espnet.nets.pytorch_backend.maskctc.mask.square_mask

source

espnet.nets.pytorch_backend.maskctc.mask.square_mask(ys_in_pad, ignore_id)

Create attention mask to avoid attending on padding tokens.

• Parameters:
  • ys_pad (torch.Tensor) – batch of padded target sequences (B, Lmax)
  • ignore_id (int) – index of padding
  • dtype (torch.dtype) – result dtype
• Return type: torch.Tensor (B, Lmax, Lmax)