def __init__(self,
args,
from_pretrained="",
generator_path=None,
discriminator_path=None,
generator_optimizer_path=None,
discriminator_optimizer_path=None):
super().__init__()
self.args = args
self.relu = nn.ReLU()
self.l1_loss = nn.L1Loss()
self.ssim = SSIM()
self.generator = Generator()
self.discriminator = Discriminator()
self.discriminator_optimizer = torch.optim.Adam(
self.discriminator.parameters(),
lr=args.lr,
betas=(args.b1, args.b2))
self.generator_optimizer = torch.optim.Adam(
self.generator.parameters(), lr=args.lr, betas=(args.b1, args.b2))
if from_pretrained:
if generator_path is None or discriminator_path is None or \
generator_optimizer_path is None or discriminator_optimizer_path is None:
raise ValueError("To train from pretrain provide all paths.")
checkpoint = torch.load(from_pretrained)
self.generator.load_state_dict(checkpoint['generator'])
self.discriminator.load_state_dict(checkpoint['discriminator'])
self.generator_optimizer.load_state_dict(
checkpoint['generator_optimizer'])
self.discriminator_optimizer.load_state_dict(
checkpoint['discriminator_optimizer'])
gt_path = input_path.replace("test", "ground_truth").replace(
".png", "_mask.png")
# input = cv2.imread(input_path)
gt = cv2.imread(gt_path, cv2.IMREAD_GRAYSCALE)
# cv2.namedWindow('image')
# cv2.imshow('image',input)
filename = os.path.basename(os.path.normpath(input_path))
filename, ext = filename.split(".")
input = Image.open(input_path)
input = torchvision.transforms.ToTensor()(input).unsqueeze(
0).to(device)
# input_torch = torch.from_numpy(input).permute(2,0,1).unsqueeze(0)
output = model(input)
diff_avg = 1 - SSIM(input, output)[1]
diff_avg = torch.mean(diff_avg, dim=1, keepdim=True)
diff_avg = channelwised_normalize(diff_avg).detach().cpu()
# enhanced_avg = enhanceMorph(diff_avg.numpy())
# folder = f"{eval_folder}/{defect}"
# cv2.imwrite(f"{folder}/{filename}_gt.{ext}", gt)
# os.makedirs(folder, exist_ok=True)
# torchvision.utils.save_image(diff_avg, f"{folder}/{filename}.{ext}")
# torchvision.utils.save_image(enhanced_avg, f"{folder}/{filename}_enh.{ext}")
gt = (gt / 255.0).reshape(-1)
diff_avg = diff_avg.numpy().squeeze(0).squeeze(0).reshape(-1)
auc = roc_auc_score(gt, diff_avg)
print(f"file {input_path}, {defect}: {auc}")
log.write(f"file {input_path}, {defect}: {auc}\n")
def test_on_mixed_samples(model,
test_loader,
loss_op,
writer,
results_folder,
n_saved_results=5,
epoch=0):
"""
Perform evaluation on the test set
Returns: average MSE error on the whole test set
"""
print("Testing on mixed set...")
model.eval()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
test_epoch_loss = 0
test_images = None
test_images2 = None
if len(test_loader) > n_saved_results:
chosen_sample_i = torch.multinomial(torch.Tensor(
range(len(test_loader))),
num_samples=n_saved_results,
replacement=False)
else:
chosen_sample_i = range(len(test_loader))
n_output_channels = 3
with torch.no_grad():
for index, (img, gt) in enumerate(tqdm(test_loader)):
# img, _ = data
img = img.to(device)
n_output_channels = img.shape[1]
gt = gt.to(device)
output = model(img)
# diff = torch.abs(img - output)
# Grayscaled diff image (average of 3 channels)
diff_avg = 1 - SSIM(img, output)[1]
diff_avg = torch.mean(diff_avg, dim=1, keepdim=True)
diff_avg = channelwised_normalize(diff_avg)
# diff = channelwised_normalize(diff)
th_diff, gth_diff, otsu_diff = binarize(diff_avg,
n_output_channels)
# Make the grayscale image 3-channeled
# diff_avg = diff_avg
loss = 1 - loss_op(diff_avg, gt)
test_epoch_loss += loss.item()
# Save the results if requested
if index in chosen_sample_i:
io_pair = torch.cat((img, output), dim=3)
gt_pair = torch.cat((gt, diff_avg), dim=3)
gt_pair = gt_pair.squeeze(0)
gt_pair = transforms.ToPILImage()(gt_pair.cpu())
draw = ImageDraw.Draw(gt_pair)
font = ImageFont.truetype(font="BebasNeue-Regular.ttf",
size=150)
# font = ImageFont.truetype("sans-serif.ttf", 16)
draw.text((0, 0), f"{loss.item():.3f}", (0), font=font)
draw.text((0, 25), f"{loss.item():.3f}", (255), font=font)
gt_pair = transforms.ToTensor()(gt_pair).unsqueeze(0).expand(
-1, n_output_channels, -1, -1).to(device)
image = torch.cat((io_pair.to(device), gt_pair.to(device),
th_diff.to(device), gth_diff.to(device),
otsu_diff.to(device)), 0)
if test_images is None:
test_images = image
else:
test_images = torch.cat((test_images, image), dim=0)
#####DIRTY ADDITION DIFF MAP RESULT#######
# diff = torch.abs(img - output)
# # Grayscaled diff image (average of 3 channels)
# diff_avg = torch.mean(diff, dim=1, keepdim=True)
# diff_avg = channelwised_normalize(diff_avg)
# # diff = channelwised_normalize(diff)
# th_diff, gth_diff, otsu_diff = binarize(diff_avg, n_output_channels)
# # Make the grayscale image 3-channeled
# # diff_avg = diff_avg
# loss = nn.MSELoss()(diff_avg, gt)
# # Save the results if requested
# if index in chosen_sample_i:
# gt_pair = torch.cat((gt, diff_avg), dim=3)
# gt_pair = gt_pair.squeeze(0)
# gt_pair = transforms.ToPILImage()(gt_pair.cpu())
# draw = ImageDraw.Draw(gt_pair)
# font = ImageFont.truetype(font="BebasNeue-Regular.ttf", size=150)
# # font = ImageFont.truetype("sans-serif.ttf", 16)
# draw.text((0,0),f"{loss.item():.3f}", (0), font=font)
# draw.text((0,25),f"{loss.item():.3f}",(255), font=font)
# gt_pair = transforms.ToTensor()(gt_pair).unsqueeze(0).expand(-1, n_output_channels, -1, -1).to(device)
# image = torch.cat((io_pair.to(device), gt_pair.to(device), th_diff.to(device), gth_diff.to(device), otsu_diff.to(device)), 0)
# if test_images2 is None:
# test_images2 = image
# else:
# test_images2 = torch.cat((test_images2, image), dim=0)
test_epoch_loss = test_epoch_loss / len(test_loader)
test_images = torchvision.utils.make_grid(test_images, nrow=5)
test_images = test_images.unsqueeze(0)
test_images = F.interpolate(test_images, scale_factor=0.1)
result_image = os.path.join(results_folder, f"val_{epoch}.png")
torchvision.utils.save_image(test_images, result_image)
print(f"Test images saved at {results_folder}")
# test_images2 = torchvision.utils.make_grid(test_images2, nrow=5)
# test_images2 = test_images2.unsqueeze(0)
# test_images2 = F.interpolate(test_images2, scale_factor=0.1)
# result_image = os.path.join(results_folder, f"val_{epoch}_more.png")
# torchvision.utils.save_image(test_images2, result_image)
# print(f"Additional diff map images saved at {results_folder}")
# write to tensorboard
if writer:
test_images = test_images.squeeze(0)
writer.add_image('Test images', test_images, global_step=epoch)
writer.add_scalar(f"{loss_op.__class__.__name__}/Test",
test_epoch_loss,
global_step=epoch)
return test_epoch_loss
def train():
if not os.path.exists(Config['checkpoint_path']):
os.makedirs(Config['checkpoint_path'])
os.makedirs(os.path.join(Config['checkpoint_path'], 'generators'))
os.makedirs(os.path.join(Config['checkpoint_path'], 'discriminators'))
device = torch.device('cuda:0' if torch.cuda.is_available()
and Config['use_cuda'] else 'cpu')
# print(Config['img_size']*Config['scale'])
transform = torchvision.transforms.Compose([
torchvision.transforms.RandomCrop(
(Config['img_size'][0] * Config['scale'],
Config['img_size'][1] * Config['scale'])),
torchvision.transforms.ToTensor()
])
normalize = torchvision.transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
#print(Config['img_size'])
scale = torchvision.transforms.Compose([
torchvision.transforms.ToPILImage(),
torchvision.transforms.Resize(Config['img_size']),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
dataset = torchvision.datasets.ImageFolder(root=Config['train_set_path'],
transform=transform)
data_loader = torch.utils.data.DataLoader(
dataset,
batch_size=Config['batch_size'],
shuffle=True,
num_workers=Config['num_workers'])
generator = Generator(Config)
if Config['generator_checkpoint']:
generator.load_state_dict(torch.load(Config['generator_checkpoint']))
discriminator = Discriminator(Config)
if Config['discriminator_checkpoint']:
discriminator.load_state_dict(
torch.load(Config['discriminator_checkpoint']))
feat_extractor = get_feat_extractor()
content_loss = nn.MSELoss()
adversarial_loss = nn.BCELoss()
ones_const = torch.autograd.Variable(torch.ones(Config['batch_size'], 1))
generator.to(device)
discriminator.to(device)
feat_extractor.to('cpu')
ones_const.to(device)
opt_generator = torch.optim.Adam(generator.parameters(),
lr=Config['generator_lr'])
opt_discriminator = torch.optim.Adam(discriminator.parameters(),
lr=Config['discriminator_lr'])
if Config['tensorboard_log']:
writer_pretrain = SummaryWriter(
os.path.join(Config['checkpoint_path'], 'pretrain'))
writer = SummaryWriter(Config['checkpoint_path'])
low_res = torch.FloatTensor(Config['batch_size'], 3, Config['img_size'][0],
