"""File with GradCAM algorithm explainer classes.
Based on https://github.com/pytorch/captum/blob/master/captum/attr/_core/guided_grad_cam.py
and https://github.com/pytorch/captum/blob/master/captum/attr/_core/layer/grad_cam.py.
"""
from abc import ABC, abstractmethod
from typing import Any, Dict, List, Optional, Union
import torch
import torch.nn.functional as F
from captum._utils.typing import TargetType
from captum.attr import GuidedGradCam
from foxai.array_utils import validate_result
from foxai.explainer.base_explainer import Explainer
from foxai.explainer.computer_vision.model_utils import (
get_last_conv_model_layer,
modify_modules,
)
from foxai.explainer.computer_vision.object_detection.base_object_detector import (
BaseObjectDetector,
)
from foxai.explainer.computer_vision.object_detection.types import ObjectDetectionOutput
[docs]class LayerBaseGradCAM(ABC):
"""Layer GradCAM for object detection task."""
def __init__(
self,
target_layer: torch.nn.Module,
):
self._gradients: Dict[str, torch.Tensor] = {}
self._activations: Dict[str, torch.Tensor] = {}
self.target_layer = target_layer
def backward_hook(
module, # pylint: disable = (unused-argument)
grad_input, # pylint: disable = (unused-argument)
grad_output,
):
self._gradients["value"] = grad_output[0]
def forward_hook(
module, # pylint: disable = (unused-argument)
input, # pylint: disable = (unused-argument,redefined-builtin)
output,
):
self._activations["value"] = output
self.target_layer.register_forward_hook(forward_hook)
self.target_layer.register_backward_hook(backward_hook)
@property
def activations(self) -> torch.Tensor:
return self._activations["value"]
@property
def gradients(self) -> torch.Tensor:
return self._gradients["value"]
[docs] @abstractmethod
def forward(
self,
input_img: torch.Tensor,
) -> Union[torch.Tensor, ObjectDetectionOutput]:
"""Forward pass of GradCAM aglorithm.
Args:
input_img: Input image with shape of (B, C, H, W).
Returns:
ObjectDetectionOutput object for object detection and tensor with saliency
map for classification.
"""
[docs] def get_saliency_map(
self,
height: int,
width: int,
gradients: torch.Tensor,
activations: torch.Tensor,
) -> torch.Tensor:
"""Generate saliency map.
Args:
height: Original image height.
width: Original image width.
gradients: Layer gradients.
activations: Layer activations.
Returns:
Saliency map.
"""
b, k, _, _ = gradients.size()
alpha = gradients.view(b, k, -1).mean(2)
weights = alpha.view(b, k, 1, 1)
saliency_map = (weights * activations).sum(1, keepdim=True)
saliency_map = F.relu(saliency_map)
saliency_map = F.upsample(
saliency_map, size=(height, width), mode="bilinear", align_corners=False
)
saliency_map_min, saliency_map_max = saliency_map.min(), saliency_map.max()
saliency_map = (
(saliency_map - saliency_map_min)
.div(saliency_map_max - saliency_map_min)
.data
)
return saliency_map
def __call__(
self,
input_img: torch.Tensor,
) -> Union[torch.Tensor, ObjectDetectionOutput]:
return self.forward(input_img)
[docs]class LayerGradCAMObjectDetection(LayerBaseGradCAM):
"""Layer GradCAM for object detection task."""
def __init__(
self,
model: torch.nn.Module,
target_layer: torch.nn.Module,
):
super().__init__(target_layer=target_layer)
self.model = model
[docs] def forward(
self,
input_img: torch.Tensor,
) -> torch.Tensor:
"""Forward pass of GradCAM aglorithm.
Args:
input_img: Input image with shape of (B, C, H, W).
Returns:
Tensor with saliency map.
"""
saliency_maps: List[torch.Tensor] = []
_, _, height, width = input_img.size()
result_list = self.model.forward(input_img)
for result in result_list:
score = result.max()
# clear gradients
self.model.zero_grad()
# calculate gradients
score.backward(retain_graph=True)
saliency_maps.append(
self.get_saliency_map(
height=height,
width=width,
gradients=self.gradients,
activations=self.activations,
)
)
return torch.cat(saliency_maps)
[docs]class BaseGradCAMCVExplainer(Explainer):
"""Base GradCAM algorithm explainer."""
