFromFolder('data/VOC2012/val', upscale_factor=upscale_factor)
#
train_loader = DataLoader(dataset=train_set, num_workers=num_workers, batch_size=batch_size, shuffle=True)
val_loader = DataLoader(dataset=val_set, num_workers=num_workers, batch_size=1, shuffle=False)
netG = Generator(upscale_factor)
print('# generator parameters:', sum(param.numel() for param in netG.parameters()))
netD = Discriminator()
print('# discriminator parameters:', sum(param.numel() for param in netD.parameters()))
generator_criterion = GeneratorLoss()
if torch.cuda.is_available():
netG.cuda()
netD.cuda()
generator_criterion.cuda()
optimizerG = optim.Adam(netG.parameters())
optimizerD = optim.Adam(netD.parameters())
results = {'d_loss': [], 'g_loss': [], 'd_score': [], 'g_score': [], 'psnr': [], 'ssim': []}
for epoch in range(1, num_epochs + 1):
train_bar = tqdm(train_loader)
running_results = {'batch_sizes': 0, 'd_loss': 0, 'g_loss': 0, 'd_score': 0, 'g_score': 0}
netG.train()
netD.train()
# FIXME: seperate function for epoch training
for data, target in train_bar:
g_update_first = True
batch_size = data.size(0)
#
# img_hr = target.numpy().transpose((0, 2, 3, 1))[0]
# img_lr = data.numpy().transpose((0, 2, 3, 1))[0]
# img_lr_x4 = cv2.resize(img_lr, img_hr.shape[:2], interpolation=cv2.INTER_CUBIC)
# #
# plt.subplot(1, 3, 1)
# plt.imshow(img_hr)
# plt.subplot(1, 3, 2)
# plt.imshow(img_lr)
# plt.subplot(1, 3, 3)
# plt.imshow(img_lr_x4)
# plt.show()
running_results['batch_sizes'] += batch_size
############################
# (1) Update D network: maximize D(x)-1-D(G(z))
###########################
# real_img = Variable(target)
# if torch.cuda.is_available():
# real_img = real_img.cuda()
# z = Variable(data)
# if torch.cuda.is_available():
# z = z.cuda()
z = data.to(to_device)
real_img = target.to(to_device)
fake_img = netG(z)
netD.zero_grad()
real_out = netD(real_img).mean()
fake_out = netD(fake_img).mean()
d_loss = 1 - real_out + fake_out
d_loss.backward(retain_graph=True)
optimizerD.step()
############################
# (2) Update G network: minimize 1-D(G(z)) + Perception Loss + Image Loss + TV Loss
###########################
netG.zero_grad()
g_loss = generator_criterion(fake_out, fake_img, real_img)
g_loss.backward()
optimizerG.step()
fake_img = netG(z)
fake_out = netD(fake_img).mean()
g_loss = generator_criterion(fake_out, fake_img, real_img)
running_results['g_loss'] += float(g_loss) * batch_size
d_loss = 1 - real_out + fake_out
running_results['d_loss'] += float(d_loss) * batch_size
running_results['d_score'] += float(real_out) * batch_size
running_results['g_score'] += float(fake_out) * batch_size
train_bar.set_description(desc='[%d/%d] Loss_D: %.4f Loss_G: %.4f D(x): %.4f D(G(z)): %.4f' % (
epoch, num_epochs, running_results['d_loss'] / running_results['batch_sizes'],
running_results['g_loss'] / running_results['batch_sizes'],
running_results['d_score'] / running_results['batch_sizes'],
running_results['g_score'] / running_results['batch_sizes']))
netG.eval()
#FIXME: seperate function for epoch validation
with torch.no_grad():
out_path = 'training_results/SRF_' + str(upscale_factor) + '/'
if not os.path.exists(out_path):
os.makedirs(out_path)
val_bar = tqdm(val_loader)
valing_results = {'mse': 0, 'ssims': 0, 'psnr': 0, 'ssim': 0, 'batch_sizes': 0}
val_images = []
for val_lr, val_hr_restore, val_hr in val_bar:
batch_size = val_lr.size(0)
valing_results['batch_sizes'] += batch_size
# lr = Variable(val_lr, volatile=True)
# hr = Variable(val_hr, volatile=True)
# if torch.cuda.is_available():
# lr = lr.cuda()
# hr = hr.cuda()
lr = val_lr.to(to_device)
hr = val_hr.to(to_device)
sr = netG(lr)
batch_mse = ((sr - hr) ** 2).mean()
valing_results['mse'] += float(batch_mse) * batch_size
batch_ssim = float(pytorch_ssim.ssim(sr, hr)) #.data[0]
valing_results['ssims'] += batch_ssim * batch_size
valing_results['psnr'] = 10 * log10(1 / (valing_results['mse'] / valing_results['batch_sizes']))
valing_results['ssim'] = valing_results['ssims'] / valing_results['batch_sizes']
val_bar.set_description(
desc='[converting LR images to SR images] PSNR: %.4f dB SSIM: %.4f' % (