Config['img_size'][1])
for epoch in range(int(Config['batch_size'] * 0.3)):
mean_generator_content_loss = 0.0
for i, data in enumerate(data_loader):
high_res_real, _ = data
if high_res_real.shape[0] != Config['batch_size']: continue
for j in range(Config['batch_size']):
low_res[j] = scale(high_res_real[j])
high_res_real[j] = normalize(high_res_real[j])
high_res_real = torch.autograd.Variable(high_res_real.to(device))
high_res_fake = generator(
torch.autograd.Variable(low_res).to(device))
generator.zero_grad()
#print(high_res_fake.shape, high_res_real.shape)
generator_content_loss = content_loss(high_res_fake, high_res_real)
#print(generator_content_loss)
mean_generator_content_loss += generator_content_loss.data
generator_content_loss.backward()
opt_generator.step()
print(
f'Epoch {epoch} Iter {i/len(data_loader)}: MSE Loss {generator_content_loss.data}'
)
writer.add_figure('Pretrain SR Generator',
visualize(low_res,
high_res_real.cpu().data,
high_res_fake.cpu().data),
global_step=epoch)
writer.add_scalar('generator_mse_loss',
mean_generator_content_loss / len(data_loader),
epoch)
torch.save(
generator.state_dict(),
os.path.join(Config['checkpoint_path'],
'generators/generator_pretrain.pth'))
opt_generator = torch.optim.Adam(generator.parameters(),
lr=Config['generator_lr'] * 0.1)
opt_discriminator = torch.optim.Adam(discriminator.parameters(),
lr=Config['discriminator_lr'] * 0.1)
for epoch in range(Config['epochs']):
mean_generator_content_loss = 0.0
mean_generator_adversarial_loss = 0.0
mean_generator_total_loss = 0.0
mean_discriminator_loss = 0.0
for i, data in enumerate(data_loader):
high_res_real, _ = data
psnrs = []
ssims = []
for j in range(Config['batch_size']):
low_res[j] = scale(high_res_real[j])
high_res_real[j] = normalize(high_res_real[j])
low_res.to(device)
high_res_real = torch.autograd.Variable(high_res_real.to(device))
target_real = torch.autograd.Variable(
torch.rand(Config['batch_size'], 1) * 0.5 + 0.7).to(device)
target_fake = torch.autograd.Variable(
torch.rand(Config['batch_size'], 1) * 0.3).to(device)
high_res_fake = generator(
torch.autograd.Variable(low_res).to(device))
discriminator.zero_grad()
discriminator_loss = adversarial_loss(
discriminator(high_res_real), target_real) + adversarial_loss(
discriminator(high_res_fake), target_fake)
mean_discriminator_loss += discriminator_loss.data
discriminator_loss.backward(retain_graph=True)
opt_discriminator.step()
generator.zero_grad()
real_features = torch.autograd.Variable(
feat_extractor(high_res_real.to('cpu')).to(device).data)
fake_features = feat_extractor(high_res_fake.to('cpu')).to(device)
generator_content_loss = content_loss(
high_res_fake, high_res_real) + 0.006 * content_loss(
fake_features, real_features)
mean_generator_content_loss += generator_content_loss.data
generator_adversarial_loss = adversarial_loss(
discriminator(high_res_fake), ones_const.to(device))
mean_generator_adversarial_loss += generator_adversarial_loss.data
generator_total_loss = generator_content_loss + 1e-3 * generator_adversarial_loss
mean_generator_total_loss += generator_total_loss.data
generator_total_loss.backward()
opt_generator.step()
for j in range(Config['batch_size']):
psnrs.append(PSNR()(high_res_real[j], high_res_fake[j]))
ssims.append(SSIM()(high_res_real[j], high_res_fake[j]))
print(
f'Epoch {epoch} Iter {i/len(data_loader)}: Discriminator Loss {discriminator_loss.data}, Generator Loss: {generator_content_loss.data}/{generator_adversarial_loss.data}/{generator_total_loss.data}'
)
writer.add_figure('Training Super Resolution',
visualize(low_res,
high_res_real.cpu().data,
high_res_fake.cpu().data),
global_step=epoch)
psnr = np.array(psnrs).mean()
ssim = np.array(ssims).mean()
writer.add_scalar('discriminator_loss',
mean_discriminator_loss / len(data_loader),
global_step=epoch)
writer.add_scalar('generator_content_loss',
mean_generator_content_loss / len(data_loader),
global_step=epoch)
writer.add_scalar('generator_adversarial_loss',
mean_generator_adversarial_loss / len(data_loader),
global_step=epoch)
writer.add_scalar('generator_total_loss',
mean_generator_total_loss / len(data_loader),
global_step=epoch)
writer.add_scalar('PSNR', psnr, global_step=epoch)
writer.add_scalar('SSIM', ssim, global_step=epoch)
torch.save(
generator.state_dict(),
os.path.join(Config['checkpoint_path'],
'generators/generator_final.pth'))
torch.save(
discriminator.state_dict(),
os.path.join(Config['checkpoint_path'],
'discriminators/discriminator_final.pth'))
# predict by pretrained model
result = model.predict(input_data, batch_size=1, verbose=1)
result = np.squeeze(result)
# inverse Wavelet transform
rA, rH, rV, rD = result[:, :, 0], result[:, :, 1], result[:, :,
2], result[:, :,
3]
x_sr = pywt.idwt2((rA, (rH, rV, rD)), 'haar')
# compute metrics, remove border first
psnr_val = PSNR(
x[scale:-scale, scale:-scale],
x_sr[scale:-scale, scale:-scale] + x_bic[scale:-scale, scale:-scale])
ssim_val = SSIM(
x[scale:-scale, scale:-scale],
x_sr[scale:-scale, scale:-scale] + x_bic[scale:-scale, scale:-scale])
psnr_list.append(psnr_val)
ssim_list.append(ssim_val)
print('%s\tPSNR: %.4f\tSSIM: %.4f' % (image_name, psnr_val, ssim_val))
# display images
if option.display:
plt.figure()
plt.subplot(221)
plt.imshow(x_bic)
plt.title('bicubic'), plt.axis('off')
plt.subplot(222)
plt.imshow(x_sr)
plt.title('SR'), plt.axis('off')
plt.subplot(223)
def __init__(self, CONFIG):
if CONFIG.mode==1 :
gen_model = get_model('G', CONFIG.gen_model)
dis_model = get_model('D', CONFIG.dis_model)
VGG = tl.models.vgg19(pretrained=True, end_with='pool4', mode='static')
lr_init = 1e-4
lr_v = tf1.Variable(lr_init)
beta1 = 0.9
n_epoch_init = CONFIG.init_epoch # n_epoch_init =20 #
n_epoch = CONFIG.total_epoch # n_epoch = 100 #
batch_size = CONFIG.batch_size # batch_size = 8 #
decay_every = int(n_epoch/ 2)
lr_decay = 0.1
resume_epoch = 0
if CONFIG.load_weights:
resume_epoch = CONFIG.model_epoch
if CONFIG.gan_init:
gen_model.load_weights('Checkpoints/GAN_INIT_{}_EPID_{}.h5'.format(CONFIG.gen_model, CONFIG.model_epoch))
resume_epoch = 0
else:
gen_model.load_weights('Checkpoints/GAN_{}_EPID_{}.h5'.format(CONFIG.gen_model, CONFIG.model_epoch))
dis_model.load_weights('Checkpoints/DIS_{}_GAN_{}_EPID_{}.h5'.format(CONFIG.dis_model, CONFIG.gen_model, CONFIG.model_epoch))
g_optimizer_init = tf2.optimizers.Adam(lr_v, beta_1=0.9, beta_2=0.999, epsilon=1e-07)
g_optimizer = tf2.optimizers.Adam(lr_v, beta_1=0.9, beta_2=0.999, epsilon=1e-07)
d_optimizer = tf2.optimizers.Adam(lr_v, beta_1=0.9, beta_2=0.999, epsilon=1e-07)
gen_model.train()
dis_model.train()
VGG.train()
train_ds = get_train_dataset(CONFIG)
if not CONFIG.load_weights or CONFIG.gan_init:
print('## initial learning (G)')
for epoch in range(n_epoch_init):
for step, (lr_patchs, hr_patchs) in enumerate(train_ds):
if lr_patchs.shape[0] != batch_size:
break
step_time = time.time()
with tf1.GradientTape() as tape:
out_bicu = generate_bicubic_samples(lr_patchs.numpy(), CONFIG)
#print( lr_patchs.shape, out_bicu.shape)
fake_patchs = gen_model([lr_patchs, out_bicu])
mse_loss = tl.cost.mean_squared_error(fake_patchs, hr_patchs, is_mean=True)
grad = tape.gradient(mse_loss, gen_model.trainable_weights)
g_optimizer_init.apply_gradients(zip(grad, gen_model.trainable_weights))
print("Epoch: [{}/{}] step: [{}/{}] time: {:.3f}s, mse: {:.6f} ".format(
epoch+1+resume_epoch, resume_epoch+n_epoch_init, step+1, CONFIG.no_of_batches, time.time() - step_time, mse_loss))
path = 'Training_outputs/gan_init_{}_train_{}.png'.format(CONFIG.gen_model, epoch+1+resume_epoch)
tl.vis.save_images(fake_patchs.numpy(), [2, CONFIG.batch_size//2], path)
if ((epoch+1+resume_epoch) % CONFIG.save_interval) == 0:
gen_model.save_weights('Checkpoints/GAN_INIT_{}_EPID_{}.h5'.format(CONFIG.gen_model, epoch+1+resume_epoch))
gen_model.save_weights('Checkpoints/GAN_INIT_{}_EPID_{}.h5'.format(CONFIG.gen_model, n_epoch_init + resume_epoch))
print('## adversarial learning (G, D)')
for epoch in range(n_epoch):
for step, (lr_patchs, hr_patchs) in enumerate(train_ds):
if lr_patchs.shape[0] != batch_size: # if the remaining data in this epoch < batch_size
break
step_time = time.time()
with tf1.GradientTape(persistent=True) as tape:
#out_bicu = generate_bicubic_samples(np.squeeze(lr_patchs,axis=3), CONFIG.scale)
out_bicu = generate_bicubic_samples(lr_patchs.numpy(), CONFIG)
#print( lr_patchs.shape, out_bicu.shape)
fake_patchs = gen_model([lr_patchs, out_bicu])
logits_fake = dis_model(fake_patchs)
logits_real = dis_model(hr_patchs)
feature_fake = VGG((fake_patchs+1)/2.) # the pre-trained VGG uses the input range of [0, 1]
feature_real = VGG((hr_patchs+1)/2.)
