def test_nistems():
mp4exc = stempeg.cmds.find_cmd("MP4Box")
stems, rate = stempeg.read_stems(stempeg.example_stem_path())
with tmp.NamedTemporaryFile(delete=False, suffix='.m4a') as tempfile:
stempeg.write_stems(tempfile.name,
stems,
sample_rate=rate,
writer=stempeg.NIStemsWriter())
callArgs = [mp4exc]
callArgs.extend(["-dump-udta", "0:stem", tempfile.name])
sp.check_call(callArgs)
root, ext = os.path.splitext(tempfile.name)
udtaFile = root + "_stem.udta"
with open(stempeg.default_metadata()) as f:
d_metadata = json.load(f)
try:
fileObj = codecs.open(udtaFile, encoding="utf-8")
fileObj.seek(8)
l_metadata = json.load(fileObj)
except json.decoder.JSONDecodeError:
with open(udtaFile) as json_file:
l_metadata = json.load(json_file)
assert ordered(l_metadata) == ordered(d_metadata)
def test_shape(nb_samples):
R = np.random.random((5, nb_samples, 2))
stempeg.write_stems("./random.stem.m4a", R, writer=stempeg.StreamsWriter())
S, rate = stempeg.read_stems("./random.stem.m4a")
assert S.shape[0] == R.shape[0]
assert S.shape[2] == R.shape[2]
assert S.shape[1] % 1024 == 0
def test_multifileformats(audio, multifile_format, nb_stems):
with tmp.NamedTemporaryFile(delete=False,
suffix='.' + multifile_format) as tempfile:
stem_names = [str(k) for k in range(nb_stems)]
stempeg.write_stems(tempfile.name,
audio,
sample_rate=44100,
writer=stempeg.FilesWriter(stem_names=stem_names))
def test_shape(nb_samples):
R = np.random.random((5, nb_samples, 2))
stempeg.write_stems(R, "./random.stem.mp4")
S, rate = stempeg.read_stems("./random.stem.mp4")
assert S.shape[0] == R.shape[0]
assert S.shape[2] == R.shape[2]
assert S.shape[1] % 1024 == 0
def test_ffmpeg_errors(audio):
if audio.ndim == 3:
with pytest.raises(RuntimeError):
with tmp.NamedTemporaryFile(delete=False,
suffix='.wav') as tempfile:
stempeg.write_stems(tempfile.name,
audio,
sample_rate=44100,
writer=stempeg.StreamsWriter())
def test_multichannel_containers(audio, nb_channels, multichannel_format):
with tmp.NamedTemporaryFile(delete=False,
suffix='.' + multichannel_format) as tempfile:
stempeg.write_stems(tempfile.name,
audio,
sample_rate=44100,
writer=ChannelsWriter())
loaded_audio, rate = stempeg.read_stems(
tempfile.name,
always_3d=True,
reader=stempeg.ChannelsReader(nb_channels=nb_channels))
assert audio.shape == loaded_audio.shape
def test_multistream_containers(audio, multistream_format, nb_stems):
if nb_stems > 1:
with tmp.NamedTemporaryFile(delete=False,
suffix='.' +
multistream_format) as tempfile:
stem_names = [str(k) for k in range(nb_stems)]
stempeg.write_stems(tempfile.name,
audio,
sample_rate=44100,
writer=stempeg.StreamsWriter(
codec='aac', stem_names=stem_names))
loaded_audio, rate = stempeg.read_stems(tempfile.name,
always_3d=True)
assert audio.shape == loaded_audio.shape
if multistream_format == "m4a":
info = stempeg.Info(tempfile.name)
loaded_stem_names = info.title_streams
# check if titles could be extracted
assert all(
[a == b for a, b in zip(stem_names, loaded_stem_names)])
import stempeg
# 0 - The mixture,
# 1 - The drums,
# 2 - The bass,
# 3 - The rest of the accompaniment,
# 4 - The vocals.
# example
S, rate = stempeg.read_stems(stempeg.example_stem_path())
stempeg.write_stems(
"output.mp4",
S,
sample_rate=rate,
writer=stempeg.StreamsWriter())
S, rate = stempeg.read_stems("C:/Users/hahla/Downloads/output.mp4", stem_id=[0])
stems_folder = "D:/Development/github/GAN-tests/audio_files_split/audio_files_001"
filename_mix = "D:/Development/github/GAN-tests/audio_files_split/audio_file_mixture_0002.wav"
filename_drums = "D:/Development/github/GAN-tests/audio_files_split/audio_file_hits_0002.wav"
filename_bass = "D:/Development/github/GAN-tests/audio_files_split/audio_file_soundless_audio_0002.wav"
filename_other = "D:/Development/github/GAN-tests/audio_files_split/audio_file_background_0002.wav"
filename_vocals = "D:/Development/github/GAN-tests/audio_files_split/audio_file_soundless_audio_0002.wav"
S_filename_mix, rate = stempeg.read_stems(filename_mix, stem_id=0)
def test_write():
S, rate = stempeg.read_stems(
"tests/data/The Easton Ellises - Falcon 69.stem.mp4"
)
stempeg.write_stems(S, "./stems.mp4")
args = parser.parse_args()
# load stems
stems, rate = stempeg.read_stems(args.input)
# load stems,
# resample to 96000 Hz,
# use multiprocessing
stems, rate = stempeg.read_stems(args.input,
sample_rate=96000,
multiprocess=True)
# --> stems now has `shape=(stem x samples x channels)``
# save stems from tensor as multi-stream mp4
stempeg.write_stems("test.stem.m4a", stems, sample_rate=96000)
# save stems as dict for convenience
stems = {
"mix": stems[0],
"drums": stems[1],
"bass": stems[2],
"other": stems[3],
"vocals": stems[4],
}
# keys will be automatically used
# from dict as files
stempeg.write_stems("test.stem.m4a", data=stems, sample_rate=96000)
# `write_stems` is a preset for the following settings
def test_write():
S, rate = stempeg.read_stems(stempeg.example_stem_path())
stempeg.write_stems(S, "./stems.mp4")
def separate():
parser = argparse.ArgumentParser(
description="UMX Inference",
add_help=True,
formatter_class=argparse.RawDescriptionHelpFormatter,
)
parser.add_argument("input",
type=str,
nargs="+",
help="List of paths to wav/flac files.")
