MuS: Multiplicative Smoothing for Stability Guarantees

[Paper] [Blog post]

Official implementation for "Stability Guarantees for Feature Attributions with Multiplicative Smoothing".

Authors: Anton Xue, Rajeev Alur, Eric Wong.

Usage

Just copy the my_models/mus.py where convenient.

Sample usages found in my_models/imagenet_utils.py and my_models/tweet_utils.py.

Getting Started with Training and Experiments

Create the conda environment.

conda env create -f environment.yml
conda activate mus

Experiment Setup

• Download TweetEval into this directory: https://github.com/cardiffnlp/tweeteval
• Make sure you have ImageNet1K somewhere
• Modify paths in qheader.py as needed
• mkdir -p dump/q1_theory dump/q1_boxatk dump/q2 dump/q3_sparsity
• mkdir -p notebooks/images
• mkdir -p saved_models

Training

To train Vision Transformer with lambd = 0.6 and lambd = 0.3 masking for 2 epochs,

• python3 -i fine_tuning/imagenet_fine_tuning -m vit16 -d /path/to/imagenet/root -s saved_models/ -l 0.6 0.3 -e 2 --go To train Vision Transformer with lambd = 0.6 and lambd = 0.3 masking for 2 epochs,
• python3 -i fine_tuning/imagenet_fine_tuning -m resnet50 -d /path/to/imagenet/root -s saved_models/ -l 0.6 0.3 -e 2 --go To train TweetEval classifiers with lambd = 0.6 and lambd = 0.3 masking for 2 epochs, do:
• python3 -i fine_tuning/tweet_fine_tuning.py -s saved_models/ -l 0.6 0.3 -e 2 --go

Experiment Scripts

The following scripts will put data into the dump/ directory. Some of these experiments take a long time, and the data in dump/ have been pre-generated.

• E1: Code is in q1_theory.py and q1_boxatk.py.
  • python3 -i q1_theory.py and run:
    • q1t_run_stuff(configs)
  • python3 -i q1_boxatk.py and run:
    • q1e_run_stuff("vit16", configs)
    • q1e_run_stuff("resnet50", configs)
    • q1e_run_stuff("roberta", configs)
• E2: Code is in q2.py
  • python3 -i q2.py and run:
    • q2_run_stuff(configs)
• E3: Code is in q3_sparsity.py
  • python3 -i q3_sparsity and run:
    • q3_run_stuff(configs)

Generate Plots

• Load up notebooks/neurips_plots.ipynb and run the entire file.

Citation

If you find our work helpful, please cite:

@article{xue2023stability,
  title={Stability Guarantees for Feature Attributions with Multiplicative Smoothing},
  author={Xue, Anton and Alur, Rajeev and Wong, Eric},
  journal={arXiv preprint arXiv:2307.05902},
  year={2023}
}