Tutorial: ship your first model to the field
Follow one full cycle through dataset, training, approval, serving, edge deployment, drift and retraining (about 30 minutes)
The example is a detection model that finds whether people wear safety helmets. Done in order, it takes about 30 minutes.
What you need
- The platform address and an account. The screenshots in this tutorial were taken in the tenant
DEMO. - Each step needs a different role. If one person has the admin role, they can do everything; if roles are split, hand off to a colleague with the right role.
| Page | What you do | Role required |
|---|---|---|
| 1. Sign-in and tenant | Sign in, check the tenant | Anyone |
| 2. Create a dataset | Create a dataset, upload files | Developer, operator, admin |
| 3. New training | Start training with the 11-step wizard | Developer, operator, admin |
| 4. Watch training progress | Six-step timeline, MLflow metrics | Anyone |
| 5. Compare runs and register a model | Compare runs, register in the model registry | Registering: developer, operator, admin |
| 6. Promotion request and approval | Production promotion request, approval | Approver, admin |
| 7. Serving | Container image build, API service, inference console | Operator, admin |
| 8. Edge deployment | Deployment record, sending a model to a device | Operator, admin |
| 9. Monitoring and retraining | Monitoring, drift, retraining | Viewing: anyone; retraining: developer, operator, admin |
Split by role, the tutorial goes like this:
- Developer: 1 → 5 (dataset, training, registration)
- Approver: 6 (approval)
- Operator: 7 → 9 (serving, deployment, operations)
Common tasks (stopping and resuming training, rollback, silencing alerts and more) are in How-to.