Training is how a camera learns to recognise something new — a defect, a part, a label. You show it a dataset of images you have already marked up, and it produces a model: the file the camera uses to make its own decisions.

The Training tab lists every training run this data source has done, and how each one went.

Before you start

You need a dataset with labelled images, saved as a version. A version is a frozen snapshot of the dataset — training always runs against one, so the run can be repeated later with exactly the same images.

See Datasets for how to create one.

Start a run

1. Open the data source and go to Datasets.
2. Open the versions list for the dataset you want to train on.
3. Choose Run training on the version.
4. Fill in the Train a model form (below) and start it.
5. Open the Training tab to watch it.

You can close the page once the run has started. It keeps going in the cloud, and anyone on your team sees it in this tab.

The "Train a model" form

Only one field is required — the name. Everything else already has a setting that works for most jobs.

What you are making

How it trains

Also build a file the camera can run

Training produces a model the cloud can read. Cameras need a smaller, converted copy, and this builds it at the same time. Leave it on unless you only want the training result.

If you see a warning that the chosen format does not run on the camera app yet, the file will still be built and stored — you just will not be able to send it to a device until the app supports it.

Advanced

Three collapsed sections — how it learns, which images it uses, and which cloud machines run the job. Nothing in there needs changing for a normal run.

Inside them you can also override the image variations the run generates — small changes to brightness, angle and framing that help a model cope with a real production line. The defaults are tuned already; change them only when you know the variation you want to add.

Follow a run

Each row in the Training tab shows the model being made, what it started from, how far it has got and how long it has taken.

Click a row to open the full detail:

• Progress — which pass it is on, and a rough estimate of the time left.
• Charts and epoch log — how the model improved pass by pass, including a per-class view so you can watch one defect catch up (or fall behind) while the run is still going. Switch between the two scoring styles to see whether a class is merely being found or being placed accurately.
• Image variations — the variations this run generated from your images, listed so you can see what the model was trained against.
• Dataset — how many images went into training, validation and testing, plus a warning if the same image appears in more than one of them (which makes the scores look better than they are).
• Final scores — per class, how often the model was right when it fired, and how much it found.
• Dataset sources — if the dataset was merged from several others, which ones and in what split.
• Restore points — snapshots taken during the run. You can turn any of them into a model.

"Stopped early" is normal

Most healthy runs end this way. It means the model stopped improving, so the run finished instead of burning through the remaining passes. The result is still the best version it found.

When it finishes

The new model appears in the Models tab. From there you can send it to a device, or fine-tune from it to make the next version.

Read next

• Datasets
• Models
• Validation