A minimal federated learning project to experiment with Flower before implementing FSC, FHR, and PCFL algorithms.

flower-fcl-sandbox/
├── README.md
├── pyproject.toml               # Project config and dependencies
├── data/
│   └── generate_synthetic.py    # Create non-IID synthetic health data
├── models/
│   └── risk_model.py            # Simple risk prediction model
├── clients/
│   └── client.py                # Flower client implementation
├── server/
│   └── strategies/
│       └── fsc_prototype.py     # Your FSC algorithm skeleton
├── simulation.py                # Run federated simulation
└── experiments/
    └── run_experiment.py        # Experiment runner with logging

# Install uv if you haven't already
curl -LsSf https://astral.sh/uv/install.sh | sh

# Clone and enter project
git clone https://github.com/YOUR_USERNAME/flower-fcl-sandbox.git
cd flower-fcl-sandbox

# Create venv and install dependencies (uv does both in one step)
uv sync

# Run simulation with 5 clients
uv run python simulation.py --num_clients 5 --num_rounds 10

# Run with custom strategy
uv run python simulation.py --strategy fsc --num_clients 5

# uv will automatically create venv and install deps on first run
uv run python simulation.py --num_clients 5 --num_rounds 10

# Add a new package
uv add requests

# Add dev dependency
uv add --dev pytest

# Add optional dependency group
uv add --optional privacy opacus

1. Day 1: Run the basic simulation, understand client-server flow
2. Day 2: Modify the model, observe convergence behavior
3. Day 3: Implement non-IID data splits, see heterogeneity effects
4. Day 4: Add EWC regularization to the client (FSC foundation)
5. Day 5: Implement Fisher information aggregation in custom strategy

This sandbox uses a synthetic "CKD risk" dataset with features similar to what FLIP-IT will use:

• Age, blood pressure, creatinine, GFR estimates
• Binary classification (high risk / low risk)
• Non-IID splits simulate different practice populations

uv run pytest
uv run pytest --cov=.  # with coverage

uv run ruff check .     # linting
uv run ruff format .    # formatting
uv run mypy .           # type checking

After mastering this sandbox:

1. Replace synthetic data with MIMIC-IV subset
2. Add differential privacy (uv add --optional privacy opacus)
3. Implement temporal distribution shift simulation
4. Scale to multi-GPU with Flower simulation