d_loss1 = tl.cost.sigmoid_cross_entropy(logits_real, tf1.ones_like(logits_real))
d_loss2 = tl.cost.sigmoid_cross_entropy(logits_fake, tf1.zeros_like(logits_fake))
d_loss = d_loss1 + d_loss2
g_gan_loss = 1e-3 * tl.cost.sigmoid_cross_entropy(logits_fake, tf1.ones_like(logits_fake))
mse_loss = tl.cost.mean_squared_error(fake_patchs, hr_patchs, is_mean=True)
vgg_loss = 2e-6 * tl.cost.mean_squared_error(feature_fake, feature_real, is_mean=True)
g_loss = mse_loss + vgg_loss + g_gan_loss
grad = tape.gradient(g_loss, gen_model.trainable_weights)
g_optimizer.apply_gradients(zip(grad, gen_model.trainable_weights))
grad = tape.gradient(d_loss, dis_model.trainable_weights)
d_optimizer.apply_gradients(zip(grad, dis_model.trainable_weights))
print("Epoch: [{}/{}] step: [{}/{}] time: {:.3f}s, g_loss(mse:{:.6f}, vgg:{:.6f}, adv:{:.6f}) d_loss: {:.6f}".format(
epoch+1+resume_epoch, resume_epoch + n_epoch, step+1, CONFIG.no_of_batches, time.time() - step_time, mse_loss, vgg_loss, g_gan_loss, d_loss))
# update the learning rate
'''if (epoch+resume_epoch) % decay_every == 0:
new_lr_decay = lr_decay**((epoch+resume_epoch)// decay_every)
lr_v.assign(lr_init * new_lr_decay)
log = " ** new learning rate: %f (for GAN)" % (lr_init * new_lr_decay)
print(log)
'''
if (epoch+1+resume_epoch)% CONFIG.save_interval == 0:
gen_model.save_weights('Checkpoints/GAN_{}_EPID_{}.h5'.format(CONFIG.gen_model, epoch+1+resume_epoch))
dis_model.save_weights('Checkpoints/DIS_{}_GAN_{}_EPID_{}.h5'.format(CONFIG.dis_model, CONFIG.gen_model, epoch+1+resume_epoch))
print("Save time: {}content".format(time.asctime( time.localtime(time.time()))))
for i in range(CONFIG.batch_size):
if CONFIG.gen_model==1:
lrimg = np.squeeze(lr_patchs[i], axis =-1)
lrimg = np.pad(lrimg, ((64, 64), (64, 64)), constant_values=(255.0))
#opimg = cast_uint8(fake_patchs[i].numpy())
opimg = fake_patchs[i].numpy()
combine_imgs = np.concatenate((lrimg[:,:,np.newaxis], out_bicu[i], opimg, hr_patchs[i]), axis = 1)
else:
lrimg = np.pad(lr_patchs[i], ((192, 192), (192, 192), (0, 0)), constant_values=(255.0))
#opimg = cast_uint8(fake_patchs[i].numpy())
opimg = fake_patchs[i].numpy()
combine_imgs = np.concatenate((lrimg, out_bicu[i], opimg, hr_patchs[i]), axis = 1)
path = 'Training_outputs/id_{}_gan_{}_train_{}.png'.format(i+1, CONFIG.gen_model, epoch+1+resume_epoch)
tl.vis.save_image(combine_imgs,path)
gen_model.save_weights('Checkpoints/GAN_{}_FINAL.h5'.format(CONFIG.gen_model))
dis_model.save_weights('Checkpoints/DIS_{}_GAN_{}_FINAL.h5'.format(CONFIG.dis_model, CONFIG.gen_model))
elif CONFIG.mode==2: ## Validation
model = get_model('G', CONFIG.gen_model)
model.load_weights('Checkpoints/GAN_{}_EPID_{}.h5'.format(CONFIG.gen_model, CONFIG.model_epoch))
model.eval() ## disable dropout, batch norm moving avg ...
save_time = time.time()
## Reading Validation dataset
lrimg_file_list = tl.files.load_file_list(path=CONFIG.dir_val_in, regx='.*.png', printable=False)
hrimg_file_list = tl.files.load_file_list(path=CONFIG.dir_val_target, regx='.*.png', printable=False)
lrimg_file_list.sort(key=tl.files.natural_keys)
hrimg_file_list.sort(key=tl.files.natural_keys)
lrimg_list = np.array(tl.vis.read_images(lrimg_file_list, path=CONFIG.dir_val_in, n_threads=32))
hrimg_list = np.array(tl.vis.read_images(hrimg_file_list, path=CONFIG.dir_val_target, n_threads=32))
if CONFIG.gen_model==1:
lrimg_list = lrimg_list[:,:,:,np.newaxis]
hrimg_list = hrimg_list[:,:,:,np.newaxis]
bcimg_list = generate_bicubic_samples(lrimg_list,CONFIG)
opimg_list = model([tf1.cast(lrimg_list,tf1.float32), tf1.cast(bcimg_list,tf1.float32)])
opimg_list = opimg_list.numpy()
bicubic_psnr, model_psnr = PSNR (hrimg_list , bcimg_list, opimg_list)
bicubic_ssim, model_ssim = SSIM (hrimg_list , bcimg_list, opimg_list)
for i in range(lrimg_list.shape[0]):
name= lrimg_file_list[i].split('/')[-1].split('.')[0]
if CONFIG.gen_model==1:
lrimg = np.pad(lrimg_list[i], ((64, 64), (64, 64),(0, 0)), constant_values=(255.0))
else:
lrimg = np.pad(lrimg_list[i], ((192, 192), (192, 192), (0, 0)), constant_values=(255.0))
combine_imgs = np.concatenate((lrimg, bcimg_list[i], opimg_list[i], hrimg_list[i]), axis = 1)
path = 'Validation_outputs/{}_gan_{}_val_{}.png'.format(name, CONFIG.gen_model, CONFIG.model_epoch)
tl.vis.save_image(combine_imgs, path)
print(np.stack((model_psnr, bicubic_psnr), axis=-1))
print(np.stack((model_ssim, bicubic_ssim), axis=-1))
print(np.subtract(model_psnr, bicubic_psnr))
print('SUM(PSNR DIFF): {}'.format(np.sum(np.subtract(model_psnr, bicubic_psnr))))
print('AVG MODEL PSNR: {}, AVG BICUBIC PSNR: {}'.format(np.sum(model_psnr)/lrimg_list.shape[0], np.sum(bicubic_psnr)/lrimg_list.shape[0]))
print('SUM(SSIM DIFF): {}'.format(np.sum(np.subtract(model_ssim, bicubic_ssim))))
print('AVG MODEL SSIM: {}, AVG BICUBIC SSIM: {}'.format(np.sum(model_ssim)/lrimg_list.shape[0], np.sum(bicubic_ssim)/lrimg_list.shape[0]))
print((time.time()-save_time)/10)
def run():
torch.ops.load_library("./Renderer.dll")
#########################
# CONFIGURATION
#########################
if 0:
OUTPUT_FOLDER = "../result-stats/adaptiveIso2/"
DATASET_PREFIX = "D:/VolumeSuperResolution-InputData/"
DATASETS = [
#("Ejecta", "gt-rendering-ejecta-v2-test.hdf5"),
("RM", "gt-rendering-rm-v1.hdf5"),
#("Human", "gt-rendering-human-v1.hdf5"),
#("Thorax", "gt-rendering-thorax-v1.hdf5"),
]
NETWORK_DIR = "D:/VolumeSuperResolution/adaptive-modeldir/"
NETWORKS = [ #suffixed with _importance.pt and _recon.pt
#("adaptive011", "adaptive011_epoch500"), #title, file prefix
("adaptive019", "adaptive019_epoch470"),
("adaptive023", "adaptive023_epoch300")
]
SAMPLING_FILE = "D:/VolumeSuperResolution-InputData/samplingPattern.hdf5"
SAMPLING_PATTERNS = ['halton', 'plastic', 'random']
HEATMAP_MIN = [0.01, 0.05, 0.2]
HEATMAP_MEAN = [0.05, 0.1, 0.2, 0.5]
UPSCALING = 8 # = networkUp * postUp
IMPORTANCE_BORDER = 8
LOSS_BORDER = 32
BATCH_SIZE = 8
elif 0:
OUTPUT_FOLDER = "../result-stats/adaptiveIsoEnhance6/"
DATASET_PREFIX = "D:/VolumeSuperResolution-InputData/"
DATASETS = [
("Ejecta", "gt-rendering-ejecta-v2-test.hdf5"),
#("RM", "gt-rendering-rm-v1.hdf5"),
#("Human", "gt-rendering-human-v1.hdf5"),
#("Thorax", "gt-rendering-thorax-v1.hdf5"),
#("Head", "gt-rendering-head.hdf5"),
]
NETWORK_DIR = "D:/VolumeSuperResolution/adaptive-modeldir/"
NETWORKS = [ #suffixed with _importance.pt and _recon.pt
#title, file prefix
#("U-Net (5-4)", "sizes/size5-4_epoch500"),
#("Enhance-Net (epoch 50)", "enhance2_imp050_epoch050"),
#("Enhance-Net (epoch 400)", "enhance2_imp050_epoch400"),
#("Enhance-Net (Thorax)", "enhance_imp050_Thorax_epoch200"),
#("Enhance-Net (RM)", "enhance_imp050_RM_epoch200"),
#("Imp100", "enhance4_imp100_epoch300"),
#("Imp100res", "enhance4_imp100res_epoch230"),
("Imp100res+N", "enhance4_imp100res+N_epoch300"),
("Imp100+N", "enhance4_imp100+N_epoch300"),
#("Imp100+N-res", "enhance4_imp100+N-res_epoch300"),
#("Imp100+N-resInterp", "enhance4_imp100+N-resInterp_epoch300"),
#("U-Net (5-4)", "size5-4_epoch500"),
#("U-Net (5-3)", "size5-3_epoch500"),
#("U-Net (4-4)", "size4-4_epoch500"),
]
# Test if it is better to post-train with dense networks and PDE inpainting
POSTTRAIN_NETWORK_DIR = "D:/VolumeSuperResolution/dense-modeldir/"
POSTTRAIN_NETWORKS = [
# title, file suffix to POSTTRAIN_NETWORK_DIR, inpainting {'fast', 'pde'}
#("Enhance PDE (post)", "inpHv2-pde05-epoch200.pt", "pde")
]
SAMPLING_FILE = "D:/VolumeSuperResolution-InputData/samplingPattern.hdf5"
SAMPLING_PATTERNS = ['plastic']
HEATMAP_MIN = [0.002]
HEATMAP_MEAN = [
0.05
] #[0.01, 0.02, 0.03, 0.04, 0.06, 0.08, 0.1, 0.2, 0.3, 0.5, 0.8, 1.0]
USE_BINARY_SEARCH_ON_MEAN = True
UPSCALING = 8 # = networkUp * postUp
IMPORTANCE_BORDER = 8
LOSS_BORDER = 32
BATCH_SIZE = 4
elif 0:
OUTPUT_FOLDER = "../result-stats/adaptiveIsoEnhance5Sampling/"
DATASET_PREFIX = "D:/VolumeSuperResolution-InputData/"
DATASETS = [
