Hi Folks, welcome to this repositiry on neural style transfer, this is the implementation og gatys paper of 2015. The coherent idea here is to basically create a network which can be freezed by its trained weights over imagenet or pascal, and then optimize the input images from noise to fit content and stylise it. A few important underatnding and assumptions in this regard are:
- Rather than starting from niose, I have selected a replica of starting image to start of
- alpha is taken at 1 and beta at 0.01 which is different from the actual paper
- The loss function is slightly modified to fit well.
- Generall 6000-12000 epochs are trained for each sample of two images.
- I donot have gpu so it was trained on kaggle and other open source libraries.
- The use of jupyter file is to elaborate learning and understanding internally than a package of complete work.
If you like my work please upvote this Notebook
