def _gradient_matching_test_assert(self, model: Module,
output_layer: Module,
test_input: Tensor) -> None:
out = _forward_layer_eval(model, test_input, output_layer)
# Select first element of tuple
out = out[0]
gradient_attrib = NeuronGradient(model, output_layer)
self.assertFalse(gradient_attrib.multiplies_by_inputs)
for i in range(cast(Tuple[int, ...], out.shape)[1]):
neuron: Tuple[int, ...] = (i, )
while len(neuron) < len(out.shape) - 1:
neuron = neuron + (0, )
input_attrib = Saliency(lambda x: _forward_layer_eval(
model, x, output_layer, grad_enabled=True)[0][
(slice(None), *neuron)])
sal_vals = input_attrib.attribute(test_input, abs=False)
grad_vals = gradient_attrib.attribute(test_input, neuron)
# Verify matching sizes
self.assertEqual(grad_vals.shape, sal_vals.shape)
self.assertEqual(grad_vals.shape, test_input.shape)
assertArraysAlmostEqual(
sal_vals.reshape(-1).tolist(),
grad_vals.reshape(-1).tolist(),
delta=0.001,
)
def forward_eval_layer_with_inputs_helper(self, model, inputs_to_test):
# hard coding for simplicity
# 0 if using args, 1 if using kwargs
# => no 0s after first 1 (left to right)
#
# used to test utilization of args/kwargs
use_args_or_kwargs = [
[[0], [1]],
[
[0, 0],
[0, 1],
[1, 1],
],
]
model = ModelInputWrapper(model)
def forward_func(*args, args_or_kwargs=None):
# convert to args or kwargs to test *args and **kwargs wrapping behavior
new_args = []
new_kwargs = {}
for args_or_kwarg, name, inp in zip(args_or_kwargs,
inputs_to_test.keys(), args):
if args_or_kwarg:
new_kwargs[name] = inp
else:
new_args.append(inp)
return model(*new_args, **new_kwargs)
for args_or_kwargs in use_args_or_kwargs[len(inputs_to_test) - 1]:
with self.subTest(args_or_kwargs=args_or_kwargs):
inputs = _forward_layer_eval(
functools.partial(forward_func,
args_or_kwargs=args_or_kwargs),
inputs=tuple(inputs_to_test.values()),
layer=[
model.input_maps[name]
for name in inputs_to_test.keys()
],
)
inputs_with_attrib_to_inp = _forward_layer_eval(
functools.partial(forward_func,
args_or_kwargs=args_or_kwargs),
inputs=tuple(inputs_to_test.values()),
layer=[
model.input_maps[name]
for name in inputs_to_test.keys()
],
attribute_to_layer_input=True,
)
for i1, i2, i3 in zip(inputs, inputs_with_attrib_to_inp,
inputs_to_test.values()):
self.assertTrue((i1[0] == i2[0]).all())
self.assertTrue((i1[0] == i3).all())
def neuron_forward_func(*args: Any):
with torch.no_grad():
layer_eval = _forward_layer_eval(
self.forward_func,
args,
self.layer,
device_ids=self.device_ids,
attribute_to_layer_input=attribute_to_neuron_input,
)
return _verify_select_neuron(layer_eval, neuron_selector)
def attribute(
self,
inputs: Union[Tensor, Tuple[Tensor, ...]],
baselines: BaselineType = None,
target: TargetType = None,
additional_forward_args: Any = None,
n_steps: int = 50,
method: str = "gausslegendre",
internal_batch_size: Union[None, int] = None,
return_convergence_delta: bool = False,
attribute_to_layer_input: bool = False,
) -> Union[Union[Tensor, Tuple[Tensor, ...], List[Union[Tensor, Tuple[
Tensor, ...]]]], Tuple[Union[Tensor, Tuple[
Tensor, ...], List[Union[Tensor, Tuple[Tensor, ...]]]],
Tensor, ], ]:
r"""
This method attributes the output of the model with given target index
(in case it is provided, otherwise it assumes that output is a
scalar) to layer inputs or outputs of the model, depending on whether
`attribute_to_layer_input` is set to True or False, using the approach
described above.
In addition to that it also returns, if `return_convergence_delta` is
set to True, integral approximation delta based on the completeness
property of integrated gradients.
