Closed XianzheMa closed 3 months ago
All modified and coverable lines are covered by tests :white_check_mark:
Project coverage is 82.84%. Comparing base (
72ffdca
) to head (9804c85
).
:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.
( % to main) ( % to main)
The changes look good but cause a problem in the integration test. When this is fixed, feel free to merge
The failure is because we have a _assert_data_size
logic added when merging the batch accumulation. This assertion happens regardless of we are doing batch accumulation or not. When we set drop_last
to False, naturally the last batch has fewer samples and does not conform to the assertion. Let's hope this time CI passes
The reason for this PR is that, for il model training, I don't want to drop the last batch, as the holdout set is already super small (and on il model training no sampling is used, so it does not harm not to drop last batch). But for main model training, usually we default
drop_last
to True.