Phase 3: Train a starter and capture more

At the end of this phase, a starter model is training in the cloud and a dataset named dice-train holds about 70 new, unlabeled images. The starter model won’t be very good, since it learns from only 30 images. That’s fine, because its only job is to give you a head start on labeling. In Phase 4 it draws a first guess at every box in the 70 new images, and you fix its mistakes, which is much faster than drawing all of those boxes by hand.

Start the starter training job

  1. On the DATASETS page, open dice-starter.
  2. Click Train model.
  3. Select TFLite as the framework.
  4. Name the model dice-starter-model.
  5. Select Object Detection as the task type.
  6. Select all six labels, 1 through 6.
  7. Click Train model.

Checkpoint

On the DATA tab, click MODELS and expand Active Training. The dice-starter-model job shows Preparing, Waiting for GPU or Training. If the job fails to start with a message about too few annotations, a label has fewer than 10 boxes; add and label more images for it, as in Phase 2.

The job runs in the cloud, so you don’t need to wait for it. A small job like this one usually finishes within 20 minutes, but it can take longer when cloud compute is busy. Keep going with the next section.

Capture the rest of the training set

Thirty images are enough for a rough starter model, but not for a model you can rely on. The models you train in Phases 5 and 6 learn from a training set of about 100 images: the 30 starter images plus about 70 new ones. You capture the new images now, while the starter model trains, and leave them unlabeled; in Phase 4 the starter model labels them for you, and then you add the 30 starter images. They go in a new dataset, dice-train, so dice-starter stays fully labeled.

  1. On the DATA tab, click DATASETS, then click Create dataset.
  2. Name it dice-train, leave Data type set to Binary Data, and click Create dataset.
  3. On the CONFIGURE tab, open the Test section of the camera card.
  4. Roll, click Add to dataset, select dice-train, and confirm.
  5. Repeat until you’ve added about 70 images.

Vary the rolls as you did in Phase 1: one to five dice, different positions, some touching. Keep the camera at the same height and angle. Other conditions, such as a different height or dim light, are for the hard test dataset in Phase 7, where they show how well each model handles scenes it hasn’t seen.

Checkpoint

dice-train on the DATASETS page shows about 70 images. If the count is lower, check that you selected dice-train, not dice-starter, each time. If an image shows no dice or a blurry roll, remove it from the dataset.

Troubleshooting

The training job shows Failed
On the DATA tab, click MODELS, expand Active Training, click the job, and read its logs. The most common cause is a label with fewer than 10 boxes. See Train a model for other causes.

In the next phase, once the starter model finishes training, you use it to label the 70 new images.