| | |
|---|
| | | xs_pad, masks = self.embed(xs_pad, masks) |
|---|
| | | else: |
|---|
| | | xs_pad = self.embed(xs_pad) |
|---|
| | | pdb.set_trace() |
|---|
| | | |
|---|
| | | intermediate_outs = [] |
|---|
| | | if len(self.interctc_layer_idx) == 0: |
|---|
| | | xs_pad, masks = self.encoders(xs_pad, masks) |
|---|
| | |
|---|
| | | xs_pad = (x, pos_emb) |
|---|
| | | else: |
|---|
| | | xs_pad = xs_pad + self.conditioning_layer(ctc_out) |
|---|
| | | pdb.set_trace() |
|---|
| | | |
|---|
| | | if isinstance(xs_pad, tuple): |
|---|
| | | xs_pad = xs_pad[0] |
|---|
| | | if self.normalize_before: |
|---|
| | | xs_pad = self.after_norm(xs_pad) |
|---|
| | | pdb.set_trace() |
|---|
| | | |
|---|
| | | olens = masks.squeeze(1).sum(1) |
|---|
| | | if len(intermediate_outs) > 0: |
|---|
| | | return (xs_pad, intermediate_outs), olens, None |
|---|
| | |
|---|
| | | |
|---|
| | | if intermediate_outs is not None: |
|---|
| | | return (encoder_out, intermediate_outs), encoder_out_lens |
|---|
| | | pdb.set_trace() |
|---|
| | | return encoder_out, encoder_out_lens |
|---|
| | | |
|---|
| | | def _calc_att_loss( |
|---|
| | |
|---|
| | | encoder_out, encoder_out_lens = self.encode(speech, speech_lengths) |
|---|
| | | if isinstance(encoder_out, tuple): |
|---|
| | | encoder_out = encoder_out[0] |
|---|
| | | |
|---|
| | | pdb.set_trace() |
|---|
| | | ocr = ocr_sample_list[0] |
|---|
| | | ocr_lengths = ocr.new_full([1], dtype=torch.long, fill_value=ocr.size(1)) |
|---|
| | | ocr, ocr_lens, _ = self.text_encoder(ocr, ocr_lengths) |
|---|
| | | pdb.set_trace() |
|---|
| | | # c. Passed the encoder result and the beam search |
|---|
| | | nbest_hyps = self.beam_search( |
|---|
| | | x=encoder_out[0], maxlenratio=kwargs.get("maxlenratio", 0.0), minlenratio=kwargs.get("minlenratio", 0.0) |
|---|
