From 54931dd4e1a099d7d6f144c4e12e5453deb3aa26 Mon Sep 17 00:00:00 2001
From: 雾聪 <wucong.lyb@alibaba-inc.com>
Date: 星期三, 28 六月 2023 10:41:57 +0800
Subject: [PATCH] Merge branch 'main' of https://github.com/alibaba-damo-academy/FunASR into main
---
funasr/tasks/asr.py | 455 +++++++++++++++++++++++---------------------------------
1 files changed, 185 insertions(+), 270 deletions(-)
diff --git a/funasr/tasks/asr.py b/funasr/tasks/asr.py
index d218902..7338513 100644
--- a/funasr/tasks/asr.py
+++ b/funasr/tasks/asr.py
@@ -38,6 +38,7 @@
from funasr.models.decoder.transformer_decoder import ParaformerDecoderSAN
from funasr.models.decoder.transformer_decoder import TransformerDecoder
from funasr.models.decoder.contextual_decoder import ContextualParaformerDecoder
+from funasr.models.decoder.transformer_decoder import SAAsrTransformerDecoder
from funasr.models.e2e_asr import ASRModel
from funasr.models.decoder.rnnt_decoder import RNNTDecoder
from funasr.models.joint_net.joint_network import JointNetwork
@@ -45,6 +46,7 @@
from funasr.models.e2e_asr_contextual_paraformer import NeatContextualParaformer
from funasr.models.e2e_tp import TimestampPredictor
from funasr.models.e2e_asr_mfcca import MFCCA
+from funasr.models.e2e_sa_asr import SAASRModel
from funasr.models.e2e_uni_asr import UniASR
from funasr.models.e2e_asr_transducer import TransducerModel, UnifiedTransducerModel
from funasr.models.encoder.abs_encoder import AbsEncoder
@@ -54,6 +56,7 @@
from funasr.models.encoder.sanm_encoder import SANMEncoder, SANMEncoderChunkOpt
from funasr.models.encoder.transformer_encoder import TransformerEncoder
from funasr.models.encoder.mfcca_encoder import MFCCAEncoder
+from funasr.models.encoder.resnet34_encoder import ResNet34Diar
from funasr.models.frontend.abs_frontend import AbsFrontend
from funasr.models.frontend.default import DefaultFrontend
from funasr.models.frontend.default import MultiChannelFrontend
@@ -132,6 +135,9 @@
neatcontextual_paraformer=NeatContextualParaformer,
mfcca=MFCCA,
timestamp_prediction=TimestampPredictor,
+ rnnt=TransducerModel,
+ rnnt_unified=UnifiedTransducerModel,
+ sa_asr=SAASRModel,
),
type_check=FunASRModel,
default="asr",
@@ -173,6 +179,27 @@
type_check=AbsEncoder,
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,
+ ),
+ type_check=AbsEncoder,
+ default="rnn",
+)
+spk_encoder_choices = ClassChoices(
+ "spk_encoder",
+ classes=dict(
+ resnet34_diar=ResNet34Diar,
+ ),
+ default="resnet34_diar",
+)
postencoder_choices = ClassChoices(
name="postencoder",
classes=dict(
@@ -195,6 +222,7 @@
paraformer_decoder_sanm=ParaformerSANMDecoder,
paraformer_decoder_san=ParaformerDecoderSAN,
contextual_paraformer_decoder=ContextualParaformerDecoder,
+ sa_decoder=SAAsrTransformerDecoder,
),
type_check=AbsDecoder,
default="rnn",
@@ -222,6 +250,15 @@
),
type_check=RNNTDecoder,
default="rnnt",
+)
+
+joint_network_choices = ClassChoices(
+ name="joint_network",
+ classes=dict(
+ joint_network=JointNetwork,
+ ),
+ default="joint_network",
+ optional=True,
)
predictor_choices = ClassChoices(
@@ -290,6 +327,8 @@
predictor_choices2,
# --stride_conv and --stride_conv_conf
stride_conv_choices,
+ # --rnnt_decoder and --rnnt_decoder_conf
+ rnnt_decoder_choices,
]
# If you need to modify train() or eval() procedures, change Trainer class here
@@ -315,6 +354,12 @@
type=str2bool,
default=True,
help="whether to split text using <space>",
+ )
+ group.add_argument(
+ "--max_spk_num",
+ type=int_or_none,
+ default=None,
+ help="A text mapping int-id to token",
)
group.add_argument(
"--seg_dict_file",
@@ -349,12 +394,6 @@
action=NestedDictAction,
default=get_default_kwargs(CTC),
help="The keyword arguments for CTC class.",
- )
- group.add_argument(
- "--joint_net_conf",
- action=NestedDictAction,
- default=None,
- help="The keyword arguments for joint network class.",
)
group = parser.add_argument_group(description="Preprocess related")
@@ -1360,254 +1399,22 @@
return retval
-class ASRTransducerTask(AbsTask):
+class ASRTransducerTask(ASRTask):
"""ASR Transducer Task definition."""
