rbtsbg/pgig

Pattern-Guided Integrated Gradients

(This repo is based on the visual-attribution (07c79...) frame work by Yulong Wang (2018).)

Code accompanying the paper Pattern-Guided Integrated Gradients, presented at the ICML 2020 Workshop on Human Interpretability in MachineLearning (WHI 2020) by Robert Schwarzenberg and Steffen Castle (equal contribution):

@inproceedings{pgig2020,
  title={Pattern-Guided Integrated Gradients},
  author={Schwarzenberg, Robert and Steffen Castle},
  booktitle={Proc. of the ICML 2020 Workshop on Human Interpretability in Machine Learning (WHI)},
  year={2020}
}

For a quick overview, see the synthetic experiments.

Addendum: We now also present arguments against the implementation invariance of PA and PGIG.

Input	Integrated Gradients	PatternAttribution	Pattern-Guided Integrated Gradients

Requirements

We ran the experiments on Ubuntu 16.04.6 LTS, with Python 3.6.9 installed, on GeForce GTX 1080 Tis, using cuda10.0+cudnn7.6.2

The minimal requirements are

torch==1.2.0
torchvision==0.4.0a0+6b959ee

scikit-learn==0.22.1

All other dependencies are listed in dependencies.txt

Data

Weights and Patterns

Weights and patterns can be downloaded using the download_patterns.sh script in the weights directory. In case the URLs become invalid, there are backups here.

Degradation Experiment

The degradation experiment can be executed by running experiments/degradation_test.py.

ImageNet

To run experiments, the validation set of ImageNet 2012 must already be downloaded and extracted locally, and the ImageNet directory must be specified using the --data_dir argument. The --imagenet_download_key argument and functionality to automatically download the ImageNet dataset is deprecated.

Arguments

Argument	Description
--methods	List of methods to evaluate
--batch_size	Batch size, default is 50
--data_dir	Directory where imagenet validation set is downloaded, or where to download it
--val_size	Number of samples to evaluate, default is the whole ImageNet validation set
--n_patches	Number of patches to degrade, default is 100
--imagenet_download_key	Optional URL string for ImageNet download (deprecated)

Available methods

Method	Method argument string
Vanilla Gradient	vanilla_grad
PatternAttribution	pattern_vanilla_grad
Gradient times Input	grad_x_input
Integrated Gradients	integrate_grad
Guided Backprop	guided_backprop
SmoothGrad	smooth_grad
SmoothGrad2	smooth_grad2
SmoothGrad-Integrated Gradients	sg_ig
VarGrad	var_grad
Expected Gradients	exp_grad
Pattern-Guided Integrated Gradients	pattern_integrate_grad

Example

python3 experiments/degradation_test.py --methods vanilla_grad pattern_vanilla_grad integrate_grad pattern_integrate_grad --batch_size 100 --val_size 10000 --n_patches 100 --data_dir /mnt/hdd/datasets/imagenet/

References

[1] Kindermans, Pieter-Jan, et al. "Learning how to explain neural networks: Patternnet and patternattribution." arXiv preprint arXiv:1705.05598 (2017).

[2] Sundararajan, Mukund, Ankur Taly, and Qiqi Yan. "Axiomatic attribution for deep networks." arXiv preprint arXiv:1703.01365 (2017).