/private_flower

This repo contains implementations of differentially private federated learning using Flower framwork.

Primary LanguagePython

Federated Learning and Few-Shot Training

This repository contains code and utilities for federated learning using the Flower framework and implementing few-shot training for models that were initially trained on out-of-distribution (OOD) data.

Overview

Federated learning is a decentralized machine learning approach where multiple clients collaboratively train a global model without sharing their raw data. This repository demonstrates a federated learning setup using the Flower framework.

Additionally, the repository provides a few-shot training mechanism, aiming to fine-tune models that were previously trained on out-of-distribution data, enhancing their performance on specific tasks.

Features

  • Federated learning using Flower framework
  • Few-shot training for models trained on out-of-distribution data
  • Utilization of CIFAR-10 dataset
  • Client and server nodes for federated training

Usage

To use this repository for federated learning and few-shot training, follow these steps:

  1. Setup Environment:

    Ensure you have Poetry installed to manage the project dependencies.

    # Install project dependencies using Poetry
poetry install

  2. Data Preparation:

    Prepare the CIFAR-10 dataset by following the data loading instructions provided in the dataloaders/cifar10_loader.py file.

  3. Training:

    • Run the federated training process using the poetry run command and main.py script. Adjust the parameters as needed for your specific use case.

    • Optionally, enable few-shot training by specifying the path to pre-trained models using the --load-from argument.

    # Run federated training
poetry run python main.py --num-clients 5 --num-rounds 20 --batch-size 128

  4. Fine-Tuning:

    • Fine-tune the models using the fine_tune_train function from trainers/fine_tune_train.py.

  5. Evaluation:

    • Evaluate the trained and fine-tuned models using the evaluate_on_loaders function from trainers/simple_trainer.py.

Directory Structure

  • dataloaders/: Contains data loading utilities, including CIFAR-10 dataset loader.
  • models/: Defines the neural network architecture (e.g., CNN model).
  • multi_process_federated/: Includes federated learning-related components like federated trainer and node types.
  • trainers/: Contains training and evaluation scripts, including fine-tuning and simple training mechanisms.