PTB-XL Federated Learning — Simple Run Instructions

1. Environment

---

python -m venv .venv

Windows: .venv\Scripts\activate

macOS/Linux: source .venv/bin/activate

pip install -r requirements.txt

2. Data placement (PTB-XL v1.0.3)

---

Place files like this (example): /dataset/ptbxl/ ptbxl_database.csv scp_statements.csv records100/ records500/ # optional

3. Configure (src/config.py)

---

• Set DATA_DIR to your dataset folder path, e.g.: DATA_DIR = "/dataset/ptbxl"
  • Keep defaults unless you know you need changes. Common edits: SAMPLE_RATE = 100 CLIENTS = 4 ROUNDS = 15 EPOCHS_LOCAL = 2 BATCH_SIZE = 64 LR = 1e-3 FREEZE_ENABLED = False SPLITS = {"train": 0.70, "val": 0.15, "test": 0.15} MODEL["type"] = "cnn" TUNING["enabled"] = False ANOVA_FSCORE_THRESHOLD: float = 300.0 ANOVA_FALLBACK_LEADS: int = 6

4. Run order (from project root)

---

(A) Optional: EDA & basic features

python -m src.eda

(B) Centralized baselines

python -m src.Centralized

(C) Federated training (start server, then clients)

python -m src.Server

In separate terminals (0..3 for 4 clients)

python -m src.Client --cid 0 python -m src.Client --cid 1 python -m src.Client --cid 2 python -m src.Client --cid 3

(D) Optional: Grid search (5-fold CV if enabled in config)

python -m src.tuning

(E) Plots & comparison (e.g., frozen vs non-frozen runs)

python -m src.results_visualization

5. Outputs (results/)

---

<run_name>.csv # global + per-client round metrics <run_name>_perclass.csv # per-class accuracy across rounds <run_name>_cm.csv # confusion matrix (long) viz/ # PNG plots (accuracy, per-class, confusion)

6. Quick fixes

---

• "Data not found": Check DATA_DIR and that ptbxl_database.csv, scp_statements.csv, and records100/ exist.
• "Clients can't connect": Start Server first; leave defaults for local run (127.0.0.1:8080).
• "GPU not used": python -c "import torch; print(torch.cuda.is_available())"