def test_cuda_reset_distribution(self):
samples = np.array([1.0, 0.5, 0.25, 0.125, 0.01], dtype=np.float32)
samples_batch = np.vstack([samples] * 10000)
sample_rate = 16000
augment = Gain(min_gain_in_db=-6, max_gain_in_db=6, p=0.5).cuda()
# Change the parameters after init
augment.min_gain_in_db = -18
augment.max_gain_in_db = 3
processed_samples = (augment(
samples=torch.from_numpy(samples_batch).cuda(),
sample_rate=sample_rate).cpu().numpy())
self.assertEqual(processed_samples.dtype, np.float32)
actual_gains_in_db = []
for i in range(processed_samples.shape[0]):
if not np.allclose(processed_samples[i], samples_batch[i]):
estimated_gain_factor = np.mean(processed_samples[i] /
samples_batch[i])
estimated_gain_factor_in_db = convert_amplitude_ratio_to_decibels(
torch.tensor(estimated_gain_factor)).item()
self.assertGreaterEqual(estimated_gain_factor_in_db, -18)
self.assertLessEqual(estimated_gain_factor_in_db, 3)
actual_gains_in_db.append(estimated_gain_factor_in_db)
mean_gain_in_db = np.mean(actual_gains_in_db)
self.assertGreater(mean_gain_in_db, (-18 + 3) / 2 - 1)
self.assertLess(mean_gain_in_db, (-18 + 3) / 2 + 1)
def test_variability_within_batch_with_p_mode_per_batch(self):
samples = np.array([[1.0, 0.5, 0.25, 0.125, 0.01]], dtype=np.float32)
samples_batch = np.stack([samples] * 100, axis=0)
sample_rate = 16000
augment = Gain(
min_gain_in_db=-6,
max_gain_in_db=6,
p=0.5,
p_mode="per_batch",
output_type="dict",
)
num_processed_batches = 0
for _ in range(100):
processed_samples = augment(
samples=torch.from_numpy(samples_batch),
sample_rate=sample_rate).samples.numpy()
self.assertEqual(processed_samples.dtype, np.float32)
if np.allclose(processed_samples, samples_batch):
continue
else:
num_processed_batches += 1
num_unprocessed_examples = 0
num_processed_examples = 0
actual_gains_in_db = []
for i in range(processed_samples.shape[0]):
if np.allclose(processed_samples[i], samples_batch[i]):
num_unprocessed_examples += 1
else:
num_processed_examples += 1
estimated_gain_factor = np.mean(processed_samples[i] /
samples_batch[i])
estimated_gain_factor_in_db = convert_amplitude_ratio_to_decibels(
torch.tensor(estimated_gain_factor)).item()
self.assertGreaterEqual(estimated_gain_factor_in_db, -6)
self.assertLessEqual(estimated_gain_factor_in_db, 6)
actual_gains_in_db.append(estimated_gain_factor_in_db)
mean_gain_in_db = np.mean(actual_gains_in_db)
self.assertGreater(mean_gain_in_db, -1.5)
self.assertLess(mean_gain_in_db, 1.5)
# Should be 0 and 100, but I give some slack due to possible numerical issues
self.assertLessEqual(num_unprocessed_examples, 1)
self.assertGreaterEqual(num_processed_examples, 99)
self.assertGreater(num_processed_batches, 10)
self.assertLess(num_processed_batches, 90)
def test_variability_within_batch(self):
samples = np.array([[1.0, 0.5, 0.25, 0.125, 0.01]], dtype=np.float32)
samples_batch = np.stack([samples] * 10000, axis=0)
sample_rate = 16000
augment = Gain(min_gain_in_db=-6,
max_gain_in_db=6,
p=0.5,
output_type="dict")
processed_samples = augment(samples=torch.from_numpy(samples_batch),
sample_rate=sample_rate).samples.numpy()
self.assertEqual(processed_samples.dtype, np.float32)
num_unprocessed_examples = 0
num_processed_examples = 0
actual_gains_in_db = []
for i in range(processed_samples.shape[0]):
if np.allclose(processed_samples[i], samples_batch[i]):
num_unprocessed_examples += 1
else:
num_processed_examples += 1
estimated_gain_factor = np.mean(processed_samples[i] /
samples_batch[i])
estimated_gain_factor_in_db = convert_amplitude_ratio_to_decibels(
