Bayes_Opt

Much of my work is still in private repositories not readily visible. However, my boss has kindly allowed me to post some of the code I've written so I can get some credit before the papers are published. This is one of those repositories.

Bayesian Optimization

This is an implementation of Bayesian Optimization, which functions by approximating our true response curve with a Gaussian process (random function). This is referred to as a surrogate model. As more and more data points are sampled we achieve a better approximation to the response curve.

The other component of Bayesian optimization is the acqusition function, which decides where we should next sample from. This seeks to balance exploration (maybe radically new parameters yield better results) with exploitation (fine-tuning parameters that have worked well previously. The combination of the surrogate model and acquisition function determines how well the model performs. Different choices of parameters and functions can radically change Bayesian optimization from quickly converging to the best value to being worse than random sampling. The only way to determine the best Bayes opt parameters is empirically.

Tutorials:

For the mathematically minded this paper provides a good summary of Bayesian optimization:

https://arxiv.org/pdf/1807.02811.pdf

For those less mathematically inclined

https://towardsdatascience.com/the-beauty-of-bayesian-optimization-explained-in-simple-terms-81f3ee13b10f

Contributors

The initial implementation/repository design was done by Austin Talbot (firstname . lastname 1993 @ gmail). Corey Keller has the final say on any decisions regarding this repository.

Dependencies

The dependencies are listed in requirements_botorch.txt, but listed here they are

1. jupyter - allows for Jupyter notebooks to demonstrate functioning and store results
2. gym - allows for interactive simulations
3. tqdm - trange allows for progress measurement in for loops
4. botorch - Facebook's Bayesian optimization implementation
5. matplotlib - visualization library analagous to MATLAB (MATplotlib)
6. torch - pytorch, basis of botorch
7. torchvision - potentially unnecessary
8. torchaudio - potentially unnecessary

Setup instructions

Make sure anaconda is downloaded, with instructions found at

https://www.anaconda.com/

Having installed Anaconda, you can create a conda environment to use the scripts. This can be done in terminal using the following commands:

conda create -n bayes_opt python=3.10
conda activate bayes_opt
pip install -r requirements_botorch.txt

Having installed the necessary packages you can now use this environment whenever you want to run the code. Type

conda activate bayes_opt

to open this environment and type

conda deactivate

once you have concluded your work. This will prevent you from accidentally installing extra packages in the environment that can potentially break the implementation.

Components

Code - this directory has the Bayesian optimization code. At this moment the only implementation uses BoTorch.

Demos - this directory has a series of examples, demonstrating how the code can be used. This will be similar to the scripts in UnitTests, except it will be nicely formatted so that anyone can understand how to use these scripts with minimal python experience.

UnitTests - this directory has a list of unit tests to ensure proper functioning of the code. These are example scripts where the behavior is known by default to ensure proper functioning of any code additions/modifications.

Version control and development practices

Every script in UnitTests/SmokeTest should be run before commiting to the main repository. This will ensure that the basic functions compile and that changes do not affect basic structure. This is simple to do on MACOS or Linux! In Terminal simply type the commands

chmod 777 runAll.sh
./runAll.sh

After you do this once the first line is unnecessary! There is simply no excuse for not running this every time before a commit. It will print a bunch of text and if

!!!!!!!!!!!
!!!!!!!!!!!
Test Failed
!!!!!!!!!!!
!!!!!!!!!!!

appears anywhere in the output figure out what's wrong before commiting!

Every commit should include the version number (which should be incremented, and the associated string version included in the code files changed accordingly). Therefore, when any saved model is loaded it is straightforward to see which version of the code that model used.

Code documentation

The code should be well-commented, meaning that the comments in the code should be informative enough that anyone even without a technical background should be able to understand how the code works by the comments alone. The first lines in every file should be a triple-quoted string that documents the file so that anyone can type

import filename
print(filename.__doc__)

and have the contents of that file summarized. It should include a paragraph(s) summary, a list of any objects, methods, a contributor list, and a version history.

Any function should also have its first line be a triple-quoted string so that the lines of code

from filename import func
from filename import ExampleObject
myObject = ExampleObject(params)
print(func.__doc__)
print(myObject.method.__doc__)

provide information on what the method does. It should provide a line summary of the method purpose, a list of parameters (if any), and what the method returns. All of these should have type and default value (if any) included.

Any object methods that are not meant to be used by a biologist should begin with an underscore.

Every single object should have the __repr__ implemented so that the line

print(myObject)

should actually provide meaningful information to the user. At a minimum it should provide the version number and creation date of the object.