FunASR/funasr/tasks/data2vec.py

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 funasr.datasets.collate_fn import CommonCollateFn
from funasr.datasets.preprocessor import CommonPreprocessor
from funasr.layers.abs_normalize import AbsNormalize
from funasr.layers.global_mvn import GlobalMVN
from funasr.layers.utterance_mvn import UtteranceMVN
from funasr.models.data2vec import Data2VecPretrainModel
from funasr.models.encoder.abs_encoder import AbsEncoder
from funasr.models.encoder.data2vec_encoder import Data2VecEncoder
from funasr.models.frontend.abs_frontend import AbsFrontend
from funasr.models.frontend.default import DefaultFrontend
from funasr.models.frontend.windowing import SlidingWindow
from funasr.models.preencoder.abs_preencoder import AbsPreEncoder
from funasr.models.preencoder.sinc import LightweightSincConvs
from funasr.models.specaug.abs_specaug import AbsSpecAug
from funasr.models.specaug.specaug import SpecAug
from funasr.tasks.abs_task import AbsTask
from funasr.tokenizer.phoneme_tokenizer import g2p_choices
from funasr.torch_utils.initialize import initialize
from funasr.train.class_choices import ClassChoices
from funasr.train.trainer import Trainer
from funasr.utils.types import float_or_none
from funasr.utils.types import int_or_none
from funasr.utils.types import str2bool
from funasr.utils.types import str_or_none

frontend_choices = ClassChoices(
    name="frontend",
    classes=dict(default=DefaultFrontend, sliding_window=SlidingWindow),
    type_check=AbsFrontend,
    default="default",
)
specaug_choices = ClassChoices(
    name="specaug",
    classes=dict(specaug=SpecAug),
    type_check=AbsSpecAug,
    default=None,
    optional=True,
)
normalize_choices = ClassChoices(
    "normalize",
    classes=dict(
        global_mvn=GlobalMVN,
        utterance_mvn=UtteranceMVN,
    ),
    type_check=AbsNormalize,
    default=None,
    optional=True,
)
preencoder_choices = ClassChoices(
    name="preencoder",
    classes=dict(
        sinc=LightweightSincConvs,
    ),
    type_check=AbsPreEncoder,
    default=None,
    optional=True,
)
encoder_choices = ClassChoices(
    "encoder",
    classes=dict(
        data2vec_encoder=Data2VecEncoder,
    ),
    type_check=AbsEncoder,
    default="data2vec_encoder",
)
model_choices = ClassChoices(
    "model",
    classes=dict(
        data2vec=Data2VecPretrainModel,
    ),
    default="data2vec",
)


class Data2VecTask(AbsTask):
    # 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,
        # --preencoder and --preencoder_conf
        preencoder_choices,
        # --encoder and --encoder_conf
        encoder_choices,
        # --model and --model_conf
        model_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")

        # NOTE(kamo): add_arguments(..., required=True) can't be used
        # to provide --print_config mode. Instead of it, do as
        group.add_argument(
            "--token_list",
            type=str_or_none,
            default=None,
            help="A text mapping int-id to token",
        )
        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 = parser.add_argument_group(description="Preprocess related")
        group.add_argument(
            "--use_preprocessor",
            type=str2bool,
            default=True,
            help="Apply preprocessing to data or not",
        )
        group.add_argument(
            "--token_type",
            type=str,
            default=None,
            choices=["bpe", "char", "word", "phn"],
            help="The text will be tokenized " "in the specified level token",
        )

        group.add_argument(
            "--feats_type",
            type=str,
            default='fbank',
            help="feats type, e.g. fbank, wav, ark_wav(needed to be scale normalization)",
        )

        group.add_argument(
            "--bpemodel",
            type=str_or_none,
            default=None,
            help="The model file of sentencepiece",
        )
        parser.add_argument(
            "--non_linguistic_symbols",
            type=str_or_none,
            help="non_linguistic_symbols file path",
        )
        parser.add_argument(
            "--cleaner",
            type=str_or_none,
            choices=[None, "tacotron", "jaconv", "vietnamese"],
            default=None,
            help="Apply text cleaning",
        )
        parser.add_argument(
            "--g2p",
            type=str_or_none,
            choices=g2p_choices,
            default=None,
            help="Specify g2p method if --token_type=phn",
        )
        parser.add_argument(
            "--speech_volume_normalize",
            type=float_or_none,
            default=None,
            help="Scale the maximum amplitude to the given value.",
        )
        parser.add_argument(
            "--rir_scp",
            type=str_or_none,
            default=None,
            help="The file path of rir scp file.",
        )
        parser.add_argument(
            "--rir_apply_prob",
            type=float,
            default=1.0,
            help="THe probability for applying RIR convolution.",
        )
        parser.add_argument(
            "--noise_scp",
            type=str_or_none,
            default=None,
            help="The file path of noise scp file.",
        )
        parser.add_argument(
            "--noise_apply_prob",
            type=float,
            default=1.0,
            help="The probability applying Noise adding.",
        )
        parser.add_argument(
            "--noise_db_range",
            type=str,
            default="13_15",
            help="The range of noise decibel level.",
        )
        parser.add_argument(
            "--pred_masked_weight",
            type=float,
            default=1.0,
            help="weight for predictive loss for masked frames",
        )
        parser.add_argument(
            "--pred_nomask_weight",
            type=float,
            default=0.0,
            help="weight for predictive loss for unmasked frames",
        )
        parser.add_argument(
            "--loss_weights",
            type=float,
            default=0.0,
            help="weights for additional loss terms (not first one)",
        )

        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]],
    ]:
        return CommonCollateFn(clipping=True)

    @classmethod
    def build_preprocess_fn(
            cls, args: argparse.Namespace, train: bool
    ) -> Optional[Callable[[str, Dict[str, np.array]], Dict[str, np.ndarray]]]:
        if args.use_preprocessor:
            retval = CommonPreprocessor(
                train=train,
                bpemodel=args.bpemodel,
                non_linguistic_symbols=args.non_linguistic_symbols,
                text_cleaner=args.cleaner,
                g2p_type=args.g2p,
                # NOTE(kamo): Check attribute existence for backward compatibility
                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
        return retval

    @classmethod
    def required_data_names(
            cls, train: bool = True, inference: bool = False
    ) -> Tuple[str, ...]:
        # for pre-training
        retval = ("speech",)
        return retval

    @classmethod
    def optional_data_names(
            cls, train: bool = True, inference: bool = False
    ) -> Tuple[str, ...]:
        retval = ()
        return retval

    @classmethod
    def build_model(cls, args: argparse.Namespace):

        # 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. 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

        # 4. Pre-encoder input block
        # NOTE(kan-bayashi): Use getattr to keep the compatibility
        if getattr(args, "preencoder", None) is not None:
            preencoder_class = preencoder_choices.get_class(args.preencoder)
            preencoder = preencoder_class(**args.preencoder_conf)
            input_size = preencoder.output_size()
        else:
            preencoder = None

        # 5. Encoder
        encoder_class = encoder_choices.get_class(args.encoder)
        encoder = encoder_class(
            input_size=input_size,
            **args.encoder_conf,
        )

        # 6. Build model
        try:
            model_class = model_choices.get_class(args.model)
        except AttributeError:
            model_class = model_choices.get_class("data2vec")
        model = model_class(
            frontend=frontend,
            specaug=specaug,
            normalize=normalize,
            preencoder=preencoder,
            encoder=encoder,
        )

        # 7. Initialize
        if args.init is not None:
            initialize(model, args.init)

        return model