def test_randomized_smoothing_schedule():
audio = torch.zeros(16, 16000, dtype=torch.float32)
tfnm = RandomizedSmoothing(sigma=[(0, 0.01), (100, 0.5)], p=0.8)
audio_aug = tfnm(audio)
# Shapes are the same
assert audio.shape == audio_aug.shape
# All samples are different than the input audio
assert (audio != audio_aug).any()
# Different batch samples receive different augmentation:
# we sum along the time axis and compare the summed values;
# if all examples got the same augmentation,
# there would have been just one unique value.
assert len(set(audio_aug.sum(dim=1).tolist())) > 1
tfnm.step = 1000
audio_aug2 = tfnm(audio)
# The schedule kicked in and the abs magnitudes should be larger.
assert audio_aug2.abs().sum() > audio_aug.abs().sum()
def test_randomized_smoothing_p1():
audio = torch.zeros(64, 4000, dtype=torch.float32)
tfnm = RandomizedSmoothing(sigma=0.1, p=1.0)
audio_aug = tfnm(audio)
# Shapes are the same
assert audio.shape == audio_aug.shape
# Some (most) samples are different than the input audio
assert (audio != audio_aug).any()
# Different batch samples receive different augmentation
assert (audio_aug[0] != audio_aug[1]).any()
def test_randomized_smoothing_p0():
audio = torch.zeros(64, 4000, dtype=torch.float32)
tfnm = RandomizedSmoothing(sigma=0.1, p=0.0)
audio_aug = tfnm(audio)
# Shapes are the same
assert audio.shape == audio_aug.shape
# Audio is unaffacted
assert (audio == audio_aug).all()
# Audio is unaffacted across batches
assert (audio_aug[0] == audio_aug[1]).all()
def test_randomized_smoothing(sample_sigma):
audio = torch.zeros(64, 4000, dtype=torch.float32)
tfnm = RandomizedSmoothing(sigma=0.1, sample_sigma=sample_sigma, p=0.8)
audio_aug = tfnm(audio)
# Shapes are the same
assert audio.shape == audio_aug.shape
# All samples are different than the input audio
assert (audio != audio_aug).any()
# Different batch samples receive different augmentation:
# we sum along the time axis and compare the summed values;
# if all examples got the same augmentation,
# there would have been just one unique value.
assert len(set(audio_aug.sum(dim=1).tolist())) > 1
def test_k2_speech_recognition_on_the_fly_feature_extraction_with_randomized_smoothing(
k2_cut_set, ):
dataset = K2SpeechRecognitionDataset(
input_strategy=OnTheFlyFeatures(extractor=Fbank(), ))
rs_dataset = K2SpeechRecognitionDataset(input_strategy=OnTheFlyFeatures(
extractor=Fbank(),
# Use p=1.0 to ensure that smoothing is applied in this test.
wave_transforms=[RandomizedSmoothing(sigma=0.5, p=1.0)],
))
sampler = SingleCutSampler(k2_cut_set, shuffle=False, max_cuts=1)
for cut_ids in sampler:
batch = dataset[cut_ids]
rs_batch = rs_dataset[cut_ids]
# Additive noise should cause the energies to go up
assert (rs_batch["inputs"] - batch["inputs"]).sum() > 0
