espnet2.legacy.nets.pytorch_backend.transformer.label_smoothing_loss.LabelSmoothingLoss

source

class espnet2.legacy.nets.pytorch_backend.transformer.label_smoothing_loss.LabelSmoothingLoss(size, padding_idx, smoothing, normalize_length=False, criterion=KLDivLoss())

Bases: Module

Label-smoothing loss.

• Parameters:
  • size (int) – the number of class
  • padding_idx (int) – ignored class id
  • smoothing (float) – smoothing rate (0.0 means the conventional CE)
  • normalize_length (bool) – normalize loss by sequence length if True
  • criterion (torch.nn.Module) – loss function to be smoothed

Construct an LabelSmoothingLoss object.

forward(x, target)

Compute loss between x and target.

• Parameters:
  • x (torch.Tensor) – prediction (batch, seqlen, class)
  • target (torch.Tensor) – target signal masked with self.padding_id (batch, seqlen)
• Returns: scalar float value

:rtype torch.Tensor