Cache-DFL

A cache-based Distributed Federated Learning framework (experimental) for "Decentralized Federated Learning with Model Caching on Mobile Agents.". This repository explores new strategies for caching and data loading in distributed or federated learning scenarios.

---

Table of Contents

• Features
• Installation
• Usage
• Project Structure
• Configuration
• Examples
• Contributing
• License
• Acknowledgments

---

Features

• Federated Learning Trainer: A customizable trainer for simulating federated learning across multiple nodes.
• Caching Mechanism: Implements a caching module to efficiently manage repeated data access.
• Data Handling Utilities: Convenient data loaders for local or remote datasets.
• Extensible Design: Easily modify or extend the core classes to fit your own research.

Installation

1. Clone the repository:

   git clone https://github.com/shawnxiaoyuwang/Cache-DFL.git
   cd Cache-DFL
   

2. Install dependencies (if using requirements.txt):

   pip install -r requirements.txt

   Or if you prefer a virtual environment:

   python -m venv venv
   source venv/bin/activate  # On Windows: venv\Scripts\activate
   pip install -r requirements.txt
   

Usage

1. Configure your settings:
   • Update config.yaml or config.json with your desired parameters (learning rate, federation rounds, caching strategy, etc.).
2. Run the main script:

   python main.py --config config.yaml

Project Structure

Cache-DFL/
├── cache_dfl/                # Main Python package
│   ├── aggregation.py
│   ├── cache_algorithm.py    #
│   ├── trainer_single.py     # Contains main training logic on single thread
│   ├── trainer_mpi.py        # Contains main training logic on multiple threads/machines (using mpi)
│   ├── data_loader.py        # Data loading and caching logic
│   ├── model.py              # Models or model architectures
│   ├── road_sim.py           # Road simulation
│   ├── seed_setter.py        # Random seeds
│   ├── utils.py              # Utility functions
│   └── utils_cnn.py          # Utility functions for CNN
|
├── examples/
│   └── demo.py (To be develop)
│   └── ...
|
├── tests/
│   ├── test_trainer.py (To be develop)
│   └── ...
├── config.json               # Example config file
├── requirements.txt          # Python dependencies
├── LICENSE
├── main.py                   # Entry point for running the federated learning pipeline using mpi process (to be developed, currently you can excute trainer_single.py ot trainer_mpi.py with argparse directly)
└── README.md

• cachedfl/: Core Python package containing the main source code.
• examples/: Example scripts that show how to use the code.
• tests/: Automated tests using pytest or unittest.
• requirements.txt: Python dependencies list.
• README.md: This file.

Configuration

• Hyperparameters: Learning rate, batch size, number of epochs, and federated rounds can be modified in the config file or via command-line flags.
• Caching Options: Customize parameters such as cache size or eviction policies.
• Distributed Settings: Simulation on single thread or multiple threads/machines (using mpi)

Examples

to be developed, currently you can excute trainer_single.py ot trainer_mpi.py with argparse directly

python examples/demo.py

License

• This project is licensed under the MIT License – see the LICENSE file for details.

Acknowledgments

• Libraries/Frameworks: PyTorch, NumPy, or others as relevant.
• Contributors: Xiaoyu Wang, New York University
• References: Wang, Xiaoyu, Guojun Xiong, Houwei Cao, Jian Li, and Yong Liu. "Decentralized Federated Learning with Model Caching on Mobile Agents." arXiv preprint arXiv:2408.14001 (2024). https://arxiv.org/abs/2408.14001 (Oral Presentation at AAAI 2025)