#!/usr/bin/env python3
# pyre-strict
from typing import Callable, cast, Generic, List, Optional, Tuple, Type, Union
import torch
import torch.nn.functional as F
from captum._utils.common import (
_format_additional_forward_args,
_format_tensor_into_tuples,
_run_forward,
_validate_target,
)
from captum._utils.gradient import compute_gradients
from captum._utils.typing import ModuleOrModuleList, TargetType
from captum.attr._utils.common import (
_format_input_baseline,
_sum_rows,
_tensorize_baseline,
_validate_input,
)
from captum.log import log_usage
from torch import Tensor
from torch.nn import Module
# pyre-fixme[13]: Attribute `attribute` is never initialized.
# pyre-fixme[13]: Attribute `compute_convergence_delta` is never initialized.
[docs]
class Attribution:
r"""
All attribution algorithms extend this class. It enforces its child classes
to extend and override core `attribute` method.
"""
# pyre-fixme[24]: Generic type `Callable` expects 2 type parameters.
def __init__(self, forward_func: Callable) -> None:
r"""
Args:
forward_func (Callable or torch.nn.Module): This can either be an instance
of pytorch model or any modification of model's forward
function.
"""
self.forward_func = forward_func
# pyre-fixme[24]: Generic type `Callable` expects 2 type parameters.
# pyre-fixme[13]: Attribute `attribute` is never initialized.
attribute: Callable
r"""
This method computes and returns the attribution values for each input tensor.
Deriving classes are responsible for implementing its logic accordingly.
Specific attribution algorithms that extend this class take relevant
arguments.
Args:
inputs (Tensor or tuple[Tensor, ...]): Input for which attribution
is computed. It can be provided as a single tensor or
a tuple of multiple tensors. If multiple input tensors
are provided, the batch sizes must be aligned across all
tensors.
Returns:
*Tensor* or *tuple[Tensor, ...]* of **attributions**:
- **attributions** (*Tensor* or *tuple[Tensor, ...]*):
Attribution values for each
input tensor. The `attributions` have the same shape and
dimensionality as the inputs.
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.
"""
# pyre-fixme[24] Generic type `Callable` expects 2 type parameters.
# pyre-fixme[13]: Attribute `attribute_future` is never initialized.
attribute_future: Callable
r"""
This method computes and returns a Future of attribution values for each input
tensor. Deriving classes are responsible for implementing its logic accordingly.
Specific attribution algorithms that extend this class take relevant
arguments.
Args:
inputs (Tensor or tuple[Tensor, ...]): Input for which attribution
is computed. It can be provided as a single tensor or
a tuple of multiple tensors. If multiple input tensors
are provided, the batch sizes must be aligned across all
tensors.
Returns:
*Future[Tensor]* or *Future[tuple[Tensor, ...]]* of **attributions**:
- **attributions** (*Future[Tensor]* or *Future[tuple[Tensor, ...]]*):
Future of attribution values for each input tensor.
The results should be the same as the attribute
method, except that the results are returned as a Future.
If a single tensor is provided as inputs, a single Future tensor
is returned. If a tuple is provided for inputs, a Future of a
tuple of corresponding sized tensors is returned.
"""
@property
def multiplies_by_inputs(self) -> bool:
return False
[docs]
def has_convergence_delta(self) -> bool:
r"""
This method informs the user whether the attribution algorithm provides
a convergence delta (aka an approximation error) or not. Convergence
delta may serve as a proxy of correctness of attribution algorithm's
approximation. If deriving attribution class provides a
`compute_convergence_delta` method, it should
override both `compute_convergence_delta` and `has_convergence_delta` methods.
Returns:
bool:
Returns whether the attribution algorithm
provides a convergence delta (aka approximation error) or not.