[docs] @abstractmethod
def create_explainer(
self, model: torch.nn.Module, layer: torch.nn.Module, **kwargs
) -> Union[GuidedGradCam, LayerBaseGradCAM]:
"""Create explainer object.
Args:
model: The forward function of the model or any
modification of it.
layer: Layer for which attributions are computed.
Output size of attribute matches this layer's input or
output dimensions, depending on whether we attribute to
the inputs or outputs of the layer, corresponding to
attribution of each neuron in the input or output of
this layer.
Returns:
Explainer object.
"""
[docs] @abstractmethod
def calculate_features(
self,
model: torch.nn.Module,
input_data: torch.Tensor,
pred_label_idx: TargetType = None,
additional_forward_args: Any = None,
attribute_to_layer_input: bool = False,
**kwargs,
) -> torch.Tensor:
"""Generate features image with GradCAM algorithm explainer.
Args:
model: The forward function of the model or any
modification of it.
input_data: Input for which attributions
are computed. If forward_func takes a single
tensor as input, a single input tensor should be provided.
pred_label_idx: 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: 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: Indicates whether to
compute the attribution with respect to the layer input
or output in `LayerGradCam`.
If `attribute_to_layer_input` is set to True
then the attributions will be computed with respect to
layer inputs, otherwise it will be computed with respect
to layer outputs.
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:
Element-wise product of (upsampled) GradCAM
and/or Guided Backprop attributions.
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.
Attributions will be the same size as the provided inputs,
with each value providing the attribution of the
corresponding input index.
If the GradCAM attributions cannot be upsampled to the shape
of a given input tensor, None is returned in the corresponding
index position.
"""
[docs]class GuidedGradCAMCVExplainer(BaseGradCAMCVExplainer):
"""GuidedGradCAM algorithm explainer."""
[docs] def create_explainer(
self,
model: torch.nn.Module,
layer: torch.nn.Module,
**kwargs,
) -> Union[GuidedGradCam, LayerBaseGradCAM]:
"""Create explainer object.
Args:
model: The forward function of the model or any
modification of it.
layer: Layer for which attributions are computed.
Output size of attribute matches this layer's input or
output dimensions, depending on whether we attribute to
the inputs or outputs of the layer, corresponding to
attribution of each neuron in the input or output of
this layer.
Returns:
Explainer object.
"""
model = modify_modules(model)
return GuidedGradCam(model=model, layer=layer)
[docs] def calculate_features(
self,
model: torch.nn.Module,
input_data: torch.Tensor,
pred_label_idx: TargetType = None,
additional_forward_args: Any = None,
attribute_to_layer_input: bool = False,
interpolate_mode: str = "nearest",
**kwargs,
) -> torch.Tensor:
"""Generate model's attributes with GradCAM algorithm explainer.
Args:
model: The forward function of the model or any
modification of it.
input_data: Input for which attributions
are computed. If forward_func takes a single
tensor as input, a single input tensor should be provided.
pred_label_idx: 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: 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
interpolate_mode: 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 chosen layer
output (which must also match the number of
dimensions for the input tensor). Note that
the original GradCAM paper uses "bilinear"
interpolation, but we default to "nearest" for
applicability to any of 3D, 4D or 5D tensors.
Default: "nearest"
attribute_to_layer_input: Indicates whether to
compute the attribution with respect to the layer input
or output in `LayerGradCam`.
If `attribute_to_layer_input` is set to True
then the attributions will be computed with respect to
layer inputs, otherwise it will be computed with respect
to layer outputs.
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:
Element-wise product of (upsampled) GradCAM
and/or Guided Backprop attributions.
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.
Attributions will be the same size as the provided inputs,
with each value providing the attribution of the
corresponding input index.
If the GradCAM attributions cannot be upsampled to the shape
of a given input tensor, None is returned in the corresponding
index position.
Raises:
ValueError: if model does not contain conv layers.
RuntimeError: if attributions has shape (0)
"""
layer: Optional[torch.nn.Module] = kwargs.get("layer", None)
if layer is None:
layer = get_last_conv_model_layer(model=model)
guided_cam = self.create_explainer(model=model, layer=layer)
if isinstance(guided_cam, GuidedGradCam):
attributions = guided_cam.attribute(
input_data,
target=pred_label_idx,
additional_forward_args=additional_forward_args,
interpolate_mode=interpolate_mode,
attribute_to_layer_input=attribute_to_layer_input,
)
validate_result(attributions=attributions)
return attributions
[docs]class LayerGradCAMCVExplainer(BaseGradCAMCVExplainer):
"""Layer GradCAM algorithm explainer."""
