aishell example

游雁

2024-02-19 94de39dde2e616a01683c518023d0fab72b4e103

 funasr/models/uniasr/model.py

@@ -14,14 +14,13 @@
from funasr.utils import postprocess_utils
from funasr.metrics.compute_acc import th_accuracy
from funasr.utils.datadir_writer import DatadirWriter
from funasr.models.paraformer.search import Hypothesis
from funasr.models.paraformer.cif_predictor import mae_loss
from funasr.train_utils.device_funcs import force_gatherable
from funasr.losses.label_smoothing_loss import LabelSmoothingLoss
from funasr.models.transformer.utils.add_sos_eos import add_sos_eos
from funasr.models.transformer.utils.nets_utils import make_pad_mask, pad_list
from funasr.utils.load_utils import load_audio_text_image_video, extract_fbank

from funasr.models.scama.utils import sequence_mask

@tables.register("model_classes", "UniASR")
class UniASR(torch.nn.Module):
@@ -31,19 +30,37 @@

    def __init__(
        self,
        specaug: Optional[str] = None,
        specaug_conf: Optional[Dict] = None,
        specaug: str = None,
        specaug_conf: dict = None,
        normalize: str = None,
        normalize_conf: Optional[Dict] = None,
        normalize_conf: dict = None,
        encoder: str = None,
        encoder_conf: Optional[Dict] = None,
        encoder_conf: dict = None,
        encoder2: str = None,
        encoder2_conf: dict = None,
        decoder: str = None,
        decoder_conf: Optional[Dict] = None,
        ctc: str = None,
        ctc_conf: Optional[Dict] = None,
        decoder_conf: dict = None,
        decoder2: str = None,
        decoder2_conf: dict = None,
        predictor: str = None,
        predictor_conf: Optional[Dict] = None,
        predictor_conf: dict = None,
        predictor_bias: int = 0,
        predictor_weight: float = 0.0,
        predictor2: str = None,
        predictor2_conf: dict = None,
        predictor2_bias: int = 0,
        predictor2_weight: float = 0.0,
        ctc: str = None,
        ctc_conf: dict = None,
        ctc_weight: float = 0.5,
        ctc2: str = None,
        ctc2_conf: dict = None,
        ctc2_weight: float = 0.5,
        decoder_attention_chunk_type: str = 'chunk',
        decoder_attention_chunk_type2: str = 'chunk',
        stride_conv=None,
        stride_conv_conf: dict = None,
        loss_weight_model1: float = 0.5,
        input_size: int = 80,
        vocab_size: int = -1,
        ignore_id: int = -1,
@@ -52,60 +69,72 @@
        eos: int = 2,
        lsm_weight: float = 0.0,
        length_normalized_loss: bool = False,
        # report_cer: bool = True,
        # report_wer: bool = True,
        # sym_space: str = "<space>",
        # sym_blank: str = "<blank>",
        # extract_feats_in_collect_stats: bool = True,
        # predictor=None,
        predictor_weight: float = 0.0,
        predictor_bias: int = 0,
        sampling_ratio: float = 0.2,
        share_embedding: bool = False,
        # preencoder: Optional[AbsPreEncoder] = None,
        # postencoder: Optional[AbsPostEncoder] = None,
        use_1st_decoder_loss: bool = False,
        encoder1_encoder2_joint_training: bool = True,
        **kwargs,
        
    ):
        assert 0.0 <= ctc_weight <= 1.0, ctc_weight
        assert 0.0 <= interctc_weight < 1.0, interctc_weight

        super().__init__()
        self.blank_id = 0
        self.sos = 1
        self.eos = 2

        if specaug is not None:
            specaug_class = tables.specaug_classes.get(specaug)
            specaug = specaug_class(**specaug_conf)
        if normalize is not None:
            normalize_class = tables.normalize_classes.get(normalize)
            normalize = normalize_class(**normalize_conf)
            
        encoder_class = tables.encoder_classes.get(encoder)
        encoder = encoder_class(input_size=input_size, **encoder_conf)
        encoder_output_size = encoder.output_size()

