DeepHV Code

Installing required packages

We provide a list of used packages in environment.yaml. If you run a linux based operating system you can install the environment using:

conda env create -f environment.yaml

Generating a dataset

Code used to generate the datasets can be found in generate_data_geometric.py To generate a dataset, set parameters in main() function and run:

python generate_data_geometric.py

Not that this code can be lengthy to run for high dimensional cases. We also provide faster matlab code, however this requires that you have Matlab and the Matlab Python SDK installed. We provided a (zipped) dataset for the setting of M=5.

Training a model

To train a model run:

python run_mape_batched.py 5 128

This will train a model for M=5 with 128 channels per layer. It will either load in the existing dataset if it's present, or generate the dataset. We have provided this dataset in the processed directory.

Running benchmarks

We provide a number of trained models in the models directory.

Evolutionary algorithm benchmarks

To run the Evolutionary algorithm benchmarks, check the settings to select a model, problems, and dimensions and run_pymoo_experiments.py and run by:

python run_pymoo_experiments.py

Results are stored in pymoo_results/

Bayesian optimization benchmarks

To run the Bayesian optimization benchmarks run the following files in the botorch_code directory, check settings to select a model, problems, dimensions, etc.

python qehvi.py
python qparego.py
python baseline_deephv_batched.py

Results are stored in botorch_results/

Timing benchmark

To run timing benchmarks run:

python time_comparisons.py