("Ejecta", "gt-rendering-ejecta-v2-test.hdf5"),
]
NETWORK_DIR = "D:/VolumeSuperResolution/adaptive-modeldir/"
NETWORKS = [ #suffixed with _importance.pt and _recon.pt
#title, file prefix
("Enhance-Net (epoch 400)", "enhance2_imp050_epoch400"),
]
# Test if it is better to post-train with dense networks and PDE inpainting
POSTTRAIN_NETWORK_DIR = "D:/VolumeSuperResolution/dense-modeldir/"
POSTTRAIN_NETWORKS = [
# title, file suffix to POSTTRAIN_NETWORK_DIR, inpainting {'fast', 'pde'}
#("Enhance PDE (post)", "inpHv2-pde05-epoch200.pt", "pde")
]
SAMPLING_FILE = "D:/VolumeSuperResolution-InputData/samplingPattern.hdf5"
SAMPLING_PATTERNS = ['halton', 'plastic', 'random', 'regular']
#SAMPLING_PATTERNS = ['regular']
HEATMAP_MIN = [0.002]
HEATMAP_MEAN = [0.05]
USE_BINARY_SEARCH_ON_MEAN = True
UPSCALING = 8 # = networkUp * postUp
IMPORTANCE_BORDER = 8
LOSS_BORDER = 32
BATCH_SIZE = 4
elif 1:
OUTPUT_FOLDER = "../result-stats/adaptiveIsoEnhance8Sampling/"
DATASET_PREFIX = "D:/VolumeSuperResolution-InputData/"
DATASETS = [
("Ejecta", "gt-rendering-ejecta-v2-test.hdf5"),
#("RM", "gt-rendering-rm-v1.hdf5"),
#("Human", "gt-rendering-human-v1.hdf5"),
#("Thorax", "gt-rendering-thorax-v1.hdf5"),
#("Head", "gt-rendering-head.hdf5"),
]
NETWORK_DIR = "D:/VolumeSuperResolution/adaptive-modeldir/"
NETWORKS = [ #suffixed with _importance.pt and _recon.pt
#title, file prefix
("regular", "enhance7_regular_epoch190"),
("random", "enhance7_random_epoch190"),
("halton", "enhance7_halton_epoch190"),
("plastic", "enhance7_plastic_epoch190"),
]
# Test if it is better to post-train with dense networks and PDE inpainting
POSTTRAIN_NETWORK_DIR = "D:/VolumeSuperResolution/dense-modeldir/"
POSTTRAIN_NETWORKS = [
# title, file suffix to POSTTRAIN_NETWORK_DIR, inpainting {'fast', 'pde'}
#("Enhance PDE (post)", "inpHv2-pde05-epoch200.pt", "pde")
]
SAMPLING_FILE = "D:/VolumeSuperResolution-InputData/samplingPattern.hdf5"
SAMPLING_PATTERNS = ['regular', 'random', 'halton', 'plastic']
HEATMAP_MIN = [0.002]
HEATMAP_MEAN = [
0.05
] #[0.01, 0.02, 0.03, 0.04, 0.06, 0.08, 0.1, 0.2, 0.3, 0.5, 0.8, 1.0]
USE_BINARY_SEARCH_ON_MEAN = True
UPSCALING = 8 # = networkUp * postUp
IMPORTANCE_BORDER = 8
LOSS_BORDER = 32
BATCH_SIZE = 4
elif 0:
OUTPUT_FOLDER = "../result-stats/adaptiveIsoImp/"
DATASET_PREFIX = "D:/VolumeSuperResolution-InputData/"
DATASETS = [
("Ejecta", "gt-rendering-ejecta-v2-test.hdf5"),
#("RM", "gt-rendering-rm-v1.hdf5"),
#("Human", "gt-rendering-human-v1.hdf5"),
#("Thorax", "gt-rendering-thorax-v1.hdf5"),
]
NETWORK_DIR = "D:/VolumeSuperResolution/adaptive-modeldir/imp/"
NETWORKS = [ #suffixed with _importance.pt and _recon.pt
#("adaptive011", "adaptive011_epoch500"), #title, file prefix
("imp005", "imp005_epoch500"),
("imp010", "imp010_epoch500"),
("imp020", "imp020_epoch500"),
("imp050", "imp050_epoch500"),
]
SAMPLING_FILE = "D:/VolumeSuperResolution-InputData/samplingPattern.hdf5"
SAMPLING_PATTERNS = ['halton']
HEATMAP_MIN = [0.002]
HEATMAP_MEAN = [0.005, 0.01, 0.02, 0.05, 0.1]
USE_BINARY_SEARCH_ON_MEAN = True
UPSCALING = 8 # = networkUp * postUp
IMPORTANCE_BORDER = 8
LOSS_BORDER = 32
BATCH_SIZE = 16
#########################
# LOADING
#########################
device = torch.device("cuda")
# Load Networks
IMPORTANCE_BASELINE1 = "ibase1"
IMPORTANCE_BASELINE2 = "ibase2"
IMPORTANCE_BASELINE3 = "ibase3"
RECON_BASELINE = "rbase"
# load importance model
print("load importance networks")
class ImportanceModel:
def __init__(self, file):
if file == IMPORTANCE_BASELINE1:
self._net = importance.UniformImportanceMap(1, 0.5)
self._upscaling = 1
self._name = "constant"
self.disableTemporal = True
self._requiresPrevious = False
elif file == IMPORTANCE_BASELINE2:
self._net = importance.GradientImportanceMap(
1, (1, 1), (2, 1), (3, 1))
self._upscaling = 1
self._name = "curvature"
self.disableTemporal = True
self._requiresPrevious = False
else:
self._name = file[0]
file = os.path.join(NETWORK_DIR, file[1] + "_importance.pt")
extra_files = torch._C.ExtraFilesMap()
extra_files['settings.json'] = ""
self._net = torch.jit.load(file,
map_location=device,
_extra_files=extra_files)
settings = json.loads(extra_files['settings.json'])
self._upscaling = settings['networkUpscale']
self._requiresPrevious = settings.get("requiresPrevious",
False)
self.disableTemporal = settings.get("disableTemporal", True)
def networkUpscaling(self):
return self._upscaling
def name(self):
return self._name
def __repr__(self):
return self.name()
def call(self, input, prev_warped_out):
if self._requiresPrevious:
input = torch.cat([
input,
models.VideoTools.flatten_high(prev_warped_out,
self._upscaling)
],
dim=1)
input = F.pad(input, [IMPORTANCE_BORDER] * 4, 'constant', 0)
output = self._net(input) # the network call
output = F.pad(output, [-IMPORTANCE_BORDER * self._upscaling] * 4,
'constant', 0)
return output
class LuminanceImportanceModel:
def __init__(self):
self.disableTemporal = True
def setTestFile(self, filename):
importance_file = filename[:-5] + "-luminanceImportance.hdf5"
if os.path.exists(importance_file):
self._exist = True
self._file = h5py.File(importance_file, 'r')
self._dset = self._file['importance']
else:
self._exist = False
self._file = None
self._dset = None
def isAvailable(self):
return self._exist
def setIndices(self, indices: torch.Tensor):
assert len(indices.shape) == 1
self._indices = list(indices.cpu().numpy())
def setTime(self, time):
self._time = time
def networkUpscaling(self):
return UPSCALING
def name(self):
return "luminance-contrast"
def __repr__(self):
return self.name()
def call(self, input, prev_warped_out):
B, C, H, W = input.shape
if not self._exist:
return torch.ones(B,
1,
H,
W,
dtype=input.dtype,
device=input.device)
outputs = []
for idx in self._indices:
outputs.append(
torch.from_numpy(self._dset[idx, self._time,
...]).to(device=input.device))
return torch.stack(outputs, dim=0)
importanceBaseline1 = ImportanceModel(IMPORTANCE_BASELINE1)
importanceBaseline2 = ImportanceModel(IMPORTANCE_BASELINE2)
importanceBaselineLuminance = LuminanceImportanceModel()
importanceModels = [ImportanceModel(f) for f in NETWORKS]
# load reconstruction networks
print("load reconstruction networks")
class ReconstructionModel:
def __init__(self, file):
if file == RECON_BASELINE:
class Inpainting(nn.Module):
def forward(self, x, mask):
input = x[:, 0:6, :, :].contiguous(
) # mask, normal xyz, depth, ao
mask = x[:, 6, :, :].contiguous()
return torch.ops.renderer.fast_inpaint(mask, input)
self._net = Inpainting()
self._upscaling = 1
self._name = "inpainting"
self.disableTemporal = True
else:
self._name = file[0]
file = os.path.join(NETWORK_DIR, file[1] + "_recon.pt")
extra_files = torch._C.ExtraFilesMap()
extra_files['settings.json'] = ""
self._net = torch.jit.load(file,
map_location=device,
_extra_files=extra_files)
self._settings = json.loads(extra_files['settings.json'])
self.disableTemporal = False
requiresMask = self._settings.get('expectMask', False)
if self._settings.get("interpolateInput", False):
self._originalNet = self._net
class Inpainting2(nn.Module):
def __init__(self, orignalNet, requiresMask):
super().__init__()
self._n = orignalNet
self._requiresMask = requiresMask
def forward(self, x, mask):
input = x[:, 0:6, :, :].contiguous(
) # mask, normal xyz, depth, ao
mask = x[:, 6, :, :].contiguous()
inpainted = torch.ops.renderer.fast_inpaint(
mask, input)
x[:, 0:6, :, :] = inpainted
if self._requiresMask:
return self._n(x, mask)
else:
return self._n(x)
self._net = Inpainting2(self._originalNet, requiresMask)
def name(self):
return self._name
def __repr__(self):
return self.name()
def call(self, input, mask, prev_warped_out):
input = torch.cat([input, prev_warped_out], dim=1)
output = self._net(input, mask)
return output
class ReconstructionModelPostTrain:
"""
Reconstruction model that are trained as dense reconstruction networks
after the adaptive training.