valing_results['psnr'], valing_results['ssim']))
val_images.extend(
[display_transform()(val_hr_restore.squeeze(0)), display_transform()(hr.data.cpu().squeeze(0)),
display_transform()(sr.data.cpu().squeeze(0))])
val_images = torch.stack(val_images)
val_images = torch.chunk(val_images, val_images.size(0) // 15)
val_save_bar = tqdm(val_images, desc='[saving training results]')
index = 1
for image in val_save_bar:
image = utils.make_grid(image, nrow=3, padding=5)
utils.save_image(image, out_path + 'epoch_%d_index_%d.png' % (epoch, index), padding=5)
index += 1
# save model parameters
torch.save(netG.state_dict(), 'epochs/netG_epoch_%d_%d.pth' % (upscale_factor, epoch))
torch.save(netD.state_dict(), 'epochs/netD_epoch_%d_%d.pth' % (upscale_factor, epoch))
# save loss\scores\psnr\ssim
results['d_loss'].append(running_results['d_loss'] / running_results['batch_sizes'])
results['g_loss'].append(running_results['g_loss'] / running_results['batch_sizes'])
results['d_score'].append(running_results['d_score'] / running_results['batch_sizes'])
results['g_score'].append(running_results['g_score'] / running_results['batch_sizes'])
results['psnr'].append(valing_results['psnr'])
results['ssim'].append(valing_results['ssim'])
if epoch % 10 == 0 and epoch != 0:
out_path = 'statistics/'
data_frame = pd.DataFrame(
data={'Loss_D': results['d_loss'], 'Loss_G': results['g_loss'], 'Score_D': results['d_score'],
'Score_G': results['g_score'], 'PSNR': results['psnr'], 'SSIM': results['ssim']},
index=range(1, epoch + 1))
data_frame.to_csv(out_path + 'srf_' + str(upscale_factor) + '_train_results.csv', index_label='Epoch')
def test_benchmark(upscale_factor, epoch_num):
model_name = 'netG_epoch_{}_100.pth'.format(upscale_factor)
results = {
'Set5': {
'psnr': [],
'ssim': []
},
'Set14': {
'psnr': [],
'ssim': []
},
'BSD100': {
'psnr': [],
'ssim': []
},
'Urban100': {
'psnr': [],
'ssim': []
},
'SunHays80': {
'psnr': [],
'ssim': []
}
}
model = Generator(upscale_factor).eval()
if torch.cuda.is_available():
model = model.cuda()
model.load_state_dict(
torch.load('epochs/' + model_name, map_location=torch.device('cpu')))
test_set = TestDatasetFromFolder('data/test',
upscale_factor=upscale_factor)
test_loader = DataLoader(dataset=test_set,
num_workers=4,
batch_size=1,
shuffle=False)
test_bar = tqdm(test_loader, desc='[testing benchmark datasets]')
out_path = 'benchmark_results/SRF_' + str(upscale_factor) + '/'
if not os.path.exists(out_path):
os.makedirs(out_path)
for image_name, lr_image, hr_restore_img, hr_image in test_bar:
image_name = image_name[0]
# volatile is no longer available
# lr_image = Variable(lr_image, volatile=True)
# hr_image = Variable(hr_image, volatile=True)
# image = Variable(ToTensor()(lr_image), volatile=True).unsqueeze(0)
with torch.no_grad():
lr_image = Variable(lr_image)
hr_image = Variable(hr_image)
if torch.cuda.is_available():
lr_image = lr_image.cuda()
hr_image = hr_image.cuda()
sr_image = model(lr_image)
mse = ((hr_image - sr_image)**2).data.mean()
psnr = 10 * log10(1 / mse)
ssim = pytorch_ssim.ssim(sr_image, hr_image).data.item()
test_images = torch.stack([
display_transform()(hr_restore_img.squeeze(0)),
display_transform()(hr_image.data.cpu().squeeze(0)),
display_transform()(sr_image.data.cpu().squeeze(0))
])
image = utils.make_grid(test_images, nrow=3, padding=5)
utils.save_image(image,
out_path + image_name.split('.')[0] +
'_psnr_%.4f_ssim_%.4f.' % (psnr, ssim) +
image_name.split('.')[-1],
padding=5)
# save psnr\ssim
results[image_name.split('_')[0]]['psnr'].append(psnr)
results[image_name.split('_')[0]]['ssim'].append(ssim)
out_path = 'statistics/'
saved_results = {'psnr': [], 'ssim': []}
for item in results.values():
psnr = np.array(item['psnr'])
ssim = np.array(item['ssim'])
if (len(psnr) == 0) or (len(ssim) == 0):
psnr = 'N/A'
ssim = 'N/A'
else:
psnr = psnr.mean()
ssim = ssim.mean()
saved_results['psnr'].append(psnr)
saved_results['ssim'].append(ssim)
data_frame = pd.DataFrame(saved_results, results.keys())
data_frame.to_csv(out_path + 'srf_' + str(upscale_factor) +
'_test_results.csv',
index_label='DataSet')
return data_frame
