Grad-CAM uses the gradient information flowing into any particular convolutional layer to produce a coarse 
localization map highlighting the important regions in the image to understand each neuron for a decision of interest.

First one is the normal image and the second one has gradCAM implemented over its last convolutional layer.
The third one has highlighted regions displayed over the initial image.

This library is not an official Google product, although it is built for Tensorflow 2.0 models.
tf-explain is completely built over the new TF2.0 API, and is primarily based on tf.keras when possible.
Implementing it over the classical dance images shows the results as: