def construct_transforms(dataset_args, transforms_args):
transforms_dict = {
"merge_two_signals":
TransformSignal(merge_two_signals, transforms_args),
"identity":
TransformSignal(identity, transforms_args),
"identity_ml":
TransformSignal(identity_ml, transforms_args),
"delay_signal":
TransformSignal(delay_signal, transforms_args),
"add_noise_to_signal":
TransformSignal(add_noise_to_signal, transforms_args),
"add_noise_to_signal_only_one_signal":
TransformSignal(add_noise_to_signal_only_one_signal, transforms_args),
"add_noise_to_signal_with_proba":
TransformSignal(add_noise_to_signal_with_proba, transforms_args),
"merge_two_emd":
TransformSignal(merge_two_emd, transforms_args),
"randaugment":
TransformSignal(construct_randaugment, transforms_args),
"mask_along_time":
TransformFFT(mask_along_time, transforms_args),
"mask_along_frequency":
TransformFFT(mask_along_frequency, transforms_args),
}
transform_list = []
for transform_name in dataset_args["transform_list"]:
transform = []
for operation in transform_name:
transform.append(transforms_dict[operation])
transform_list.append(transform)
return transform_list
def test_transform_construction():
transforms = {
"transform_list": [["identity"], ["mask_along_time", "identity"]]
}
constructed_transforms = [[TransformSignal(identity, transforms_args)],
[
TransformFFT(mask_along_time,
transforms_args),
TransformSignal(identity, transforms_args)
]]
compare(construct_transforms(transforms, transforms_args),
constructed_transforms)
def test_data_recovery():
train_sample, _, _ = get_dummy_sample()
X = train_sample[0][0]
datum = Datum(X=X, y=train_sample[0][1])
datum = TransformSignal(identity)(TransformFFT(identity)(datum))
assert_almost_equal(datum.X.numpy(), X, decimal=3)
def test_delay_signal():
train_sample, _, _ = get_dummy_sample()
transf = TransformSignal(delay_signal, transforms_args)
X = train_sample[0][0]
datum = Datum(X=X, y=train_sample[0][1])
datum = transf(datum)
value_cutoff = int(np.round(transforms_args["magnitude"] * X.shape[1]))
assert_array_equal(X[:, -value_cutoff:], datum.X[:, :value_cutoff])
assert_array_equal(X[:, :-value_cutoff], datum.X[:, value_cutoff:])
def test_noise_addition():
transforms_args["magnitude"] = 0.1
train_sample, _, _ = get_dummy_sample()
transf = TransformSignal(add_noise_to_signal, transforms_args)
X = train_sample[0][0]
scale = torch.mean(torch.abs(X)).item()
datum = Datum(X=X, y=train_sample[0][1])
datum = transf(datum)
remains = (datum.X - X) / (transforms_args["magnitude"] * 0.1)
assert_almost_equal(np.mean(remains.numpy()) / scale, 0, 1)
assert_almost_equal(np.var(remains.numpy()) / (scale * scale), 1, 1)
def test_randaugment():
train_sample, _, _ = get_dummy_sample()
subset_aug_sample, subset_aug_labels = get_sample(train_sample,
[["randaugment"]],
sample_size=1,
random_state=1)
transforms_args["train_sample"] = train_sample
transforms_args["label_index_dict"] = create_label_index_dict(
subset_aug_sample, subset_aug_labels)
transf = TransformSignal(randaugment, transforms_args)
X = train_sample[0][0]
datum = Datum(X=X, y=train_sample[0][1])
datum = transf(datum)
def test_signal_addition():
train_sample, _, _ = get_dummy_sample()
subset_aug_sample, subset_aug_labels = get_sample(train_sample,
[["merge_two_signals"]],
sample_size=1,
random_state=1)
transforms_args["train_sample"] = train_sample
transforms_args["label_index_dict"] = create_label_index_dict(
subset_aug_sample, subset_aug_labels)
transf = TransformSignal(merge_two_signals, transforms_args)
X = train_sample[0][0]
datum = Datum(X=X, y=train_sample[0][1])
datum = transf(datum)
assert_array_almost_equal(X, datum.X, 5)
def test_add_noise_to_signal_with_proba():
transforms_args["magnitude"] = 0.1
train_sample, _, _ = get_dummy_sample()
transf = TransformSignal(add_noise_to_signal_with_proba, transforms_args)
random.seed(1) # returns 0.13
X = train_sample[0][0]
datum = Datum(X=X, y=train_sample[0][1])
datum = transf(datum)
assert_almost_equal(datum.X.numpy(), X)
random.seed(0) # return 0.88
datum = transf(datum)
scale = torch.mean(torch.abs(X)).item()
remains = (datum.X - X) / (transforms_args["magnitude"] * 0.1)
assert_almost_equal(np.mean(remains.numpy()) / scale, 0, 1)
assert_almost_equal(np.var(remains.numpy()) / (scale * scale), 1, 1)
def test_add_noise_to_signal_only_one_signal():
transforms_args["magnitude"] = 0.1
train_sample, _, _ = get_dummy_sample()
transf = TransformSignal(add_noise_to_signal_only_one_signal,
transforms_args)
random.seed(1) # returns 0.1
X = train_sample[0][0]
datum = Datum(X=X, y=train_sample[0][1])
datum = transf(datum)
assert_almost_equal(datum.X[1, :].numpy(), X[1, :])
random.seed(0) # returns 0.8
X = train_sample[0][0]
datum = Datum(X=X, y=train_sample[0][1])
datum = transf(datum)
assert_almost_equal(datum.X[0, :].numpy(), X[0, :])
def create_transforms_and_subset(train_dataset, dataset_args, sample_size,
transforms_args, i):
train_subset = train_dataset
train_dataset.change_transform_list([[TransformSignal(identity)]])
# then, find what will be the labels of augmented data, without constructing transforms
subset_aug_sample, subset_aug_labels = get_sample(
train_dataset,
dataset_args["transform_list"],
sample_size,
random_state=i)
# Define everything needed to construct transforms, even if "train_subset" will be replaced
# afterwards.
transforms_args["train_sample"] = train_dataset
transforms_args["label_index_dict"] = create_label_index_dict(
subset_aug_sample, subset_aug_labels)
# Constructs transforms
dataset_args["constructed_transform_list"] = construct_transforms(
dataset_args, transforms_args)
# Update train dataset
train_dataset.change_transform_list(
dataset_args["constructed_transform_list"])
train_subset = Subset(dataset=train_dataset, indices=subset_aug_sample)
return dataset_args, train_subset