Args:
inputs (tensor or tuple of tensors): Input for which layer integrated
gradients 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 (scalar, tensor, tuple of scalars or tensors, optional):
Baselines define the starting point from which integral
is computed and can be provided as:
- a single tensor, if inputs is a single tensor, with
exactly the same dimensions as inputs or the first
dimension is one and the remaining dimensions match
with inputs.
- a single scalar, if inputs is a single tensor, which will
be broadcasted for each input value in input tensor.
- a tuple of tensors or scalars, the baseline corresponding
to each tensor in the inputs' tuple can be:
- either a tensor with matching dimensions to
corresponding tensor in the inputs' tuple
or the first dimension is one and the remaining
dimensions match with the corresponding
input tensor.
- or a scalar, corresponding to a tensor in the
inputs' tuple. This scalar value is broadcasted
for corresponding input tensor.
In the cases when `baselines` is not provided, we internally
use zero scalar corresponding to each input tensor.
Default: None
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 must be either a single additional
argument of a Tensor or arbitrary (non-tuple) type or a
tuple containing multiple additional arguments including
tensors or any arbitrary python types. These arguments
are provided to forward_func in order following the
arguments in inputs.
For a tensor, the first dimension of the tensor must
correspond to the number of examples. It will be
repeated for each of `n_steps` along the integrated
path. For all other types, the given argument is used
for all forward evaluations.
Note that attributions are not computed with respect
to these arguments.
Default: None
n_steps (int, optional): The number of steps used by the approximation
method. Default: 50.
method (string, optional): Method for approximating the integral,
one of `riemann_right`, `riemann_left`, `riemann_middle`,
`riemann_trapezoid` or `gausslegendre`.
Default: `gausslegendre` if no method is provided.
internal_batch_size (int, optional): Divides total #steps * #examples
data points into chunks of size at most internal_batch_size,
which are computed (forward / backward passes)
sequentially. internal_batch_size must be at least equal to
#examples.
For DataParallel models, each batch is split among the
available devices, so evaluations on each available
device contain internal_batch_size / num_devices examples.
If internal_batch_size is None, then all evaluations are
processed in one batch.
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
attribute_to_layer_input (bool, optional): Indicates whether to
compute the attribution with respect to the layer input
or output. If `attribute_to_layer_input` is set to True
then the attributions will be computed with respect to
layer input, otherwise it will be computed with respect
to layer output.
Note that currently it is assumed that either the input
or the output of internal layer, depending on whether we
attribute to the input or output, is a single tensor.
Support for multiple tensors will be added later.
Default: False
Returns:
**attributions** or 2-element tuple of **attributions**, **delta**:
- **attributions** (*tensor*, tuple of *tensors* or tuple of *tensors*):
Integrated gradients with respect to `layer`'s inputs or
outputs. Attributions will always be the same size and
dimensionality as the input or output of the given layer,
depending on whether we attribute to the inputs or outputs
of the layer which is decided by the input flag
`attribute_to_layer_input`.
For a single layer, attributions are returned in a tuple if
the layer inputs / outputs contain multiple tensors,
otherwise a single tensor is returned.
For multiple layers, attributions will always be
returned as a list. Each element in this list will be
equivalent to that of a single layer output, i.e. in the
case that one layer, in the given layers, inputs / outputs
multiple tensors: the corresponding output element will be
a tuple of tensors. The ordering of the outputs will be
the same order as the layers given in the constructor.
- **delta** (*tensor*, returned if return_convergence_delta=True):
The difference between the total approximated and true
integrated gradients. This is computed using the property
that the total sum of forward_func(inputs) -
forward_func(baselines) must equal the total sum of the
integrated gradient.
Delta is calculated per example, meaning that the number of
elements in returned delta tensor is equal to the number of
of examples in inputs.
Examples::
>>> # ImageClassifier takes a single input tensor of images Nx3x32x32,
>>> # and returns an Nx10 tensor of class probabilities.
>>> # It contains an attribute conv1, which is an instance of nn.conv2d,
>>> # and the output of this layer has dimensions Nx12x32x32.
>>> net = ImageClassifier()
>>> lig = LayerIntegratedGradients(net, net.conv1)
>>> input = torch.randn(2, 3, 32, 32, requires_grad=True)
>>> # Computes layer integrated gradients for class 3.
