#!/usr/bin/env python3
# pyre-strict
import functools
import warnings
from typing import (
Any,
Callable,
cast,
Dict,
List,
Literal,
Optional,
overload,
Tuple,
Union,
)
import torch
from captum._utils.common import (
_extract_device,
_format_additional_forward_args,
_format_outputs,
)
from captum._utils.gradient import _forward_layer_eval, _run_forward
from captum._utils.typing import BaselineType, ModuleOrModuleList, TargetType
from captum.attr._core.integrated_gradients import IntegratedGradients
from captum.attr._utils.attribution import GradientAttribution, LayerAttribution
from captum.attr._utils.common import (
_format_input_baseline,
_tensorize_baseline,
_validate_input,
)
from captum.log import log_usage
from torch import Tensor
from torch.nn.parallel.scatter_gather import scatter
[docs]
class LayerIntegratedGradients(LayerAttribution, GradientAttribution):
r"""
Layer Integrated Gradients is a variant of Integrated Gradients that assigns
an importance score to layer inputs or outputs, depending on whether we
attribute to the former or to the latter one.
Integrated Gradients is an axiomatic model interpretability algorithm that
attributes / assigns an importance score to each input feature by approximating
the integral of gradients of the model's output with respect to the inputs
along the path (straight line) from given baselines / references to inputs.
Baselines can be provided as input arguments to attribute method.
To approximate the integral we can choose to use either a variant of
Riemann sum or Gauss-Legendre quadrature rule.
More details regarding the integrated gradients method can be found in the
original paper:
https://arxiv.org/abs/1703.01365
"""
def __init__(
self,
# pyre-fixme[24]: Generic type `Callable` expects 2 type parameters.
forward_func: Callable,
layer: ModuleOrModuleList,
device_ids: Union[None, List[int]] = None,
multiply_by_inputs: bool = True,
) -> None:
r"""
Args:
forward_func (Callable): The forward function of the model or any
modification of it
layer (ModuleOrModuleList): Layer or list of layers for which attributions
are computed. For each layer the output size of the attribute
matches this layer's input or output dimensions, depending on
whether we attribute to the inputs or outputs of the
layer, corresponding to the attribution of each neuron
in the input or output of this layer.
Please note that layers to attribute on cannot be
dependent on each other. That is, a subset of layers in
`layer` cannot produce the inputs for another layer.
For example, if your model is of a simple linked-list
based graph structure (think nn.Sequence), e.g. x -> l1
-> l2 -> l3 -> output. If you pass in any one of those
layers, you cannot pass in another due to the
dependence, e.g. if you pass in l2 you cannot pass in
l1 or l3.
device_ids (list[int]): Device ID list, necessary only if forward_func
applies a DataParallel model. This allows reconstruction of
intermediate outputs from batched results across devices.
If forward_func is given as the DataParallel model itself,
then it is not necessary to provide this argument.
multiply_by_inputs (bool, optional): Indicates whether to factor
model inputs' multiplier in the final attribution scores.
In the literature this is also known as local vs global
attribution. If inputs' multiplier isn't factored in,
then this type of attribution method is also called local
attribution. If it is, then that type of attribution
method is called global.
More detailed can be found here:
https://arxiv.org/abs/1711.06104
In case of layer integrated gradients, if `multiply_by_inputs`
is set to True, final sensitivity scores are being multiplied by
layer activations for inputs - layer activations for baselines.
"""
LayerAttribution.__init__(self, forward_func, layer, device_ids=device_ids)
GradientAttribution.__init__(self, forward_func)
self.ig = IntegratedGradients(forward_func, multiply_by_inputs)
if isinstance(layer, list) and len(layer) > 1:
warnings.warn(
"Multiple layers provided. Please ensure that each layer is"
"**not** solely dependent on the outputs of"
"another layer. Please refer to the documentation for more"
"detail.",
stacklevel=2,
)
def _make_gradient_func(
self,
num_outputs_cumsum: Tensor,
attribute_to_layer_input: bool,
grad_kwargs: Optional[Dict[str, Any]],
) -> Callable[..., Tuple[Tensor, ...]]:
def _gradient_func(
# pyre-fixme[24]: Generic type `Callable` expects 2 type parameters.
forward_fn: Callable,
inputs: Union[Tensor, Tuple[Tensor, ...]],
target_ind: TargetType = None,
additional_forward_args: Optional[object] = 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
