Dev lyh (#645)

yhliang

2023-06-16 e8528b8f6208cee52ed9c02ecfa9185f84706502

 funasr/build_utils/build_asr_model.py

@@ -21,8 +21,10 @@
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.decoder.transformer_decoder import SAAsrTransformerDecoder
from funasr.models.e2e_asr import ASRModel
from funasr.models.e2e_asr_mfcca import MFCCA
from funasr.models.e2e_sa_asr import SAASRModel
from funasr.models.e2e_asr_paraformer import Paraformer, ParaformerOnline, ParaformerBert, BiCifParaformer, ContextualParaformer
from funasr.models.e2e_tp import TimestampPredictor
from funasr.models.e2e_uni_asr import UniASR
@@ -30,6 +32,7 @@
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
from funasr.models.encoder.resnet34_encoder import ResNet34Diar
from funasr.models.encoder.rnn_encoder import RNNEncoder
from funasr.models.encoder.sanm_encoder import SANMEncoder, SANMEncoderChunkOpt
from funasr.models.encoder.transformer_encoder import TransformerEncoder
@@ -90,6 +93,8 @@
        timestamp_prediction=TimestampPredictor,
        rnnt=TransducerModel,
        rnnt_unified=UnifiedTransducerModel,
        sa_asr=SAASRModel,

    ),
    default="asr",
)
@@ -106,6 +111,27 @@
        chunk_conformer=ConformerChunkEncoder,
    ),
    default="rnn",
)
asr_encoder_choices = ClassChoices(
    "asr_encoder",
    classes=dict(
        conformer=ConformerEncoder,
        transformer=TransformerEncoder,
        rnn=RNNEncoder,
        sanm=SANMEncoder,
        sanm_chunk_opt=SANMEncoderChunkOpt,
        data2vec_encoder=Data2VecEncoder,
        mfcca_enc=MFCCAEncoder,
    ),
    default="rnn",
)

spk_encoder_choices = ClassChoices(
    "spk_encoder",
    classes=dict(
        resnet34_diar=ResNet34Diar,
    ),
    default="resnet34_diar",
)
encoder_choices2 = ClassChoices(
    "encoder2",
@@ -131,6 +157,7 @@
        paraformer_decoder_sanm=ParaformerSANMDecoder,
        paraformer_decoder_san=ParaformerDecoderSAN,
        contextual_paraformer_decoder=ContextualParaformerDecoder,
        sa_decoder=SAAsrTransformerDecoder,
    ),
    default="rnn",
)
@@ -222,6 +249,10 @@
    rnnt_decoder_choices,
    # --joint_network and --joint_network_conf
    joint_network_choices,
    # --asr_encoder and --asr_encoder_conf
    asr_encoder_choices,
    # --spk_encoder and --spk_encoder_conf
    spk_encoder_choices,
]


@@ -239,7 +270,7 @@
    # frontend
    if args.input_size is None:
        frontend_class = frontend_choices.get_class(args.frontend)
        if args.frontend == 'wav_frontend':
        if args.frontend == 'wav_frontend' or args.frontend == 'multichannelfrontend':
            frontend = frontend_class(cmvn_file=args.cmvn_file, **args.frontend_conf)
        else:
            frontend = frontend_class(**args.frontend_conf)
@@ -413,6 +444,33 @@
            joint_network=joint_network,
            **args.model_conf,
        )
    elif args.model == "sa_asr":
        asr_encoder_class = asr_encoder_choices.get_class(args.asr_encoder)
        asr_encoder = asr_encoder_class(input_size=input_size, **args.asr_encoder_conf)
        spk_encoder_class = spk_encoder_choices.get_class(args.spk_encoder)
        spk_encoder = spk_encoder_class(input_size=input_size, **args.spk_encoder_conf)
        decoder = decoder_class(
            vocab_size=vocab_size,
            encoder_output_size=asr_encoder.output_size(),
            **args.decoder_conf,
        )
        ctc = CTC(
            odim=vocab_size, encoder_output_size=asr_encoder.output_size(), **args.ctc_conf
        )

        model_class = model_choices.get_class(args.model)
        model = model_class(
            vocab_size=vocab_size,
            frontend=frontend,
            specaug=specaug,
            normalize=normalize,
            asr_encoder=asr_encoder,
            spk_encoder=spk_encoder,
            decoder=decoder,
            ctc=ctc,
            token_list=token_list,
            **args.model_conf,
        )

    else:
        raise NotImplementedError("Not supported model: {}".format(args.model))