DGDA-VR: Decentralized Gradient Decent Ascent - Variance Reduction method

This document provides instructions on how to reproduce the experimental results from the paper: "Jointly Improving the Sample and Communication Complexities in Decentralized Minimax Optimization". Covered in this document are:

• package requirements utilized in the experiments
• instructions on how to reproduce the results from the paper (i.e. hyperparameter settings, etc.)
• a description of the main components in this repository, their use, and how to modify them for new use cases

Experiment set-up

Experiments are divided into two categories: those that require a GPU (located in the gpu folder) and those that do not (located in the cpu) folder

The CPU experiments were ran on a 2019 MacBook Pro with a 1.7GHz Quad-Core Intel processor and 16 GB of RAM. The computing environment used is Python 3.8 and the following packages are required:

numpy==1.19.1
scikit-learn==0.23.1

The GPU experiments were ran on clusters of 8 NVIDIA Tesla V100's (each with 32 GiB HBM) connected by dual 100 Gb EDR Infiniband. The operating system utilized is CentOS 7 and all experiments are ran within a conda version 4.9.2 environment. All code is written in Python version 3.7.9, using PyTorch version 1.6.0 with CUDA version 10.2; for instructions on how to install PyTorch with CUDA see here. The GCC version of the system utilized is 4.8.5. To perform neighbor communication, mpi4py was utilized; see here for instructions on how to install this package. A complete list of packages necessary for completing these experiments is located in the requirements.txt file. Comprehensive installation instructions can be found in Install.md

Reproducing the experiments

The cpu/run_trials.py script reproduces the robust non-convex linear regression experiments from the paper. The learning rates are set in lines 120-137 and follow the values we used in our experiments. When the code is running properly, you will see the following output (using our method as an example):

========================= DGDA-VR TRAINING =========================
[COMMS]: 5000
[GRAPH]: N = 20 | rho = 0.9674
[TIME]: 9.13 sec (0.0018 sec / iteration)
========================= DGDA-VR FINISHED =========================

The gpu/main.py script reproduces the PL game and robust neural network training experiments from the paper. The learning rates are set using argparse

Structure of the code and adding new methods

The code is organized as follows:

cpu/
    data/
    mixing_matrices/
    models/
gpu/
    data/
    mixing_matrices/
    models/

models

The models folder contains the base class which each method inherits from and helper functions utilized by all methods. Each folder contains a class in the base.py script; to implement a new method, simply let the method inherit the base class and modify the one_step method accordingly

For the GPU based experiments, we have the helper function which aids in managing the PyTorch tensor updates:

1. replace_weights.py: a custom PyTorch optimizer that simply replaces model parameters from a LIST of PyTorch tensors; since all methods update many variables, this is a straightforward way to minimize conflicts when updating model parameters. The .step() method requires two parameters:
   • weights: a LIST of PyTorch tensors of length k where k represents the number of model parameters
   • device: the torch device (i.e. GPU) the model is saved on

data

Contains the relevant data. To add new datasets, just make sure the run_trials.py and main.py scripts load it appropriately (lines 58-59 + 87-90 and 43-44 + 63-103, respectively)

mixing_matrices

The mixing_matrices folder contains Numpy arrays of size N x N where each (i,j) entry corresponds to agent i's weighting of agent j's information

To run experiments with different weight matrices/communication patterns or sizes, save the corresponding mixing matrix in this directory

Citation

Please cite our paper if you use this code in your work:

@inproceedings{zhang2024jointly,
   title={Jointly Improving the Sample and Communication Complexities in Decentralized Stochastic Minimax Optimization},
   author={Zhang, Xuan and Mancino-Ball, Gabriel and Aybat, Necdet Serhat and Xu, Yangyang},
   journal={Proceedings of the AAAI Conference on Artificial Intelligence},
   year={2024}
}