This repo is a fork of mil-ad/snip that I modified to my needs. The actual implementation is not mine, but Milad Alizadeh's. I just adapted it to my needs. The corresponding paper to the code is by Lee et al. (2018).
For my master's thesis , I rewrote the SNIP method in a more efficient way that makes use of the prune function that is built in to PyTorch. I'm using this code as a sanity-checking method, to compare my results see whether my implementation works. To be able to do this, I applied the following changes to this repo:
- Moved loss to
crossentropy - Moved to
timmfor models - Moved to my dataloaders for fair comparison
- Included a stylized print function for all layers
- Removed anything to do with training as I don't need it
- Added a main to demonstrate the functionality