def test_sensitivity_additional_forward_args_multi_args(self) -> None:
model = BasicModel4_MultiArgs()
input1 = torch.tensor([[1.5, 2.0, 3.3]])
input2 = torch.tensor([[3.0, 3.5, 2.2]])
args = torch.tensor([[1.0, 3.0, 4.0]])
ig = DeepLift(model)
sensitivity1 = self.sensitivity_max_assert(
ig.attribute,
(input1, input2),
torch.zeros(1),
additional_forward_args=args,
n_perturb_samples=1,
max_examples_per_batch=1,
perturb_func=_perturb_func,
)
sensitivity2 = self.sensitivity_max_assert(
ig.attribute,
(input1, input2),
torch.zeros(1),
additional_forward_args=args,
n_perturb_samples=4,
max_examples_per_batch=2,
perturb_func=_perturb_func,
)
assertTensorAlmostEqual(self, sensitivity1, sensitivity2, 0.0)
def test_gradient_additional_args(self) -> None:
model = BasicModel4_MultiArgs()
input1 = torch.tensor([[10.0]], requires_grad=True)
input2 = torch.tensor([[8.0]], requires_grad=True)
grads = compute_gradients(model, (input1, input2),
additional_forward_args=(2, ))
assertArraysAlmostEqual(grads[0].squeeze(0).tolist(), [1.0],
delta=0.01)
assertArraysAlmostEqual(grads[1].squeeze(0).tolist(), [-0.5],
delta=0.01)
def _assert_multi_argument(
self, type: str, approximation_method: str = "gausslegendre"
) -> None:
model = BasicModel4_MultiArgs()
self._compute_attribution_and_evaluate(
model,
(
torch.tensor([[1.5, 2.0, 34.3]], requires_grad=True),
torch.tensor([[3.0, 3.5, 23.2]], requires_grad=True),
),
baselines=(0.0, torch.zeros((1, 3))),
additional_forward_args=torch.arange(1.0, 4.0).reshape(1, 3),
type=type,
approximation_method=approximation_method,
)
# uses batching with an integer variable and nd-tensors as
# additional forward arguments
self._compute_attribution_and_evaluate(
model,
(
torch.tensor([[1.5, 2.0, 34.3], [3.4, 1.2, 2.0]], requires_grad=True),
torch.tensor([[3.0, 3.5, 23.2], [2.3, 1.2, 0.3]], requires_grad=True),
),
baselines=(torch.zeros((2, 3)), 0.0),
additional_forward_args=(torch.arange(1.0, 7.0).reshape(2, 3), 1),
type=type,
approximation_method=approximation_method,
)
# uses batching with an integer variable and python list
# as additional forward arguments
model = BasicModel5_MultiArgs()
self._compute_attribution_and_evaluate(
model,
(
torch.tensor([[1.5, 2.0, 34.3], [3.4, 1.2, 2.0]], requires_grad=True),
torch.tensor([[3.0, 3.5, 23.2], [2.3, 1.2, 0.3]], requires_grad=True),
),
baselines=(0.0, 0.00001),
additional_forward_args=([2, 3], 1),
type=type,
approximation_method=approximation_method,
)
# similar to previous case plus baseline consists of a tensor and
# a single example
self._compute_attribution_and_evaluate(
model,
(
torch.tensor([[1.5, 2.0, 34.3], [3.4, 1.2, 2.0]], requires_grad=True),
torch.tensor([[3.0, 3.5, 23.2], [2.3, 1.2, 0.3]], requires_grad=True),
),
baselines=(torch.zeros((1, 3)), 0.00001),
additional_forward_args=([2, 3], 1),
type=type,
approximation_method=approximation_method,
)
def test_basic_infidelity_additional_forward_args1(self) -> None:
model = BasicModel4_MultiArgs()
input1 = torch.tensor([[1.5, 2.0, 3.3]])
input2 = torch.tensor([[3.0, 3.5, 2.2]])
args = torch.tensor([[1.0, 3.0, 4.0]])
ig = IntegratedGradients(model)
infidelity1 = self.basic_model_global_assert(
ig,
model,
(input1, input2),
torch.zeros(1),
additional_args=args,
n_perturb_samples=1,
max_batch_size=1,
perturb_func=_global_perturb_func1,
)
infidelity2 = self.basic_model_global_assert(
ig,
model,
(input1, input2),
torch.zeros(1),
additional_args=args,
n_perturb_samples=5,
max_batch_size=2,
perturb_func=_global_perturb_func1,
)
infidelity2_w_custom_pert_func = self.basic_model_global_assert(
ig,
model,
(input1, input2),
torch.zeros(1),
additional_args=args,
n_perturb_samples=5,
max_batch_size=2,
perturb_func=_global_perturb_func1_default,
)
assertTensorAlmostEqual(self, infidelity1, infidelity2, 0.0)
assertTensorAlmostEqual(self, infidelity2_w_custom_pert_func,
infidelity2, 0.0)