Skip to content

Code for the experiments of the ICML 2021 Interpretability workshop paper "This Looks Like That... Does it? Shortcomings of Latent Space Prototype Interpretability in Deep Networks"

Notifications You must be signed in to change notification settings

fanconic/this-does-not-look-like-that

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

51 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

This Looks Like That... Does it? Shortcomings of Latent Space Prototype Interpretability in Deep Networks

This repository contains code for reproducing experiments and models from the paper "This Looks Like That... Does it? Shortcomings of Latent Space Prototype Interpretability in Deep Networks" by Adrian Hoffmann*, Claudio Fanconi*, Rahul Rade* and Jonas Kohler.

A short version of the paper has been accepted at the ICML 2021 Workshop on Theoretic Foundation, Criticism, and Application Trend of Explainable AI.

Head-On-Stomach Experiment JPEG Experiment

Setup

Installation

Clone this repository.

$ git clone https://gitlab.ethz.ch/fanconic/this-does-not-look-like-that
$ cd this-does-not-look-like-that

We suggest to create a virtual environment and install the required packages.

$ conda create --name experiment_env
$ conda activate experiment_env
$ conda install --file requirements.txt

Dataset

Download the CUB-200-2011 dataset from http://www.vision.caltech.edu/datasets/cub_200_2011/ and extract the files in a suitable folder.

Repository Structure

  • run_training.sh: Train ProtoPNet on CUB-200-2011 dataset.
  • run_jpeg_training.sh: Train ProtoPNet on altered CUB-200-2011 dataset.
  • local_analysis_attack1.ipynb: Head-On-Stomach Experiment.
  • JPEG_experiment_analysis: JPEG Experiment.
  • local_analysis_attack3.ipynb: Attack 2 (Make head disappear).
  • adv_training_attack1.ipynb: Head-On-Stomach Experiment Success Rate.
  • run_fgsm_training.sh : Remedy for Head-On-Stomach Experiment (Train ProtoPNet on CUB-200-2011 dataset via fast FGSM training).
  • run_jpeg_shuffled.sh : Remedy for JPEG Experiment (Train ProtoPNet on CUB-200-2011 dataset with JPEG augmentation).

Source Code Directory Tree

.
└── src                 # Source code for the experiments
    ├── data                # Data setup, preprocessing, augmentation 
    ├── models              # ProtoPNet and various features of it
    ├── training            # Training and test scripts
    └── utils               # Helpers, visualization and local analysis

Training ProtoPNets

Please set the variable base_architecture in settings.py to the backbone which you want to use for the ProtoPNet. Also, set the path to the downloaded dataset in src/data/setup.py.

Training ProtoPNets used in the Head-On-Stomach experiment.

$ ./run_training.sh

Training ProtoPNets used in the JPEG experiment.

$ ./run_jpeg_training.sh

Reproducing Results of Our Experiments

To recreate the results of the experiments, you can use the provided Jupyter Notebooks.

Download the pretrained model weights of both experiments from here: Google Drive

Finally, follow the instructions in the provided notebooks to reproduce our experiments.

Contributors

Reference

@misc{hoffmann2021looks,
    title={This Looks Like That... Does it? Shortcomings of Latent Space Prototype Interpretability in Deep Networks}, 
    author={Adrian Hoffmann and Claudio Fanconi and Rahul Rade and Jonas Kohler},
    year={2021},
    booktitle={ICML 2021 Workshop on Theoretic Foundation, Criticism, and Application Trend of Explainable AI},
}

Acknowledgement

This code is partly borrowed from cfchen-duke/ProtoPNet. We would like to thank the authors of ProtoPNet for making their code openly available.

About

Code for the experiments of the ICML 2021 Interpretability workshop paper "This Looks Like That... Does it? Shortcomings of Latent Space Prototype Interpretability in Deep Networks"

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 4

  •  
  •  
  •  
  •