update funasr-onnx-offline

雾聪

2023-06-02 fae856e23d45fd27d5fd55fd036e8e3fc7b24915

 funasr/bin/asr_infer.py

@@ -9,6 +9,7 @@
import time
import copy
import os
import re
import codecs
import tempfile
import requests
@@ -304,6 +305,7 @@
            nbest: int = 1,
            frontend_conf: dict = None,
            hotword_list_or_file: str = None,
            decoding_ind: int = 0,
            **kwargs,
    ):
        assert check_argument_types()
@@ -414,6 +416,7 @@
        self.nbest = nbest
        self.frontend = frontend
        self.encoder_downsampling_factor = 1
        self.decoding_ind = decoding_ind
        if asr_train_args.encoder == "data2vec_encoder" or asr_train_args.encoder_conf["input_layer"] == "conv2d":
            self.encoder_downsampling_factor = 4

@@ -451,7 +454,7 @@
        batch = to_device(batch, device=self.device)

        # b. Forward Encoder
        enc, enc_len = self.asr_model.encode(**batch)
        enc, enc_len = self.asr_model.encode(**batch, ind=self.decoding_ind)
        if isinstance(enc, tuple):
            enc = enc[0]
        # assert len(enc) == 1, len(enc)
@@ -488,15 +491,20 @@

                nbest_hyps = nbest_hyps[: self.nbest]
            else:
                yseq = am_scores.argmax(dim=-1)
                score = am_scores.max(dim=-1)[0]
                score = torch.sum(score, dim=-1)
                # pad with mask tokens to ensure compatibility with sos/eos tokens
                yseq = torch.tensor(
                    [self.asr_model.sos] + yseq.tolist() + [self.asr_model.eos], device=yseq.device
                )
                if pre_token_length[i] == 0:
                    yseq = torch.tensor(
                        [self.asr_model.sos] + [self.asr_model.eos], device=yseq.device
                    )
                    score = torch.tensor(0.0, device=yseq.device)
                else:
                    yseq = am_scores.argmax(dim=-1)
                    score = am_scores.max(dim=-1)[0]
                    score = torch.sum(score, dim=-1)
                    # pad with mask tokens to ensure compatibility with sos/eos tokens
                    yseq = torch.tensor(
                        [self.asr_model.sos] + yseq.tolist() + [self.asr_model.eos], device=yseq.device
                    )
                nbest_hyps = [Hypothesis(yseq=yseq, score=score)]

            for hyp in nbest_hyps:
                assert isinstance(hyp, (Hypothesis)), type(hyp)

@@ -823,9 +831,16 @@

                # Change integer-ids to tokens
                token = self.converter.ids2tokens(token_int)
                token = " ".join(token)

                results.append(token)
                postprocessed_result = ""
                for item in token:
                    if item.endswith('@@'):
                        postprocessed_result += item[:-2]
                    elif re.match('^[a-zA-Z]+$', item):
                        postprocessed_result += item + " "
                    else:
                        postprocessed_result += item
                        
                results.append(postprocessed_result)

        # assert check_return_type(results)
        return results
@@ -1497,8 +1512,13 @@
        if isinstance(speech, np.ndarray):
            speech = torch.tensor(speech)
        
        feats = speech.unsqueeze(0).to(getattr(torch, self.dtype))
        feats_lengths = feats.new_full([1], dtype=torch.long, fill_value=feats.size(1))
        if self.frontend is not None:
            speech = torch.unsqueeze(speech, axis=0)
            speech_lengths = speech.new_full([1], dtype=torch.long, fill_value=speech.size(1))
            feats, feats_lengths = self.frontend(speech, speech_lengths)
        else:                
            feats = speech.unsqueeze(0).to(getattr(torch, self.dtype))
            feats_lengths = feats.new_full([1], dtype=torch.long, fill_value=feats.size(1))
        
        if self.asr_model.normalize is not None:
            feats, feats_lengths = self.asr_model.normalize(feats, feats_lengths)
@@ -1523,14 +1543,19 @@
        
        if isinstance(speech, np.ndarray):
            speech = torch.tensor(speech)
        
        feats = speech.unsqueeze(0).to(getattr(torch, self.dtype))
        feats_lengths = feats.new_full([1], dtype=torch.long, fill_value=feats.size(1))

        if self.frontend is not None:
            speech = torch.unsqueeze(speech, axis=0)
            speech_lengths = speech.new_full([1], dtype=torch.long, fill_value=speech.size(1))
            feats, feats_lengths = self.frontend(speech, speech_lengths)
        else:                
            feats = speech.unsqueeze(0).to(getattr(torch, self.dtype))
            feats_lengths = feats.new_full([1], dtype=torch.long, fill_value=feats.size(1))
        
        feats = to_device(feats, device=self.device)
        feats_lengths = to_device(feats_lengths, device=self.device)
        
        enc_out, _ = self.asr_model.encoder(feats, feats_lengths)
        enc_out, _, _ = self.asr_model.encoder(feats, feats_lengths)
        
        nbest_hyps = self.beam_search(enc_out[0])