Add custom grad cam

app.py

@@ -19,7 +19,8 @@ from gradio_blocks import build_video_to_camvideo
 from Nets import CustomResNet18
 from PIL import Image, ImageDraw, ImageFont
 
-from pytorch_grad_cam import GradCAM, HiResCAM, GradCAMPlusPlus, AblationCAM, XGradCAM, EigenCAM, FullGrad
+# from pytorch_grad_cam import GradCAM, HiResCAM, GradCAMPlusPlus, AblationCAM, XGradCAM, EigenCAM, FullGrad
+from custom_code.custom_grad_cam import GradCAM, HiResCAM, GradCAMPlusPlus, AblationCAM, XGradCAM, EigenCAM, FullGrad
 from pytorch_grad_cam.utils.model_targets import ClassifierOutputTarget
 from pytorch_grad_cam.utils.image import show_cam_on_image
 

src/custom_code/custom_grad_cam/__init__.py

@@ -0,0 +1,20 @@
+from pytorch_grad_cam.grad_cam import GradCAM
+from pytorch_grad_cam.hirescam import HiResCAM
+from pytorch_grad_cam.grad_cam_elementwise import GradCAMElementWise
+from pytorch_grad_cam.ablation_layer import AblationLayer, AblationLayerVit, AblationLayerFasterRCNN
+from pytorch_grad_cam.ablation_cam import AblationCAM
+from pytorch_grad_cam.xgrad_cam import XGradCAM
+from pytorch_grad_cam.grad_cam_plusplus import GradCAMPlusPlus
+from pytorch_grad_cam.score_cam import ScoreCAM
+from pytorch_grad_cam.layer_cam import LayerCAM
+from pytorch_grad_cam.eigen_cam import EigenCAM
+from pytorch_grad_cam.eigen_grad_cam import EigenGradCAM
+from pytorch_grad_cam.random_cam import RandomCAM
+from pytorch_grad_cam.fullgrad_cam import FullGrad
+from pytorch_grad_cam.guided_backprop import GuidedBackpropReLUModel
+from pytorch_grad_cam.activations_and_gradients import ActivationsAndGradients
+from pytorch_grad_cam.feature_factorization.deep_feature_factorization import DeepFeatureFactorization, run_dff_on_image
+import pytorch_grad_cam.utils.model_targets
+import pytorch_grad_cam.utils.reshape_transforms
+import pytorch_grad_cam.metrics.cam_mult_image
+import pytorch_grad_cam.metrics.road

src/custom_code/custom_grad_cam/ablation_cam.py

@@ -0,0 +1,148 @@
+import numpy as np
+import torch
+import tqdm
+from typing import Callable, List
+from pytorch_grad_cam.base_cam import BaseCAM
+from pytorch_grad_cam.utils.find_layers import replace_layer_recursive
+from pytorch_grad_cam.ablation_layer import AblationLayer
+
+
+""" Implementation of AblationCAM
+https://openaccess.thecvf.com/content_WACV_2020/papers/Desai_Ablation-CAM_Visual_Explanations_for_Deep_Convolutional_Network_via_Gradient-free_Localization_WACV_2020_paper.pdf
+
+Ablate individual activations, and then measure the drop in the target score.
+
+In the current implementation, the target layer activations is cached, so it won't be re-computed.
+However layers before it, if any, will not be cached.
+This means that if the target layer is a large block, for example model.featuers (in vgg), there will
+be a large save in run time.
+
+Since we have to go over many channels and ablate them, and every channel ablation requires a forward pass,
+it would be nice if we could avoid doing that for channels that won't contribute anwyay, making it much faster.
+The parameter ratio_channels_to_ablate controls how many channels should be ablated, using an experimental method
+(to be improved). The default 1.0 value means that all channels will be ablated.
+"""
+
+
+class AblationCAM(BaseCAM):
+    def __init__(self,
+                 model: torch.nn.Module,
+                 target_layers: List[torch.nn.Module],
+                 use_cuda: bool = False,
+                 reshape_transform: Callable = None,
+                 ablation_layer: torch.nn.Module = AblationLayer(),
+                 batch_size: int = 32,
+                 ratio_channels_to_ablate: float = 1.0) -> None:
+
+        super(AblationCAM, self).__init__(model,
+                                          target_layers,
+                                          use_cuda,
+                                          reshape_transform,
+                                          uses_gradients=False)
+        self.batch_size = batch_size
+        self.ablation_layer = ablation_layer
+        self.ratio_channels_to_ablate = ratio_channels_to_ablate
+
+    def save_activation(self, module, input, output) -> None:
+        """ Helper function to save the raw activations from the target layer """
+        self.activations = output
+
+    def assemble_ablation_scores(self,
+                                 new_scores: list,
+                                 original_score: float,
+                                 ablated_channels: np.ndarray,
+                                 number_of_channels: int) -> np.ndarray:
+        """ Take the value from the channels that were ablated,
+            and just set the original score for the channels that were skipped """
+
+        index = 0
+        result = []
+        sorted_indices = np.argsort(ablated_channels)
+        ablated_channels = ablated_channels[sorted_indices]
+        new_scores = np.float32(new_scores)[sorted_indices]
+
+        for i in range(number_of_channels):
+            if index < len(ablated_channels) and ablated_channels[index] == i:
+                weight = new_scores[index]
+                index = index + 1
+            else:
+                weight = original_score
+            result.append(weight)
+
+        return result
+
+    def get_cam_weights(self,
+                        input_tensor: torch.Tensor,
+                        target_layer: torch.nn.Module,
+                        targets: List[Callable],
+                        activations: torch.Tensor,
+                        grads: torch.Tensor) -> np.ndarray:
+
+        # Do a forward pass, compute the target scores, and cache the
+        # activations
+        handle = target_layer.register_forward_hook(self.save_activation)
+        with torch.no_grad():
+            outputs = self.model(input_tensor)
+            handle.remove()
+            original_scores = np.float32(
+                [target(output).cpu().item() for target, output in zip(targets, outputs)])
+
+        # Replace the layer with the ablation layer.
+        # When we finish, we will replace it back, so the original model is
+        # unchanged.
+        ablation_layer = self.ablation_layer
+        replace_layer_recursive(self.model, target_layer, ablation_layer)
+
+        number_of_channels = activations.shape[1]
+        weights = []
+        # This is a "gradient free" method, so we don't need gradients here.
+        with torch.no_grad():
+            # Loop over each of the batch images and ablate activations for it.
+            for batch_index, (target, tensor) in enumerate(
+                    zip(targets, input_tensor)):
+                new_scores = []
+                batch_tensor = tensor.repeat(self.batch_size, 1, 1, 1)
+
+                # Check which channels should be ablated. Normally this will be all channels,
+                # But we can also try to speed this up by using a low
+                # ratio_channels_to_ablate.
+                channels_to_ablate = ablation_layer.activations_to_be_ablated(
+                    activations[batch_index, :], self.ratio_channels_to_ablate)
+                number_channels_to_ablate = len(channels_to_ablate)
+
+                for i in tqdm.tqdm(
+                    range(
+                        0,
+                        number_channels_to_ablate,
+                        self.batch_size)):
+                    if i + self.batch_size > number_channels_to_ablate:
+                        batch_tensor = batch_tensor[:(
+                            number_channels_to_ablate - i)]
+
+                    # Change the state of the ablation layer so it ablates the next channels.
+                    # TBD: Move this into the ablation layer forward pass.
+                    ablation_layer.set_next_batch(
+                        input_batch_index=batch_index,
+                        activations=self.activations,
+                        num_channels_to_ablate=batch_tensor.size(0))
+                    score = [target(o).cpu().item()
+                             for o in self.model(batch_tensor)]
+                    new_scores.extend(score)
+                    ablation_layer.indices = ablation_layer.indices[batch_tensor.size(
+                        0):]
+
+                new_scores = self.assemble_ablation_scores(
+                    new_scores,
+                    original_scores[batch_index],
+                    channels_to_ablate,
+                    number_of_channels)
+                weights.extend(new_scores)
+
+        weights = np.float32(weights)
+        weights = weights.reshape(activations.shape[:2])
+        original_scores = original_scores[:, None]
+        weights = (original_scores - weights) / original_scores
+
+        # Replace the model back to the original state
+        replace_layer_recursive(self.model, ablation_layer, target_layer)
+        return weights

