5. Compare runs and register a model
Compare runs on the Experiments screen and register the best one in the model registry
Required permission: Register modelDEVELOPER · OPERATOR · ADMIN
One training run becomes one MLflow run. If you train several times with the same experiment name, the runs gather under one experiment, so you can put results from different settings side by side and choose.
- Open Experiments (실험) in the sidebar. Items are grouped as project → experiment ① → run, and each run shows its status and a headline metric (for example mAP50). When you tick runs to compare, the count appears on the Compare (비교) ② button at the top right. Select two or more and press it.
Experiments — ① runs under an experiment; tick them and press ② Compare - On the Metrics (메트릭) tab of Run comparison (Run 비교), view the values side by side. Cells whose values differ between runs are highlighted, and Show differences only (차이만 보기) keeps only the differing items. The Parameters (파라미터) tab shows setting differences, and the Charts (그래프) tab overlays the curves.
Run comparison — the Metrics tab puts both runs side by side - Click the name of the run you chose to open its run detail, and press Register model (모델 등록). Enter the registry name (for example
helmet-detector) in Model name (모델 이름) and press Register (등록). If the name already exists, the model becomes the next version number.Register model on the run detail — enter a model name and the new version enters Staging
When registration finishes, View version (버전 보기) in the "Model registered (모델 등록 완료)" notification takes you to the new version screen. A registered version goes straight into the Staging stage.
The run detail has Overview (개요), Metrics (메트릭), Parameters (파라미터), Tags (태그) and Artifacts (아티팩트) tabs, and if the run was created by platform training, View training run (학습 실행 보기) takes you back to the training detail.