[roedoejet]

EveryVoiceTTS/EveryVoice#887

[semanticdiff-com]

Review changes with  

Changed Files

File	Status
fs2/loss.py	69% smaller
fs2/tests/test_loss.py	0% smaller

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 49.89%. Comparing base (b3ee268) to head (e392e22).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #166      +/-   ##
==========================================
+ Coverage   47.92%   49.89%   +1.96%     
==========================================
  Files          25       25              
  Lines        1980     1980              
  Branches      229      229              
==========================================
+ Hits          949      988      +39     
+ Misses        996      950      -46     
- Partials       35       42       +7     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[joanise]

Nothing jumps out at me as problematic. I understand neither * in this context nor masked_select, but you've added unit tests, so presumably those confirm it's doing the right thing.

Do these changes invalidate old models at all, or just improve training by calculating loss better?

Oh, but once again you reached for unittest.TestCase. Stop doing that please.

[joanise]

use

assert losses_before[key].item() == pytest.approx(loses_after[key].item())

places=4 is replaced by either rel= or abs= depending on what makes sense.

See https://docs.pytest.org/en/stable/reference/reference.html#pytest-approx

[joanise]

Looks good, thanks for convert the unit tests.