A playground implementation of GradCAM algorithm.
Gradient-weighted Class Activation Mapping (Grad-CAM), uses the gradients of any target class flowing into the final convolutional layer to produce a coarse localization map highlighting the important regions in the image for predicting the class. Reference
Here is a implementation of Grad-CAM that you can play with. You can choose your model (VGG16, ResNet152V2, InceptionV3, Xception), your desired layer (the final conv layer or any other layer), and define the number of top predictions. Increase alpha for more intense overlay of the heatmap on the original image. Simply, open the notebook in colab, upload your image and enjoy!