src/custom_code/custom_grad_cam/ablation_cam_multilayer.py

@@ -0,0 +1,136 @@
+import cv2
+import numpy as np
+import torch
+import tqdm
+from pytorch_grad_cam.base_cam import BaseCAM
+
+
+class AblationLayer(torch.nn.Module):
+    def __init__(self, layer, reshape_transform, indices):
+        super(AblationLayer, self).__init__()
+
+        self.layer = layer
+        self.reshape_transform = reshape_transform
+        # The channels to zero out:
+        self.indices = indices
+
+    def forward(self, x):
+        self.__call__(x)
+
+    def __call__(self, x):
+        output = self.layer(x)
+
+        # Hack to work with ViT,
+        # Since the activation channels are last and not first like in CNNs
+        # Probably should remove it?
+        if self.reshape_transform is not None:
+            output = output.transpose(1, 2)
+
+        for i in range(output.size(0)):
+
+            # Commonly the minimum activation will be 0,
+            # And then it makes sense to zero it out.
+            # However depending on the architecture,
+            # If the values can be negative, we use very negative values
+            # to perform the ablation, deviating from the paper.
+            if torch.min(output) == 0:
+                output[i, self.indices[i], :] = 0
+            else:
+                ABLATION_VALUE = 1e5
+                output[i, self.indices[i], :] = torch.min(
+                    output) - ABLATION_VALUE
+
+        if self.reshape_transform is not None:
+            output = output.transpose(2, 1)
+
+        return output
+
+
+def replace_layer_recursive(model, old_layer, new_layer):
+    for name, layer in model._modules.items():
+        if layer == old_layer:
+            model._modules[name] = new_layer
+            return True
+        elif replace_layer_recursive(layer, old_layer, new_layer):
+            return True
+    return False
+
+
+class AblationCAM(BaseCAM):
+    def __init__(self, model, target_layers, use_cuda=False,
+                 reshape_transform=None):
+        super(AblationCAM, self).__init__(model, target_layers, use_cuda,
+                                          reshape_transform)
+
+        if len(target_layers) > 1:
+            print(
+                "Warning. You are usign Ablation CAM with more than 1 layers. "
+                "This is supported only if all layers have the same output shape")
+
+    def set_ablation_layers(self):
+        self.ablation_layers = []
+        for target_layer in self.target_layers:
+            ablation_layer = AblationLayer(target_layer,
+                                           self.reshape_transform, indices=[])
+            self.ablation_layers.append(ablation_layer)
+            replace_layer_recursive(self.model, target_layer, ablation_layer)
+
+    def unset_ablation_layers(self):
+        # replace the model back to the original state
+        for ablation_layer, target_layer in zip(
+                self.ablation_layers, self.target_layers):
+            replace_layer_recursive(self.model, ablation_layer, target_layer)
+
+    def set_ablation_layer_batch_indices(self, indices):
+        for ablation_layer in self.ablation_layers:
+            ablation_layer.indices = indices
+
+    def trim_ablation_layer_batch_indices(self, keep):
+        for ablation_layer in self.ablation_layers:
+            ablation_layer.indices = ablation_layer.indices[:keep]
+
+    def get_cam_weights(self,
+                        input_tensor,
+                        target_category,
+                        activations,
+                        grads):
+        with torch.no_grad():
+            outputs = self.model(input_tensor).cpu().numpy()
+            original_scores = []
+            for i in range(input_tensor.size(0)):
+                original_scores.append(outputs[i, target_category[i]])
+        original_scores = np.float32(original_scores)
+
+        self.set_ablation_layers()
+
+        if hasattr(self, "batch_size"):
+            BATCH_SIZE = self.batch_size
+        else:
+            BATCH_SIZE = 32
+
+        number_of_channels = activations.shape[1]
+        weights = []
+
+        with torch.no_grad():
+            # Iterate over the input batch
+            for tensor, category in zip(input_tensor, target_category):
+                batch_tensor = tensor.repeat(BATCH_SIZE, 1, 1, 1)
+                for i in tqdm.tqdm(range(0, number_of_channels, BATCH_SIZE)):
+                    self.set_ablation_layer_batch_indices(
+                        list(range(i, i + BATCH_SIZE)))
+
+                    if i + BATCH_SIZE > number_of_channels:
+                        keep = number_of_channels - i
+                        batch_tensor = batch_tensor[:keep]
+                        self.trim_ablation_layer_batch_indices(self, keep)
+                    score = self.model(batch_tensor)[:, category].cpu().numpy()
+                    weights.extend(score)
+
+        weights = np.float32(weights)
+        weights = weights.reshape(activations.shape[:2])
+        original_scores = original_scores[:, None]
+        weights = (original_scores - weights) / original_scores
+
+        # replace the model back to the original state
+        self.unset_ablation_layers()
+        return weights

src/custom_code/custom_grad_cam/ablation_layer.py

@@ -0,0 +1,155 @@
+import torch
+from collections import OrderedDict
+import numpy as np
+from pytorch_grad_cam.utils.svd_on_activations import get_2d_projection
+
+
+class AblationLayer(torch.nn.Module):
+    def __init__(self):
+        super(AblationLayer, self).__init__()
+
+    def objectiveness_mask_from_svd(self, activations, threshold=0.01):
+        """ Experimental method to get a binary mask to compare if the activation is worth ablating.
+            The idea is to apply the EigenCAM method by doing PCA on the activations.
+            Then we create a binary mask by comparing to a low threshold.
+            Areas that are masked out, are probably not interesting anyway.
+        """
+
+        projection = get_2d_projection(activations[None, :])[0, :]
+        projection = np.abs(projection)
+        projection = projection - projection.min()
+        projection = projection / projection.max()
+        projection = projection > threshold
+        return projection
+
+    def activations_to_be_ablated(
+            self,
+            activations,
+            ratio_channels_to_ablate=1.0):
+        """ Experimental method to get a binary mask to compare if the activation is worth ablating.
+            Create a binary CAM mask with objectiveness_mask_from_svd.
+            Score each Activation channel, by seeing how much of its values are inside the mask.
+            Then keep the top channels.
+
+        """
+        if ratio_channels_to_ablate == 1.0:
+            self.indices = np.int32(range(activations.shape[0]))
+            return self.indices
+
+        projection = self.objectiveness_mask_from_svd(activations)
+
+        scores = []
+        for channel in activations:
+            normalized = np.abs(channel)
+            normalized = normalized - normalized.min()
+            normalized = normalized / np.max(normalized)
+            score = (projection * normalized).sum() / normalized.sum()
+            scores.append(score)
+        scores = np.float32(scores)
+
+        indices = list(np.argsort(scores))
+        high_score_indices = indices[::-
+                                     1][: int(len(indices) *
+                                              ratio_channels_to_ablate)]
+        low_score_indices = indices[: int(
+            len(indices) * ratio_channels_to_ablate)]
+        self.indices = np.int32(high_score_indices + low_score_indices)
+        return self.indices
+
+    def set_next_batch(
+            self,
+            input_batch_index,
+            activations,
+            num_channels_to_ablate):
+        """ This creates the next batch of activations from the layer.
+            Just take corresponding batch member from activations, and repeat it num_channels_to_ablate times.
+        """
+        self.activations = activations[input_batch_index, :, :, :].clone(
+        ).unsqueeze(0).repeat(num_channels_to_ablate, 1, 1, 1)
+
+    def __call__(self, x):
+        output = self.activations
+        for i in range(output.size(0)):
+            # Commonly the minimum activation will be 0,
+            # And then it makes sense to zero it out.
+            # However depending on the architecture,
+            # If the values can be negative, we use very negative values
+            # to perform the ablation, deviating from the paper.
+            if torch.min(output) == 0:
+                output[i, self.indices[i], :] = 0
+            else:
+                ABLATION_VALUE = 1e7
+                output[i, self.indices[i], :] = torch.min(
+                    output) - ABLATION_VALUE
+
+        return output
+
+
+class AblationLayerVit(AblationLayer):
+    def __init__(self):
+        super(AblationLayerVit, self).__init__()
+
+    def __call__(self, x):
+        output = self.activations
+        output = output.transpose(1, len(output.shape) - 1)
+        for i in range(output.size(0)):
+
+            # Commonly the minimum activation will be 0,
+            # And then it makes sense to zero it out.
+            # However depending on the architecture,
+            # If the values can be negative, we use very negative values
+            # to perform the ablation, deviating from the paper.
+            if torch.min(output) == 0:
+                output[i, self.indices[i], :] = 0
+            else:
+                ABLATION_VALUE = 1e7
+                output[i, self.indices[i], :] = torch.min(
+                    output) - ABLATION_VALUE
+
+        output = output.transpose(len(output.shape) - 1, 1)
+
+        return output
+
+    def set_next_batch(
+            self,
+            input_batch_index,
+            activations,
+            num_channels_to_ablate):
+        """ This creates the next batch of activations from the layer.
+            Just take corresponding batch member from activations, and repeat it num_channels_to_ablate times.
+        """
+        repeat_params = [num_channels_to_ablate] + \
+            len(activations.shape[:-1]) * [1]
+        self.activations = activations[input_batch_index, :, :].clone(
+        ).unsqueeze(0).repeat(*repeat_params)
+
+
+class AblationLayerFasterRCNN(AblationLayer):
+    def __init__(self):
+        super(AblationLayerFasterRCNN, self).__init__()
+
+    def set_next_batch(
+            self,
+            input_batch_index,
+            activations,
+            num_channels_to_ablate):
+        """ Extract the next batch member from activations,
+            and repeat it num_channels_to_ablate times.
+        """
+        self.activations = OrderedDict()
+        for key, value in activations.items():
+            fpn_activation = value[input_batch_index,
+                                   :, :, :].clone().unsqueeze(0)
+            self.activations[key] = fpn_activation.repeat(
+                num_channels_to_ablate, 1, 1, 1)
+
+    def __call__(self, x):
+        result = self.activations
+        layers = {0: '0', 1: '1', 2: '2', 3: '3', 4: 'pool'}
+        num_channels_to_ablate = result['pool'].size(0)
+        for i in range(num_channels_to_ablate):
+            pyramid_layer = int(self.indices[i] / 256)
+            index_in_pyramid_layer = int(self.indices[i] % 256)
+            result[layers[pyramid_layer]][i,
+                                          index_in_pyramid_layer, :, :] = -1000
+        return result

src/custom_code/custom_grad_cam/activations_and_gradients.py

@@ -0,0 +1,46 @@
+class ActivationsAndGradients:
+    """ Class for extracting activations and
+    registering gradients from targetted intermediate layers """
+
+    def __init__(self, model, target_layers, reshape_transform):
+        self.model = model
+        self.gradients = []
+        self.activations = []
+        self.reshape_transform = reshape_transform
+        self.handles = []
+        for target_layer in target_layers:
+            self.handles.append(
+                target_layer.register_forward_hook(self.save_activation))
+            # Because of https://github.com/pytorch/pytorch/issues/61519,
+            # we don't use backward hook to record gradients.
+            self.handles.append(
+                target_layer.register_forward_hook(self.save_gradient))
+
+    def save_activation(self, module, input, output):
+        activation = output
+
+        if self.reshape_transform is not None:
+            activation = self.reshape_transform(activation)
+        self.activations.append(activation.cpu().detach())
+
+    def save_gradient(self, module, input, output):
+        if not hasattr(output, "requires_grad") or not output.requires_grad:
+            # You can only register hooks on tensor requires grad.
+            return
+
+        # Gradients are computed in reverse order
+        def _store_grad(grad):
+            if self.reshape_transform is not None:
+                grad = self.reshape_transform(grad)
+            self.gradients = [grad.cpu().detach()] + self.gradients
+
+        output.register_hook(_store_grad)
+
+    def __call__(self, x):
+        self.gradients = []
+        self.activations = []
+        return self.model(x)
+
+    def release(self):
+        for handle in self.handles:
+            handle.remove()

src/custom_code/custom_grad_cam/base_cam.py

@@ -0,0 +1,205 @@
+import numpy as np
+import torch
+import ttach as tta
+from typing import Callable, List, Tuple
+from pytorch_grad_cam.activations_and_gradients import ActivationsAndGradients
+from pytorch_grad_cam.utils.svd_on_activations import get_2d_projection
+from pytorch_grad_cam.utils.image import scale_cam_image
+from pytorch_grad_cam.utils.model_targets import ClassifierOutputTarget
+
+
+class BaseCAM:
+    def __init__(self,
+                 model: torch.nn.Module,
+                 target_layers: List[torch.nn.Module],
+                 use_cuda: bool = False,
+                 reshape_transform: Callable = None,
+                 compute_input_gradient: bool = False,
+                 uses_gradients: bool = True) -> None:
+        self.model = model.eval()
+        self.target_layers = target_layers
+        self.cuda = use_cuda
+        if self.cuda:
+            self.model = model.cuda()
+        self.reshape_transform = reshape_transform
+        self.compute_input_gradient = compute_input_gradient
+        self.uses_gradients = uses_gradients
+        self.activations_and_grads = ActivationsAndGradients(
+            self.model, target_layers, reshape_transform)
+        self.outputs = None
+
+    """ Get a vector of weights for every channel in the target layer.
+        Methods that return weights channels,
+        will typically need to only implement this function. """
+
+    def get_cam_weights(self,
+                        input_tensor: torch.Tensor,
+                        target_layers: List[torch.nn.Module],
+                        targets: List[torch.nn.Module],
+                        activations: torch.Tensor,
+                        grads: torch.Tensor) -> np.ndarray:
+        raise Exception("Not Implemented")
+
+    def get_cam_image(self,
+                      input_tensor: torch.Tensor,
+                      target_layer: torch.nn.Module,
+                      targets: List[torch.nn.Module],
+                      activations: torch.Tensor,
+                      grads: torch.Tensor,
+                      eigen_smooth: bool = False) -> np.ndarray:
+
+        weights = self.get_cam_weights(input_tensor,
+                                       target_layer,
+                                       targets,
+                                       activations,
+                                       grads)
+        weighted_activations = weights[:, :, None, None] * activations
+        if eigen_smooth:
+            cam = get_2d_projection(weighted_activations)
+        else:
+            cam = weighted_activations.sum(axis=1)
+        return cam
+
+    def forward(self,
+                input_tensor: torch.Tensor,
+                targets: List[torch.nn.Module],
+                eigen_smooth: bool = False) -> np.ndarray:
+
+        if self.cuda:
+            input_tensor = input_tensor.cuda()
+
+        if self.compute_input_gradient:
+            input_tensor = torch.autograd.Variable(input_tensor,
+                                                   requires_grad=True)
+
+        outputs = self.activations_and_grads(input_tensor)
+        self.outputs = outputs
+        if targets is None:
+            target_categories = np.argmax(outputs.cpu().data.numpy(), axis=-1)
+            targets = [ClassifierOutputTarget(
+                category) for category in target_categories]
+
+        if self.uses_gradients:
+            self.model.zero_grad()
+            loss = sum([target(output)
+                       for target, output in zip(targets, outputs)])
+            loss.backward(retain_graph=True)
+
+        # In most of the saliency attribution papers, the saliency is
+        # computed with a single target layer.
+        # Commonly it is the last convolutional layer.
+        # Here we support passing a list with multiple target layers.
+        # It will compute the saliency image for every image,
+        # and then aggregate them (with a default mean aggregation).
+        # This gives you more flexibility in case you just want to
+        # use all conv layers for example, all Batchnorm layers,
+        # or something else.
+        cam_per_layer = self.compute_cam_per_layer(input_tensor,
+                                                   targets,
+                                                   eigen_smooth)
+        return self.aggregate_multi_layers(cam_per_layer)
+
+    def get_target_width_height(self,
+                                input_tensor: torch.Tensor) -> Tuple[int, int]:
+        width, height = input_tensor.size(-1), input_tensor.size(-2)
+        return width, height
+
+    def compute_cam_per_layer(
+            self,
+            input_tensor: torch.Tensor,
+            targets: List[torch.nn.Module],
+            eigen_smooth: bool) -> np.ndarray:
+        activations_list = [a.cpu().data.numpy()
+                            for a in self.activations_and_grads.activations]
+        grads_list = [g.cpu().data.numpy()
+                      for g in self.activations_and_grads.gradients]
+        target_size = self.get_target_width_height(input_tensor)
+
+        cam_per_target_layer = []
+        # Loop over the saliency image from every layer
+        for i in range(len(self.target_layers)):
+            target_layer = self.target_layers[i]
+            layer_activations = None
+            layer_grads = None
+            if i < len(activations_list):
+                layer_activations = activations_list[i]
+            if i < len(grads_list):
+                layer_grads = grads_list[i]
+
+            cam = self.get_cam_image(input_tensor,
+                                     target_layer,
+                                     targets,
+                                     layer_activations,
+                                     layer_grads,
+                                     eigen_smooth)
+            cam = np.maximum(cam, 0)
+            scaled = scale_cam_image(cam, target_size)
+            cam_per_target_layer.append(scaled[:, None, :])
+
+        return cam_per_target_layer
+
+    def aggregate_multi_layers(
+            self,
+            cam_per_target_layer: np.ndarray) -> np.ndarray:
+        cam_per_target_layer = np.concatenate(cam_per_target_layer, axis=1)
+        cam_per_target_layer = np.maximum(cam_per_target_layer, 0)
+        result = np.mean(cam_per_target_layer, axis=1)
+        return scale_cam_image(result)
+
+    def forward_augmentation_smoothing(self,
+                                       input_tensor: torch.Tensor,
+                                       targets: List[torch.nn.Module],
+                                       eigen_smooth: bool = False) -> np.ndarray:
+        transforms = tta.Compose(
+            [
+                tta.HorizontalFlip(),
+                tta.Multiply(factors=[0.9, 1, 1.1]),
+            ]
+        )
+        cams = []
+        for transform in transforms:
+            augmented_tensor = transform.augment_image(input_tensor)
+            cam = self.forward(augmented_tensor,
+                               targets,
+                               eigen_smooth)
+
+            # The ttach library expects a tensor of size BxCxHxW
+            cam = cam[:, None, :, :]
+            cam = torch.from_numpy(cam)
+            cam = transform.deaugment_mask(cam)
+
+            # Back to numpy float32, HxW
+            cam = cam.numpy()
+            cam = cam[:, 0, :, :]
+            cams.append(cam)
+
+        cam = np.mean(np.float32(cams), axis=0)
+        return cam
+
+    def __call__(self,
+                 input_tensor: torch.Tensor,
+                 targets: List[torch.nn.Module] = None,
+                 aug_smooth: bool = False,
+                 eigen_smooth: bool = False) -> np.ndarray:
+
+        # Smooth the CAM result with test time augmentation
+        if aug_smooth is True:
+            return self.forward_augmentation_smoothing(
+                input_tensor, targets, eigen_smooth)
+
+        return self.forward(input_tensor,
+                            targets, eigen_smooth)
+
+    def __del__(self):
+        self.activations_and_grads.release()
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, exc_type, exc_value, exc_tb):
+        self.activations_and_grads.release()
+        if isinstance(exc_value, IndexError):
+            # Handle IndexError here...
+            print(
+                f"An exception occurred in CAM with block: {exc_type}. Message: {exc_value}")
+            return True

src/custom_code/custom_grad_cam/eigen_cam.py

@@ -0,0 +1,23 @@
+from pytorch_grad_cam.base_cam import BaseCAM
+from pytorch_grad_cam.utils.svd_on_activations import get_2d_projection
+
+# https://arxiv.org/abs/2008.00299
+
+
+class EigenCAM(BaseCAM):
+    def __init__(self, model, target_layers, use_cuda=False,
+                 reshape_transform=None):
+        super(EigenCAM, self).__init__(model,
+                                       target_layers,
+                                       use_cuda,
+                                       reshape_transform,
+                                       uses_gradients=False)
+
+    def get_cam_image(self,
+                      input_tensor,
+                      target_layer,
+                      target_category,
+                      activations,
+                      grads,
+                      eigen_smooth):
+        return get_2d_projection(activations)

src/custom_code/custom_grad_cam/eigen_grad_cam.py

@@ -0,0 +1,21 @@
+from pytorch_grad_cam.base_cam import BaseCAM
+from pytorch_grad_cam.utils.svd_on_activations import get_2d_projection
+
+# Like Eigen CAM: https://arxiv.org/abs/2008.00299
+# But multiply the activations x gradients
+
+
+class EigenGradCAM(BaseCAM):
+    def __init__(self, model, target_layers, use_cuda=False,
+                 reshape_transform=None):
+        super(EigenGradCAM, self).__init__(model, target_layers, use_cuda,
+                                           reshape_transform)
+
+    def get_cam_image(self,
+                      input_tensor,
+                      target_layer,
+                      target_category,
+                      activations,
+                      grads,
+                      eigen_smooth):
+        return get_2d_projection(grads * activations)

src/custom_code/custom_grad_cam/feature_factorization/deep_feature_factorization.py

