Train a deep learning model

Available with Image Server

In Deep Learning Studio, the Train model step creates the deep learning model that can be used throughout ArcGIS. The deep learning package appears in ArcGIS Enterprise as an item. The tool creates a deep learning package that uses the labeled image chips based on training samples. Image chips are small images containing the feature or object of interest used to train the deep learning model.

Note:

For the Train model step to be completed, image chips must be created or registered in the Prepare training data step.

Create a deep learning model

In the Train model step, you create a deep learning model by iterating through the training process based on defined criteria. The model training process can be performed multiple times, with each time referred to as an iteration. Based on the parameters set up during the iteration, the model creation process is performed a specified number of times with different parameters controlling the process.

The first step in the deep learning model training is to create an iteration.

Note:

To complete this step, sign in to Deep Learning Studio and open a project if one is not already open.

1. On the What would you like to do page, click Train model to open the model training tool.
2. Click the New iteration button to open a new model creation iteration.
3. On the New training iteration page, select the image chips that were created previously or registered to the project.

Caution:

If there are no created or registered image chips, a dialog box appears prompting you to export or register image chips. Choose Yes to open the Manage image chips substep in the Prepare training data step.

4. On the Choose image chips tab, check the check box next to the image chips to be used for the training.
Once the selection is made, previews of the image chips are visible.
5. Click Next to move to the model parameters.
6. Choose the parameters to use to create the model.

Note:

The configuration defaults are based on the project type that was specified when the project was created. For more information about the model parameters, see the description in the Train model step parameter descriptions.

7. Click Start training.

The progress of the training is indicated in the step. When model creation is complete, a graphic appears with information about the training process, a comparison of the training and the validation loss, and the average precision score. There are also some ground truth and prediction comparisons.

On the iteration information page, the following three options for the next step are available:

• Train new interation from this one—Further train the existing deep learning model to improve it.
• Run inference—Use the deep learning model in the next step of the workflow.
• Publish as Deep Learning Package—Share the deep learning model as an item in ArcGIS Enterprise for use in other applications.

8. Consider the next step and click one of the buttons to proceed.

Workaround:

If the model training results are not satisfactory, you can go back to the Prepare training data step to collect additional or different training samples. Unsatisfactory results could be if objects in the ground truth visualizations are missing or mislabeled, a low average precision score, or if the train and validation loss graph indicates a problem with the model.

The Train model step creates a deep learning model that can be used in Deep Learning Studio or in other ArcGIS applications. The Run inference step requires a deep learning model for processing. Depending on the results of this step, the next step may be to use the model in a deep learning analysis or continue to refine the model.