def test_simple_ablation_int_to_int_nt(self) -> None:
ablation_algo = NoiseTunnel(FeatureAblation(BasicModel()))
inp = torch.tensor([[-3, 1, 2]]).float()
self._ablation_test_assert(
ablation_algo,
inp,
[[-3.0, 0.0, 0.0]],
perturbations_per_eval=(1, 2, 3),
stdevs=1e-10,
)
def _input_x_gradient_classification_assert(self, nt_type: str = "vanilla") -> None:
num_in = 5
input = torch.tensor([[0.0, 1.0, 2.0, 3.0, 4.0]], requires_grad=True)
target = torch.tensor(5)
# 10-class classification model
model = SoftmaxModel(num_in, 20, 10)
input_x_grad = InputXGradient(model.forward)
if nt_type == "vanilla":
attributions = input_x_grad.attribute(input, target)
output = model(input)[:, target]
output.backward()
expected = input.grad * input
assertTensorAlmostEqual(self, attributions, expected, 0.00001, "max")
else:
nt = NoiseTunnel(input_x_grad)
attributions = nt.attribute(
input, nt_type=nt_type, nt_samples=10, stdevs=1.0, target=target
)
self.assertEqual(attributions.shape, input.shape)
def test_batched_input_smoothgrad_wo_mutliplying_by_inputs(self) -> None:
model = BasicModel_MultiLayer()
inputs = torch.tensor(
[[1.5, 2.0, 1.3], [0.5, 0.1, 2.3], [1.5, 2.0, 1.3]], requires_grad=True
)
ig_wo_mutliplying_by_inputs = IntegratedGradients(
model, multiply_by_inputs=False
)
nt_wo_mutliplying_by_inputs = NoiseTunnel(ig_wo_mutliplying_by_inputs)
ig = IntegratedGradients(model)
nt = NoiseTunnel(ig)
n_samples = 5
target = 0
type = "smoothgrad"
attributions_wo_mutliplying_by_inputs = nt_wo_mutliplying_by_inputs.attribute(
inputs,
nt_type=type,
nt_samples=n_samples,
stdevs=0.0,
target=target,
n_steps=500,
)
with self.assertWarns(DeprecationWarning):
attributions = nt.attribute(
inputs,
nt_type=type,
n_samples=n_samples,
stdevs=0.0,
target=target,
n_steps=500,
)
assertTensorAlmostEqual(
self, attributions_wo_mutliplying_by_inputs * inputs, attributions
)
def _input_x_gradient_base_assert(
self,
model: Module,
inputs: TensorOrTupleOfTensorsGeneric,
expected_grads: TensorOrTupleOfTensorsGeneric,
additional_forward_args: Any = None,
nt_type: str = "vanilla",
) -> None:
input_x_grad = InputXGradient(model)
attributions: TensorOrTupleOfTensorsGeneric
if nt_type == "vanilla":
attributions = input_x_grad.attribute(
inputs, additional_forward_args=additional_forward_args)
else:
nt = NoiseTunnel(input_x_grad)
attributions = nt.attribute(
inputs,
nt_type=nt_type,
n_samples=10,
stdevs=0.0002,
additional_forward_args=additional_forward_args,
)
if isinstance(attributions, tuple):
for input, attribution, expected_grad in zip(
inputs, attributions, expected_grads):
if nt_type == "vanilla":
assertArraysAlmostEqual(attribution.reshape(-1),
(expected_grad *
input).reshape(-1))
self.assertEqual(input.shape, attribution.shape)
elif isinstance(attributions, Tensor):
if nt_type == "vanilla":
assertArraysAlmostEqual(
attributions.reshape(-1),
(expected_grads * inputs).reshape(-1),
delta=0.5,
)
self.assertEqual(inputs.shape, attributions.shape)
def _input_x_gradient_base_assert(
self,
model: Module,
inputs: TensorOrTupleOfTensorsGeneric,
expected_grads: TensorOrTupleOfTensorsGeneric,
additional_forward_args: Any = None,
nt_type: str = "vanilla",
) -> None:
input_x_grad = InputXGradient(model)
self.assertTrue(input_x_grad.multiplies_by_inputs)
attributions: TensorOrTupleOfTensorsGeneric
if nt_type == "vanilla":
attributions = input_x_grad.attribute(
