Code to run the DVR algorithm

This repository contains the code for the NeurIPS paper "Dual-Free Stochastic Decentralized Optimization with Variance Reduction", which introduces the DVR algorithm. Details and theory for the algorithm can be found in the paper (DVR.pdf file).

Requirements

Create a new python 3 environment:

conda create --name dvr_env python=3.8

Install the following packages:

conda install -c conda-forge openmpi

conda install matplotlib numba scikit-learn mpi4py numpy scipy tensorflow

Run the code

To run the code, and plot the results, use the command:

mpirun -n nb_nodes python main.py --plot

with nb_nodes the number of processors on the 2D grid (4, 9, 16...)

For instance:

mpirun -n 4 python main.py --plot

Configuration

By default, DVR is run using the config.json file in the current directory. Another file can be specified using the --config_file option. A sample configuration file is provided that the user can modify to test different options.

In particular, it is necessary to change the path_to_data and output_path options to specify the dataset that should be used.