Lucas Vazquez

Investigate why u-net performs poorly with style transfer

4

Theoretically it should be way better than `TransformerNet`. It performs really well for superres (which is almost the same thing). It's a more appropriate architecture image to image problems overall.

swap two parts of the image

1

Select two boxes at random and swap their places

When doing the "increase resolution as training proceeded trick", we can use the output of the just finished trained model as the input of the new one (instead of the...

Just need to apply an edge detection algorithm to the input, like canny

Hat is very small when the cat is laying on it's side

Similar to #11 can have a vector weights defining what hat to use

Paper [link](https://arxiv.org/pdf/1610.07629.pdf). Very simple concept: Use a vector of weights together with input image to choose how much of each style is wanted. Should be very easy to implement with...

Refactor FeatModels

1

I feel like there must be a better way of architecturing this. Not so clear right now, also not important. No need to worry about this for now.

Right now it's just a copy-paste from `pytorch-examples`