Rename 'res' in line 514 to avoid with naming conflict with line 365

liugz18

2024-07-18 d80ac2fd2df4e7fb8a28acfa512bb11472b5cc99

 funasr/models/whisper_lid/model.py

@@ -22,7 +22,6 @@
class OpenAIWhisperModel(nn.Module):
    """CTC-attention hybrid Encoder-Decoder model"""

    
    def __init__(
        self,
        specaug: str = None,
@@ -71,14 +70,12 @@
            decoder_class = tables.decoder_classes.get(decoder)
            decoder = decoder_class(decoder_conf)
        if ctc_weight > 0.0:
            

            if ctc_conf is None:
                ctc_conf = {}
            
            ctc = CTC(
                odim=vocab_size, encoder_output_size=encoder_output_size, **ctc_conf
            )
    

            ctc = CTC(odim=vocab_size, encoder_output_size=encoder_output_size, **ctc_conf)

        self.blank_id = blank_id
        self.sos = sos if sos is not None else vocab_size - 1
        self.eos = eos if eos is not None else vocab_size - 1
@@ -102,7 +99,7 @@
            self.decoder = None
        else:
            self.decoder = decoder
        

        self.criterion_att = LabelSmoothingLoss(
            size=vocab_size,
            padding_idx=ignore_id,
@@ -120,14 +117,14 @@
            self.ctc = None
        else:
            self.ctc = ctc
            

        self.share_embedding = share_embedding
        if self.share_embedding:
            self.decoder.embed = None
        

        self.length_normalized_loss = length_normalized_loss
        self.beam_search = None
    

    def forward(
        self,
        speech: torch.Tensor,
@@ -143,36 +140,34 @@
                text: (Batch, Length)
                text_lengths: (Batch,)
        """
        # import pdb;
        # pdb.set_trace()
        if len(text_lengths.size()) > 1:
            text_lengths = text_lengths[:, 0]
        if len(speech_lengths.size()) > 1:
            speech_lengths = speech_lengths[:, 0]
        

        batch_size = speech.shape[0]
        

        # 1. Encoder
        encoder_out, encoder_out_lens = self.encode(speech, speech_lengths)
        intermediate_outs = None
        if isinstance(encoder_out, tuple):
            intermediate_outs = encoder_out[1]
            encoder_out = encoder_out[0]
        

        loss_att, acc_att, cer_att, wer_att = None, None, None, None
        loss_ctc, cer_ctc = None, None
        stats = dict()
        

        # decoder: CTC branch
        if self.ctc_weight != 0.0:
            loss_ctc, cer_ctc = self._calc_ctc_loss(
                encoder_out, encoder_out_lens, text, text_lengths
            )
            

            # Collect CTC branch stats
            stats["loss_ctc"] = loss_ctc.detach() if loss_ctc is not None else None
            stats["cer_ctc"] = cer_ctc
        

        # Intermediate CTC (optional)
        loss_interctc = 0.0
        if self.interctc_weight != 0.0 and intermediate_outs is not None:
@@ -183,25 +178,23 @@
                    intermediate_out, encoder_out_lens, text, text_lengths
                )
                loss_interctc = loss_interctc + loss_ic
                

                # Collect Intermedaite CTC stats
                stats["loss_interctc_layer{}".format(layer_idx)] = (
                    loss_ic.detach() if loss_ic is not None else None
                )
                stats["cer_interctc_layer{}".format(layer_idx)] = cer_ic
            

            loss_interctc = loss_interctc / len(intermediate_outs)
            

            # calculate whole encoder loss
            loss_ctc = (
                           1 - self.interctc_weight
                       ) * loss_ctc + self.interctc_weight * loss_interctc
        
            loss_ctc = (1 - self.interctc_weight) * loss_ctc + self.interctc_weight * loss_interctc

        # decoder: Attention decoder branch
        loss_att, acc_att, cer_att, wer_att = self._calc_att_loss(
            encoder_out, encoder_out_lens, text, text_lengths
        )
        

        # 3. CTC-Att loss definition
        if self.ctc_weight == 0.0:
            loss = loss_att
@@ -209,25 +202,27 @@
            loss = loss_ctc
        else:
            loss = self.ctc_weight * loss_ctc + (1 - self.ctc_weight) * loss_att
        

        # Collect Attn branch stats
        stats["loss_att"] = loss_att.detach() if loss_att is not None else None
        stats["acc"] = acc_att
        stats["cer"] = cer_att
        stats["wer"] = wer_att
        