        decoder_class = tables.decoder_classes.get(decoder)
        decoder = decoder_class(
            vocab_size=vocab_size,
            encoder_output_size=encoder_output_size,
            **decoder_conf,
        )
        predictor_class = tables.predictor_classes.get(predictor)
        predictor = predictor_class(**predictor_conf)
        

        
        from funasr.models.transformer.utils.subsampling import Conv1dSubsampling
        stride_conv = Conv1dSubsampling(**stride_conv_conf, idim=input_size + encoder_output_size,
                                        odim=input_size + encoder_output_size)
        stride_conv_output_size = stride_conv.output_size()

        encoder_class = tables.encoder_classes.get(encoder2)
        encoder2 = encoder_class(input_size=stride_conv_output_size, **encoder2_conf)
        encoder2_output_size = encoder2.output_size()

        decoder_class = tables.decoder_classes.get(decoder2)
        decoder2 = decoder_class(
            vocab_size=vocab_size,
            encoder_output_size=encoder2_output_size,
            **decoder2_conf,
        )
        predictor_class = tables.predictor_classes.get(predictor2)
        predictor2 = predictor_class(**predictor2_conf)


        
        self.blank_id = blank_id
        self.sos = sos
        self.eos = eos
        self.vocab_size = vocab_size
        self.ignore_id = ignore_id
        self.ctc_weight = ctc_weight
        self.interctc_weight = interctc_weight
        self.token_list = token_list.copy()
        self.ctc2_weight = ctc2_weight

        self.frontend = frontend
        self.specaug = specaug
        self.normalize = normalize
        self.preencoder = preencoder
        self.postencoder = postencoder
        
        self.encoder = encoder

        if not hasattr(self.encoder, "interctc_use_conditioning"):
            self.encoder.interctc_use_conditioning = False
        if self.encoder.interctc_use_conditioning:
            self.encoder.conditioning_layer = torch.nn.Linear(
                vocab_size, self.encoder.output_size()
            )

        self.error_calculator = None

        # we set self.decoder = None in the CTC mode since
        # self.decoder parameters were never used and PyTorch complained
        # and threw an Exception in the multi-GPU experiment.
        # thanks Jeff Farris for pointing out the issue.
        if ctc_weight == 1.0:
            self.decoder = None
        else:
            self.decoder = decoder
        self.decoder = decoder
        self.ctc = None
        self.ctc2 = None

        self.criterion_att = LabelSmoothingLoss(
            size=vocab_size,
@@ -113,22 +142,13 @@
            smoothing=lsm_weight,
            normalize_length=length_normalized_loss,
        )

        if report_cer or report_wer:
            self.error_calculator = ErrorCalculator(
                token_list, sym_space, sym_blank, report_cer, report_wer
            )

        if ctc_weight == 0.0:
            self.ctc = None
        else:
            self.ctc = ctc

        self.extract_feats_in_collect_stats = extract_feats_in_collect_stats
        
        self.predictor = predictor
        self.predictor_weight = predictor_weight
        self.criterion_pre = mae_loss(normalize_length=length_normalized_loss)
        self.step_cur = 0
        self.encoder1_encoder2_joint_training = kwargs.get("encoder1_encoder2_joint_training", True)
        

        if self.encoder.overlap_chunk_cls is not None:
            from funasr.models.scama.chunk_utilis import build_scama_mask_for_cross_attention_decoder
            self.build_scama_mask_for_cross_attention_decoder_fn = build_scama_mask_for_cross_attention_decoder
@@ -136,14 +156,10 @@

        self.encoder2 = encoder2
        self.decoder2 = decoder2
        self.ctc_weight2 = ctc_weight2
        if ctc_weight2 == 0.0:
            self.ctc2 = None
        else:
            self.ctc2 = ctc2
        self.interctc_weight2 = interctc_weight2
        self.ctc2_weight = ctc2_weight

        self.predictor2 = predictor2
        self.predictor_weight2 = predictor_weight2
        self.predictor2_weight = predictor2_weight
        self.decoder_attention_chunk_type2 = decoder_attention_chunk_type2
        self.stride_conv = stride_conv
        self.loss_weight_model1 = loss_weight_model1
@@ -152,10 +168,10 @@
            self.build_scama_mask_for_cross_attention_decoder_fn2 = build_scama_mask_for_cross_attention_decoder
            self.decoder_attention_chunk_type2 = decoder_attention_chunk_type2