>>> # attribution size matches layer output, Nx12x32x32
>>> attribution = lig.attribute(input, target=3)
"""
inps, baselines = _format_input_baseline(inputs, baselines)
_validate_input(inps, baselines, n_steps, method)
baselines = _tensorize_baseline(inps, baselines)
additional_forward_args = _format_additional_forward_args(
additional_forward_args)
def flatten_tuple(tup):
return tuple(
sum((list(x) if isinstance(x, (tuple, list)) else [x]
for x in tup), []))
if self.device_ids is None:
self.device_ids = getattr(self.forward_func, "device_ids", None)
inputs_layer = _forward_layer_eval(
self.forward_func,
inps,
self.layer,
device_ids=self.device_ids,
additional_forward_args=additional_forward_args,
attribute_to_layer_input=attribute_to_layer_input,
)
# if we have one output
if not isinstance(self.layer, list):
inputs_layer = (inputs_layer, )
num_outputs = [
1 if isinstance(x, Tensor) else len(x) for x in inputs_layer
]
num_outputs_cumsum = torch.cumsum(
torch.IntTensor([0] + num_outputs),
dim=0 # type: ignore
)
inputs_layer = flatten_tuple(inputs_layer)
baselines_layer = _forward_layer_eval(
self.forward_func,
baselines,
self.layer,
device_ids=self.device_ids,
additional_forward_args=additional_forward_args,
attribute_to_layer_input=attribute_to_layer_input,
)
baselines_layer = flatten_tuple(baselines_layer)
# inputs -> these inputs are scaled
def gradient_func(
forward_fn: Callable,
inputs: Union[Tensor, Tuple[Tensor, ...]],
target_ind: TargetType = None,
additional_forward_args: Any = None,
) -> Tuple[Tensor, ...]:
if self.device_ids is None or len(self.device_ids) == 0:
scattered_inputs = (inputs, )
else:
# scatter method does not have a precise enough return type in its
# stub, so suppress the type warning.
scattered_inputs = scatter( # type:ignore
inputs, target_gpus=self.device_ids)
scattered_inputs_dict = {
scattered_input[0].device: scattered_input
for scattered_input in scattered_inputs
}
with torch.autograd.set_grad_enabled(True):
def layer_forward_hook(module,
hook_inputs,
hook_outputs=None,
layer_idx=0):
device = _extract_device(module, hook_inputs, hook_outputs)
is_layer_tuple = (
isinstance(hook_outputs, tuple)
# hook_outputs is None if attribute_to_layer_input == True
if hook_outputs is not None else isinstance(
hook_inputs, tuple))
if is_layer_tuple:
return scattered_inputs_dict[device][
num_outputs_cumsum[layer_idx]:num_outputs_cumsum[
layer_idx + 1]]
return scattered_inputs_dict[device][
num_outputs_cumsum[layer_idx]]
hooks = []
try:
layers = self.layer
if not isinstance(layers, list):
layers = [self.layer]
for layer_idx, layer in enumerate(layers):
hook = None
# TODO:
# Allow multiple attribute_to_layer_input flags for
# each layer, i.e. attribute_to_layer_input[layer_idx]
if attribute_to_layer_input:
hook = layer.register_forward_pre_hook(
functools.partial(layer_forward_hook,
layer_idx=layer_idx))
else:
hook = layer.register_forward_hook(
functools.partial(layer_forward_hook,
layer_idx=layer_idx))
hooks.append(hook)
output = _run_forward(self.forward_func, tuple(),
target_ind, additional_forward_args)
finally:
for hook in hooks:
if hook is not None:
hook.remove()
assert output[0].numel() == 1, (
"Target not provided when necessary, cannot"
" take gradient with respect to multiple outputs.")