        # Collect total loss stats
        stats["loss"] = torch.clone(loss.detach())
        

        # force_gatherable: to-device and to-tensor if scalar for DataParallel
        if self.length_normalized_loss:
            batch_size = int((text_lengths + 1).sum())
        loss, stats, weight = force_gatherable((loss, stats, batch_size), loss.device)
        return loss, stats, weight
    

    def encode(
        self, speech: torch.Tensor, speech_lengths: torch.Tensor, **kwargs,
        self,
        speech: torch.Tensor,
        speech_lengths: torch.Tensor,
        **kwargs,
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """Frontend + Encoder. Note that this method is used by asr_inference.py
        Args:
@@ -240,30 +235,28 @@
            # Data augmentation
            if self.specaug is not None and self.training:
                speech, speech_lengths = self.specaug(speech, speech_lengths)
            

            # Normalization for feature: e.g. Global-CMVN, Utterance-CMVN
            if self.normalize is not None:
                speech, speech_lengths = self.normalize(speech, speech_lengths)
        

        # Forward encoder
        # feats: (Batch, Length, Dim)
        # -> encoder_out: (Batch, Length2, Dim2)
        if self.encoder.interctc_use_conditioning:
            encoder_out, encoder_out_lens, _ = self.encoder(
                speech, speech_lengths, ctc=self.ctc
            )
            encoder_out, encoder_out_lens, _ = self.encoder(speech, speech_lengths, ctc=self.ctc)
        else:
            encoder_out, encoder_out_lens, _ = self.encoder(speech, speech_lengths)
        intermediate_outs = None
        if isinstance(encoder_out, tuple):
            intermediate_outs = encoder_out[1]
            encoder_out = encoder_out[0]
        

        if intermediate_outs is not None:
            return (encoder_out, intermediate_outs), encoder_out_lens
        

        return encoder_out, encoder_out_lens
    

    def _calc_att_loss(
        self,
        encoder_out: torch.Tensor,
@@ -273,12 +266,10 @@
    ):
        ys_in_pad, ys_out_pad = add_sos_eos(ys_pad, self.sos, self.eos, self.ignore_id)
        ys_in_lens = ys_pad_lens + 1
        

        # 1. Forward decoder
        decoder_out, _ = self.decoder(
            encoder_out, encoder_out_lens, ys_in_pad, ys_in_lens
        )
        
        decoder_out, _ = self.decoder(encoder_out, encoder_out_lens, ys_in_pad, ys_in_lens)

        # 2. Compute attention loss
        loss_att = self.criterion_att(decoder_out, ys_out_pad)
        acc_att = th_accuracy(
@@ -286,16 +277,16 @@
            ys_out_pad,
            ignore_label=self.ignore_id,
        )
        

        # Compute cer/wer using attention-decoder
        if self.training or self.error_calculator is None:
            cer_att, wer_att = None, None
        else:
            ys_hat = decoder_out.argmax(dim=-1)
            cer_att, wer_att = self.error_calculator(ys_hat.cpu(), ys_pad.cpu())
        

        return loss_att, acc_att, cer_att, wer_att
    

    def _calc_ctc_loss(
        self,
        encoder_out: torch.Tensor,
@@ -305,41 +296,39 @@
    ):
        # Calc CTC loss
        loss_ctc = self.ctc(encoder_out, encoder_out_lens, ys_pad, ys_pad_lens)
        

        # Calc CER using CTC
        cer_ctc = None
        if not self.training and self.error_calculator is not None:
            ys_hat = self.ctc.argmax(encoder_out).data
            cer_ctc = self.error_calculator(ys_hat.cpu(), ys_pad.cpu(), is_ctc=True)
        return loss_ctc, cer_ctc
    
    def init_beam_search(self,
                         **kwargs,
                         ):

    def init_beam_search(
        self,
        **kwargs,
    ):
        from funasr.models.transformer.search import BeamSearch
        from funasr.models.transformer.scorers.ctc import CTCPrefixScorer
        from funasr.models.transformer.scorers.length_bonus import LengthBonus
    

        # 1. Build ASR model
        scorers = {}
        

        if self.ctc != None:
            ctc = CTCPrefixScorer(ctc=self.ctc, eos=self.eos)
            scorers.update(
                ctc=ctc
            )
            scorers.update(ctc=ctc)
        token_list = kwargs.get("token_list")
        scorers.update(
            decoder=self.decoder,
            length_bonus=LengthBonus(len(token_list)),
        )

