def forward(self, sample_low, previous_warped_flattened):
inputs = sample_low[:, 0:self.input_channels, :, :]
resized_inputs = F.interpolate(
inputs,
size=[
inputs.shape[2] * self.upscale_factor,
inputs.shape[3] * self.upscale_factor
],
mode=self.upsample)
prediction = torch.cat([
resized_inputs,
torch.ones(resized_inputs.shape[0],
self.output_channels - self.input_channels,
resized_inputs.shape[2],
resized_inputs.shape[3],
dtype=resized_inputs.dtype,
device=resized_inputs.device)
],
dim=1)
prediction[:, 0:1, :, :] = torch.clamp(prediction[:, 0:1, :, :], -1,
+1) #mask
prediction[:, 1:4, :, :] = ScreenSpaceShading.normalize(
prediction[:, 1:4, :, :], dim=1) # normal
prediction[:, 4:6, :, :] = torch.clamp(prediction[:, 4:6, :, :], 0,
+1) # depth+ao
color = self.shading(prediction)
return color, prediction
def colorize_and_pad(tensor):
assert tensor.shape[1]==6
mask = tensor[:,0:1,:,:]
normal = ScreenSpaceShading.normalize(tensor[:,1:4,:,:], dim=1)
depth_ao = tensor[:,4:6,:,:]
ao = tensor[:,5:6,:,:]
color = self.shading(torch.cat([mask, normal, depth_ao], dim=1))
tensor_with_color = torch.cat([mask, normal, ao, color], dim=1)
return LossNetUnshaded.pad(tensor_with_color, self.padding)
def forward(self, sample_low, previous_warped_flattened):
single_input = torch.cat((sample_low, previous_warped_flattened),
dim=1)
prediction, _ = self.mdl(single_input)
prediction[:, 0:1, :, :] = torch.clamp(prediction[:, 0:1, :, :], -1,
+1) #mask
prediction[:, 1:4, :, :] = ScreenSpaceShading.normalize(
prediction[:, 1:4, :, :], dim=1) # normal
prediction[:, 4:6, :, :] = torch.clamp(prediction[:, 4:6, :, :], 0,
+1) # depth+ao
color = self.shading(prediction)
return color, prediction
pass # nothing to do
else: # network
if not m['temporal']: previous_image = None
imageInput = np.copy(image)
#image = cv.resize(image.transpose((2, 1, 0)),
# dsize=None,
# fx=UPSCALING,
# fy=UPSCALING,
# interpolation=cv.INTER_LINEAR).transpose((2, 1, 0))
if models[k].unshaded:
# unshaded input
imageRaw = models[k].inference(imageInput, previous_image)
imageRaw = torch.cat([
torch.clamp(imageRaw[:, 0:1, :, :], -1, +1),
ScreenSpaceShading.normalize(imageRaw[:, 1:4, :, :],
dim=1),
torch.clamp(imageRaw[:, 4:, :, :], 0, 1)
],
dim=1)
previous_image = imageRaw
image = image.transpose((2, 1, 0))
image = cv.resize(image,
dsize=None,
fx=UPSCALING,
fy=UPSCALING,
interpolation=cv.INTER_LINEAR)
image = image.transpose((2, 1, 0))
base_mask = np.copy(image[3, :, :])
imageRawCpu = imageRaw.cpu()
image[0:3, :, :] = shading(imageRawCpu)[0].numpy()
def add_timestep_sample(self, pred_mnda, gt_mnda, input_mnda):
"""
adds a timestep sample:
pred_mnda: prediction: mask, normal, depth, AO
pred_color: shaded color
gt_mnda: ground truth: mask, normal, depth, AO
gt_color: shaded ground truth
"""
#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)
input_color_noAO = shading(input_mnda)
#apply border
BORDER2 = BORDER * UPSCALING
pred_mnda = pred_mnda[:, :, BORDER2:-BORDER2, BORDER2:-BORDER2]
pred_color_withAO = pred_color_withAO[:, :, BORDER2:-BORDER2,
BORDER2:-BORDER2]
pred_color_noAO = pred_color_noAO[:, :, BORDER2:-BORDER2,
BORDER2:-BORDER2]
gt_mnda = gt_mnda[:, :, BORDER2:-BORDER2, BORDER2:-BORDER2]
gt_color_withAO = gt_color_withAO[:, :, BORDER2:-BORDER2,
BORDER2:-BORDER2]
gt_color_noAO = gt_color_noAO[:, :, BORDER2:-BORDER2, BORDER2:-BORDER2]
input_mnda = input_mnda[:, :, BORDER:-BORDER, BORDER:-BORDER]
input_color_noAO = input_color_noAO[:, :, BORDER:-BORDER,
BORDER:-BORDER]
#self.psnr_normal += 10 * math.log10(1 / torch.mean((pred_mnda[0,1:4,:,:]-gt_mnda[0,1:4,:,:])**2).item())
#self.psnr_ao += 10 * math.log10(1 / torch.mean((pred_mnda[0,5:6,:,:]-gt_mnda[0,5:6,:,:])**2).item())
#self.psnr_color += 10 * math.log10(1 / torch.mean((pred_color-gt_color)**2).item())
# check fill rate
mask = gt_mnda[:, 0:1, :, :] * 0.5 + 0.5
B, C, H, W = mask.shape
factor = torch.sum(mask).item() / (H * W)
if factor < MIN_FILLING:
#print("Ignore, too few filled points")
return
self.n += 1
# PSNR
self.psnr_normal += psnrLoss(pred_mnda[:, 1:4, :, :],
gt_mnda[:, 1:4, :, :],
mask=mask).item()
self.psnr_depth += psnrLoss(pred_mnda[:, 4:5, :, :],
gt_mnda[:, 4:5, :, :],
mask=mask).item()
self.psnr_ao += psnrLoss(pred_mnda[:, 5:6, :, :],
gt_mnda[:, 5:6, :, :],
mask=mask).item()
self.psnr_color_withAO += psnrLoss(pred_color_withAO,
gt_color_withAO,
mask=mask).item()
self.psnr_color_noAO += psnrLoss(pred_color_noAO,
gt_color_noAO,
mask=mask).item()
# SSIM
pred_mnda = gt_mnda + mask * (pred_mnda - gt_mnda)
self.ssim_normal += ssimLoss(pred_mnda[:, 1:4, :, :],
gt_mnda[:, 1:4, :, :]).item()
self.ssim_depth += ssimLoss(pred_mnda[:, 4:5, :, :],
gt_mnda[:, 4:5, :, :]).item()
self.ssim_ao += ssimLoss(pred_mnda[:, 5:6, :, :],
gt_mnda[:, 5:6, :, :]).item()
self.ssim_color_withAO += ssimLoss(pred_color_withAO,
gt_color_withAO).item()
self.ssim_color_noAO += ssimLoss(pred_color_noAO, gt_color_noAO).item()
# Downsample Loss
ds_normal = self.downsample_loss(
input_mnda[:, 1:4, :, :],
ScreenSpaceShading.normalize(self.downsample(pred_mnda[:,
1:4, :, :]),
dim=1))
ds_color = self.downsample_loss(input_color_noAO,
self.downsample(pred_color_noAO))
self.l2ds_normal_mean += torch.mean(ds_normal).item()
self.l2ds_normal_max = max(self.l2ds_normal_max,
torch.max(ds_normal).item())
self.l2ds_colorNoAO_mean += torch.mean(ds_color).item()
