This repo implements algorithms for decentralised optimisation from the paper https://arxiv.org/abs/2405.18031

Specifically:

• Centralised Gradient Descent (for comparison)
• Decentralised Gradient Descent
• Decentralised Communication Sliding
• ZO-SADOM
• ProxNSADOM (new)

Requirements:

    python >= 3.10

To reproduce them follow the instructions:

clone the repo, create the virtual env and install the libraries

(i prefer uv, but feel free to use your favorite package manager)

    git clone https://github.com/somvy/decentralised_optimisation && cd decentralised_optimisation
    python -m pip install uv 
    uv venv && . .venv/bin/activate
    uv pip install -r requirements.txt
    wandb login

run experiments on some method

logs are saved to wandb

specify desirable configuration (method, task, network topology, number of nodes etc) in config.yaml

export PYTHONPATH=.
python run.py

Or, if you prefer the Jupyter format, checkout the L1_regression.ipynb