SoftTargetCrossEntropy

class SoftTargetCrossEntropy(*args, **kwargs)

Bases: Module

Cross-entropy loss with soft targets (from DeiT III / timm).

Works with Mixup/CutMix soft labels by computing -sum(target * log_softmax(logits)) per sample, then averaging over the batch. Avoids the multi-class gradient dilution issue that occurs with BCEWithLogitsLoss(reduction=’mean’).

Parameters:

• args (Any)

• kwargs (Any)

forward(logits, targets)

Compute soft-target cross-entropy: -sum(target * log_softmax(logits)).

Parameters:

• logits (Tensor)

• targets (Tensor)

Return type:

Tensor