They don't recive the sampling mask as input, but can start with PDE-based inpainting
"""
def __init__(self, name: str, model_path: str, inpainting: str):
assert inpainting == 'fast' or inpainting == 'pde', "inpainting must be either 'fast' or 'pde', but got %s" % inpainting
self._inpainting = inpainting
self._name = name
file = os.path.join(POSTTRAIN_NETWORK_DIR, model_path)
extra_files = torch._C.ExtraFilesMap()
extra_files['settings.json'] = ""
self._net = torch.jit.load(file,
map_location=device,
_extra_files=extra_files)
self._settings = json.loads(extra_files['settings.json'])
assert self._settings.get(
'upscale_factor', None) == 1, "selected file is not a 1x SRNet"
self.disableTemporal = False
def name(self):
return self._name
def __repr__(self):
return self.name()
def call(self, input, prev_warped_out):
# no sampling and no AO
input_no_sampling = input[:, 0:5, :, :].contiguous(
) # mask, normal xyz, depth
sampling_mask = input[:, 6, :, :].contiguous()
# perform inpainting
if self._inpainting == 'pde':
inpainted = torch.ops.renderer.pde_inpaint(
sampling_mask,
input_no_sampling,
200,
1e-4,
5,
2, # m0, epsilon, m1, m2
0, # mc -> multigrid recursion count. =0 disables the multigrid hierarchy
9,
0) # ms, m3
else:
inpainted = torch.ops.renderer.fast_inpaint(
sampling_mask, input_no_sampling)
# run network
input = torch.cat([inpainted, prev_warped_out], dim=1)
output = self._net(input)
if isinstance(output, tuple):
output = output[0]
return output
reconBaseline = ReconstructionModel(RECON_BASELINE)
reconModels = [ReconstructionModel(f) for f in NETWORKS]
reconPostModels = [
ReconstructionModelPostTrain(name, file, inpainting)
for (name, file, inpainting) in POSTTRAIN_NETWORKS
]
allReconModels = reconModels + reconPostModels
NETWORK_COMBINATIONS = \
[(importanceBaseline1, reconBaseline), (importanceBaseline2, reconBaseline)] + \
[(importanceBaselineLuminance, reconBaseline)] + \
[(importanceBaseline1, reconNet) for reconNet in allReconModels] + \
[(importanceBaseline2, reconNet) for reconNet in allReconModels] + \
[(importanceBaselineLuminance, reconNet) for reconNet in allReconModels] + \
[(importanceNet, reconBaseline) for importanceNet in importanceModels] + \
list(zip(importanceModels, reconModels)) + \
[(importanceNet, reconPostModel) for importanceNet in importanceModels for reconPostModel in reconPostModels]
#NETWORK_COMBINATIONS = list(zip(importanceModels, reconModels))
print("Network combinations:")
for (i, r) in NETWORK_COMBINATIONS:
print(" %s - %s" % (i.name(), r.name()))
# load sampling patterns
print("load sampling patterns")
with h5py.File(SAMPLING_FILE, 'r') as f:
sampling_pattern = dict([(name, torch.from_numpy(f[name][...]).to(device)) \
for name in SAMPLING_PATTERNS])
# create shading
shading = ScreenSpaceShading(device)
shading.fov(30)
shading.ambient_light_color(np.array([0.1, 0.1, 0.1]))
shading.diffuse_light_color(np.array([1.0, 1.0, 1.0]))
shading.specular_light_color(np.array([0.0, 0.0, 0.0]))
shading.specular_exponent(16)
shading.light_direction(np.array([0.1, 0.1, 1.0]))
shading.material_color(np.array([1.0, 0.3, 0.0]))
AMBIENT_OCCLUSION_STRENGTH = 1.0
shading.ambient_occlusion(1.0)
shading.inverse_ao = False
#heatmap
HEATMAP_CFG = [(min, mean) for min in HEATMAP_MIN for mean in HEATMAP_MEAN
if min < mean]
print("heatmap configs:", HEATMAP_CFG)
#########################
# DEFINE STATISTICS
#########################
ssimLoss = SSIM(size_average=False)
ssimLoss.to(device)
psnrLoss = PSNR()
psnrLoss.to(device)
lpipsColor = lpips.PerceptualLoss(model='net-lin',
net='alex',
use_gpu=True)
MIN_FILLING = 0.05
NUM_BINS = 200
class Statistics:
def __init__(self):
self.histogram_color_withAO = np.zeros(NUM_BINS, dtype=np.float64)
self.histogram_color_noAO = np.zeros(NUM_BINS, dtype=np.float64)
self.histogram_depth = np.zeros(NUM_BINS, dtype=np.float64)
self.histogram_normal = np.zeros(NUM_BINS, dtype=np.float64)
self.histogram_mask = np.zeros(NUM_BINS, dtype=np.float64)
self.histogram_ao = np.zeros(NUM_BINS, dtype=np.float64)
self.histogram_counter = 0
def create_datasets(self, hdf5_file: h5py.File, stats_name: str,
histo_name: str, num_samples: int,
extra_info: dict):
self.expected_num_samples = num_samples
stats_shape = (num_samples, len(list(StatField)))
self.stats_file = hdf5_file.require_dataset(stats_name,
stats_shape,
dtype='f',
exact=True)
self.stats_file.attrs['NumFields'] = len(list(StatField))
for field in list(StatField):
self.stats_file.attrs['Field%d' % field.value] = field.name
for key, value in extra_info.items():
self.stats_file.attrs[key] = value
self.stats_index = 0
histo_shape = (NUM_BINS, len(list(HistoField)))
self.histo_file = hdf5_file.require_dataset(histo_name,
histo_shape,
dtype='f',
exact=True)
self.histo_file.attrs['NumFields'] = len(list(HistoField))
for field in list(HistoField):
self.histo_file.attrs['Field%d' % field.value] = field.name
for key, value in extra_info.items():
self.histo_file.attrs[key] = value
def add_timestep_sample(self, pred_mnda, gt_mnda, sampling_mask):
"""
adds a timestep sample:
pred_mnda: prediction: mask, normal, depth, AO
gt_mnda: ground truth: mask, normal, depth, AO
"""
B = pred_mnda.shape[0]
#shading
shading.ambient_occlusion(AMBIENT_OCCLUSION_STRENGTH)
pred_color_withAO = shading(pred_mnda)
gt_color_withAO = shading(gt_mnda)
shading.ambient_occlusion(0.0)
pred_color_noAO = shading(pred_mnda)
gt_color_noAO = shading(gt_mnda)
#apply border
pred_mnda = pred_mnda[:, :, LOSS_BORDER:-LOSS_BORDER,
LOSS_BORDER:-LOSS_BORDER]
pred_color_withAO = pred_color_withAO[:, :,
LOSS_BORDER:-LOSS_BORDER,
LOSS_BORDER:-LOSS_BORDER]
pred_color_noAO = pred_color_noAO[:, :, LOSS_BORDER:-LOSS_BORDER,
LOSS_BORDER:-LOSS_BORDER]
gt_mnda = gt_mnda[:, :, LOSS_BORDER:-LOSS_BORDER,
LOSS_BORDER:-LOSS_BORDER]
gt_color_withAO = gt_color_withAO[:, :, LOSS_BORDER:-LOSS_BORDER,
LOSS_BORDER:-LOSS_BORDER]
gt_color_noAO = gt_color_noAO[:, :, LOSS_BORDER:-LOSS_BORDER,
LOSS_BORDER:-LOSS_BORDER]
mask = gt_mnda[:, 0:1, :, :] * 0.5 + 0.5
# PSNR
psnr_mask = psnrLoss(pred_mnda[:, 0:1, :, :],
gt_mnda[:, 0:1, :, :]).cpu().numpy()
psnr_normal = psnrLoss(pred_mnda[:, 1:4, :, :],
gt_mnda[:, 1:4, :, :],
mask=mask).cpu().numpy()
psnr_depth = psnrLoss(pred_mnda[:, 4:5, :, :],
gt_mnda[:, 4:5, :, :],
mask=mask).cpu().numpy()
psnr_ao = psnrLoss(pred_mnda[:, 5:6, :, :],
gt_mnda[:, 5:6, :, :],
mask=mask).cpu().numpy()
psnr_color_withAO = psnrLoss(pred_color_withAO,
gt_color_withAO,
mask=mask).cpu().numpy()
psnr_color_noAO = psnrLoss(pred_color_noAO,
gt_color_noAO,
mask=mask).cpu().numpy()
# SSIM
ssim_mask = ssimLoss(pred_mnda[:, 0:1, :, :],
gt_mnda[:, 0:1, :, :]).cpu().numpy()
pred_mnda = gt_mnda + mask * (pred_mnda - gt_mnda)
ssim_normal = ssimLoss(pred_mnda[:, 1:4, :, :],
gt_mnda[:, 1:4, :, :]).cpu().numpy()
ssim_depth = ssimLoss(pred_mnda[:, 4:5, :, :],
gt_mnda[:, 4:5, :, :]).cpu().numpy()
ssim_ao = ssimLoss(pred_mnda[:, 5:6, :, :],
gt_mnda[:, 5:6, :, :]).cpu().numpy()
ssim_color_withAO = ssimLoss(pred_color_withAO,
gt_color_withAO).cpu().numpy()
ssim_color_noAO = ssimLoss(pred_color_noAO,
gt_color_noAO).cpu().numpy()
# Perceptual
lpips_color_withAO = torch.cat([
lpipsColor(
pred_color_withAO[b], gt_color_withAO[b], normalize=True)
for b in range(B)
],
dim=0).cpu().numpy()
lpips_color_noAO = torch.cat([
lpipsColor(
pred_color_noAO[b], gt_color_noAO[b], normalize=True)
for b in range(B)
],
dim=0).cpu().numpy()
# Samples
samples = torch.mean(sampling_mask, dim=(1, 2, 3)).cpu().numpy()
# Write samples to file
for b in range(B):
assert self.stats_index < self.expected_num_samples, "Adding more samples than specified"
self.stats_file[self.stats_index, :] = np.array([
psnr_mask[b], psnr_normal[b], psnr_depth[b], psnr_ao[b],
psnr_color_noAO[b], psnr_color_withAO[b], ssim_mask[b],
ssim_normal[b], ssim_depth[b], ssim_ao[b],
ssim_color_noAO[b], ssim_color_withAO[b],
lpips_color_noAO[b], lpips_color_withAO[b], samples[b]
],
dtype='f')
self.stats_index += 1
# Histogram
self.histogram_counter += 1
mask_diff = F.l1_loss(gt_mnda[:, 0, :, :],
pred_mnda[:, 0, :, :],
reduction='none')
histogram, _ = np.histogram(mask_diff.cpu().numpy(),
bins=NUM_BINS,
range=(0, 1),
density=True)
self.histogram_mask += (
histogram /
(NUM_BINS * B) - self.histogram_mask) / self.histogram_counter
#normal_diff = (-F.cosine_similarity(gt_mnda[0,1:4,:,:], pred_mnda[0,1:4,:,:], dim=0)+1)/2
normal_diff = F.l1_loss(gt_mnda[:, 1:4, :, :],
pred_mnda[:, 1:4, :, :],
reduction='none').sum(dim=0) / 6
histogram, _ = np.histogram(normal_diff.cpu().numpy(),
bins=NUM_BINS,
range=(0, 1),
density=True)
self.histogram_normal += (histogram /
(NUM_BINS * B) - self.histogram_normal
) / self.histogram_counter
depth_diff = F.l1_loss(gt_mnda[:, 4, :, :],
pred_mnda[:, 4, :, :],
reduction='none')