inputs,
additional_forward_args=additional_forward_args,
)
else:
nt = NoiseTunnel(input_x_grad)
attributions = nt.attribute(
inputs,
nt_type=nt_type,
nt_samples=10,
stdevs=0.0002,
additional_forward_args=additional_forward_args,
)
if isinstance(attributions, tuple):
for input, attribution, expected_grad in zip(
inputs, attributions, expected_grads
):
if nt_type == "vanilla":
self._assert_attribution(expected_grad, input, attribution)
self.assertEqual(input.shape, attribution.shape)
elif isinstance(attributions, Tensor):
if nt_type == "vanilla":
self._assert_attribution(expected_grads, inputs, attributions)
self.assertEqual(
cast(Tensor, inputs).shape, cast(Tensor, attributions).shape
)
def _saliency_classification_assert(self, nt_type="vanilla"):
num_in = 5
input = torch.tensor([[0.0, 1.0, 2.0, 3.0, 4.0]], requires_grad=True)
target = torch.tensor(5)
# 10-class classification model
model = SoftmaxModel(num_in, 20, 10)
saliency = Saliency(model)
if nt_type == "vanilla":
attributions = saliency.attribute(input, target)
output = model(input)[:, target]
output.backward()
expected = torch.abs(input.grad)
self.assertEqual(
expected.detach().numpy().tolist(),
attributions.detach().numpy().tolist(),
)
else:
nt = NoiseTunnel(saliency)
attributions = nt.attribute(
input, nt_type=nt_type, n_samples=10, stdevs=0.0002, target=target
)
self.assertEqual(input.shape, attributions.shape)
def test_multi_input_ablation_with_mask_nt(self) -> None:
ablation_algo = NoiseTunnel(
FeatureAblation(BasicModel_MultiLayer_MultiInput()))
inp1 = torch.tensor([[23.0, 100.0, 0.0], [20.0, 50.0, 30.0]])
inp2 = torch.tensor([[20.0, 50.0, 30.0], [0.0, 100.0, 0.0]])
inp3 = torch.tensor([[0.0, 100.0, 10.0], [2.0, 10.0, 3.0]])
mask1 = torch.tensor([[1, 1, 1], [0, 1, 0]])
mask2 = torch.tensor([[0, 1, 2]])
mask3 = torch.tensor([[0, 1, 2], [0, 0, 0]])
expected = (
[[492.0, 492.0, 492.0], [200.0, 200.0, 200.0]],
[[80.0, 200.0, 120.0], [0.0, 400.0, 0.0]],
[[0.0, 400.0, 40.0], [60.0, 60.0, 60.0]],
)
self._ablation_test_assert(
ablation_algo,
(inp1, inp2, inp3),
expected,
additional_input=(1, ),
feature_mask=(mask1, mask2, mask3),
stdevs=1e-10,
)
self._ablation_test_assert(
ablation_algo,
(inp1, inp2),
expected[0:1],
additional_input=(inp3, 1),
feature_mask=(mask1, mask2),
perturbations_per_eval=(1, 2, 3),
stdevs=1e-10,
)
expected_with_baseline = (
[[468.0, 468.0, 468.0], [184.0, 192.0, 184.0]],
[[68.0, 188.0, 108.0], [-12.0, 388.0, -12.0]],
[[-16.0, 384.0, 24.0], [12.0, 12.0, 12.0]],
)
self._ablation_test_assert(
ablation_algo,
(inp1, inp2, inp3),
expected_with_baseline,
additional_input=(1, ),
feature_mask=(mask1, mask2, mask3),
baselines=(2, 3.0, 4),
perturbations_per_eval=(1, 2, 3),
stdevs=1e-10,
)
def hook_removal_test_assert(self) -> None:
attr_method: Attribution
expect_error = False
if layer is not None:
if mode is HookRemovalMode.invalid_module:
expect_error = True
if isinstance(layer, list):
_set_deep_layer_value(model, layer[0], ErrorModule())
else:
_set_deep_layer_value(model, layer, ErrorModule())
target_layer = get_target_layer(model, layer)
internal_algorithm = cast(Type[InternalAttribution], algorithm)
attr_method = internal_algorithm(model, target_layer)
else:
attr_method = algorithm(model)
if noise_tunnel:
attr_method = NoiseTunnel(attr_method)
if mode is HookRemovalMode.incorrect_target_or_neuron:
