fix bug, ys_pad_masked in sampler of paraformer

游雁

2023-01-30 cfb2fda87c29db780e595b75f2de1c7710ebadd2

fix bug, ys_pad_masked in sampler of paraformer

 funasr/models/e2e_asr_paraformer.py

@@ -499,11 +499,11 @@
    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