dist = ((y.data * 256).floor() -
(t.data * 256).floor()).pow(2.).sum()
running_dist += dist
# print('batch{:3d} dist: {:.4f}'.format(batch, dist))
# print('\n')
# for i in range(0,8):
# print((y.data[i]-t.data[i]).pow(2).mean())
if epoch % 15 == 0:
np_y = (y * 256).data.transpose(1, 3).transpose(
1, 2).squeeze().cpu().numpy()
for i, image in enumerate(np_y):
cv2.imwrite(
'./result/{}/{:d}.bmp'.format(phase, idx[i]),
image)
mse = running_dist / (y.data.cpu().numpy().size * batch)
psnr = 10 * ((255 * 255 / mse).log() / tensor(10.).log().cuda())
print('running psnr: {:.4f}\n'.format(psnr))
torch.save(module.state_dict(), './autosave/module.pth')
torch.save(optimizer.state_dict(), './autosave/optimizer.pth')
torch.save(optimizer_lam.state_dict(), './autosave/optimizer_lam.pth')
if epoch % 50 == 0:
torch.save(module.state_dict(),
'./pth/module_iter_{:d}.pth'.format(iterations))
print('saved - pth/module_iter_{:d}.pth'.format(iterations))
iterations += 1
torch.save(module.state_dict(), 'module.pth')
