Spaces:
Runtime error
Runtime error
File size: 5,795 Bytes
5019931 |
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 |
import argparse
import os
import time
from concurrent.futures import ProcessPoolExecutor
from typing import NoReturn
import h5py
import librosa
import musdb
import numpy as np
from bytesep.utils import float32_to_int16
# Source types of the MUSDB18 dataset.
SOURCE_TYPES = ["vocals", "drums", "bass", "other", "accompaniment"]
def pack_audios_to_hdf5s(args) -> NoReturn:
r"""Pack (resampled) audio files into hdf5 files to speed up loading.
Args:
dataset_dir: str
subset: str, 'train' | 'test'
split: str, '' | 'train' | 'valid'
hdf5s_dir: str, directory to write out hdf5 files
sample_rate: int
channels_num: int
mono: bool
Returns:
NoReturn
"""
# arguments & parameters
dataset_dir = args.dataset_dir
subset = args.subset
split = None if args.split == "" else args.split
hdf5s_dir = args.hdf5s_dir
sample_rate = args.sample_rate
channels = args.channels
mono = True if channels == 1 else False
source_types = SOURCE_TYPES
resample_type = "kaiser_fast"
# Paths
os.makedirs(hdf5s_dir, exist_ok=True)
# Dataset of corresponding subset and split.
mus = musdb.DB(root=dataset_dir, subsets=[subset], split=split)
print("Subset: {}, Split: {}, Total pieces: {}".format(subset, split, len(mus)))
params = [] # A list of params for multiple processing.
for track_index in range(len(mus.tracks)):
param = (
dataset_dir,
subset,
split,
track_index,
source_types,
mono,
sample_rate,
resample_type,
hdf5s_dir,
)
params.append(param)
# Uncomment for debug.
# write_single_audio_to_hdf5(params[0])
# os._exit(0)
pack_hdf5s_time = time.time()
with ProcessPoolExecutor(max_workers=None) as pool:
# Maximum works on the machine
pool.map(write_single_audio_to_hdf5, params)
print("Pack hdf5 time: {:.3f} s".format(time.time() - pack_hdf5s_time))
def write_single_audio_to_hdf5(param) -> NoReturn:
r"""Write single audio into hdf5 file."""
(
dataset_dir,
subset,
split,
track_index,
source_types,
mono,
sample_rate,
resample_type,
hdf5s_dir,
) = param
# Dataset of corresponding subset and split.
mus = musdb.DB(root=dataset_dir, subsets=[subset], split=split)
track = mus.tracks[track_index]
# Path to write out hdf5 file.
hdf5_path = os.path.join(hdf5s_dir, "{}.h5".format(track.name))
with h5py.File(hdf5_path, "w") as hf:
hf.attrs.create("audio_name", data=track.name.encode(), dtype="S100")
hf.attrs.create("sample_rate", data=sample_rate, dtype=np.int32)
for source_type in source_types:
audio = track.targets[source_type].audio.T
# (channels_num, audio_samples)
# Preprocess audio to mono / stereo, and resample.
audio = preprocess_audio(
audio, mono, track.rate, sample_rate, resample_type
)
# audio = load_audio(audio_path=audio_path, mono=mono, sample_rate=sample_rate)
# (channels_num, audio_samples) | (audio_samples,)
hf.create_dataset(
name=source_type, data=float32_to_int16(audio), dtype=np.int16
)
# Mixture
audio = track.audio.T
# (channels_num, audio_samples)
# Preprocess audio to mono / stereo, and resample.
audio = preprocess_audio(audio, mono, track.rate, sample_rate, resample_type)
# (channels_num, audio_samples)
hf.create_dataset(name="mixture", data=float32_to_int16(audio), dtype=np.int16)
print("{} Write to {}, {}".format(track_index, hdf5_path, audio.shape))
def preprocess_audio(audio, mono, origin_sr, sr, resample_type) -> np.array:
r"""Preprocess audio to mono / stereo, and resample.
Args:
audio: (channels_num, audio_samples), input audio
mono: bool
origin_sr: float, original sample rate
sr: float, target sample rate
resample_type: str, e.g., 'kaiser_fast'
Returns:
output: ndarray, output audio
"""
if mono:
audio = np.mean(audio, axis=0)
# (audio_samples,)
output = librosa.core.resample(
audio, orig_sr=origin_sr, target_sr=sr, res_type=resample_type
)
# (audio_samples,) | (channels_num, audio_samples)
if output.ndim == 1:
output = output[None, :]
# (1, audio_samples,)
return output
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument(
"--dataset_dir",
type=str,
required=True,
help="Directory of the MUSDB18 dataset.",
)
parser.add_argument(
"--subset",
type=str,
required=True,
choices=["train", "test"],
help="Train subset: 100 pieces; test subset: 50 pieces.",
)
parser.add_argument(
"--split",
type=str,
required=True,
choices=["", "train", "valid"],
help="Use '' to use all 100 pieces to train. Use 'train' to use 86 \
pieces for train, and use 'test' to use 14 pieces for valid.",
)
parser.add_argument(
"--hdf5s_dir",
type=str,
required=True,
help="Directory to write out hdf5 files.",
)
parser.add_argument("--sample_rate", type=int, required=True, help="Sample rate.")
parser.add_argument(
"--channels", type=int, required=True, help="Use 1 for mono, 2 for stereo."
)
# Parse arguments.
args = parser.parse_args()
# Pack audios into hdf5 files.
pack_audios_to_hdf5s(args)
|