Is it possible to choose the layers ?
zifken opened this issue · 2 comments
Is it possible to choose wich layers of the vgg decoders corresponds to contents and syle while using a pretrained decoder ?
It sounds interesting. Maybe we could do some experiment on it. However, in my intuition, each layer of the decoder contains features of both content and style. That means there doesn't exist a layer only holds content features or style features. But I am sure that some layers hold content features with a higher weight, while the other layers hold "more" style features. So maybe we could use some strategy to split the content features and style features. After all, your idea is great and it is worth to do some research.
Thank you for your answer.
to be moer precise : This idea came from other implementation mainly The Leon A. Gatys implementions of style transfer.
In those implementations we use low level layers to compute the content loss and hight level layer and the gram matrix of the pastiche .