initial commit

README.md

@@ -0,0 +1,12 @@
+---
+title: SDSdemo
+emoji: 👩‍⚕️
+colorFrom: pink
+colorTo: indigo
+sdk: gradio
+app_file: runSDSdemo.py
+pinned: false
+license: afl-3.0
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces#reference

checkpoints/ham10k_checkpoint_mobile_0.82_epoch24.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7c671f87198d4c5bb3a442ca9bf810dad45f755f6f7045b0b7b5186df9f767db
+size 50851683

pytorch_grad_cam/__init__.py

@@ -0,0 +1,14 @@
+from pytorch_grad_cam.grad_cam import GradCAM
+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.fullgrad_cam import FullGrad
+from pytorch_grad_cam.guided_backprop import GuidedBackpropReLUModel
+from pytorch_grad_cam.activations_and_gradients import ActivationsAndGradients
+import pytorch_grad_cam.utils.model_targets
+import pytorch_grad_cam.utils.reshape_transforms

pytorch_grad_cam/ablation_cam.py

@@ -0,0 +1,134 @@
+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

pytorch_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

pytorch_grad_cam/ablation_layer.py

@@ -0,0 +1,124 @@
+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, 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 = 1e7
+                output[i, self.indices[i], :] = torch.min(
+                    output) - ABLATION_VALUE
+
+        output = output.transpose(2, 1)
+
+        return output
+
+
+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

pytorch_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()

pytorch_grad_cam/base_cam.py

@@ -0,0 +1,199 @@
+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)
+
+    """ 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)
+        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

pytorch_grad_cam/eigen_cam.py

@@ -0,0 +1,20 @@
+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)
+
+    def get_cam_image(self,
+                      input_tensor,
+                      target_layer,
+                      target_category,
+                      activations,
+                      grads,
+                      eigen_smooth):
+        return get_2d_projection(activations)

pytorch_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)

pytorch_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)

pytorch_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))

pytorch_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

pytorch_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

pytorch_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

pytorch_grad_cam/score_cam.py

@@ -0,0 +1,63 @@
+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)
+
+        if len(target_layers) > 0:
+            print("Warning: You are using ScoreCAM with target layers, "
+                  "however ScoreCAM will ignore them.")
+
+    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

pytorch_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

pytorch_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

pytorch_grad_cam/utils/image.py

@@ -0,0 +1,73 @@
+import cv2
+import numpy as np
+import torch
+from torchvision.transforms import Compose, Normalize, ToTensor
+
+
+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) -> 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.
+    :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]")
+
+    cam = heatmap + img
+    cam = cam / np.max(cam)
+    return np.uint8(255 * cam)
+
+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

pytorch_grad_cam/utils/model_targets.py

@@ -0,0 +1,61 @@
+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 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

pytorch_grad_cam/utils/reshape_transforms.py

@@ -0,0 +1,27 @@
+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

pytorch_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)

