# Copyright      2021  Piotr Żelasko
#                2021  Xiaomi Corp.        (authors: Fangjun Kuang)
#
# See ../../../../LICENSE for clarification regarding multiple authors
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import argparse
import logging
from functools import lru_cache
from pathlib import Path
from typing import List

from lhotse import CutSet, Fbank, FbankConfig, load_manifest_lazy
from lhotse.dataset import (
    CutConcatenate,
    DynamicBucketingSampler,
    K2SpeechRecognitionDataset,
    PrecomputedFeatures,
    SingleCutSampler,
)
from lhotse.dataset.input_strategies import OnTheFlyFeatures
from torch.utils.data import DataLoader

from icefall.dataset.datamodule import DataModule
from icefall.utils import str2bool


class YesNoAsrDataModule(DataModule):
    """
    DataModule for k2 ASR experiments.
    It assumes there is always one train dataloader,
    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
    and test-other).

    It contains all the common data pipeline modules used in ASR
    experiments, e.g.:
    - dynamic batch size,
    - bucketing samplers,
    - cut concatenation,
    - augmentation,
    - on-the-fly feature extraction
    """

    @classmethod
    def add_arguments(cls, parser: argparse.ArgumentParser):
        super().add_arguments(parser)
        group = parser.add_argument_group(
            title="ASR data related options",
            description="These options are used for the preparation of "
            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
            "effective batch sizes, sampling strategies, applied data "
            "augmentations, etc.",
        )
        group.add_argument(
            "--feature-dir",
            type=Path,
            default=Path("data/fbank"),
            help="Path to directory with train/test cuts.",
        )
        group.add_argument(
            "--max-duration",
            type=int,
            default=30.0,
            help="Maximum pooled recordings duration (seconds) in a "
            "single batch. You can reduce it if it causes CUDA OOM.",
        )
        group.add_argument(
            "--bucketing-sampler",
            type=str2bool,
            default=False,
            help="When enabled, the batches will come from buckets of "
            "similar duration (saves padding frames).",
        )
        group.add_argument(
            "--num-buckets",
            type=int,
            default=10,
            help="The number of buckets for the DynamicBucketingSampler"
            "(you might want to increase it for larger datasets).",
        )
        group.add_argument(
            "--concatenate-cuts",
            type=str2bool,
            default=False,
            help="When enabled, utterances (cuts) will be concatenated "
            "to minimize the amount of padding.",
        )
        group.add_argument(
            "--duration-factor",
            type=float,
            default=1.0,
            help="Determines the maximum duration of a concatenated cut "
            "relative to the duration of the longest cut in a batch.",
        )
        group.add_argument(
            "--gap",
            type=float,
            default=1.0,
            help="The amount of padding (in seconds) inserted between "
            "concatenated cuts. This padding is filled with noise when "
            "noise augmentation is used.",
        )
        group.add_argument(
            "--on-the-fly-feats",
            type=str2bool,
            default=False,
            help="When enabled, use on-the-fly cut mixing and feature "
            "extraction. Will drop existing precomputed feature manifests "
            "if available.",
        )
        group.add_argument(
            "--shuffle",
            type=str2bool,
            default=False,
            help="When enabled (=default), the examples will be "
            "shuffled for each epoch.",
        )
        group.add_argument(
            "--return-cuts",
            type=str2bool,
            default=True,
            help="When enabled, each batch will have the "
            "field: batch['supervisions']['cut'] with the cuts that "
            "were used to construct it.",
        )

        group.add_argument(
            "--num-workers",
            type=int,
            default=2,
            help="The number of training dataloader workers that "
            "collect the batches.",
        )

    def train_dataloaders(self) -> DataLoader:
        logging.info("About to get train cuts")
        cuts_train = self.train_cuts()

        logging.info("About to create train dataset")
        transforms = []
        if self.args.concatenate_cuts:
            logging.info(
                f"Using cut concatenation with duration factor "
                f"{self.args.duration_factor} and gap {self.args.gap}."
            )
            # Cut concatenation should be the first transform in the list,
            # so that if we e.g. mix noise in, it will fill the gaps between
            # different utterances.
            transforms = [
                CutConcatenate(
                    duration_factor=self.args.duration_factor, gap=self.args.gap
                )
            ] + transforms

        train = K2SpeechRecognitionDataset(
            cut_transforms=transforms,
            return_cuts=self.args.return_cuts,
        )

        if self.args.on_the_fly_feats:
            # NOTE: the PerturbSpeed transform should be added only if we
            # remove it from data prep stage.
            # Add on-the-fly speed perturbation; since originally it would
            # have increased epoch size by 3, we will apply prob 2/3 and use
            # 3x more epochs.
            # Speed perturbation probably should come first before
            # concatenation, but in principle the transforms order doesn't have
            # to be strict (e.g. could be randomized)
            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
            # Drop feats to be on the safe side.
            train = K2SpeechRecognitionDataset(
                cut_transforms=transforms,
                input_strategy=OnTheFlyFeatures(
                    FbankConfig(sampling_rate=8000, num_mel_bins=23)
                ),
                return_cuts=self.args.return_cuts,
            )

        if self.args.bucketing_sampler:
            logging.info("Using DynamicBucketingSampler.")
            train_sampler = DynamicBucketingSampler(
                cuts_train,
                max_duration=self.args.max_duration,
                shuffle=self.args.shuffle,
                num_buckets=self.args.num_buckets,
                drop_last=True,
            )
        else:
            logging.info("Using SingleCutSampler.")
            train_sampler = SingleCutSampler(
                cuts_train,
                max_duration=self.args.max_duration,
                shuffle=self.args.shuffle,
            )
        logging.info("About to create train dataloader")

        train_dl = DataLoader(
            train,
            sampler=train_sampler,
            batch_size=None,
            num_workers=self.args.num_workers,
            persistent_workers=False,
        )

        return train_dl

    def test_dataloaders(self) -> DataLoader:
        logging.info("About to get test cuts")
        cuts_test = self.test_cuts()

        logging.debug("About to create test dataset")
        test = K2SpeechRecognitionDataset(
            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=23)))
            if self.args.on_the_fly_feats
            else PrecomputedFeatures(),
            return_cuts=self.args.return_cuts,
        )
        sampler = DynamicBucketingSampler(
            cuts_test,
            max_duration=self.args.max_duration,
            shuffle=False,
        )
        logging.debug("About to create test dataloader")
        test_dl = DataLoader(
            test,
            batch_size=None,
            sampler=sampler,
            num_workers=self.args.num_workers,
        )
        return test_dl

    @lru_cache()
    def train_cuts(self) -> CutSet:
        logging.info("About to get train cuts")
        cuts_train = load_manifest_lazy(
            self.args.feature_dir / "yesno_cuts_train.jsonl.gz"
        )
        return cuts_train

    @lru_cache()
    def test_cuts(self) -> List[CutSet]:
        logging.info("About to get test cuts")
        cuts_test = load_manifest_lazy(
            self.args.feature_dir / "yesno_cuts_test.jsonl.gz"
        )
        return cuts_test