This project implements federated learning for binary chest x-ray classification across 3 hospital silos using Flower framework and EfficientNetB0.

• Task - Binary classification (No Finding vs Any Finding)
• Metric - AUROC
• Dataset - NIH Chest X-Ray (non-IID split across 3 hospitals)
• Model - EfficientNetB0 with custom Layer

coldstart/
├── pyproject.toml          # Project configuration and Flower federation settings
├── README.md               # This file
├── coldstart/
│   ├── __init__.py
│   ├── models.py           # EfficientNetB0 model architecture
│   ├── dataset.py          # Dataset loading and preprocessing
│   ├── client_app.py       # Federated client logic
│   ├── server_app.py       # Federated server and strategy
│   └── utils.py            # Utility functions (metrics, etc.)
└── requirements.txt        # Python dependencies

##PPT

PPT

Applied Differential learning rates - classifier (12x), first conv (2.5x), backbone (0.12x)

Learning rate scheduler - 15% linear warmup + cosine decay (min_lr=0.05)

Optimizer - AdamW with weight_decay=1e-4, amsgrad=True

Focal Loss (gamma=2.0, alpha=1.0) with class weighting

Adaptive gradient clipping - 3.0 (warmup) → 1.5 (after)

Mixed precision training (FP16) with GradScaler

Batch size increased from 16 to 32

1. Install dependencies:

pip install -e .

2. Run simulation locally:

flwr run . local-simulation

1. Sync to cluster:

scp -P 32605 -r coldstart/ team04@129.212.178.168:~/

2. SSH into cluster:

ssh team04@129.212.178.168 -p 32605

3. Submit GPU job:

cd ~/coldstart
./submit-job.sh "flwr run . cluster-gpu" --gpu --name xray-fed

4. Monitor logs:

tail -f ~/logs/job*_xray-fed.out

• Backbone: EfficientNetB0 (pretrained on ImageNet)
• Input: 128x128 grayscale images
• Output: Binary classification (sigmoid activation)
• Loss: Binary Cross-Entropy
• Optimizer: Adam with learning rate scheduling

• Algorithm - FedAvg with weighted averaging
• Aggregation - Based on number of samples per client
• Client sampling - All 3 hospitals per round

• Random horizontal flip
• Random rotation (±10 degrees)
• Random brightness/contrast adjustment
• Normalization (ImageNet stats)