Build with Ultralytics Platform
Run a real detection project on Ultralytics Platform: budget it, ingest and label data, prove the splits are clean, train and compare runs, deploy a dedicated endpoint, monitor accuracy, and automate the retraining loop.
By Ultralytics Academy
Ship one vision project end to end on Ultralytics Platform and leave with the scripts and runbook that keep it healthy: a pre-upload check, a labeling-threshold sweep, a leak audit, a promotion gate, a resilient endpoint client, and scheduled health and holdout checks.
Map your project onto Platform and budget it before you spend a credit: training, annotation, and endpoint uptime.
Get data in so the dataset reaches Ready on the first try, including deliberate video sampling.
Pick a labeling model and confidence threshold by measuring review work, then verify Batch Annotation with an edit rate.
Prove a dataset version has no cross-split duplicates, no camera leaks, and enough examples of every class.
Launch budgeted cloud runs, judge them from their curves, and fine-tune without the optimizer silently ignoring your learning rate.
Decide whether a new run beats the baseline with pre-declared gates, deploy it, and call it with a client that survives cold starts.
Alert on endpoint health and holdout accuracy, handle regions and personal data, and automate retraining and release.
A project plan with a lifecycle map and a monthly budget.
A Ready dataset with sampled video, reviewed labels, audited splits, and a numbered version.
A trained model that passed a scripted promotion gate against its baseline.
A dedicated endpoint with a retrying client, a scheduled health check, and a holdout accuracy history.
A retraining runbook with trigger, steps, owner, rollback, and an automated release step.
An Ultralytics Platform account. The Free plan and its signup credit cover the course; connected cloud storage needs Pro.
Python 3.11+ with
ultralytics>=8.4.120andultralytics-platformto run the worked examples; theultralytics-platformSDK used in the API lessons is what needs 3.11 (Ultralytics itself runs on 3.10+).Familiarity with Ultralytics YOLO — recommended: Computer Vision Foundations, Building High-Performance YOLO Datasets, and Train your first YOLO model.
A small image dataset (50–500 images) and one short video from your own domain, or the public datasets the examples use.
Course content
4 modules · 10 lessonsModule 1
Module 2
Add and Prepare Data
Check a dataset locally before upload, sample video deliberately, and get a Platform dataset to Ready on the first try.
Smart Annotation and Batch Labeling
Choose a labeling model and confidence threshold by measuring review work, then run Batch Annotation and verify it with an edit rate.
Curate, Dedup, and Split
Find cross-split duplicates and class gaps with two short scripts, split by the unit that matters, and freeze an audited dataset version.
Module 3
Train, Read, and Fine-Tune Runs
Launch a budgeted cloud run, judge a run from its per-epoch curves, and fine-tune without the default optimizer ignoring your learning rate.
Compare Runs and Pick a Winner
Compare runs in a Platform project, record why each run exists, and decide whether a candidate ships with a scripted promotion gate.
Module 4
Deploy and Call an Endpoint
Create a dedicated endpoint in one of 42 regions, size it for your traffic, and call it with a client that handles cold starts, 429s, and bad keys.
Monitor Health and Accuracy
Read Platform's endpoint metrics and prediction statistics, alert from a scheduled health check, and catch accuracy drift with a holdout job.
Privacy, Regions, and Compliance
Choose data and deployment regions from your contracts, keep data private, anonymize images before upload, and answer a pre-launch compliance checklist.
Automate the Retraining Loop
Turn one deployment into a repeatable loop: a runbook, a scripted promotion gate, an automated release with rollback, and a weekly review.