@@ -0,0 +1,131 @@
+import numpy as np
+from PIL import Image
+import torch
+from typing import Callable, List, Tuple, Optional
+from sklearn.decomposition import NMF
+from pytorch_grad_cam.activations_and_gradients import ActivationsAndGradients
+from pytorch_grad_cam.utils.image import scale_cam_image, create_labels_legend, show_factorization_on_image
+
+
+def dff(activations: np.ndarray, n_components: int = 5):
+    """ Compute Deep Feature Factorization on a 2d Activations tensor.
+
+    :param activations: A numpy array of shape batch x channels x height x width
+    :param n_components: The number of components for the non negative matrix factorization
+    :returns: A tuple of the concepts (a numpy array with shape channels x components),
+              and the explanation heatmaps (a numpy arary with shape batch x height x width)
+    """
+
+    batch_size, channels, h, w = activations.shape
+    reshaped_activations = activations.transpose((1, 0, 2, 3))
+    reshaped_activations[np.isnan(reshaped_activations)] = 0
+    reshaped_activations = reshaped_activations.reshape(
+        reshaped_activations.shape[0], -1)
+    offset = reshaped_activations.min(axis=-1)
+    reshaped_activations = reshaped_activations - offset[:, None]
+
+    model = NMF(n_components=n_components, init='random', random_state=0)
+    W = model.fit_transform(reshaped_activations)
+    H = model.components_
+    concepts = W + offset[:, None]
+    explanations = H.reshape(n_components, batch_size, h, w)
+    explanations = explanations.transpose((1, 0, 2, 3))
+    return concepts, explanations
+
+
+class DeepFeatureFactorization:
+    """ Deep Feature Factorization: https://arxiv.org/abs/1806.10206
+        This gets a model andcomputes the 2D activations for a target layer,
+        and computes Non Negative Matrix Factorization on the activations.
+
+        Optionally it runs a computation on the concept embeddings,
+        like running a classifier on them.
+
+        The explanation heatmaps are scalled to the range [0, 1]
+        and to the input tensor width and height.
+     """
+
+    def __init__(self,
+                 model: torch.nn.Module,
+                 target_layer: torch.nn.Module,
+                 reshape_transform: Callable = None,
+                 computation_on_concepts=None
+                 ):
+        self.model = model
+        self.computation_on_concepts = computation_on_concepts
+        self.activations_and_grads = ActivationsAndGradients(
+            self.model, [target_layer], reshape_transform)
+
+    def __call__(self,
+                 input_tensor: torch.Tensor,
+                 n_components: int = 16):
+        batch_size, channels, h, w = input_tensor.size()
+        _ = self.activations_and_grads(input_tensor)
+
+        with torch.no_grad():
+            activations = self.activations_and_grads.activations[0].cpu(
+            ).numpy()
+
+        concepts, explanations = dff(activations, n_components=n_components)
+
+        processed_explanations = []
+
+        for batch in explanations:
+            processed_explanations.append(scale_cam_image(batch, (w, h)))
+
+        if self.computation_on_concepts:
+            with torch.no_grad():
+                concept_tensors = torch.from_numpy(
+                    np.float32(concepts).transpose((1, 0)))
+                concept_outputs = self.computation_on_concepts(
+                    concept_tensors).cpu().numpy()
+            return concepts, processed_explanations, concept_outputs
+        else:
+            return concepts, processed_explanations
+
+    def __del__(self):
+        self.activations_and_grads.release()
+
+    def __exit__(self, exc_type, exc_value, exc_tb):
+        self.activations_and_grads.release()
+        if isinstance(exc_value, IndexError):
+            # Handle IndexError here...
+            print(
+                f"An exception occurred in ActivationSummary with block: {exc_type}. Message: {exc_value}")
+            return True
+
+
+def run_dff_on_image(model: torch.nn.Module,
+                     target_layer: torch.nn.Module,
+                     classifier: torch.nn.Module,
+                     img_pil: Image,
+                     img_tensor: torch.Tensor,
+                     reshape_transform=Optional[Callable],
+                     n_components: int = 5,
+                     top_k: int = 2) -> np.ndarray:
+    """ Helper function to create a Deep Feature Factorization visualization for a single image.
+        TBD: Run this on a batch with several images.
+    """
+    rgb_img_float = np.array(img_pil) / 255
+    dff = DeepFeatureFactorization(model=model,
+                                   reshape_transform=reshape_transform,
+                                   target_layer=target_layer,
+                                   computation_on_concepts=classifier)
+
+    concepts, batch_explanations, concept_outputs = dff(
+        img_tensor[None, :], n_components)
+
+    concept_outputs = torch.softmax(
+        torch.from_numpy(concept_outputs),
+        axis=-1).numpy()
+    concept_label_strings = create_labels_legend(concept_outputs,
+                                                 labels=model.config.id2label,
+                                                 top_k=top_k)
+    visualization = show_factorization_on_image(
+        rgb_img_float,
+        batch_explanations[0],
+        image_weight=0.3,
+        concept_labels=concept_label_strings)
+
+    result = np.hstack((np.array(img_pil), visualization))
+    return result

src/custom_code/custom_grad_cam/feature_factorization/utils.py

@@ -0,0 +1,19 @@
+import requests
+import numpy as np
+from typing import Dict
+
+
+def create_labels_legend(concept_scores: np.ndarray,
+                         labels: Dict[int, str],
+                         top_k=2):
+    concept_categories = np.argsort(concept_scores, axis=1)[:, ::-1][:, :top_k]
+    concept_labels_topk = []
+    for concept_index in range(concept_categories.shape[0]):
+        categories = concept_categories[concept_index, :]
+        concept_labels = []
+        for category in categories:
+            score = concept_scores[concept_index, category]
+            label = f"{labels[category].split(',')[0]}:{score:.2f}"
+            concept_labels.append(label)
+        concept_labels_topk.append("\n".join(concept_labels))
+    return concept_labels_topk

src/custom_code/custom_grad_cam/fullgrad_cam.py

@@ -0,0 +1,95 @@
+import numpy as np
+import torch
+from pytorch_grad_cam.base_cam import BaseCAM
+from pytorch_grad_cam.utils.find_layers import find_layer_predicate_recursive
+from pytorch_grad_cam.utils.svd_on_activations import get_2d_projection
+from pytorch_grad_cam.utils.image import scale_accross_batch_and_channels, scale_cam_image
+
+# https://arxiv.org/abs/1905.00780
+
+
+class FullGrad(BaseCAM):
+    def __init__(self, model, target_layers, use_cuda=False,
+                 reshape_transform=None):
+        if len(target_layers) > 0:
+            print(
+                "Warning: target_layers is ignored in FullGrad. All bias layers will be used instead")
+
+        def layer_with_2D_bias(layer):
+            bias_target_layers = [torch.nn.Conv2d, torch.nn.BatchNorm2d]
+            if type(layer) in bias_target_layers and layer.bias is not None:
+                return True
+            return False
+        target_layers = find_layer_predicate_recursive(
+            model, layer_with_2D_bias)
+        super(
+            FullGrad,
+            self).__init__(
+            model,
+            target_layers,
+            use_cuda,
+            reshape_transform,
+            compute_input_gradient=True)
+        self.bias_data = [self.get_bias_data(
+            layer).cpu().numpy() for layer in target_layers]
+
+    def get_bias_data(self, layer):
+        # Borrowed from official paper impl:
+        # https://github.com/idiap/fullgrad-saliency/blob/master/saliency/tensor_extractor.py#L47
+        if isinstance(layer, torch.nn.BatchNorm2d):
+            bias = - (layer.running_mean * layer.weight
+                      / torch.sqrt(layer.running_var + layer.eps)) + layer.bias
+            return bias.data
+        else:
+            return layer.bias.data
+
+    def compute_cam_per_layer(
+            self,
+            input_tensor,
+            target_category,
+            eigen_smooth):
+        input_grad = input_tensor.grad.data.cpu().numpy()
+        grads_list = [g.cpu().data.numpy() for g in
+                      self.activations_and_grads.gradients]
+        cam_per_target_layer = []
+        target_size = self.get_target_width_height(input_tensor)
+
+        gradient_multiplied_input = input_grad * input_tensor.data.cpu().numpy()
+        gradient_multiplied_input = np.abs(gradient_multiplied_input)
+        gradient_multiplied_input = scale_accross_batch_and_channels(
+            gradient_multiplied_input,
+            target_size)
+        cam_per_target_layer.append(gradient_multiplied_input)
+
+        # Loop over the saliency image from every layer
+        assert(len(self.bias_data) == len(grads_list))
+        for bias, grads in zip(self.bias_data, grads_list):
+            bias = bias[None, :, None, None]
+            # In the paper they take the absolute value,
+            # but possibily taking only the positive gradients will work
+            # better.
+            bias_grad = np.abs(bias * grads)
+            result = scale_accross_batch_and_channels(
+                bias_grad, target_size)
+            result = np.sum(result, axis=1)
+            cam_per_target_layer.append(result[:, None, :])
+        cam_per_target_layer = np.concatenate(cam_per_target_layer, axis=1)
+        if eigen_smooth:
+            # Resize to a smaller image, since this method typically has a very large number of channels,
+            # and then consumes a lot of memory
+            cam_per_target_layer = scale_accross_batch_and_channels(
+                cam_per_target_layer, (target_size[0] // 8, target_size[1] // 8))
+            cam_per_target_layer = get_2d_projection(cam_per_target_layer)
+            cam_per_target_layer = cam_per_target_layer[:, None, :, :]
+            cam_per_target_layer = scale_accross_batch_and_channels(
+                cam_per_target_layer,
+                target_size)
+        else:
+            cam_per_target_layer = np.sum(
+                cam_per_target_layer, axis=1)[:, None, :]
+
+        return cam_per_target_layer
+
+    def aggregate_multi_layers(self, cam_per_target_layer):
+        result = np.sum(cam_per_target_layer, axis=1)
+        return scale_cam_image(result)

