export/vits/data_utils.py

import os
import random

import torch
import torchaudio
import torch.utils.data

import commons
from mel_processing import spectrogram_torch
from utils import load_filepaths_and_text


class TextAudioSpeakerLoader(torch.utils.data.Dataset):
    """
        1) loads audio, speaker_id, text pairs
        2) normalizes text and converts them to sequences of integers
        3) computes spectrograms from audio files.
    """
    def __init__(self, audiopaths_sid_text, hparams):
        self.audiopaths_sid_text = load_filepaths_and_text(audiopaths_sid_text)
        # self.text_cleaners = hparams.text_cleaners
        self.max_wav_value = hparams.max_wav_value
        self.sampling_rate = hparams.sampling_rate
        self.filter_length = hparams.filter_length
        self.hop_length = hparams.hop_length
        self.win_length = hparams.win_length
        self.sampling_rate = hparams.sampling_rate
        self.src_sampling_rate = getattr(hparams, "src_sampling_rate",
                                         self.sampling_rate)

        self.cleaned_text = getattr(hparams, "cleaned_text", False)

        self.add_blank = hparams.add_blank
        self.min_text_len = getattr(hparams, "min_text_len", 1)
        self.max_text_len = getattr(hparams, "max_text_len", 190)

        phone_file = getattr(hparams, "phone_table", None)
        self.phone_dict = None
        if phone_file is not None:
            self.phone_dict = {}
            with open(phone_file) as fin:
                for line in fin:
                    arr = line.strip().split()
                    self.phone_dict[arr[0]] = int(arr[1])

        speaker_file = getattr(hparams, "speaker_table", None)
        self.speaker_dict = None
        if speaker_file is not None:
            self.speaker_dict = {}
            with open(speaker_file) as fin:
                for line in fin:
                    arr = line.strip().split()
                    self.speaker_dict[arr[0]] = int(arr[1])

        random.seed(1234)
        random.shuffle(self.audiopaths_sid_text)
        self._filter()

    def _filter(self):
        """
        Filter text & store spec lengths
        """
        # Store spectrogram lengths for Bucketing
        # wav_length ~= file_size / (wav_channels * Bytes per dim) = file_size / (1 * 2)
        # spec_length = wav_length // hop_length

        audiopaths_sid_text_new = []
        lengths = []
        for item in self.audiopaths_sid_text:
            audiopath = item[0]
            # filename|text or filename|speaker|text
            text = item[1] if len(item) == 2 else item[2]
            if self.min_text_len <= len(text) and len(
                    text) <= self.max_text_len:
                audiopaths_sid_text_new.append(item)
                lengths.append(
                    int(
                        os.path.getsize(audiopath) * self.sampling_rate /
                        self.src_sampling_rate) // (2 * self.hop_length))
        self.audiopaths_sid_text = audiopaths_sid_text_new
        self.lengths = lengths

    def get_audio_text_speaker_pair(self, audiopath_sid_text):
        audiopath = audiopath_sid_text[0]
        if len(audiopath_sid_text) == 2:  # filename|text
            sid = 0
            text = audiopath_sid_text[1]
        else:  # filename|speaker|text
            sid = self.speaker_dict[audiopath_sid_text[1]]
            text = audiopath_sid_text[2]
        text = self.get_text(text)
        spec, wav = self.get_audio(audiopath)
        sid = self.get_sid(sid)
        return (text, spec, wav, sid)

    def get_audio(self, filename):
        audio, sampling_rate = torchaudio.load(filename, normalize=False)
        if sampling_rate != self.sampling_rate:
            audio = audio.to(torch.float)
            audio = torchaudio.transforms.Resample(sampling_rate,
                                                   self.sampling_rate)(audio)
            audio = audio.to(torch.int16)
        audio = audio[0]  # Get the first channel
        audio_norm = audio / self.max_wav_value
        audio_norm = audio_norm.unsqueeze(0)
        spec = spectrogram_torch(audio_norm,
                                 self.filter_length,
                                 self.sampling_rate,
                                 self.hop_length,
                                 self.win_length,
                                 center=False)
        spec = torch.squeeze(spec, 0)
        return spec, audio_norm

    def get_text(self, text):
        text_norm = [self.phone_dict[phone] for phone in text.split()]
        if self.add_blank:
            text_norm = commons.intersperse(text_norm, 0)
        text_norm = torch.LongTensor(text_norm)
        return text_norm

    def get_sid(self, sid):
        sid = torch.LongTensor([int(sid)])
        return sid

    def __getitem__(self, index):
        return self.get_audio_text_speaker_pair(
            self.audiopaths_sid_text[index])

    def __len__(self):
        return len(self.audiopaths_sid_text)


class TextAudioSpeakerCollate():
    """ Zero-pads model inputs and targets
    """
    def __init__(self, return_ids=False):
        self.return_ids = return_ids

    def __call__(self, batch):
        """Collate's training batch from normalized text, audio and speaker identities
        PARAMS
        ------
        batch: [text_normalized, spec_normalized, wav_normalized, sid]
        """
        # Right zero-pad all one-hot text sequences to max input length
        _, ids_sorted_decreasing = torch.sort(torch.LongTensor(
            [x[1].size(1) for x in batch]),
            dim=0,
            descending=True)

