def evaluate(self, current_epoch=0):
self.logger.info('Epoch ' + str(current_epoch) +
' evaluation process...')
if (self.args.dataset == 'HMDB_FRAMES'):
self.model.eval()
with torch.no_grad():
psnr, ssim, cnt = 0., 0., 0
for i_batch, sample_batched in enumerate(
self.dataloader['test']['1']):
cnt += 1
sample_batched = self.prepare(sample_batched)
lr = sample_batched['LR']
lr_sr = sample_batched['LR_sr']
hr = sample_batched['HR']
ref = sample_batched['Ref']
ref_sr = sample_batched['Ref_sr']
hr = hr.repeat(1, 3, 1, 1)
# sr, _, _, _, _ = naive_averaging(
# self.model, lr, lr_sr, hr, ref, ref_sr) TODO uncomment for modified fusion model
sr, _, _, _, _ = self.model(lr=lr,
lrsr=lr_sr,
ref=ref,
refsr=ref_sr)
if (self.args.eval_save_results):
sr_save = (sr + 1.) * 127.5
sr_save = np.transpose(
sr_save.squeeze().round().cpu().numpy(),
(1, 2, 0)).astype(np.uint8)
imsave(
os.path.join(self.args.save_dir, 'save_results',
str(i_batch).zfill(5) + '.png'),
sr_save)
# calculate psnr and ssim
# sr = sr.squeeze(0)
# hr = hr.squeeze(0) TODO uncomment for modified fusion model
_psnr, _ssim = calc_psnr_and_ssim(sr.detach(), hr.detach())
psnr += _psnr
ssim += _ssim
psnr_ave = psnr / cnt
ssim_ave = ssim / cnt
self.logger.info('Ref PSNR (now): %.3f \t SSIM (now): %.4f' %
(psnr_ave, ssim_ave))
if (psnr_ave > self.max_psnr):
self.max_psnr = psnr_ave
self.max_psnr_epoch = current_epoch
if (ssim_ave > self.max_ssim):
self.max_ssim = ssim_ave
self.max_ssim_epoch = current_epoch
self.logger.info(
'Ref PSNR (max): %.3f (%d) \t SSIM (max): %.4f (%d)' %
(self.max_psnr, self.max_psnr_epoch, self.max_ssim,
self.max_ssim_epoch))
self.logger.info('Evaluation over.')
def test_FixedRef(self, lr_path, ref_path, save_path):
### LR and LR_sr
HR = imread(lr_path)
h1, w1 = HR.shape[:2]
h1, w1 = h1 // 4 * 4, w1 // 4 * 4
HR = HR[:h1, :w1, :]
LR = np.array(
Image.fromarray(HR).resize((w1 // 4, h1 // 4), Image.BICUBIC))
LR_sr = np.array(Image.fromarray(LR).resize((w1, h1), Image.BICUBIC))
### Ref and Ref_sr
Ref = imread(ref_path)
h2, w2 = Ref.shape[:2]
h2, w2 = h2 // 4 * 4, w2 // 4 * 4
Ref = Ref[:h2, :w2, :]
Ref_sr = np.array(
Image.fromarray(Ref).resize((w2 // 4, h2 // 4), Image.BICUBIC))
Ref_sr = np.array(
Image.fromarray(Ref_sr).resize((w2, h2), Image.BICUBIC))
### change type
HR = HR.astype(np.float32)
LR = LR.astype(np.float32)
LR_sr = LR_sr.astype(np.float32)
Ref = Ref.astype(np.float32)
Ref_sr = Ref_sr.astype(np.float32)
### rgb range to [-1, 1]
HR = HR / 127.5 - 1.
LR = LR / 127.5 - 1.
LR_sr = LR_sr / 127.5 - 1.
Ref = Ref / 127.5 - 1.
Ref_sr = Ref_sr / 127.5 - 1.
### to tensor
HR_t = torch.from_numpy(HR.transpose(
(2, 0, 1))).unsqueeze(0).float().to(self.device)
LR_t = torch.from_numpy(LR.transpose(
(2, 0, 1))).unsqueeze(0).float().to(self.device)
LR_sr_t = torch.from_numpy(LR_sr.transpose(
(2, 0, 1))).unsqueeze(0).float().to(self.device)
Ref_t = torch.from_numpy(Ref.transpose(
(2, 0, 1))).unsqueeze(0).float().to(self.device)
Ref_sr_t = torch.from_numpy(Ref_sr.transpose(
(2, 0, 1))).unsqueeze(0).float().to(self.device)
self.model.eval()
with torch.no_grad():
sr, _, _, _, _ = self.model(lr=LR_t,
lrsr=LR_sr_t,
ref=Ref_t,
refsr=Ref_sr_t)
sr_save = (sr + 1.) * 127.5
sr_save = np.transpose(sr_save.squeeze().round().cpu().numpy(),
(1, 2, 0)).astype(np.uint8)
imsave(save_path, sr_save)
psnr, ssim = calc_psnr_and_ssim(sr.detach(), HR_t.detach())
return psnr, ssim
def evaluate(self, current_epoch=0):
self.logger.info('Epoch ' + str(current_epoch) +
' evaluation process...')
if (
self.args.dataset == 'IMM'
): #Dataset name has to be manually altered here.. better Solution!
