dcos

游雁

2023-06-29 bc723ea200144bd6fa8a5dff4b9a780feda144fc

 funasr/build_utils/build_asr_model.py

@@ -20,16 +20,22 @@
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.decoder.transformer_decoder import SAAsrTransformerDecoder
from funasr.models.e2e_asr import ASRModel
from funasr.models.e2e_asr_contextual_paraformer import NeatContextualParaformer
from funasr.models.e2e_asr_mfcca import MFCCA
from funasr.models.e2e_asr_paraformer import Paraformer, ParaformerBert, BiCifParaformer, ContextualParaformer

from funasr.models.e2e_asr_transducer import TransducerModel, UnifiedTransducerModel

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
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
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
@@ -39,6 +45,7 @@
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 +89,17 @@
        asr=ASRModel,
        uniasr=UniASR,
        paraformer=Paraformer,
        paraformer_online=ParaformerOnline,
        paraformer_bert=ParaformerBert,
        bicif_paraformer=BiCifParaformer,
        contextual_paraformer=ContextualParaformer,
        neatcontextual_paraformer=NeatContextualParaformer,
        mfcca=MFCCA,
        timestamp_prediction=TimestampPredictor,
        rnnt=TransducerModel,
        rnnt_unified=UnifiedTransducerModel,
        sa_asr=SAASRModel,

    ),
    default="asr",
)
@@ -103,6 +116,27 @@
        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",
@@ -128,6 +162,7 @@
        paraformer_decoder_sanm=ParaformerSANMDecoder,
        paraformer_decoder_san=ParaformerDecoderSAN,
        contextual_paraformer_decoder=ContextualParaformerDecoder,
        sa_decoder=SAAsrTransformerDecoder,
    ),
    default="rnn",
)
@@ -219,24 +254,33 @@
    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,
]


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:
        token_list = None
        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':
        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)
@@ -245,7 +289,7 @@
        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:
@@ -257,7 +301,10 @@
    # normalization layer
    if args.normalize is not None:
        normalize_class = normalize_choices.get_class(args.normalize)
        normalize = normalize_class(**args.normalize_conf)
        if args.model == "mfcca":
            normalize = normalize_class(stats_file=args.cmvn_file, **args.normalize_conf)
        else:
            normalize = normalize_class(**args.normalize_conf)
    else:
        normalize = None

@@ -291,7 +338,8 @@
            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", "neatcontextual_paraformer"]:
        # predictor
        predictor_class = predictor_choices.get_class(args.predictor)
        predictor = predictor_class(**args.predictor_conf)
@@ -360,14 +408,19 @@
            **args.model_conf,
        )
    elif args.model == "timestamp_prediction":
        # predictor
        predictor_class = predictor_choices.get_class(args.predictor)
        predictor = predictor_class(**args.predictor_conf)
        
        model_class = model_choices.get_class(args.model)
        model = model_class(
            frontend=frontend,
            encoder=encoder,
            predictor=predictor,
            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 +449,48 @@
            **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,
        )
    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
        )

        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,
            )
        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))