I read the A neural Algorithm of artistic style recently and while trying to implement I referred pytorch's tutorial but it used only top layers rather than full model as suggested by paper
From same input images :
Style image
Context image:
This is result from tutorial :
This is the result after full model :
As you can see my results captured context as well as style better than current version
the algorithm is expected to work with transfer learning from VGG19 model which will have 19 conv layers but the current tutorial only have 5 layers which is why results aren't up to the mark.
Changes:
Changed in implantation of model
max pool to avg pool as suggested by author in paper
VGG takes 224*224 input image which is not the case with current tutorial.
multiplying individual loss by a factor
due to normalisation on gram matrix style loss tends to zero
I read the A neural Algorithm of artistic style recently and while trying to implement I referred pytorch's tutorial but it used only top layers rather than full model as suggested by paper
From same input images :
Style image
Context image:
This is result from tutorial :
This is the result after full model :
As you can see my results captured context as well as style better than current version the algorithm is expected to work with transfer learning from VGG19 model which will have 19 conv layers but the current tutorial only have 5 layers which is why results aren't up to the mark.
Changes:
I would like to contribute to this issue