Why vgg19 updates?

[AlbertZhangHIT]

I'm confused whether we need to update the vgg network. The example in pytorch tutorial keeps the parameters in vgg16 untouched. But as far as I understand, your implementation updates the parameters. Did I miss something important?

[twtygqyy]

Hi @AlbertZhangHIT I didn't update the params in VGG network, just apply the loss for back propagating the sr network.

[AlbertZhangHIT]

God my silly question. The parameters should only be updated by optimizers. So the content_loss will be backpropagated through VGG, and then the Generator since one candidate of the content loss criterion, the content_input, comes from the input(which requires gradient), do I get the point? Thanks for your reply.

Edit: Your implementation is different from others' that I viewed where the total loss is usually in a manner of total_loss = loss1 + loss2 + ... and then backpropagated. It was a little difficult for me to understand your implementation at the first glimpse because of my poor knowledge about PyTorch. But your implementation is quite great. Thanks again!