[ Quick Overview | Docs | Original TPS Repo]

The code that came with the FOMM paper became a bit of a standard in this area of research, but the code lacks annotation as to what kind of objects it passes around.

This lack of annotation and documentation makes it harder to understand the code and scarier to tweak it.

This project is here to document what’s going on there. Key focus is on type annotaoitn and use of dataclasses to be more explicit as to what is being passed around and einops to be clearer about the shapes of tensors.

Eventually you should get something like this:

During inference we can skip part of the copmutations for the source image. Most notably, we can skip the encoder part for the inpainting network. Here I split the inpainting forward pass to encoder and decoder and run the encoder once during inference. This seems to speed up inference FPS by ~10%-20% (don't trust this number, I got it by eyeballing the tqdm its/sec for several runs)

• data loading functionality
• training loop documentation