def test_improper_stride(self) -> None:
net = BasicModel_ConvNet_One_Conv()
inp = torch.arange(16, dtype=torch.float).view(1, 1, 4, 4)
occ = Occlusion(net)
# Check error when too few stride dimensions
with self.assertRaises(AssertionError):
_ = occ.attribute(inp,
sliding_window_shapes=(1, 2, 2),
strides=(1, 2),
target=0)
# Check error when too many stride dimensions
with self.assertRaises(AssertionError):
_ = occ.attribute(
(inp, inp),
sliding_window_shapes=((1, 1, 2), (1, 2, 2)),
strides=((1, 1, 2), (2, 1, 2, 2)),
target=0,
)
# Check error when too many stride tuples
with self.assertRaises(AssertionError):
_ = occ.attribute(
(inp, inp),
sliding_window_shapes=((1, 1, 2), (1, 2, 2)),
strides=((1, 1, 2), (1, 2, 2), (1, 2, 2)),
target=0,
)
def _occlusion_test_assert(
self,
model: Callable,
test_input: TensorOrTupleOfTensorsGeneric,
expected_ablation: Union[float, List[float], List[List[float]],
Tuple[Union[Tensor, List[float],
List[List[float]]], ...], ],
sliding_window_shapes: Union[Tuple[int, ...], Tuple[Tuple[int, ...],
...]],
target: TargetType = 0,
additional_input: Any = None,
perturbations_per_eval: Tuple[int, ...] = (1, ),
baselines: BaselineType = None,
strides: Union[None, int, Tuple[Union[int, Tuple[int, ...]],
...]] = None,
) -> None:
for batch_size in perturbations_per_eval:
ablation = Occlusion(model)
attributions = ablation.attribute(
test_input,
sliding_window_shapes=sliding_window_shapes,
target=target,
additional_forward_args=additional_input,
baselines=baselines,
perturbations_per_eval=batch_size,
strides=strides,
)
if isinstance(expected_ablation, tuple):
for i in range(len(expected_ablation)):
assertTensorAlmostEqual(self, attributions[i],
expected_ablation[i])
else:
assertTensorAlmostEqual(self, attributions, expected_ablation)
def _occlusion_test_assert(
self,
model,
test_input,
expected_ablation,
sliding_window_shapes=None,
target=0,
additional_input=None,
ablations_per_eval=(1, ),
baselines=None,
strides=None,
):
for batch_size in ablations_per_eval:
ablation = Occlusion(model)
attributions = ablation.attribute(
test_input,
sliding_window_shapes=sliding_window_shapes,
target=target,
additional_forward_args=additional_input,
baselines=baselines,
ablations_per_eval=batch_size,
strides=strides,
)
if isinstance(expected_ablation, tuple):
for i in range(len(expected_ablation)):
assertTensorAlmostEqual(self, attributions[i],
expected_ablation[i])
else:
assertTensorAlmostEqual(self, attributions, expected_ablation)
def test_too_large_stride(self) -> None:
net = BasicModel_ConvNet_One_Conv()
inp = torch.arange(16, dtype=torch.float).view(1, 1, 4, 4)
occ = Occlusion(net)
with self.assertRaises(AssertionError):
_ = occ.attribute(
inp, sliding_window_shapes=((1, 1, 2),), strides=2, target=0
)
with self.assertRaises(AssertionError):
_ = occ.attribute(
(inp, inp),
sliding_window_shapes=((1, 1, 2), (1, 4, 2)),
strides=(2, (1, 2, 3)),
target=0,
)
with self.assertRaises(AssertionError):
_ = occ.attribute(
inp, sliding_window_shapes=((2, 1, 2),), strides=2, target=0
)
def test_improper_window_shape(self):
net = BasicModel_ConvNet_One_Conv()
inp = torch.arange(16).view(1, 1, 4, 4).type(torch.FloatTensor)
occ = Occlusion(net)
# Check error when too few sliding window dimensions
with self.assertRaises(AssertionError):
_ = occ.attribute(inp, sliding_window_shapes=((1, 2), ), target=0)
# Check error when too many sliding window dimensions
with self.assertRaises(AssertionError):
_ = occ.attribute((inp, inp),
sliding_window_shapes=((1, 1, 2), (1, 1, 1, 2)),
target=0)
# Check error when too many sliding window tuples
with self.assertRaises(AssertionError):
_ = occ.attribute(
(inp, inp),
sliding_window_shapes=((1, 1, 2), (1, 1, 2), (1, 1, 2)),
target=0,
)