The purpose of reference model training is to generate a model that can be used for auto labelling for a specific application or for hardware deployment.

• For auto-annotation choose a YoloV8 for training the reference model.
• For Hardware deployment choose modelpack model.

To train the following steps are required:

Prepare Dataset for Training

Training required a Training Set, a Validation set and Optionally a Test set. These datasets can be independently imported.  To generate these from a single dataset, the split dataset option can also be used. 

• To use datasets for training, either split the dataset into groups- usually named as train and val but any name can work or dynamically split into training and validation in the trainer dialog.
• The dataset must have a ground truth annotation set (with any name)

Train a New Model From Annotated Datasets

Based on previously annotated datasets, we can train new models for export into edge devices or to auto-annotate future datasets.  To create a new model, we must first select a project and confirm the project is selected in the Project field in the Navigation Bar (or "Nav Bar") on the top of website.

Then go to the Trainer Workspace using the Workspace button on the right side of the Nav Bar.

Click on the "Create" button to create a new Training Experiment.  Training Experiments can be considered as user-defined folders for multiple model-training sessions.

In the "Create New Experiment" popup, give the experiment a name and optional description and click the "Create New Experiment" button.

Confirm a new experiment is created on the Training Workspace.

Create a new Training Session

Each trainer experiment can have multiple training sessions. Creating mutiple training sessions in one experiment allows the user to compare the results on a common ground.

To create a new session in an experiment, click on the Menu button and select "New Training Session".

In the "Training Sessions" pop-up, you can configure the training session.  Fill in the name and optional description.  You can choose any training dataset from the project.

You can create a new training validation split here ir use existing groups in a dataset and training validation split.

For this specific test, we are going to use the groups from the Dataset:

As well, you can configure some of the high-end parameters of the model, such as which kind of model (ModelPack or Yolo V8), what task (detection or segmentation), import weights from a previously trained model or the default weights, the number of epochs to train, the batch size, and the image width and height.

For this testing, we'll use the following settings:

For advanced users, the learning rate parameters can be configured as well:

Once these are configured, you can press the "Start Session" button to start training.  This will take us back to the Training Windows and the session will initiate:

When the model has completed training:

Clicking on the session will take us to the session modal:

Here we can download the floating point Keras (H5) model, or convert and download the model to ONNX, TFLite, quantized RKNN, or quantized RTM model.  The console window can present the metrics of the model, parameters used in training, and the console log of the training session (shown above).

Going back to the Experiments Window, we can also review the training outputs of sessions within an experiment by clicking on the Training Analytics button in the top right of the Experiments Card.

This will take us to the Training Analytics window.  Here, we have TensorBoard visualization of all the sessions within a run.

One of the noteworthy outputs will be the validation outputs of each epoch of training, which can be viewed by clicking on the Text link on the orange, TensorBoard navigation bar.