self.l2ds_colorNoAO_max = max(self.l2ds_colorNoAO_max,
torch.max(ds_color).item())
# Histogram
self.histogram_counter += 1
mask_diff = F.l1_loss(gt_mnda[0, 0, :, :],
pred_mnda[0, 0, :, :],
reduction='none')
histogram, _ = np.histogram(mask_diff.cpu().numpy(),
bins=NUM_BINS,
range=(0, 1),
density=True)
self.histogram_mask += (histogram / NUM_BINS -
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[0, 1:4, :, :],
pred_mnda[0, 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 - self.histogram_normal
) / self.histogram_counter
depth_diff = F.l1_loss(gt_mnda[0, 4, :, :],
pred_mnda[0, 4, :, :],
reduction='none')
histogram, _ = np.histogram(depth_diff.cpu().numpy(),
bins=NUM_BINS,
range=(0, 1),
density=True)
self.histogram_depth += (histogram / NUM_BINS -
self.histogram_depth) / self.histogram_counter
ao_diff = F.l1_loss(gt_mnda[0, 5, :, :],
pred_mnda[0, 5, :, :],
reduction='none')
histogram, _ = np.histogram(ao_diff.cpu().numpy(),
bins=NUM_BINS,
range=(0, 1),
density=True)
self.histogram_ao += (histogram / NUM_BINS -
self.histogram_ao) / self.histogram_counter
color_diff = F.l1_loss(gt_color_withAO[0, 0, :, :],
pred_color_withAO[0, 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 -
self.histogram_color_withAO) / self.histogram_counter
color_diff = F.l1_loss(gt_color_noAO[0, 0, :, :],
pred_color_noAO[0, 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 -
self.histogram_color_noAO) / self.histogram_counter
sample_flow[j - 1:j, :, :, :],
UPSCALING,
special_mask=True)
#previous_warped = previous_output
previous_warped_flattened = models.VideoTools.flatten_high(
previous_warped, UPSCALING)
single_input = torch.cat((sample_low[j:j + 1, :, :, :],
previous_warped_flattened),
dim=1)
# run model
prediction, _ = model_list[model_index](single_input)
prediction[:, 0:1, :, :] = torch.clamp(
prediction[:, 0:1, :, :], -1, +1) #mask
prediction[:,
1:4, :, :] = ScreenSpaceShading.normalize(
prediction[:,
1:4, :, :], dim=1) # normal
prediction[:, 4:6, :, :] = torch.clamp(
prediction[:, 4:6, :, :], 0, +1) # depth+ao
##DEBUG: save the images
#img_rgb = (prediction[:,1:4,:,:] * 0.5 + 0.5)[0].cpu().numpy()
##img_rgb = img_rgb.transpose((2,1,0))
#scipy.misc.toimage(img_rgb, cmin=0.0, cmax=1.0).save(
# os.path.join(OUTPUT_FOLDER, "Img_%s_%s_%d.png"%(dataset_name, MODELS[model_index]['name'], j)))
#if model_index==0:
# img_rgb = (sample_high[j,1:4,:,:] * 0.5 + 0.5).cpu().numpy()
# scipy.misc.toimage(img_rgb, cmin=0.0, cmax=1.0).save(
# os.path.join(OUTPUT_FOLDER, "Img_%s_GT_%d.png"%(dataset_name,j)))
# STATISTICS
def forward(self, gt, pred, input, prev_input_warped, prev_pred_warped):
"""
gt: ground truth high resolution image (B x C=output_channels x 4W x 4H)
pred: predicted high resolution image (B x C=output_channels x 4W x 4H)
input: upsampled low resolution input image (B x C=input_channels x 4W x 4H)
Only used for the discriminator, can be None if only the other losses are used
prev_input_warped: upsampled, warped previous input image
prev_pred_warped: predicted image from the previous frame warped by the flow
Shape: B x Cout x 4W x 4H
Only used for temporal losses, can be None if only the other losses are used
"""
B, Cout, Hhigh, Whigh = gt.shape
assert Cout == 6
assert gt.shape == pred.shape
# zero border padding
gt = LossNetUnshaded.pad(gt, self.padding)
pred = LossNetUnshaded.pad(pred, self.padding)
if prev_pred_warped is not None:
prev_pred_warped = LossNetUnshaded.pad(prev_pred_warped, self.padding)
# extract mask and normal
gt_mask = gt[:,0:1,:,:]
gt_mask_clamp = torch.clamp(gt_mask*0.5 + 0.5, 0, 1)
gt_normal = ScreenSpaceShading.normalize(gt[:,1:4,:,:], dim=1)
gt_depth = gt[:,4:5,:,:]
gt_ao = gt[:,5:6,:,:]
pred_mask = pred[:,0:1,:,:]
pred_mask_clamp = torch.clamp(pred_mask*0.5 + 0.5, 0, 1)
pred_normal = ScreenSpaceShading.normalize(pred[:,1:4,:,:], dim=1)
pred_depth = pred[:,4:5,:,:]
pred_ao = pred[:,5:6,:,:]
input_mask = input[:,0:1,:,:]
input_mask_clamp = torch.clamp(input_mask*0.5 + 0.5, 0, 1)
input_normal = ScreenSpaceShading.normalize(input[:,1:4,:,:], dim=1)
input_depth = input[:,4:5,:,:]
input_ao = None #not available
# compute color output
gt_color = self.shading(gt)
pred_color = self.shading(pred)
input_color = self.shading(input)
generator_loss = 0.0
loss_values = {}
# normal, simple losses, uses gt+pred
for name in ['mse','l1']:
if (name,'mask') in self.weight_dict.keys():
loss = self.loss_dict[name](gt_mask, pred_mask)
loss_values[(name,'mask')] = loss.item()
generator_loss += self.weight_dict[(name,'mask')] * loss
if (name,'normal') in self.weight_dict.keys():
loss = self.loss_dict[name](gt_normal*gt_mask_clamp, pred_normal*gt_mask_clamp)
loss_values[(name,'normal')] = loss.item()
generator_loss += self.weight_dict[(name,'normal')] * loss
if (name,'ao') in self.weight_dict.keys():
loss = self.loss_dict[name](gt_ao*gt_mask_clamp, pred_ao*gt_mask_clamp)
loss_values[(name,'ao')] = loss.item()
generator_loss += self.weight_dict[(name,'ao')] * loss
if (name,'depth') in self.weight_dict.keys():
loss = self.loss_dict[name](gt_depth*gt_mask_clamp, pred_depth*gt_mask_clamp)
loss_values[(name,'depth')] = loss.item()
generator_loss += self.weight_dict[(name,'depth')] * loss
if (name,'color') in self.weight_dict.keys():
loss = self.loss_dict[name](gt_color, pred_color)
loss_values[(name,'color')] = loss.item()
