update repo

嘉渊

2023-06-14 adb8997b8023776ad1389ca8e9adcbaa1de4821e

 funasr/build_utils/build_asr_model.py

@@ -6,6 +6,7 @@
from funasr.models.decoder.abs_decoder import AbsDecoder
from funasr.models.decoder.contextual_decoder import ContextualParaformerDecoder
from funasr.models.decoder.rnn_decoder import RNNDecoder
from funasr.models.decoder.rnnt_decoder import RNNTDecoder
from funasr.models.decoder.sanm_decoder import ParaformerSANMDecoder, FsmnDecoderSCAMAOpt
from funasr.models.decoder.transformer_decoder import (
    DynamicConvolution2DTransformerDecoder,  # noqa: H301
@@ -19,14 +20,13 @@
)
from funasr.models.decoder.transformer_decoder import ParaformerDecoderSAN
from funasr.models.decoder.transformer_decoder import TransformerDecoder
from funasr.models.decoder.rnnt_decoder import RNNTDecoder
from funasr.models.joint_net.joint_network import JointNetwork
from funasr.models.e2e_asr import ASRModel
from funasr.models.e2e_asr_mfcca import MFCCA
from funasr.models.e2e_asr_paraformer import Paraformer, ParaformerBert, BiCifParaformer, ContextualParaformer
from funasr.models.e2e_asr_paraformer import Paraformer, ParaformerOnline, ParaformerBert, BiCifParaformer, \
    ContextualParaformer
from funasr.models.e2e_asr_transducer import TransducerModel, UnifiedTransducerModel
from funasr.models.e2e_tp import TimestampPredictor
from funasr.models.e2e_uni_asr import UniASR
from funasr.models.e2e_asr_transducer import TransducerModel, UnifiedTransducerModel
from funasr.models.encoder.conformer_encoder import ConformerEncoder, ConformerChunkEncoder
from funasr.models.encoder.data2vec_encoder import Data2VecEncoder
from funasr.models.encoder.mfcca_encoder import MFCCAEncoder
@@ -39,6 +39,7 @@
from funasr.models.frontend.s3prl import S3prlFrontend
from funasr.models.frontend.wav_frontend import WavFrontend
from funasr.models.frontend.windowing import SlidingWindow
from funasr.models.joint_net.joint_network import JointNetwork
from funasr.models.predictor.cif import CifPredictor, CifPredictorV2, CifPredictorV3
from funasr.models.specaug.specaug import SpecAug
from funasr.models.specaug.specaug import SpecAugLFR
@@ -82,11 +83,14 @@
        asr=ASRModel,
        uniasr=UniASR,
        paraformer=Paraformer,
        paraformer_online=ParaformerOnline,
        paraformer_bert=ParaformerBert,
        bicif_paraformer=BiCifParaformer,
        contextual_paraformer=ContextualParaformer,
        mfcca=MFCCA,
        timestamp_prediction=TimestampPredictor,
        rnnt=TransducerModel,
        rnnt_unified=UnifiedTransducerModel,
    ),
    default="asr",
)
@@ -224,17 +228,21 @@

def build_asr_model(args):
    # token_list
    if args.token_list is not None:
        with open(args.token_list) as f:
    if isinstance(args.token_list, str):
        with open(args.token_list, encoding="utf-8") as f:
            token_list = [line.rstrip() for line in f]
        args.token_list = list(token_list)
        vocab_size = len(token_list)
        logging.info(f"Vocabulary size: {vocab_size}")
    elif isinstance(args.token_list, (tuple, list)):
        token_list = list(args.token_list)
        vocab_size = len(token_list)
        logging.info(f"Vocabulary size: {vocab_size}")
    else:
        vocab_size = None

    # frontend
    if args.input_size is None:
    if hasattr(args, "input_size") and args.input_size is None:
        frontend_class = frontend_choices.get_class(args.frontend)
        if args.frontend == 'wav_frontend':
            frontend = frontend_class(cmvn_file=args.cmvn_file, **args.frontend_conf)
@@ -245,7 +253,7 @@
        args.frontend = None
        args.frontend_conf = {}
        frontend = None
        input_size = args.input_size
        input_size = args.input_size if hasattr(args, "input_size") else None

    # data augmentation for spectrogram
    if args.specaug is not None:
@@ -291,7 +299,8 @@
            token_list=token_list,
            **args.model_conf,
        )
    elif args.model in ["paraformer", "paraformer_bert", "bicif_paraformer", "contextual_paraformer"]:
    elif args.model in ["paraformer", "paraformer_online", "paraformer_bert", "bicif_paraformer",
                        "contextual_paraformer"]:
        # predictor
        predictor_class = predictor_choices.get_class(args.predictor)
        predictor = predictor_class(**args.predictor_conf)
@@ -367,7 +376,7 @@
            token_list=token_list,
            **args.model_conf,
        )
    elif args.model == "rnnt":
    elif args.model == "rnnt" or args.model == "rnnt_unified":
        # 5. Decoder
        encoder_output_size = encoder.output_size()

@@ -396,34 +405,21 @@
            **args.joint_network_conf,
        )

        model_class = model_choices.get_class(args.model)
        # 7. Build model
        if hasattr(encoder, 'unified_model_training') and encoder.unified_model_training:
            model = UnifiedTransducerModel(
                vocab_size=vocab_size,
                token_list=token_list,
                frontend=frontend,
                specaug=specaug,
                normalize=normalize,
                encoder=encoder,
                decoder=decoder,
                att_decoder=att_decoder,
                joint_network=joint_network,
                **args.model_conf,
            )
        model = model_class(
            vocab_size=vocab_size,
            token_list=token_list,
            frontend=frontend,
            specaug=specaug,
            normalize=normalize,
            encoder=encoder,
            decoder=decoder,
            att_decoder=att_decoder,
            joint_network=joint_network,
            **args.model_conf,
        )

        else:
            model = TransducerModel(
                vocab_size=vocab_size,
                token_list=token_list,
                frontend=frontend,
                specaug=specaug,
                normalize=normalize,
                encoder=encoder,
                decoder=decoder,
                att_decoder=att_decoder,
                joint_network=joint_network,
                **args.model_conf,
            )
    else:
        raise NotImplementedError("Not supported model: {}".format(args.model))