histogram, _ = np.histogram(depth_diff.cpu().numpy(),
bins=NUM_BINS,
range=(0, 1),
density=True)
self.histogram_depth += (
histogram /
(NUM_BINS * B) - self.histogram_depth) / self.histogram_counter
ao_diff = F.l1_loss(gt_mnda[:, 5, :, :],
pred_mnda[:, 5, :, :],
reduction='none')
histogram, _ = np.histogram(ao_diff.cpu().numpy(),
bins=NUM_BINS,
range=(0, 1),
density=True)
self.histogram_ao += (histogram / (NUM_BINS * B) -
self.histogram_ao) / self.histogram_counter
color_diff = F.l1_loss(gt_color_withAO[:, 0, :, :],
pred_color_withAO[:, 0, :, :],
reduction='none')
histogram, _ = np.histogram(color_diff.cpu().numpy(),
bins=NUM_BINS,
range=(0, 1),
density=True)
self.histogram_color_withAO += (
histogram / (NUM_BINS * B) -
self.histogram_color_withAO) / self.histogram_counter
color_diff = F.l1_loss(gt_color_noAO[:, 0, :, :],
pred_color_noAO[:, 0, :, :],
reduction='none')
histogram, _ = np.histogram(color_diff.cpu().numpy(),
bins=NUM_BINS,
range=(0, 1),
density=True)
self.histogram_color_noAO += (
histogram / (NUM_BINS * B) -
self.histogram_color_noAO) / self.histogram_counter
def close_stats_file(self):
self.stats_file.attrs['NumEntries'] = self.stats_index
def write_histogram(self):
"""
After every sample for the current dataset was processed, write
a histogram of the errors in a new file
"""
for i in range(NUM_BINS):
self.histo_file[i, :] = np.array([
i / NUM_BINS, (i + 1) / NUM_BINS, self.histogram_mask[i],
self.histogram_normal[i], self.histogram_depth[i],
self.histogram_ao[i], self.histogram_color_withAO[i],
self.histogram_color_noAO[i]
])
#########################
# DATASET
#########################
class FullResDataset(torch.utils.data.Dataset):
def __init__(self, file):
self.hdf5_file = h5py.File(file, 'r')
self.dset = self.hdf5_file['gt']
print("Dataset shape:", self.dset.shape)
def __len__(self):
return self.dset.shape[0]
def num_timesteps(self):
return self.dset.shape[1]
def __getitem__(self, idx):
return (self.dset[idx, ...], np.array(idx))
#########################
# COMPUTE STATS for each dataset
#########################
for dataset_name, dataset_file in DATASETS:
dataset_file = os.path.join(DATASET_PREFIX, dataset_file)
print("Compute statistics for", dataset_name)
# init luminance importance map
importanceBaselineLuminance.setTestFile(dataset_file)
if importanceBaselineLuminance.isAvailable():
print("Luminance-contrast importance map is available")
# create output file
os.makedirs(OUTPUT_FOLDER, exist_ok=True)
output_file = os.path.join(OUTPUT_FOLDER, dataset_name + '.hdf5')
print("Save to", output_file)
with h5py.File(output_file, 'a') as output_hdf5_file:
# load dataset
set = FullResDataset(dataset_file)
data_loader = torch.utils.data.DataLoader(set,
batch_size=BATCH_SIZE,
shuffle=False)
# define statistics
StatsCfg = collections.namedtuple(
"StatsCfg", "stats importance recon heatmin heatmean pattern")
statistics = []
for (inet, rnet) in NETWORK_COMBINATIONS:
for (heatmin, heatmean) in HEATMAP_CFG:
for pattern in SAMPLING_PATTERNS:
stats_info = {
'importance': inet.name(),
'reconstruction': rnet.name(),
'heatmin': heatmin,
'heatmean': heatmean,
'pattern': pattern
}
stats_filename = "Stats_%s_%s_%03d_%03d_%s" % (
inet.name(), rnet.name(), heatmin * 100,
heatmean * 100, pattern)
histo_filename = "Histogram_%s_%s_%03d_%03d_%s" % (
inet.name(), rnet.name(), heatmin * 100,
heatmean * 100, pattern)
s = Statistics()
s.create_datasets(output_hdf5_file, stats_filename,
histo_filename,
len(set) * set.num_timesteps(),
stats_info)
statistics.append(
StatsCfg(stats=s,
importance=inet,
recon=rnet,
heatmin=heatmin,
heatmean=heatmean,
pattern=pattern))
print(len(statistics),
" different combinations are performed per sample")
# compute statistics
try:
with torch.no_grad():
num_minibatch = len(data_loader)
pg = ProgressBar(num_minibatch, 'Evaluation', length=50)
for iteration, (batch, batch_indices) in enumerate(
data_loader, 0):
pg.print_progress_bar(iteration)
batch = batch.to(device)
importanceBaselineLuminance.setIndices(batch_indices)
B, T, C, H, W = batch.shape
# try out each combination
for s in statistics:
#print(s)
# get input to evaluation
importanceNetUpscale = s.importance.networkUpscaling(
)
importancePostUpscale = UPSCALING // importanceNetUpscale
crop_low = torch.nn.functional.interpolate(
batch.reshape(B * T, C, H, W),
scale_factor=1 / UPSCALING,
mode='area').reshape(B, T, C, H // UPSCALING,
W // UPSCALING)
pattern = sampling_pattern[s.pattern][:H, :W]
crop_high = batch
# loop over timesteps
pattern = pattern.unsqueeze(0).unsqueeze(0)
previous_importance = None
previous_output = None
reconstructions = []
for j in range(T):
importanceBaselineLuminance.setTime(j)
# extract flow (always the last two channels of crop_high)
flow = crop_high[:, j, C - 2:, :, :]
# compute importance map
importance_input = crop_low[:, j, :5, :, :]
if j == 0 or s.importance.disableTemporal:
previous_input = torch.zeros(
B,
1,
importance_input.shape[2] *
importanceNetUpscale,
importance_input.shape[3] *
importanceNetUpscale,
dtype=crop_high.dtype,
device=crop_high.device)
else:
flow_low = F.interpolate(
flow,
scale_factor=1 / importancePostUpscale)
previous_input = models.VideoTools.warp_upscale(
previous_importance, flow_low, 1,
False)
importance_map = s.importance.call(
importance_input, previous_input)
if len(importance_map.shape) == 3:
importance_map = importance_map.unsqueeze(
1)
previous_importance = importance_map
target_mean = s.heatmean
if USE_BINARY_SEARCH_ON_MEAN:
# For regular sampling, the normalization does not work properly,
# use binary search on the heatmean instead
def f(x):
postprocess = importance.PostProcess(
s.heatmin, x,
importancePostUpscale,
LOSS_BORDER //
importancePostUpscale, 'basic')
importance_map2 = postprocess(
importance_map)[0].unsqueeze(1)
sampling_mask = (
importance_map2 >= pattern).to(
dtype=importance_map.dtype)
samples = torch.mean(
sampling_mask).item()
return samples
target_mean = binarySearch(
f, s.heatmean, s.heatmean, 10, 0, 1)
#print("Binary search for #samples, mean start={}, result={} with samples={}, original={}".
# format(s.heatmean, s.heatmean, f(target_mean), f(s.heatmean)))
# normalize and upscale importance map
postprocess = importance.PostProcess(
s.heatmin, target_mean,
importancePostUpscale,
LOSS_BORDER // importancePostUpscale,
'basic')
importance_map = postprocess(
importance_map)[0].unsqueeze(1)
#print("mean:", torch.mean(importance_map).item())
# create samples
sample_mask = (importance_map >= pattern).to(
dtype=importance_map.dtype)
reconstruction_input = torch.cat(
(
sample_mask *
crop_high[:, j, 0:
5, :, :], # mask, normal x, normal y, normal z, depth
sample_mask * torch.ones(
B,
1,
H,
W,
dtype=crop_high.dtype,
device=crop_high.device), # ao
sample_mask), # sample mask
dim=1)
# warp previous output
if j == 0 or s.recon.disableTemporal:
previous_input = torch.zeros(
B,
6,
H,
W,
dtype=crop_high.dtype,
device=crop_high.device)
else:
previous_input = models.VideoTools.warp_upscale(
previous_output, flow, 1, False)
# run reconstruction network
reconstruction = s.recon.call(
reconstruction_input, sample_mask,
previous_input)
# clamp
reconstruction_clamped = torch.cat(
[
torch.clamp(
reconstruction[:, 0:1, :, :], -1,
+1), # mask
ScreenSpaceShading.normalize(
reconstruction[:, 1:4, :, :],
dim=1),
torch.clamp(
reconstruction[:, 4:5, :, :], 0,
+1), # depth
torch.clamp(reconstruction[:,
5:6, :, :],
0, +1) # ao
],
dim=1)
reconstructions.append(reconstruction_clamped)
# save for next frame
previous_output = reconstruction_clamped
#endfor: timesteps
# compute statistics
reconstructions = torch.cat(reconstructions, dim=0)
crops_high = torch.cat(
[crop_high[:, j, :6, :, :] for j in range(T)],
dim=0)
sample_masks = torch.cat([sample_mask] * T, dim=0)
s.stats.add_timestep_sample(
reconstructions, crops_high, sample_masks)
# endfor: statistic
# endfor: batch
pg.print_progress_bar(num_minibatch)
# end no_grad()
finally:
# close files
for s in statistics:
s.stats.write_histogram()
s.stats.close_stats_file()
def run():
torch.ops.load_library("./Renderer.dll")
#########################
# CONFIGURATION
#########################
if 1:
OUTPUT_FOLDER = "../result-stats/adaptiveDvr3/"
DATASET_PREFIX = "D:/VolumeSuperResolution-InputData/"
DATASETS = [
("Ejecta", "gt-dvr-ejecta6-test.hdf5",
"gt-dvr-ejecta6-test-screen8x.hdf5"),
("RM", "gt-dvr-rm1-test.hdf5", "gt-dvr-rm1-test-screen8x.hdf5"),
("Thorax", "gt-dvr-thorax2-test.hdf5",
"gt-dvr-thorax2-test-screen8x.hdf5"),
]
NETWORK_DIR = "D:/VolumeSuperResolution/adaptive-dvr-modeldir/"
NETWORKS = [ #suffixed with _importance.pt and _recon.pt
#title, file prefix
#("v5-temp001", "adapDvr5-rgb-temp001-epoch300"),
#("v5-temp010", "adapDvr5-rgb-temp010-epoch300"),
#("v5-temp100", "adapDvr5-rgb-temp100-epoch300"),
