Selected reusable utilities extracted from the Kaggle training pipeline. These are not the full training notebooks — see KAGGLE_NOTEBOOK.md for complete training code.
| Script | Purpose |
|---|---|
parse_keypoints.py |
Parse 2DCoordinates.log → JSON with 33 valid keypoints |
compute_pck.py |
Compute PCK, MAE, RMSE from numpy prediction arrays |
# Parse annotations
python scripts/parse_keypoints.py path/to/2DCoordinates.log output/parsed.json
# Evaluate metrics (demo with random data)
python scripts/compute_pck.py- Full training pipeline (model, DataLoader, GPU training loop) runs on Kaggle
- Adapt file paths when running locally — Kaggle uses
/kaggle/input/and/kaggle/working/ - Install dependencies:
pip install -r requirements.txt