        
        # 3. Build ngram model
        # ngram is not supported now
        ngram = None
        scorers["ngram"] = ngram
        

        weights = dict(
            decoder=1.0 - kwargs.get("decoding_ctc_weight", 0.5),
            ctc=kwargs.get("decoding_ctc_weight", 0.5),
@@ -359,19 +348,20 @@
        )

        self.beam_search = beam_search
        
    def inference(self,
             data_in,
             data_lengths=None,
             key: list=None,
             tokenizer=None,
             frontend=None,
             **kwargs,
             ):
        

    def inference(
        self,
        data_in,
        data_lengths=None,
        key: list = None,
        tokenizer=None,
        frontend=None,
        **kwargs,
    ):

        if kwargs.get("batch_size", 1) > 1:
            raise NotImplementedError("batch decoding is not implemented")
        

        # init beamsearch
        if self.beam_search is None:
            logging.info("enable beam_search")
@@ -379,7 +369,9 @@
            self.nbest = kwargs.get("nbest", 1)

        meta_data = {}
        if isinstance(data_in, torch.Tensor) and kwargs.get("data_type", "sound") == "fbank":  # fbank
        if (
            isinstance(data_in, torch.Tensor) and kwargs.get("data_type", "sound") == "fbank"
        ):  # fbank
            speech, speech_lengths = data_in, data_lengths
            if len(speech.shape) < 3:
                speech = speech[None, :, :]
@@ -388,16 +380,23 @@
        else:
            # extract fbank feats
            time1 = time.perf_counter()
            audio_sample_list = load_audio_text_image_video(data_in, fs=frontend.fs, audio_fs=kwargs.get("fs", 16000),
                                                            data_type=kwargs.get("data_type", "sound"),
                                                            tokenizer=tokenizer)
            audio_sample_list = load_audio_text_image_video(
                data_in,
                fs=frontend.fs,
                audio_fs=kwargs.get("fs", 16000),
                data_type=kwargs.get("data_type", "sound"),
                tokenizer=tokenizer,
            )
            time2 = time.perf_counter()
            meta_data["load_data"] = f"{time2 - time1:0.3f}"
            speech, speech_lengths = extract_fbank(audio_sample_list, data_type=kwargs.get("data_type", "sound"),
                                                   frontend=frontend)
            speech, speech_lengths = extract_fbank(
                audio_sample_list, data_type=kwargs.get("data_type", "sound"), frontend=frontend
            )
            time3 = time.perf_counter()
            meta_data["extract_feat"] = f"{time3 - time2:0.3f}"
            meta_data["batch_data_time"] = speech_lengths.sum().item() * frontend.frame_shift * frontend.lfr_n / 1000
            meta_data["batch_data_time"] = (
                speech_lengths.sum().item() * frontend.frame_shift * frontend.lfr_n / 1000
            )

        speech = speech.to(device=kwargs["device"])
        speech_lengths = speech_lengths.to(device=kwargs["device"])
@@ -405,14 +404,15 @@
        encoder_out, encoder_out_lens = self.encode(speech, speech_lengths)
        if isinstance(encoder_out, tuple):
            encoder_out = encoder_out[0]
        

        # 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)
            x=encoder_out[0],
            maxlenratio=kwargs.get("maxlenratio", 0.0),
            minlenratio=kwargs.get("minlenratio", 0.0),
        )
        
        nbest_hyps = nbest_hyps[: self.nbest]

        nbest_hyps = nbest_hyps[: self.nbest]

        results = []
        b, n, d = encoder_out.size()
@@ -424,29 +424,33 @@
                    if not hasattr(self, "writer"):
                        self.writer = DatadirWriter(kwargs.get("output_dir"))
                    ibest_writer = self.writer[f"{nbest_idx + 1}best_recog"]
                    

                # remove sos/eos and get results
                last_pos = -1
                if isinstance(hyp.yseq, list):
                    token_int = hyp.yseq[1:last_pos]
                else:
                    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 != self.eos and x != self.sos and x != self.blank_id, token_int))
                
                token_int = list(
                    filter(
                        lambda x: x != self.eos and x != self.sos and x != self.blank_id, token_int
                    )
                )

                # Change integer-ids to tokens
                token = tokenizer.ids2tokens(token_int)
                text = tokenizer.tokens2text(token)
                

                text_postprocessed, _ = postprocess_utils.sentence_postprocess(token)
                result_i = {"key": key[i], "token": token, "text": text_postprocessed}
                results.append(result_i)
                

                if ibest_writer is not None:
                    ibest_writer["token"][key[i]] = " ".join(token)
                    ibest_writer["text"][key[i]] = text_postprocessed
        

        return results, meta_data


@@ -455,18 +459,18 @@
    """WhisperEncoder and EResNet based LID Model"""

