Important

This repo is a fork of mil-ad/prospr that I modified to my needs. The actual implementation is not mine, but the original author's. The corresponding paper to the method can be found on ArXiV: Alizadeh et al. (2022).

Changes I made

For my master's thesis , I rewrote the ProsPr method in a more efficient way that makes use of the prune function that is built in to PyTorch. I'm using the original code as a sanity-checking method, to compare my results at the original results and see whether Im implementation works. To be able to do this, I applied the following changes to this repo:

• Moved from python=3.8 to python=3.10.8
• Moved from pytorch=1.9.1 to pytorch=1.13.1
  • Required minor changes to different for loops, where instance must be checked to avoid errors
• Added some docstrings for functions I found confusing at first
• Removed anything related to structured pruning as it is irrelevant for my use case
• Removed anything after computing the pruning masks (i.e. application and training), as I don't need the overhead for checking the pruning method (which is done at initialization)
• For my project, I will only require the files from the prospr folder, but I added things to check if the method works with my data/model
  • Use timm to load the models instead of the custom-built models by the authors
  • Use the dataloader from my thesis instead of the author's
    • Since I'm CPU bound (for now), I only work with CIFAR10 data here
  • Removed CLI functionality, added main instead