YOLO-Craft

Описание и применение

YOLO‑Craft предоставляет легковесный, скриптовый подход к дообучению и развертыванию моделей YOLOv8, включая активное обучение, интерактивную разметку, экспорт на Edge‑устройства и визуальную оценку.

Задача

Fine‑tuning YOLOv8 models for custom computer vision projects was hindered by heavy MLOps tools that bundled large frameworks, complex wrappers, or required cloud setups.

Результат

YOLO‑Craft enables lightweight, end‑to‑end fine‑tuning and deployment of YOLOv8 models without heavy frameworks, improving workflow efficiency and device compatibility.

Рабочий процесс

• 1. Run watcher.py to capture low‑confidence frames from a stream or webcam.
• 2. Use labler.py to draw bounding boxes interactively, switching target classes with hotkeys and handling background‑only frames.
• 3. Train or fine‑tune YOLOv8 on the labeled data.
• 4. Export the trained model with export_model.py to TensorRT, TFLite, CoreML, or ONNX for deployment on edge devices.
• 5. Generate visual metrics and performance reports with report.py.

Доказательства

When fine‑tuning YOLOv8 models for custom computer vision projects, I kept running into the same issue: most MLOps tools are packed with heavy frameworks, complex wrappers, or forced cloud setups.

Источник

Active Learning Loop (watcher.py): Automatically captures low-confidence frames from streams or webcams to continuously improve your model where it struggles most.

Источник

Multi-Class Labeler (labler.py): Interactive bounding box drawing with quick hotkeys (like using l to switch target classes) and built-in support for background-only frames.

Источник

Universal Edge Export (export_model.py): Exports trained models directly to TensorRT, TFLite, CoreML, or ONNX for real-time deployment on devices like Raspberry Pi or NVIDIA Jetson.

Источник

Visual Evaluation & Reports (report.py): Generates automated visual metrics and before-and-after performance comparisons.

Источник