num_optimizers: int = 1
class_choices_list = [
+ model_choices,
frontend_choices,
specaug_choices,
normalize_choices,
encoder_choices,
rnnt_decoder_choices,
+ joint_network_choices,
]
trainer = Trainer
-
- @classmethod
- def add_task_arguments(cls, parser: argparse.ArgumentParser):
- """Add Transducer task arguments.
- Args:
- cls: ASRTransducerTask object.
- parser: Transducer arguments parser.
- """
- group = parser.add_argument_group(description="Task related.")
-
- # required = parser.get_default("required")
- # required += ["token_list"]
-
- group.add_argument(
- "--token_list",
- type=str_or_none,
- default=None,
- help="Integer-string mapper for tokens.",
- )
- group.add_argument(
- "--split_with_space",
- type=str2bool,
- default=True,
- help="whether to split text using <space>",
- )
- group.add_argument(
- "--input_size",
- type=int_or_none,
- default=None,
- help="The number of dimensions for input features.",
- )
- group.add_argument(
- "--init",
- type=str_or_none,
- default=None,
- help="Type of model initialization to use.",
- )
- group.add_argument(
- "--model_conf",
- action=NestedDictAction,
- default=get_default_kwargs(TransducerModel),
- help="The keyword arguments for the model class.",
- )
- # group.add_argument(
- # "--encoder_conf",
- # action=NestedDictAction,
- # default={},
- # help="The keyword arguments for the encoder class.",
- # )
- group.add_argument(
- "--joint_network_conf",
- action=NestedDictAction,
- default={},
- help="The keyword arguments for the joint network class.",
- )
- group = parser.add_argument_group(description="Preprocess related.")
- group.add_argument(
- "--use_preprocessor",
- type=str2bool,
- default=True,
- help="Whether to apply preprocessing to input data.",
- )
- group.add_argument(
- "--token_type",
- type=str,
- default="bpe",
- choices=["bpe", "char", "word", "phn"],
- help="The type of tokens to use during tokenization.",
- )
- group.add_argument(
- "--bpemodel",
- type=str_or_none,
- default=None,
- help="The path of the sentencepiece model.",
- )
- parser.add_argument(
- "--non_linguistic_symbols",
- type=str_or_none,
- help="The 'non_linguistic_symbols' file path.",
- )
- parser.add_argument(
- "--cleaner",
- type=str_or_none,
- choices=[None, "tacotron", "jaconv", "vietnamese"],
- default=None,
- help="Text cleaner to use.",
- )
- parser.add_argument(
- "--g2p",
- type=str_or_none,
- choices=g2p_choices,
- default=None,
- help="g2p method to use if --token_type=phn.",
- )
- parser.add_argument(
- "--speech_volume_normalize",
- type=float_or_none,
- default=None,
- help="Normalization value for maximum amplitude scaling.",
- )
- parser.add_argument(
- "--rir_scp",
- type=str_or_none,
- default=None,
- help="The RIR SCP file path.",
- )
- parser.add_argument(
- "--rir_apply_prob",
- type=float,
- default=1.0,
- help="The probability of the applied RIR convolution.",
- )
- parser.add_argument(
- "--noise_scp",
- type=str_or_none,
- default=None,
- help="The path of noise SCP file.",
- )
- parser.add_argument(
- "--noise_apply_prob",
- type=float,
- default=1.0,
- help="The probability of the applied noise addition.",
- )
- parser.add_argument(
- "--noise_db_range",
- type=str,
- default="13_15",
- help="The range of the noise decibel level.",
- )
- for class_choices in cls.class_choices_list:
- # Append --<name> and --<name>_conf.
