| import pdb | |
| import torch | |
| import torch.nn as nn | |
| import torch.nn.functional as F | |
| class SeqKD(nn.Module): | |
| """ | |
| NLL loss with label smoothing. | |
| """ | |
| def __init__(self, T=1): | |
| super(SeqKD, self).__init__() | |
| self.kdloss = nn.KLDivLoss(reduction='batchmean') | |
| self.T = T | |
| def forward(self, prediction_logits, ref_logits, use_blank=True): | |
| start_idx = 0 if use_blank else 1 | |
| prediction_logits = F.log_softmax(prediction_logits[:, :, start_idx:]/self.T, dim=-1) \ | |
| .view(-1, ref_logits.shape[2] - start_idx) | |
| ref_probs = F.softmax(ref_logits[:, :, start_idx:]/self.T, dim=-1) \ | |
| .view(-1, ref_logits.shape[2] - start_idx) | |
| loss = self.kdloss(prediction_logits, ref_probs)*self.T*self.T | |
| # mask_probs = F.softmax(ref_logits[:, :, 1:], dim=-1).view(-1, ref_logits.shape[2] - 1) | |
| # mask = torch.max(mask_probs, dim=1)[0] > 0.5 | |
| # if torch.sum(mask) != 0: | |
| # loss = torch.sum(torch.sum(loss, dim=1) * mask) / torch.sum(mask) | |
| # else: | |
| # loss = torch.sum(torch.sum(loss, dim=1) * mask) | |
| return loss | |