kirthevasank/gp-parallel-ts

gp-parallel-ts

This is a python implementation of parallelised Bayesian optimisation via Thompson sampling. For more details, see our paper below.

Download

You can download the code from github

$ git clone https://github.com/kirthevasank/gp-parallel-ts

Installation & Getting Started

Run source set_up_thompson to set up all environment variables.
You also need to build the direct fortran library. For this cd into direct_fortran and run bash make_direct.sh. You will need a fortran compiler such as gnu95. Once this is done, you can run simple_direct_test.py to make sure that it was installed correctly.
To test the installation, run bash run_all_tests.sh. Some of the tests are probabilistic and could fail at times. If this happens, run the same test several times and make sure it is not consistently failing.

Demo

Check demo_simtime.py for a demo on how to run experiments with simulated time values (see paper below for experiments).
We will also include a demo for real time experiments soon, but for now you can check out resnet_experiments/resnet_function_caller.py to see how to set it up. You need to use the RealWorkerManager class under bo/worker_manager.py

Some notes

We choose the GP hyper-parameters every 25 iterations via marginal likelihood maximisation for each GP. The chosen values are printed out.
We report progress on the optimisation every 20 iterations.

Citation

If you use any part of this code in your work, please cite our Arxiv paper:

@article{kandasamy2017asynchronous,
  title={Asynchronous Parallel Bayesian Optimisation via Thompson Sampling},
  author={Kandasamy, Kirthevasan and Krishnamurthy, Akshay and Schneider, Jeff and Poczos, Barnabas},
  journal={arXiv preprint arXiv:1705.09236},
  year={2017}
}

License

This software is released under the MIT license. For more details, please refer LICENSE.txt.

For questions and bug reports please email kandasamy@cs.cmu.edu

kirthevasank / gp-parallel-ts