Model lifecycle

A model in AcuSight isn’t a one-off file — it moves through a lifecycle, and every stage is versioned so you can always see where a running model came from and roll back if you need to. This page explains how the stages fit together. The Train & evaluate and Deploy guides cover the how-to.

​

Dataset versions are the starting point

You train on a dataset version — an immutable snapshot of images and labels, split into train / validation / test. Because it’s frozen, a model is always traceable to the exact data it learned from, and two training runs on the same version are comparable. Curating those versions is covered in Dataset versioning.

​

Training produces a model version

A training run turns one dataset version into a model version — a specific set of learned weights, recorded with its configuration and metrics. Each run is its own version; nothing is overwritten, so you accumulate a history you can compare across.

​

Evaluation decides what’s good

Before a model runs on the line, you compare it against the others — precision, recall, and mAP, plus the training curves — to decide whether it’s an improvement. See Model evaluation.

​

Champion and baseline

AcuSight marks models with aliases rather than renaming files:

• champion — the active model for a project. When you promote a new model, it becomes champion.
• baseline — the previous champion, kept so you can roll back with one step if the new one underperforms.

Deployments target the alias, so promoting or rolling back is a change of pointer, not a redeploy from scratch.

​

Deployment to the fleet

The champion is deployed to your devices, where the agent runs it locally on the camera feed.

​

The loop closes

Once deployed, telemetry and model-monitoring show how the model behaves on real production data. When you spot drift or gather better data, you curate a new dataset version and the cycle repeats — each turn traceable back through champion, model version, and dataset version.

​

Next steps