generator_loss += self.weight_dict[(name,'color')] * loss
# downsample loss, use input+pred
# TODO: input is passed in upsampled version, but this introduces more errors
# Better: input low-res input and use the 'low_res_gt=True' option in downsample_loss
# This requires the input to be upsampled here for the GAN.
for name in ['l2-ds', 'l1-ds']:
if (name,'mask') in self.weight_dict.keys():
loss = self.loss_dict[name](input_mask, pred_mask)
loss_values[(name,'mask')] = loss.item()
generator_loss += self.weight_dict[(name,'mask')] * loss
if (name,'normal') in self.weight_dict.keys():
loss = self.loss_dict[name](input_normal*input_mask_clamp, pred_normal*input_mask_clamp)
loss_values[(name,'normal')] = loss.item()
generator_loss += self.weight_dict[(name,'normal')] * loss
if (name,'ao') in self.weight_dict.keys():
loss = self.loss_dict[name](input_ao*input_mask_clamp, pred_ao*input_mask_clamp)
loss_values[(name,'ao')] = loss.item()
generator_loss += self.weight_dict[(name,'ao')] * loss
if (name,'depth') in self.weight_dict.keys():
loss = self.loss_dict[name](input_depth*input_mask_clamp, pred_depth*input_mask_clamp)
loss_values[(name,'depth')] = loss.item()
generator_loss += self.weight_dict[(name,'depth')] * loss
if (name,'color') in self.weight_dict.keys():
loss = self.loss_dict[name](input_color, pred_color)
loss_values[(name,'color')] = loss.item()
generator_loss += self.weight_dict[(name,'color')] * loss
# special losses: perceptual+texture, uses gt+pred
def compute_perceptual(target, in_pred, in_gt):
if ('perceptual',target) in self.weight_dict.keys() \
or ('texture',target) in self.weight_dict.keys():
style_weight=self.weight_dict.get(('texture',target), 0)
content_weight=self.weight_dict.get(('perceptual',target), 0)
style_score = 0
content_score = 0
input_images = torch.cat([in_gt, in_pred], dim=0)
self.pt_loss(input_images)
for sl in self.style_losses:
style_score += sl.loss
for cl in self.content_losses:
content_score += cl.loss
if ('perceptual',target) in self.weight_dict.keys():
loss_values[('perceptual',target)] = content_score.item()
if ('texture',target) in self.weight_dict.keys():
loss_values[('texture',target)] = style_score.item()
return style_weight * style_score + content_weight * content_score
return 0
generator_loss += compute_perceptual('mask', pred_mask.expand(-1,3,-1,-1)*0.5+0.5, gt_mask.expand(-1,3,-1,-1)*0.5+0.5)
generator_loss += compute_perceptual('normal', (pred_normal*gt_mask_clamp)*0.5+0.5, (gt_normal*gt_mask_clamp)*0.5+0.5)
generator_loss += compute_perceptual('color', pred_color, gt_color)
generator_loss += compute_perceptual('ao', pred_ao.expand(-1,3,-1,-1), gt_ao.expand(-1,3,-1,-1))
generator_loss += compute_perceptual('depth', pred_depth.expand(-1,3,-1,-1), gt_depth.expand(-1,3,-1,-1))
# special: discriminator, uses input+pred+prev_pred_warped
if self.has_discriminator:
pred_with_color = torch.cat([
pred_mask,
pred_normal,
pred_color,
pred_ao], dim=1)
prev_pred_normal = ScreenSpaceShading.normalize(prev_pred_warped[:,1:4,:,:], dim=1)
prev_pred_with_color = torch.cat([
prev_pred_warped[:,0:1,:,:],
prev_pred_normal,
self.shading(torch.cat([
prev_pred_warped[:,0:1,:,:],
prev_pred_normal,
prev_pred_warped[:,4:6,:,:]
], dim=1)),
prev_pred_warped[:,5:6,:,:]
], dim=1)
input_pad = LossNetUnshaded.pad(input, self.padding)
prev_input_warped_pad = LossNetUnshaded.pad(prev_input_warped, self.padding)
pred_with_color_pad = LossNetUnshaded.pad(pred_with_color, self.padding)
prev_pred_warped_pad = LossNetUnshaded.pad(prev_pred_with_color, self.padding)
if ('adv','all') in self.weight_dict.keys(): # spatial-temporal
discr_input = torch.cat([input_pad, prev_input_warped_pad, pred_with_color_pad, prev_pred_warped_pad], dim=1)
gen_loss = self.adv_loss(self.discriminator(discr_input))
loss_values['discr_pred'] = gen_loss.item()
generator_loss += self.weight_dict[('adv','all')] * gen_loss
if ('tgan','all') in self.weight_dict.keys(): #temporal
discr_input = torch.cat([pred_with_color_pad, prev_pred_warped_pad], dim=1)
gen_loss = self.temp_adv_loss(self.temp_discriminator(discr_input))
loss_values['temp_discr_pred'] = gen_loss.item()
generator_loss += self.weight_dict[('tgan','all')] * gen_loss
if ('sgan','all') in self.weight_dict.keys(): #spatial
discr_input = torch.cat([input_pad, pred_with_color_pad], dim=1)
gen_loss = self.spatial_adv_loss(self.spatial_discriminator(discr_input))
loss_values['spatial_discr_pred'] = gen_loss.item()
generator_loss += self.weight_dict[('sgan','all')] * gen_loss
# special: temporal l2 loss, uses input (for the mask) + pred + prev_warped
if self.has_temporal_l2_loss:
prev_pred_mask = prev_pred_warped[:,0:1,:,:]
prev_pred_normal = ScreenSpaceShading.normalize(prev_pred_warped[:,1:4,:,:], dim=1)
if ('temp-l2','mask') in self.weight_dict.keys():
loss = self.loss_dict['temp-l2'](pred_mask, prev_pred_mask)
loss_values[('temp-l2','mask')] = loss.item()
generator_loss += self.weight_dict[('temp-l2','mask')] * loss
if ('temp-l2','normal') in self.weight_dict.keys():
loss = self.loss_dict['temp-l2'](
pred_normal*gt_mask_clamp,
prev_pred_normal*gt_mask_clamp)
loss_values[('temp-l2','normal')] = loss.item()
generator_loss += self.weight_dict[('temp-l2','normal')] * loss
if ('temp-l2','ao') in self.weight_dict.keys():
prev_pred_ao = prev_pred_warped[:,5:6,:,:]
loss = self.loss_dict['temp-l2'](
pred_ao*gt_mask_clamp,
prev_pred_ao*gt_mask_clamp)
loss_values[('temp-l2','ao')] = loss.item()
generator_loss += self.weight_dict[('temp-l2','ao')] * loss
if ('temp-l2','depth') in self.weight_dict.keys():
prev_pred_depth = prev_pred_warped[:,4:5,:,:]
loss = self.loss_dict['temp-l2'](
pred_depth*gt_mask_clamp,
prev_pred_depth*gt_mask_clamp)
loss_values[('temp-l2','depth')] = loss.item()
generator_loss += self.weight_dict[('temp-l2','depth')] * loss
if ('temp-l2','color') in self.weight_dict.keys():
prev_pred_color = self.shading(prev_pred_warped)
loss = self.loss_dict['temp-l2'](pred_color, prev_pred_color)
loss_values[('temp-l2','color')] = loss.item()
generator_loss += self.weight_dict[('temp-l2','color')] * loss
return generator_loss, loss_values
def trainAdv_v2(epoch):
"""
Second version of adverserial training,
for each batch, train both discriminator and generator.
