This is the official repository accompanying the NeurIPS 2021 paper:

R. Hesse, S. Schaub-Meyer, and S. Roth. Fast axiomatic attribution for neural networks. NeurIPS, 2021.

Paper | Preprint (arXiv) | Project Page | Video

The repository contains:

• Pre-trained -DNN (X-DNN) variants of popular image classification models obtained by removing the bias term of each layer
• Detailed information on how to easily compute axiomatic attributions in closed form for your own project
• PyTorch code to reproduce the main experiments in the paper

Removing the bias from different image classification models has a surpringly minor impact on the predictive accuracy of the models while allowing to efficiently compute axiomatic attributions. Results of popular models with and without bias term (regular vs. X-) on the ImageNet validation split are:

In the following we illustrate how to efficiently compute axiomatic attributions for X-DNNs. For a detailed example please see demo.ipynb.

First, make sure that requires_grad of your input is set to True and run a forward pass:

inputs.requires_grad = True

# forward pass
outputs = model(inputs)

Next, you can compute X-Gradient via:

# compute attribution
target_outputs = torch.gather(outputs, 1, target.unsqueeze(-1))
gradients = torch.autograd.grad(torch.unbind(target_outputs), inputs, create_graph=True)[0] # set to false if attribution is only used for evaluation
xgradient_attributions = inputs * gradients

If the attribution is only used for evaluation you can set create_graph to False. If you want to use the attribution for training, e.g., for training with attribution priors, you can define attribution_prior() and update the weights of your model:

loss1 = criterion(outputs, target) # standard loss
loss2 = attribution_prior(xgradient_attributions) # attribution prior    

loss = loss1 + lambda * loss2 # set weighting factor for loss2

optimizer.zero_grad()
loss.backward()
optimizer.step()

The code and a README with detailed instructions on how to reproduce the results from experiments in Sec 4.1, Sec 4.2, and Sec 4.4. of our paper can be found in the imagenet folder. To reproduce the results from the experiment in Sec 4.3. please refer to the sparsity folder.

• Clone the repository: git clone https://github.com/visinf/fast-axiomatic-attribution.git
• Set up environment
  • add the required conda channels and create new environment:
  • conda config --add channels pytorch
  • conda config --add channels anaconda
  • conda config --add channels pipy
  • conda config --add channels conda-forge
  • conda create --name fast-axiomatic-attribution --file requirements.txt
• download ImageNet (ILSVRC2012)

We would like to thank the contributors of the following repositories for using parts of their publicly available code:

• https://github.com/suinleelab/attributionpriors
• https://github.com/pytorch/examples/tree/master/imagenet
• https://github.com/hongyi-zhang/Fixup

If you find our work helpful please consider citing

@inproceedings{Hesse:2021:FAA,
  title     = {Fast Axiomatic Attribution for Neural Networks},
  author    = {Hesse, Robin and Schaub-Meyer, Simone and Roth, Stefan},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  volume    = {34},
  year      = {2021}
}