MockingBird/models/vocoder/fregan/meldataset.py

import math
import os
import random
import torch
import torch.utils.data
import numpy as np
from librosa.util import normalize
from scipy.io.wavfile import read
from utils.audio_utils import mel_spectrogram

MAX_WAV_VALUE = 32768.0


def load_wav(full_path):
    sampling_rate, data = read(full_path)
    return data, sampling_rate


def get_dataset_filelist(a):
    #with open(a.input_training_file, 'r', encoding='utf-8') as fi:
    #    training_files = [os.path.join(a.input_wavs_dir, x.split('|')[0] + '.wav')
    #                      for x in fi.read().split('\n') if len(x) > 0]

    #with open(a.input_validation_file, 'r', encoding='utf-8') as fi:
    #   validation_files = [os.path.join(a.input_wavs_dir, x.split('|')[0] + '.wav')
    #                        for x in fi.read().split('\n') if len(x) > 0]
    files = os.listdir(a.input_wavs_dir)
    random.shuffle(files)
    files = [os.path.join(a.input_wavs_dir, f) for f in files]
    training_files = files[: -int(len(files) * 0.05)]
    validation_files = files[-int(len(files) * 0.05):]
    return training_files, validation_files


class MelDataset(torch.utils.data.Dataset):
    def __init__(self, training_files, segment_size, n_fft, num_mels,
                 hop_size, win_size, sampling_rate,  fmin, fmax, split=True, shuffle=True, n_cache_reuse=1,
                 device=None, fmax_loss=None, fine_tuning=False, base_mels_path=None):
        self.audio_files = training_files
        random.seed(1234)
        if shuffle:
            random.shuffle(self.audio_files)
        self.segment_size = segment_size
        self.sampling_rate = sampling_rate
        self.split = split
        self.n_fft = n_fft
        self.num_mels = num_mels
        self.hop_size = hop_size
        self.win_size = win_size
        self.fmin = fmin
        self.fmax = fmax
        self.fmax_loss = fmax_loss
        self.cached_wav = None
        self.n_cache_reuse = n_cache_reuse
        self._cache_ref_count = 0
        self.device = device
        self.fine_tuning = fine_tuning
        self.base_mels_path = base_mels_path

    def __getitem__(self, index):
        filename = self.audio_files[index]
        if self._cache_ref_count == 0:
            #audio, sampling_rate = load_wav(filename)
            #audio = audio / MAX_WAV_VALUE
            audio = np.load(filename)
            if not self.fine_tuning:
                audio = normalize(audio) * 0.95
            self.cached_wav = audio
            #if sampling_rate != self.sampling_rate:
            #    raise ValueError("{} SR doesn't match target {} SR".format(
            #        sampling_rate, self.sampling_rate))
            self._cache_ref_count = self.n_cache_reuse
        else:
            audio = self.cached_wav
            self._cache_ref_count -= 1

        audio = torch.FloatTensor(audio)
        audio = audio.unsqueeze(0)

        if not self.fine_tuning:
            if self.split:
                if audio.size(1) >= self.segment_size:
                    max_audio_start = audio.size(1) - self.segment_size
                    audio_start = random.randint(0, max_audio_start)
                    audio = audio[:, audio_start:audio_start+self.segment_size]
                else:
                    audio = torch.nn.functional.pad(audio, (0, self.segment_size - audio.size(1)), 'constant')

            mel = mel_spectrogram(audio, self.n_fft, self.num_mels,
                                  self.sampling_rate, self.hop_size, self.win_size, self.fmin, self.fmax,
                                  center=False)
        else:
            mel_path = os.path.join(self.base_mels_path, "mel" + "-" + filename.split("/")[-1].split("-")[-1])
            mel = np.load(mel_path).T
            #mel = np.load(
            #    os.path.join(self.base_mels_path, os.path.splitext(os.path.split(filename)[-1])[0] + '.npy'))
            mel = torch.from_numpy(mel)

            if len(mel.shape) < 3:
                mel = mel.unsqueeze(0)

            if self.split:
                frames_per_seg = math.ceil(self.segment_size / self.hop_size)

                if audio.size(1) >= self.segment_size:
                    mel_start = random.randint(0, mel.size(2) - frames_per_seg - 1)
                    mel = mel[:, :, mel_start:mel_start + frames_per_seg]
                    audio = audio[:, mel_start * self.hop_size:(mel_start + frames_per_seg) * self.hop_size]
                else:
                    mel = torch.nn.functional.pad(mel, (0, frames_per_seg - mel.size(2)), 'constant')
                    audio = torch.nn.functional.pad(audio, (0, self.segment_size - audio.size(1)), 'constant')

        mel_loss = mel_spectrogram(audio, self.n_fft, self.num_mels,
                                   self.sampling_rate, self.hop_size, self.win_size, self.fmin, self.fmax_loss,
                                   center=False)

        return (mel.squeeze(), audio.squeeze(0), filename, mel_loss.squeeze())

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