def test_apply_frequency_mask_stereo(self):
np.random.seed(42)
sample_len = 1024
samples_in = np.random.uniform(low=-0.5, high=0.5, size=(2, sample_len)).astype(
np.float32
)
samples_in_copy = np.copy(samples_in)
sample_rate = 16000
augmenter = FrequencyMask(min_frequency_band=0.3, max_frequency_band=0.3, p=1.0)
samples_out = augmenter(samples=samples_in, sample_rate=sample_rate)
self.assertEqual(samples_out.dtype, np.float32)
self.assertEqual(samples_out.shape, samples_in.shape)
# Check that the input is left untouched
assert_array_equal(samples_in, samples_in_copy)
std_in = np.mean(np.abs(samples_in))
std_out = np.mean(np.abs(samples_out))
self.assertLess(std_out, std_in)
augmenter.freeze_parameters()
samples_only_1st_channel = augmenter(
samples=samples_in[0, :], sample_rate=sample_rate
)
assert_array_equal(samples_out[0, :], samples_only_1st_channel)
def test_dynamic_length(self):
sample_len = 1024
samples_in = np.random.normal(0, 1, size=sample_len).astype(np.float32)
sample_rate = 16000
augmenter = Compose(
[FrequencyMask(min_frequency_band=0.3, max_frequency_band=0.5, p=1.0)]
)
samples_out = augmenter(samples=samples_in, sample_rate=sample_rate)
self.assertEqual(samples_out.dtype, np.float32)
self.assertEqual(len(samples_out), sample_len)
std_in = np.mean(np.abs(samples_in))
std_out = np.mean(np.abs(samples_out))
self.assertLess(std_out, std_in)
def test_apply_frequency_mask(self):
sample_len = 1024
samples_in = np.random.normal(0, 1, size=sample_len).astype(np.float32)
samples_in_copy = np.copy(samples_in)
sample_rate = 16000
augmenter = Compose(
[FrequencyMask(min_frequency_band=0.3, max_frequency_band=0.5, p=1.0)]
)
samples_out = augmenter(samples=samples_in, sample_rate=sample_rate)
self.assertEqual(samples_out.dtype, np.float32)
self.assertEqual(len(samples_out), sample_len)
# Check that the input is left untouched
assert_array_equal(samples_in, samples_in_copy)
std_in = np.mean(np.abs(samples_in))
std_out = np.mean(np.abs(samples_out))
self.assertLess(std_out, std_in)
def test_filter_instability(self):
"""
An early implementation of FrequencyMask had a problem with filter instability
sometimes. That would lead to NaN values in the result. This test checks whether or not
the problem currently exists.
"""
np.random.seed(42)
sample_len = 32000
samples_in = np.random.uniform(-1, 1, size=sample_len).astype(np.float32)
sample_rate = 16000
augmenter = Compose([FrequencyMask(p=1.0)])
augmenter.transforms[0].randomize_parameters(samples_in, sample_rate)
augmenter.transforms[0].parameters["bandwidth"] = 600
augmenter.transforms[0].parameters["freq_start"] = 17
augmenter.freeze_parameters()
samples_out = augmenter(samples=samples_in, sample_rate=sample_rate)
self.assertFalse(np.isnan(samples_out).any())
def test_randomize_parameters_and_apply(self):
samples = 1.0 / np.arange(1, 21, dtype=np.float32)
sample_rate = 44100
augmenter = Compose([
AddBackgroundNoise(
sounds_path=os.path.join(DEMO_DIR, "background_noises"),
min_snr_in_db=15,
max_snr_in_db=35,
p=1.0,
),
ClippingDistortion(p=0.5),
FrequencyMask(min_frequency_band=0.3,
max_frequency_band=0.5,
p=0.5),
TimeMask(min_band_part=0.2, max_band_part=0.5, p=0.5),
Shift(min_fraction=0.5, max_fraction=0.5, p=0.5),
])
augmenter.freeze_parameters()
augmenter.randomize_parameters(samples=samples,
sample_rate=sample_rate)
parameters = [
transform.parameters for transform in augmenter.transforms
]
perturbed_samples1 = augmenter(samples=samples,
sample_rate=sample_rate)
perturbed_samples2 = augmenter(samples=samples,
sample_rate=sample_rate)
assert_array_equal(perturbed_samples1, perturbed_samples2)
augmenter.unfreeze_parameters()
for transform_parameters, transform in zip(parameters,
augmenter.transforms):
self.assertTrue(transform_parameters == transform.parameters)
self.assertFalse(transform.are_parameters_frozen)