# torch.unbind(forward_out) is a list of scalar tensor tuples and
# contains batch_size * #steps elements
grads = torch.autograd.grad(torch.unbind(output), inputs)
return grads
self.ig.gradient_func = gradient_func
all_inputs = ((inps + additional_forward_args)
if additional_forward_args is not None else inps)
attributions = self.ig.attribute.__wrapped__( # type: ignore
self.ig, # self
inputs_layer,
baselines=baselines_layer,
target=target,
additional_forward_args=all_inputs,
n_steps=n_steps,
method=method,
internal_batch_size=internal_batch_size,
return_convergence_delta=False,
)
# handle multiple outputs
output: List[Tuple[Tensor, ...]] = [
tuple(attributions[int(num_outputs_cumsum[i]
):int(num_outputs_cumsum[i + 1])])
for i in range(len(num_outputs))
]
if return_convergence_delta:
start_point, end_point = baselines, inps
# computes approximation error based on the completeness axiom
delta = self.compute_convergence_delta(
attributions,
start_point,
end_point,
additional_forward_args=additional_forward_args,
target=target,
)
return _format_outputs(isinstance(self.layer, list), output), delta
return _format_outputs(isinstance(self.layer, list), output)
def attribute( # type: ignore
self,
inputs: Union[Tensor, Tuple[Tensor, ...]],
baselines: Union[Tensor, Tuple[Tensor, ...]],
target: TargetType = None,
additional_forward_args: Any = None,
return_convergence_delta: bool = False,
attribute_to_layer_input: bool = False,
) -> Union[Tensor, Tuple[Tensor, ...], Tuple[Union[Tensor, Tuple[
Tensor, ...]], Tensor]]:
inputs, baselines = _format_input_baseline(inputs, baselines)
rand_coefficient = torch.tensor(
np.random.uniform(0.0, 1.0, inputs[0].shape[0]),
device=inputs[0].device,
dtype=inputs[0].dtype,
)
input_baseline_scaled = tuple(
_scale_input(input, baseline, rand_coefficient)
for input, baseline in zip(inputs, baselines))
grads, _ = compute_layer_gradients_and_eval(
self.forward_func,
self.layer,
input_baseline_scaled,
target,
additional_forward_args,
device_ids=self.device_ids,
attribute_to_layer_input=attribute_to_layer_input,
)
attr_baselines = _forward_layer_eval(
self.forward_func,
baselines,
self.layer,
additional_forward_args=additional_forward_args,
device_ids=self.device_ids,
attribute_to_layer_input=attribute_to_layer_input,
)
attr_inputs = _forward_layer_eval(
self.forward_func,
inputs,
self.layer,
additional_forward_args=additional_forward_args,
device_ids=self.device_ids,
attribute_to_layer_input=attribute_to_layer_input,
)
if self.multiplies_by_inputs:
input_baseline_diffs = tuple(
input - baseline
for input, baseline in zip(attr_inputs, attr_baselines))
attributions = tuple(input_baseline_diff * grad
for input_baseline_diff, grad in zip(
input_baseline_diffs, grads))
else:
attributions = grads
return _compute_conv_delta_and_format_attrs(
self,
return_convergence_delta,
attributions,
baselines,
inputs,
additional_forward_args,
target,
cast(Union[Literal[True], Literal[False]],
len(attributions) > 1),
)
def attribute(
self,
inputs: Union[Tensor, Tuple[Tensor, ...]],
layer_baselines: BaselineType = None,
target: TargetType = None,
additional_forward_args: Any = None,
layer_mask: Union[None, Tensor, Tuple[Tensor, ...]] = None,
attribute_to_layer_input: bool = False,
perturbations_per_eval: int = 1,
) -> Union[Tensor, Tuple[Tensor, ...]]:
r"""
Args:
inputs (tensor or tuple of tensors): Input for which layer
attributions 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.
layer_baselines (scalar, tensor, tuple of scalars or tensors, optional):
Layer baselines define reference values which replace each
layer input / output value when ablated.
Layer baselines should be a single tensor with dimensions
matching the input / output of the target layer (or
broadcastable to match it), based
on whether we are attributing to the input or output
of the target layer.
In the cases when `baselines` is not provided, we internally
use zero as the baseline for each neuron.
Default: None
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 must be either a single additional
argument of a Tensor or arbitrary (non-tuple) type or a
tuple containing multiple additional arguments including
tensors or any arbitrary python types. These arguments
are provided to forward_func in order following the
arguments in inputs.
Note that attributions are not computed with respect
to these arguments.