"""
return False
# pyre-fixme[24]: Generic type `Callable` expects 2 type parameters.
# pyre-fixme[13]: Attribute `compute_convergence_delta` is never initialized.
compute_convergence_delta: Callable
r"""
The attribution algorithms which derive `Attribution` class and provide
convergence delta (aka approximation error) should implement this method.
Convergence delta can be computed based on certain properties of the
attribution alogrithms.
Args:
attributions (Tensor or tuple[Tensor, ...]): Attribution scores that
are precomputed by an attribution algorithm.
Attributions can be provided in form of a single tensor
or a tuple of those. It is assumed that attribution
tensor's dimension 0 corresponds to the number of
examples, and if multiple input tensors are provided,
the examples must be aligned appropriately.
*args (Any, optional): Additonal arguments that are used by the
sub-classes depending on the specific implementation
of `compute_convergence_delta`.
Returns:
*Tensor* of **deltas**:
- **deltas** (*Tensor*):
Depending on specific implementaion of
sub-classes, convergence delta can be returned per
sample in form of a tensor or it can be aggregated
across multuple samples and returned in form of a
single floating point tensor.
"""
[docs]
@classmethod
def get_name(cls: Type["Attribution"]) -> str:
r"""
Create readable class name by inserting a space before any capital
characters besides the very first.
Returns:
str: a readable class name
Example:
for a class called IntegratedGradients, we return the string
'Integrated Gradients'
"""
return "".join(
[
char if char.islower() or idx == 0 else " " + char
for idx, char in enumerate(cls.__name__)
]
)
[docs]
class GradientAttribution(Attribution):
r"""
All gradient based attribution algorithms extend this class. It requires a
forward function, which most commonly is the forward function of the model
that we want to interpret or the model itself.
"""
# pyre-fixme[24]: Generic type `Callable` expects 2 type parameters.
def __init__(self, forward_func: Callable) -> None:
r"""
Args:
forward_func (Callable or torch.nn.Module): This can either be an instance
of pytorch model or any modification of model's forward
function.
"""
Attribution.__init__(self, forward_func)
# pyre-fixme[4]: Attribute must be annotated.
self.gradient_func = compute_gradients
[docs]
@log_usage()
def compute_convergence_delta(
self,
attributions: Union[Tensor, Tuple[Tensor, ...]],
start_point: Union[
None, int, float, Tensor, Tuple[Union[int, float, Tensor], ...]
],
end_point: Union[Tensor, Tuple[Tensor, ...]],
target: TargetType = None,
additional_forward_args: Optional[object] = None,
) -> Tensor:
r"""
Here we provide a specific implementation for `compute_convergence_delta`
which is based on a common property among gradient-based attribution algorithms.
In the literature sometimes it is also called completeness axiom. Completeness
axiom states that the sum of the attribution must be equal to the differences of
NN Models's function at its end and start points. In other words:
sum(attributions) - (F(end_point) - F(start_point)) is close to zero.
Returned delta of this method is defined as above stated difference.
This implementation assumes that both the `start_point` and `end_point` have
the same shape and dimensionality. It also assumes that the target must have
the same number of examples as the `start_point` and the `end_point` in case
it is provided in form of a list or a non-singleton tensor.
Args:
attributions (Tensor or tuple[Tensor, ...]): Precomputed attribution
scores. The user can compute those using any attribution
algorithm. It is assumed the shape and the
dimensionality of attributions must match the shape and
the dimensionality of `start_point` and `end_point`.
It also assumes that the attribution tensor's
dimension 0 corresponds to the number of
examples, and if multiple input tensors are provided,
the examples must be aligned appropriately.
start_point (Tensor or tuple[Tensor, ...], optional): `start_point`
is passed as an input to model's forward function. It
is the starting point of attributions' approximation.
It is assumed that both `start_point` and `end_point`
have the same shape and dimensionality.
end_point (Tensor or tuple[Tensor, ...]): `end_point`
is passed as an input to model's forward function. It
is the end point of attributions' approximation.
