Interpretable Concept groups on CNNs

Introduction

This repository is based upon NetDissect, which contains the demo code for the work Network Dissection: Quantifying Interpretability of Deep Visual Representations.

The code is written in pytorch and python3.6, tested on Ubuntu 16.04. Please install Pytorch in python36 and Torchvision first to work with the code.

Download

Clone the code of Network Dissection Lite from github

    git clone https://github.com/srb-cv/cgl.git

Training

In order to see all the flags to train the model:

python Places365/train_places_cnn.py --help

Example: Training an alexnet model on Places 365 model with different L-p norm regularizer, group activation loss, and spaltial loss:

python Places365/train_places_cnn.py /path-to-the-dataset -a alexnet --lr 0.01 --wd 0 --penalty 0.005 --activation-penalty 0 --spatial-penalty 0 --batch-size 256 --num_classes 365 --workers 7 --save /path-to-save-checkpoints --groups 5

Here, the loss function to train a model can be defined as : $$L = L_d(w) + \lambda_{bn} R_{bn}(w) + \lambda_{g}L_g(\psi) + \lambda_{s}L_s(\psi)$$ where,

$\lambda_{bn}$ for the General Block Norm can be set using the flag --penalty
$\lambda_{g}$ for the Group Activation Loss can be set using the flag --activation-penalty
$\lambda_{s}$ for the Spatial Loss can be set using the flag --spatial-penalty

Evaluation

Download the Broden dataset (~1GB space) and the example pretrained model. If you already download this, you can create a symbolic link to your original dataset.

    ./script/dlbroden.sh
    ./script/dlzoo_example.sh

Run NetDissect

    python main.py

At the end of the dissection script, a report will be generated inside result folder that summarizes the interpretable units of the tested network. These are, respectively, the HTML-formatted report, the semantics of the units of the layer summarized as a bar graph, visualizations of all the units of the layer (using zero-indexed unit numbers), and a CSV file containing raw scores of the top matching semantic concepts in each category for each unit of the layer.

Reference

If you find the codes useful, please cite these papers

@inproceedings{ijcai2021-147,
  title     = {Learning Interpretable Concept Groups in CNNs},
  author    = {Varshneya, Saurabh and Ledent, Antoine and Vandermeulen, Robert A. and Lei, Yunwen and Enders, Matthias and Borth, Damian and Kloft, Marius},
  booktitle = {Proceedings of the Thirtieth International Joint Conference on
               Artificial Intelligence, {IJCAI-21}},
  publisher = {International Joint Conferences on Artificial Intelligence Organization},
  editor    = {Zhi-Hua Zhou},
  pages     = {1061--1067},
  year      = {2021},
  month     = {8},
  note      = {Main Track}
  doi       = {10.24963/ijcai.2021/147},
  url       = {https://doi.org/10.24963/ijcai.2021/147},
}

@inproceedings{netdissect2017,
  title={Network Dissection: Quantifying Interpretability of Deep Visual Representations},
  author={Bau, David and Zhou, Bolei and Khosla, Aditya and Oliva, Aude and Torralba, Antonio},
  booktitle={Computer Vision and Pattern Recognition},
  year={2017}
}

=======

Learning Interpretable Concept Groups in CNNs