yuzu/spleeter
1
0
Fork 0
mirror of https://github.com/YuzuZensai/spleeter.git synced 2026-01-31 14:58:23 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #444 from deezer/pad_waveform

Browse Source 
Added padding at the begining to avoid tf STFT reconstruction error
This commit is contained in:
Moussallam
2020-07-24 15:28:30 +02:00
committed by GitHub
parent 4744ffbc84 3fcc4ea28f
commit ca5cdd7d28
4 changed files with 80 additions and 35 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
spleeter/commands/__init__.py
View File

@@ -170,6 +170,7 @@ def _create_evaluate_parser(parser_factory):
parser.add_argument('-o', '--output_path', **OPT_OUTPUT) parser.add_argument('-o', '--output_path', **OPT_OUTPUT)
parser.add_argument('--mus_dir', **OPT_MUSDB) parser.add_argument('--mus_dir', **OPT_MUSDB)
parser.add_argument('-m', '--mwf', **OPT_MWF) parser.add_argument('-m', '--mwf', **OPT_MWF)
parser.add_argument('-B', '--stft-backend', **OPT_STFT_BACKEND)
return parser return parser

2
spleeter/commands/evaluate.py
View File

@@ -77,7 +77,7 @@ def _separate_evaluation_dataset(arguments, musdb_root_directory, params):
bitrate='128k', bitrate='128k',
MWF=arguments.MWF, MWF=arguments.MWF,
verbose=arguments.verbose, verbose=arguments.verbose,
stft_backend="auto"), stft_backend=arguments.stft_backend),
params) params)
return audio_output_directory return audio_output_directory

11
spleeter/model/__init__.py
View File

@@ -275,9 +275,16 @@ class EstimatorSpecBuilder(object):
spec_name = self.spectrogram_name spec_name = self.spectrogram_name
if stft_name not in self._features: if stft_name not in self._features:
# pad input with a frame of zeros
waveform = tf.concat([
tf.zeros((self._frame_length, self._n_channels)),
self._features['waveform']
],
0
)
stft_feature = tf.transpose( stft_feature = tf.transpose(
stft( stft(
tf.transpose(self._features['waveform']), tf.transpose(waveform),
self._frame_length, self._frame_length,
self._frame_step, self._frame_step,
window_fn=lambda frame_length, dtype: ( window_fn=lambda frame_length, dtype: (
@@ -341,7 +348,7 @@ class EstimatorSpecBuilder(object):
reshaped = tf.transpose(inversed) reshaped = tf.transpose(inversed)
if time_crop is None: if time_crop is None:
time_crop = tf.shape(self._features['waveform'])[0] time_crop = tf.shape(self._features['waveform'])[0]
return reshaped[:time_crop, :] return reshaped[self._frame_length:self._frame_length+time_crop, :]
def _build_mwf_output_waveform(self): def _build_mwf_output_waveform(self):
""" Perform separation with multichannel Wiener Filtering using Norbert. """ Perform separation with multichannel Wiener Filtering using Norbert.

101
tests/test_eval.py
View File

@@ -25,33 +25,64 @@ from spleeter.commands import evaluate
from spleeter.utils.configuration import load_configuration from spleeter.utils.configuration import load_configuration
res_4stems = { "vocals": { BACKENDS = ["tensorflow", "librosa"]
"SDR": -0.007, TEST_CONFIGURATIONS = {el:el for el in BACKENDS}
"SAR": -19.231,
"SIR": -4.528, res_4stems = {
"ISR": 0.000 "librosa": {
"vocals": {
"SDR": -0.007,
"SAR": -19.231,
"SIR": -4.528,
"ISR": 0.000
},
"drums": {
"SDR": -0.071,
"SAR": -14.496,
"SIR": -4.987,
"ISR": 0.001
},
"bass":{
"SDR": -0.001,
"SAR": -12.426,
"SIR": -7.198,
"ISR": -0.001
},
"other":{
"SDR": -1.453,
"SAR": -14.899,
"SIR": -4.678,
"ISR": -0.015
}
}, },
"drums": { "tensorflow": {
"SDR": -0.071, "vocals": {
"SAR": -14.496, "SDR": 3.25e-05,
"SIR": -4.987, "SAR": -11.153575,
"ISR": 0.001 "SIR": -1.3849,
}, "ISR": 2.75e-05
"bass":{ },
"SDR": -0.001, "drums": {
"SAR": -12.426, "SDR": -0.079505,
"SIR": -7.198, "SAR": -15.7073575,
"ISR": -0.001 "SIR": -4.972755,
}, "ISR": 0.0013575
"other":{ },
"SDR": -1.453, "bass":{
"SAR": -14.899, "SDR": 2.5e-06,
"SIR": -4.678, "SAR": -10.3520575,
"ISR": -0.015 "SIR": -4.272325,
"ISR": 2.5e-06
},
"other":{
"SDR": -1.359175,
"SAR": -14.7076775,
"SIR": -4.761505,
"ISR": -0.01528
}
} }
} }
def generate_fake_eval_dataset(path): def generate_fake_eval_dataset(path):
aa = get_default_audio_adapter() aa = get_default_audio_adapter()
n_songs = 2 n_songs = 2
@@ -68,12 +99,18 @@ def generate_fake_eval_dataset(path):
aa.save(filename, data, fs) aa.save(filename, data, fs)
def test_evaluate(path="FAKE_MUSDB_DIR"): @pytest.mark.parametrize('backend', TEST_CONFIGURATIONS)
generate_fake_eval_dataset(path) def test_evaluate(backend):
p = create_argument_parser() with TemporaryDirectory() as directory:
arguments = p.parse_args(["evaluate", "-p", "spleeter:4stems", "--mus_dir", path])
params = load_configuration(arguments.configuration) generate_fake_eval_dataset(directory)
metrics = evaluate.entrypoint(arguments, params) p = create_argument_parser()
for instrument, metric in metrics.items(): arguments = p.parse_args(["evaluate", "-p", "spleeter:4stems", "--mus_dir", directory, "-B", backend])
for metric, value in metric.items(): params = load_configuration(arguments.configuration)
assert np.allclose(np.median(value), res_4stems[instrument][metric], atol=1e-3) metrics = evaluate.entrypoint(arguments, params)
for instrument, metric in metrics.items():
for metric, value in metric.items():
assert np.allclose(np.median(value), res_4stems[backend][instrument][metric], atol=1e-3)
# test_evaluate("tensorflow")