#("v5-temp001-perc", "adapDvr5-rgb-temp001-perc01-epoch300"),
("v5-perc01+bn", "adapDvr5-rgb-perc01-bn-epoch500"),
("v5-perc01-bn", "adapDvr5-rgb-temp001-perc01-epoch500")
]
# Test if it is better to post-train with dense networks and PDE inpainting
POSTTRAIN_NETWORK_DIR = "D:/VolumeSuperResolution/adaptive-dvr-modeldir/"
POSTTRAIN_NETWORKS = [
# title, file suffix to POSTTRAIN_NETWORK_DIR, inpainting {'fast', 'pde'}
#("Enhance PDE (post)", "inpHv2-pde05-epoch200.pt", "pde")
("v6pr2-noTemp",
"ejecta6pr2-plastic05-lpips-noTempCon-epoch500_recon.pt", "fast",
False),
("v6pr2-tl2-100",
"ejecta6pr2-plastic05-lpips-tl2-100-epoch500_recon.pt", "fast",
True)
]
SAMPLING_FILE = "D:/VolumeSuperResolution-InputData/samplingPattern.hdf5"
SAMPLING_PATTERNS = ['plastic']
HEATMAP_MIN = [0.002]
HEATMAP_MEAN = [
0.02, 0.05, 0.1, 0.2
] #[0.01, 0.02, 0.03, 0.04, 0.06, 0.08, 0.1, 0.2, 0.3, 0.5, 0.8, 1.0]
USE_BINARY_SEARCH_ON_MEAN = True
UPSCALING = 8 # = networkUp * postUp
IMPORTANCE_BORDER = 8
LOSS_BORDER = 32
BATCH_SIZE = 1 #2
#########################
# LOADING
#########################
device = torch.device("cuda")
# Load Networks
IMPORTANCE_BASELINE1 = "ibase1"
IMPORTANCE_BASELINE2 = "ibase2"
RECON_BASELINE = "rbase"
# load importance model
print("load importance networks")
class ImportanceModel:
def __init__(self, file):
if file == IMPORTANCE_BASELINE1:
self._net = importance.UniformImportanceMap(1, 0.5)
self._upscaling = 1
self._name = "constant"
self.disableTemporal = True
self._requiresPrevious = False
elif file == IMPORTANCE_BASELINE2:
self._net = importance.GradientImportanceMap(
1, (0, 1), (1, 1), (2, 1))
self._upscaling = 1
self._name = "curvature"
self.disableTemporal = True
self._requiresPrevious = False
else:
self._name = file[0]
file = os.path.join(NETWORK_DIR, file[1] + "_importance.pt")
extra_files = torch._C.ExtraFilesMap()
extra_files['settings.json'] = ""
self._net = torch.jit.load(file,
map_location=device,
_extra_files=extra_files)
settings = json.loads(extra_files['settings.json'])
self._upscaling = settings['networkUpscale']
self._requiresPrevious = settings.get("requiresPrevious",
False)
self.disableTemporal = settings.get("disableTemporal", True)
def networkUpscaling(self):
return self._upscaling
def name(self):
return self._name
def __repr__(self):
return self.name()
def call(self, input, prev_warped_out):
if self._requiresPrevious:
input = torch.cat([
input,
models.VideoTools.flatten_high(prev_warped_out,
self._upscaling)
],
dim=1)
input = F.pad(input, [IMPORTANCE_BORDER] * 4, 'constant', 0)
output = self._net(input) # the network call
output = F.pad(output, [-IMPORTANCE_BORDER * self._upscaling] * 4,
'constant', 0)
return output
importanceBaseline1 = ImportanceModel(IMPORTANCE_BASELINE1)
importanceBaseline2 = ImportanceModel(IMPORTANCE_BASELINE2)
importanceModels = [ImportanceModel(f) for f in NETWORKS]
# load reconstruction networks
print("load reconstruction networks")
class ReconstructionModel:
def __init__(self, file):
if file == RECON_BASELINE:
class Inpainting(nn.Module):
def forward(self, x, mask):
input = x[:, 0:4, :, :].contiguous(
) # rgba, don't use normal xyz, depth
mask = x[:, 8, :, :].contiguous()
return torch.ops.renderer.fast_inpaint(mask, input)
self._net = Inpainting()
self._upscaling = 1
self._name = "inpainting"
self.disableTemporal = True
else:
self._name = file[0]
file = os.path.join(NETWORK_DIR, file[1] + "_recon.pt")
extra_files = torch._C.ExtraFilesMap()
extra_files['settings.json'] = ""
self._net = torch.jit.load(file,
map_location=device,
_extra_files=extra_files)
self._settings = json.loads(extra_files['settings.json'])
self.disableTemporal = False
requiresMask = self._settings.get('expectMask', False)
if self._settings.get("interpolateInput", False):
self._originalNet = self._net
class Inpainting2(nn.Module):
def __init__(self, orignalNet, requiresMask):
super().__init__()
self._n = orignalNet
self._requiresMask = requiresMask
def forward(self, x, mask):
input = x[:, 0:8, :, :].contiguous(
) # rgba, normal xyz, depth
mask = x[:, 8, :, :].contiguous()
inpainted = torch.ops.renderer.fast_inpaint(
mask, input)
x[:, 0:8, :, :] = inpainted
if self._requiresMask:
return self._n(x, mask)
else:
return self._n(x)
self._net = Inpainting2(self._originalNet, requiresMask)
def name(self):
return self._name
def __repr__(self):
return self.name()
def call(self, input, mask, prev_warped_out):
input = torch.cat([input, prev_warped_out], dim=1)
output = self._net(input, mask)
return output
class ReconstructionModelPostTrain:
"""
Reconstruction model that are trained as dense reconstruction networks
after the adaptive training.
They don't recive the sampling mask as input, but can start with PDE-based inpainting
"""
def __init__(self, name: str, model_path: str, inpainting: str,
has_temporal: bool):
assert inpainting == 'fast' or inpainting == 'pde', "inpainting must be either 'fast' or 'pde', but got %s" % inpainting
self._inpainting = inpainting
self._name = name
file = os.path.join(POSTTRAIN_NETWORK_DIR, model_path)
extra_files = torch._C.ExtraFilesMap()
extra_files['settings.json'] = ""
self._net = torch.jit.load(file,
map_location=device,
_extra_files=extra_files)
self._settings = json.loads(extra_files['settings.json'])
assert self._settings.get(
'upscale_factor', None) == 1, "selected file is not a 1x SRNet"
self.disableTemporal = not has_temporal
def name(self):
return self._name
def __repr__(self):
return self.name()
def call(self, input, mask, prev_warped_out):
# no sampling and no AO
input_no_sampling = input[:, 0:8, :, :].contiguous(
) # rgba, normal xyz, depth
sampling_mask = mask[:, 0, :, :].contiguous()
# perform inpainting
if self._inpainting == 'pde':
inpainted = torch.ops.renderer.pde_inpaint(
sampling_mask,
input_no_sampling,
200,
1e-4,
5,
2, # m0, epsilon, m1, m2
0, # mc -> multigrid recursion count. =0 disables the multigrid hierarchy
9,
0) # ms, m3
else:
inpainted = torch.ops.renderer.fast_inpaint(
sampling_mask, input_no_sampling)
# run network
if self.disableTemporal:
prev_warped_out = torch.zeros_like(prev_warped_out)
input = torch.cat([inpainted, prev_warped_out], dim=1)
output = self._net(input)
if isinstance(output, tuple):
output = output[0]
return output
reconBaseline = ReconstructionModel(RECON_BASELINE)
reconModels = [ReconstructionModel(f) for f in NETWORKS]
reconPostModels = [
ReconstructionModelPostTrain(name, file, inpainting, has_temporal)
for (name, file, inpainting, has_temporal) in POSTTRAIN_NETWORKS
]
allReconModels = reconModels + reconPostModels
NETWORK_COMBINATIONS = \
[(importanceBaseline1, reconBaseline), (importanceBaseline2, reconBaseline)] + \
[(importanceBaseline1, reconNet) for reconNet in allReconModels] + \
[(importanceBaseline2, reconNet) for reconNet in allReconModels] + \
[(importanceNet, reconBaseline) for importanceNet in importanceModels] + \
list(zip(importanceModels, reconModels)) + \
[(importanceNet, reconPostModel) for importanceNet in importanceModels for reconPostModel in reconPostModels]
#NETWORK_COMBINATIONS = list(zip(importanceModels, reconModels))
print("Network combinations:")
for (i, r) in NETWORK_COMBINATIONS:
print(" %s - %s" % (i.name(), r.name()))
# load sampling patterns
print("load sampling patterns")
with h5py.File(SAMPLING_FILE, 'r') as f:
sampling_pattern = dict([(name, torch.from_numpy(f[name][...]).to(device)) \
for name in SAMPLING_PATTERNS])
#heatmap
HEATMAP_CFG = [(min, mean) for min in HEATMAP_MIN for mean in HEATMAP_MEAN
if min < mean]
print("heatmap configs:", HEATMAP_CFG)
#########################
# DEFINE STATISTICS
#########################
ssimLoss = SSIM(size_average=False)
ssimLoss.to(device)
psnrLoss = PSNR()
psnrLoss.to(device)
lpipsColor = lpips.PerceptualLoss(model='net-lin',
net='alex',
use_gpu=True)
MIN_FILLING = 0.05
NUM_BINS = 200
class Statistics:
def __init__(self):
self.histogram_color = np.zeros(NUM_BINS, dtype=np.float64)
self.histogram_alpha = np.zeros(NUM_BINS, dtype=np.float64)
self.histogram_counter = 0
def create_datasets(self, hdf5_file: h5py.File, stats_name: str,
histo_name: str, num_samples: int,
extra_info: dict):
self.stats_name = stats_name
self.expected_num_samples = num_samples
stats_shape = (num_samples, len(list(StatFieldDvr)))
self.stats_file = hdf5_file.require_dataset(stats_name,
stats_shape,
dtype='f',
exact=True)
self.stats_file.attrs['NumFields'] = len(list(StatFieldDvr))
for field in list(StatFieldDvr):
self.stats_file.attrs['Field%d' % field.value] = field.name
for key, value in extra_info.items():
self.stats_file.attrs[key] = value
self.stats_index = 0
histo_shape = (NUM_BINS, len(list(HistoFieldDvr)))
self.histo_file = hdf5_file.require_dataset(histo_name,
histo_shape,
dtype='f',
exact=True)
