Skip to content

chore(cl): remove EWC — not yet validated, dropping until results land#21

Open
miya-mizumitsuha wants to merge 1 commit into
AlphaBrainGroup:mainfrom
miya-mizumitsuha:chore/cl-drop-ewc
Open

chore(cl): remove EWC — not yet validated, dropping until results land#21
miya-mizumitsuha wants to merge 1 commit into
AlphaBrainGroup:mainfrom
miya-mizumitsuha:chore/cl-drop-ewc

Conversation

@miya-mizumitsuha
Copy link
Copy Markdown
Contributor

EWC was scaffolded as part of the original pluggable CL algorithm family but never produced paper-grade results across a meaningful λ sweep on QwenGR00T-3B × LIBERO-Goal. Rather than ship an under-tested method under the same banner as ER / MIR, we remove it from the public surface until a working configuration is found and re-introduce it then.

Changes:

  • Delete algorithms/regularization_based/ (only contained EWC) along with the empty package itself.
  • Drop the ewc branch in build_cl_algorithm, the EWC re-export from continual_learning/__init__.py + algorithms/__init__.py, and the EWC entry in the algorithm registry docstrings.
  • Strip EWC mentions from train.py comments, base.py docstrings, the docs quickstart page, and the run_continual_learning_scripts README (algorithm tree, hook-protocol table).

Result: ER and MIR remain the two shipped, validated CL algorithms.

EWC was scaffolded as part of the original pluggable CL algorithm family
but never produced paper-grade results across a meaningful λ sweep on
QwenGR00T-3B × LIBERO-Goal. Rather than ship an under-tested method
under the same banner as ER / MIR, we remove it from the public surface
until a working configuration is found and re-introduce it then.

Changes:
- Delete `algorithms/regularization_based/` (only contained EWC) along
  with the empty package itself.
- Drop the `ewc` branch in `build_cl_algorithm`, the EWC re-export from
  `continual_learning/__init__.py` + `algorithms/__init__.py`, and the
  EWC entry in the algorithm registry docstrings.
- Strip EWC mentions from `train.py` comments, `base.py` docstrings,
  the docs quickstart page, and the run_continual_learning_scripts
  README (algorithm tree, hook-protocol table).

Result: ER and MIR remain the two shipped, validated CL algorithms.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant