Training ModelPack

This tutorial describes the steps to train ModelPack Vision models in EdgeFirst Studio. For a tutorial to train Fusion models, see Training Fusion Models.

Verify Dataset

First ensure that the dataset is ready to be used for training. This means that the dataset is properly annotated and the dataset is properly split into training and validation groups. The section in Verify Dataset will show what to look for in a dataset before deploying it for training.

Specify Project Experiments

From the projects page, choose the project that contains the dataset you plan to use. In this example, the project chosen is the "Object Detection" project which was created in the Quickstart Guide. Next click the "Model Experiments" button as indicated in red.

Create Model Experiment

You will be greeted with the "Model Experiments" page. A new project will not have any experiments as shown below. You will need to first create a model experiment. As mentioned in the Model Experiments Dashboard, model experiments will contain both training and validation sessions.

Click on the "New Experiment" button as shown on the top right corner of the page.

Enter the name and the description of the experiment marked by the fields shown below. Click on the "Create New Experiment" button to create your experiment.

Your created experiment will appear like the following below. At the start, this experiment will contain zero training and validation sessions. The next step will show how to start your first training session on this experiment using the dataset in the project.

Create Training Session

In the experiment card, click the "Training Sessions" button as indicated in red below.

You will be greeted to the "Training Sessions" page as shown below.

Start a training session by clicking on the "New Session" button on the top right corner of the page.

You will be greeted with a training session dialog. In this dialog, specify the "Trainer Type" to "ModelPack" and provide a name and description of the training session as shown below. Next specify the dataset to be used with training and validation partitions. In this example, the dataset specified is the "Coffee Cup" dataset which was created in the Getting Started. Next specify the training parameters. By default, an object detection (bounding boxes) model will be trained. However, you can specify either "Segmentation" or "Multitask" as shown below. This model will output both bounding boxes and segmentation masks. Additional information on these parameters are provided by hovering over the info button .

Input Resolution

We recommend changing the input resolution to 640x360 to maximize detection rates on small datasets.

Large Batch Size

For small datasets, a large batch size may produce poor results. Use a batch size of 4 or 8.

For more information on available "Data Augmentations" please see Vision Augmentations.

Once the configurations have been made, go ahead and click on the "Start Session" button on the bottom right of the window. This will start the training session which will train the model for the number of epochs specified.

Session Progress

Once the training session has started, the progress with the stages will be shown on the left and additional information and status is shown on the right.

The training process begins with cloud instance initialization. Then the dataset is downloaded and cached. Training starts afterwards. At the end of the training process, ModelPack quantizes the model and publishes the checkpoints.

Completed Session

The completed session will look as follows with the status set to "Complete".

The attributes of the training sessions in EdgeFirst Studio are labeled below.

Training Outcomes

Once the training session completes, you can view the training charts by clicking the "View Training Charts" button on the top of the session card.

You can go back to the training session card by pressing the "Back" button as indicated in red below on the top left corner of the page.

The trained model artifacts can be downloaded by clicking the "View Additional Details" button on the training session card in EdgeFirst Studio. This will open the session details and the models are listed under the "Artifacts" tab as shown below. Click on the downward arrows to download the models to your PC.

Session Details	Artifacts

It is also possible to compare the training metrics for multiple sessions. See Training Sessions in the Model Experiments Dashboard for further details.

Netron

You can visualize the architecture of these models using https://netron.app/.

Next Steps

Now that you have generated your Vision model, follow along the next steps for validating your model either through managed or user-managed validation sessions.