# Overwriting target and neuron index arguments to
# incorrect values.
if "target" in args:
args["target"] = (9999, ) * 20
expect_error = True
if "neuron_selector" in args:
args["neuron_selector"] = (9999, ) * 20
expect_error = True
if expect_error:
with self.assertRaises(AssertionError):
attr_method.attribute(**args)
else:
attr_method.attribute(**args)
def check_leftover_hooks(module):
self.assertEqual(len(module._forward_hooks), 0)
self.assertEqual(len(module._backward_hooks), 0)
self.assertEqual(len(module._forward_pre_hooks), 0)
model.apply(check_leftover_hooks)
def data_parallel_test_assert(self) -> None:
# Construct cuda_args, moving all tensor inputs in args to CUDA device
cuda_args = {}
for key in args:
if isinstance(args[key], Tensor):
cuda_args[key] = args[key].cuda()
elif isinstance(args[key], tuple):
cuda_args[key] = tuple(
elem.cuda() if isinstance(elem, Tensor) else elem
for elem in args[key])
else:
cuda_args[key] = args[key]
alt_device_ids = None
cuda_model = copy.deepcopy(model).cuda()
# Initialize models based on DataParallelCompareMode
if mode is DataParallelCompareMode.cpu_cuda:
model_1, model_2 = model, cuda_model
args_1, args_2 = args, cuda_args
elif mode is DataParallelCompareMode.data_parallel_default:
model_1, model_2 = (
cuda_model,
torch.nn.parallel.DataParallel(cuda_model),
)
args_1, args_2 = cuda_args, cuda_args
elif mode is DataParallelCompareMode.data_parallel_alt_dev_ids:
alt_device_ids = [0] + [
x for x in range(torch.cuda.device_count() - 1, 0, -1)
]
model_1, model_2 = (
cuda_model,
torch.nn.parallel.DataParallel(cuda_model,
device_ids=alt_device_ids),
)
args_1, args_2 = cuda_args, cuda_args
else:
raise AssertionError(
"DataParallel compare mode type is not valid.")
attr_method_1: Attribution
attr_method_2: Attribution
if target_layer:
internal_algorithm = cast(Type[InternalAttribution], algorithm)
attr_method_1 = internal_algorithm(
model_1, _get_deep_layer_name(model_1, target_layer))
# cuda_model is used to obtain target_layer since DataParallel
# adds additional wrapper.
# model_2 is always either the CUDA model itself or DataParallel
if alt_device_ids is None:
attr_method_2 = internal_algorithm(
model_2, _get_deep_layer_name(cuda_model,
target_layer))
else:
