Create model training

Model Training offers a comprehensive and easy-to-use solution for training a model using labeled data.

This article lists the steps to create a model training project and start a model training session. A project can hold multiple training sessions for experiments with different dataset versions and model parameters.

Refer to the following video to start a model training session.

1. In the left navigation panel, navigate to Model Copilot > Training Projects.
   

2. Click the Add Project button.

3. In the Add Project window:

   1. Enter Name and Description for the project.

   2. Select the type of model, and click Submit.
      Currently, Classification and Object Detection types of models are supported. For illustration purposes, Object Detection type is selected here. Click Submit.

      The project is now created. You will now need to create the training session within the project.

4. Enter Name and Description for the training session, and click Continue.
   

5. Add images to be used for the model training. The images are added by running a catalog query.

   1. In the training session screen, click the Add Data button on the top-right corner.
      You will need to add image data for the Train Set, Val Set, and Test Set. The data is versioned, and the project can hold multiple versions of the data.
      

   2. Click Create to open the Import from Catalog screen.
      

   3. Select Dataset, View, and the column with the ground truth and bounding box information.

      1. For Object detection, the column should be of type boundingbox2dgt or boundingbox2dpred virtual column types.

      2. For Classification, the column should be of type classlabelgt or classlabelpred type.

         Refer to Virtual Columns for details on virtual column types.
         Refer to Create view with classification data and Create view with object detection data to create views with Classification and Object Detection data, respectively.

   4. In the Max Frame drop down, select a number or enter a custom number for the images to be imported for the training.
      
      
      The images are added to the training session screen.

6. Click Import on the top-right corner, and specify the split ratio for the Train, Validation, and Test sets.

7. Select the applicable split-mode, and click Done.
   
   
   For object detection, only random split is supported.

8. Review the images in each set.

   1. Scroll through each section to review the images.

   2. Click the graph icon on the top right of each set to view the class distribution of the images.

   3. Click the expand icon adjacent to the graph icon to view images of the selected set on a single page.
      

      You can add more images by going through the Add Data steps again.

9. Click Save.

10. In the Save Version window, enter a Name and click Save to save this dataset version.
    

11. Click the Back to Session button.

12. On the session creation page, select the data version that you just created, and click Next.
    

13. Under the Select Model tab, select the Model Size, Dataset Type, and Max Epochs for the training.

    These parameters will be used to select a suitable base model automatically.
    

14. Enable Early stop, if needed, and set up the early stop parameters.

    If there is not enough change in the loss or accuracy value within a given range for a specified number of epochs, you can trigger an early stop for the process to stop and generate results.  

    1. Minimum Epoch: Define the minimum number of epochs that must run before applying early stop condition.

    2. Patience Window: Define the number of epochs to monitor for changes in validation loss or validation accuracy to trigger an early stop.

    3. Minimum Loss Delta and Minimum Accuracy Delta: Specify the threshold value such that early stop is triggered if the improvement in validation loss and validation accuracy is below this value over a epochs specified as ‘Patience Window’. The default values are suitable for most cases.

       

15. Click Next.

16. Review the data version and model parameters, and click Submit to initiate a training session.
    
    

    You can track the status of the session on the session listing page.

All the sessions are displayed in the listing page, along with the status. In case you had triggered an early stop for the training, it is indicated in the listing page as shown below: