pytorch-grad-cam

Why we included this project

Training or shipping a PyTorch vision model almost always means answering one question at some point: what is the network actually looking at? This library packages a large collection of pixel-attribution methods, from the original GradCAM through newer options like ScoreCAM, LayerCAM, and HiResCAM, behind a consistent API that lets you swap techniques by changing a few lines. It is not limited to image classifiers; the same code handles object detection, semantic segmentation, embedding similarity, and even CLIP-style text prompts, so it is useful for debugging real workflows, not just research notebooks. There is also built-in smoothing, full batched inference, and faithfulness metrics for checking whether the explanations you generate can be trusted. If you do applied computer vision work, this is one of the first libraries to reach for when a model misbehaves.