# LayerDeepLift and LayerDeepLiftShap do not take device ids
# as a parameter, since they must always have the DataParallel
# model object directly.
# Some neuron methods and GuidedGradCAM also require the
# model and cannot take a forward function.
if issubclass(
internal_algorithm,
(
LayerDeepLift,
LayerDeepLiftShap,
NeuronDeepLift,
NeuronDeepLiftShap,
NeuronDeconvolution,
NeuronGuidedBackprop,
GuidedGradCam,
),
):
attr_method_2 = internal_algorithm(
model_2,
_get_deep_layer_name(cuda_model, target_layer))
else:
attr_method_2 = internal_algorithm(
model_2.forward,
_get_deep_layer_name(cuda_model, target_layer),
device_ids=alt_device_ids,
)
else:
attr_method_1 = algorithm(model_1)
attr_method_2 = algorithm(model_2)
if noise_tunnel:
attr_method_1 = NoiseTunnel(attr_method_1)
attr_method_2 = NoiseTunnel(attr_method_2)
if attr_method_1.has_convergence_delta():
attributions_1, delta_1 = attr_method_1.attribute(
return_convergence_delta=True, **args_1)
self.setUp()
attributions_2, delta_2 = attr_method_2.attribute(
return_convergence_delta=True, **args_2)
assertTensorTuplesAlmostEqual(self,
attributions_1,
attributions_2,
mode="max",
delta=dp_delta)
assertTensorTuplesAlmostEqual(self,
delta_1,
delta_2,
mode="max",
delta=dp_delta)
else:
attributions_1 = attr_method_1.attribute(**args_1)
self.setUp()
attributions_2 = attr_method_2.attribute(**args_2)
assertTensorTuplesAlmostEqual(self,
attributions_1,
attributions_2,
mode="max",
delta=dp_delta)
def attribute(
self,
inputs: TensorOrTupleOfTensorsGeneric,
baselines: Union[TensorOrTupleOfTensorsGeneric,
Callable[..., TensorOrTupleOfTensorsGeneric]],
n_samples: int = 5,
stdevs: Union[float, Tuple[float, ...]] = 0.0,
target: TargetType = None,
additional_forward_args: Any = None,
return_convergence_delta: bool = False,
) -> Union[TensorOrTupleOfTensorsGeneric, Tuple[
TensorOrTupleOfTensorsGeneric, Tensor]]:
r"""
Args:
inputs (tensor or tuple of tensors): Input for which SHAP attribution
values are computed. If `forward_func` takes a single
tensor as input, a single input tensor should be provided.
If `forward_func` takes multiple tensors as input, a tuple
of the input tensors should be provided. It is assumed
that for all given input tensors, dimension 0 corresponds
to the number of examples, and if multiple input tensors
are provided, the examples must be aligned appropriately.
baselines (tensor, tuple of tensors, callable):
Baselines define the starting point from which expectation
is computed and can be provided as:
- a single tensor, if inputs is a single tensor, with
the first dimension equal to the number of examples
in the baselines' distribution. The remaining dimensions
must match with input tensor's dimension starting from
the second dimension.
- a tuple of tensors, if inputs is a tuple of tensors,
with the first dimension of any tensor inside the tuple
equal to the number of examples in the baseline's
distribution. The remaining dimensions must match
the dimensions of the corresponding input tensor
starting from the second dimension.
- callable function, optionally takes `inputs` as an
argument and either returns a single tensor
or a tuple of those.
It is recommended that the number of samples in the baselines'
tensors is larger than one.
n_samples (int, optional): The number of randomly generated examples
per sample in the input batch. Random examples are
generated by adding gaussian random noise to each sample.
Default: `5` if `n_samples` is not provided.
stdevs (float, or a tuple of floats optional): The standard deviation
of gaussian noise with zero mean that is added to each
input in the batch. If `stdevs` is a single float value
then that same value is used for all inputs. If it is
a tuple, then it must have the same length as the inputs
tuple. In this case, each stdev value in the stdevs tuple
corresponds to the input with the same index in the inputs
tuple.
Default: 0.0
target (int, tuple, tensor or list, optional): Output indices for
which gradients are computed (for classification cases,
this is usually the target class).
If the network returns a scalar value per example,
no target index is necessary.
For general 2D outputs, targets can be either:
- a single integer or a tensor containing a single
integer, which is applied to all input examples
- a list of integers or a 1D tensor, with length matching
the number of examples in inputs (dim 0). Each integer
is applied as the target for the corresponding example.
For outputs with > 2 dimensions, targets can be either:
- A single tuple, which contains #output_dims - 1
elements. This target index is applied to all examples.
- A list of tuples with length equal to the number of
examples in inputs (dim 0), and each tuple containing
#output_dims - 1 elements. Each tuple is applied as the
target for the corresponding example.
Default: None
additional_forward_args (any, optional): If the forward function
requires additional arguments other than the inputs for
which attributions should not be computed, this argument
can be provided. It can contain a tuple of ND tensors or
any arbitrary python type of any shape.
In case of the ND tensor the first dimension of the
tensor must correspond to the batch size. It will be
repeated for each `n_steps` for each randomly generated
input sample.
Note that the gradients are not computed with respect
to these arguments.
Default: None
return_convergence_delta (bool, optional): Indicates whether to return
convergence delta or not. If `return_convergence_delta`
is set to True convergence delta will be returned in
a tuple following attributions.
Default: False
Returns:
**attributions** or 2-element tuple of **attributions**, **delta**:
- **attributions** (*tensor* or tuple of *tensors*):
Attribution score computed based on GradientSHAP with respect
to each input feature. Attributions will always be
the same size as the provided inputs, with each value
providing the attribution of the corresponding input index.