self.model.eval()
with torch.no_grad():
psnr, ssim, cnt = 0., 0., 0
for i_batch, sample_batched in enumerate(
self.dataloader['test']['1']):
cnt += 1
sample_batched = self.prepare(sample_batched)
lr = sample_batched['LR']
lr_sr = sample_batched['LR_sr']
hr = sample_batched['HR']
ref = sample_batched['Ref']
ref_sr = sample_batched['Ref_sr']
sr, _, _, _, _ = self.model(lr=lr,
lrsr=lr_sr,
ref=ref,
refsr=ref_sr)
if (self.args.eval_save_results):
sr_save = (sr + 1.) * 127.5
sr_save = np.transpose(
sr_save.squeeze().round().cpu().numpy(), (1, 2, 0)
).astype(
np.uint8
) # squeeze delets all 1 values --- round() rounds to closest integer -- .cpu() moves object to cpu ---.numpy transforms object to numpy array--- transform to np.unit8 type
imsave(
os.path.join(self.args.save_dir, 'save_results',
str(i_batch).zfill(5) + '.tif'),
sr_save)
### calculate psnr and ssim
_psnr, _ssim = calc_psnr_and_ssim(sr.detach(), hr.detach())
psnr += _psnr
ssim += _ssim
psnr_ave = psnr / cnt
ssim_ave = ssim / cnt
self.logger.info('Ref PSNR (now): %.3f \t SSIM (now): %.4f' %
(psnr_ave, ssim_ave))
if (psnr_ave > self.max_psnr):
self.max_psnr = psnr_ave
self.max_psnr_epoch = current_epoch
if (ssim_ave > self.max_ssim):
self.max_ssim = ssim_ave
self.max_ssim_epoch = current_epoch
self.logger.info(
'Ref PSNR (max): %.3f (%d) \t SSIM (max): %.4f (%d)' %
(self.max_psnr, self.max_psnr_epoch, self.max_ssim,
self.max_ssim_epoch))
self.logger.info('Evaluation over.')
def evaluate(self, filepath):
input_image = align_image(load_image(filepath), self.scale)
input_y_image = resize_image(convert_rgb_to_y(input_image),
1 / self.scale)
input_scaled_y_image = resize_image(input_y_image, self.scale)
output_y_image = self.run(input_y_image, input_scaled_y_image)
ground_truth_y_image = convert_rgb_to_y(input_image)
return calc_psnr_and_ssim(ground_truth_y_image,
output_y_image,
border=self.scale)
def evaluate_flownet(self, args, current_epoch=0):
self.logger.info('Epoch ' + str(current_epoch) +
' evaluation process...')
if (self.args.dataset == 'HMDB_FLOWNET'):
self.model.eval()
with torch.no_grad():
psnr, ssim, cnt = 0., 0., 0
for i_batch, sample_batched in enumerate(
self.dataloader['test']['1']):
cnt += 1
sample_batched = self.prepare(sample_batched)
lr = sample_batched['LR'].float(
) / 127.5 - 1 # [5, 3, 40, 40]
lr_sr = sample_batched['LR_sr'].float(
) / 127.5 - 1 # [5, 3, 160, 160]
hr = sample_batched['HR'].float(
) / 127.5 - 1 # [1, 3, 160, 160]
ref = sample_batched['Ref'].float(
) / 127.5 - 1 # [1, 3, 160, 160]
ref_sr = sample_batched['Ref_sr'].float(
) / 127.5 - 1 # [1, 3, 160, 160]
if args.train_style == "normal":
sr, S, T_lv3, T_lv2, T_lv1 = self.model(
lr=lr[:, 2, :, :, :],
lrsr=lr_sr[:, 2, :, :, :],
ref=ref[:, :, :, :],
refsr=ref_sr[:, :, :, :])
elif args.train_style == "average":
sr, S, T_lv3, T_lv2, T_lv1 = flownet_naive_averaging(
self.model, lr, lr_sr, hr, ref, ref_sr)
else:
sr, S, T_lv3, T_lv2, T_lv1 = flownet_conv3d_1x1(
self.model, lr, lr_sr, hr, ref, ref_sr)
if (self.args.eval_save_results):
sr_save = (sr + 1.) * 127.5
sr_save = np.transpose(
sr_save.squeeze().round().cpu().numpy(),
(1, 2, 0)).astype(np.uint8)
imsave(
os.path.join(self.args.save_dir, 'save_results',
str(i_batch).zfill(5) + '.png'),
sr_save)
# calculate psnr and ssim
# sr = sr.squeeze(0)
# hr = hr.squeeze(0) TODO uncomment for modified fusion model
_psnr, _ssim = calc_psnr_and_ssim(sr.detach(), hr.detach())
psnr += _psnr
ssim += _ssim
psnr_ave = psnr / cnt
ssim_ave = ssim / cnt
self.logger.info('Ref PSNR (now): %.3f \t SSIM (now): %.4f' %
(psnr_ave, ssim_ave))
if (psnr_ave > self.max_psnr):
self.max_psnr = psnr_ave
self.max_psnr_epoch = current_epoch
if (ssim_ave > self.max_ssim):
self.max_ssim = ssim_ave
self.max_ssim_epoch = current_epoch
self.logger.info(
'Ref PSNR (max): %.3f (%d) \t SSIM (max): %.4f (%d)' %
(self.max_psnr, self.max_psnr_epoch, self.max_ssim,
self.max_ssim_epoch))
self.logger.info('Evaluation over.')