src/custom_code/custom_grad_cam/grad_cam.py

@@ -0,0 +1,22 @@
+import numpy as np
+from pytorch_grad_cam.base_cam import BaseCAM
+
+
+class GradCAM(BaseCAM):
+    def __init__(self, model, target_layers, use_cuda=False,
+                 reshape_transform=None):
+        super(
+            GradCAM,
+            self).__init__(
+            model,
+            target_layers,
+            use_cuda,
+            reshape_transform)
+
+    def get_cam_weights(self,
+                        input_tensor,
+                        target_layer,
+                        target_category,
+                        activations,
+                        grads):
+        return np.mean(grads, axis=(2, 3))

src/custom_code/custom_grad_cam/grad_cam_elementwise.py

@@ -0,0 +1,30 @@
+import numpy as np
+from pytorch_grad_cam.base_cam import BaseCAM
+from pytorch_grad_cam.utils.svd_on_activations import get_2d_projection
+
+
+class GradCAMElementWise(BaseCAM):
+    def __init__(self, model, target_layers, use_cuda=False,
+                 reshape_transform=None):
+        super(
+            GradCAMElementWise,
+            self).__init__(
+            model,
+            target_layers,
+            use_cuda,
+            reshape_transform)
+
+    def get_cam_image(self,
+                      input_tensor,
+                      target_layer,
+                      target_category,
+                      activations,
+                      grads,
+                      eigen_smooth):
+        elementwise_activations = np.maximum(grads * activations, 0)
+
+        if eigen_smooth:
+            cam = get_2d_projection(elementwise_activations)
+        else:
+            cam = elementwise_activations.sum(axis=1)
+        return cam

src/custom_code/custom_grad_cam/grad_cam_plusplus.py

@@ -0,0 +1,32 @@
+import numpy as np
+from pytorch_grad_cam.base_cam import BaseCAM
+
+# https://arxiv.org/abs/1710.11063
+
+
+class GradCAMPlusPlus(BaseCAM):
+    def __init__(self, model, target_layers, use_cuda=False,
+                 reshape_transform=None):
+        super(GradCAMPlusPlus, self).__init__(model, target_layers, use_cuda,
+                                              reshape_transform)
+
+    def get_cam_weights(self,
+                        input_tensor,
+                        target_layers,
+                        target_category,
+                        activations,
+                        grads):
+        grads_power_2 = grads**2
+        grads_power_3 = grads_power_2 * grads
+        # Equation 19 in https://arxiv.org/abs/1710.11063
+        sum_activations = np.sum(activations, axis=(2, 3))
+        eps = 0.000001
+        aij = grads_power_2 / (2 * grads_power_2 +
+                               sum_activations[:, :, None, None] * grads_power_3 + eps)
+        # Now bring back the ReLU from eq.7 in the paper,
+        # And zero out aijs where the activations are 0
+        aij = np.where(grads != 0, aij, 0)
+
+        weights = np.maximum(grads, 0) * aij
+        weights = np.sum(weights, axis=(2, 3))
+        return weights

src/custom_code/custom_grad_cam/guided_backprop.py

@@ -0,0 +1,100 @@
+import numpy as np
+import torch
+from torch.autograd import Function
+from pytorch_grad_cam.utils.find_layers import replace_all_layer_type_recursive
+
+
+class GuidedBackpropReLU(Function):
+    @staticmethod
+    def forward(self, input_img):
+        positive_mask = (input_img > 0).type_as(input_img)
+        output = torch.addcmul(
+            torch.zeros(
+                input_img.size()).type_as(input_img),
+            input_img,
+            positive_mask)
+        self.save_for_backward(input_img, output)
+        return output
+
+    @staticmethod
+    def backward(self, grad_output):
+        input_img, output = self.saved_tensors
+        grad_input = None
+
+        positive_mask_1 = (input_img > 0).type_as(grad_output)
+        positive_mask_2 = (grad_output > 0).type_as(grad_output)
+        grad_input = torch.addcmul(
+            torch.zeros(
+                input_img.size()).type_as(input_img),
+            torch.addcmul(
+                torch.zeros(
+                    input_img.size()).type_as(input_img),
+                grad_output,
+                positive_mask_1),
+            positive_mask_2)
+        return grad_input
+
+
+class GuidedBackpropReLUasModule(torch.nn.Module):
+    def __init__(self):
+        super(GuidedBackpropReLUasModule, self).__init__()
+
+    def forward(self, input_img):
+        return GuidedBackpropReLU.apply(input_img)
+
+
+class GuidedBackpropReLUModel:
+    def __init__(self, model, use_cuda):
+        self.model = model
+        self.model.eval()
+        self.cuda = use_cuda
+        if self.cuda:
+            self.model = self.model.cuda()
+
+    def forward(self, input_img):
+        return self.model(input_img)
+
+    def recursive_replace_relu_with_guidedrelu(self, module_top):
+
+        for idx, module in module_top._modules.items():
+            self.recursive_replace_relu_with_guidedrelu(module)
+            if module.__class__.__name__ == 'ReLU':
+                module_top._modules[idx] = GuidedBackpropReLU.apply
+        print("b")
+
+    def recursive_replace_guidedrelu_with_relu(self, module_top):
+        try:
+            for idx, module in module_top._modules.items():
+                self.recursive_replace_guidedrelu_with_relu(module)
+                if module == GuidedBackpropReLU.apply:
+                    module_top._modules[idx] = torch.nn.ReLU()
+        except BaseException:
+            pass
+
+    def __call__(self, input_img, target_category=None):
+        replace_all_layer_type_recursive(self.model,
+                                         torch.nn.ReLU,
+                                         GuidedBackpropReLUasModule())
+
+        if self.cuda:
+            input_img = input_img.cuda()
+
+        input_img = input_img.requires_grad_(True)
+
+        output = self.forward(input_img)
+
+        if target_category is None:
+            target_category = np.argmax(output.cpu().data.numpy())
+
+        loss = output[0, target_category]
+        loss.backward(retain_graph=True)
+
+        output = input_img.grad.cpu().data.numpy()
+        output = output[0, :, :, :]
+        output = output.transpose((1, 2, 0))
+
+        replace_all_layer_type_recursive(self.model,
+                                         GuidedBackpropReLUasModule,
+                                         torch.nn.ReLU())
+
+        return output

src/custom_code/custom_grad_cam/hirescam.py

@@ -0,0 +1,32 @@
+import numpy as np
+from pytorch_grad_cam.base_cam import BaseCAM
+from pytorch_grad_cam.utils.svd_on_activations import get_2d_projection
+
+
+class HiResCAM(BaseCAM):
+    def __init__(self, model, target_layers, use_cuda=False,
+                 reshape_transform=None):
+        super(
+            HiResCAM,
+            self).__init__(
+            model,
+            target_layers,
+            use_cuda,
+            reshape_transform)
+
+    def get_cam_image(self,
+                      input_tensor,
+                      target_layer,
+                      target_category,
+                      activations,
+                      grads,
+                      eigen_smooth):
+        elementwise_activations = grads * activations
+
+        if eigen_smooth:
+            print(
+                "Warning: HiResCAM's faithfulness guarantees do not hold if smoothing is applied")
+            cam = get_2d_projection(elementwise_activations)
+        else:
+            cam = elementwise_activations.sum(axis=1)
+        return cam

src/custom_code/custom_grad_cam/layer_cam.py

@@ -0,0 +1,36 @@
+import numpy as np
+from pytorch_grad_cam.base_cam import BaseCAM
+from pytorch_grad_cam.utils.svd_on_activations import get_2d_projection
+
+# https://ieeexplore.ieee.org/document/9462463
+
+
+class LayerCAM(BaseCAM):
+    def __init__(
+            self,
+            model,
+            target_layers,
+            use_cuda=False,
+            reshape_transform=None):
+        super(
+            LayerCAM,
+            self).__init__(
+            model,
+            target_layers,
+            use_cuda,
+            reshape_transform)
+
+    def get_cam_image(self,
+                      input_tensor,
+                      target_layer,
+                      target_category,
+                      activations,
+                      grads,
+                      eigen_smooth):
+        spatial_weighted_activations = np.maximum(grads, 0) * activations
+
+        if eigen_smooth:
+            cam = get_2d_projection(spatial_weighted_activations)
+        else:
+            cam = spatial_weighted_activations.sum(axis=1)
+        return cam

src/custom_code/custom_grad_cam/metrics/cam_mult_image.py

@@ -0,0 +1,37 @@
+import torch
+import numpy as np
+from typing import List, Callable
+from pytorch_grad_cam.metrics.perturbation_confidence import PerturbationConfidenceMetric
+
+
+def multiply_tensor_with_cam(input_tensor: torch.Tensor,
+                             cam: torch.Tensor):
+    """ Multiply an input tensor (after normalization)
+        with a pixel attribution map
+    """
+    return input_tensor * cam
+
+
+class CamMultImageConfidenceChange(PerturbationConfidenceMetric):
+    def __init__(self):
+        super(CamMultImageConfidenceChange,
+              self).__init__(multiply_tensor_with_cam)
+
+
+class DropInConfidence(CamMultImageConfidenceChange):
+    def __init__(self):
+        super(DropInConfidence, self).__init__()
+
+    def __call__(self, *args, **kwargs):
+        scores = super(DropInConfidence, self).__call__(*args, **kwargs)
+        scores = -scores
+        return np.maximum(scores, 0)
+
+
+class IncreaseInConfidence(CamMultImageConfidenceChange):
+    def __init__(self):
+        super(IncreaseInConfidence, self).__init__()
+
+    def __call__(self, *args, **kwargs):
+        scores = super(IncreaseInConfidence, self).__call__(*args, **kwargs)
+        return np.float32(scores > 0)