pytorch_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

runSDSdemo.py

@@ -0,0 +1,96 @@
+# import pytorch related dependencies
+import torch
+from PIL import Image
+from torch import nn
+import numpy as np
+import torchvision as torchvision
+import torchvision.transforms as transforms
+from pytorch_grad_cam import GradCAM, ScoreCAM, GradCAMPlusPlus, AblationCAM, XGradCAM, EigenCAM, FullGrad
+from pytorch_grad_cam.utils.image import show_cam_on_image
+import gradio as gr
+
+# model setup
+device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+classes = [ 'actinic keratoses', 'basal cell carcinoma', 'benign keratosis-like lesions', 
+           'dermatofibroma','melanoma', 'melanocytic nevi', 'vascular lesions']
+model = torchvision.models.mobilenet_v3_large(pretrained = False) # This is a very well known network but it is designed for 1000 classes and not just cats and dogs this is why we need the next line
+model.classifier[3] = nn.Linear(1280, 7)
+#state_dict_trained = torch.hub.load_state_dict_from_url("https://github.com/tobiascz/demotime/raw/main/checkpoints/ham10k_checkpoint_mobile_0.82_epoch24.pt", model_dir=".", map_location = device)
+import os
+print(os.getcwd())
+state_dict_trained = torch.load('checkpoints/ham10k_checkpoint_mobile_0.82_epoch24.pt', map_location=torch.device('cpu'))
+model.load_state_dict(state_dict_trained["model_state_dict"]) ## Here we load the trained weights (state_dict) in our model 
+model.eval() # This
+
+# image pre-processing
+norm_mean = (0.4914, 0.4822, 0.4465)
+norm_std = (0.2023, 0.1994, 0.2010)
+transform = transforms.Compose([ # resize image to the network input size
+                  transforms.CenterCrop((400,400)),
+                  transforms.ToTensor(),
+                  transforms.Normalize(norm_mean, norm_std)
+                  ])
+# convert tensot to numpy array
+def tensor2npimg(tensor, mean, std):
+  # inverse of normalization
+  tensor = tensor.clone()
+  mean_tensor = torch.as_tensor(list(mean), dtype=tensor.dtype, device=tensor.device).view(-1,1,1)
+  std_tensor = torch.as_tensor(list(std), dtype=tensor.dtype, device=tensor.device).view(-1,1,1)
+  tensor.mul_(std_tensor).add_(mean_tensor)
+  # convert tensor to numpy format for plt presentation
+  npimg = tensor.numpy()
+  npimg = np.transpose(npimg,(1,2,0)) # C*H*W => H*W*C
+  return npimg
+
+
+# draw Grad-CAM on image
+# target layer could be any layer before the final attention block
+# Some common choices are:
+#   FasterRCNN: model.backbone
+#   Resnet18 and 50: model.layer4[-1]
+#   VGG and densenet161: model.features[-1]
+#   mnasnet1_0: model.layers[-1]
+#   ViT: model.blocks[-1].norm1
+#   SwinT: model.layers[-1].blocks[-1].norm1
+def image_grad_cam(model, input_tensor, input_float_np, target_layers):
+  cam = GradCAM(model=model, target_layers=target_layers, use_cuda=False)
+  grayscale_cam = cam(input_tensor=input_tensor, aug_smooth=True, eigen_smooth=True)
+  grayscale_cam = grayscale_cam[0, :]
+  return show_cam_on_image(input_float_np, grayscale_cam, use_rgb=True)
+
+
+# config the predict function for Gradio, input type of image is numpy.nparray
+def predict(input_img):
+  # numpy.nparray -> PIL.Image
+  leasionExample = Image.fromarray(input_img.astype('uint8'), 'RGB')
+  # normalize the image to fit the input size of our model
+  leasion_tensor = transform(leasionExample)
+  input_float_np = tensor2npimg(leasion_tensor, norm_mean, norm_std)
+  leasion_tensor = leasion_tensor.unsqueeze(dim=0)
+  # predict
+  with torch.no_grad():
+    outputs = model(leasion_tensor)
+  outputs = torch.exp(outputs) 
+  # probabilities of all classes
+  pred_softmax = torch.softmax(outputs, dim=1).cpu().numpy()[0]
+  # class with hightest probability
+  pred = torch.argmax(outputs, dim=1).cpu().numpy()
+  # diagnostic suggestions
+  if pred == 1 or pred == 4:
+    suggestion = "CHECK WITH YOUR MD!"
+  else:
+    suggestion = "Nothing to be worried about."
+  # grad_cam image
+  target_layers = model.features[-1]
+  output_img = image_grad_cam(model,leasion_tensor,input_float_np,target_layers)
+  # return label dict and suggestion
+  return {classes[i]: float(pred_softmax[i]) for i in range(len(classes))}, suggestion, output_img
+
+# start gradio application
+gr.Interface(
+        fn=predict, 
+        inputs=gr.inputs.Image(), 
+        outputs=[gr.outputs.Label(label="Predict Result"), gr.outputs.Textbox(type="str", label="Recommendation"), gr.outputs.Image(label="GradCAM")],
+        examples=[['sample1.png'],['sample2.jpg'],['sample3.jpg']],
+        title="Skin Lesion Classifier"
+      ).launch()