        max_text_len = max([len(x[0]) for x in batch])
        max_spec_len = max([x[1].size(1) for x in batch])
        max_wav_len = max([x[2].size(1) for x in batch])

        text_lengths = torch.LongTensor(len(batch))
        spec_lengths = torch.LongTensor(len(batch))
        wav_lengths = torch.LongTensor(len(batch))
        sid = torch.LongTensor(len(batch))

        text_padded = torch.LongTensor(len(batch), max_text_len)
        spec_padded = torch.FloatTensor(len(batch), batch[0][1].size(0),
                                        max_spec_len)
        wav_padded = torch.FloatTensor(len(batch), 1, max_wav_len)
        text_padded.zero_()
        spec_padded.zero_()
        wav_padded.zero_()
        for i in range(len(ids_sorted_decreasing)):
            row = batch[ids_sorted_decreasing[i]]

            text = row[0]
            text_padded[i, :text.size(0)] = text
            text_lengths[i] = text.size(0)

            spec = row[1]
            spec_padded[i, :, :spec.size(1)] = spec
            spec_lengths[i] = spec.size(1)

            wav = row[2]
            wav_padded[i, :, :wav.size(1)] = wav
            wav_lengths[i] = wav.size(1)

            sid[i] = row[3]

        if self.return_ids:
            return (text_padded, text_lengths, spec_padded, spec_lengths,
                    wav_padded, wav_lengths, sid, ids_sorted_decreasing)
        return (text_padded, text_lengths, spec_padded, spec_lengths,
                wav_padded, wav_lengths, sid)


class DistributedBucketSampler(torch.utils.data.distributed.DistributedSampler
                               ):
    """
    Maintain similar input lengths in a batch.
    Length groups are specified by boundaries.
    Ex) boundaries = [b1, b2, b3] -> any batch is included either
    {x | b1 < length(x) <=b2} or {x | b2 < length(x) <= b3}.

    It removes samples which are not included in the boundaries.
    Ex) boundaries = [b1, b2, b3] -> any x s.t. length(x) <= b1
    or length(x) > b3 are discarded.
    """
    def __init__(self,
                 dataset,
                 batch_size,
                 boundaries,
                 num_replicas=None,
                 rank=None,
                 shuffle=True):
        super().__init__(dataset,
                         num_replicas=num_replicas,
                         rank=rank,
                         shuffle=shuffle)
        self.lengths = dataset.lengths
        self.batch_size = batch_size
        self.boundaries = boundaries

        self.buckets, self.num_samples_per_bucket = self._create_buckets()
        self.total_size = sum(self.num_samples_per_bucket)
        self.num_samples = self.total_size // self.num_replicas

    def _create_buckets(self):
        buckets = [[] for _ in range(len(self.boundaries) - 1)]
        for i in range(len(self.lengths)):
            length = self.lengths[i]
            idx_bucket = self._bisect(length)
            if idx_bucket != -1:
                buckets[idx_bucket].append(i)

        for i in range(len(buckets) - 1, 0, -1):
            if len(buckets[i]) == 0:
                buckets.pop(i)
                self.boundaries.pop(i + 1)

        num_samples_per_bucket = []
        for i in range(len(buckets)):
            len_bucket = len(buckets[i])
            total_batch_size = self.num_replicas * self.batch_size
            rem = (total_batch_size -
                   (len_bucket % total_batch_size)) % total_batch_size
            num_samples_per_bucket.append(len_bucket + rem)
        return buckets, num_samples_per_bucket

    def __iter__(self):
        # deterministically shuffle based on epoch
        g = torch.Generator()
        g.manual_seed(self.epoch)

        indices = []
        if self.shuffle:
            for bucket in self.buckets:
                indices.append(
                    torch.randperm(len(bucket), generator=g).tolist())
        else:
            for bucket in self.buckets:
                indices.append(list(range(len(bucket))))

        batches = []
        for i in range(len(self.buckets)):
            bucket = self.buckets[i]
            len_bucket = len(bucket)
            ids_bucket = indices[i]
            num_samples_bucket = self.num_samples_per_bucket[i]

            # add extra samples to make it evenly divisible
            rem = num_samples_bucket - len_bucket
            ids_bucket = ids_bucket + ids_bucket * (
                rem // len_bucket) + ids_bucket[:(rem % len_bucket)]

            # subsample
            ids_bucket = ids_bucket[self.rank::self.num_replicas]

            # batching
            for j in range(len(ids_bucket) // self.batch_size):
                batch = [
                    bucket[idx]
                    for idx in ids_bucket[j * self.batch_size:(j + 1) *
                                          self.batch_size]
                ]
                batches.append(batch)

        if self.shuffle:
            batch_ids = torch.randperm(len(batches), generator=g).tolist()
            batches = [batches[i] for i in batch_ids]
        self.batches = batches

        assert len(self.batches) * self.batch_size == self.num_samples
        return iter(self.batches)

    def _bisect(self, x, lo=0, hi=None):
        if hi is None:
            hi = len(self.boundaries) - 1

        if hi > lo:
            mid = (hi + lo) // 2
            if self.boundaries[mid] < x and x <= self.boundaries[mid + 1]:
                return mid
            elif x <= self.boundaries[mid]:
                return self._bisect(x, lo, mid)
            else:
                return self._bisect(x, mid + 1, hi)
        else:
            return -1

    def __len__(self):
        return self.num_samples // self.batch_size