Sampler. Args: encoder_out: Encoder output tensor. encoder_out_lens: Encoder output lengths. ys_pad: TODO. ys_pad_lens: Lengths of ys_pad. pre_acoustic_embeds: TODO.
(self, encoder_out, encoder_out_lens, ys_pad, ys_pad_lens, pre_acoustic_embeds)
| 406 | return loss_att, acc_att, cer_att, wer_att, loss_pre, pre_loss_att |
| 407 | |
| 408 | def sampler(self, encoder_out, encoder_out_lens, ys_pad, ys_pad_lens, pre_acoustic_embeds): |
| 409 | |
| 410 | """Sampler. |
| 411 | |
| 412 | Args: |
| 413 | encoder_out: Encoder output tensor. |
| 414 | encoder_out_lens: Encoder output lengths. |
| 415 | ys_pad: TODO. |
| 416 | ys_pad_lens: Lengths of ys_pad. |
| 417 | pre_acoustic_embeds: TODO. |
| 418 | """ |
| 419 | tgt_mask = (~make_pad_mask(ys_pad_lens, maxlen=ys_pad_lens.max())[:, :, None]).to( |
| 420 | ys_pad.device |
| 421 | ) |
| 422 | ys_pad_masked = ys_pad * tgt_mask[:, :, 0] |
| 423 | if self.share_embedding: |
| 424 | ys_pad_embed = self.decoder.output_layer.weight[ys_pad_masked] |
| 425 | else: |
| 426 | ys_pad_embed = self.decoder.embed(ys_pad_masked) |
| 427 | with torch.no_grad(): |
| 428 | decoder_outs = self.decoder( |
| 429 | encoder_out, encoder_out_lens, pre_acoustic_embeds, ys_pad_lens |
| 430 | ) |
| 431 | decoder_out, _ = decoder_outs[0], decoder_outs[1] |
| 432 | pred_tokens = decoder_out.argmax(-1) |
| 433 | nonpad_positions = ys_pad.ne(self.ignore_id) |
| 434 | seq_lens = (nonpad_positions).sum(1) |
| 435 | same_num = ((pred_tokens == ys_pad) & nonpad_positions).sum(1) |
| 436 | input_mask = torch.ones_like(nonpad_positions) |
| 437 | bsz, seq_len = ys_pad.size() |
| 438 | for li in range(bsz): |
| 439 | target_num = ( |
| 440 | ((seq_lens[li] - same_num[li].sum()).float()) * self.sampling_ratio |
| 441 | ).long() |
| 442 | if target_num > 0: |
| 443 | input_mask[li].scatter_( |
| 444 | dim=0, |
| 445 | index=torch.randperm(seq_lens[li])[:target_num].to(input_mask.device), |
| 446 | value=0, |
| 447 | ) |
| 448 | input_mask = input_mask.eq(1) |
| 449 | input_mask = input_mask.masked_fill(~nonpad_positions, False) |
| 450 | input_mask_expand_dim = input_mask.unsqueeze(2).to(pre_acoustic_embeds.device) |
| 451 | |
| 452 | sematic_embeds = pre_acoustic_embeds.masked_fill( |
| 453 | ~input_mask_expand_dim, 0 |
| 454 | ) + ys_pad_embed.masked_fill(input_mask_expand_dim, 0) |
| 455 | return sematic_embeds * tgt_mask, decoder_out * tgt_mask |
| 456 | |
| 457 | def _calc_ctc_loss( |
| 458 | self, |
no test coverage detected