Integrated Gradients

This is the pytorch implementation of "Axiomatic Attribution for Deep Networks".

Requirements

  • python-3.5.2
  • pytorch-0.4.1
  • opencv-python

TODO List

  • [ ] add more functions as the original code.
  • [ ] finetune the results, make them close to the original paper.

Instructions

Highly recommend to use GPU to accelerate the computation. If you use CPU, I will recommend to select some small networks, such as resnet18. You also need to put your images under examples/.

Lists of networks that support (of course, you can add any networks by yourself)

  • inception
  • resnet18
  • resnet152
  • vgg19

Run the code

python main.py --cuda --model-type='inception' --img='01.jpg'

Results

Results are slightly different from the original paper, it may have some bugs or need to do some adjustments. I will keep updating it, any contributions are welcome!

Inception-v3

inception

ResNet-18

resnet18

ResNet-152

resnet152

VGG-19

vgg19

GitHub

https://github.com/TianhongDai/integrated-gradient-pytorch