def test_silent_sources(X, V):
V[..., :] = 0.0
Y = norbert.softmask(V, X)
assert X.shape == Y.shape[:-1]
Y = norbert.wiener(V, X)
assert X.shape == Y.shape[:-1]
def test_softmask_copy(X, V):
X0 = X.clone()
V0 = V.clone()
_ = norbert.softmask(V, X)
assert torch.allclose(X0, X)
assert torch.allclose(V0, V)
def test_softmask_copy(X, V):
X0 = np.copy(X)
V0 = np.copy(V)
_ = norbert.softmask(V, X)
assert np.allclose(X0, X)
assert np.allclose(V0, V)
def test_shapes(X, V):
Y = norbert.wiener(V, X)
assert X.shape == Y.shape[:-1]
Y = norbert.softmask(V, X)
assert X.shape == Y.shape[:-1]
def test_shapes(V, X):
Y = norbert.residual(V, X)
assert X.shape == Y.shape[:-1]
Y = norbert.wiener(V, X)
assert X.shape == Y.shape[:-1]
Y = norbert.softmask(V, X)
assert X.shape == Y.shape[:-1]
def oracle(track):
# compute the mixture complex tf transform
x = stft(torch.from_numpy(track.audio.T)).transpose(0, 2)
v = []
for name, value in track.sources.items():
v_j = stft(torch.from_numpy(value.audio.T)).transpose(0, 2).abs()**2
v += [v_j]
v = torch.stack(v, 3)
y = norbert.softmask(v, x).permute(3, 2, 1, 0)
estimates = {}
for j, (name, value) in enumerate(track.sources.items()):
audio_hat = istft(y[j]).numpy().T
estimates[name] = audio_hat
# Evaluate using museval
scores = museval.eval_mus_track(track, estimates, output_dir=None)
print(scores)
return estimates
def test_softmask(V, X):
X = X.shape[-1] * np.ones(X.shape)
Y = norbert.softmask(V, X)
assert np.allclose(Y.sum(-1), X)
def test_softmask(V, X):
X = (X.shape[-1] * torch.ones(X.shape)).to(torch.complex128)
Y = norbert.softmask(V, X)
assert torch.allclose(Y.sum(-1), X)
Y.sum().backward()
def closure(v, x):
return norbert.softmask(torch.from_numpy(v),
torch.from_numpy(x)).numpy()
def test_softmask(V, X):
X = (X.shape[-1] * np.ones(X.shape)).astype(np.complex128)
Y = norbert.softmask(V, X)
assert np.allclose(Y.sum(-1), X)
