Training YOLO vision models on Kaggle datasets (github.com)

This repo is a ready-to-use template for training and running YOLO object-detection models directly on Kaggle datasets, lowering the friction for experimenters and practitioners who want to move from dataset to model quickly. It bundles command-line scripts, an interactive notebook, and a Streamlit web app so you can fetch a Kaggle dataset by handle, run optional data preprocessing/augmentation, train a YOLO model, and perform inference on images, video, or webcam. That combination makes it convenient for rapid prototyping, benchmarking, teaching, and demos without rebuilding dataset pipelines from scratch. Key technical pieces: scripts/main.py trains with flags like --dataset <kaggle-handle>, --nc <num-classes>, --names <class-names>, plus --preprocess, --augment-only and --preprocess-config to control preprocessing. scripts/inference.py handles inference (--model <model-path> --input <image/video/webcam>), and streamlit_app.py provides an interactive front end (run with streamlit run). Preprocessing supports cleaning (remove corrupted images, validate annotations, check bbox validity, min/max bbox size) and augmentation (augment_factor, flips/rotations/color/noise) while correctly updating YOLO labels; defaults are configurable in preprocessing_config.yaml. The repo also includes example datasets, a requirements.txt, and contribution docs — useful for reproducible experiments and accelerating dataset-to-deployment workflows in the AI/ML community.