Not full epoch for each seperately
"""
print("===> Epoch %d Training" % epoch)
discr_scheduler.step()
writer.add_scalar('train/lr_discr', discr_scheduler.get_lr()[0], epoch)
gen_scheduler.step()
writer.add_scalar('train/lr_gen', gen_scheduler.get_lr()[0], epoch)
disc_steps = opt.advDiscrInitialSteps if opt.advDiscrInitialSteps is not None and epoch == 1 else opt.advDiscrMaxSteps
gen_steps = opt.advGenMaxSteps
num_minibatch = len(training_data_loader)
model.train()
criterion.discr_train()
total_discr_loss = 0
total_gen_loss = 0
total_gt_score = 0
total_pred_score = 0
pg = ProgressBar(num_minibatch, 'Train', length=50)
for iteration, batch in enumerate(training_data_loader):
pg.print_progress_bar(iteration)
input, flow, target = batch[0].to(device), batch[1].to(
device), batch[2].to(device)
B, _, Cout, Hhigh, Whigh = target.shape
_, _, Cin, H, W = input.shape
# DISCRIMINATOR
for _ in range(disc_steps):
discr_optimizer.zero_grad()
gen_optimizer.zero_grad()
loss = 0
#iterate over all timesteps
for j in range(dataset_data.num_frames):
# prepare input for the generator
if j == 0 or opt.disableTemporal:
previous_warped = initialImage(input[:, 0, :, :, :], Cout,
opt.initialImage, False,
opt.upscale_factor)
# loss takes the ground truth current image as warped previous image,
# to not introduce a bias and big loss for the first image
previous_warped_loss = target[:, 0, :, :, :]
previous_input = F.interpolate(input[:, 0, :, :, :],
size=(Hhigh, Whigh),
mode=opt.upsample)
else:
previous_warped = models.VideoTools.warp_upscale(
previous_output,
flow[:, j - 1, :, :, :],
opt.upscale_factor,
special_mask=True)
previous_warped_loss = previous_warped
previous_input = F.interpolate(input[:, j - 1, :, :, :],
size=(Hhigh, Whigh),
mode=opt.upsample)
previous_input = models.VideoTools.warp_upscale(
previous_input,
flow[:, j - 1, :, :, :],
opt.upscale_factor,
special_mask=True)
previous_warped_flattened = models.VideoTools.flatten_high(
previous_warped, opt.upscale_factor)
single_input = torch.cat(
(input[:, j, :, :, :], previous_warped_flattened), dim=1)
#evaluate generator
with torch.no_grad():
prediction, _ = model(single_input)
#prepare input for the discriminator
gt_prev_warped = models.VideoTools.warp_upscale(
target[:, j - 1, :, :, :],
flow[:, j - 1, :, :, :],
opt.upscale_factor,
special_mask=True)
#evaluate discriminator
input_high = F.interpolate(input[:, j, :, :, :],
size=(Hhigh, Whigh),
mode=opt.upsample)
disc_loss, gt_score, pred_score = criterion.train_discriminator(
input_high, target[:, j, :, :, :], previous_input,
gt_prev_warped, prediction, previous_warped_loss)
loss += disc_loss
total_gt_score += float(gt_score)
total_pred_score += float(pred_score)
# save output
previous_output = torch.cat(
[
torch.clamp(prediction[:, 0:1, :, :], -1, +1), # mask
ScreenSpaceShading.normalize(prediction[:, 1:4, :, :],
dim=1),
torch.clamp(prediction[:, 4:5, :, :], 0, +1), # depth
torch.clamp(prediction[:, 5:6, :, :], 0, +1) # ao
],
dim=1)
loss.backward()
discr_optimizer.step()
total_discr_loss += loss.item()
# GENERATOR
for _ in range(disc_steps):
discr_optimizer.zero_grad()
gen_optimizer.zero_grad()
loss = 0
#iterate over all timesteps
for j in range(dataset_data.num_frames):
# prepare input for the generator
if j == 0 or opt.disableTemporal:
previous_warped = initialImage(input[:, 0, :, :, :], Cout,
opt.initialImage, False,
opt.upscale_factor)
# loss takes the ground truth current image as warped previous image,
# to not introduce a bias and big loss for the first image
previous_warped_loss = target[:, 0, :, :, :]
previous_input = F.interpolate(input[:, 0, :, :, :],
size=(Hhigh, Whigh),
mode=opt.upsample)
else:
previous_warped = models.VideoTools.warp_upscale(
previous_output,
flow[:, j - 1, :, :, :],
opt.upscale_factor,
special_mask=True)
previous_warped_loss = previous_warped
previous_input = F.interpolate(input[:, j - 1, :, :, :],
size=(Hhigh, Whigh),
mode=opt.upsample)
previous_input = models.VideoTools.warp_upscale(
previous_input,
flow[:, j - 1, :, :, :],
opt.upscale_factor,
special_mask=True)
previous_warped_flattened = models.VideoTools.flatten_high(
previous_warped, opt.upscale_factor)
single_input = torch.cat(
(input[:, j, :, :, :], previous_warped_flattened), dim=1)
#evaluate generator
prediction, _ = model(single_input)
#evaluate loss
input_high = F.interpolate(input[:, j, :, :, :],
size=(Hhigh, Whigh),
mode=opt.upsample)
loss0, map = criterion(target[:, j, :, :, :], prediction,
input_high, previous_input,
previous_warped_loss)
loss += loss0
# save output
previous_output = torch.cat(
[
torch.clamp(prediction[:, 0:1, :, :], -1, +1), # mask
ScreenSpaceShading.normalize(prediction[:, 1:4, :, :],
dim=1),
torch.clamp(prediction[:, 4:5, :, :], 0, +1), # depth
torch.clamp(prediction[:, 5:6, :, :], 0, +1) # ao
],
dim=1)
loss.backward()
gen_optimizer.step()
total_gen_loss += loss.item()
pg.print_progress_bar(num_minibatch)
total_discr_loss /= num_minibatch * dataset_data.num_frames
total_gen_loss /= num_minibatch * dataset_data.num_frames
total_gt_score /= num_minibatch * dataset_data.num_frames
total_pred_score /= num_minibatch * dataset_data.num_frames
writer.add_scalar('train/discr_loss', total_discr_loss, epoch)
writer.add_scalar('train/gen_loss', total_gen_loss, epoch)
writer.add_scalar('train/gt_score', total_gt_score, epoch)
writer.add_scalar('train/pred_score', total_pred_score, epoch)
print("===> Epoch {} Complete".format(epoch))
def render(rerender=True, resuperres=True):
"""
Main render function
rerender: if True, the volumes are retraced. If false, the previous images are kept
resuperres: if True, the superresolution is performed again. If False, the previous result is used
"""
global oldFile, frameIndex, global_depth_max, global_depth_min, previous_frames
# check if file was changed
if oldFile != scene.file:
oldFile = scene.file
renderer.load(os.path.join(DATA_DIR_GPU, scene.file))
# send render parameters
currentOrigin = camera.getOrigin()
currentLookAt = camera.getLookAt()
currentUp = camera.getUp()
renderer.send_command(
"cameraOrigin", "%5.3f,%5.3f,%5.3f" %
(currentOrigin[0], currentOrigin[1], currentOrigin[2]))
renderer.send_command(
"cameraLookAt", "%5.3f,%5.3f,%5.3f" %
(currentLookAt[0], currentLookAt[1], currentLookAt[2]))
renderer.send_command(
"cameraUp", "%5.3f,%5.3f,%5.3f" %
(currentUp[0], currentUp[1], currentUp[2]))
renderer.send_command("cameraFoV", "%.3f" % shading.get_fov())
renderer.send_command("isovalue", "%5.3f" % float(scene.isovalue))
renderer.send_command("aoradius", "%5.3f" % float(scene.aoRadius))
if PREVIEW:
renderer.send_command("aosamples", "0")
else:
renderer.send_command("aosamples", "%d" % scene.aoSamples)
if rerender:
# render low resolution
renderer.send_command(
"resolution",