        self.enable_maas_finetune = enable_maas_finetune
        self.freeze_encoder2 = freeze_encoder2
        self.encoder1_encoder2_joint_training = encoder1_encoder2_joint_training
        self.length_normalized_loss = length_normalized_loss
        self.enable_maas_finetune = kwargs.get("enable_maas_finetune", False)
        self.freeze_encoder2 = kwargs.get("freeze_encoder2", False)
        self.beam_search = None

    def forward(
        self,
@@ -163,7 +179,7 @@
        speech_lengths: torch.Tensor,
        text: torch.Tensor,
        text_lengths: torch.Tensor,
        decoding_ind: int = None,
        **kwargs,
    ) -> Tuple[torch.Tensor, Dict[str, torch.Tensor], torch.Tensor]:
        """Frontend + Encoder + Decoder + Calc loss
        Args:
@@ -172,19 +188,14 @@
                        text: (Batch, Length)
                        text_lengths: (Batch,)
        """
        assert text_lengths.dim() == 1, text_lengths.shape
        # Check that batch_size is unified
        assert (
            speech.shape[0]
            == speech_lengths.shape[0]
            == text.shape[0]
            == text_lengths.shape[0]
        ), (speech.shape, speech_lengths.shape, text.shape, text_lengths.shape)
        decoding_ind = kwargs.get("decoding_ind", None)
        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]

        # for data-parallel
        text = text[:, : text_lengths.max()]
        speech = speech[:, :speech_lengths.max()]

        ind = self.encoder.overlap_chunk_cls.random_choice(self.training, decoding_ind)
        # 1. Encoder
@@ -194,10 +205,6 @@
        else:
            speech_raw, encoder_out, encoder_out_lens = self.encode(speech, speech_lengths, ind=ind)

        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
@@ -210,62 +217,12 @@
            # 1. CTC branch
            if self.enable_maas_finetune:
                with torch.no_grad():
                    if self.ctc_weight != 0.0:
                        if self.encoder.overlap_chunk_cls is not None:
                            encoder_out_ctc, encoder_out_lens_ctc = self.encoder.overlap_chunk_cls.remove_chunk(encoder_out,
                                                                                                                encoder_out_lens,
                                                                                                                chunk_outs=None)
                        loss_ctc, cer_ctc = self._calc_ctc_loss(
                            encoder_out_ctc, encoder_out_lens_ctc, 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
                    loss_att, acc_att, cer_att, wer_att, loss_pre = self._calc_att_predictor_loss(
                        encoder_out, encoder_out_lens, text, text_lengths
                    )

                    # Intermediate CTC (optional)
                    loss_interctc = 0.0
                    if self.interctc_weight != 0.0 and intermediate_outs is not None:
                        for layer_idx, intermediate_out in intermediate_outs:
                            # we assume intermediate_out has the same length & padding
                            # as those of encoder_out
                            if self.encoder.overlap_chunk_cls is not None:
                                encoder_out_ctc, encoder_out_lens_ctc = \
                                    self.encoder.overlap_chunk_cls.remove_chunk(
                                        intermediate_out,
                                        encoder_out_lens,
                                        chunk_outs=None)
                            loss_ic, cer_ic = self._calc_ctc_loss(
                                encoder_out_ctc, encoder_out_lens_ctc, 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

                    # 2b. Attention decoder branch
                    if self.ctc_weight != 1.0:
                        loss_att, acc_att, cer_att, wer_att, loss_pre = self._calc_att_predictor_loss(
                            encoder_out, encoder_out_lens, text, text_lengths
                        )

                    # 3. CTC-Att loss definition
                    if self.ctc_weight == 0.0:
                        loss = loss_att + loss_pre * self.predictor_weight
                    elif self.ctc_weight == 1.0:
                        loss = loss_ctc
                    else:
                        loss = self.ctc_weight * loss_ctc + (1 - self.ctc_weight) * loss_att + loss_pre * self.predictor_weight
                    loss = loss_att + loss_pre * self.predictor_weight

