def test_two_equal_inputs_equal_noise(self):
size = (2, 2, 10000)
x = np.random.normal(0, 1, size)
noise = np.random.normal(0, 1, (2, 10000))
SDR, SIR, SNR = output_sxr(x, noise)
nptest.assert_almost_equal(SDR, -3., decimal=1)
nptest.assert_almost_equal(SIR, 0, decimal=1)
nptest.assert_almost_equal(SNR, 0, decimal=1)
def test_output_sxr(self):
image_contribution = morph(
'kKt->kKt',
np.asarray([[10 * self.s1, 1 * self.s2],
[0 * self.s1, 10 * self.s2]]))
noise_contribution = morph('Kt->Kt',
np.asarray([10 * self.n, 0 * self.n]))
sdr, sir, snr = output_sxr(
image_contribution,
noise_contribution,
average_sources=False,
)
assert all(snr == output_sxr(
image_contribution,
noise_contribution,
average_sources=False,
).snr)
np.testing.assert_allclose(sdr, [0, 20], atol=1e-6)
np.testing.assert_allclose(sir, [np.inf, 20], atol=1e-6)
np.testing.assert_allclose(snr, [0, np.inf], atol=1e-6)
def test_for_single_input(self):
expected_snr = np.array(10)
size = (1, 1, 10000)
x = np.random.normal(0, 1, size)
noise = 10**(-expected_snr / 20) * np.random.normal(0, 1, size[1:])
SDR, SIR, SNR = output_sxr(x, noise)
nptest.assert_almost_equal(SDR, expected_snr, decimal=1)
self.assertTrue(np.isinf(SIR))
nptest.assert_almost_equal(SNR, expected_snr, decimal=1)
def invasive_sxr(self):
from pb_bss.evaluation.sxr_module import output_sxr
invasive_sxr = output_sxr(
rearrange(self.speech_contribution,
'sources targets samples -> sources targets samples')
[:, self.selection, :],
rearrange(self.noise_contribution,
'targets samples -> targets samples')[self.selection, :],
average_sources=False,
return_dict=True,
)
return invasive_sxr
def test_output_sxr_more_outputs_than_sources(self):
sdr, sir, snr = output_sxr(
morph(
'kKt->kKt',
np.asarray([[10 * self.s1, 1 * self.s2, 0 * self.n],
[0 * self.s1, 10 * self.s2, 0 * self.n]])),
morph('Kt->Kt', np.asarray([10 * self.n, 0 * self.n, 0 * self.n])),
average_sources=False,
)
np.testing.assert_allclose(sdr, [0, 20], atol=1e-6)
np.testing.assert_allclose(sir, [np.inf, 20], atol=1e-6)
np.testing.assert_allclose(snr, [0, np.inf], atol=1e-6)
def test_output_sxr_more_outputs_than_sources_inf(self):
sdr, sir, snr = output_sxr(
morph(
'kKt->kKt',
np.asarray([[1 * self.s1, 0 * self.s2, 0 * self.n],
[0 * self.s1, 1 * self.s2, 0 * self.n]])),
morph('Kt->Kt', np.asarray([0 * self.n, 0 * self.n, 1 * self.n])),
average_sources=False,
)
np.testing.assert_allclose(sdr, 2 * [np.inf])
np.testing.assert_allclose(sir, 2 * [np.inf])
np.testing.assert_allclose(snr, 2 * [np.inf])
def test_two_equal_inputs_no_noise(self):
expected_sir = np.array(20)
size = (2, 2, 10000)
x = np.random.normal(0, 1, size)
x[0, 1, :] = 10**(-expected_sir / 20) * x[0, 0, :]
x[1, 1, :] = x[0, 1, :]
x[1, 0, :] = 10**(-expected_sir / 20) * x[0, 1, :]
noise = np.zeros((2, 10000))
SDR, SIR, SNR = output_sxr(x, noise)
nptest.assert_almost_equal(SDR, 20., decimal=1)
nptest.assert_almost_equal(SIR, 20., decimal=1)
self.assertTrue(np.isinf(SNR).all())