src/custom_code/custom_grad_cam/metrics/perturbation_confidence.py

@@ -0,0 +1,109 @@
+import torch
+import numpy as np
+from typing import List, Callable
+
+import numpy as np
+import cv2
+
+
+class PerturbationConfidenceMetric:
+    def __init__(self, perturbation):
+        self.perturbation = perturbation
+
+    def __call__(self, input_tensor: torch.Tensor,
+                 cams: np.ndarray,
+                 targets: List[Callable],
+                 model: torch.nn.Module,
+                 return_visualization=False,
+                 return_diff=True):
+
+        if return_diff:
+            with torch.no_grad():
+                outputs = model(input_tensor)
+                scores = [target(output).cpu().numpy()
+                          for target, output in zip(targets, outputs)]
+                scores = np.float32(scores)
+
+        batch_size = input_tensor.size(0)
+        perturbated_tensors = []
+        for i in range(batch_size):
+            cam = cams[i]
+            tensor = self.perturbation(input_tensor[i, ...].cpu(),
+                                       torch.from_numpy(cam))
+            tensor = tensor.to(input_tensor.device)
+            perturbated_tensors.append(tensor.unsqueeze(0))
+        perturbated_tensors = torch.cat(perturbated_tensors)
+
+        with torch.no_grad():
+            outputs_after_imputation = model(perturbated_tensors)
+        scores_after_imputation = [
+            target(output).cpu().numpy() for target, output in zip(
+                targets, outputs_after_imputation)]
+        scores_after_imputation = np.float32(scores_after_imputation)
+
+        if return_diff:
+            result = scores_after_imputation - scores
+        else:
+            result = scores_after_imputation
+
+        if return_visualization:
+            return result, perturbated_tensors
+        else:
+            return result
+
+
+class RemoveMostRelevantFirst:
+    def __init__(self, percentile, imputer):
+        self.percentile = percentile
+        self.imputer = imputer
+
+    def __call__(self, input_tensor, mask):
+        imputer = self.imputer
+        if self.percentile != 'auto':
+            threshold = np.percentile(mask.cpu().numpy(), self.percentile)
+            binary_mask = np.float32(mask < threshold)
+        else:
+            _, binary_mask = cv2.threshold(
+                np.uint8(mask * 255), 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
+
+        binary_mask = torch.from_numpy(binary_mask)
+        binary_mask = binary_mask.to(mask.device)
+        return imputer(input_tensor, binary_mask)
+
+
+class RemoveLeastRelevantFirst(RemoveMostRelevantFirst):
+    def __init__(self, percentile, imputer):
+        super(RemoveLeastRelevantFirst, self).__init__(percentile, imputer)
+
+    def __call__(self, input_tensor, mask):
+        return super(RemoveLeastRelevantFirst, self).__call__(
+            input_tensor, 1 - mask)
+
+
+class AveragerAcrossThresholds:
+    def __init__(
+        self,
+        imputer,
+        percentiles=[
+            10,
+            20,
+            30,
+            40,
+            50,
+            60,
+            70,
+            80,
+            90]):
+        self.imputer = imputer
+        self.percentiles = percentiles
+
+    def __call__(self,
+                 input_tensor: torch.Tensor,
+                 cams: np.ndarray,
+                 targets: List[Callable],
+                 model: torch.nn.Module):
+        scores = []
+        for percentile in self.percentiles:
+            imputer = self.imputer(percentile)
+            scores.append(imputer(input_tensor, cams, targets, model))
+        return np.mean(np.float32(scores), axis=0)

src/custom_code/custom_grad_cam/metrics/road.py

@@ -0,0 +1,181 @@
+# A Consistent and Efficient Evaluation Strategy for Attribution Methods
+# https://arxiv.org/abs/2202.00449
+# Taken from https://raw.githubusercontent.com/tleemann/road_evaluation/main/imputations.py
+# MIT License
+
+# Copyright (c) 2022 Tobias Leemann
+
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+
+# Implementations of our imputation models.
+import torch
+import numpy as np
+from scipy.sparse import lil_matrix, csc_matrix
+from scipy.sparse.linalg import spsolve
+from typing import List, Callable
+from pytorch_grad_cam.metrics.perturbation_confidence import PerturbationConfidenceMetric, \
+    AveragerAcrossThresholds, \
+    RemoveMostRelevantFirst, \
+    RemoveLeastRelevantFirst
+
+# The weights of the surrounding pixels
+neighbors_weights = [((1, 1), 1 / 12),
+                     ((0, 1), 1 / 6),
+                     ((-1, 1), 1 / 12),
+                     ((1, -1), 1 / 12),
+                     ((0, -1), 1 / 6),
+                     ((-1, -1), 1 / 12),
+                     ((1, 0), 1 / 6),
+                     ((-1, 0), 1 / 6)]
+
+
+class NoisyLinearImputer:
+    def __init__(self,
+                 noise: float = 0.01,
+                 weighting: List[float] = neighbors_weights):
+        """
+                Noisy linear imputation.
+                noise: magnitude of noise to add (absolute, set to 0 for no noise)
+                weighting: Weights of the neighboring pixels in the computation.
+                List of tuples of (offset, weight)
+        """
+        self.noise = noise
+        self.weighting = neighbors_weights
+
+    @staticmethod
+    def add_offset_to_indices(indices, offset, mask_shape):
+        """ Add the corresponding offset to the indices.
+    Return new indices plus a valid bit-vector. """
+        cord1 = indices % mask_shape[1]
+        cord0 = indices // mask_shape[1]
+        cord0 += offset[0]
+        cord1 += offset[1]
+        valid = ((cord0 < 0) | (cord1 < 0) |
+                 (cord0 >= mask_shape[0]) |
+                 (cord1 >= mask_shape[1]))
+        return ~valid, indices + offset[0] * mask_shape[1] + offset[1]
+
+    @staticmethod
+    def setup_sparse_system(mask, img, neighbors_weights):
+        """ Vectorized version to set up the equation system.
+                mask: (H, W)-tensor of missing pixels.
+                Image: (H, W, C)-tensor of all values.
+                Return (N,N)-System matrix, (N,C)-Right hand side for each of the C channels.
+        """
+        maskflt = mask.flatten()
+        imgflat = img.reshape((img.shape[0], -1))
+    # Indices that are imputed in the flattened mask:
+        indices = np.argwhere(maskflt == 0).flatten()
+        coords_to_vidx = np.zeros(len(maskflt), dtype=int)
+        coords_to_vidx[indices] = np.arange(len(indices))
+        numEquations = len(indices)
+    # System matrix:
+        A = lil_matrix((numEquations, numEquations))
+        b = np.zeros((numEquations, img.shape[0]))
+    # Sum of weights assigned:
+        sum_neighbors = np.ones(numEquations)
+        for n in neighbors_weights:
+            offset, weight = n[0], n[1]
+            # Take out outliers
+            valid, new_coords = NoisyLinearImputer.add_offset_to_indices(
+                indices, offset, mask.shape)
+            valid_coords = new_coords[valid]
+            valid_ids = np.argwhere(valid == 1).flatten()
+            # Add values to the right hand-side
+            has_values_coords = valid_coords[maskflt[valid_coords] > 0.5]
+            has_values_ids = valid_ids[maskflt[valid_coords] > 0.5]
+            b[has_values_ids, :] -= weight * imgflat[:, has_values_coords].T
+            # Add weights to the system (left hand side)
+# Find coordinates in the system.
+            has_no_values = valid_coords[maskflt[valid_coords] < 0.5]
+            variable_ids = coords_to_vidx[has_no_values]
+            has_no_values_ids = valid_ids[maskflt[valid_coords] < 0.5]
+            A[has_no_values_ids, variable_ids] = weight
+            # Reduce weight for invalid
+            sum_neighbors[np.argwhere(valid == 0).flatten()] = \
+                sum_neighbors[np.argwhere(valid == 0).flatten()] - weight
+
+        A[np.arange(numEquations), np.arange(numEquations)] = -sum_neighbors
+        return A, b
+
+    def __call__(self, img: torch.Tensor, mask: torch.Tensor):
+        """ Our linear inputation scheme. """
+        """
+		This is the function to do the linear infilling
+		img: original image (C,H,W)-tensor;
+		mask: mask; (H,W)-tensor
+
+		"""
+        imgflt = img.reshape(img.shape[0], -1)
+        maskflt = mask.reshape(-1)
+    # Indices that need to be imputed.
+        indices_linear = np.argwhere(maskflt == 0).flatten()
+        # Set up sparse equation system, solve system.
+        A, b = NoisyLinearImputer.setup_sparse_system(
+            mask.numpy(), img.numpy(), neighbors_weights)
+        res = torch.tensor(spsolve(csc_matrix(A), b), dtype=torch.float)
+
+        # Fill the values with the solution of the system.
+        img_infill = imgflt.clone()
+        img_infill[:, indices_linear] = res.t() + self.noise * \
+            torch.randn_like(res.t())
+
+        return img_infill.reshape_as(img)
+
+
+class ROADMostRelevantFirst(PerturbationConfidenceMetric):
+    def __init__(self, percentile=80):
+        super(ROADMostRelevantFirst, self).__init__(
+            RemoveMostRelevantFirst(percentile, NoisyLinearImputer()))
+
+
+class ROADLeastRelevantFirst(PerturbationConfidenceMetric):
+    def __init__(self, percentile=20):
+        super(ROADLeastRelevantFirst, self).__init__(
+            RemoveLeastRelevantFirst(percentile, NoisyLinearImputer()))
+
+
+class ROADMostRelevantFirstAverage(AveragerAcrossThresholds):
+    def __init__(self, percentiles=[10, 20, 30, 40, 50, 60, 70, 80, 90]):
+        super(ROADMostRelevantFirstAverage, self).__init__(
+            ROADMostRelevantFirst, percentiles)
+
+
+class ROADLeastRelevantFirstAverage(AveragerAcrossThresholds):
+    def __init__(self, percentiles=[10, 20, 30, 40, 50, 60, 70, 80, 90]):
+        super(ROADLeastRelevantFirstAverage, self).__init__(
+            ROADLeastRelevantFirst, percentiles)
+
+
+class ROADCombined:
+    def __init__(self, percentiles=[10, 20, 30, 40, 50, 60, 70, 80, 90]):
+        self.percentiles = percentiles
+        self.morf_averager = ROADMostRelevantFirstAverage(percentiles)
+        self.lerf_averager = ROADLeastRelevantFirstAverage(percentiles)
+
+    def __call__(self,
+                 input_tensor: torch.Tensor,
+                 cams: np.ndarray,
+                 targets: List[Callable],
+                 model: torch.nn.Module):
+
+        scores_lerf = self.lerf_averager(input_tensor, cams, targets, model)
+        scores_morf = self.morf_averager(input_tensor, cams, targets, model)
+        return (scores_lerf - scores_morf) / 2

