Chainer implementation of Grad-CAM [1]. Grad-CAM can localize and highlight important region in the image for predicting the concept without changing the model architecture. Currently, this implementation supports AlexNet, VGGNet and ResNet.
| Grad-CAM | Guided Backpropagation | Guided Grad-CAM | |
|---|---|---|---|
| Boxer (242) |  |
 |
 |
| Tiger Cat (282) |  |
 |
 |
- Chainer
- Cupy (for GPU support)
- OpenCV
python run.py --input images/dog_cat.png --label 242 --layer conv5_3 --gpu 0
python run.py --input images/dog_cat.png --label 282 --layer conv5_3 --gpu 0
- [1] Ramprasaath R. Selvaraju, Abhishek Das, Ramakrishna Vedantam, Michael Cogswell, Devi Parikh, Dhruv Batra, "Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization", https://arxiv.org/abs/1610.02391