This is the pytorch implementation of "Axiomatic Attribution for Deep Networks". The original tensorflow version could be found here.
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/
.
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](figures/inception.png)
### ResNet-18
![resnet18](figures/resnet18.png)
### ResNet-152
![resnet152](figures/resnet152.png)
### VGG-19
![vgg19](figures/vgg19.png)