Skip to main content

Train Page

Route: /train/:jobId — Source: ui/src/pages/TrainPage.tsx

Configure and launch YOLO26 training runs against a finished labeling job. Live logs stream from the trainer worker over WebSocket.

Train page

Dataset quality pre-flight​

Before you see the config form, the Train page pulls a dataset quality report from GET /api/v1/train/dataset-stats/{job_id} and renders it above the configuration. Catching a bad dataset here is cheaper than waiting an hour for a run to finish with poor metrics.

What it shows:

  • Hero stats — annotated frames / total frames, total annotations, class count, and the imbalance ratio (max_class_count / min_class_count). Imbalance ≥5× is highlighted in the warning color.
  • Class balance bars — one row per class with raw counts and percentage share, sorted by frequency. A 10× imbalance between pothole and manhole_cover is immediately visible.
  • Warnings — concrete, actionable lines generated server-side from the stats. Examples:
    • "Only 42 annotated frames. Aim for 100+ per class for reliable training."
    • "Class imbalance 18× between 'pothole' and 'streetlight'. Collect more rare-class examples or use class weights."
    • "38% of objects are smaller than 1% of the frame. Training at 1280px is recommended so detail survives the backbone."
    • "Rarest class has only 4 instance(s). YOLO needs at least ~10–20 per class to learn it."
  • Recommended settings — a one-line summary of the variant / epochs / batch / imgsz / augmentation the server would pick for a dataset this shape. Click Apply recommended to copy those values into the configuration form.

The recommendations are dataset-size aware: under 50 frames gets yolo26n + aggressive augmentation + 150 epochs; 200–1000 gets yolo26s + standard; 1000+ gets yolo26m. imgsz bumps to 1280 when the small-object ratio is ≥25%.

Configuration​

FieldNotes
ModelYOLO26 (yolo26n/s/m/l/x) — accuracy ↔ speed tradeoff
TaskDetection or segmentation
PretrainedStart from COCO, Open Images, or a model already in your registry
Image size640 default; 1280 for small objects
Batch sizeAuto if blank (Ultralytics picks based on VRAM)
Epochs50 is a sensible default; bump to 100+ for hard datasets
Augmentation presetminimal / standard / aggressive
Learning rateAuto by default; override for fine-tunes from a strong base

The Standard preset is highlighted in the UI for a reason — it covers most cases. Pick aggressive only when you have lots of frames and want extra robustness.

Live progress​

While a run is active, the page shows:

  • Status badge — queued, running, completed, failed, stopped
  • Epoch counter + progress bar with ETA
  • Loss chart — box, cls, dfl losses streaming in
  • Validation metrics — mAP@50 and mAP@50:95 per epoch
  • Per-class performance card — precision / recall / mAP50 / mAP50-95 per class, populated from the Ultralytics validator. The metrics_streamer helper pulls validator.metrics.box.{p, r, ap50, ap} every epoch and writes them under per_class/{class}/{metric} keys, which the UI reads directly. Final values are persisted into training_run.best_metrics at on_train_end so completed runs still show per-class breakdowns.
  • Live log tail with auto-scroll (toggle to pause)

You can stop a running training job from the same page (POST /api/v1/train/{run_id}/stop) — the trainer flushes the latest checkpoint and registers it.

After training​

Successful runs auto-register a new model in the registry. From there you can:

  • Activate it (POST /api/v1/models/{model_id}/activate) so the default /predict/* endpoints use it.
  • Promote it to a labeled alias (production, staging, canary).
  • Export to ONNX, CoreML, TFLite, or Edge TPU (POST /api/v1/models/{model_id}/export).
  • Deploy to a named endpoint or push to an edge device.