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 #426 from deezer/fix_librosa_istft_edge

Browse Source 
Fixing gltches issues with Istft
This commit is contained in:
Moussallam
2020-07-24 17:05:58 +02:00
committed by GitHub
parent ca5cdd7d28 c6cc510069
commit 47b990e5f2
4 changed files with 60 additions and 58 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
CHANGELOG.md
View File

@@ -1,5 +1,12 @@
# Changelog History # Changelog History
## 1.5.4
First release, July 24th 2020
Add some padding of the input waveform to avoid separation artefacts on the edges due to unstabilities in the inverse fourier transforms.
Also add tests to ensure both librosa and tensorflow backends have same outputs.
## 1.5.2 ## 1.5.2
First released, May 15th 2020 First released, May 15th 2020

8
spleeter/separator.py
View File

@@ -123,14 +123,18 @@ class Separator(object):
data = np.asfortranarray(data) data = np.asfortranarray(data)
N = self._params["frame_length"] N = self._params["frame_length"]
H = self._params["frame_step"] H = self._params["frame_step"]
win = hann(N, sym=False) win = hann(N, sym=False)
fstft = istft if inverse else stft fstft = istft if inverse else stft
win_len_arg = {"win_length": None, "length": length} if inverse else {"n_fft": N} win_len_arg = {"win_length": None,
"length": None} if inverse else {"n_fft": N}
n_channels = data.shape[-1] n_channels = data.shape[-1]
out = [] out = []
for c in range(n_channels): for c in range(n_channels):
d = data[:, :, c].T if inverse else data[:, c] d = np.concatenate((np.zeros((N, )), data[:, c], np.zeros((N, )))) if not inverse else data[:, :, c].T
s = fstft(d, hop_length=H, window=win, center=False, **win_len_arg) s = fstft(d, hop_length=H, window=win, center=False, **win_len_arg)
if inverse:
s = s[N:N+length]
s = np.expand_dims(s.T, 2-inverse) s = np.expand_dims(s.T, 2-inverse)
out.append(s) out.append(s)
if len(out) == 1: if len(out) == 1:

80
tests/test_eval.py
View File

@@ -29,57 +29,29 @@ BACKENDS = ["tensorflow", "librosa"]
TEST_CONFIGURATIONS = {el:el for el in BACKENDS} TEST_CONFIGURATIONS = {el:el for el in BACKENDS}
res_4stems = { res_4stems = {
"librosa": { "vocals": {
"vocals": { "SDR": 3.25e-05,
"SDR": -0.007, "SAR": -11.153575,
"SAR": -19.231, "SIR": -1.3849,
"SIR": -4.528, "ISR": 2.75e-05
"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
}
}, },
"tensorflow": { "drums": {
"vocals": { "SDR": -0.079505,
"SDR": 3.25e-05, "SAR": -15.7073575,
"SAR": -11.153575, "SIR": -4.972755,
"SIR": -1.3849, "ISR": 0.0013575
"ISR": 2.75e-05 },
}, "bass":{
"drums": { "SDR": 2.5e-06,
"SDR": -0.079505, "SAR": -10.3520575,
"SAR": -15.7073575, "SIR": -4.272325,
"SIR": -4.972755, "ISR": 2.5e-06
"ISR": 0.0013575 },
}, "other":{
"bass":{ "SDR": -1.359175,
"SDR": 2.5e-06, "SAR": -14.7076775,
"SAR": -10.3520575, "SIR": -4.761505,
"SIR": -4.272325, "ISR": -0.01528
"ISR": 2.5e-06
},
"other":{
"SDR": -1.359175,
"SAR": -14.7076775,
"SIR": -4.761505,
"ISR": -0.01528
}
} }
} }
@@ -102,15 +74,11 @@ def generate_fake_eval_dataset(path):
@pytest.mark.parametrize('backend', TEST_CONFIGURATIONS) @pytest.mark.parametrize('backend', TEST_CONFIGURATIONS)
def test_evaluate(backend): def test_evaluate(backend):
with TemporaryDirectory() as directory: with TemporaryDirectory() as directory:
generate_fake_eval_dataset(directory) generate_fake_eval_dataset(directory)
p = create_argument_parser() p = create_argument_parser()
arguments = p.parse_args(["evaluate", "-p", "spleeter:4stems", "--mus_dir", directory, "-B", backend]) arguments = p.parse_args(["evaluate", "-p", "spleeter:4stems", "--mus_dir", directory, "-B", backend])
params = load_configuration(arguments.configuration) params = load_configuration(arguments.configuration)
metrics = evaluate.entrypoint(arguments, params) metrics = evaluate.entrypoint(arguments, params)
for instrument, metric in metrics.items(): for instrument, metric in metrics.items():
for metric, value in metric.items(): for m, value in metric.items():
assert np.allclose(np.median(value), res_4stems[backend][instrument][metric], atol=1e-3) assert np.allclose(np.median(value), res_4stems[instrument][m], atol=1e-3)
# test_evaluate("tensorflow")

23
tests/test_separator.py
View File

@@ -39,6 +39,29 @@ TEST_CONFIGURATIONS = list(itertools.product(TEST_AUDIO_DESCRIPTORS, MODELS, BAC
print("RUNNING TESTS WITH TF VERSION {}".format(tf.__version__)) print("RUNNING TESTS WITH TF VERSION {}".format(tf.__version__))
@pytest.mark.parametrize('test_file', TEST_AUDIO_DESCRIPTORS)
def test_separator_backends(test_file):
adapter = get_default_audio_adapter()
waveform, _ = adapter.load(test_file)
separator_lib = Separator("spleeter:2stems", stft_backend="librosa")
separator_tf = Separator("spleeter:2stems", stft_backend="tensorflow")
# Test the stft and inverse stft provides exact reconstruction
stft_matrix = separator_lib._stft(waveform)
reconstructed = separator_lib._stft(
stft_matrix, inverse=True, length=waveform.shape[0])
assert np.allclose(reconstructed, waveform, atol=3e-2)
# compare both separation, it should be close
out_tf = separator_tf._separate_tensorflow(waveform, test_file)
out_lib = separator_lib._separate_librosa(waveform, test_file)
for instrument in out_lib.keys():
# test that both outputs are close everywhere
assert np.allclose(out_tf[instrument], out_lib[instrument], atol=1e-5)
@pytest.mark.parametrize('test_file, configuration, backend', TEST_CONFIGURATIONS) @pytest.mark.parametrize('test_file, configuration, backend', TEST_CONFIGURATIONS)
def test_separate(test_file, configuration, backend): def test_separate(test_file, configuration, backend):
""" Test separation from raw data. """ """ Test separation from raw data. """