                    # Collect Attn branch stats
                    stats["loss_att"] = loss_att.detach() if loss_att is not None else None
@@ -274,62 +231,13 @@
                    stats["wer"] = wer_att
                    stats["loss_pre"] = loss_pre.detach().cpu() if loss_pre is not None else None
            else:
                if self.ctc_weight != 0.0:
                    if self.encoder.overlap_chunk_cls is not None:
                        encoder_out_ctc, encoder_out_lens_ctc = self.encoder.overlap_chunk_cls.remove_chunk(encoder_out,
                                                                                                            encoder_out_lens,
                                                                                                            chunk_outs=None)
                    loss_ctc, cer_ctc = self._calc_ctc_loss(
                        encoder_out_ctc, encoder_out_lens_ctc, text, text_lengths
                    )
                
                loss_att, acc_att, cer_att, wer_att, loss_pre = self._calc_att_predictor_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:
                    for layer_idx, intermediate_out in intermediate_outs:
                        # we assume intermediate_out has the same length & padding
                        # as those of encoder_out
                        if self.encoder.overlap_chunk_cls is not None:
                            encoder_out_ctc, encoder_out_lens_ctc = \
                                self.encoder.overlap_chunk_cls.remove_chunk(
                                    intermediate_out,
                                    encoder_out_lens,
                                    chunk_outs=None)
                        loss_ic, cer_ic = self._calc_ctc_loss(
                            encoder_out_ctc, encoder_out_lens_ctc, 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

                # 2b. Attention decoder branch
                if self.ctc_weight != 1.0:
                    loss_att, acc_att, cer_att, wer_att, loss_pre = self._calc_att_predictor_loss(
                        encoder_out, encoder_out_lens, text, text_lengths
                    )

                # 3. CTC-Att loss definition
                if self.ctc_weight == 0.0:
                    loss = loss_att + loss_pre * self.predictor_weight
                elif self.ctc_weight == 1.0:
                    loss = loss_ctc
                else:
                    loss = self.ctc_weight * loss_ctc + (1 - self.ctc_weight) * loss_att + loss_pre * self.predictor_weight
                loss = loss_att + loss_pre * self.predictor_weight

                # Collect Attn branch stats
                stats["loss_att"] = loss_att.detach() if loss_att is not None else None
@@ -354,67 +262,14 @@
            if isinstance(encoder_out, tuple):
                intermediate_outs = encoder_out[1]
                encoder_out = encoder_out[0]
            # CTC2
            if self.ctc_weight2 != 0.0:
                if self.encoder2.overlap_chunk_cls is not None:
                    encoder_out_ctc, encoder_out_lens_ctc = \
                        self.encoder2.overlap_chunk_cls.remove_chunk(
                            encoder_out,
                            encoder_out_lens,
                            chunk_outs=None,
                        )
                loss_ctc, cer_ctc = self._calc_ctc_loss2(
                    encoder_out_ctc, encoder_out_lens_ctc, text, text_lengths
                )

                # Collect CTC branch stats
                stats["loss_ctc2"] = loss_ctc.detach() if loss_ctc is not None else None
                stats["cer_ctc2"] = cer_ctc

            # Intermediate CTC (optional)
            loss_interctc = 0.0
            if self.interctc_weight2 != 0.0 and intermediate_outs is not None:
                for layer_idx, intermediate_out in intermediate_outs:
                    # we assume intermediate_out has the same length & padding
                    # as those of encoder_out
                    if self.encoder2.overlap_chunk_cls is not None:
                        encoder_out_ctc, encoder_out_lens_ctc = \
                            self.encoder2.overlap_chunk_cls.remove_chunk(
                                intermediate_out,
                                encoder_out_lens,
                                chunk_outs=None)
                    loss_ic, cer_ic = self._calc_ctc_loss2(
                        encoder_out_ctc, encoder_out_lens_ctc, text, text_lengths
                    )
                    loss_interctc = loss_interctc + loss_ic
            loss_att, acc_att, cer_att, wer_att, loss_pre = self._calc_att_predictor_loss2(
                encoder_out, encoder_out_lens, text, text_lengths
            )