Default: None
layer_mask (tensor or tuple of tensors, optional):
layer_mask defines a mask for the layer, grouping
elements of the layer input / output which should be
ablated together.
layer_mask should be a single tensor with dimensions
matching the input / output of the target layer (or
broadcastable to match it), based
on whether we are attributing to the input or output
of the target layer. layer_mask
should contain integers in the range 0 to num_groups
- 1, and all elements with the same value are
considered to be in the same group.
If None, then a layer mask is constructed which assigns
each neuron within the layer as a separate group, which
is ablated independently.
Default: None
attribute_to_layer_input (bool, optional): Indicates whether to
compute the attributions with respect to the layer input
or output. If `attribute_to_layer_input` is set to True
then the attributions will be computed with respect to
layer's inputs, otherwise it will be computed with respect
to layer's outputs.
Note that currently it is assumed that either the input
or the output of the layer, depending on whether we
attribute to the input or output, is a single tensor.
Support for multiple tensors will be added later.
Default: False
perturbations_per_eval (int, optional): Allows ablation of multiple
neuron (groups) to be processed simultaneously in one
call to forward_fn.
Each forward pass will contain a maximum of
perturbations_per_eval * #examples samples.
For DataParallel models, each batch is split among the
available devices, so evaluations on each available
device contain at most
(perturbations_per_eval * #examples) / num_devices
samples.
Default: 1
Returns:
*tensor* or tuple of *tensors* of **attributions**:
- **attributions** (*tensor* or tuple of *tensors*):
Attribution of each neuron in given layer input or
output. Attributions will always be the same size as
the input or output of the given layer, depending on
whether we attribute to the inputs or outputs
of the layer which is decided by the input flag
`attribute_to_layer_input`
Attributions are returned in a tuple if
the layer inputs / outputs contain multiple tensors,
otherwise a single tensor is returned.
Examples::
>>> # SimpleClassifier takes a single input tensor of size Nx4x4,
>>> # and returns an Nx3 tensor of class probabilities.
>>> # It contains an attribute conv1, which is an instance of nn.conv2d,
>>> # and the output of this layer has dimensions Nx12x3x3.
>>> net = SimpleClassifier()
>>> # Generating random input with size 2 x 4 x 4
>>> input = torch.randn(2, 4, 4)
>>> # Defining LayerFeatureAblation interpreter
>>> ablator = LayerFeatureAblation(net, net.conv1)
>>> # Computes ablation attribution, ablating each of the 108
>>> # neurons independently.
>>> attr = ablator.attribute(input, target=1)
>>> # Alternatively, we may want to ablate neurons in groups, e.g.
>>> # grouping all the layer outputs in the same row.
>>> # This can be done by creating a layer mask as follows, which
>>> # defines the groups of layer inputs / outouts, e.g.:
>>> # +---+---+---+
>>> # | 0 | 0 | 0 |
>>> # +---+---+---+
>>> # | 1 | 1 | 1 |
>>> # +---+---+---+
>>> # | 2 | 2 | 2 |
>>> # +---+---+---+
>>> # With this mask, all the 36 neurons in a row / channel are ablated
>>> # simultaneously, and the attribution for each neuron in the same
>>> # group (0 - 2) per example are the same.
>>> # The attributions can be calculated as follows:
>>> # layer mask has dimensions 1 x 3 x 3
>>> layer_mask = torch.tensor([[[0,0,0],[1,1,1],
>>> [2,2,2]]])
>>> attr = ablator.attribute(input, target=1,
>>> layer_mask=layer_mask)
"""
def layer_forward_func(*args):
layer_length = args[-1]
layer_input = args[:layer_length]
original_inputs = args[layer_length:-1]
device_ids = self.device_ids
if device_ids is None:
device_ids = getattr(self.forward_func, "device_ids", None)
all_layer_inputs = {}
if device_ids is not None:
scattered_layer_input = scatter(layer_input,
target_gpus=device_ids)
for device_tensors in scattered_layer_input:
all_layer_inputs[device_tensors[0].device] = device_tensors
else:
all_layer_inputs[layer_input[0].device] = layer_input
def forward_hook(module, inp, out=None):
device = _extract_device(module, inp, out)
is_layer_tuple = (isinstance(out, tuple) if out is not None
else isinstance(inp, tuple))
if device not in all_layer_inputs:
raise AssertionError(
"Layer input not placed on appropriate "
"device. If using a DataParallel model, either provide the "
"DataParallel model as forward_func or provide device ids"
" to the constructor.")
