Advanced Deep Learning with Python

Gradient-weighted class activation mapping

To understand gradient-weighted class activation mapping (Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization, https://arxiv.org/abs/1610.02391), let's quote the source paper itself:

"Grad-CAM uses the gradients of any target concept (say, the logits for 'dog' or even a caption) flowing into the final convolutional layer to produce a coarse localization map highlighting the important regions in the image for predicting the concept."

The following screenshot shows the Grad-CAM algorithm:

Grad-CAM schema; Source: https://arxiv.org/abs/1610.02391

Now, let's look at how it ...