                    # Collect Intermedaite CTC stats
                    stats["loss_interctc_layer{}2".format(layer_idx)] = (
                        loss_ic.detach() if loss_ic is not None else None
                    )
                    stats["cer_interctc_layer{}2".format(layer_idx)] = cer_ic

                loss_interctc = loss_interctc / len(intermediate_outs)

                # calculate whole encoder loss
                loss_ctc = (
                               1 - self.interctc_weight2
                           ) * loss_ctc + self.interctc_weight2 * loss_interctc

            # 2b. Attention decoder branch
            if self.ctc_weight2 != 1.0:
                loss_att, acc_att, cer_att, wer_att, loss_pre = self._calc_att_predictor_loss2(
                    encoder_out, encoder_out_lens, text, text_lengths
                )

            # 3. CTC-Att loss definition
            if self.ctc_weight2 == 0.0:
                loss = loss_att + loss_pre * self.predictor_weight2
            elif self.ctc_weight2 == 1.0:
                loss = loss_ctc
            else:
                loss = self.ctc_weight2 * loss_ctc + (
                    1 - self.ctc_weight2) * loss_att + loss_pre * self.predictor_weight2
            loss = loss_att + loss_pre * self.predictor2_weight

            # Collect Attn branch stats
            stats["loss_att2"] = loss_att.detach() if loss_att is not None else None
@@ -422,6 +277,7 @@
            stats["cer2"] = cer_att
            stats["wer2"] = wer_att
            stats["loss_pre2"] = loss_pre.detach().cpu() if loss_pre is not None else None
        
        loss2 = loss

        loss = loss1 * self.loss_weight_model1 + loss2 * (1 - self.loss_weight_model1)
@@ -456,60 +312,30 @@
        return {"feats": feats, "feats_lengths": feats_lengths}

    def encode(
        self, speech: torch.Tensor, speech_lengths: torch.Tensor, ind: int = 0,
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        self, speech: torch.Tensor, speech_lengths: torch.Tensor, **kwargs,
    ):
        """Frontend + Encoder. Note that this method is used by asr_inference.py
        Args:
                        speech: (Batch, Length, ...)
                        speech_lengths: (Batch, )
        """
        ind = kwargs.get("ind", 0)
        with autocast(False):
            # 1. Extract feats
            feats, feats_lengths = self._extract_feats(speech, speech_lengths)

            # 2. Data augmentation
            # Data augmentation
            if self.specaug is not None and self.training:
                feats, feats_lengths = self.specaug(feats, feats_lengths)

            # 3. Normalization for feature: e.g. Global-CMVN, Utterance-CMVN
                speech, speech_lengths = self.specaug(speech, speech_lengths)
    
            # Normalization for feature: e.g. Global-CMVN, Utterance-CMVN
            if self.normalize is not None:
                feats, feats_lengths = self.normalize(feats, feats_lengths)
        speech_raw = feats.clone().to(feats.device)
        # Pre-encoder, e.g. used for raw input data
        if self.preencoder is not None:
            feats, feats_lengths = self.preencoder(feats, feats_lengths)
                speech, speech_lengths = self.normalize(speech, speech_lengths)
                
        speech_raw = speech.clone().to(speech.device)


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

        # Post-encoder, e.g. NLU
        if self.postencoder is not None:
            encoder_out, encoder_out_lens = self.postencoder(
                encoder_out, encoder_out_lens
            )

        assert encoder_out.size(0) == speech.size(0), (
            encoder_out.size(),
            speech.size(0),
        )
        assert encoder_out.size(1) <= encoder_out_lens.max(), (
            encoder_out.size(),
            encoder_out_lens.max(),
        )

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

        return speech_raw, encoder_out, encoder_out_lens

@@ -519,28 +345,15 @@
        encoder_out_lens: torch.Tensor,
        speech: torch.Tensor,
        speech_lengths: torch.Tensor,
        ind: int = 0,
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        **kwargs,
    ):
        """Frontend + Encoder. Note that this method is used by asr_inference.py
        Args:
                        speech: (Batch, Length, ...)
                        speech_lengths: (Batch, )
        """
        # with autocast(False):
        #     # 1. Extract feats
        #     feats, feats_lengths = self._extract_feats(speech, speech_lengths)
        #
        #     # 2. Data augmentation
        #     if self.specaug is not None and self.training:
        #         feats, feats_lengths = self.specaug(feats, feats_lengths)
        #
        #     # 3. Normalization for feature: e.g. Global-CMVN, Utterance-CMVN
        #     if self.normalize is not None:
        #         feats, feats_lengths = self.normalize(feats, feats_lengths)

