README.md

Structured Gradient-based Interpretations via Norm-Regularized Adversarial Training

Implementation for CVPR 2024 paper Structured Gradient-based Interpretations via Norm-Regularized Adversarial Training by Shizhan Gong, Qi Dou, and Farzan Farnia.

Sample Results

Setup

Our code is easily runable on most pytorch environment without special dependency. Or readers can install the environment via pip install -r requirements.txt.

Dataset

Imagenette

The data can be downloaded from the official page. To use our dataloader, put the data split file dataset/Imagette/train.csv (val.csv) under the data directory and organize the folder as

imagenette2/
	└── train/ 
	└── val/
	└── train.csv
	└── val.csv

CUB-GHA

The data can be downloaded from the official page. To use our dataloader, put the data split file dataset/CUB-GHA/train.csv (test.csv) under the data directory and organize the folder as

CUB_200_2011/
	└── images/ 
	└── CUB_GHA/
	└── train.csv
	└── test.csv

Training

Training on Imagenette

python imagenet_train.py --mode elastic --training fast --epsilon 0.01 --epsilon2 0.05 --data_dir ./imagenette2 --snapshot_path checkpoint

--mode: the variants of adversarial training, can be one of standard, l1, elastic, group.

--training: to use ietrative optimization or one-step optimization, can be fast or iterative.

--epsilon: the coefficient $\epsilon$ in l1 norm and group norm mode, or $\epsilon_1$ in elastic net mode.

--epsilon2: the coefficient $\epsilon_2$ in elastic net mode.

--data_dir: directory of the training data.

--snapshot_path: directory to store the trained model.

Training on CUB-GHA

python cub_train.py --epsilon 0.5 --data_dir ./CUB_200_2011 --snapshot_path checkpoint_cub

Interpretation

We use simple gradient for interpretation.

from utils import simplegrad, agg_clip

saliency_map = simplegrad(net, x, label)
saliency_map = agg_clip(saliency_map)
plt.imshow(saliency_map)

where net is the trained neural network, x is an image of shape [1,3,224,224] and label is can be ground truth label or the predicted label.

Bibtex

If you find this work helpful, you can cite our paper as follows:

@inproceedings{gong2024structured,
  title={Structured Gradient-based Interpretations via Norm-Regularized Adversarial Training},
  author={Gong, Shizhan and Dou, Qi and Farnia, Farzan},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={11009--11018},
  year={2024}
}