keras-vis

Why we included this project

Keras-vis is a small toolkit for looking inside trained Keras models. It frames activation maximization, saliency maps, and class activation maps as energy minimization problems behind one API, so you can plug in custom loss functions and regularizers without much ceremony. If you are debugging an image classifier and want to know which parts of an input the network actually relies on, or what a given filter has learned, the saliency and class activation maps give you a direct look. The examples, including a VGG16 walkthrough that renders an animated GIF of the optimization as it runs, make it easy to adapt the approach to your own model. It is a focused interpretability library rather than a general framework, which suits anyone who just needs to see inside a trained vision model.