        # Pre-encoder, e.g. used for raw input data
        # if self.preencoder is not None:
        #     feats, feats_lengths = self.preencoder(feats, feats_lengths)
        ind = kwargs.get("ind", 0)
        encoder_out_rm, encoder_out_lens_rm = self.encoder.overlap_chunk_cls.remove_chunk(
            encoder_out,
            encoder_out_lens,
@@ -557,55 +370,14 @@
        # 4. Forward encoder
        # feats: (Batch, Length, Dim)
        # -> encoder_out: (Batch, Length2, Dim2)
        if self.encoder2.interctc_use_conditioning:
            encoder_out, encoder_out_lens, _ = self.encoder2(
                speech, speech_lengths, ctc=self.ctc2, ind=ind
            )
        else:
            encoder_out, encoder_out_lens, _ = self.encoder2(speech, speech_lengths, ind=ind)
        intermediate_outs = None

        encoder_out, encoder_out_lens, _ = self.encoder2(speech, speech_lengths, ind=ind)
        if isinstance(encoder_out, tuple):
            intermediate_outs = encoder_out[1]
            encoder_out = encoder_out[0]

        # # Post-encoder, e.g. NLU
        # if self.postencoder is not None:
        #     encoder_out, encoder_out_lens = self.postencoder(
        #         encoder_out, encoder_out_lens
        #     )

        assert encoder_out.size(0) == speech.size(0), (
            encoder_out.size(),
            speech.size(0),
        )
        assert encoder_out.size(1) <= encoder_out_lens.max(), (
            encoder_out.size(),
            encoder_out_lens.max(),
        )

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

        return encoder_out, encoder_out_lens

    def _extract_feats(
        self, speech: torch.Tensor, speech_lengths: torch.Tensor
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        assert speech_lengths.dim() == 1, speech_lengths.shape

        # for data-parallel
        speech = speech[:, : speech_lengths.max()]

        if self.frontend is not None:
            # Frontend
            #  e.g. STFT and Feature extract
            #       data_loader may send time-domain signal in this case
            # speech (Batch, NSamples) -> feats: (Batch, NFrames, Dim)
            feats, feats_lengths = self.frontend(speech, speech_lengths)
        else:
            # No frontend and no feature extract
            feats, feats_lengths = speech, speech_lengths
        return feats, feats_lengths

    def nll(
        self,
@@ -1024,36 +796,152 @@

        return pre_acoustic_embeds, pre_token_length, predictor_alignments, predictor_alignments_len, scama_mask

    def _calc_ctc_loss(
        self,
        encoder_out: torch.Tensor,
        encoder_out_lens: torch.Tensor,
        ys_pad: torch.Tensor,
        ys_pad_lens: torch.Tensor,
    ):
        # Calc CTC loss
        loss_ctc = self.ctc(encoder_out, encoder_out_lens, ys_pad, ys_pad_lens)
    def init_beam_search(self,
                         **kwargs,
                         ):
        from funasr.models.uniasr.beam_search import BeamSearchScama
        from funasr.models.transformer.scorers.ctc import CTCPrefixScorer
        from funasr.models.transformer.scorers.length_bonus import LengthBonus