If a single tensor is provided as inputs, a single tensor is
returned. If a tuple is provided for inputs, a tuple of
corresponding sized tensors is returned.
- **delta** (*tensor*, returned if return_convergence_delta=True):
This is computed using the property that the total
sum of forward_func(inputs) - forward_func(baselines)
must be very close to the total sum of the attributions
based on GradientSHAP.
Delta is calculated for each example in the input after adding
`n_samples` times gaussian noise to each of them. Therefore,
the dimensionality of the deltas tensor is equal to the
`number of examples in the input` * `n_samples`
The deltas are ordered by each input example and `n_samples`
noisy samples generated for it.
Examples::
>>> # ImageClassifier takes a single input tensor of images Nx3x32x32,
>>> # and returns an Nx10 tensor of class probabilities.
>>> net = ImageClassifier()
>>> gradient_shap = GradientShap(net)
>>> input = torch.randn(3, 3, 32, 32, requires_grad=True)
>>> # choosing baselines randomly
>>> baselines = torch.randn(20, 3, 32, 32)
>>> # Computes gradient shap for the input
>>> # Attribution size matches input size: 3x3x32x32
>>> attribution = gradient_shap.attribute(input, baselines,
target=5)
"""
# since `baselines` is a distribution, we can generate it using a function
# rather than passing it as an input argument
baselines = _format_callable_baseline(baselines, inputs)
assert isinstance(baselines[0], torch.Tensor), (
"Baselines distribution has to be provided in a form "
"of a torch.Tensor {}.".format(baselines[0]))
input_min_baseline_x_grad = InputBaselineXGradient(
self.forward_func, self.multiplies_by_inputs)
input_min_baseline_x_grad.gradient_func = self.gradient_func
nt = NoiseTunnel(input_min_baseline_x_grad)
# NOTE: using attribute.__wrapped__ to not log
attributions = nt.attribute.__wrapped__(
nt, # self
inputs,
nt_type="smoothgrad",
nt_samples=n_samples,
stdevs=stdevs,
draw_baseline_from_distrib=True,
baselines=baselines,
target=target,
additional_forward_args=additional_forward_args,
return_convergence_delta=return_convergence_delta,
)
return attributions
def jit_test_assert(self) -> None:
model_1 = model
attr_args = args
if (mode is JITCompareMode.data_parallel_jit_trace
or JITCompareMode.data_parallel_jit_script):
if not torch.cuda.is_available() or torch.cuda.device_count(
) == 0:
raise unittest.SkipTest(
"Skipping GPU test since CUDA not available.")
# Construct cuda_args, moving all tensor inputs in args to CUDA device
cuda_args = {}
for key in args:
if isinstance(args[key], Tensor):
cuda_args[key] = args[key].cuda()
elif isinstance(args[key], tuple):
cuda_args[key] = tuple(
elem.cuda() if isinstance(elem, Tensor) else elem
for elem in args[key])
else:
cuda_args[key] = args[key]
attr_args = cuda_args
model_1 = model_1.cuda()
# Initialize models based on JITCompareMode
if (mode is JITCompareMode.cpu_jit_script
or JITCompareMode.data_parallel_jit_script):
model_2 = torch.jit.script(model_1) # type: ignore
elif (mode is JITCompareMode.cpu_jit_trace
or JITCompareMode.data_parallel_jit_trace):
all_inps = _format_input(args["inputs"]) + (
_format_additional_forward_args(
args["additional_forward_args"])
if "additional_forward_args" in args and
args["additional_forward_args"] is not None else tuple())
model_2 = torch.jit.trace(model_1, all_inps) # type: ignore
else:
raise AssertionError("JIT compare mode type is not valid.")
