Local Model Training

Using DEEPCRAFT™ Studio, you can train machine learning models locally on your own PC or on a Linux-based server within your network. The server may be an on-premises machine, a virtual machine, or a cloud instance accessible via SSH. Local training supports containerized environments using Podman or Docker, allowing you to train models consistently across supported systems.

How Local Training Works

You can train models in one of the following ways:

• Local machine training: Runs the training process on your PC using Podman or Docker.
• Remote machine training: Runs the training process on a Linux-based server connected through SSH.

Before you Begin

Before starting local training—on either your local machine or a remote server, perform the following steps:

Open the Local Training window

1. Open the project file (.improj) and click the Training tab on the left pane.

2. Click the Model Generation button and select Generate Models to open the Model Wizard window and generate the model list for training. Refer Generating model to know more about generating model.

3. Click the Start Training button and select Local to train the model locally. The Local Training window opens in a new tab.

   
   

Set the Training Configuration

• Use GPU for training: Select this checkbox to run training on an NVIDIA GPU instead of the CPU. When enabled, DEEPCRAFT™ Studio uses the CUDA-based container image and passes the GPU through to the training container.

  To use this option, the training host must have an NVIDIA GPU with the required GPU support installed — the nvidia-container-toolkit for Docker, or a configured CDI specification for Podman. When training on a remote machine over SSH, these requirements apply to the remote server, since that is where the container runs.

  Clear this checkbox to train the model on the CPU. The CPU image is smaller to download (roughly a third of the GPU image size) and is well suited to non-vision models.

• Imagifier Image: Specifies the container image used for the training container. By default, this parameter is set to the standard image for your project type, so you do not need to change the image in most cases.

  To use a different image, enter your own image reference, or click Browse to select a saved .tar archive from your local machine. This is useful when you want to train with a custom or offline container image instead of the default. Clear the field to restore the default image. DEEPCRAFT™ Studio will automatically reapply the image for your current project type.

Choose your Training Workflow

You can now choose to run the training on your local machine or on a remote server.

Train on your Local Machine

Follow the steps below to train on your local machine:

1. Install Container Runtime: Podman or Docker
2. Set up the MLflow configuration to view the training statistics and results locally.
3. Start the local training on your PC
4. Evaluate the models using MLflow

Install Container Runtime: Podman or Docker

You can install either Podman or Docker, but we recommend using Podman. To install Docker, read the Install Docker Desktop on Windows  guide. If you have already installed Podman or Docker, skip this step.

To install Podman follow the steps:

1. Click the podman_install.ps1 to download the PowerShell script to download the podman.

   
   
   

2. Open the terminal as an administrator.

   
   
   

3. In the terminal, navigate to the folder where you downloaded the script and install Podman by running the command: .\podman_install.ps1

   
   
   

4. Initialize and starting the machine and Login into the container registry to pulling the training container image are handled automatically. Podman is now installed on your PC.

Train on a Remote Machine

Follow the steps below to train on a Linux-based server connected through SSH. The remote machine can be an on-premises server, a virtual machine, or a cloud instance.

1. Check the prerequisites.
2. Configure the remote execution configuration to start the training on the remote server.
3. Set up the MLflow configuration to view the training statistics and results locally.
4. Start the local training on remote server.
5. Evaluate the models using MLflow

Prerequisites:

Before starting remote training, ensure that:

• An SSH connection is available between your local PC and the remote server.
• Podman  or Docker  is installed on the remote machine
• The remote machine has sufficient disk space for training data and output files.
• AWS CLI is configured on the remote machine if private ECR images are used.

Remote Execution Configuration

Configure the following fields before starting training on a remote server:

After setting the prerequisites, configure the remote execution configuration to start the training on the remote server.

1. Remote Path: Enter the path where the training data should be saved on the remote server. By default, the training data is saved in ~Imagimob/local-training.

2. SSH Key: If the .pem file is located in the \Users\xxx\.ssh\ folder, leave the SSH Key Path field blank. If the .pem file is stored in a different location, enter the full path in the SSH Key Path field. When left blank, the system will automatically detect supported SSH keys such as id_ed25519, id_ecdsa, or id_rsa from the .ssh folder.

3. SSH URL: Enter the SSH URL of the remote server. Click Check URL to verify the connection between the remote server and your local PC.

If you want to track training statistics and results, configure MLflow before starting the training job.

ML Flow Configuration

The MLflow configuration is used to track model training and view training statistics and results locally.

• MLflow URL: If MLflow is already installed, enter the MLflow URL and click Check URL to verify the url. If MLflow is not installed, enable MLflow Tracking and leave the URL field empty. The system will run MLflow locally at localhost:5000.

• Enable MLflow Tracking: Select this checkbox to use MLflow for viewing training statistics and results locally.

After downloading the runtime container and setting up the MLflow configuration, you can start the training on your local machine. Refer Start the Training Job to know more about starting the training job.

Start the Training Job

1. Click the Start Training button to begin training locally. The pipeline will run through several stages, including validating the project files, initializing and detecting the container, starting the ML server, and starting the training process. You can monitor the status of each stage in the Pipeline Progress window.

   
   
   

2. After the training is complete, click the Open Output Folder button to open the folder where the trained models are placed locally. The trained models and preprocessing files are saved in the training folder.

   
   
   

3. Click the Copy to Workspace button to copy the trained models and other files to your workspace and evaluate the models using MLflow. You can also view the performance metrics of the trained model in the logs displayed in the Output window.

Evaluate the models using MLflow

1. Click Open MLflow Browser to view the model statistics in the MLflow UI.

2. Go to Runs and search for your project. The project will display all trained models. Open the run for the trained model you want to evaluate.

3. Under Model Metrics, review key metrics such as accuracy, loss, validation accuracy, and validation loss. You can review the statistics for each trained model, compare their performance, and select the best-performing model.

After copying the models to the workspace, refer to the relevant evaluation section based on your model type, such as classification, regression, or object detection. After evaluating the model, refer to the code generation section to generate code for deployment.