Vision Train
A local, all-in-one workbench for vision models. Data prep, annotation, training, evaluation, export and deployment templates — unified in one Web UI.
PROJECT / 2026A local, all-in-one workbench for vision models. Data preparation, annotation, training, evaluation, export and deployment templates — unified in a single Web UI around the loop of "project → dataset → annotation → training workflow → evaluation / export / deployment templates", across detect / classify / segment / pose tasks.
What it does
Data preparation
- Project management: isolate data and training records per project.
- Import & organize: import existing data, split train sets, clean duplicates, merge datasets.
- Video processing: upload, preview, extract frames into datasets.
- Annotation: box targets, edit classes, bulk clean; draft annotations with an existing model, then review by hand.
- Augmentation: generate reinforcing data for under-represented classes to ease imbalance.
- Dataset versioning: publish / restore / auto-guided dataset snapshots.
Training & deployment
- Training: workflow-centric management of multiple training runs.
- Evaluation: inspect test results and judge whether they meet expectations.
- Export & deploy: generate model files for different environments, plus FastAPI / Python SDK / batch-processing template packs from
ptor exported artifacts. - Task center: unified status for frame extraction, annotation, training, testing and export jobs.
Technical notes
Front-end Vue 3 + Pinia + Vite + Tailwind CSS; back-end Flask + SQLAlchemy + Ultralytics (PyTorch) + OpenCV + OpenVINO; YOLO data format; jobs run in worker subprocesses with progress written back to a SQLite task table.
Single-machine, single-project first. MIT licensed, open source: github.com/noconfuse/vision-train · Docs: vision-train-docs.netlify.app