/DeepSeer

Code for CHI'23 - DeepSeer

Primary LanguagePythonMIT LicenseMIT

DeepSeer: Interactive RNN Explanation and Debugging via State Abstraction

basic view

DeepSeer is an interactive system for developers to interpret and debug recurrent (stateful) neural networks. This repository contains the official implementation of our related paper:

DeepSeer: Interactive RNN Explanation and Debugging via State Abstraction

Zhijie Wang, Yuheng Huang, Da Song, Lei Ma, Tianyi Zhang

2023 ACM CHI Conference on Human Factors in Computing Systems (CHI 2023)

Getting Started

If you only want to try our tool with provided models, you can easily follow the instruction below to setup. Note that, only necessary scrips about state abstraction techniques are included in the backend (server/deepstellar_backend).

If you want to use DeepSeer with your own model, please follow the folder ml_dev for more details.

Environments Set-up

Python >= 3.6

We suggest use virtual environment to avoid messing up your own environments.

$ python -m venv ./venv
$ source ./venv/bin/activate
$ pip install -r ./server/deepstellar_backend/requirements.txt

NPM >= 7

$ cd ./web/dashboard
$ npm install

Basic Usage

Quick start

$ npm start

Copy the url and open it in browser.

Start backend

$ source ./venv/bin/activate
$ cd ./server/deepstellar_backend
$ python server.py

Switch between different models

The main branch was set to a binary classification RNN model trained on Toxic dataset. To switch, please checkout different branches (remember to restart backend).

Citation

If you found our paper/code useful in your research, please consider citing:

@inproceedings{wang2023deepseer,
 author = {Wang, Zhijie and Huang, Yuheng and Song, Da and Ma, Lei and Zhang, Tianyi},
 title = {DeepSeer: Interactive RNN Explanation and Debugging via State Abstraction}, 
 year = {2023}, 
 isbn = {9781450394215}, 
 publisher = {Association for Computing Machinery}, 
 address = {New York, NY, USA},
 doi = {10.1145/3544548.3580852},
 booktitle = {Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems}, 
 articleno = {740}, 
 numpages = {20}, 
 location = {Hamburg, Germany}, 
 series = {CHI '23}
} 

License

This project is released under the MIT license.

Acknowledgement

Special thanks to our pilot study participants for helping us improve DeepSeer.