CAFE — Centralized and Federated Experiments

CAFE (Centralized And Federated Experiments) is a unified framework designed to run, compare, and analyze centralized and federated training experiments using YOLOv8 for object detection tasks.

The repository provides reproducible pipelines, environment-driven configuration, and evaluation workflows for benchmarking edge–cloud, distributed, and federated learning scenarios.

---

📁 Repository Structure

CAFE/
├── centralized_training/
│   ├── centralized_training.py
│   ├── example.env
│   ├── requirements.txt
│   └── README.md
│
├── federated_training/
│   ├── server.py
│   ├── client.py
│   ├── server.env
│   ├── client.env
│   ├── requirements.txt
│   └── README.md
│
└── README.md   <-- (this file)

---

🎯 Project Goals

CAFE provides a controlled environment for comparing:

• Centralized vs Federated training performance
• CPU, GPU, and Jetson-edge devices
• Different federated learning strategies (FedAvg, FedAvgM, FedAdam, FedYogi)
• Training time, communication cost, accuracy, and scalability

The project enables reproducible experiments through:

• .env-based configuration
• deterministic training options
• unified logging and CSV output
• consistent dataset management for both approaches

---

⚙️ 1. Centralized Training

The folder centralized_training/ contains:

• centralized_training.py — YOLOv8 centralized training script
• .env — configuration file
• automatic CSV metrics output
• training time profiling

▶️ Run centralized training

cd centralized_training
python centralized_training.py

---

⚙️ 2. Federated Training

The folder federated_training/ contains:

• server.py — Flower federated learning server
• client.py — YOLOv8 federated client
• server.env — server configuration
• client.env — client configuration
• Jetson-safe operations
• final aggregated model + CSV logs

▶️ Start server

cd federated_training
python server.py

▶️ Start a client

python client.py

To launch multiple clients, duplicate client.env and modify:

CLIENT_ID=client2
DATA=dataset_client2/fasdd.yaml

---

📊 Results & Outputs

Both centralized and federated modules generate:

• YOLOv8 model checkpoints (.pt)
• per-class and global metrics
• CSV logs
• timing information
• (Federated) final aggregated model
• (Federated) round-by-round time measurements

---