Captum

Captum is a PyTorch library dedicated to model interpretability — helping developers and researchers understand exactly why their models make the predictions they do. It provides a comprehensive set of tools and techniques to attribute model predictions across various modalities, including vision and text, supporting most PyTorch model architectures out of the box. Captum makes it straightforward to implement and benchmark new interpretability algorithms, making it the go-to resource for anyone working on trustworthy, transparent AI. Whether you're debugging a model, building explainable AI systems, or conducting research into how neural networks reason, Captum gives you the visibility you need.