Ideas for New Features

[OFSkean]

Hello! I'm a big fan of this library and have been using it in my research for over a year now. At this point, I have a fork with several new methods and some new features. I want to fold my changes back into the main repository, so are you still accepting new PRs?

New Methods:

• FroSSL Disclaimer: this is my work
• CorInfomax
• MMCR
• EMP-SSL
• AugSelf
• Data Augmentation Disentanglement

New Features

• Using the sampled augmentation parameters (like was_flipped or Gaussian blur strength) in the loss function. This is necessary for the AugSelf and Data Augmentation Disentanglement methods.
• New evaluation metrics like RankMe, NCC and CDNV, and non-linear probe (MLP). Adding these metrics will be a nice complement to the standard linear probe evaluation.

[vturrisi]

Hey! Yes, please feel free to create PRs for the things that you have implemented. I currently don't have much time to implement new features, but I would be happy to review your PRs.

Note: create one separate PR for each feature so that it's easier for me to review the individual components.

[OFSkean]

Great! I'll try to submit a few PRs this weekend

[OFSkean]

Hey @vturrisi! I just submitted a PR here that implements FroSSL. I'll start with this one for now, and I'll do the others once I figure out how the CI and tests work. By the way, how do I run tests locally?

[vturrisi]

Thanks, I'll check it in the course of this week. To run the tests, you can do pytest --cov=solo tests/args tests/backbones tests/data tests/losses tests/methods tests/utils (simply doing pytest --cov=solo tests/ also works, but it will also run the dali tests). You will also need to install some libraries (pip install -e .[umap,h5] codecov mypy pytest-cov).