Merge branch 'main' of https://github.com/alibaba-damo-academy/FunASR into ...

雾聪

2023-06-28 54931dd4e1a099d7d6f144c4e12e5453deb3aa26

 funasr/models/e2e_uni_asr.py

@@ -25,7 +25,7 @@
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

@@ -38,9 +38,9 @@
        yield


class UniASR(AbsESPnetModel):
class UniASR(FunASRModel):
    """
    Author: Speech Lab, Alibaba Group, China
    Author: Speech Lab of DAMO Academy, Alibaba Group
    """

    def __init__(
@@ -50,9 +50,7 @@
        frontend: Optional[AbsFrontend],
        specaug: Optional[AbsSpecAug],
        normalize: Optional[AbsNormalize],
        preencoder: Optional[AbsPreEncoder],
        encoder: AbsEncoder,
        postencoder: Optional[AbsPostEncoder],
        decoder: AbsDecoder,
        ctc: CTC,
        ctc_weight: float = 0.5,
@@ -80,6 +78,8 @@
        loss_weight_model1: float = 0.5,
        enable_maas_finetune: bool = False,
        freeze_encoder2: bool = False,
        preencoder: Optional[AbsPreEncoder] = None,
        postencoder: Optional[AbsPostEncoder] = None,
        encoder1_encoder2_joint_training: bool = True,
    ):
        assert check_argument_types()
@@ -179,7 +179,6 @@
        decoding_ind: int = None,
    ) -> Tuple[torch.Tensor, Dict[str, torch.Tensor], torch.Tensor]:
        """Frontend + Encoder + Decoder + Calc loss

        Args:
                        speech: (Batch, Length, ...)
                        speech_lengths: (Batch, )
@@ -206,7 +205,7 @@
            with torch.no_grad():
                speech_raw, encoder_out, encoder_out_lens = self.encode(speech, speech_lengths, ind=ind)
        else:
            speech_raw, 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):
@@ -469,7 +468,6 @@
        self, speech: torch.Tensor, speech_lengths: torch.Tensor, ind: int = 0,
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """Frontend + Encoder. Note that this method is used by asr_inference.py

        Args:
                        speech: (Batch, Length, ...)
                        speech_lengths: (Batch, )
@@ -533,7 +531,6 @@
        ind: int = 0,
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """Frontend + Encoder. Note that this method is used by asr_inference.py

        Args:
                        speech: (Batch, Length, ...)
                        speech_lengths: (Batch, )
@@ -627,9 +624,7 @@
        ys_pad_lens: torch.Tensor,
    ) -> torch.Tensor:
        """Compute negative log likelihood(nll) from transformer-decoder

        Normally, this function is called in batchify_nll.

        Args:
                        encoder_out: (Batch, Length, Dim)
                        encoder_out_lens: (Batch,)
@@ -666,7 +661,6 @@
        batch_size: int = 100,
    ):
        """Compute negative log likelihood(nll) from transformer-decoder

        To avoid OOM, this fuction seperate the input into batches.
        Then call nll for each batch and combine and return results.
        Args:
@@ -1072,4 +1066,3 @@
            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