Training · Quick start
Train from a published dataset, then evaluate the saved result.
Before you start
A published version with Train and Val, available images and annotations, a ready training environment and an output destination.
Quick start
- Open Training & Run History and choose the dataset version.
- Check preparation. Box datasets use Create ML; contour datasets offer mask preparation and Segmentation.
- Set the model/run identity, training budget and destination. Leave Augmentation Off unless the variations suit your task.
- Start the run and monitor progress. Evaluate held-out images before registering or using a result.
Functions and choices
Explore the guide
Use numbered tabs and topic buttons to explore. Overview returns to the introduction; Manual opens help.
Choose a workspace
The section menu opens Guide, Training & Run History, Overview, Evaluation and Comparison.
Model Journal
Model Journal is available in the header on every Training page. It opens the current project’s model records.
Follow the walkthrough
Chapter 08, Tutorials, groups the two offline training films with playback controls, seeking and English captions. The first film covers initial training; the second adds fresh reviewed data, a new dataset version, augmentation and model comparison.
Check readiness
Train is disabled while required data, preparation, configuration or environment is missing.
Follow progress
During augmentation, the counter shows how many Train images have been prepared. Total elapsed time includes preparation, augmentation, training, evaluation and model verification.
Judge model quality
Lower training loss alone does not prove model quality.
Result
A saved training run with its settings and result evidence.
Next step
Choose Val/Test before evaluating, then compare models on the same reference images.