attr_method_1 = algorithm(model_1)
attr_method_2 = algorithm(model_2)
if noise_tunnel:
attr_method_1 = NoiseTunnel(attr_method_1)
attr_method_2 = NoiseTunnel(attr_method_2)
if attr_method_1.has_convergence_delta():
attributions_1, delta_1 = attr_method_1.attribute(
return_convergence_delta=True, **attr_args)
self.setUp()
attributions_2, delta_2 = attr_method_2.attribute(
return_convergence_delta=True, **attr_args)
assertTensorTuplesAlmostEqual(self,
attributions_1,
attributions_2,
mode="max")
assertTensorTuplesAlmostEqual(self,
delta_1,
delta_2,
mode="max")
else:
attributions_1 = attr_method_1.attribute(**attr_args)
self.setUp()
attributions_2 = attr_method_2.attribute(**attr_args)
assertTensorTuplesAlmostEqual(self,
attributions_1,
attributions_2,
mode="max")
def target_test_assert(self) -> None:
attr_method: Attribution
if target_layer:
internal_algorithm = cast(Type[InternalAttribution], algorithm)
attr_method = internal_algorithm(model, target_layer)
else:
attr_method = algorithm(model)
if noise_tunnel:
attr_method = NoiseTunnel(attr_method)
attributions_orig = attr_method.attribute(**args)
self.setUp()
for i in range(num_examples):
args["target"] = (original_targets[i] if len(original_targets)
== num_examples else original_targets)
args["inputs"] = (original_inputs[i:i + 1] if isinstance(
original_inputs, Tensor) else tuple(
original_inp[i:i + 1]
for original_inp in original_inputs))
if original_additional_forward_args is not None:
args["additional_forward_args"] = tuple(
single_add_arg[i:i + 1] if isinstance(
single_add_arg, Tensor) else single_add_arg
for single_add_arg in original_additional_forward_args)
if replace_baselines:
if isinstance(original_inputs, Tensor):
args["baselines"] = original_baselines[i:i + 1]
elif isinstance(original_baselines, tuple):
args["baselines"] = tuple(
single_baseline[i:i + 1] if isinstance(
single_baseline, Tensor) else single_baseline
for single_baseline in original_baselines)
# Since Lime methods compute attributions for a batch
# sequentially, random seed should not be reset after
# each example after the first.
if not issubclass(algorithm, Lime):
self.setUp()
single_attr = attr_method.attribute(**args)
current_orig_attributions = (
attributions_orig[i:i + 1] if isinstance(
attributions_orig, Tensor) else tuple(
single_attrib[i:i + 1]
for single_attrib in attributions_orig))
assertTensorTuplesAlmostEqual(
self,
current_orig_attributions,
single_attr,
delta=target_delta,
mode="max",
)
if (not issubclass(algorithm, Lime)
and len(original_targets) == num_examples):
# If original_targets contained multiple elements, then
# we also compare with setting targets to a list with
# a single element.
args["target"] = original_targets[i:i + 1]
self.setUp()
single_attr_target_list = attr_method.attribute(**args)
assertTensorTuplesAlmostEqual(
self,
current_orig_attributions,
single_attr_target_list,
delta=target_delta,
mode="max",
)
def _compute_attribution_and_evaluate(
self,
model: Module,
inputs: TensorOrTupleOfTensorsGeneric,
baselines: BaselineType = None,
target: Union[None, int] = None,
additional_forward_args: Any = None,
type: str = "vanilla",
approximation_method: str = "gausslegendre",
multiply_by_inputs=True,
) -> Tuple[Tensor, ...]:
r"""
attrib_type: 'vanilla', 'smoothgrad', 'smoothgrad_sq', 'vargrad'
"""
ig = IntegratedGradients(model, multiply_by_inputs=multiply_by_inputs)
self.assertEquals(ig.multiplies_by_inputs, multiply_by_inputs)
if not isinstance(inputs, tuple):
inputs = (inputs,) # type: ignore
inputs: Tuple[Tensor, ...]