        # 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
        decoding_mode = kwargs.get("decoding_mode", "model1")
        if decoding_mode == "model1":
            decoder = self.decoder
        else:
            decoder = self.decoder2
        # 1. Build ASR model
        scorers = {}
    
        if self.ctc != None:
            ctc = CTCPrefixScorer(ctc=self.ctc, eos=self.eos)
            scorers.update(
                ctc=ctc
            )
        token_list = kwargs.get("token_list")
        scorers.update(
            decoder=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", 0.0),
            ctc=kwargs.get("decoding_ctc_weight", 0.0),
            lm=kwargs.get("lm_weight", 0.0),
            ngram=kwargs.get("ngram_weight", 0.0),
            length_bonus=kwargs.get("penalty", 0.0),
        )
        beam_search = BeamSearchScama(
            beam_size=kwargs.get("beam_size", 5),
            weights=weights,
            scorers=scorers,
            sos=self.sos,
            eos=self.eos,
            vocab_size=len(token_list),
            token_list=token_list,
            pre_beam_score_key=None if self.ctc_weight == 1.0 else "full",
        )
        
        self.beam_search = beam_search

    def _calc_ctc_loss2(
        self,
        encoder_out: torch.Tensor,
        encoder_out_lens: torch.Tensor,
        ys_pad: torch.Tensor,
        ys_pad_lens: torch.Tensor,
    ):
        # Calc CTC loss
        loss_ctc = self.ctc2(encoder_out, encoder_out_lens, ys_pad, ys_pad_lens)
    def inference(self,
                  data_in,
                  data_lengths=None,
                  key: list = None,
                  tokenizer=None,
                  frontend=None,
                  **kwargs,
                  ):

        # Calc CER using CTC
        cer_ctc = None
        if not self.training and self.error_calculator is not None:
            ys_hat = self.ctc2.argmax(encoder_out).data
            cer_ctc = self.error_calculator(ys_hat.cpu(), ys_pad.cpu(), is_ctc=True)
        return loss_ctc, cer_ctc
        decoding_model = kwargs.get("decoding_model", "normal")
        token_num_relax = kwargs.get("token_num_relax", 5)
        if decoding_model == "fast":
            decoding_ind = 0
            decoding_mode = "model1"
        elif decoding_model == "offline":
            decoding_ind = 1
            decoding_mode = "model2"
        else:
            decoding_ind = 0
            decoding_mode = "model2"
        # init beamsearch
        
        if self.beam_search is None:
            logging.info("enable beam_search")
            self.init_beam_search(decoding_mode=decoding_mode, **kwargs)
            self.nbest = kwargs.get("nbest", 1)
    
        meta_data = {}
        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, :, :]
            if speech_lengths is None:
                speech_lengths = speech.shape[1]
        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)
            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)
            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
    
        speech = speech.to(device=kwargs["device"])
        speech_lengths = speech_lengths.to(device=kwargs["device"])
        speech_raw = speech.clone().to(device=kwargs["device"])
        # Encoder
        _, encoder_out, encoder_out_lens = self.encode(speech, speech_lengths, ind=decoding_ind)
        if decoding_mode == "model1":
            predictor_outs = self.calc_predictor_mask(encoder_out, encoder_out_lens)
        else:
            encoder_out, encoder_out_lens = self.encode2(encoder_out, encoder_out_lens, speech_raw, speech_lengths, ind=decoding_ind)
            predictor_outs = self.calc_predictor_mask2(encoder_out, encoder_out_lens)


        scama_mask = predictor_outs[4]
        pre_token_length = predictor_outs[1]
        pre_acoustic_embeds = predictor_outs[0]
        maxlen = pre_token_length.sum().item() + token_num_relax
        minlen = max(0, pre_token_length.sum().item() - token_num_relax)
        # c. Passed the encoder result and the beam search
        nbest_hyps = self.beam_search(
            x=encoder_out[0], scama_mask=scama_mask, pre_acoustic_embeds=pre_acoustic_embeds, maxlenratio=0.0,
            minlenratio=0.0, maxlen=int(maxlen), minlen=int(minlen),
        )

        nbest_hyps = nbest_hyps[: self.nbest]

        results = []
        for hyp in nbest_hyps:

            # 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 != 0, token_int))


            # Change integer-ids to tokens
            token = tokenizer.ids2tokens(token_int)
            text_postprocessed = tokenizer.tokens2text(token)
            if not hasattr(tokenizer, "bpemodel"):
                text_postprocessed, _ = postprocess_utils.sentence_postprocess(token)
    

            result_i = {"key": key[0], "text": text_postprocessed}
            results.append(result_i)

        return results, meta_data