Phase 5: Train a built-in model

At the end of this phase, a built-in TFLite model is training on all of dice-train. It’s the first of two models you’ll compare. Built-in training needs no code, and the TFLite model it produces runs well on small machines. While this model trains, you can move on to Phase 6 and start the second model.

Start the built-in training job

This is the same built-in TFLite training you used for the starter model in Phase 3, but this time it learns from the full training set: about 100 images, all labeled and reviewed. With more than three times as many images, expect it to be more accurate than the starter. In Phase 8 you compare it with the YOLO model from Phase 6 to see which one reads dice better.

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

For the other options and the CLI equivalent, see Train a model.

Checkpoint

On the DATA tab, click MODELS and expand Active Training. The dice-tflite job shows Preparing, Waiting for GPU or Training, and later Completed. If the job won’t start, the message names the problem: usually a label with fewer than 10 boxes, or more than 20% of images unlabeled. Fix it in dice-train as in Phase 4 and try again.

In the next phase, you start a second training job on the same dataset with a YOLO training script from the registry.