Grad-Cam and Grad-Cam++ implemented in tf.keras 2.X (tensorflow 2.X).
Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization by Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, Dhruv Batra (https://arxiv.org/abs/1610.02391).
Grad-CAM++: Generalized Gradient-based Visual Explanations for Deep Convolutional Networks by Aditya Chattopadhyay, Anirban Sarkar, Prantik Howlader, Vineeth N Balasubramanian (https://arxiv.org/abs/1710.11063).
Adapted and optimized code from https://github.com/totti0223/gradcamplusplus.
Resolve the problem of using eager mode in tf.keras, and almost follow the formula in grad-cam++ paper.
For more results, check the images in results folder.
- Execute following command in terminal:
git clone https://github.com/samson6460/tf.keras-gradcamplusplus.git
cd tf.keras-gradcamplusplus
- Create a new python file and import
utilsandgradcam. - Pass your model and image array to
grad_cam()orgrad_cam_plus()func, and it will return a heatmap. - Pass image path and heatmap to
show_imgwithheat()func, and it will show a superimposed image.
Here's an example model that can classify bone X-rays into three categories: wrist, shoulder and elbow based on VGG16.
The model was pretrained on ImageNet and finetuned on MURA dataset.
Get the model by calling vgg16_mura_model(destination_path).If it's the first time it will download the weights automatically.
Get the MURA(musculoskeletal radiographs) dataset from https://stanfordmlgroup.github.io/competitions/mura/.
Or test the model with no copyright images in images folder.
Images source:
Run example.py, you will understand more.