This project involves the implementation of CCBO based on Gaussian Process (GP) models using GPyTorch and BoTorch for guided experimentation.
This code repository includes two notebooks demonstrating the implementation of CCBO for benchmarking and guiding lab experiment. For detailed information, see the Getting Started section.
CCBO requires python environment to work. To get started with CCBO, follow these steps:
The source code provided for CCBO does not require any specific hardware. For ideal performance, we recommend RAM of 16 GB and at least a 4-core CPU.
The code haven been tested on macOS Sonoma (14.4.1), Windows 10/11, and Linux (Ubuntu 16.04). It is recommended to have package management software such as Conda.
To run the notebooks and scripts in this project, you will need the following Python dependencies:
python = 3.9botorch == 0.10.0gpytorch == 1.11numpy == 1.23pytorchpandasseabornjoblib- (optional)
tqdm
Note: NumPy versions above 2.0 are not supported by BoTorch/PyTorch (2024.11.26)
The typical installation time should be less than 10 min. To begin, install the required dependencies by running the following commands in your Terminal (make sure you have Conda already installed).
Step 1: Create a new virtual environment:
conda create -n ccbo python=3.9
conda activate ccboStep 2: Install dependencies
-
Option A - Install with Conda (tested on Win/Linux OS):
- Install PyTorch first (here is an example of CPU only version of PyTorch on Win)
conda install pytorch cpuonly -c pytorch
- Then, install BoTorch and dependencies
conda install botorch==0.10 seaborn joblib pandas -c gpytorch -c conda-forge
-
Option B - Install with pip (tested on Mac)
- Use
pipto easily install BoTorch and all dependencies. Be aware to use the pip from the new virtual environment, you can refer to this post to make sure you did not install BoTorch into your global Python.
python -m pip install botorch==0.10.0 seaborn numpy==1.23 pandas joblib
- Use
Step 3: clone this repository to your local machine:
git clone https://github.com/FrankWanger/CCBO.gitThere are two demo notebooks to reproduce the experiments included in the manuscript. If you are not familiar with Jupyter notebooks, please refer to this guide.
CCBO_benchmark.ipynb A notebook showing benchmarking CCBO against vanilla BO, constrained BO, and random baseline with a synthetic electrospray problem. The same initial dataset has been provided in the notebook.
The parameters that are related to reproduce the results in the manuscript are:
ITERS, the total number of iterations, the default is10TRIALS, the number of repetition (of the iterations) to obtain confidence interval, the default is 20PARA_EVAL, a boolean that sets whether to usejoblibto parallel the different trials. This will significantly affect benchmark time, the default isTrue.
Notably, the run time for reproducing the benchmark result with the default parameters on a PC with 8-core CPU (i7-10700F, 2.9 GHz) and 32 GB RAM was ~1 h.
CCBO_guide_exp.ipynb An example notebook for using CCBO to guide laboratory electrospray experiments, with SOBEL initialization. The wet-lab experiment results were accompanied to the notebook. Albeit efforts have been made to set all random seeds in PyTorch and Numpy libraries, the authors are aware that the individual experiment recommended may vary between different installations of BoTorch. We believe that this does not weaken the conclusions as the benchmark results have been tested to show consistency across platforms.
There are a few core and helper functions developed in ccbo/core.py
sim_espray_constrained: The simulation function for the constrained electrospray problem.optimize_acqf_and_get_recommendation: The main function to optimize the acquisition function and get the next experiment recommendation.
This project is licensed under the MIT License. See the LICENSE file for more details.
For any questions or issues, please contact Fanjin Wang at [email protected]