Model training

Create a model

The Create model option is used to create a new AI model and configure it for training.

As a requirement to create a model, you must have already uploaded a data set and saved at least one version of that data set. Only versioned data sets can be used for training.

General model information

• Name: Unique name of the model.
• Description: Optional description of the model and its intended use.

Data set configuration

• Data set: Choice of the data set to be used for training. Only data sets for which a version has already been created are displayed.
• Data set version: Choice of the data set version to use. This version represents a fixed and reproducible state of the data set. Changes made to the data set after revision control do not affect existing training configurations.

Runtime and target system configuration

• Latency threshold in ms: Defines the maximum allowed execution time for the AI model on the target system. This value is taken into account during model creation in order to make a choice of a model that is suitable in terms of accuracy and runtime performance.
• Target device: Choice of the target hardware on which the model will later be executed.
• CPU: Choice of the CPU configuration used by the target system.
• GPU option: Choice of supported GPU configurations.
• Target software: Defines the TwinCAT target environment for subsequent inference, for example:
• TF3820 | TwinCAT 3 Machine Learning Server
• TF7810 | TwinCAT 3 Vision Neural Network
• Execution provider (EP): Defines the execution environment for the Machine Learning Server inference engine, such as CPU or CUDA® (NVIDIA® GPU).
• Number of cores used for inference: Specifies the number of CPU cores that will be used for subsequent model execution. When using TF7810, this value refers to isolated CPU cores within the Vision Job Pool. When using TF3820 with Execution Provider = CPU, the value corresponds to the number of available CPU cores in the user mode process of the TwinCAT Machine Learning Server. Only user mode cores are used in this process. On systems with a hybrid architecture, only performance cores (P-cores) are taken into account.

Create a model

• Click Submit to save the training configuration. Training must then be started on the model card.
• Clicking Cancel closes the dialog box without making any changes.

Model card

The model card displays the current state as well as the most important configuration and performance data for an AI model.

The card's title matches the model name.

Model actions

Various actions are available in the upper section:

• Open analysis: Opens the model's analysis and explainability view.
• Download model: Downloads the trained model.
• Edit model: Opens the model configuration to edit the metadata
• Delete model: Removes the model from the project.
• Start training: Starts a training run for the model. Please note the rules regarding compute hours billing.

Model information

• State: Displays the actual state of the model. Possible states include, for example:
• Submitted
• Waiting
• Running
• Finished (including a timestamp indicating when the model reached this state)
• Error
• Description: Optional description of the model
• Data set: Displays the data set used, including the data set version. You can open the corresponding data set by clicking the search icon.
• Labels: Displays the classes or target categories contained in the data set.
• Target device: Displays the configured target hardware, including the CPU, GPU, and execution provider configurations.
• Target software: Displays the configured TwinCAT target environment for later inference.
• Target latency threshold: Displays the configured maximum target latency for model execution.

Training and performance information

• Performance: Displays the achieved model quality as a percentage. The metric shown is based on the test results calculated during validation.
• Training time: Total duration of the training process.
• Estimated latency: Estimated runtime of the model on the configured target hardware.
• Training history: The "Loss History" feature allows you to visualize the progress of the training process. This represents the training metrics for individual epochs.