DARPA Age Group Prediction - Submission Package

Contents

This submission package contains all necessary files to generate predictions on the test dataset.

Files Structure

submission/
├── generate_submission.py     # Main prediction script
├── best_model.pth            # Trained model checkpoint
├── aortaP_test_data.csv      # Test data (AortaP signals)
├── brachP_test_data.csv      # Test data (BrachP signals)
├── requirements.txt          # Python dependencies
├── config/
│   └── config.yaml          # Model configuration
└── src/
    ├── data/
    │   ├── dataset.py       # Dataset preprocessing
    │   └── preprocessing.py # Signal filtering utilities
    ├── models/
    │   └── age_group_predictor.py  # Model architecture
    └── utils/
        └── config.py        # Configuration utilities

Requirements

• Python 3.8+
• PyTorch
• NumPy
• Pandas
• SciPy
• scikit-learn
• PyYAML

Install dependencies:

pip install -r requirements.txt

Usage

Generate Predictions

python generate_submission.py \
    --model best_model.pth \
    --aorta-data aortaP_test_data.csv \
    --brach-data brachP_test_data.csv \
    --output OCA-SENTINEL_output.json \
    --device auto

Arguments

• --model: Path to trained model checkpoint (default: best_model.pth)
• --aorta-data: Path to AortaP test data CSV
• --brach-data: Path to BrachP test data CSV
• --output: Output JSON file path
• --device: Device for inference (auto, cuda, or cpu)
• --config: Path to config file (default: config/config.yaml)
• --team-name: Team name for filename validation (OCA-SENTINEL)

Device Options

• auto (default): Automatically detects GPU availability
• cuda: Force GPU usage (falls back to CPU if unavailable)
• cpu: Force CPU usage

Output Format

The script generates a JSON file with predictions in the format:

{
  "0": 2,
  "1": 4,
  "2": 1,
  ...
}

Where keys are sample indices (0-874) and values are predicted age group classes (0-5).

Model Architecture

• Encoder: Transformer-based encoders for AortaP and BrachP signals
• Features: Handles missing data with attention masking
• Preprocessing: Butterworth low-pass filtering + normalization
• Classes: 6 age groups (0-5)

Notes

• The script automatically preprocesses signals (filtering + normalization)
• Predictions are validated before saving
• Progress updates are shown during inference
• Class distribution statistics are displayed after generation