eth-easl / modyn

Modyn is a research-platform for training ML models on growing datasets.
MIT License
22 stars 3 forks source link

RHO-LOSS Step 4: Implement `RHOLossRemoteDownsampler` #491

Closed XianzheMa closed 1 month ago

XianzheMa commented 1 month ago

This PR implements the remote RHO-LOSS downsampler at trainer server, thus finishing the implementation of RHO-LOSS.

How to review

The number of touched files is huge but most of them are nominal changes. Please first take a look at remote_rho_loss_downsampling.py, then at irreducible_loss_producer.py, then at their corresponding unit test files, then at rest files.

What's left

I still need an integration test to connect all pieces together as this selection policy is non-trivial.

github-actions[bot] commented 1 month ago

Line Coverage: -% ( % to main) Branch Coverage: -% ( % to main)

codecov[bot] commented 1 month ago

Codecov Report

Attention: Patch coverage is 96.85039% with 4 lines in your changes missing coverage. Please review.

Project coverage is 82.03%. Comparing base (4373fe2) to head (3b91ef5).

Files Patch % Lines
...trainer_server/internal/trainer/pytorch_trainer.py 33.33% 4 Missing :warning:
Additional details and impacted files ```diff @@ Coverage Diff @@ ## main #491 +/- ## ========================================== - Coverage 83.90% 82.03% -1.88% ========================================== Files 209 214 +5 Lines 9633 9948 +315 ========================================== + Hits 8083 8161 +78 - Misses 1550 1787 +237 ```

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

MaxiBoether commented 1 month ago

feel free to reduce the required coverage to 82%, if that's the blocking issue here