"%d,%d" % (RESOLUTION_LOW[0], RESOLUTION_LOW[1]))
renderer.send_command(
"viewport", "%d,%d,%d,%d" %
(0, 0, RESOLUTION_LOW[0], RESOLUTION_LOW[1]))
renderer.render_direct(rendered_low)
# render high resolution
if not PREVIEW:
renderer.send_command(
"resolution", "%d,%d" % (RESOLUTION[0], RESOLUTION[1]))
renderer.send_command(
"viewport",
"%d,%d,%d,%d" % (0, 0, RESOLUTION[0], RESOLUTION[1]))
renderer.render_direct(rendered_high)
# preprocessing
def preprocess(input):
input = input.to(
cpuDevice if CPU_SUPERRES else cudaDevice).permute(
2, 0, 1)
output = torch.unsqueeze(input, 0)
output = torch.cat(
(
output[:, 0:3, :, :],
output[:, 3:4, :, :] * 2 -
1, #transform mask into -1,+1
output[:, 4:, :, :]),
dim=1)
#image_shaded_input = torch.cat((output[:,3:4,:,:], output[:,4:8,:,:], output[:,10:11,:,:]), dim=1)
#image_shaded = torch.clamp(shading(image_shaded_input), 0, 1)
#output[:,0:3,:,:] = image_shaded
return output
processed_low = preprocess(rendered_low)
processed_high = preprocess(rendered_high)
# image now contains all channels:
# 0:3 - color (shaded)
# 3:4 - mask in -1,+1
# 4:7 - normal
# 7:8 - depth
# 8:10 - flow
# 10:11 - AO
# prepare bounds for depth
depthForBounds = processed_low[:, 7:8, :, :]
maxDepth = torch.max(depthForBounds)
minDepth = torch.min(
depthForBounds +
torch.le(depthForBounds, 1e-5).type_as(depthForBounds))
global_depth_max = max(global_depth_max, maxDepth.item())
global_depth_min = min(global_depth_min, minDepth.item())
if scene.depthMin is not None:
minDepth = scene.depthMin
if scene.depthMax is not None:
maxDepth = scene.depthMax
# mask
if PREVIEW:
base_mask = F.interpolate(processed_low,
scale_factor=UPSCALING,
mode='bilinear')[:, 3:4, :, :]
else:
base_mask = processed_high[:, 3:4, :, :]
base_mask = (base_mask * 0.5 + 0.5)
# loop through the models
for model_idx, model in enumerate(MODELS):
# perform super-resolution
if model['path'] == MODEL_NEAREST:
image = F.interpolate(processed_low,
scale_factor=UPSCALING,
mode='nearest')
elif model['path'] == MODEL_BILINEAR:
image = F.interpolate(processed_low,
scale_factor=UPSCALING,
mode='bilinear')
elif model['path'] == MODEL_BICUBIC:
image = F.interpolate(processed_low,
scale_factor=UPSCALING,
mode='bicubic')
elif model['path'] == MODEL_GROUND_TRUTH:
image = processed_high
else:
# NETWROK
if resuperres:
# previous frame
if scene.temporalConsistency:
previous_frame = previous_frames[model_idx]
else:
previous_frame = None
# apply network
imageRaw = models[model_idx].inference(
processed_low, previous_frame)
# post-process
imageRaw = torch.cat([
torch.clamp(imageRaw[:, 0:1, :, :], -1, +1),
ScreenSpaceShading.normalize(
imageRaw[:, 1:4, :, :], dim=1),
torch.clamp(imageRaw[:, 4:, :, :], 0, 1)
],
dim=1)
previous_frames[model_idx] = imageRaw
else:
imageRaw = previous_frames[model_idx]
image = F.interpolate(processed_low,
scale_factor=UPSCALING,
mode='bilinear')
#image[:,0:3,:,:] = shading(imageRaw)
image[:, 3:8, :, :] = imageRaw[:, 0:-1, :, :]
image[:, 10, :, :] = imageRaw[:, -1, :, :]
#masking
if model['masking']:
image[:, 3:4, :, :] = base_mask * 2 - 1
#image[:,7:8,:,:] = 0 + base_mask * (image[:,7:8,:,:] - 0)
image[:, 10:11, :, :] = 1 + base_mask * (
image[:, 10:11, :, :] - 1)
# shading
image_shaded_input = torch.cat(
(image[:, 3:4, :, :], image[:, 4:8, :, :],
image[:, 10:11, :, :]),
dim=1)
image_shaded_withAO = torch.clamp(shading(image_shaded_input),
0, 1)
ao = shading._ao
shading.ambient_occlusion(0.0)
image_shaded_noAO = torch.clamp(shading(image_shaded_input), 0,
1)
shading.ambient_occlusion(ao)
# perform channel selection
for channel_idx in range(len(CHANNEL_NAMES)):
if channel_idx == CHANNEL_AO:
if SHOW_DIFFERENCE and model[
'path'] != MODEL_GROUND_TRUTH:
image[:, 10:11, :, :] = 1 - torch.abs(
image[:, 10:11, :, :])
imageRGB = torch.cat(
(image[:, 10:11, :, :], image[:, 10:11, :, :],
image[:, 10:11, :, :]),
dim=1)
elif channel_idx == CHANNEL_COLOR_NOAO:
imageRGB = image_shaded_noAO
elif channel_idx == CHANNEL_COLOR_WITHAO:
imageRGB = image_shaded_withAO
elif channel_idx == CHANNEL_DEPTH:
if SHOW_DIFFERENCE and model[
'path'] != MODEL_GROUND_TRUTH:
depthVal = torch.abs(
image[:, 7:8, :, :] -
processed_high[:, 7:8, :, :]
) # / (2*(maxDepth - minDepth))
else:
depthVal = (image[:, 7:8, :, :] -
minDepth) / (maxDepth - minDepth)
imageRGB = torch.cat((depthVal, depthVal, depthVal),
dim=1)
imageRGB = 1 - imageRGB
imageRGB[imageRGB < 0.05] = 1.0
#imageRGB = BACKGROUND[0] + base_mask * (imageRGB - BACKGROUND[0])
elif channel_idx == CHANNEL_NORMAL:
if SHOW_DIFFERENCE and model[
'path'] != MODEL_GROUND_TRUTH:
diffVal = F.cosine_similarity(
image[:, 4:7, :, :],
processed_high[:, 4:7, :, :],
dim=1) * 0.5 + 0.5
imageRGB = torch.stack((diffVal, diffVal, diffVal),
dim=1)
#imageRGB = 1 - torch.abs(image[:,4:7,:,:])
else:
imageRGB = image[:, 4:7, :, :] * 0.5 + 0.5
imageRGB = BACKGROUND[0] + base_mask * (imageRGB -
BACKGROUND[0])
imageRGB = torch.clamp(imageRGB, 0, 1)
# copy to numpy and write to video
imageRGB_cpu = imageRGB.cpu().numpy()[0].transpose(
(1, 2, 0))
imageRGB_cpu = np.clip(imageRGB_cpu * 255, 0,
255).astype(np.uint8)
if OUTPUT_FIRST_IMAGE:
scipy.misc.imsave(writers[model_idx][channel_idx],
imageRGB_cpu)
else:
writers[model_idx][channel_idx].append_data(
imageRGB_cpu)
# done with this frame
frameIndex += 1
if frameIndex % 10 == 0:
print(" %d" % frameIndex)
if OUTPUT_FIRST_IMAGE:
raise BreakRenderingException()
def test_images(epoch):
def write_image(img, filename):
out_img = img.cpu().detach().numpy()
out_img *= 255.0
out_img = out_img.clip(0, 255)
out_img = np.uint8(out_img)
writer.add_image(filename, out_img, epoch)
with torch.no_grad():
num_minibatch = len(testing_full_data_loader)
pg = ProgressBar(num_minibatch,
'Test %d Images' % num_minibatch,
length=50)
model.eval()
if criterion.has_discriminator:
criterion.discr_eval()
for i, batch in enumerate(testing_full_data_loader):
pg.print_progress_bar(i)
input, flow, target = batch[0].to(device), batch[1].to(
device), batch[2].to(device)
B, _, Cin, H, W = input.shape
Hhigh = H * upscale_factor
Whigh = W * upscale_factor
Cout = output_channels
channel_mask = [1, 2, 3] #normal
previous_output = None
for j in range(dataset_data.num_frames):
# prepare input
if j == 0 or opt.disableTemporal:
previous_warped = initialImage(input[:, 0, :, :, :],
Cout, opt.initialImage,
False, upscale_factor)
else:
previous_warped = models.VideoTools.warp_upscale(
previous_output,
flow[:, j - 1, :, :, :],
upscale_factor,
special_mask=True)
# TODO: enable temporal component again
#previous_warped_flattened = models.VideoTools.flatten_high(previous_warped, opt.upscale_factor)
#single_input = torch.cat((
# input[:,j,:,:,:],
# previous_warped_flattened),
# dim=1)
single_input = input[:, j, :, :, :]
# run generator
heatMap = model(single_input)
heatMap = postprocess(heatMap)
prediction = importance.adaptiveSmoothing(
target[:, j, :, :, :].contiguous(),
1 / heatMap.unsqueeze(1),
opt.distanceToStandardDeviation)
# write heatmap
write_image(heatMap[0].unsqueeze(0),
'image%03d/frame%03d_heatmap' % (i, j))
## write warped previous frame
#write_image(previous_warped[0, channel_mask], 'image%03d/frame%03d_warped' % (i, j))
# write predicted normals
prediction[:, 1:4, :, :] = ScreenSpaceShading.normalize(
prediction[:, 1:4, :, :], dim=1)
write_image(prediction[0, channel_mask],
'image%03d/frame%03d_prediction' % (i, j))
# write shaded image if network runs in deferredShading mode
shaded_image = shading(prediction)
write_image(shaded_image[0],
'image%03d/frame%03d_shaded' % (i, j))
# write mask
write_image(prediction[0, 0:1, :, :] * 0.5 + 0.5,
'image%03d/frame%03d_mask' % (i, j))
# write ambient occlusion
write_image(prediction[0, 5:6, :, :],
'image%03d/frame%03d_ao' % (i, j))
# save output for next frame
previous_output = prediction
pg.print_progress_bar(num_minibatch)
print("Test images sent to Tensorboard for visualization")
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 test(epoch):
avg_psnr = 0
avg_losses = defaultdict(float)
heatmap_min = 1e10
heatmap_max = -1e10
heatmap_avg = heatmap_count = 0
with torch.no_grad():
num_minibatch = len(testing_data_loader)
pg = ProgressBar(num_minibatch, 'Testing', length=50)
model.eval()
if criterion.has_discriminator:
criterion.discr_eval()
for iteration, batch in enumerate(testing_data_loader, 0):
pg.print_progress_bar(iteration)
input, flow, target = batch[0].to(device), batch[1].to(
device), batch[2].to(device)
B, _, Cout, Hhigh, Whigh = target.shape
_, _, Cin, H, W = input.shape
previous_output = None
for j in range(dataset_data.num_frames):
# prepare input
if j == 0 or opt.disableTemporal:
previous_warped = initialImage(input[:, 0, :, :, :],
Cout, opt.initialImage,
False, upscale_factor)
# loss takes the ground truth current image as warped previous image,
# to not introduce a bias and big loss for the first image
previous_warped_loss = target[:, 0, :, :, :]
previous_input = F.interpolate(input[:, 0, :, :, :],
size=(Hhigh, Whigh),
mode='bilinear')
else:
previous_warped = models.VideoTools.warp_upscale(
previous_output,
flow[:, j - 1, :, :, :],
upscale_factor,
special_mask=True)
previous_warped_loss = previous_warped
previous_input = F.interpolate(input[:,
j - 1, :, :, :],
size=(Hhigh, Whigh),
mode='bilinear')
previous_input = models.VideoTools.warp_upscale(
previous_input,
flow[:, j - 1, :, :, :],
upscale_factor,
special_mask=True)
# TODO: enable temporal component again
#previous_warped_flattened = models.VideoTools.flatten_high(previous_warped, opt.upscale_factor)
#single_input = torch.cat((
# input[:,j,:,:,:],
# previous_warped_flattened),
# dim=1)
single_input = input[:, j, :, :, :]
# run generator
heatMap = model(single_input)
heatMapCrop = heatMap[:, opt.lossBorderPadding:-opt.
lossBorderPadding,
opt.lossBorderPadding:-opt.
lossBorderPadding]
heatmap_min = min(heatmap_min,
torch.min(heatMapCrop).item())
heatmap_max = max(heatmap_max,
torch.max(heatMapCrop).item())
heatmap_avg += torch.mean(heatMapCrop).item()
heatmap_count += 1
heatMap = postprocess(heatMap)
prediction = importance.adaptiveSmoothing(
target[:, j, :, :, :].contiguous(),
1 / heatMap.unsqueeze(1),
opt.distanceToStandardDeviation)
# evaluate cost
input_high = F.interpolate(input[:, j, :, :, :],
size=(Hhigh, Whigh),
mode='bilinear')
loss0, loss_values = criterion(target[:, j, :, :, :],
prediction, input_high,
previous_input,
previous_warped_loss)
avg_losses['total_loss'] += loss0.item()
psnr = 10 * log10(
1 / max(1e-10, loss_values[('mse', 'color')]))
avg_losses['psnr'] += psnr
for key, value in loss_values.items():
avg_losses[str(key)] += value
# save output for next frame
previous_output = torch.cat(
[
torch.clamp(prediction[:, 0:1, :, :], -1,
+1), # mask
ScreenSpaceShading.normalize(
prediction[:, 1:4, :, :], dim=1),
torch.clamp(prediction[:, 4:5, :, :], 0,
+1), # depth
torch.clamp(prediction[:, 5:6, :, :], 0, +1) # ao
],
dim=1)
pg.print_progress_bar(num_minibatch)
for key in avg_losses.keys():
avg_losses[key] /= num_minibatch * dataset_data.num_frames
print("===> Avg. PSNR: {:.4f} dB".format(avg_losses['psnr']))
print(" losses:", avg_losses)
for key, value in avg_losses.items():
writer.add_scalar('test/%s' % key, value, epoch)
print(" heatmap: min=%f, max=%f, avg=%f" %
(heatmap_min, heatmap_max, heatmap_avg / heatmap_count))
writer.add_scalar('test/heatmap_min', heatmap_min, epoch)
writer.add_scalar('test/heatmap_max', heatmap_max, epoch)
writer.add_scalar('test/heatmap_avg', heatmap_avg / heatmap_count,
epoch)
def test_images(epoch):
def write_image(img, filename):
out_img = img.cpu().detach().numpy()
out_img *= 255.0
out_img = out_img.clip(0, 255)
out_img = np.uint8(out_img)
writer.add_image(filename, out_img, epoch)
with torch.no_grad():
num_minibatch = len(testing_full_data_loader)
pg = ProgressBar(num_minibatch, 'Test %d Images'%num_minibatch, length=50)
model.eval()
if criterion.has_discriminator:
criterion.discr_eval()
for i,batch in enumerate(testing_full_data_loader):
pg.print_progress_bar(i)
input, flow = batch[0].to(device), batch[1].to(device)
B, _, Cin, H, W = input.shape
Hhigh = H * opt.upscale_factor
Whigh = W * opt.upscale_factor
Cout = output_channels
channel_mask = [1, 2, 3] #normal
previous_output = None
for j in range(dataset_data.num_frames):
# prepare input
if j == 0 or opt.disableTemporal:
previous_warped = initialImage(input[:,0,:,:,:], Cout,
opt.initialImage, False, opt.upscale_factor)
else:
previous_warped = models.VideoTools.warp_upscale(
previous_output,
flow[:, j-1, :, :, :],
opt.upscale_factor,
special_mask = True)
previous_warped_flattened = models.VideoTools.flatten_high(previous_warped, opt.upscale_factor)
single_input = torch.cat((
input[:,j,:,:,:],
previous_warped_flattened),
dim=1)
# write warped previous frame
write_image(previous_warped[0, channel_mask], 'image%03d/frame%03d_warped' % (i, j))
# run generator and cost
prediction, residual = model(single_input)
# normalize normal
prediction[:,1:4,:,:] = ScreenSpaceShading.normalize(prediction[:,1:4,:,:], dim=1)
# write prediction image
write_image(prediction[0, channel_mask], 'image%03d/frame%03d_prediction' % (i, j))
# write residual image
if residual is not None:
write_image(residual[0, channel_mask], 'image%03d/frame%03d_residual' % (i, j))
# write shaded image if network runs in deferredShading mode
shaded_image = shading(prediction)
write_image(shaded_image[0], 'image%03d/frame%03d_shaded' % (i, j))
# write mask
write_image(prediction[0, 0:1, :, :]*0.5+0.5, 'image%03d/frame%03d_mask' % (i, j))
# write ambient occlusion
# write mask
write_image(prediction[0, 5:6, :, :], 'image%03d/frame%03d_ao' % (i, j))
