def test_supported_modes_property(self):
augment = Compose(transforms=[
PeakNormalization(p=1.0),
], )
assert augment.supported_modes == {
"per_batch", "per_example", "per_channel"
}
augment = Compose(
transforms=[PeakNormalization(p=1.0),
ShuffleChannels(p=1.0)], )
assert augment.supported_modes == {"per_example"}
def test_shuffle(self):
random.seed(42)
samples = np.array([[[1.0, 0.5, -0.25, -0.125, 0.0]]],
dtype=np.float32)
sample_rate = 16000
augment = Compose(
transforms=[
Gain(min_gain_in_db=-18.0, max_gain_in_db=-16.0, p=1.0),
PeakNormalization(p=1.0),
],
shuffle=True,
output_type="dict",
)
num_peak_normalization_last = 0
num_gain_last = 0
for i in range(100):
processed_samples = augment(
samples=torch.from_numpy(samples),
sample_rate=sample_rate).samples.numpy()
# Either PeakNormalization or Gain was applied last
if processed_samples[0, 0, 0] < 0.2:
num_gain_last += 1
elif processed_samples[0, 0, 0] == 1.0:
num_peak_normalization_last += 1
else:
raise AssertionError("Unexpected value!")
self.assertGreater(num_peak_normalization_last, 10)
self.assertGreater(num_gain_last, 10)
def setUp(self):
self.sample_rate = 16000
self.audio = torch.randn(1, 1, 16000)
self.transforms = [
Gain(min_gain_in_db=-6.000001, max_gain_in_db=-2, p=1.0),
PolarityInversion(p=1.0),
PeakNormalization(p=1.0),
]
Shift(p=1.0, output_type="dict"),
# Non-differentiable transforms:
# RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation:
# [torch.DoubleTensor [1, 1, 5]], which is output 0 of IndexBackward, is at version 1; expected version 0 instead.
# Hint: enable anomaly detection to find the operation that failed to compute its gradient,
# with torch.autograd.set_detect_anomaly(True).
pytest.param(
HighPassFilter(p=1.0, output_type="dict"),
marks=pytest.mark.skip("Not differentiable"),
),
pytest.param(
LowPassFilter(p=1.0, output_type="dict"),
marks=pytest.mark.skip("Not differentiable"),
),
pytest.param(
PeakNormalization(p=1.0, output_type="dict"),
marks=pytest.mark.skip("Not differentiable"),
),
],
)
def test_transform_is_differentiable(augment):
sample_rate = 16000
# Note: using float64 dtype to be compatible with AddBackgroundNoise fixtures
samples = torch.tensor([[1.0, 0.5, -0.25, -0.125, 0.0]],
dtype=torch.float64).unsqueeze(1)
samples_cpy = deepcopy(samples)
# We are going to convert the input tensor to a nn.Parameter so that we can
# track the gradients with respect to it. We'll "optimize" the input signal
# to be closer to that after the augmentation to test differentiability
# of the transform. If the signal got changed in any way, and the test
Gain(min_gain_in_db=-15.0, max_gain_in_db=5.0, p=1.0),
PolarityInversion(p=1.0),
]),
Gain(min_gain_in_db=-6.000001, max_gain_in_db=-6, p=1.0),
PolarityInversion(p=1.0),
Shift(p=1.0),
# Non-differentiable transforms:
# RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation:
# [torch.DoubleTensor [1, 1, 5]], which is output 0 of IndexBackward, is at version 1; expected version 0 instead.
# Hint: enable anomaly detection to find the operation that failed to compute its gradient,
# with torch.autograd.set_detect_anomaly(True).
pytest.param(HighPassFilter(p=1.0),
marks=pytest.mark.skip("Not differentiable")),
pytest.param(LowPassFilter(p=1.0),
marks=pytest.mark.skip("Not differentiable")),
pytest.param(PeakNormalization(p=1.0),
marks=pytest.mark.skip("Not differentiable")),
],
)
def test_transform_is_differentiable(augment):
sample_rate = 16000
# Note: using float64 dtype to be compatible with AddBackgroundNoise fixtures
samples = torch.tensor([[1.0, 0.5, -0.25, -0.125, 0.0]],
dtype=torch.float64).unsqueeze(1)
samples_cpy = deepcopy(samples)