# pyre-fixme[6]: For 1st argument expected `Tensor` but got
# `Union[Tensor, typing.Tuple[Tensor, ...]]`.
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):
# pyre-fixme[53]: Captured variable `num_outputs_cumsum` is not
# annotated.
# pyre-fixme[53]: Captured variable `scattered_inputs_dict` is not
# annotated.
# pyre-fixme[3]: Return type must be annotated.
def layer_forward_hook(
# pyre-fixme[2]: Parameter must be annotated.
module,
# pyre-fixme[2]: Parameter must be annotated.
hook_inputs,
# pyre-fixme[2]: Parameter must be annotated.
hook_outputs=None,
# pyre-fixme[2]: Parameter must be annotated.
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)
# the inputs is an empty tuple
# coz it is prepended into additional_forward_args
output = _run_forward(
self.forward_func, (), target_ind, additional_forward_args
)
finally:
for hook in hooks:
if hook is not None:
hook.remove()
# _run_forward may return future of Tensor,
# but we don't support it here now
# And it will fail before here.
output = cast(Tensor, output)
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, **grad_kwargs or {}
)
return grads
return _gradient_func
@overload
def attribute(
self,
inputs: Union[Tensor, Tuple[Tensor, ...]],
baselines: BaselineType,
target: TargetType,
additional_forward_args: Optional[object],
n_steps: int,
method: str,
internal_batch_size: Union[None, int],
return_convergence_delta: Literal[False],
attribute_to_layer_input: bool,
grad_kwargs: Optional[Dict[str, Any]],
) -> Union[Tensor, Tuple[Tensor, ...], List[Union[Tensor, Tuple[Tensor, ...]]]]: ...
@overload
def attribute( # type: ignore
self,
inputs: Union[Tensor, Tuple[Tensor, ...]],
baselines: BaselineType,
target: TargetType,
additional_forward_args: Optional[object],
n_steps: int,
method: str,
internal_batch_size: Union[None, int],
return_convergence_delta: Literal[True],
attribute_to_layer_input: bool,
grad_kwargs: Optional[Dict[str, Any]],
) -> Tuple[
Union[Tensor, Tuple[Tensor, ...], List[Union[Tensor, Tuple[Tensor, ...]]]],
Tensor,
]: ...
@overload
# pyre-fixme[43]: This definition does not have the same decorators as the
# preceding overload(s).
def attribute(
self,
inputs: Union[Tensor, Tuple[Tensor, ...]],
baselines: BaselineType = None,
target: TargetType = None,
additional_forward_args: Optional[object] = 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,
grad_kwargs: Optional[Dict[str, Any]] = None,
) -> Union[
Union[Tensor, Tuple[Tensor, ...], List[Union[Tensor, Tuple[Tensor, ...]]]],
Tuple[
Union[Tensor, Tuple[Tensor, ...], List[Union[Tensor, Tuple[Tensor, ...]]]],
Tensor,
],
]: ...
[docs]
@log_usage()
# pyre-fixme[43]: This definition does not have the same decorators as the
# preceding overload(s).
def attribute(
self,
inputs: Union[Tensor, Tuple[Tensor, ...]],
baselines: BaselineType = None,
target: TargetType = None,
additional_forward_args: Optional[object] = 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,
grad_kwargs: Optional[Dict[str, Any]] = None,
) -> 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[Tensor, ...]): 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 scalar, or Tensor, 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 (str, 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
grad_kwargs (Dict[str, Any], optional): Additional keyword
arguments for torch.autograd.grad.
Default: None
Returns:
**attributions** or 2-element tuple of **attributions**, **delta**:
- **attributions** (*Tensor* or *tuple[Tensor, ...]*):
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
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
)
# pyre-fixme[3]: Return type must be annotated.
# pyre-fixme[2]: Parameter must be annotated.
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
self.ig.gradient_func = self._make_gradient_func(
num_outputs_cumsum, attribute_to_layer_input, grad_kwargs
)
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)
[docs]
def has_convergence_delta(self) -> bool:
return True
@property
def multiplies_by_inputs(self) -> bool:
return self.ig.multiplies_by_inputs
