torch-cam

Why we included this project

When you train a CNN, it helps to know which pixels actually drove a given prediction, and this library answers that question directly. It implements a broad family of class activation mapping methods, from the classic CAM and Grad-CAM through newer variants like Score-CAM, Layer-CAM, and XGrad-CAM, all behind a consistent API. The hooks work with many common architectures out of the box, and you can wire up custom models too, which makes it a handy tool for auditing or debugging vision models before you ship them. There's also a Hugging Face Space and a Colab quickstart if you want to see the maps before adding the dependency. For teams building image classifiers, it's a practical way to sanity-check what the model focuses on and to explain a prediction to stakeholders.