self.histo_file.attrs['NumFields'] = len(list(HistoFieldDvr))
for field in list(HistoFieldDvr):
self.histo_file.attrs['Field%d' % field.value] = field.name
for key, value in extra_info.items():
self.histo_file.attrs[key] = value
def add_timestep_sample(self, pred_rgba, gt_rgba, sampling_mask):
"""
adds a timestep sample:
pred_rgba: prediction rgba
gt_rgba: ground truth rgba
"""
B = pred_rgba.shape[0]
#apply border
pred_rgba = pred_rgba[:, :, LOSS_BORDER:-LOSS_BORDER,
LOSS_BORDER:-LOSS_BORDER]
gt_rgba = gt_rgba[:, :, LOSS_BORDER:-LOSS_BORDER,
LOSS_BORDER:-LOSS_BORDER]
# PSNR
psnr_color = psnrLoss(pred_rgba[:, 0:3, :, :],
gt_rgba[:, 0:3, :, :]).cpu().numpy()
psnr_alpha = psnrLoss(pred_rgba[:, 3:4, :, :],
gt_rgba[:, 3:4, :, :]).cpu().numpy()
# SSIM
ssim_color = ssimLoss(pred_rgba[:, 0:3, :, :],
gt_rgba[:, 0:3, :, :]).cpu().numpy()
ssim_alpha = ssimLoss(pred_rgba[:, 3:4, :, :],
gt_rgba[:, 3:4, :, :]).cpu().numpy()
# Perceptual
lpips_color = torch.cat([ \
lpipsColor(pred_rgba[b, 0:3, :, :], gt_rgba[b, 0:3, :, :], normalize=True) \
for b in range(B)], dim=0).cpu().numpy()
# Samples
samples = torch.mean(sampling_mask, dim=(1, 2, 3)).cpu().numpy()
# Write samples to file
for b in range(B):
assert self.stats_index < self.expected_num_samples, "Adding more samples than specified"
self.stats_file[self.stats_index, :] = np.array([
psnr_color[b], psnr_alpha[b], ssim_color[b], ssim_alpha[b],
lpips_color[b], samples[b]
],
dtype='f')
self.stats_index += 1
# Histogram
self.histogram_counter += 1
color_diff = F.l1_loss(gt_rgba[:, 0:3, :, :],
pred_rgba[:, 0:3, :, :],
reduction='none').sum(dim=0) / 6
histogram, _ = np.histogram(color_diff.cpu().numpy(),
bins=NUM_BINS,
range=(0, 1),
density=True)
self.histogram_color += (
histogram /
(NUM_BINS * B) - self.histogram_color) / self.histogram_counter
alpha_diff = F.l1_loss(gt_rgba[:, 3, :, :],
pred_rgba[:, 3, :, :],
reduction='none')
histogram, _ = np.histogram(alpha_diff.cpu().numpy(),
bins=NUM_BINS,
range=(0, 1),
density=True)
self.histogram_alpha += (
histogram /
(NUM_BINS * B) - self.histogram_alpha) / self.histogram_counter
def close_stats_file(self):
self.stats_file.attrs['NumEntries'] = self.stats_index
def write_histogram(self):
"""
After every sample for the current dataset was processed, write
a histogram of the errors in a new file
"""
for i in range(NUM_BINS):
self.histo_file[i, :] = np.array([
i / NUM_BINS, (i + 1) / NUM_BINS, self.histogram_color[i],
self.histogram_alpha[i]
])
#########################
# DATASET
#########################
class FullResDataset(torch.utils.data.Dataset):
def __init__(self, file_high, file_low):
self.hdf5_file_high = h5py.File(file_high, 'r')
self.dset_high = self.hdf5_file_high['gt']
self.hdf5_file_low = h5py.File(file_low, 'r')
self.dset_low = self.hdf5_file_low['gt']
print("Dataset shape:", self.dset_high.shape)
def __len__(self):
return self.dset_high.shape[0]
def num_timesteps(self):
return self.dset_high.shape[1]
def __getitem__(self, idx):
return self.dset_high[idx, ...], self.dset_low[idx, ...]
#########################
# COMPUTE STATS for each dataset
#########################
for dataset_name, dataset_file_high, dataset_file_low in DATASETS:
dataset_file_high = os.path.join(DATASET_PREFIX, dataset_file_high)
dataset_file_low = os.path.join(DATASET_PREFIX, dataset_file_low)
print("Compute statistics for", dataset_name)
# create output file
os.makedirs(OUTPUT_FOLDER, exist_ok=True)
output_file = os.path.join(OUTPUT_FOLDER, dataset_name + '.hdf5')
print("Save to", output_file)
with h5py.File(output_file, 'a') as output_hdf5_file:
# load dataset
set = FullResDataset(dataset_file_high, dataset_file_low)
data_loader = torch.utils.data.DataLoader(set,
batch_size=BATCH_SIZE,
shuffle=False)
# define statistics
StatsCfg = collections.namedtuple(
"StatsCfg", "stats importance recon heatmin heatmean pattern")
statistics = []
for (inet, rnet) in NETWORK_COMBINATIONS:
for (heatmin, heatmean) in HEATMAP_CFG:
for pattern in SAMPLING_PATTERNS:
stats_info = {
'importance': inet.name(),
'reconstruction': rnet.name(),
'heatmin': heatmin,
'heatmean': heatmean,
'pattern': pattern
}
stats_filename = "Stats_%s_%s_%03d_%03d_%s" % (
inet.name(), rnet.name(), heatmin * 1000,
heatmean * 1000, pattern)
histo_filename = "Histogram_%s_%s_%03d_%03d_%s" % (
inet.name(), rnet.name(), heatmin * 1000,
heatmean * 1000, pattern)
s = Statistics()
s.create_datasets(output_hdf5_file, stats_filename,
histo_filename,
len(set) * set.num_timesteps(),
stats_info)
statistics.append(
StatsCfg(stats=s,
importance=inet,
recon=rnet,
heatmin=heatmin,
heatmean=heatmean,
pattern=pattern))
print(len(statistics),
" different combinations are performed per sample")
# compute statistics
try:
with torch.no_grad():
num_minibatch = len(data_loader)
pg = ProgressBar(num_minibatch, 'Evaluation', length=50)
for iteration, (crop_high,
crop_low) in enumerate(data_loader, 0):
pg.print_progress_bar(iteration)
crop_high = crop_high.to(device)
crop_low = crop_low.to(device)
B, T, C, H, W = crop_high.shape
_, _, _, Hlow, Wlow = crop_low.shape
assert Hlow * UPSCALING == H
# try out each combination
for s in statistics:
#print(s)
# get input to evaluation
importanceNetUpscale = s.importance.networkUpscaling(
)
importancePostUpscale = UPSCALING // importanceNetUpscale
pattern = sampling_pattern[s.pattern][:H, :W]
# loop over timesteps
pattern = pattern.unsqueeze(0).unsqueeze(0)
previous_importance = None
previous_output = None
reconstructions = []
for j in range(T):
# extract flow (always the last two channels of crop_high)
flow = crop_high[:, j, C - 2:, :, :]
# compute importance map
importance_input = crop_low[:, j, :8, :, :]
if j == 0 or s.importance.disableTemporal:
previous_input = torch.zeros(
B,
1,
importance_input.shape[2] *
importanceNetUpscale,
importance_input.shape[3] *
importanceNetUpscale,
dtype=crop_high.dtype,
device=crop_high.device)
else:
flow_low = F.interpolate(
flow,
scale_factor=1 / importancePostUpscale)
previous_input = models.VideoTools.warp_upscale(
previous_importance, flow_low, 1,
False)
importance_map = s.importance.call(
importance_input, previous_input)
if len(importance_map.shape) == 3:
importance_map = importance_map.unsqueeze(
1)
previous_importance = importance_map
target_mean = s.heatmean
if USE_BINARY_SEARCH_ON_MEAN:
# For regular sampling, the normalization does not work properly,
# use binary search on the heatmean instead
def f(x):
postprocess = importance.PostProcess(
s.heatmin, x,
importancePostUpscale,
LOSS_BORDER //
importancePostUpscale, 'basic')
importance_map2 = postprocess(
importance_map)[0].unsqueeze(1)
sampling_mask = (
importance_map2 >= pattern).to(
dtype=importance_map.dtype)
samples = torch.mean(
sampling_mask).item()
return samples
target_mean = binarySearch(
f, s.heatmean, s.heatmean, 10, 0, 1)
#print("Binary search for #samples, mean start={}, result={} with samples={}, original={}".
# format(s.heatmean, s.heatmean, f(target_mean), f(s.heatmean)))
# normalize and upscale importance map
postprocess = importance.PostProcess(
s.heatmin, target_mean,
importancePostUpscale,
LOSS_BORDER // importancePostUpscale,
'basic')
importance_map = postprocess(
importance_map)[0].unsqueeze(1)
#print("mean:", torch.mean(importance_map).item())
# create samples
sample_mask = (importance_map >= pattern).to(
dtype=importance_map.dtype)
reconstruction_input = torch.cat(
(
sample_mask *
crop_high[:, j, 0:
8, :, :], # rgba, normal xyz, depth
sample_mask), # sample mask
dim=1)
# warp previous output
if j == 0 or s.recon.disableTemporal:
previous_input = torch.zeros(
B,
4,
H,
W,
dtype=crop_high.dtype,
device=crop_high.device)
else:
previous_input = models.VideoTools.warp_upscale(
previous_output, flow, 1, False)
# run reconstruction network
reconstruction = s.recon.call(
reconstruction_input, sample_mask,
previous_input)
# clamp
reconstruction_clamped = torch.clamp(
reconstruction, 0, 1)
reconstructions.append(reconstruction_clamped)
## test
#if j==0:
# plt.figure()
# plt.imshow(reconstruction_clamped[0,0:3,:,:].cpu().numpy().transpose((1,2,0)))
# plt.title(s.stats.stats_name)
# plt.show()
# save for next frame
previous_output = reconstruction_clamped
#endfor: timesteps
# compute statistics
reconstructions = torch.cat(reconstructions, dim=0)
crops_high = torch.cat(
[crop_high[:, j, :8, :, :] for j in range(T)],
dim=0)
sample_masks = torch.cat([sample_mask] * T, dim=0)
s.stats.add_timestep_sample(
reconstructions, crops_high, sample_masks)
# endfor: statistic
# endfor: batch
pg.print_progress_bar(num_minibatch)
# end no_grad()
finally:
# close files
for s in statistics:
s.stats.write_histogram()
s.stats.close_stats_file()
def __init__(self, alpha=0):
self.alpha = 0.85
self.L2Loss = nn.L1Loss()
self.SSIMLoss = SSIM(nc=3)