    def __init__(
            self,
            vocab_size: int,
            specaug: str = None,
            specaug_conf: dict = None,
            encoder: str = None,
            encoder_conf: dict = None,
            lid_predictor:  str = None,
            lid_predictor_conf: dict = None,
            proj_dim: int = None,
            clip_frames: int = None,
            random_clip: bool = False,
            **kwargs,
        self,
        vocab_size: int,
        specaug: str = None,
        specaug_conf: dict = None,
        encoder: str = None,
        encoder_conf: dict = None,
        lid_predictor: str = None,
        lid_predictor_conf: dict = None,
        proj_dim: int = None,
        clip_frames: int = None,
        random_clip: bool = False,
        **kwargs,
    ):
        super().__init__()
        if specaug is not None:
@@ -477,7 +481,7 @@
        lid_predictor_class = tables.lid_predictor_classes.get(lid_predictor)
        lid_predictor = lid_predictor_class(**lid_predictor_conf)
        if encoder.output_size() != proj_dim:
            self.proj_layer =  torch.nn.Linear(encoder.output_size(), proj_dim)
            self.proj_layer = torch.nn.Linear(encoder.output_size(), proj_dim)
        else:
            self.proj_layer = None
        self.output_layer = torch.nn.Linear(lid_predictor.output_size(), vocab_size)
@@ -498,23 +502,32 @@
        if not hasattr(self.encoder, "interctc_use_conditioning"):
            self.encoder.interctc_use_conditioning = False

    def forward(self,
                speech: torch.Tensor,  # may be padding
                speech_lengths: torch.Tensor,  # actual length
                lid: torch.Tensor,  # lid label, (batch_size, 1)
                lid_lengths: torch.Tensor,
                ):
    def forward(
        self,
        speech: torch.Tensor,  # may be padding
        speech_lengths: torch.Tensor,  # actual length
        lid: torch.Tensor,  # lid label, (batch_size, 1)
        lid_lengths: torch.Tensor,
    ):
        assert lid.shape[1] == 1
        batch_size = speech.shape[0]
        encoder_out, encoder_out_lens = self.encode(speech, speech_lengths)

        # re-generate encoder_out
        if self.clip_frames is None:
            reduced_encoder_out = torch.zeros(batch_size, encoder_out_lens.max(), encoder_out.shape[-1]).to(encoder_out.dtype).to(encoder_out.device)
            reduced_encoder_out = (
                torch.zeros(batch_size, encoder_out_lens.max(), encoder_out.shape[-1])
                .to(encoder_out.dtype)
                .to(encoder_out.device)
            )
            for i, enc_length in enumerate(encoder_out_lens):
                reduced_encoder_out[i, :enc_length] = encoder_out[i, :enc_length]
        else:
            reduced_encoder_out = torch.zeros(batch_size, self.clip_frames, encoder_out.shape[-1]).to(encoder_out.dtype).to(encoder_out.device)
            reduced_encoder_out = (
                torch.zeros(batch_size, self.clip_frames, encoder_out.shape[-1])
                .to(encoder_out.dtype)
                .to(encoder_out.device)
            )
            if self.random_clip:
                for i, enc_length in enumerate(encoder_out_lens):
                    if enc_length <= self.clip_frames:
@@ -523,7 +536,9 @@
                    else:
                        max_start_index = enc_length.item() - self.clip_frames
                        start_index = np.random.randint(0, max_start_index + 1)
                        reduced_encoder_out[i, :self.clip_frames] = encoder_out[i, start_index:start_index + self.clip_frames]
                        reduced_encoder_out[i, : self.clip_frames] = encoder_out[
                            i, start_index : start_index + self.clip_frames
                        ]
                        encoder_out_lens[i] = self.clip_frames
            else:
                for i, enc_length in enumerate(encoder_out_lens):
@@ -548,7 +563,7 @@
        return loss, stats, weight

    def encode(
            self, speech: torch.Tensor, speech_lengths: torch.Tensor
        self, speech: torch.Tensor, speech_lengths: torch.Tensor
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """Frontend + Encoder. Note that this method is used by asr_inference.py
        Args:
@@ -573,9 +588,7 @@
        # feats: (Batch, Length, Dim)
        # -> encoder_out: (Batch, Length2, Dim2)
        if self.encoder.interctc_use_conditioning:
            encoder_out, encoder_out_lens, _ = self.encoder(
                speech, speech_lengths, ctc=self.ctc
            )
            encoder_out, encoder_out_lens, _ = self.encoder(speech, speech_lengths, ctc=self.ctc)
        else:
            encoder_out, encoder_out_lens, _ = self.encoder(speech, speech_lengths)
        intermediate_outs = None
@@ -588,20 +601,23 @@

