File size: 8,385 Bytes
ad16788 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 |
import argparse
from typing import Callable
from typing import Collection
from typing import Dict
from typing import List
from typing import Optional
from typing import Tuple
import numpy as np
import torch
from typeguard import check_argument_types
from typeguard import check_return_type
from espnet2.asr.encoder.abs_encoder import AbsEncoder
from espnet2.asr.encoder.conformer_encoder import ConformerEncoder
from espnet2.asr.encoder.rnn_encoder import RNNEncoder
from espnet2.asr.encoder.transformer_encoder import TransformerEncoder
from espnet2.asr.frontend.abs_frontend import AbsFrontend
from espnet2.asr.frontend.default import DefaultFrontend
from espnet2.asr.frontend.windowing import SlidingWindow
from espnet2.diar.decoder.abs_decoder import AbsDecoder
from espnet2.diar.decoder.linear_decoder import LinearDecoder
from espnet2.diar.espnet_model import ESPnetDiarizationModel
from espnet2.layers.abs_normalize import AbsNormalize
from espnet2.layers.global_mvn import GlobalMVN
from espnet2.layers.label_aggregation import LabelAggregate
from espnet2.layers.utterance_mvn import UtteranceMVN
from espnet2.tasks.abs_task import AbsTask
from espnet2.torch_utils.initialize import initialize
from espnet2.train.class_choices import ClassChoices
from espnet2.train.collate_fn import CommonCollateFn
from espnet2.train.preprocessor import CommonPreprocessor
from espnet2.train.trainer import Trainer
from espnet2.utils.get_default_kwargs import get_default_kwargs
from espnet2.utils.nested_dict_action import NestedDictAction
from espnet2.utils.types import int_or_none
from espnet2.utils.types import str2bool
from espnet2.utils.types import str_or_none
frontend_choices = ClassChoices(
name="frontend",
classes=dict(default=DefaultFrontend, sliding_window=SlidingWindow),
type_check=AbsFrontend,
default="default",
)
normalize_choices = ClassChoices(
"normalize",
classes=dict(
global_mvn=GlobalMVN,
utterance_mvn=UtteranceMVN,
),
type_check=AbsNormalize,
default="utterance_mvn",
optional=True,
)
label_aggregator_choices = ClassChoices(
"label_aggregator",
classes=dict(label_aggregator=LabelAggregate),
default="label_aggregator",
)
encoder_choices = ClassChoices(
"encoder",
classes=dict(
conformer=ConformerEncoder,
transformer=TransformerEncoder,
rnn=RNNEncoder,
),
type_check=AbsEncoder,
default="rnn",
)
decoder_choices = ClassChoices(
"decoder",
classes=dict(linear=LinearDecoder),
type_check=AbsDecoder,
default="linear",
)
class DiarizationTask(AbsTask):
# If you need more than one optimizer, change this value
num_optimizers: int = 1
# Add variable objects configurations
class_choices_list = [
# --frontend and --frontend_conf
frontend_choices,
# --normalize and --normalize_conf
normalize_choices,
# --encoder and --encoder_conf
encoder_choices,
# --decoder and --decoder_conf
decoder_choices,
# --label_aggregator and --label_aggregator_conf
label_aggregator_choices,
]
# If you need to modify train() or eval() procedures, change Trainer class here
trainer = Trainer
@classmethod
def add_task_arguments(cls, parser: argparse.ArgumentParser):
group = parser.add_argument_group(description="Task related")
group.add_argument(
"--num_spk",
type=int_or_none,
default=None,
help="The number fo speakers (for each recording) used in system training",
)
group.add_argument(
"--init",
type=lambda x: str_or_none(x.lower()),
default=None,
help="The initialization method",
choices=[
"chainer",
"xavier_uniform",
"xavier_normal",
"kaiming_uniform",
"kaiming_normal",
None,
],
)
group.add_argument(
"--input_size",
type=int_or_none,
default=None,
help="The number of input dimension of the feature",
)
group.add_argument(
"--model_conf",
action=NestedDictAction,
default=get_default_kwargs(ESPnetDiarizationModel),
help="The keyword arguments for model class.",
)
group = parser.add_argument_group(description="Preprocess related")
group.add_argument(
"--use_preprocessor",
type=str2bool,
default=True,
help="Apply preprocessing to data or not",
)
for class_choices in cls.class_choices_list:
# Append --<name> and --<name>_conf.
# e.g. --encoder and --encoder_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]],
]:
assert check_argument_types()
# NOTE(kamo): int value = 0 is reserved by CTC-blank symbol
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]]]:
assert check_argument_types()
if args.use_preprocessor:
# FIXME (jiatong): add more arugment here
retval = CommonPreprocessor(train=train)
else:
retval = None
assert check_return_type(retval)
return retval
@classmethod
def required_data_names(
cls, train: bool = True, inference: bool = False
) -> Tuple[str, ...]:
if not inference:
retval = ("speech", "spk_labels")
else:
# Recognition mode
retval = ("speech",)
return retval
@classmethod
def optional_data_names(
cls, train: bool = True, inference: bool = False
) -> Tuple[str, ...]:
# (Note: jiatong): no optional data names for now
retval = ()
assert check_return_type(retval)
return retval
@classmethod
def build_model(cls, args: argparse.Namespace) -> ESPnetDiarizationModel:
assert check_argument_types()
# 1. frontend
if args.input_size is None:
# Extract features in the model
frontend_class = frontend_choices.get_class(args.frontend)
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. 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
# 3. Label Aggregator layer
label_aggregator_class = label_aggregator_choices.get_class(
args.label_aggregator
)
label_aggregator = label_aggregator_class(**args.label_aggregator_conf)
# 3. Encoder
encoder_class = encoder_choices.get_class(args.encoder)
# Note(jiatong): Diarization may not use subsampling when processing
encoder = encoder_class(input_size=input_size, **args.encoder_conf)
# 4. Decoder
decoder_class = decoder_choices.get_class(args.decoder)
decoder = decoder_class(
num_spk=args.num_spk,
encoder_output_size=encoder.output_size(),
**args.decoder_conf,
)
# 5. Build model
model = ESPnetDiarizationModel(
frontend=frontend,
normalize=normalize,
label_aggregator=label_aggregator,
encoder=encoder,
decoder=decoder,
**args.model_conf,
)
# FIXME(kamo): Should be done in model?
# 6. Initialize
if args.init is not None:
initialize(model, args.init)
assert check_return_type(model)
return model
|