if baselines is not None and not isinstance(baselines, tuple):
baselines = (baselines,)
if baselines is None:
baselines = _tensorize_baseline(inputs, _zeros(inputs))
if type == "vanilla":
attributions, delta = ig.attribute(
inputs,
baselines,
additional_forward_args=additional_forward_args,
method=approximation_method,
n_steps=500,
target=target,
return_convergence_delta=True,
)
model.zero_grad()
attributions_without_delta, delta = ig.attribute(
inputs,
baselines,
additional_forward_args=additional_forward_args,
method=approximation_method,
n_steps=500,
target=target,
return_convergence_delta=True,
)
model.zero_grad()
self.assertEqual([inputs[0].shape[0]], list(delta.shape))
delta_external = ig.compute_convergence_delta(
attributions,
baselines,
inputs,
target=target,
additional_forward_args=additional_forward_args,
)
assertArraysAlmostEqual(delta, delta_external, 0.0)
else:
nt = NoiseTunnel(ig)
n_samples = 5
attributions, delta = nt.attribute(
inputs,
nt_type=type,
nt_samples=n_samples,
stdevs=0.00000002,
baselines=baselines,
target=target,
additional_forward_args=additional_forward_args,
method=approximation_method,
n_steps=500,
return_convergence_delta=True,
)
with self.assertWarns(DeprecationWarning):
attributions_without_delta = nt.attribute(
inputs,
nt_type=type,
n_samples=n_samples,
stdevs=0.00000002,
baselines=baselines,
target=target,
additional_forward_args=additional_forward_args,
method=approximation_method,
n_steps=500,
)
self.assertEquals(nt.multiplies_by_inputs, multiply_by_inputs)
self.assertEqual([inputs[0].shape[0] * n_samples], list(delta.shape))
for input, attribution in zip(inputs, attributions):
self.assertEqual(attribution.shape, input.shape)
if multiply_by_inputs:
self.assertTrue(all(abs(delta.numpy().flatten()) < 0.07))
# compare attributions retrieved with and without
# `return_convergence_delta` flag
for attribution, attribution_without_delta in zip(
attributions, attributions_without_delta
):
assertTensorAlmostEqual(
self, attribution, attribution_without_delta, delta=0.05
)
return cast(Tuple[Tensor, ...], attributions)
def _compute_attribution_and_evaluate(
self,
model,
inputs,
baselines=None,
target=None,
additional_forward_args=None,
type="vanilla",
):
r"""
attrib_type: 'vanilla', 'smoothgrad', 'smoothgrad_sq', 'vargrad'
"""
ig = IntegratedGradients(model)
if not isinstance(inputs, tuple):
inputs = (inputs,)
if baselines is not None and not isinstance(baselines, tuple):
baselines = (baselines,)
if baselines is None:
baselines = _zeros(inputs)
for method in [
"riemann_right",
"riemann_left",
"riemann_middle",
"riemann_trapezoid",
"gausslegendre",
]:
if type == "vanilla":
attributions, delta = ig.attribute(
inputs,
baselines,
additional_forward_args=additional_forward_args,
method=method,
n_steps=2000,
target=target,
return_convergence_delta=True,
)
model.zero_grad()
attributions_without_delta, delta = ig.attribute(
inputs,
baselines,
additional_forward_args=additional_forward_args,
method=method,
n_steps=2000,
target=target,
return_convergence_delta=True,
)
model.zero_grad()
self.assertEqual([inputs[0].shape[0]], list(delta.shape))
delta_external = ig.compute_convergence_delta(
attributions,
baselines,
inputs,
target=target,
additional_forward_args=additional_forward_args,
)
assertArraysAlmostEqual(delta, delta_external, 0.0)
else:
nt = NoiseTunnel(ig)
n_samples = 5
attributions, delta = nt.attribute(
inputs,
nt_type=type,
n_samples=n_samples,
stdevs=0.00000002,
baselines=baselines,
target=target,
additional_forward_args=additional_forward_args,
method=method,
n_steps=2000,
return_convergence_delta=True,
)
attributions_without_delta = nt.attribute(
inputs,
nt_type=type,
n_samples=n_samples,
stdevs=0.00000002,
baselines=baselines,
target=target,
additional_forward_args=additional_forward_args,
method=method,
n_steps=2000,
)
self.assertEqual([inputs[0].shape[0] * n_samples], list(delta.shape))
for input, attribution in zip(inputs, attributions):
self.assertEqual(attribution.shape, input.shape)
self.assertTrue(all(abs(delta.numpy().flatten()) < 0.05))
# compare attributions retrieved with and without
# `return_convergence_delta` flag
for attribution, attribution_without_delta in zip(
attributions, attributions_without_delta
):
assertTensorAlmostEqual(
self, attribution, attribution_without_delta, delta=0.05
)
return attributions
