update libtorch inference

shixian.shi

2023-03-20 a9683afbb2c89ac16b3d9e1d22b36e1e8934c5ae

update libtorch inference

 funasr/runtime/python/libtorch/torch_paraformer/paraformer_bin.py

@@ -62,26 +62,28 @@
                am_scores, valid_token_lens = outputs[0], outputs[1]
                if len(outputs) == 4:
                    # for BiCifParaformer Inference
                    us_alphas, us_cif_peak = outputs[2], outputs[3]
                    us_alphas, us_peaks = outputs[2], outputs[3]
                else:
                    us_alphas, us_cif_peak = None, None
                    us_alphas, us_peaks = None, None
            except:
                #logging.warning(traceback.format_exc())
                logging.warning("input wav is silence or noise")
                preds = ['']
            else:
                am_scores, valid_token_lens = am_scores.detach().cpu().numpy(), valid_token_lens.detach().cpu().numpy()
                preds = self.decode(am_scores, valid_token_lens)
                if us_cif_peak is None:
                if us_peaks is None:
                    for pred in preds:
                        pred = sentence_postprocess(pred)
                        asr_res.append({'preds': pred})
                else:
                    for pred, us_cif_peak_ in zip(preds, us_cif_peak):
                        text, tokens = pred
                        timestamp, timestamp_total = time_stamp_lfr6_onnx(us_cif_peak_, copy.copy(tokens))
                    for pred, us_peaks_ in zip(preds, us_peaks):
                        raw_tokens = pred
                        timestamp, timestamp_raw = time_stamp_lfr6_onnx(us_peaks_, copy.copy(raw_tokens))
                        text_proc, timestamp_proc, _ = sentence_postprocess(raw_tokens, timestamp_raw)
                        # logging.warning(timestamp)
                        if len(self.plot_timestamp_to):
                            self.plot_wave_timestamp(waveform_list[0], timestamp_total, self.plot_timestamp_to)
                        asr_res.append({'preds': text, 'timestamp': timestamp})
                            self.plot_wave_timestamp(waveform_list[0], timestamp, self.plot_timestamp_to)
                        asr_res.append({'preds': text_proc, 'timestamp': timestamp_proc, "raw_tokens": raw_tokens})
        return asr_res

    def plot_wave_timestamp(self, wav, text_timestamp, dest):
@@ -182,6 +184,6 @@
        # Change integer-ids to tokens
        token = self.converter.ids2tokens(token_int)
        token = token[:valid_token_num-self.pred_bias]
        texts = sentence_postprocess(token)
        return texts
        # texts = sentence_postprocess(token)
        return token