src/custom_code/custom_grad_cam/random_cam.py

@@ -0,0 +1,22 @@
+import numpy as np
+from pytorch_grad_cam.base_cam import BaseCAM
+
+
+class RandomCAM(BaseCAM):
+    def __init__(self, model, target_layers, use_cuda=False,
+                 reshape_transform=None):
+        super(
+            RandomCAM,
+            self).__init__(
+            model,
+            target_layers,
+            use_cuda,
+            reshape_transform)
+
+    def get_cam_weights(self,
+                        input_tensor,
+                        target_layer,
+                        target_category,
+                        activations,
+                        grads):
+        return np.random.uniform(-1, 1, size=(grads.shape[0], grads.shape[1]))

src/custom_code/custom_grad_cam/score_cam.py

@@ -0,0 +1,60 @@
+import torch
+import tqdm
+from pytorch_grad_cam.base_cam import BaseCAM
+
+
+class ScoreCAM(BaseCAM):
+    def __init__(
+            self,
+            model,
+            target_layers,
+            use_cuda=False,
+            reshape_transform=None):
+        super(ScoreCAM, self).__init__(model,
+                                       target_layers,
+                                       use_cuda,
+                                       reshape_transform=reshape_transform,
+                                       uses_gradients=False)
+
+    def get_cam_weights(self,
+                        input_tensor,
+                        target_layer,
+                        targets,
+                        activations,
+                        grads):
+        with torch.no_grad():
+            upsample = torch.nn.UpsamplingBilinear2d(
+                size=input_tensor.shape[-2:])
+            activation_tensor = torch.from_numpy(activations)
+            if self.cuda:
+                activation_tensor = activation_tensor.cuda()
+
+            upsampled = upsample(activation_tensor)
+
+            maxs = upsampled.view(upsampled.size(0),
+                                  upsampled.size(1), -1).max(dim=-1)[0]
+            mins = upsampled.view(upsampled.size(0),
+                                  upsampled.size(1), -1).min(dim=-1)[0]
+
+            maxs, mins = maxs[:, :, None, None], mins[:, :, None, None]
+            upsampled = (upsampled - mins) / (maxs - mins)
+
+            input_tensors = input_tensor[:, None,
+                                         :, :] * upsampled[:, :, None, :, :]
+
+            if hasattr(self, "batch_size"):
+                BATCH_SIZE = self.batch_size
+            else:
+                BATCH_SIZE = 16
+
+            scores = []
+            for target, tensor in zip(targets, input_tensors):
+                for i in tqdm.tqdm(range(0, tensor.size(0), BATCH_SIZE)):
+                    batch = tensor[i: i + BATCH_SIZE, :]
+                    outputs = [target(o).cpu().item()
+                               for o in self.model(batch)]
+                    scores.extend(outputs)
+            scores = torch.Tensor(scores)
+            scores = scores.view(activations.shape[0], activations.shape[1])
+            weights = torch.nn.Softmax(dim=-1)(scores).numpy()
+            return weights

src/custom_code/custom_grad_cam/sobel_cam.py

@@ -0,0 +1,11 @@
+import cv2
+
+
+def sobel_cam(img):
+    gray = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
+    grad_x = cv2.Sobel(gray, cv2.CV_64F, 1, 0, ksize=3)
+    grad_y = cv2.Sobel(gray, cv2.CV_64F, 0, 1, ksize=3)
+    abs_grad_x = cv2.convertScaleAbs(grad_x)
+    abs_grad_y = cv2.convertScaleAbs(grad_y)
+    grad = cv2.addWeighted(abs_grad_x, 0.5, abs_grad_y, 0.5, 0)
+    return grad

src/custom_code/custom_grad_cam/utils/__init__.py

@@ -0,0 +1,4 @@
+from pytorch_grad_cam.utils.image import deprocess_image
+from pytorch_grad_cam.utils.svd_on_activations import get_2d_projection
+from pytorch_grad_cam.utils import model_targets
+from pytorch_grad_cam.utils import reshape_transforms

src/custom_code/custom_grad_cam/utils/find_layers.py

@@ -0,0 +1,30 @@
+def replace_layer_recursive(model, old_layer, new_layer):
+    for name, layer in model._modules.items():
+        if layer == old_layer:
+            model._modules[name] = new_layer
+            return True
+        elif replace_layer_recursive(layer, old_layer, new_layer):
+            return True
+    return False
+
+
+def replace_all_layer_type_recursive(model, old_layer_type, new_layer):
+    for name, layer in model._modules.items():
+        if isinstance(layer, old_layer_type):
+            model._modules[name] = new_layer
+        replace_all_layer_type_recursive(layer, old_layer_type, new_layer)
+
+
+def find_layer_types_recursive(model, layer_types):
+    def predicate(layer):
+        return type(layer) in layer_types
+    return find_layer_predicate_recursive(model, predicate)
+
+
+def find_layer_predicate_recursive(model, predicate):
+    result = []
+    for name, layer in model._modules.items():
+        if predicate(layer):
+            result.append(layer)
+        result.extend(find_layer_predicate_recursive(layer, predicate))
+    return result

