jbloomAus commented 1 month ago

Description

In this PR, we support HookedSAETransformer (coming over from TransformerLens) and refactor a bunch of stuff internally. I'm sorry this PR is so large and for the number of breaking changes. I'm hoping that the future need for refactors is much smaller. We expect the changes to be fairly superficial / easy to adapt to except for those working on forks. Feel free to reach out for assistance / clarification if you are trying to update a fork. Finally, we think it's likely there are no major regressions or introduced bugs, but the test coverage now looks lower as we're showing over the entire repo and not just the training subpackage.

New Features:

HookedSAE Transformer has been ported from TransformerLens to SAE Lens! (note: use_error_term is a property of SAE not SAEConfig.
I've added more SAEs to from_pretrained (and will likely add more in the coming days).

Breaking changes:

Features we removed:

Recently, we removed training SAEs in parallel with one dataloader. This wasn't very useful in practice and made the code quite nasty in places.
We also recently removed resuming training when SAE training jobs are interrupted. This required we store much larger objects and caused a breakdown in many of the abstractions. I've kept the "save on ctrl-c" functionality because that seems useful. We might restore this in the future.

Renaming:

Classes have been renamed so that we use "SAE" instead of "SparseAutoencoder" wherever possible.
Some modules have been moved around to better reflect dependency structures.
Key config items have been renamed to remove “_point” wherever it is present. This brings naming closer in line with TransformerLens.
Config arguments dtype/device of runners are now strings and only strings. This makes things simpler when dealing with high level APIs / typing. We want cfg classes are always serializable.

Notes:

SAEs now have their own configs. The base SAE config includes the minimal information (hopefully) to run / use the SAE. You can mock this info if loading in your own SAEs but we recommend you find the appropriate values so that the SAE can be used correctly (eg: specifying the context length it was trained on or whether prompts has a bos token prepended).

Type of change

Please delete options that are not relevant.

[ ] Bug fix (non-breaking change which fixes an issue)
[ ] New feature (non-breaking change which adds functionality)
[x] Breaking change (fix or feature that would cause existing functionality to not work as expected)
[x] This change requires a documentation update

Checklist:

[x] I have commented my code, particularly in hard-to-understand areas
[x] I have made corresponding changes to the documentation
[x] My changes generate no new warnings
[x] I have added tests that prove my fix is effective or that my feature works
[x] New and existing unit tests pass locally with my changes
[x] I have not rewritten tests relating to key interfaces which would affect backward compatibility

You have tested formatting, typing and unit tests (acceptance tests not currently in use)

[x] I have run make check-ci to check format and linting. (you can run make format to format code if needed.)

Performance Check.

If you have implemented a training change, please indicate precisely how performance changes with respect to the following metrics:

[x] L0
[x] CE Loss
[x] MSE Loss
[ ] Feature Dashboard Interpretability

Please links to wandb dashboards with a control and test group.

codecov[bot] commented 1 month ago

Codecov Report

Attention: Patch coverage is 59.28500% with 410 lines in your changes are missing coverage. Please review.

Project coverage is 52.34%. Comparing base (eb9489a) to head (3faeae8).

Files	Patch %	Lines
sae_lens/training/sae.py	0.00%	172 Missing :warning:
sae_lens/sae_training_runner.py	30.76%	63 Missing :warning:
sae_lens/training/sae_trainer.py	68.13%	46 Missing and 12 partials :warning:
sae_lens/sae.py	86.74%	12 Missing and 12 partials :warning:
sae_lens/training/training_sae.py	88.05%	19 Missing :warning:
sae_lens/config.py	64.86%	13 Missing :warning:
sae_lens/analysis/hooked_sae_transformer.py	86.41%	6 Missing and 5 partials :warning:
sae_lens/analysis/neuronpedia_runner.py	0.00%	11 Missing :warning:
sae_lens/toolkit/pretrained_sae_loaders.py	57.69%	10 Missing and 1 partial :warning:
sae_lens/evals.py	66.66%	4 Missing and 3 partials :warning:
... and 6 more

Additional details and impacted files

```diff @@ Coverage Diff @@ ## main #162 +/- ## =========================================== - Coverage 67.13% 52.34% -14.80% =========================================== Files 19 26 +7 Lines 1710 2755 +1045 Branches 267 462 +195 =========================================== + Hits 1148 1442 +294 - Misses 504 1240 +736 - Partials 58 73 +15 ```

:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.