update

嘉渊

2023-04-24 df5f263e5fe3d7961b1aeb3589012400a9905a8f

 funasr/models/e2e_uni_asr.py

@@ -17,15 +17,12 @@
    LabelSmoothingLoss,  # noqa: H301
)
from funasr.models.ctc import CTC
from funasr.models.decoder.abs_decoder import AbsDecoder
from funasr.models.encoder.abs_encoder import AbsEncoder
from funasr.models.frontend.abs_frontend import AbsFrontend
from funasr.models.decoder.abs_decoder import AbsDecoder
from funasr.models.postencoder.abs_postencoder import AbsPostEncoder
from funasr.models.preencoder.abs_preencoder import AbsPreEncoder
from funasr.models.specaug.abs_specaug import AbsSpecAug
from funasr.layers.abs_normalize import AbsNormalize
from funasr.torch_utils.device_funcs import force_gatherable
from funasr.train.abs_espnet_model import AbsESPnetModel
from funasr.models.base_model import FunASRModel
from funasr.modules.streaming_utils.chunk_utilis import sequence_mask
from funasr.models.predictor.cif import mae_loss

@@ -38,7 +35,7 @@
        yield


class UniASR(AbsESPnetModel):
class UniASR(FunASRModel):
    """
    Author: Speech Lab, Alibaba Group, China
    """
@@ -47,11 +44,11 @@
        self,
        vocab_size: int,
        token_list: Union[Tuple[str, ...], List[str]],
        frontend: Optional[AbsFrontend],
        specaug: Optional[AbsSpecAug],
        normalize: Optional[AbsNormalize],
        frontend: Optional[torch.nn.Module],
        specaug: Optional[torch.nn.Module],
        normalize: Optional[torch.nn.Module],
        preencoder: Optional[AbsPreEncoder],
        encoder: AbsEncoder,
        encoder: torch.nn.Module,
        postencoder: Optional[AbsPostEncoder],
        decoder: AbsDecoder,
        ctc: CTC,
@@ -198,16 +195,15 @@

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

        ind = self.encoder.overlap_chunk_cls.random_choice(self.training, decoding_ind)
        speech_raw = speech.clone().to(speech.device)
        # 1. Encoder
        if self.enable_maas_finetune:
            with torch.no_grad():
                encoder_out, encoder_out_lens = self.encode(speech, speech_lengths, ind=ind)
                speech_raw, encoder_out, encoder_out_lens = self.encode(speech, speech_lengths, ind=ind)
        else:
            encoder_out, encoder_out_lens = self.encode(speech, speech_lengths, ind=ind)
            speech_raw, encoder_out, encoder_out_lens = self.encode(speech, speech_lengths, ind=ind)

        intermediate_outs = None
        if isinstance(encoder_out, tuple):
@@ -486,7 +482,7 @@
            # 3. 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)
@@ -523,7 +519,7 @@
        if intermediate_outs is not None:
            return (encoder_out, intermediate_outs), encoder_out_lens

        return encoder_out, encoder_out_lens
        return speech_raw, encoder_out, encoder_out_lens

    def encode2(
        self,