def test_run_sequential(bn_model, bn_criterion, bn_images, bn_labels):
attack = Attack(bn_model, bn_criterion)
advs1 = attack(bn_images, bn_labels)
advs2 = run_sequential(Attack, bn_model, bn_criterion, bn_images, bn_labels)
advs2 = np.stack([a.perturbed for a in advs2])
assert np.all(advs1 == advs2)
def test_run_sequential_invalid_arguments(bn_model, bn_criterion, bn_images,
bn_labels):
labels_wrong = bn_labels[[0]]
criteria_wrong = [bn_criterion] * (len(bn_images) + 1)
distances_wrong = [MSE] * (len(bn_images) + 1)
individual_kwargs_wrong = [{"max_epsilon": 1}] * (len(bn_images) + 1)
# test too few labels
with pytest.raises(AssertionError):
run_sequential(Attack, bn_model, bn_criterion, bn_images, labels_wrong)
# test too many criteria
with pytest.raises(AssertionError):
run_sequential(Attack, bn_model, criteria_wrong, bn_images, bn_labels)
# test too many distances
with pytest.raises(AssertionError):
run_sequential(
Attack,
bn_model,
bn_criterion,
bn_images,
bn_labels,
distance=distances_wrong,
)
# test too many kwargs
with pytest.raises(AssertionError):
run_sequential(
Attack,
bn_model,
bn_criterion,
bn_images,
bn_labels,
individual_kwargs=individual_kwargs_wrong,
)
def test_run_multiple_kwargs(bn_model, bn_criterion, bn_images, bn_labels):
attack = Attack(bn_model, bn_criterion)
advs1 = attack(
bn_images,
bn_labels,
unpack=False,
individual_kwargs=[{
"epsilons": 900
} for _ in bn_images],
)
advs1p = run_parallel(
Attack,
bn_model,
bn_criterion,
bn_images,
bn_labels,
individual_kwargs=[{
"epsilons": 900
} for _ in bn_images],
)
advs1s = run_sequential(
Attack,
bn_model,
bn_criterion,
bn_images,
bn_labels,
individual_kwargs=[{
"epsilons": 900
} for _ in bn_images],
)
advs2 = attack(
bn_images,
bn_labels,
unpack=False,
individual_kwargs=[{
"epsilons": 900
} for _ in bn_images],
max_epsilon=0.99,
)
advs2p = run_parallel(
Attack,
bn_model,
bn_criterion,
bn_images,
bn_labels,
individual_kwargs=[{
"epsilons": 900
} for _ in bn_images],
max_epsilon=0.99,
)
advs2s = run_sequential(
Attack,
bn_model,
bn_criterion,
bn_images,
bn_labels,
individual_kwargs=[{
"epsilons": 900
} for _ in bn_images],
max_epsilon=0.99,
)
advs3 = attack(bn_images, bn_labels, unpack=False, max_epsilon=0.99)
advs3p = run_parallel(Attack,
bn_model,
bn_criterion,
bn_images,
bn_labels,
max_epsilon=0.99)
advs3s = run_sequential(Attack,
bn_model,
bn_criterion,
bn_images,
bn_labels,
max_epsilon=0.99)
for i in range(len(bn_images)):
assert advs1[i].perturbed is not None
assert advs2[i].perturbed is not None
assert advs3[i].perturbed is not None
assert advs1p[i].perturbed is not None
assert advs2p[i].perturbed is not None
assert advs3p[i].perturbed is not None
assert advs1s[i].perturbed is not None
assert advs2s[i].perturbed is not None
assert advs3s[i].perturbed is not None
assert np.allclose(advs1[i].perturbed, advs1p[i].perturbed)
assert np.allclose(advs1[i].perturbed, advs1s[i].perturbed)
assert np.allclose(advs2[i].perturbed, advs2p[i].perturbed)
assert np.allclose(advs2[i].perturbed, advs2s[i].perturbed)
assert np.allclose(advs3[i].perturbed, advs3p[i].perturbed)
assert np.allclose(advs3[i].perturbed, advs3s[i].perturbed)