Merge pull request #58 from alibaba-damo-academy/dev

zhifu gao

2023-02-03 79bd015ab0ded4e5aed1b1ecf32fcbc84eefde68

 funasr/bin/asr_inference_paraformer.py

@@ -95,10 +95,13 @@
        logging.info("asr_train_args: {}".format(asr_train_args))
        asr_model.to(dtype=getattr(torch, dtype)).eval()

        ctc = CTCPrefixScorer(ctc=asr_model.ctc, eos=asr_model.eos)
        if asr_model.ctc != None:
            ctc = CTCPrefixScorer(ctc=asr_model.ctc, eos=asr_model.eos)
            scorers.update(
                ctc=ctc
            )
        token_list = asr_model.token_list
        scorers.update(
            ctc=ctc,
            length_bonus=LengthBonus(len(token_list)),
        )

@@ -166,7 +169,7 @@
        self.converter = converter
        self.tokenizer = tokenizer
        is_use_lm = lm_weight != 0.0 and lm_file is not None
        if ctc_weight == 0.0 and not is_use_lm:
        if (ctc_weight == 0.0 or asr_model.ctc == None) and not is_use_lm:
            beam_search = None
        self.beam_search = beam_search
        logging.info(f"Beam_search: {self.beam_search}")
@@ -224,6 +227,8 @@
        pre_acoustic_embeds, pre_token_length, alphas, pre_peak_index = predictor_outs[0], predictor_outs[1], \
                                                                        predictor_outs[2], predictor_outs[3]
        pre_token_length = pre_token_length.round().long()
        if torch.max(pre_token_length) < 1:
            return []
        decoder_outs = self.asr_model.cal_decoder_with_predictor(enc, enc_len, pre_acoustic_embeds, pre_token_length)
        decoder_out, ys_pad_lens = decoder_outs[0], decoder_outs[1]

@@ -259,7 +264,7 @@
                    token_int = hyp.yseq[1:last_pos].tolist()

                # remove blank symbol id, which is assumed to be 0
                token_int = list(filter(lambda x: x != 0, token_int))
                token_int = list(filter(lambda x: x != 0 and x != 2, token_int))

                # Change integer-ids to tokens
                token = self.converter.ids2tokens(token_int)
@@ -391,7 +396,7 @@
#         results = speech2text(**batch)
#         if len(results) < 1:
#             hyp = Hypothesis(score=0.0, scores={}, states={}, yseq=[])
#             results = [[" ", ["<space>"], [2], hyp, 10, 6]] * nbest
#             results = [[" ", ["sil"], [2], hyp, 10, 6]] * nbest
#         time_end = time.time()
#         forward_time = time_end - time_beg
#         lfr_factor = results[0][-1]
@@ -524,6 +529,7 @@
        nbest: int = 1,
        num_workers: int = 1,
        output_dir: Optional[str] = None,
        param_dict: dict = None,
        **kwargs,
):
    assert check_argument_types()
@@ -573,6 +579,7 @@
            data_path_and_name_and_type,
            raw_inputs: Union[np.ndarray, torch.Tensor] = None,
            output_dir_v2: Optional[str] = None,
            param_dict: dict = None,
    ):
        # 3. Build data-iterator
        if data_path_and_name_and_type is None and raw_inputs is not None:
@@ -618,7 +625,7 @@
            results = speech2text(**batch)
            if len(results) < 1:
                hyp = Hypothesis(score=0.0, scores={}, states={}, yseq=[])
                results = [[" ", ["<space>"], [2], hyp, 10, 6]] * nbest
                results = [[" ", ["sil"], [2], hyp, 10, 6]] * nbest
            time_end = time.time()
            forward_time = time_end - time_beg
            lfr_factor = results[0][-1]
@@ -650,7 +657,7 @@
                        finish_count += 1
                        # asr_utils.print_progress(finish_count / file_count)
                        if writer is not None:
                            ibest_writer["text"][key] = text
                            ibest_writer["text"][key] = text_postprocessed

                    logging.info("decoding, utt: {}, predictions: {}".format(key, text))
        rtf_avg = "decoding, feature length total: {}, forward_time total: {:.4f}, rtf avg: {:.4f}".format(length_total, forward_time_total, 100 * forward_time_total / (length_total * lfr_factor))