It is assumed that both `start_point` and `end_point`
have the same shape and dimensionality.
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.
`additional_forward_args` is used both for `start_point`
and `end_point` when computing the forward pass.
Default: None
Returns:
*Tensor* of **deltas**:
- **deltas** (*Tensor*):
This implementation returns convergence delta per
sample. Deriving sub-classes may do any type of aggregation
of those values, if necessary.
"""
end_point, start_point = _format_input_baseline(end_point, start_point)
additional_forward_args = _format_additional_forward_args(
additional_forward_args
)
# tensorizing start_point in case it is a scalar or one example baseline
# If the batch size is large we could potentially also tensorize only one
# sample and expand the output to the rest of the elements in the batch
start_point = _tensorize_baseline(end_point, start_point)
attributions = _format_tensor_into_tuples(attributions)
# verify that the attributions and end_point match on 1st dimension
for attribution, end_point_tnsr in zip(attributions, end_point):
assert end_point_tnsr.shape[0] == attribution.shape[0], (
"Attributions tensor and the end_point must match on the first"
" dimension but found attribution: {} and end_point: {}".format(
attribution.shape[0], end_point_tnsr.shape[0]
)
)
num_samples = end_point[0].shape[0]
_validate_input(end_point, start_point)
_validate_target(num_samples, target)
with torch.no_grad():
start_out_eval = _run_forward(
self.forward_func, start_point, target, additional_forward_args
)
# _run_forward may return future of Tensor,
# but we don't support it here now
# And it will fail before here.
start_out_sum = _sum_rows(cast(Tensor, start_out_eval))
end_out_eval = _run_forward(
self.forward_func, end_point, target, additional_forward_args
)
# _run_forward may return future of Tensor,
# but we don't support it here now
# And it will fail before here.
end_out_sum = _sum_rows(cast(Tensor, end_out_eval))
row_sums = [_sum_rows(attribution) for attribution in attributions]
attr_sum = torch.stack(
[cast(Tensor, sum(row_sum)) for row_sum in zip(*row_sums)]
)
_delta = attr_sum - (end_out_sum - start_out_sum)
return _delta
[docs]
class PerturbationAttribution(Attribution):
r"""
All perturbation based attribution algorithms extend this class. It requires a
forward function, which most commonly is the forward function of the model
that we want to interpret or the model itself.
"""
# pyre-fixme[24]: Generic type `Callable` expects 2 type parameters.
def __init__(self, forward_func: Callable) -> None:
r"""
Args:
forward_func (Callable or torch.nn.Module): This can either be an instance
of pytorch model or any modification of model's forward
function.
"""
Attribution.__init__(self, forward_func)
@property
def multiplies_by_inputs(self) -> bool:
return True
# mypy false positive "Free type variable expected in Generic[...]" but
# ModuleOrModuleList is a TypeVar
class InternalAttribution(Attribution, Generic[ModuleOrModuleList]): # type: ignore
r"""
Shared base class for LayerAttrubution and NeuronAttribution,
attribution types that require a model and a particular layer.
"""
layer: ModuleOrModuleList
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,
) -> None:
r"""
Args:
forward_func (Callable or torch.nn.Module): This can either be an instance
of pytorch model or any modification of model's forward
function.
layer (torch.nn.Module): Layer for which output attributions are computed.
Output size of attribute matches that of layer output.
device_ids (list[int]): Device ID list, necessary only if forward_func
applies a DataParallel model, which 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.
"""
Attribution.__init__(self, forward_func)
self.layer = layer
self.device_ids = device_ids
# pyre-fixme[24]: Generic type `InternalAttribution` expects 1 type parameter.
[docs]
class LayerAttribution(InternalAttribution):
r"""
Layer attribution provides attribution values for the given layer, quantifying
the importance of each neuron within the given layer's output. The output
attribution of calling attribute on a LayerAttribution object always matches
the size of the layer output.