[docs] def create_explainer(
self,
model: torch.nn.Module,
layer: torch.nn.Module,
**kwargs,
) -> Union[GuidedGradCam, LayerBaseGradCAM]:
"""Create explainer object.
Args:
model: The forward function of the model or any
modification of it.
layer: Layer for which attributions are computed.
Output size of attribute matches this layer's input or
output dimensions, depending on whether we attribute to
the inputs or outputs of the layer, corresponding to
attribution of each neuron in the input or output of
this layer.
Returns:
Explainer object.
"""
model = modify_modules(model)
return LayerGradCAMObjectDetection(model=model, target_layer=layer)
[docs] def calculate_features(
self,
model: torch.nn.Module,
input_data: torch.Tensor,
pred_label_idx: TargetType = None, # pylint: disable = (unused-argument)
additional_forward_args: Any = None, # pylint: disable = (unused-argument)
attribute_to_layer_input: bool = False, # pylint: disable = (unused-argument)
relu_attributions: bool = False, # pylint: disable = (unused-argument)
**kwargs,
) -> torch.Tensor:
"""Generate features image with GradCAM algorithm explainer.
Args:
model: The forward function of the model or any
modification of it.
inputs_data: Input for which 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.
pred_label_idx: 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: 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: Indicates whether to
compute the attribution with respect to the layer input
or output in `LayerGradCam`.
If `attribute_to_layer_input` is set to True
then the attributions will be computed with respect to
layer inputs, otherwise it will be computed with respect
to layer outputs.
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
relu_attributions: Indicates whether to
apply a ReLU operation on the final attribution,
returning only non-negative attributions. Setting this
flag to True matches the original GradCAM algorithm,
otherwise, by default, both positive and negative
attributions are returned.
Default: False
Returns:
Element-wise product of (upsampled) GradCAM
and/or Guided Backprop attributions.
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.
Attributions will be the same size as the provided inputs,
with each value providing the attribution of the
corresponding input index.
If the GradCAM attributions cannot be upsampled to the shape
of a given input tensor, None is returned in the corresponding
index position.
Raises:
ValueError: if model does not contain conv layers.
RuntimeError: if attributions has shape (0)
"""
layer: Optional[torch.nn.Module] = kwargs.get("layer", None)
if layer is None:
layer = get_last_conv_model_layer(model=model)
gradcam = self.create_explainer(model=model, layer=layer)
if isinstance(gradcam, LayerBaseGradCAM):
attributions = gradcam(input_data)
if not isinstance(attributions, torch.Tensor):
raise RuntimeError(
f"Saliency map is `{type(attributions)}`, but expected type is `torch.Tensor`."
)
validate_result(attributions=attributions)
return attributions
[docs]class LayerGradCAMObjectDetectionExplainer(LayerBaseGradCAM):
"""Layer GradCAM for object detection task.
Code based on https://github.com/pooya-mohammadi/yolov5-gradcam.
"""
def __init__(
self,
model: BaseObjectDetector,
target_layer: torch.nn.Module,
):
super().__init__(target_layer=target_layer)
self.model = model
[docs] def forward(
self,
input_img: torch.Tensor,
) -> ObjectDetectionOutput:
"""Forward pass of GradCAM aglorithm.
Args:
input_img: Input image with shape of (B, C, H, W).
Returns:
ObjectDetectionOutput object.
"""
saliency_maps: List[torch.Tensor] = []
_, _, height, width = input_img.size()
predictions, logits = self.model.forward(input_img)
for logit, cls in zip(logits[0], [p.class_number for p in predictions]):
score = logit[cls]
# clear gradients
self.model.zero_grad()
# calculate gradients
score.backward(retain_graph=True)
saliency_maps.append(
self.get_saliency_map(
height=height,
width=width,
gradients=self.gradients,
activations=self.activations,
)
)
return ObjectDetectionOutput(
saliency_maps=saliency_maps,
logits=logits,
predictions=predictions,
)