https://github.com/mlech26l/bayesian_nn_safety
Code repository of the NeurIPS 2021 paper Infinite Time Horizon Safety of Bayesian Neural Networks
https://github.com/mlech26l/bayesian_nn_safety
bayesian-deep-learning bayesian-neural-networks neural-networks
Last synced: about 1 month ago
JSON representation
Code repository of the NeurIPS 2021 paper Infinite Time Horizon Safety of Bayesian Neural Networks
- Host: GitHub
- URL: https://github.com/mlech26l/bayesian_nn_safety
- Owner: mlech26l
- License: apache-2.0
- Created: 2021-10-07T06:00:27.000Z (over 3 years ago)
- Default Branch: main
- Last Pushed: 2022-04-11T05:52:12.000Z (about 3 years ago)
- Last Synced: 2025-02-09T09:14:04.237Z (3 months ago)
- Topics: bayesian-deep-learning, bayesian-neural-networks, neural-networks
- Language: Python
- Homepage: https://proceedings.neurips.cc/paper/2021/file/544defa9fddff50c53b71c43e0da72be-Paper.pdf
- Size: 3.81 MB
- Stars: 0
- Watchers: 2
- Forks: 1
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# Infinite Time Horizon Safety of Bayesian Neural Networks
Code repository of the paper *Infinite Time Horizon Safety of Bayesian Neural Networks*, by Lechner, Mathias, Đorđe Žikelić, Krishnendu Chatterjee, and Thomas Henzinger published at NeurIPS 2021 [PDF](https://proceedings.neurips.cc/paper/2021/file/544defa9fddff50c53b71c43e0da72be-Paper.pdf)
## Requirements
```bash
pip3 install -U pip # Upgrade pip
pip3 install -U gym tensorflow mip seaborn gym
```
Tested on Ubuntu 16.04 and Ubuntu 20.04 and Python 3.8.
Running the experiments requires a valid Gurobi license (freely available MIP solvers scale only to the LDS example).
Make sure Gurobi is installed and added to the ```PATH``` and ```LD_LIBRARY_PATH``` environment variables
```bash
export GUROBI_HOME="/xxxxxx/gurobi/903/linux64/"
export PATH="${PATH}:${GUROBI_HOME}/bin"
export LD_LIBRARY_PATH="${LD_LIBRARY_PATH}:${GUROBI_HOME}/lib"
```
## Files
- ```lds_env.py``` Implementation of the LDS environment
- ```pend_env.py``` Implementation of the inverted pendulum environment
- ```car_env.py``` Implementation of the collision avoidance environment
- ```run_mip.py``` Implementation of the verification loop (algorithm 1 in the paper) for the LDS and the inv-pend environment
- ```mixed_integer_encoding.py``` Mixed-integer encoding module for ReLU networks according to Tjeng et al. (Evaluating Robustness of Neural Networks with Mixed Integer Programming)
- ```run_car_mip.py``` Implementation of the verification loop (algorithm 1 in the paper) for the collision avoidance environment (mostly equivalent to the file ```run_mip.py```)
- ```weights``` Folder containing the pre-trained weights of the policies
## Example usage
```bash
# Runs the verification loop with the LDS environment
python3 run_mip.py --env lds
# Runs the verification loop with the LDS environment and bootstrapping the invariant candidate network
python3 run_mip.py --env lds --bootstrap
# Runs the verification loop with the inv-pend environment with Bayesian weights in all layers
python3 run_mip.py --env pend --first_layer
# Runs the verification loop with the collision avoidance environment
python3 run_car_mip.py --bootstrap
```
## Additional information
The networks with Bayesian weights in the second hidden layer only have a Guassian distribution with a standard deviation of 0.1, while the networks with both layers Bayesian have a standard deviation of 0.05.
Here is an example of how the positive invariant changes during the learner-verifier training procedure:
## Citation
```bib
@inproceedings{lechner2021infinite,
author = {Mathias Lechner and Dorde Zikelic and Krishnendu Chatterjee and Thomas A. Henzinger},
title = {Infinite Time Horizon Safety of Bayesian Neural Networks},
booktitle = {In Conference on Neural Information Processing Systems (NeurIPS)},
year = {2021}
}
```