Fixed VGG encoder weights query about wct2 HOT 2 CLOSED

@SidShenoy Thank you for your comment. That is a very sharp observation!

Yes, indeed, changing the pooling layers affects the following feature maps. Since the output of the max-pooling is now changed to the LL filter (similar to average pooling), the feature map after the pooling layer is slightly different. (Note that the other feature maps from the other three wavelet filters do not propagate to the next layer of the encoder. They are skipped to the decoder so that the only change the encoder has to care about is due to the LL filter change from the max-pooling)

However, as we wrote in the paper, we decided not to touch the encoder but just let the decoder adapt to those changes. You can fine-tune the encoder weights by partially or entirely freeing encoder weight parameters and we actually tried some variants, such as freeing only the following convolution parameters (after pooling layer) so that the change will be dealt in the encoder as well. There was not much difference at the final outcomes so we chose to stick on the simpler training strategy.

This can be explained in two-folds; 1) It is already a well-known phenomenon and a lot of observations were consistently reported that style transfer can be done with changing the max-pooling to average pooling (even though the VGG network was trained using the max-pooling) and the effect is sometimes even better. Similarly, our encoder with LL filter, which is an average pooling with some scaling factor, shares this characteristic. 2) Since the decoder is newly trained, it has enough capacity to deal with such shiftings of the feature maps in the encoder to output a good reconstruction.

Still, your comment is very valuable and we will include our description of the training procedure more in detail to clarify the point. Thx a lot for your attention :)

from wct2.

Maybe this partial code snippet would help your understanding on what we did:

for param in self.encoder.parameters():
    param.requires_grad = False
self.dec_optim = torch.optim.Adam(
           filter(lambda p: p.requires_grad, self.decoder.parameters()),
           lr = self.lr,
           betas=(self.beta1, self.beta2)
        )
feature, skips = self.encoder(real_image)
recon_image = self.decoder(feature, skips)
feature_recon, _ = self.encoder(recon_image)
recon_loss = self.MSE_loss(recon_image, real_image)
feature_loss = torch.zeros(1).to(self.device)
feature_loss += self.MSE_loss(feature_recon, feature.detach())
loss = recon_loss * self.recon_weight + feature_loss * self.feature_weight
self.reset_grad()
loss.backward()
self.dec_optim.step()

from wct2.