        return encoder_out, encoder_out_lens

    def inference(self,
                  data_in,
                  data_lengths=None,
                  key: list = None,
                  tokenizer=None,
                  frontend=None,
                  **kwargs,
                  ):
    def inference(
        self,
        data_in,
        data_lengths=None,
        key: list = None,
        tokenizer=None,
        frontend=None,
        **kwargs,
    ):

        if kwargs.get("batch_size", 1) > 1:
            raise NotImplementedError("batch decoding is not implemented")

        meta_data = {}
        if isinstance(data_in, torch.Tensor) and kwargs.get("data_type", "sound") == "fbank":  # fbank
        if (
            isinstance(data_in, torch.Tensor) and kwargs.get("data_type", "sound") == "fbank"
        ):  # fbank
            speech, speech_lengths = data_in, data_lengths
            if len(speech.shape) < 3:
                speech = speech[None, :, :]
@@ -610,39 +626,60 @@
        else:
            # extract fbank feats
            time1 = time.perf_counter()
            audio_sample_list = load_audio_text_image_video(data_in, fs=frontend.fs, audio_fs=kwargs.get("fs", 16000),
                                                            data_type=kwargs.get("data_type", "sound"),
                                                            tokenizer=tokenizer)
            audio_sample_list = load_audio_text_image_video(
                data_in,
                fs=frontend.fs,
                audio_fs=kwargs.get("fs", 16000),
                data_type=kwargs.get("data_type", "sound"),
                tokenizer=tokenizer,
            )
            time2 = time.perf_counter()
            meta_data["load_data"] = f"{time2 - time1:0.3f}"
            speech, speech_lengths = extract_fbank(audio_sample_list, data_type=kwargs.get("data_type", "sound"),
                                                   frontend=frontend)
            speech, speech_lengths = extract_fbank(
                audio_sample_list, data_type=kwargs.get("data_type", "sound"), frontend=frontend
            )
            time3 = time.perf_counter()
            meta_data["extract_feat"] = f"{time3 - time2:0.3f}"
            meta_data["batch_data_time"] = speech_lengths.sum().item() * frontend.frame_shift * frontend.lfr_n / 1000
            meta_data["batch_data_time"] = (
                speech_lengths.sum().item() * frontend.frame_shift * frontend.lfr_n / 1000
            )

        speech = speech.to(device=kwargs["device"])
        speech_lengths = speech_lengths.to(device=kwargs["device"])
        # Encoder
        enc, enc_out_lens  = self.encode(speech, speech_lengths)
        enc, enc_out_lens = self.encode(speech, speech_lengths)

        inference_clip_length = kwargs.get("inference_clip_length", None)
        if self.clip_frames is not None:
            if inference_clip_length is None:
                reduced_enc = torch.zeros(enc.shape[0], self.clip_frames, enc.shape[-1]).to(enc.dtype).to(enc.device)
                reduced_enc = (
                    torch.zeros(enc.shape[0], self.clip_frames, enc.shape[-1])
                    .to(enc.dtype)
                    .to(enc.device)
                )
                for i, enc_length in enumerate(enc_out_lens):
                    enc_length = self.clip_frames if enc_length >= self.clip_frames else enc_length
                    reduced_enc[i, :enc_length] = enc[i, :enc_length]
                    enc_out_lens[i] = enc_length
            else:
                assert inference_clip_length > 0, "inference_clip_length must be larger than 0"
                reduced_enc = torch.zeros(enc.shape[0], inference_clip_length, enc.shape[-1]).to(enc.dtype).to(enc.device)
                reduced_enc = (
                    torch.zeros(enc.shape[0], inference_clip_length, enc.shape[-1])
                    .to(enc.dtype)
                    .to(enc.device)
                )
                for i, enc_length in enumerate(enc_out_lens):
                    enc_length = inference_clip_length if enc_length >= inference_clip_length else enc_length
                    enc_length = (
                        inference_clip_length if enc_length >= inference_clip_length else enc_length
                    )
                    reduced_enc[i, :enc_length] = enc[i, :enc_length]
                    enc_out_lens[i] = enc_length
        else:
            reduced_enc = torch.zeros(enc.shape[0], enc_out_lens.max(), enc.shape[-1]).to(enc.dtype).to(enc.device)
            reduced_enc = (
                torch.zeros(enc.shape[0], enc_out_lens.max(), enc.shape[-1])
                .to(enc.dtype)
                .to(enc.device)
            )
            for i, enc_length in enumerate(enc_out_lens):
                reduced_enc[i, :enc_length] = enc[i, :enc_length]

@@ -662,4 +699,4 @@

        results = [{"key": key[0], "lid": predicted_lid}]

        return results, meta_data
        return results, meta_data