mnist-Grad-CAM

contents

I implemented Grad-CAM and applied it to mnist datasets. if you find some bugs or anything else, feel free to open issues or PRs.

environment

Dockerfile and docker-compose.yml are settled in docker/cpu and docker/gpu. They will make the executable environment on your machine by using docker. when you use GPU, you can choose docker/gpu dir, otherwise you have to use docker/cpu. The docker/gpu contents were well worked in my env,

Ubuntu:16.04 
Geforce 1080Ti

If you find some issues, please tell it to me via issues.

docker/cpu worked fine on my local machine, which is Mac Book Pro 2016 Mid 2015.

usage

cd docker/gpu(or docker/cpu)
docker-compose up -d --build
docker-compose exec tf bash
# train
python mnist.py
# visualization: my_model.h5 created by the command above.
python mnist_visualize.py

the visualized image will be putted in mnist_cams dir, which will be automatically created.

currently, this directory and my_model.h5 names cannot be fixed by argments.
it is my future work.

#examples

original

heatmap

heatmap overrayed

guided grad cam figure

reference

CNN structure

https://github.com/keras-team/keras/blob/950e5d063320a72ca61f2082c154a65a48766239/examples/mnist_cnn.py

Grad-CAM implements

https://github.com/insikk/Grad-CAM-tensorflow
https://github.com/jacobgil/keras-grad-cam
https://github.com/ysasaki6023/imageCounting