Skip to content

README.md

Latest commit

 

History

History
59 lines (43 loc) · 2.27 KB

README.md

File metadata and controls

59 lines (43 loc) · 2.27 KB

YOLO26 Fine-Tuner

A terminal-based interactive tool for fine-tuning and running inference with YOLO26 models on custom datasets. Navigate training, validation, prediction, and hyperparameter configuration entirely through keyboard-driven menus — no config files needed.

Features

  • Train — configure epochs, batch size, optimizer, LR scheduler, and more through interactive menus
  • Validate — run validation with a custom trained or official model; all settings optional (model remembers training config)
  • Predict — run inference on a webcam or video file
  • Hyperparameter tuning — manual overrides or automatic tuning via model.tune()
  • Augmentation settings — full control over mosaic, mixup, flips, HSV shifts, and more
  • Dataset support — any dataset with a YOLO-format .yaml file; point the menu at it and go

Requirements

  • Python 3.14+
  • CUDA-capable GPU (recommended)
  • uv package manager

Setup

git clone https://github.com/LightCyan01/yolo26-trainer.git
cd yolo26-trainer
uv sync

Official model weights (YOLO26) are downloaded automatically by Ultralytics on first use.

Usage

uv run main.py

Use arrow keys to navigate, Enter to select. The main menu offers:

  • Train — select task type → pick an official model → set dataset YAML → configure settings → start
  • Validate — select task type → pick official or custom trained model → optionally override val settings → start
  • Predict — select task type → pick official or custom trained model → select webcam or video file → optionally override predict settings → start

Project Structure

src/
  menu/         # Interactive menus (train, val, predict)
  training/     # run_train(), run_val(), run_predict() wrappers
  utils/        # Settings dataclasses, validators, file dialogs, model lists
dataset/
  NEU-DET-YOLO/ # Example: NEU surface defect dataset (YOLO format)
  MVTecAD/      # Example: MVTec anomaly detection dataset
models/         # Auto-downloaded weights are cached here

Tech Stack