update onnx batch inference

维石

2024-07-23 925910c5995bac39e9c5907c40070a30edfabe60

update onnx batch inference

 runtime/python/libtorch/funasr_torch/sensevoice_bin.py

@@ -81,27 +81,94 @@
        self.lid_int_dict = {24884: 3, 24885: 4, 24888: 7, 24892: 11, 24896: 12, 24992: 13}
        self.textnorm_dict = {"withitn": 14, "woitn": 15}
        self.textnorm_int_dict = {25016: 14, 25017: 15}
        
    def _get_lid(self, lid):
        if lid in list(self.lid_dict.keys()):
            return self.lid_dict[lid]
        else:
            raise ValueError(
                f"The language {l} is not in {list(self.lid_dict.keys())}"
            )
            
    def _get_tnid(self, tnid):
        if tnid in list(self.textnorm_dict.keys()):
            return self.textnorm_dict[tnid]
        else:
            raise ValueError(
                f"The textnorm {tnid} is not in {list(self.textnorm_dict.keys())}"
            )
    
    def read_tags(self, language_input, textnorm_input):
        # handle language
        if isinstance(language_input, list):
            language_list = []
            for l in language_input:
                language_list.append(self._get_lid(l))
        elif isinstance(language_input, str):
            # if is existing file
            if os.path.exists(language_input):
                language_file = open(language_input, "r").readlines()
                language_list = [
                    self._get_lid(l.strip())
                    for l in language_file
                ]
            else:
                language_list = [self._get_lid(language_input)]
        else:
            raise ValueError(
                f"Unsupported type {type(language_input)} for language_input"
            )
        # handle textnorm
        if isinstance(textnorm_input, list):
            textnorm_list = []
            for tn in textnorm_input:
                textnorm_list.append(self._get_tnid(tn))
        elif isinstance(textnorm_input, str):
            # if is existing file
            if os.path.exists(textnorm_input):
                textnorm_file = open(textnorm_input, "r").readlines()
                textnorm_list = [
                    self._get_tnid(tn.strip())
                    for tn in textnorm_file
                ]
            else:
                textnorm_list = [self._get_tnid(textnorm_input)]
        else:
            raise ValueError(
                f"Unsupported type {type(textnorm_input)} for textnorm_input"
            )
        return language_list, textnorm_list

    def __call__(self, wav_content: Union[str, np.ndarray, List[str]], **kwargs) -> List:

        language = self.lid_dict[kwargs.get("language", "auto")]
        use_itn = kwargs.get("use_itn", False)
        textnorm = kwargs.get("text_norm", None)
        if textnorm is None:
            textnorm = "withitn" if use_itn else "woitn"
        textnorm = self.textnorm_dict[textnorm]

    def __call__(self, wav_content: Union[str, np.ndarray, List[str]], **kwargs):
        language_input = kwargs.get("language", "auto")
        textnorm_input = kwargs.get("textnorm", "woitn")
        language_list, textnorm_list = self.read_tags(language_input, textnorm_input)
        
        waveform_list = self.load_data(wav_content, self.frontend.opts.frame_opts.samp_freq)
        waveform_nums = len(waveform_list)
        
        assert len(language_list) == 1 or len(language_list) == waveform_nums, \
            "length of parsed language list should be 1 or equal to the number of waveforms"
        assert len(textnorm_list) == 1 or len(textnorm_list) == waveform_nums, \
            "length of parsed textnorm list should be 1 or equal to the number of waveforms"
        
        asr_res = []
        for beg_idx in range(0, waveform_nums, self.batch_size):
            end_idx = min(waveform_nums, beg_idx + self.batch_size)
            feats, feats_len = self.extract_feat(waveform_list[beg_idx:end_idx])
            _language_list = language_list[beg_idx:end_idx]
            _textnorm_list = textnorm_list[beg_idx:end_idx]
            B = feats.shape[0]
            if len(_language_list) == 1 and B != 1:
                _language_list = _language_list * B
            if len(_textnorm_list) == 1 and B != 1:
                _textnorm_list = _textnorm_list * B

            ctc_logits, encoder_out_lens = self.ort_infer(
                torch.Tensor(feats).to(self.device),
                torch.Tensor(feats_len).to(self.device),
                torch.tensor([language]).to(self.device),
                torch.tensor([textnorm]).to(self.device),
                torch.tensor([_language_list]).to(self.device),
                torch.tensor([_textnorm_list]).to(self.device),
            )
            # support batch_size=1 only currently
            x = ctc_logits[0, : encoder_out_lens[0].item(), :]

 runtime/python/onnxruntime/funasr_onnx/sensevoice_bin.py

@@ -88,39 +88,107 @@
        self.lid_int_dict = {24884: 3, 24885: 4, 24888: 7, 24892: 11, 24896: 12, 24992: 13}
        self.textnorm_dict = {"withitn": 14, "woitn": 15}
        self.textnorm_int_dict = {25016: 14, 25017: 15}
        