# save output for next frame
previous_output = torch.cat([
torch.clamp(prediction[:,0:1,:,:], -1, +1), # mask
prediction[:,1:4,:,:], #already normalized
torch.clamp(prediction[:,4:5,:,:], 0, +1), # depth
torch.clamp(prediction[:,5:6,:,:], 0, +1) # ao
], dim=1)
pg.print_progress_bar(num_minibatch)
print("Test images sent to Tensorboard for visualization")
def test(epoch):
avg_psnr = 0
avg_losses = defaultdict(float)
with torch.no_grad():
num_minibatch = len(testing_data_loader)
pg = ProgressBar(num_minibatch, 'Testing', length=50)
model.eval()
if criterion.has_discriminator:
criterion.discr_eval()
for iteration, batch in enumerate(testing_data_loader, 0):
pg.print_progress_bar(iteration)
input, flow, target = batch[0].to(device), batch[1].to(device), batch[2].to(device)
B, _, Cout, Hhigh, Whigh = target.shape
_, _, Cin, H, W = input.shape
previous_output = None
for j in range(dataset_data.num_frames):
# prepare input
if j == 0 or opt.disableTemporal:
previous_warped = initialImage(input[:,0,:,:,:], Cout,
opt.initialImage, False, opt.upscale_factor)
# loss takes the ground truth current image as warped previous image,
# to not introduce a bias and big loss for the first image
previous_warped_loss = target[:,0,:,:,:]
previous_input = F.interpolate(input[:,0,:,:,:], size=(Hhigh, Whigh), mode=opt.upsample)
else:
previous_warped = models.VideoTools.warp_upscale(
previous_output,
flow[:, j-1, :, :, :],
opt.upscale_factor,
special_mask = True)
previous_warped_loss = previous_warped
previous_input = F.interpolate(input[:,j-1,:,:,:], size=(Hhigh, Whigh), mode=opt.upsample)
previous_input = models.VideoTools.warp_upscale(
previous_input,
flow[:, j-1, :, :, :],
opt.upscale_factor,
special_mask = True)
previous_warped_flattened = models.VideoTools.flatten_high(previous_warped, opt.upscale_factor)
single_input = torch.cat((
input[:,j,:,:,:],
previous_warped_flattened),
dim=1)
# run generator
prediction, _ = model(single_input)
# evaluate cost
input_high = F.interpolate(input[:,j,:,:,:], size=(Hhigh, Whigh), mode=opt.upsample)
loss0, loss_values = criterion(
target[:,j,:,:,:],
prediction,
input_high,
previous_input,
previous_warped_loss)
avg_losses['total_loss'] += loss0.item()
psnr = 10 * log10(1 / max(1e-10, loss_values[('mse','color')]))
avg_losses['psnr'] += psnr
for key, value in loss_values.items():
avg_losses[str(key)] += value
## extra: evaluate discriminator on ground truth data
#if criterion.discriminator is not None:
# input_high = F.interpolate(input[:,j,:,:,:], size=(Hhigh, Whigh), mode=opt.upsample)
# if criterion.discriminator_use_previous_image:
# gt_prev_warped = models.VideoTools.warp_upscale(
# target[:,j-1,:,:,:],
# flow[:, j-1, :, :, :],
# opt.upscale_factor,
# special_mask = True)
# input_images = torch.cat([input_high, target[:,j,:,:,:], gt_prev_warped], dim=1)
# else:
# input_images = torch.cat([input_high, target[:,j,:,:,:]], dim=1)
# input_images = losses.LossNet.pad(input_images, criterion.padding)
# discr_gt = criterion.adv_loss(criterion.discriminator(input_images))
# avg_losses['discr_gt'] += discr_gt.item()
# save output for next frame
previous_output = torch.cat([
torch.clamp(prediction[:,0:1,:,:], -1, +1), # mask
ScreenSpaceShading.normalize(prediction[:,1:4,:,:], dim=1),
torch.clamp(prediction[:,4:5,:,:], 0, +1), # depth
torch.clamp(prediction[:,5:6,:,:], 0, +1) # ao
], dim=1)
pg.print_progress_bar(num_minibatch)
for key in avg_losses.keys():
avg_losses[key] /= num_minibatch * dataset_data.num_frames
print("===> Avg. PSNR: {:.4f} dB".format(avg_losses['psnr']))
print(" losses:",avg_losses)
for key, value in avg_losses.items():
writer.add_scalar('test/%s'%key, value, epoch)
def trainNormal(epoch):
epoch_loss = 0
scheduler.step()
num_minibatch = len(training_data_loader)
pg = ProgressBar(num_minibatch, 'Training', length=50)
model.train()
for iteration, batch in enumerate(training_data_loader, 0):
pg.print_progress_bar(iteration)
input, flow, target = batch[0].to(device), batch[1].to(device), batch[2].to(device)
B, _, Cout, Hhigh, Whigh = target.shape
_, _, Cin, H, W = input.shape
assert(Cout == output_channels)
assert(Cin == input_channels)
assert(H == dataset_data.crop_size)
assert(W == dataset_data.crop_size)
assert(Hhigh == dataset_data.crop_size * opt.upscale_factor)
assert(Whigh == dataset_data.crop_size * opt.upscale_factor)
optimizer.zero_grad()
previous_output = None
loss = 0
for j in range(dataset_data.num_frames):
# prepare input
if j == 0 or opt.disableTemporal:
previous_warped = initialImage(input[:,0,:,:,:], Cout,
opt.initialImage, False, opt.upscale_factor)
# loss takes the ground truth current image as warped previous image,
# to not introduce a bias and big loss for the first image
previous_warped_loss = target[:,0,:,:,:]
previous_input = F.interpolate(input[:,0,:,:,:], size=(Hhigh, Whigh), mode=opt.upsample)
else:
previous_warped = models.VideoTools.warp_upscale(
previous_output,
flow[:, j-1, :, :, :],
opt.upscale_factor,
special_mask = True)
previous_warped_loss = previous_warped
previous_input = F.interpolate(input[:,j-1,:,:,:], size=(Hhigh, Whigh), mode=opt.upsample)
previous_input = models.VideoTools.warp_upscale(
previous_input,
flow[:, j-1, :, :, :],
opt.upscale_factor,
special_mask = True)
previous_warped_flattened = models.VideoTools.flatten_high(previous_warped, opt.upscale_factor)
single_input = torch.cat((
input[:,j,:,:,:],
previous_warped_flattened),
dim=1)
# run generator
prediction, _ = model(single_input)
# evaluate cost
input_high = F.interpolate(input[:,j,:,:,:], size=(Hhigh, Whigh), mode=opt.upsample)
loss0,_ = criterion(
target[:,j,:,:,:],
prediction,
input_high,
previous_input,
previous_warped_loss)
del _
loss += loss0
epoch_loss += loss0.item()
# save output
previous_output = torch.cat([
torch.clamp(prediction[:,0:1,:,:], -1, +1), # mask
ScreenSpaceShading.normalize(prediction[:,1:4,:,:], dim=1),
torch.clamp(prediction[:,4:5,:,:], 0, +1), # depth
torch.clamp(prediction[:,5:6,:,:], 0, +1) # ao
], dim=1)
loss.backward()
optimizer.step()
pg.print_progress_bar(num_minibatch)
epoch_loss /= num_minibatch * dataset_data.num_frames
print("===> Epoch {} Complete: Avg. Loss: {:.4f}".format(epoch, epoch_loss))
writer.add_scalar('train/total_loss', epoch_loss, epoch)
writer.add_scalar('train/lr', scheduler.get_lr()[0], epoch)
def forward(self,
gt,
pred,
prev_pred_warped,
no_temporal_loss: bool,
use_checkpoints=False):
"""
gt: ground truth high resolution image (B x C x H x W)
pred: predicted high resolution image (B x C x H x W)
prev_pred_warped: predicted image from the previous frame warped by the flow
Shape: B x C x H x W
Only used for temporal losses, can be None if only the other losses are used
use_checkpoints: True if checkpointing should be used.
