search

ai-safety-foundation / sparse_autoencoder

Sparse Autoencoder for Mechanistic Interpretability
https://ai-safety-foundation.github.io/sparse_autoencoder/
MIT License
191 stars 39 forks source link

Should be a way to upload models automatically #140

Closed HoagyC closed 11 months ago

HoagyC commented 12 months ago

System should work with aws S3 or similar. should check that it can upload a dummy file at the beginning and then upload the checkpoints as they are made.

Should probably be a separate, optional set of hparams to pass to pipeline for the basics like bucket_name and access_key, and then get the private keys from a secrets.json

needs to be careful with uploading checkpoints from ../../.checkpoints as the multiple .. cause errors with my existing aws code.

I can add this, just want to see if there are any suggestions or disagreements on how to add this. Will have a look at doing it this eve.

alan-cooney commented 11 months ago

Sounds good! Maybe shove this into the abstract model class? And then we can call from the pipeline?

For referencing directories use pathlib.Path

HoagyC commented 11 months ago

Reckon this is still needed given that you've added wandb integration? Am now thinking that the S3 stuff is quite specific - it's how I'd do my handling but feels maybe a bit too user specific to put into the package?

alan-cooney commented 11 months ago

Up to you - no strong views on this