"""
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,
) -> None:
r"""
Args:
forward_func (Callable or torch.nn.Module): This can either be an instance
of pytorch model or any modification of model's forward
function.
layer (torch.nn.Module): Layer for which output attributions are computed.
Output size of attribute matches that of layer output.
device_ids (list[int]): Device ID list, necessary only if forward_func
applies a DataParallel model, which 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.
"""
InternalAttribution.__init__(self, forward_func, layer, device_ids)
[docs]
@staticmethod
def interpolate(
layer_attribution: Tensor,
interpolate_dims: Union[int, Tuple[int, ...]],
interpolate_mode: str = "nearest",
) -> Tensor:
r"""
Interpolates given 3D, 4D or 5D layer attribution to given dimensions.
This is often utilized to upsample the attribution of a convolutional layer
to the size of an input, which allows visualizing in the input space.
Args:
layer_attribution (Tensor): Tensor of given layer attributions.
interpolate_dims (int or tuple): Upsampled dimensions. The
number of elements must be the number of dimensions
of layer_attribution - 2, since the first dimension
corresponds to number of examples and the second is
assumed to correspond to the number of channels.
interpolate_mode (str): Method for interpolation, which
must be a valid input interpolation mode for
torch.nn.functional. These methods are
"nearest", "area", "linear" (3D-only), "bilinear"
(4D-only), "bicubic" (4D-only), "trilinear" (5D-only)
based on the number of dimensions of the given layer
attribution.
Returns:
*Tensor* of upsampled **attributions**:
- **attributions** (*Tensor*):
Upsampled layer attributions with first 2 dimensions matching
slayer_attribution and remaining dimensions given by
interpolate_dims.
"""
return F.interpolate(layer_attribution, interpolate_dims, mode=interpolate_mode)
# pyre-fixme[13]: Attribute `attribute` is never initialized.
# pyre-fixme[24]: Generic type `InternalAttribution` expects 1 type parameter.
[docs]
class NeuronAttribution(InternalAttribution):
r"""
Neuron attribution provides input attribution for a given neuron, quantifying
the importance of each input feature in the activation of a particular neuron.
Calling attribute on a NeuronAttribution object requires also providing
the index of the neuron in the output of the given layer for which attributions
are required.
The output attribution of calling attribute on a NeuronAttribution object
always matches the size of the input.
"""
def __init__(
self,
# pyre-fixme[24]: Generic type `Callable` expects 2 type parameters.
forward_func: Callable,
layer: Module,
device_ids: Union[None, List[int]] = None,
) -> None:
r"""
Args:
forward_func (Callable or torch.nn.Module): This can either be an instance
of pytorch model or any modification of model's forward
function.
layer (torch.nn.Module): Layer for which output attributions are computed.
Output size of attribute matches that of layer output.
device_ids (list[int]): Device ID list, necessary only if forward_func
applies a DataParallel model, which 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.
"""
InternalAttribution.__init__(self, forward_func, layer, device_ids)
# pyre-fixme[24]: Generic type `Callable` expects 2 type parameters.
# pyre-fixme[13]: Attribute `attribute` is never initialized.
attribute: Callable
r"""
This method computes and returns the neuron attribution values for each
input tensor. Deriving classes are responsible for implementing
its logic accordingly.
Specific attribution algorithms that extend this class take relevant
arguments.
Args:
inputs: A single high dimensional input tensor or a tuple of them.
neuron_selector (int or tuple): Tuple providing index of neuron in output
of given layer for which attribution is desired. Length of
this tuple must be one less than the number of
dimensions in the output of the given layer (since
dimension 0 corresponds to number of examples).
Returns:
*Tensor* or *tuple[Tensor, ...]* of **attributions**:
- **attributions** (*Tensor* or *tuple[Tensor, ...]*):
Attribution values for
each input vector. The `attributions` have the
dimensionality of inputs.
"""
