def separate_and_evaluate(
track,
targets,
model_name,
niter,
alpha,
softmask,
output_dir,
eval_dir,
device='cpu'
):
estimates = test.separate(
audio=track.audio,
targets=targets,
model_name=model_name,
niter=niter,
alpha=alpha,
softmask=softmask,
device=device
)
if output_dir:
mus.save_estimates(estimates, track, output_dir)
scores = museval.eval_mus_track(
track, estimates, output_dir=eval_dir
)
return scores
def separate_and_evaluate(track, args, ext):
estimates = test.separate(track.audio, args)
if args.out_dir:
mus.save_estimates(estimates, track, args.out_dir)
scores = museval.eval_mus_track(track, estimates, output_dir=args.out_dir)
# clear cache memory
ext.clear_memory_cache()
return scores
def separate_and_evaluate(
track,
model,
niter,
alpha,
softmask,
output_dir,
eval_dir,
):
estimates = test.separate(audio=track.audio,
model_path=model,
niter=niter,
alpha=alpha,
softmask=softmask)
if output_dir:
mus.save_estimates(estimates, track, output_dir)
scores = museval.eval_mus_track(track, estimates, output_dir=eval_dir)
return scores
root=root,
download=root is None,
subsets=subset,
is_wav=is_wav
)
# iterate over all tracks present in test folder
for track in mus.tracks:
outdir = exp_model_res_path + track.name + '/'
print(track.name)
estimates = test.separate(
audio=track.audio,
targets=targets,
model_name=model,
niter=2,
alpha=1,
softmask=False,
device=device
)
for target, estimate in estimates.items():
sf.write(
outdir / Path(target).with_suffix('.wav'),
estimate,
samplerate
)
print("SAVED SEPARATED VOCALS AND ACCOMPANIMENTS!")
audio_estimates = []
"oops, output directory {} already exists - delete this first or change outdir"
.format(basename))
sys.exit(1)
os.makedirs(basename)
# https://www.youtube.com/watch?v=DQLUygS0IAQ
# https://www.youtube.com/watch?v=J9gKyRmic20
use_cuda = torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
audio, rate = librosa.load(infile, sr=44100, mono=False)
estimates = test.separate(audio=audio.T,
targets=['vocals', 'drums', 'bass', 'other'],
device=device,
residual_model=False)
# estimates['vocals'].shape = (11953152, 2)
# audio.shape = (2, 11954040)
# import soundfile as sf
# sf.available_formats()
# librosa.output.write_wav('out_west.wav', audio, rate)
# librosa.output.write_wav('out_west.mp3', audio, rate)
librosa.output.write_wav('{}/out0_all.wav'.format(basename), audio, rate)
librosa.output.write_wav('{}/out1_vocals.wav'.format(basename),
estimates['vocals'], rate)
librosa.output.write_wav('{}/out2_drums.wav'.format(basename),