Skip to content

feat(losses): generalize noise_corrected_bce to multiclass - #322

Open
jayhesselberth wants to merge 1 commit into
mainfrom
worktree-noise-corrected-ce-multiclass
Open

jayhesselberth wants to merge 1 commit into
mainfrom
worktree-noise-corrected-ce-multiclass

Conversation

@jayhesselberth

Copy link
Copy Markdown
Member

Generalizes --loss noise_corrected_bce to multiclass (issue #321), closing the gap that blocked using the unified 20+1-class charge-aware model with the existing noise-correction loss.

🤖 Generated with Claude Code

https://claude.ai/code/session_0165BfnCSaxTps9iJqRyurKi

NoiseCorrectedCrossEntropyLoss extends the binary forward correction
(#279) to a single-sink noise structure for --num-out > 1: each
non-sink class has its own measured purity, and every class's label
noise flows only to one designated sink class (--noise-sink-class,
resolved via label_map). --label-noise-rate keys are resolved as
class labels rather than per-sample source_group lookups at num_out
> 1; all-zero rates reduce to plain cross-entropy bit-for-bit. Also
fixes train_model silently overriding an explicit --loss
noise_corrected_bce to cross_entropy whenever num_out > 1.

Closes #321.

Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_0165BfnCSaxTps9iJqRyurKi
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