if not is_layer_tuple:
return all_layer_inputs[device][0]
return all_layer_inputs[device]
hook = None
try:
if attribute_to_layer_input:
hook = self.layer.register_forward_pre_hook(forward_hook)
else:
hook = self.layer.register_forward_hook(forward_hook)
eval = _run_forward(self.forward_func,
original_inputs,
target=target)
finally:
if hook is not None:
hook.remove()
return eval
with torch.no_grad():
inputs = _format_tensor_into_tuples(inputs)
additional_forward_args = _format_additional_forward_args(
additional_forward_args)
layer_eval = _forward_layer_eval(
self.forward_func,
inputs,
self.layer,
additional_forward_args,
device_ids=self.device_ids,
attribute_to_layer_input=attribute_to_layer_input,
)
layer_eval_len = (len(layer_eval), )
all_inputs = ((inputs + additional_forward_args + layer_eval_len)
if additional_forward_args is not None else inputs +
layer_eval_len)
ablator = FeatureAblation(layer_forward_func)
layer_attribs = ablator.attribute.__wrapped__(
ablator, # self
layer_eval,
baselines=layer_baselines,
additional_forward_args=all_inputs,
feature_mask=layer_mask,
perturbations_per_eval=perturbations_per_eval,
)
_attr = _format_output(len(layer_attribs) > 1, layer_attribs)
return _attr
def attribute(
self,
inputs: Union[Tensor, Tuple[Tensor, ...]],
additional_forward_args: Any = None,
attribute_to_layer_input: bool = False,
) -> Union[Tensor, Tuple[Tensor, ...], List[Union[Tensor, Tuple[Tensor,
...]]]]:
r"""
Args:
inputs (tensor or tuple of tensors): Input for which layer
activation is 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.
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 must be either a single additional
argument of a Tensor or arbitrary (non-tuple) type or a
tuple containing multiple additional arguments including
tensors or any arbitrary python types. These arguments
are provided to forward_func in order following the
arguments in inputs.
Note that attributions are not computed with respect
to these arguments.
Default: None
attribute_to_layer_input (bool, optional): Indicates whether to
compute the attribution with respect to the layer input
or output. If `attribute_to_layer_input` is set to True
then the attributions will be computed with respect to
layer input, otherwise it will be computed with respect
to layer output.
Note that currently it is assumed that either the input
or the output of internal layer, depending on whether we
attribute to the input or output, is a single tensor.
Support for multiple tensors will be added later.
Default: False
Returns:
*tensor* or tuple of *tensors* or *list* of **attributions**:
- **attributions** (*tensor* or tuple of *tensors* or *list*):
Activation of each neuron in given layer output.
Attributions will always be the same size as the
output of the given layer.
Attributions are returned in a tuple if
the layer inputs / outputs contain multiple tensors,
otherwise a single tensor is returned.
If multiple layers are provided, attributions
are returned as a list, each element corresponding to the
activations of the corresponding layer.
Examples::
>>> # ImageClassifier takes a single input tensor of images Nx3x32x32,
>>> # and returns an Nx10 tensor of class probabilities.
>>> # It contains an attribute conv1, which is an instance of nn.conv2d,
>>> # and the output of this layer has dimensions Nx12x32x32.
>>> net = ImageClassifier()
>>> layer_act = LayerActivation(net, net.conv1)
>>> input = torch.randn(2, 3, 32, 32, requires_grad=True)
>>> # Computes layer activation.
>>> # attribution is layer output, with size Nx12x32x32
>>> attribution = layer_cond.attribute(input)
"""
with torch.no_grad():
layer_eval = _forward_layer_eval(
self.forward_func,
inputs,
self.layer,
additional_forward_args,
device_ids=self.device_ids,
attribute_to_layer_input=attribute_to_layer_input,
)
if isinstance(self.layer, Module):
return _format_output(len(layer_eval) > 1, layer_eval)
else:
return [
_format_output(len(single_layer_eval) > 1, single_layer_eval)
for single_layer_eval in layer_eval
]