src/custom_code/custom_grad_cam/utils/image.py

@@ -0,0 +1,183 @@
+import matplotlib
+from matplotlib import pyplot as plt
+from matplotlib.lines import Line2D
+import cv2
+import numpy as np
+import torch
+from torchvision.transforms import Compose, Normalize, ToTensor
+from typing import List, Dict
+import math
+
+
+def preprocess_image(
+    img: np.ndarray, mean=[
+        0.5, 0.5, 0.5], std=[
+            0.5, 0.5, 0.5]) -> torch.Tensor:
+    preprocessing = Compose([
+        ToTensor(),
+        Normalize(mean=mean, std=std)
+    ])
+    return preprocessing(img.copy()).unsqueeze(0)
+
+
+def deprocess_image(img):
+    """ see https://github.com/jacobgil/keras-grad-cam/blob/master/grad-cam.py#L65 """
+    img = img - np.mean(img)
+    img = img / (np.std(img) + 1e-5)
+    img = img * 0.1
+    img = img + 0.5
+    img = np.clip(img, 0, 1)
+    return np.uint8(img * 255)
+
+
+def show_cam_on_image(img: np.ndarray,
+                      mask: np.ndarray,
+                      use_rgb: bool = False,
+                      colormap: int = cv2.COLORMAP_JET,
+                      image_weight: float = 0.5) -> np.ndarray:
+    """ This function overlays the cam mask on the image as an heatmap.
+    By default the heatmap is in BGR format.
+
+    :param img: The base image in RGB or BGR format.
+    :param mask: The cam mask.
+    :param use_rgb: Whether to use an RGB or BGR heatmap, this should be set to True if 'img' is in RGB format.
+    :param colormap: The OpenCV colormap to be used.
+    :param image_weight: The final result is image_weight * img + (1-image_weight) * mask.
+    :returns: The default image with the cam overlay.
+    """
+    heatmap = cv2.applyColorMap(np.uint8(255 * mask), colormap)
+    if use_rgb:
+        heatmap = cv2.cvtColor(heatmap, cv2.COLOR_BGR2RGB)
+    heatmap = np.float32(heatmap) / 255
+
+    if np.max(img) > 1:
+        raise Exception(
+            "The input image should np.float32 in the range [0, 1]")
+
+    if image_weight < 0 or image_weight > 1:
+        raise Exception(
+            f"image_weight should be in the range [0, 1].\
+                Got: {image_weight}")
+
+    cam = (1 - image_weight) * heatmap + image_weight * img
+    cam = cam / np.max(cam)
+    return np.uint8(255 * cam)
+
+
+def create_labels_legend(concept_scores: np.ndarray,
+                         labels: Dict[int, str],
+                         top_k=2):
+    concept_categories = np.argsort(concept_scores, axis=1)[:, ::-1][:, :top_k]
+    concept_labels_topk = []
+    for concept_index in range(concept_categories.shape[0]):
+        categories = concept_categories[concept_index, :]
+        concept_labels = []
+        for category in categories:
+            score = concept_scores[concept_index, category]
+            label = f"{','.join(labels[category].split(',')[:3])}:{score:.2f}"
+            concept_labels.append(label)
+        concept_labels_topk.append("\n".join(concept_labels))
+    return concept_labels_topk
+
+
+def show_factorization_on_image(img: np.ndarray,
+                                explanations: np.ndarray,
+                                colors: List[np.ndarray] = None,
+                                image_weight: float = 0.5,
+                                concept_labels: List = None) -> np.ndarray:
+    """ Color code the different component heatmaps on top of the image.
+        Every component color code will be magnified according to the heatmap itensity
+        (by modifying the V channel in the HSV color space),
+        and optionally create a lagend that shows the labels.
+
+        Since different factorization component heatmaps can overlap in principle,
+        we need a strategy to decide how to deal with the overlaps.
+        This keeps the component that has a higher value in it's heatmap.
+
+    :param img: The base image RGB format.
+    :param explanations: A tensor of shape num_componetns x height x width, with the component visualizations.
+    :param colors: List of R, G, B colors to be used for the components.
+                   If None, will use the gist_rainbow cmap as a default.
+    :param image_weight: The final result is image_weight * img + (1-image_weight) * visualization.
+    :concept_labels: A list of strings for every component. If this is paseed, a legend that shows
+                     the labels and their colors will be added to the image.
+    :returns: The visualized image.
+    """
+    n_components = explanations.shape[0]
+    if colors is None:
+        # taken from https://github.com/edocollins/DFF/blob/master/utils.py
+        _cmap = plt.cm.get_cmap('gist_rainbow')
+        colors = [
+            np.array(
+                _cmap(i)) for i in np.arange(
+                0,
+                1,
+                1.0 /
+                n_components)]
+    concept_per_pixel = explanations.argmax(axis=0)
+    masks = []
+    for i in range(n_components):
+        mask = np.zeros(shape=(img.shape[0], img.shape[1], 3))
+        mask[:, :, :] = colors[i][:3]
+        explanation = explanations[i]
+        explanation[concept_per_pixel != i] = 0
+        mask = np.uint8(mask * 255)
+        mask = cv2.cvtColor(mask, cv2.COLOR_RGB2HSV)
+        mask[:, :, 2] = np.uint8(255 * explanation)
+        mask = cv2.cvtColor(mask, cv2.COLOR_HSV2RGB)
+        mask = np.float32(mask) / 255
+        masks.append(mask)
+
+    mask = np.sum(np.float32(masks), axis=0)
+    result = img * image_weight + mask * (1 - image_weight)
+    result = np.uint8(result * 255)
+
+    if concept_labels is not None:
+        px = 1 / plt.rcParams['figure.dpi']  # pixel in inches
+        fig = plt.figure(figsize=(result.shape[1] * px, result.shape[0] * px))
+        plt.rcParams['legend.fontsize'] = int(
+            14 * result.shape[0] / 256 / max(1, n_components / 6))
+        lw = 5 * result.shape[0] / 256
+        lines = [Line2D([0], [0], color=colors[i], lw=lw)
+                 for i in range(n_components)]
+        plt.legend(lines,
+                   concept_labels,
+                   mode="expand",
+                   fancybox=True,
+                   shadow=True)
+
+        plt.tight_layout(pad=0, w_pad=0, h_pad=0)
+        plt.axis('off')
+        fig.canvas.draw()
+        data = np.frombuffer(fig.canvas.tostring_rgb(), dtype=np.uint8)
+        plt.close(fig=fig)
+        data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))
+        data = cv2.resize(data, (result.shape[1], result.shape[0]))
+        result = np.hstack((result, data))
+    return result
+
+
+def scale_cam_image(cam, target_size=None):
+    result = []
+    for img in cam:
+        img = img - np.min(img)
+        img = img / (1e-7 + np.max(img))
+        if target_size is not None:
+            img = cv2.resize(img, target_size)
+        result.append(img)
+    result = np.float32(result)
+
+    return result
+
+
+def scale_accross_batch_and_channels(tensor, target_size):
+    batch_size, channel_size = tensor.shape[:2]
+    reshaped_tensor = tensor.reshape(
+        batch_size * channel_size, *tensor.shape[2:])
+    result = scale_cam_image(reshaped_tensor, target_size)
+    result = result.reshape(
+        batch_size,
+        channel_size,
+        target_size[1],
+        target_size[0])
+    return result

src/custom_code/custom_grad_cam/utils/model_targets.py

@@ -0,0 +1,103 @@
+import numpy as np
+import torch
+import torchvision
+
+
+class ClassifierOutputTarget:
+    def __init__(self, category):
+        self.category = category
+
+    def __call__(self, model_output):
+        if len(model_output.shape) == 1:
+            return model_output[self.category]
+        return model_output[:, self.category]
+
+
+class ClassifierOutputSoftmaxTarget:
+    def __init__(self, category):
+        self.category = category
+
+    def __call__(self, model_output):
+        if len(model_output.shape) == 1:
+            return torch.softmax(model_output, dim=-1)[self.category]
+        return torch.softmax(model_output, dim=-1)[:, self.category]
+
+
+class BinaryClassifierOutputTarget:
+    def __init__(self, category):
+        self.category = category
+
+    def __call__(self, model_output):
+        if self.category == 1:
+            sign = 1
+        else:
+            sign = -1
+        return torch.abs(model_output) * sign
+
+
+class SoftmaxOutputTarget:
+    def __init__(self):
+        pass
+
+    def __call__(self, model_output):
+        return torch.softmax(model_output, dim=-1)
+
+
+class RawScoresOutputTarget:
+    def __init__(self):
+        pass
+
+    def __call__(self, model_output):
+        return model_output
+
+
+class SemanticSegmentationTarget:
+    """ Gets a binary spatial mask and a category,
+        And return the sum of the category scores,
+        of the pixels in the mask. """
+
+    def __init__(self, category, mask):
+        self.category = category
+        self.mask = torch.from_numpy(mask)
+        if torch.cuda.is_available():
+            self.mask = self.mask.cuda()
+
+    def __call__(self, model_output):
+        return (model_output[self.category, :, :] * self.mask).sum()
+
+
+class FasterRCNNBoxScoreTarget:
+    """ For every original detected bounding box specified in "bounding boxes",
+        assign a score on how the current bounding boxes match it,
+            1. In IOU
+            2. In the classification score.
+        If there is not a large enough overlap, or the category changed,
+        assign a score of 0.
+
+        The total score is the sum of all the box scores.
+    """
+
+    def __init__(self, labels, bounding_boxes, iou_threshold=0.5):
+        self.labels = labels
+        self.bounding_boxes = bounding_boxes
+        self.iou_threshold = iou_threshold
+
+    def __call__(self, model_outputs):
+        output = torch.Tensor([0])
+        if torch.cuda.is_available():
+            output = output.cuda()
+
+        if len(model_outputs["boxes"]) == 0:
+            return output
+
+        for box, label in zip(self.bounding_boxes, self.labels):
+            box = torch.Tensor(box[None, :])
+            if torch.cuda.is_available():
+                box = box.cuda()
+
+            ious = torchvision.ops.box_iou(box, model_outputs["boxes"])
+            index = ious.argmax()
+            if ious[0, index] > self.iou_threshold and model_outputs["labels"][index] == label:
+                score = ious[0, index] + model_outputs["scores"][index]
+                output = output + score
+        return output

src/custom_code/custom_grad_cam/utils/reshape_transforms.py

@@ -0,0 +1,34 @@
+import torch
+
+
+def fasterrcnn_reshape_transform(x):
+    target_size = x['pool'].size()[-2:]
+    activations = []
+    for key, value in x.items():
+        activations.append(
+            torch.nn.functional.interpolate(
+                torch.abs(value),
+                target_size,
+                mode='bilinear'))
+    activations = torch.cat(activations, axis=1)
+    return activations
+
+
+def swinT_reshape_transform(tensor, height=7, width=7):
+    result = tensor.reshape(tensor.size(0),
+                            height, width, tensor.size(2))
+
+    # Bring the channels to the first dimension,
+    # like in CNNs.
+    result = result.transpose(2, 3).transpose(1, 2)
+    return result
+
+
+def vit_reshape_transform(tensor, height=14, width=14):
+    result = tensor[:, 1:, :].reshape(tensor.size(0),
+                                      height, width, tensor.size(2))
+
+    # Bring the channels to the first dimension,
+    # like in CNNs.
+    result = result.transpose(2, 3).transpose(1, 2)
+    return result

src/custom_code/custom_grad_cam/utils/svd_on_activations.py

@@ -0,0 +1,19 @@
+import numpy as np
+
+
+def get_2d_projection(activation_batch):
+    # TBD: use pytorch batch svd implementation
+    activation_batch[np.isnan(activation_batch)] = 0
+    projections = []
+    for activations in activation_batch:
+        reshaped_activations = (activations).reshape(
+            activations.shape[0], -1).transpose()
+        # Centering before the SVD seems to be important here,
+        # Otherwise the image returned is negative
+        reshaped_activations = reshaped_activations - \
+            reshaped_activations.mean(axis=0)
+        U, S, VT = np.linalg.svd(reshaped_activations, full_matrices=True)
+        projection = reshaped_activations @ VT[0, :]
+        projection = projection.reshape(activations.shape[1:])
+        projections.append(projection)
+    return np.float32(projections)

src/custom_code/custom_grad_cam/xgrad_cam.py

@@ -0,0 +1,31 @@
+import numpy as np
+from pytorch_grad_cam.base_cam import BaseCAM
+
+
+class XGradCAM(BaseCAM):
+    def __init__(
+            self,
+            model,
+            target_layers,
+            use_cuda=False,
+            reshape_transform=None):
+        super(
+            XGradCAM,
+            self).__init__(
+            model,
+            target_layers,
+            use_cuda,
+            reshape_transform)
+
+    def get_cam_weights(self,
+                        input_tensor,
+                        target_layer,
+                        target_category,
+                        activations,
+                        grads):
+        sum_activations = np.sum(activations, axis=(2, 3))
+        eps = 1e-7
+        weights = grads * activations / \
+            (sum_activations[:, :, None, None] + eps)
+        weights = weights.sum(axis=(2, 3))
+        return weights