Installation

Requires Python 3.12 or above. Would recommend JAX 0.4.17 because 0.4.18 to 0.4.23 appears to sometimes yield CUDA internal errors randomly at runtime.

User

Internet

pip install git+https://github.com/cemlyn007/gpbo.git#egg=gpbo

Locally

Clone the repository at https://github.com/cemlyn007/gpbo and run:

pip install .

Development

Clone the repository at https://github.com/cemlyn007/gpbo and run: pip install -e .[dev] If using Visual Studio Code with Pylance, you will want to run instead:

pip install -e .[dev] --config-settings editable_mode=strict

To get the type hinter to find the package.

Examples

You will need to install the example dependencies by running: pip install -e .[example]

Optimizing a Gaussian Process

Univariate

python examples/gp.py --objective_function=univariate --use_x64 --iterations 48

Six Hump Camel

python examples/gp.py --objective_function=six_hump_camel --use_x64 --iterations 256

MNIST Log Learning Rate vs Negative Accuracy

python examples/gp.py --objective_function=mnist_1d --use_x64 --iterations 256

MNIST Log Learning Rate and Log Momentum vs Negative Accuracy

python examples/gp.py --objective_function=mnist_2d --iterations 135

Bayesian Optimization

Univariate

python examples/bo.py --objective_function=univariate --use_x64 --iterations 16

Six Hump Camel

python examples/bo.py --objective_function=six_hump_camel --use_x64 --iterations 48

Generate Results

All commands are expected to be run from the root of the repository.

Via Shell

Run this command to generate the MNIST dataset:

./scripts/make_data.sh

Run this command to generate the results, passing gpu or cpu as the first argument to specify the device to use:

./scripts/run_experiments.sh gpu

To export the 3D results to be visualised using VTK via something like Paraview:

./scripts/export_3d.sh

Via Slurm

Run this command to generate the MNIST dataset:

sbatch --chdir=$(pwd) ./scripts/make_data.sh

Run this command to generate the results, passing gpu or cpu as the first argument to specify the device to use:

sbatch --chdir=$(pwd) ./scripts/run_experiments.sh gpu