This does not apply here since we don't have GANs or perceptual losses
"""
# TODO: loss that penalizes deviation from the input samples
B, C, H, W = gt.shape
if self.has_flow:
assert C == 8
else:
assert C == 6
assert gt.shape == pred.shape
# zero border padding
gt = LossNetSparse.pad(gt, self.padding)
pred = LossNetSparse.pad(pred, self.padding)
if prev_pred_warped is not None:
prev_pred_warped = LossNetSparse.pad(prev_pred_warped,
self.padding)
# extract mask and normal
gt_mask = gt[:, 0:1, :, :]
gt_mask_clamp = torch.clamp(gt_mask * 0.5 + 0.5, 0, 1)
gt_normal = ScreenSpaceShading.normalize(gt[:, 1:4, :, :], dim=1)
gt_depth = gt[:, 4:5, :, :]
gt_ao = gt[:, 5:6, :, :]
if self.has_flow:
gt_flow = gt[:, 6:8, :, :]
pred_mask = pred[:, 0:1, :, :]
pred_mask_clamp = torch.clamp(pred_mask * 0.5 + 0.5, 0, 1)
pred_normal = ScreenSpaceShading.normalize(pred[:, 1:4, :, :], dim=1)
pred_depth = pred[:, 4:5, :, :]
pred_ao = pred[:, 5:6, :, :]
if self.has_flow:
pred_flow = pred[:, 6:8, :, :]
if prev_pred_warped is not None and self.has_temporal_l2_loss:
prev_pred_mask = prev_pred_warped[:, 0:1, :, :]
prev_pred_mask_clamp = torch.clamp(prev_pred_mask * 0.5 + 0.5, 0,
1)
prev_pred_normal = ScreenSpaceShading.normalize(
prev_pred_warped[:, 1:4, :, :], dim=1)
prev_pred_depth = prev_pred_warped[:, 4:5, :, :]
prev_pred_ao = prev_pred_warped[:, 5:6, :, :]
if self.has_flow:
prev_pred_flow = prev_pred_warped[:, 6:8, :, :]
generator_loss = 0.0
loss_values = {}
# normal, simple losses, uses gt+pred
for name in ['mse', 'l1', 'ssim', 'dssim', 'lpips']:
if (name, 'mask') in self.weight_dict.keys():
loss = self.loss_dict[name](gt_mask, pred_mask)
loss_values[(name, 'mask')] = loss.item()
generator_loss += self.weight_dict[(name, 'mask')] * loss
if (name, 'normal') in self.weight_dict.keys():
loss = self.loss_dict[name](gt_normal * gt_mask_clamp,
pred_normal * gt_mask_clamp)
if torch.isnan(loss).item():
test = self.loss_dict[name](gt_normal * gt_mask_clamp,
pred_normal * gt_mask_clamp)
loss_values[(name, 'normal')] = loss.item()
generator_loss += self.weight_dict[(name, 'normal')] * loss
if (name, 'ao') in self.weight_dict.keys():
loss = self.loss_dict[name](gt_ao * gt_mask_clamp,
pred_ao * gt_mask_clamp)
loss_values[(name, 'ao')] = loss.item()
generator_loss += self.weight_dict[(name, 'ao')] * loss
if (name, 'depth') in self.weight_dict.keys():
loss = self.loss_dict[name](gt_depth * gt_mask_clamp,
pred_depth * gt_mask_clamp)
loss_values[(name, 'depth')] = loss.item()
generator_loss += self.weight_dict[(name, 'depth')] * loss
if (name, 'flow') in self.weight_dict.keys():
# note: flow is not restricted to inside regions
loss = self.loss_dict[name](gt_flow, pred_flow)
loss_values[(name, 'flow')] = loss.item()
generator_loss += self.weight_dict[(name, 'flow')] * loss
if (name, 'color') in self.weight_dict.keys():
gt_color = self.shading(gt)
pred_color = self.shading(pred)
loss = self.loss_dict[name](gt_color, pred_color)
loss_values[(name, 'color')] = loss.item()
generator_loss += self.weight_dict[(name, 'color')] * loss
if ("bounded", "mask") in self.weight_dict.keys():
# penalizes if the mask diverges too far away from [0,1]
zero = torch.zeros(1,
1,
1,
1,
dtype=pred_mask.dtype,
device=pred_mask.device)
loss = torch.mean(torch.max(zero, pred_mask * pred_mask - 2))
loss_values[("bounded", "mask")] = loss.item()
generator_loss += self.weight_dict[("bounded", "mask")] * loss
if ("bce", "mask") in self.weight_dict.keys():
# binary cross entry loss between the unclamped masks
loss = F.binary_cross_entropy_with_logits(pred_mask * 0.5 + 0.5,
gt_mask * 0.5 + 0.5)
loss_values[("bce", "mask")] = loss.item()
generator_loss += self.weight_dict[("bce", "mask")] * loss
# temporal l2 loss, uses input (for the mask) + pred + prev_warped
if self.has_temporal_l2_loss and not no_temporal_loss:
assert prev_pred_warped is not None
for name in ['temp-l2', 'temp-l1']:
if (name, 'mask') in self.weight_dict.keys():
loss = self.loss_dict[name](pred_mask, prev_pred_mask)
loss_values[(name, 'mask')] = loss.item()
generator_loss += self.weight_dict[(name, 'mask')] * loss
if (name, 'normal') in self.weight_dict.keys():
loss = self.loss_dict[name](pred_normal * gt_mask_clamp,
prev_pred_normal *
gt_mask_clamp)
loss_values[(name, 'normal')] = loss.item()
generator_loss += self.weight_dict[(name, 'normal')] * loss
if (name, 'ao') in self.weight_dict.keys():
prev_pred_ao = prev_pred_warped[:, 5:6, :, :]
loss = self.loss_dict[name](pred_ao * gt_mask_clamp,
prev_pred_ao * gt_mask_clamp)
loss_values[(name, 'ao')] = loss.item()
generator_loss += self.weight_dict[(name, 'ao')] * loss
if (name, 'depth') in self.weight_dict.keys():
prev_pred_depth = prev_pred_warped[:, 4:5, :, :]
loss = self.loss_dict[name](pred_depth * gt_mask_clamp,
prev_pred_depth *
gt_mask_clamp)
loss_values[(name, 'depth')] = loss.item()
generator_loss += self.weight_dict[(name, 'depth')] * loss
if (name, 'flow') in self.weight_dict.keys():
prev_pred_depth = prev_pred_warped[:, 4:5, :, :]
loss = self.loss_dict[name](
pred_flow, #note: no restriction to inside areas
prev_pred_flow)
loss_values[(name, 'flow')] = loss.item()
generator_loss += self.weight_dict[(name, 'flow')] * loss
if (name, 'color') in self.weight_dict.keys():
prev_pred_color = self.shading(prev_pred_warped)
loss = self.loss_dict[name](pred_color, prev_pred_color)
loss_values[(name, 'color')] = loss.item()
generator_loss += self.weight_dict[(name, 'color')] * loss
return generator_loss, loss_values