Implement the Grad-CAM algorithm using PyTorch. Please refer to the paper for more details: Selvaraja et al. "Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization", IJCV, 2019.
Gradient Class Activation Map (Grad-CAM) is useful for analyzing which part of an image (or feature map) is most important to determine a specific class. In other words, we can look into the black box of convolutional neural networks to some extent. The idea is that the importance of a pixel in the feature can be measured by the product of the gradient of the classification score (logits) w.r.t the pixel and the feature response in the position of this pixel.
The algorithm is implemented as a single function grad_cam.grad_cam
from grad_cam import grad_cam
cam = grad_cam(model, img, feature_layer='layer4.2.relu', score_layer=None, class_index=None)where model is a PyTorch neural network model (module), img is a torch.Tensor
representing one image (or a batch of image) with size (N, C, H, W). feature_layer
is the layer producing the feature map. Similarly, we can specify the layer which produce the
classification score (logits). score_layer will be default to the last layer of the model if
not specified. class_index specify which category we want to analyze, will be default to the
category with the highest classification score if not specified.