Register custom model / loops outside of the VISSL directory

[pclucas14]

🚀 Feature

I would like to be able to integrate custom code, such as custom model head / trunks, or custom training loops, without having to put them directly in the vissl directory

Motivation & Examples

Like many users, I like to download my packages via pip or conda. Right now I have to manually copy my new models to the vissl directory within my environment. The main issue I have with this is that I have to manually keep track of my added files should I do a vissl update.

Note

There is already something similar already established for datasets, where I can register a new one outside of the main codebase. If there would exist something like this for models that would be great.

[prigoyal]

Hi @pclucas14 , the new components can be registered by importing registeries of vissl https://github.com/facebookresearch/vissl/blob/master/vissl/trainer/train_steps/__init__.py#L19 for example importing this function. However, the current VISSL packages don't expose the register functions in __all__ method. But we welcome contributions :) In the next VISSL release in few weeks, we will look into incorporating this feature request :)

[pclucas14]

ok great :) thanks for the heads up

[pclucas14]

Along similar lines, is it possible to add new hooks without modifying the hook init file ?

[prigoyal]

yep! we will open all the register functions along with the ones mentioned here! :)

[pclucas14]

Great :) Sorry to add yet another point to this, but one thing to keep in mind for future PRs would be to ensure the compatibility with the distributed slurm launcher.

Currently I added new import lines in my version of run_distributed_engines.py which takes care of the registration. However when using the slurm launcher, launch_distributed from vissl/utils/distributed_launcher.py is called, which bypasses these imports. An easy hack for now is to modify distributed_launcher.py. In a perfect world if users could avoid modifying internal files usability would be even better.

Thanks again for the great support!