torch.tensor(estimated_gain_factor)).item()
self.assertGreaterEqual(estimated_gain_factor_in_db, -6)
self.assertLessEqual(estimated_gain_factor_in_db, 6)
actual_gains_in_db.append(estimated_gain_factor_in_db)
mean_gain_in_db = np.mean(actual_gains_in_db)
self.assertGreater(mean_gain_in_db, -1)
self.assertLess(mean_gain_in_db, 1)
self.assertEqual(num_unprocessed_examples + num_processed_examples,
10000)
self.assertGreater(num_processed_examples, 2000)
self.assertLess(num_processed_examples, 8000)
def test_gain_per_channel(self):
samples = np.array(
[[1.0, 0.5, 0.25, 0.125, 0.01], [0.95, 0.5, 0.25, 0.12, 0.011]],
dtype=np.float32,
)
samples_batch = np.stack([samples] * 1000, axis=0)
sample_rate = 16000
augment = Gain(min_gain_in_db=-30.0,
max_gain_in_db=0.0,
mode="per_channel",
p=0.5)
processed_samples = augment(samples=torch.from_numpy(samples_batch),
sample_rate=sample_rate).numpy()
num_unprocessed_channels = 0
num_processed_channels = 0
perturbation_type_counter = {
"zero_channels_changed": 0,
"only_the_first_channel_changed": 0,
"only_the_second_channel_changed": 0,
"both_channels_changed_differently": 0,
"both_channels_changed_equally": 0,
}
for i in range(processed_samples.shape[0]):
num_perturbed_channels = 0
est_gain_factors = []
for chn_index in range(processed_samples.shape[1]):
if np.allclose(processed_samples[i, chn_index],
samples_batch[i, chn_index]):
num_unprocessed_channels += 1
else:
num_processed_channels += 1
num_perturbed_channels += 1
estimated_gain_factors = (processed_samples[i, chn_index] /
samples_batch[i, chn_index])
self.assertAlmostEqual(
np.amin(estimated_gain_factors),
np.amax(estimated_gain_factors),
places=6,
)
estimated_gain_factor = np.mean(estimated_gain_factors)
estimated_gain_factor_in_db = convert_amplitude_ratio_to_decibels(
torch.tensor(estimated_gain_factor)).item()
est_gain_factors.append(estimated_gain_factor_in_db)
self.assertGreaterEqual(estimated_gain_factor_in_db, -30)
self.assertLessEqual(estimated_gain_factor_in_db, 0)
if num_perturbed_channels == 0:
perturbation_type_counter["zero_channels_changed"] += 1
elif num_perturbed_channels == 1:
if np.allclose(processed_samples[i, 0], samples_batch[i, 0]):
perturbation_type_counter[
"only_the_first_channel_changed"] += 1
else:
perturbation_type_counter[
"only_the_second_channel_changed"] += 1
elif num_perturbed_channels == 2:
if np.allclose(est_gain_factors[0], est_gain_factors[1]):
# This should be very unlikely
perturbation_type_counter[
"both_channels_changed_equally"] += 1
else:
perturbation_type_counter[
"both_channels_changed_differently"] += 1
self.assertGreater(perturbation_type_counter["zero_channels_changed"],
100)
self.assertLess(perturbation_type_counter["zero_channels_changed"],
500)
self.assertGreater(
perturbation_type_counter["only_the_first_channel_changed"], 100)
self.assertLess(
perturbation_type_counter["only_the_first_channel_changed"], 500)
self.assertGreater(
perturbation_type_counter["only_the_second_channel_changed"], 100)
self.assertLess(
perturbation_type_counter["only_the_second_channel_changed"], 500)
self.assertLess(
perturbation_type_counter["both_channels_changed_equally"], 10)
self.assertGreater(
perturbation_type_counter["both_channels_changed_differently"],
100)
self.assertLess(
perturbation_type_counter["both_channels_changed_differently"],
500)
self.assertEqual(num_unprocessed_channels + num_processed_channels,
1000 * 2)
self.assertGreater(num_processed_channels, 200 * 2)
self.assertLess(num_processed_channels, 800 * 2)
self.assertEqual(processed_samples.dtype, np.float32)
