File size: 6,229 Bytes
7bc29af |
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 |
# Copyright (c) Facebook, Inc. and its affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import argparse
import os
from collections import defaultdict, namedtuple
from pathlib import Path
import musdb
import numpy as np
import torch as th
import tqdm
from torch.utils.data import DataLoader
from .audio import AudioFile
ChunkInfo = namedtuple("ChunkInfo", ["file_index", "offset", "local_index"])
class Rawset:
"""
Dataset of raw, normalized, float32 audio files
"""
def __init__(self, path, samples=None, stride=None, channels=2, streams=None):
self.path = Path(path)
self.channels = channels
self.samples = samples
if stride is None:
stride = samples if samples is not None else 0
self.stride = stride
entries = defaultdict(list)
for root, folders, files in os.walk(self.path, followlinks=True):
folders.sort()
files.sort()
for file in files:
if file.endswith(".raw"):
path = Path(root) / file
name, stream = path.stem.rsplit('.', 1)
entries[(path.parent.relative_to(self.path), name)].append(int(stream))
self._entries = list(entries.keys())
sizes = []
self._lengths = []
ref_streams = sorted(entries[self._entries[0]])
assert ref_streams == list(range(len(ref_streams)))
if streams is None:
self.streams = ref_streams
else:
self.streams = streams
for entry in sorted(entries.keys()):
streams = entries[entry]
assert sorted(streams) == ref_streams
file = self._path(*entry)
length = file.stat().st_size // (4 * channels)
if samples is None:
sizes.append(1)
else:
if length < samples:
self._entries.remove(entry)
continue
sizes.append((length - samples) // stride + 1)
self._lengths.append(length)
if not sizes:
raise ValueError(f"Empty dataset {self.path}")
self._cumulative_sizes = np.cumsum(sizes)
self._sizes = sizes
def __len__(self):
return self._cumulative_sizes[-1]
@property
def total_length(self):
return sum(self._lengths)
def chunk_info(self, index):
file_index = np.searchsorted(self._cumulative_sizes, index, side='right')
if file_index == 0:
local_index = index
else:
local_index = index - self._cumulative_sizes[file_index - 1]
return ChunkInfo(offset=local_index * self.stride,
file_index=file_index,
local_index=local_index)
def _path(self, folder, name, stream=0):
return self.path / folder / (name + f'.{stream}.raw')
def __getitem__(self, index):
chunk = self.chunk_info(index)
entry = self._entries[chunk.file_index]
length = self.samples or self._lengths[chunk.file_index]
streams = []
to_read = length * self.channels * 4
for stream_index, stream in enumerate(self.streams):
offset = chunk.offset * 4 * self.channels
file = open(self._path(*entry, stream=stream), 'rb')
file.seek(offset)
content = file.read(to_read)
assert len(content) == to_read
content = np.frombuffer(content, dtype=np.float32)
content = content.copy() # make writable
streams.append(th.from_numpy(content).view(length, self.channels).t())
return th.stack(streams, dim=0)
def name(self, index):
chunk = self.chunk_info(index)
folder, name = self._entries[chunk.file_index]
return folder / name
class MusDBSet:
def __init__(self, mus, streams=slice(None), samplerate=44100, channels=2):
self.mus = mus
self.streams = streams
self.samplerate = samplerate
self.channels = channels
def __len__(self):
return len(self.mus.tracks)
def __getitem__(self, index):
track = self.mus.tracks[index]
return (track.name, AudioFile(track.path).read(channels=self.channels,
seek_time=0,
streams=self.streams,
samplerate=self.samplerate))
def build_raw(mus, destination, normalize, workers, samplerate, channels):
destination.mkdir(parents=True, exist_ok=True)
loader = DataLoader(MusDBSet(mus, channels=channels, samplerate=samplerate),
batch_size=1,
num_workers=workers,
collate_fn=lambda x: x[0])
for name, streams in tqdm.tqdm(loader):
if normalize:
ref = streams[0].mean(dim=0) # use mono mixture as reference
streams = (streams - ref.mean()) / ref.std()
for index, stream in enumerate(streams):
open(destination / (name + f'.{index}.raw'), "wb").write(stream.t().numpy().tobytes())
def main():
parser = argparse.ArgumentParser('rawset')
parser.add_argument('--workers', type=int, default=10)
parser.add_argument('--samplerate', type=int, default=44100)
parser.add_argument('--channels', type=int, default=2)
parser.add_argument('musdb', type=Path)
parser.add_argument('destination', type=Path)
args = parser.parse_args()
build_raw(musdb.DB(root=args.musdb, subsets=["train"], split="train"),
args.destination / "train",
normalize=True,
channels=args.channels,
samplerate=args.samplerate,
workers=args.workers)
build_raw(musdb.DB(root=args.musdb, subsets=["train"], split="valid"),
args.destination / "valid",
normalize=True,
samplerate=args.samplerate,
channels=args.channels,
workers=args.workers)
if __name__ == "__main__":
main()
|