Higher-order Quasi-Monte Carlo Training of Deep Neural Networks

This repository contains the implementation to reproduce the numerical experiments of the paper Higher-order Quasi-Monte Carlo Training of Deep Neural Networks

Installation

Please make sure you have installed the requirements before executing the python scripts.

pip install numpy
pip install matplotlib
pip install pytorch

Data

The data required for each experiment is stored in the data folder and handled by Git LFS (Large File Storage), as it exceeds the maximum storage capacity of github.

Excecuting the code

The code to reproduce each experiment in the paper can be found in the source folder.

In each experiment directory run

 python plot_results.py 

to plot the stored results of the already trained neural network ensemble.

To retrain the ensemble, run

python train_ensemble.py <id>

where <id> corresponce to the id of each neural network in the ensemble specified in the correspodning run_script.py file. The new results of the trained network ensemble will be stored in new_results directory. In order to plot the new results, change the data source path in the corresponding plot_results.py file accordingly.

Note, that retraining the ensemble sequentially may require a very long time. Instead, if applicable, one may run the whole ensemble in parallel.

If applicable, simply run

bash run.sh

for each experiment. This will generate and submit a job-array, where each element corresponce to a neural network in the ensemble.