# We are going to convert the input tensor to a nn.Parameter so that we can
# track the gradients with respect to it. We'll "optimize" the input signal
# to be closer to that after the augmentation to test differentiability
# of the transform. If the signal got changed in any way, and the test
# didn't crash, it means it works.
),
"num_runs": 5,
},
{
"instance": ApplyImpulseResponse(
ir_paths=TEST_FIXTURES_DIR / "ir", mode=mode, p=1.0
),
"num_runs": 1,
},
{
"instance": Compose(
transforms=[
Gain(
min_gain_in_db=-18.0, max_gain_in_db=-16.0, mode=mode, p=1.0
),
PeakNormalization(mode=mode, p=1.0),
],
shuffle=True,
),
"name": "Shuffled Compose with Gain and PeakNormalization",
"num_runs": 5,
},
{
"instance": Compose(
transforms=[
Gain(
min_gain_in_db=-18.0, max_gain_in_db=-16.0, mode=mode, p=0.5
),
PolarityInversion(mode=mode, p=0.5),
],
shuffle=True,
sample_rates=[16000],
devices=["cpu", "cuda"],
)
if not torch.cuda.is_available():
params["devices"].remove("cuda")
devices = {
device_name: torch.device(device_name)
for device_name in params["devices"]
}
transforms = [
Gain(p=1.0),
PolarityInversion(p=1.0),
PeakNormalization(p=1.0),
Shift(p=1.0),
]
perf_objects = []
for device_name in params["devices"]:
device = devices[device_name]
for batch_size in tqdm(params["batch_sizes"]):
for num_channels in params["channels"]:
for duration in params["durations"]:
for sample_rate in params["sample_rates"]:
for transform in transforms:
perf_objects += measure_execution_time(
transform,
batch_size,
mono=False)
samples = np.stack((samples1, samples2), axis=0)
samples = torch.from_numpy(samples)
modes = ["per_batch", "per_example", "per_channel"]
for mode in modes:
transforms = [
{
"instance":
Compose(
transforms=[
Gain(min_gain_in_db=-18.0,
max_gain_in_db=-16.0,
mode=mode,
p=1.0),
PeakNormalization(mode=mode, p=1.0),
],
shuffle=True,
),
"name":
"Shuffled Compose with Gain and PeakNormalization",
"num_runs":
5,
},
{
"instance":
Compose(
transforms=[
Gain(min_gain_in_db=-18.0,
max_gain_in_db=-16.0,
mode=mode,
),
"num_runs": 5,
},
{
"get_instance": lambda: ApplyImpulseResponse(
ir_paths=TEST_FIXTURES_DIR / "ir", mode=mode, p=1.0
),
"num_runs": 1,
},
{
"get_instance": lambda: Compose(
transforms=[
Gain(
min_gain_in_db=-18.0, max_gain_in_db=-16.0, mode=mode, p=1.0
),
PeakNormalization(mode=mode, p=1.0),
],
shuffle=True,
),
"name": "Shuffled Compose with Gain and PeakNormalization",
"num_runs": 5,
},
{
"get_instance": lambda: Compose(
transforms=[
Gain(
min_gain_in_db=-18.0, max_gain_in_db=-16.0, mode=mode, p=0.5
),
PolarityInversion(mode=mode, p=0.5),
],
shuffle=True,
"get_instance": lambda: BandPassFilter(mode=mode, p=1.0),
"num_runs": 5
},
{
"get_instance": lambda: BandStopFilter(mode=mode, p=1.0),
"num_runs": 5
},
{
"get_instance":
lambda: Compose(
transforms=[
Gain(min_gain_in_db=-18.0,
max_gain_in_db=-16.0,
mode=mode,
p=1.0),
PeakNormalization(mode=mode, p=1.0),
],
shuffle=True,
),
"name":
"Shuffled Compose with Gain and PeakNormalization",
"num_runs":
5,
},
{
"get_instance":
lambda: Compose(
transforms=[
Gain(min_gain_in_db=-18.0,
max_gain_in_db=-16.0,
mode=mode,