parser.add_argument(
"--model",
default="umxhq",
type=str,
help="path to mode base directory of pretrained models",
)
parser.add_argument(
"--targets",
nargs="+",
type=str,
help="provide targets to be processed. \
If none, all available targets will be computed",
)
parser.add_argument(
"--outdir",
type=str,
help="Results path where audio evaluation results are stored",
)
parser.add_argument(
"--ext",
type=str,
default=".wav",
help="Output extension which sets the audio format",
)
parser.add_argument("--start",
type=float,
default=0.0,
help="Audio chunk start in seconds")
parser.add_argument(
"--duration",
type=float,
help="Audio chunk duration in seconds, negative values load full track",
)
parser.add_argument("--no-cuda",
action="store_true",
default=False,
help="disables CUDA inference")
parser.add_argument(
"--audio-backend",
type=str,
default="sox_io",
help="Set torchaudio backend "
"(`sox_io`, `sox`, `soundfile` or `stempeg`), defaults to `sox_io`",
)
parser.add_argument(
"--niter",
type=int,
default=1,
help="number of iterations for refining results.",
)
parser.add_argument(
"--wiener-win-len",
type=int,
default=300,
help="Number of frames on which to apply filtering independently",
)
parser.add_argument(
"--residual",
type=str,
default=None,
help="if provided, build a source with given name"
"for the mix minus all estimated targets",
)
parser.add_argument(
"--aggregate",
type=str,
default=None,
help="if provided, must be a string containing a valid expression for "
"a dictionary, with keys as output target names, and values "
"a list of targets that are used to build it. For instance: "
'\'{"vocals":["vocals"], "accompaniment":["drums",'
'"bass","other"]}\'',
)
parser.add_argument(
"--filterbank",
type=str,
default="torch",
help="filterbank implementation method. "
"Supported: `['torch', 'asteroid']`. `torch` is ~30% faster"
"compared to `asteroid` on large FFT sizes such as 4096. However"
"asteroids stft can be exported to onnx, which makes is practical"
"for deployment.",
)
args = parser.parse_args()
if args.audio_backend != "stempeg":
torchaudio.set_audio_backend(args.audio_backend)
use_cuda = not args.no_cuda and torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
print("Using ", device)
# parsing the output dict
aggregate_dict = None if args.aggregate is None else json.loads(
args.aggregate)
# create separator only once to reduce model loading
# when using multiple files
separator = utils.load_separator(
model_str_or_path=args.model,
targets=args.targets,
niter=args.niter,
residual=args.residual,
wiener_win_len=args.wiener_win_len,
device=device,
pretrained=True,
filterbank=args.filterbank,
)
separator.freeze()
separator.to(device)
if args.audio_backend == "stempeg":
try:
import stempeg
except ImportError:
raise RuntimeError("Please install pip package `stempeg`")
# loop over the files
for input_file in args.input:
if args.audio_backend == "stempeg":
audio, rate = stempeg.read_stems(
input_file,
start=args.start,
duration=args.duration,
sample_rate=separator.sample_rate,
dtype=np.float32,
)
audio = torch.tensor(audio)
else:
audio, rate = data.load_audio(input_file,
start=args.start,
dur=args.duration)
estimates = predict.separate(
audio=audio,
rate=rate,
aggregate_dict=aggregate_dict,
separator=separator,
device=device,
)
if not args.outdir:
model_path = Path(args.model)
if not model_path.exists():
outdir = Path(Path(input_file).stem + "_" + args.model)
else:
outdir = Path(Path(input_file).stem + "_" + model_path.stem)
else:
outdir = Path(args.outdir) / Path(input_file).stem
outdir.mkdir(exist_ok=True, parents=True)
# write out estimates
if args.audio_backend == "stempeg":
target_path = str(outdir / Path("target").with_suffix(args.ext))
# convert torch dict to numpy dict
estimates_numpy = {}
for target, estimate in estimates.items():
estimates_numpy[target] = torch.squeeze(
estimate).detach().cpu().numpy().T
stempeg.write_stems(
target_path,
estimates_numpy,
sample_rate=separator.sample_rate,
writer=stempeg.FilesWriter(multiprocess=True,
output_sample_rate=rate),
)
else:
for target, estimate in estimates.items():
target_path = str(outdir / Path(target).with_suffix(args.ext))
torchaudio.save(
target_path,
torch.squeeze(estimate).to("cpu"),
sample_rate=separator.sample_rate,
)
"""Opens a stem file and saves (reencodes) back to a stem file
"""
import argparse
import stempeg
import numpy as np
from os import path as op
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('input', )
args = parser.parse_args()
# read stems
stems, rate = stempeg.read_stems(args.input)
print(stems.shape)
stempeg.write_stems(stems, "stems.mp4")
stems2, rate = stempeg.read_stems("stems.mp4")
print(stems2.shape)