- # e.g. --decoder and --decoder_conf
- class_choices.add_arguments(group)
-
- @classmethod
- def build_collate_fn(
- cls, args: argparse.Namespace, train: bool
- ) -> Callable[
- [Collection[Tuple[str, Dict[str, np.ndarray]]]],
- Tuple[List[str], Dict[str, torch.Tensor]],
- ]:
- """Build collate function.
- Args:
- cls: ASRTransducerTask object.
- args: Task arguments.
- train: Training mode.
- Return:
- : Callable collate function.
- """
- assert check_argument_types()
-
- return CommonCollateFn(float_pad_value=0.0, int_pad_value=-1)
-
- @classmethod
- def build_preprocess_fn(
- cls, args: argparse.Namespace, train: bool
- ) -> Optional[Callable[[str, Dict[str, np.array]], Dict[str, np.ndarray]]]:
- """Build pre-processing function.
- Args:
- cls: ASRTransducerTask object.
- args: Task arguments.
- train: Training mode.
- Return:
- : Callable pre-processing function.
- """
- assert check_argument_types()
-
- if args.use_preprocessor:
- retval = CommonPreprocessor(
- train=train,
- token_type=args.token_type,
- token_list=args.token_list,
- bpemodel=args.bpemodel,
- non_linguistic_symbols=args.non_linguistic_symbols,
- text_cleaner=args.cleaner,
- g2p_type=args.g2p,
- split_with_space=args.split_with_space if hasattr(args, "split_with_space") else False,
- rir_scp=args.rir_scp if hasattr(args, "rir_scp") else None,
- rir_apply_prob=args.rir_apply_prob
- if hasattr(args, "rir_apply_prob")
- else 1.0,
- noise_scp=args.noise_scp if hasattr(args, "noise_scp") else None,
- noise_apply_prob=args.noise_apply_prob
- if hasattr(args, "noise_apply_prob")
- else 1.0,
- noise_db_range=args.noise_db_range
- if hasattr(args, "noise_db_range")
- else "13_15",
- speech_volume_normalize=args.speech_volume_normalize
- if hasattr(args, "rir_scp")
- else None,
- )
- else:
- retval = None
-
- assert check_return_type(retval)
- return retval
-
- @classmethod
- def required_data_names(
- cls, train: bool = True, inference: bool = False
- ) -> Tuple[str, ...]:
- """Required data depending on task mode.
- Args:
- cls: ASRTransducerTask object.
- train: Training mode.
- inference: Inference mode.
- Return:
- retval: Required task data.
- """
- if not inference:
- retval = ("speech", "text")
- else:
- retval = ("speech",)
-
- return retval
-
- @classmethod
- def optional_data_names(
- cls, train: bool = True, inference: bool = False
- ) -> Tuple[str, ...]:
- """Optional data depending on task mode.
- Args:
- cls: ASRTransducerTask object.
- train: Training mode.
- inference: Inference mode.
- Return:
- retval: Optional task data.
