transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) ? about ccpl HOT 2 CLOSED

Hi, thank you for your greate work! I find that your code was not normalized with the pre trained ImageNet value, as shown in the title. So I add this to the code, but when I use the trained weight to test, the performance is not good. In contrast, I remove regularization and get the same result as yours. So why this difference？ Can you tell me the reason? thank you!

Hello, Zhang! Thanks for your appreciation. Our code is based on the pytorch implementation of the AdaIN paper. As you can see in issue 13, the original implementation does not involve input normalization, too. I guess it might not be necessary to pre-normalize the input images in stylization tasks. You could try to add pre-normalization during and see how it works in the inference stage.

from ccpl.

Hi, thank you for your greate work! I find that your code was not normalized with the pre trained ImageNet value, as shown in the title. So I add this to the code, but when I use the trained weight to test, the performance is not good. In contrast, I remove regularization and get the same result as yours. So why this difference？ Can you tell me the reason? thank you!

Hello, Zhang! Thanks for your appreciation. Our code is based on the pytorch implementation of the AdaIN paper. As you can see in issue 13, the original implementation does not involve input normalization, too. I guess it might not be necessary to pre-normalize the input images in stylization tasks. You could try to add pre-normalization during and see how it works in the inference stage.

Thank you for your reply. I guess that I have found why the difference occuried.
def denormalzation(tensor, device): mean = torch.tensor([0.485, 0.456, 0.406]).reshape(-1, 1, 1).to(device) std = torch.tensor([0.229, 0.224, 0.225]).reshape(-1, 1, 1).to(device) tensor = torch.clamp(tensor * std + mean, 0, 1) return tensor
I made a mistake that I used the torch.clamp(tensor * mean + std). Thank you again. It's really not necessary. I will verify the modified results.

from ccpl.