def test(dataset, loader, model, args, device, tag=''):
psnr = AverageMeter()
# Set the model to evaluation mode
model.eval()
with tqdm(total=len(dataset)) as t:
t.set_description(tag)
for data in loader:
lr, hr = data
lr = lr.to(device)
hr = hr.to(device)
if args.self_ensemble:
sr = forward_x8(lr)
else:
sr = forward(lr)
# Quantize results
sr = quantize(sr, args.rgb_range)
# Update PSNR
psnr.update(calc_psnr(sr, hr, scale=args.scale, max_value=args.rgb_range[1]), lr.shape[0])
t.update(lr.shape[0])
print('DIV2K (val) PSNR: {:.4f} dB'.format(psnr.avg))
def test(dataset, loader, model, device, args, tag=''):
psnr = AverageMeter()
# Set the model to evaluation mode
model.eval()
count = 0
with tqdm(total=len(dataset)) as t:
t.set_description(tag)
for data in loader:
count = count + 1
lr, hr = data
# pdb.set_trace()
lr = lr.to(device)
hr = hr.to(device)
if args.self_ensemble:
sr = forward_x8(lr)
else:
sr = forward(lr)
# pdb.set_trace()
# (Pdb) lr.size()
# torch.Size([1, 3, 678, 1020])
# (Pdb) sr.size()
# torch.Size([1, 3, 1356, 2040])
# Quantize results
sr = quantize(sr, args.rgb_range)
# Update PSNR
psnr.update(
calc_psnr(sr,
hr,
scale=args.scale,
max_value=args.rgb_range[1]), lr.shape[0])
t.update(lr.shape[0])
# if count > 100:
# srimg = torchvision.transforms.ToPILImage()(sr.squeeze().div(255).cpu())
# srimg.save("result/{:03d}.png".format(count - 100))
# # srimg.show()
# # pdb.set_trace()
print('SDR (val) PSNR: {:.4f} dB'.format(psnr.avg))
def test(dataset, loader, model, args, device, tag=''):
psnr = AverageMeter()
# Set the model to evaluation mode
model.eval()
with tqdm(total=len(dataset)) as t:
t.set_description(tag)
count = 0
for data in loader:
lr, hr = data
lr = lr.to(device)
hr = hr.to(device)
if args.self_ensemble:
sr = forward_x8(lr, model)
else:
sr = forward(lr, model)
# Quantize results
sr = quantize(sr, args.rgb_range)
if args.output_dir:
im = Image.fromarray(np.uint8(sr.squeeze().cpu().numpy().transpose(2, 1, 0))).transpose(Image.ROTATE_270)
im_lr = Image.fromarray(np.uint8(lr.squeeze().cpu().numpy().transpose(2, 1, 0))).transpose(Image.ROTATE_270)
im.save(args.output_dir + str(count) + ".png")
im_lr.save(args.output_dir + str(count) + "_lr.png")
count += 1
# to save activation layers
# act = activation['deconv'].squeeze()
#
# for idx in range(50):
# plt.imsave('results/output/act/activation%d_.png' % idx, act[idx], vmin=act[idx].min(), vmax=act[idx].max())
# Update PSNR
# psnr.update(calc_psnr(sr, hr, scale=args.scale, max_value=args.rgb_range[1]), lr.shape[0])
t.update(lr.shape[0])
print('DIV2K (val) PSNR: {:.4f} dB'.format(psnr.avg))
def test(dataset, loader, model, criterion, args, tag=''):
losses = AverageMeter()
psnr = AverageMeter()
# Set the model to evaluation mode
model.eval()
with tqdm(total=len(dataset)) as t:
t.set_description(tag)
for data in loader:
lr, hr = data
lr = lr.to(device)
hr = hr.to(device)
# Predict results without calculating gradients
with torch.no_grad():
sr = model(lr)
loss = criterion(sr, hr)
# Update loss
losses.update(loss.item(), lr.shape[0])
# Quantize results
sr = quantize(sr, args.rgb_range)
# Update PSNR
psnr.update(
calc_psnr(sr,
hr,
scale=args.scale,
max_value=args.rgb_range[1]), lr.shape[0])
t.set_postfix(loss='{:.4f}'.format(losses.avg))
t.update(lr.shape[0])
return losses.avg, psnr.avg