shaoxiongji/federated-learning
Federated Learning
This project is a research-oriented platform designed for simulating decentralized machine learning environments. It provides a framework for training models across multiple client nodes while keeping raw data localized, enabling the evaluation of model convergence and performance under various distributed network conditions.
The system utilizes a parameter-server architecture to coordinate training, where a central coordinator manages the global model state and aggregates weight updates from distributed participants. By decoupling the training orchestration logic from the underlying neural network definitions, the framework allows for the testing of diverse model architectures and data distributions.
The platform includes tools for configuring and executing experiments, such as defining client participation strategies, managing dataset partitioning, and setting training parameters. It supports the simulation of heterogeneous data environments and provides utilities for handling standard datasets, all built to ensure compatibility with the PyTorch ecosystem.
Features
- Federated Learning Frameworks - Provides a research-oriented platform for training decentralized machine learning models using PyTorch and global model averaging.
- Parameter Servers - Coordinates global model state and aggregates weight updates from distributed clients using a central parameter server.
- Federated Learning - Enables research into training machine learning models across decentralized data sources while preserving local data privacy.
- Distributed Training Coordination - Coordinates multiple simulated clients to perform local updates and aggregate them into a single global model.
- Weighted Parameter Averaging - Combines local model parameters using mathematical averaging to update the global model after each training round.
- Decentralized Machine Learning - Simulates network conditions and client participation to evaluate global model convergence on private datasets.
- Model Training Frameworks - Coordinates decentralized model training across multiple clients to update a global model using averaging techniques.
- Configurable Training Loops - Executes local training loops on private data subsets before transmitting weight updates to the central server.
- Data Partitioning - Splits datasets across simulated client nodes to mimic non-IID data distributions in federated networks.