    def _get_lid(self, lid):
        if lid in list(self.lid_dict.keys()):
            return self.lid_dict[lid]
        else:
            raise ValueError(
                f"The language {l} is not in {list(self.lid_dict.keys())}"
            )
            
    def _get_tnid(self, tnid):
        if tnid in list(self.textnorm_dict.keys()):
            return self.textnorm_dict[tnid]
        else:
            raise ValueError(
                f"The textnorm {tnid} is not in {list(self.textnorm_dict.keys())}"
            )
    
    def read_tags(self, language_input, textnorm_input):
        # handle language
        if isinstance(language_input, list):
            language_list = []
            for l in language_input:
                language_list.append(self._get_lid(l))
        elif isinstance(language_input, str):
            # if is existing file
            if os.path.exists(language_input):
                language_file = open(language_input, "r").readlines()
                language_list = [
                    self._get_lid(l.strip())
                    for l in language_file
                ]
            else:
                language_list = [self._get_lid(language_input)]
        else:
            raise ValueError(
                f"Unsupported type {type(language_input)} for language_input"
            )
        # handle textnorm
        if isinstance(textnorm_input, list):
            textnorm_list = []
            for tn in textnorm_input:
                textnorm_list.append(self._get_tnid(tn))
        elif isinstance(textnorm_input, str):
            # if is existing file
            if os.path.exists(textnorm_input):
                textnorm_file = open(textnorm_input, "r").readlines()
                textnorm_list = [
                    self._get_tnid(tn.strip())
                    for tn in textnorm_file
                ]
            else:
                textnorm_list = [self._get_tnid(textnorm_input)]
        else:
            raise ValueError(
                f"Unsupported type {type(textnorm_input)} for textnorm_input"
            )
        return language_list, textnorm_list

    def __call__(self, wav_content: Union[str, np.ndarray, List[str]], **kwargs):

        language = self.lid_dict[kwargs.get("language", "auto")]
        use_itn = kwargs.get("use_itn", False)
        textnorm = kwargs.get("text_norm", None)
        if textnorm is None:
            textnorm = "withitn" if use_itn else "woitn"
        textnorm = self.textnorm_dict[textnorm]

        language_input = kwargs.get("language", "auto")
        textnorm_input = kwargs.get("textnorm", "woitn")
        language_list, textnorm_list = self.read_tags(language_input, textnorm_input)
        
        waveform_list = self.load_data(wav_content, self.frontend.opts.frame_opts.samp_freq)
        waveform_nums = len(waveform_list)
        
        assert len(language_list) == 1 or len(language_list) == waveform_nums, \
            "length of parsed language list should be 1 or equal to the number of waveforms"
        assert len(textnorm_list) == 1 or len(textnorm_list) == waveform_nums, \
            "length of parsed textnorm list should be 1 or equal to the number of waveforms"
        
        asr_res = []
        for beg_idx in range(0, waveform_nums, self.batch_size):
            end_idx = min(waveform_nums, beg_idx + self.batch_size)
            feats, feats_len = self.extract_feat(waveform_list[beg_idx:end_idx])
            _language_list = language_list[beg_idx:end_idx]
            _textnorm_list = textnorm_list[beg_idx:end_idx]
            B = feats.shape[0]
            if len(_language_list) == 1 and B != 1:
                _language_list = _language_list * B
            if len(_textnorm_list) == 1 and B != 1:
                _textnorm_list = _textnorm_list * B
            ctc_logits, encoder_out_lens = self.infer(
                feats,
                feats_len,
                np.array([language], dtype=np.int32),
                np.array([textnorm], dtype=np.int32),
                np.array(_language_list, dtype=np.int32),
                np.array(_textnorm_list, dtype=np.int32),
            )
            # back to torch.Tensor
            ctc_logits = torch.from_numpy(ctc_logits).float()
            # support batch_size=1 only currently
            x = ctc_logits[0, : encoder_out_lens[0].item(), :]
            yseq = x.argmax(dim=-1)
            yseq = torch.unique_consecutive(yseq, dim=-1)
            for b in range(feats.shape[0]):
                # back to torch.Tensor
                if isinstance(ctc_logits, np.ndarray):
                    ctc_logits = torch.from_numpy(ctc_logits).float()
                # support batch_size=1 only currently
                x = ctc_logits[b, : encoder_out_lens[b].item(), :]
                yseq = x.argmax(dim=-1)
                yseq = torch.unique_consecutive(yseq, dim=-1)

            mask = yseq != self.blank_id
            token_int = yseq[mask].tolist()
                mask = yseq != self.blank_id
                token_int = yseq[mask].tolist()

            asr_res.append(self.tokenizer.decode(token_int))
                asr_res.append(self.tokenizer.decode(token_int))

        return asr_res