Fix BiC bias-correction loss to use logits

[marcus-campbell]

Description

During BiC’s bias-correction stage, _run applies torch.softmax to the bias-corrected logits before passing them to F.cross_entropy. However, F.cross_entropy expects unnormalized logits and already applies log-softmax internally. Passing normalized probabilities as its input therefore changes the loss and the gradients used to train the bias-layer parameters.

This PR passes the bias-corrected logits directly to F.cross_entropy. The same call form is already used by BiC’s stage-one training path.

Validation

If you look at the original BiC formulation:

• Equations 4 and 5 of the BiC paper define affine-corrected logits q and optimize -log(softmax(q)).
• The paper authors’ reference implementation passes its corrected logits directly to TensorFlow’s softmax_cross_entropy.

The patched BiC._run(..., stage="bias_correction") path produces exactly the paper-form gradients in a focused PyTorch check. The current "softmax before cross-entropy" expression produced different loss and gradients.

References:

• Large Scale Incremental Learning (BiC)
• BiC authors’ reference implementation
• PyTorch cross_entropy documentation

[caoshuai888]

这是来自QQ邮箱的假期自动回复邮件。   您好，我最近正在休假中，无法亲自回复您的邮件。我将在假期结束后，尽快给您回复。