Merge pull request #47 from alibaba-damo-academy/dev
fix bug, ys_pad_masked in sampler of paraformer
| | |
|---|
| | | def sampler(self, encoder_out, encoder_out_lens, ys_pad, ys_pad_lens, pre_acoustic_embeds): |
|---|
| | | |
|---|
| | | tgt_mask = (~make_pad_mask(ys_pad_lens, maxlen=ys_pad_lens.max())[:, :, None]).to(ys_pad.device) |
|---|
| | | ys_pad = ys_pad * tgt_mask[:, :, 0] |
|---|
| | | ys_pad_masked = ys_pad * tgt_mask[:, :, 0] |
|---|
| | | if self.share_embedding: |
|---|
| | | ys_pad_embed = self.decoder.output_layer.weight[ys_pad] |
|---|
| | | ys_pad_embed = self.decoder.output_layer.weight[ys_pad_masked] |
|---|
| | | else: |
|---|
| | | ys_pad_embed = self.decoder.embed(ys_pad) |
|---|
| | | ys_pad_embed = self.decoder.embed(ys_pad_masked) |
|---|
| | | with torch.no_grad(): |
|---|
| | | decoder_outs = self.decoder( |
|---|
| | | encoder_out, encoder_out_lens, pre_acoustic_embeds, ys_pad_lens |
|---|
