Models and deployments

Models in the main sidebar is where trained and pretrained models live. It has four tabs: Registry, Catalog, Benchmarks and Deployments.

Registry

The registry holds your workspace’s models. A model is a family — a name, a task such as object detection, a framework — and it has versions. Every training run adds a version. A version keeps everything needed to use and trust it:

• the checkpoint files, downloadable as one archive,
• the ordered class list it predicts, pinned to an ontology version,
• its metrics from training, and the training report,
• the training and validation datasets and the base model it started from,
• its training parameters and any exports,
• evaluations: every evaluation node that scored this version records its result here, with a link to the report.

That’s the lineage a model registry is for: open a version and you see what data it was trained on, what it scored, and on which test sets.

Models get into the registry in three ways: a training node creates or extends a model family, the Upload model node brings in weights you trained elsewhere, and Add to registry copies a pretrained model from the catalog.

Latest and aliases

@latest always points at the newest ready version of a model. You can add your own aliases to a version, such as production or candidate. An alias names one version at a time: giving it to another version moves it there.

In a node, the model picker works like the dataset picker. Choose the model marked @latest and the node uses the newest ready version every time it runs — so a prediction node picks up the model from the latest training run on its own. Choose a specific version to keep using exactly that one.

Catalog

The catalog lists pretrained models that come with OvalBee’s training and prediction nodes: detectors trained on COCO, classifiers trained on ImageNet and more. Browse them by task, framework and class. Prediction nodes can run a catalog model directly. To fine-tune one, use Add to registry to copy it into your workspace — training nodes work on registry models, and each run adds a version to the copy.

Benchmarks

Every evaluation also lands in Benchmarks, grouped by the dataset it was measured on. Open a dataset to get a leaderboard of every model scored on it, ranked by the metric you choose, with a link to each evaluation report. OvalBee warns you when the runs being compared used different class sets, because their scores aren’t directly comparable.

Deployments

A deployment runs a model as a long-lived service with an HTTP endpoint. The model loads once and stays in memory until you stop the deployment, and the machine it runs on stays reserved for it.

Start one with Deploy on a model in the registry, or Deploy model on the Deployments tab. OvalBee offers the serving templates that can serve that model’s framework and task, and you give the deployment a name and pick the model version (@latest is resolved each time the deployment starts).

Open Info on a running deployment to get its base URL and a ready curl example for each route. Requests go through OvalBee, which checks the caller’s API token before forwarding them. For a YOLO detector, a prediction on images already in the workspace looks like this:

export OVALBEE_API_TOKEN='your-api-token'
curl -X POST "https://ovalbee.example.com/api/deployments/DEPLOYMENT_ID/invoke/predict" \
  -H "Authorization: Bearer $OVALBEE_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"asset_ids": ["IMAGE_ASSET_ID"], "settings": {"conf": 0.25, "iou": 0.7}}'

Routes and request bodies depend on the serving template, which is why Info shows the examples for your deployment. Deployments can be stopped and started again, and their logs and report are one click away in the table.

Related pages

• Train a detector — produce your first model version.
• Evaluate and compare models — fill the Evaluations and Benchmarks tables.
• Reports — training and evaluation reports.