grad-cam

Why we included this project

This is the reference implementation of Grad-CAM, the method that turns a CNN's decision into a heatmap showing which image regions drove the prediction. Written in Torch/Lua and built around Caffe models, it is best treated as a historical and pedagogical resource rather than something to drop into a modern PyTorch or TensorFlow stack. The repo covers classification, visual question answering, and captioning, with simple command-line flags for selecting a layer and saving heatmaps, which makes it easy to experiment with the technique. For anyone studying interpretability or reproducing the ICCV 2017 results, having the authors' own code is a convenient, readable way to see how the method works before porting it.