search

f-dangel / backpack

BackPACK - a backpropagation package built on top of PyTorch which efficiently computes quantities other than the gradient.
https://backpack.pt/
MIT License
562 stars 56 forks source link

Feature for backpack on VAEs #319

Open shubham0704 opened 1 year ago

shubham0704 commented 1 year ago

Hi All! I want to contribute to backpack with an implementation that works for VAEs. Please let me know if anyone else is working on this or I can carry out this task. I need to do this for my own research and figured you might also want this feature. Thanks a lot for building this amazing library!

f-dangel commented 1 year ago

Hi,

sounds like a feature that would be useful for others! I am not aware of anyone working on this.

Disclaimer: I don't have much expertise on VAEs. So its hard for me to foresee potential design problems in BackPACK that might complicate realizing this feature, and to provide informed feedback. If you want though, we can have a virtual meeting to talk details.

Felix

shubham0704 commented 1 year ago

Got it! My email for correspondence is shubham.bhardwaj@utexas.edu Currently, I am looking to see how to add custom loss function and then going for checking how can we incorporate models using torch.distributions package. What I can do initially is to create a colab notebook with the fork of this repository with modifications that would enable me to have a better discussion with you on where I am stuck.