- """
- retval = ()
- assert check_return_type(retval)
-
- return retval
@classmethod
def build_model(cls, args: argparse.Namespace) -> TransducerModel:
@@ -1675,7 +1482,7 @@
decoder_output_size = decoder.output_size
if getattr(args, "decoder", None) is not None:
- att_decoder_class = decoder_choices.get_class(args.att_decoder)
+ att_decoder_class = decoder_choices.get_class(args.decoder)
att_decoder = att_decoder_class(
vocab_size=vocab_size,
@@ -1693,35 +1500,23 @@
)
# 7. Build model
+ try:
+ model_class = model_choices.get_class(args.model)
+ except AttributeError:
+ model_class = model_choices.get_class("rnnt_unified")
- 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,
- )
-
- 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 = 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,
+ )
# 8. Initialize model
if args.init is not None:
raise NotImplementedError(
@@ -1732,3 +1527,123 @@
#assert check_return_type(model)
return model
+
+
+class ASRTaskSAASR(ASRTask):
+ # If you need more than one optimizers, change this value
+ num_optimizers: int = 1
+
+ # Add variable objects configurations
+ class_choices_list = [
+ # --frontend and --frontend_conf
+ frontend_choices,
+ # --specaug and --specaug_conf
+ specaug_choices,
+ # --normalize and --normalize_conf
+ normalize_choices,
+ # --model and --model_conf
+ model_choices,
+ # --preencoder and --preencoder_conf
+ preencoder_choices,
+ # --encoder and --encoder_conf
+ # --asr_encoder and --asr_encoder_conf
+ asr_encoder_choices,
+ # --spk_encoder and --spk_encoder_conf
+ spk_encoder_choices,
+ # --decoder and --decoder_conf
+ decoder_choices,
+ ]
+
+ # If you need to modify train() or eval() procedures, change Trainer class here
+ trainer = Trainer
+
+ @classmethod
+ def build_model(cls, args: argparse.Namespace):
+ assert check_argument_types()
+ if isinstance(args.token_list, str):
+ with open(args.token_list, encoding="utf-8") as f:
+ token_list = [line.rstrip() for line in f]
+
+ # Overwriting token_list to keep it as "portable".
+ args.token_list = list(token_list)
+ elif isinstance(args.token_list, (tuple, list)):
+ token_list = list(args.token_list)
+ else:
+ raise RuntimeError("token_list must be str or list")
+ vocab_size = len(token_list)
+ logging.info(f"Vocabulary size: {vocab_size}")
+
+ # 1. frontend
+ if args.input_size is None:
+ # Extract features in the model
+ frontend_class = frontend_choices.get_class(args.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)
+ input_size = frontend.output_size()
+ else:
+ # Give features from data-loader
+ args.frontend = None
+ args.frontend_conf = {}
+ frontend = None
+ input_size = args.input_size
+
+ # 2. Data augmentation for spectrogram
+ if args.specaug is not None:
+ specaug_class = specaug_choices.get_class(args.specaug)
+ specaug = specaug_class(**args.specaug_conf)
+ else:
+ specaug = None
+
+ # 3. Normalization layer
+ if args.normalize is not None:
+ normalize_class = normalize_choices.get_class(args.normalize)
+ normalize = normalize_class(**args.normalize_conf)
+ else:
+ normalize = None
+
+ # 5. Encoder
+ 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)
+
+ # 7. Decoder
+ decoder_class = decoder_choices.get_class(args.decoder)
+ decoder = decoder_class(
+ vocab_size=vocab_size,
+ encoder_output_size=asr_encoder.output_size(),
+ **args.decoder_conf,
+ )
+
+ # 8. CTC
+ ctc = CTC(
+ odim=vocab_size, encoder_output_size=asr_encoder.output_size(), **args.ctc_conf
+ )
+
+ # import ipdb;ipdb.set_trace()
+ # 9. Build model
+ try:
+ model_class = model_choices.get_class(args.model)
+ except AttributeError:
+ model_class = model_choices.get_class("asr")
+ 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,
+ )
+
+ # 10. Initialize
+ if args.init is not None:
+ initialize(model, args.init)
+
+ assert check_return_type(model)
+ return model
--
Gitblit v1.9.1