Launch Initial Training

The XNAT ML release was intended to start from a pre-trained model. We can exploit some code behavior to let you start from scratch. Future versions will support training from scratch natively.

Launching Initial training requires that you have the following:

• An installed training container image, with the training command enabled in your project. See Install Docker Images and Commands and Enabling Installed Commands
• An empty model in your project. See Installing a Model.  To train from scratch, you will need to create and upload an empty zip file instead of the checkpoint files referenced in fine-tune training.
• An installed training configuration in your project. See Creating a Training Config for your Model
• A saved dataset in your project with files to train. See Validating and Saving your Dataset from Project Data

Once you have all your data in order and your model configured and ready, you can launch a training run using XNAT and your connected processing resources. Follow the following steps:

• Navigate to the Machine Learning Dashboard. (i.e. From the project page, open the "Machine Learning" tab and click "Open Machine Learning Dashboard". Ensure that the "Installed Models" tab is active, and you will see a panel that lists all models in your project, and all training runs associated with each model.

• Find your model in the list of Installed Models, and click "Launch Training". A training dialog will open up. 

• Select the appropriate command for your training. (i.e. "clara-v3-training")
• Select the training config and dataset you want to use. The dialog will reveal inputs for each parameter that can be set at runtime.

• Adjust your parameters as necessary and click "Launch Training". After configuring the launch, a success dialog will display.