JSON Settings

[source,json]

{ "trainingData": "C:/Users/user_name/Desktop/recordings/raw", "validationData": "C:/Users/user_name/Desktop/recordings/raw", "exportDirectory": "C:/Users/user_name/Desktop/recordings/network", "finetuneDataRoot": "C:/Users/user_name/Desktop/recordings/raw", "classes": [ "Arrow", "Disk" ], "nonRotables": [ "Disk" ], "imageScale": 0.5, "networkBase": null, "networkType": "ORTMS", "schedule": "Full", "trainingProfile": "B8x1", "deploymentConfiguration": "LumiScanVGR", "tmpFiles": "tmp" } -

These are standard parameters of the JSON configuration. Their meaning is explained in Table 2. The required parameters depend on the application and training workflow.

JSON Parameters Description

When specifying paths, use forward slashes (/) instead of backslashes (\).
If no value is specified for a parameter that supports an empty value, use null exactly as shown. Do not use quotation marks or capital letters.
Table 1. Table 2: LumiScan Training JSON Settings
Parameter Description

trainingData

Path to the training data.
This folder is typically generated by LumiScan Annotation and contains the images and corresponding annotation data in CSV or JSON format. Each image should have a corresponding annotation file.

validationData

Path to the validation data used to evaluate the network during training.
This can be set to the same path as trainingData.

testData

Optional. Path to the data set used when Start Testing is selected.

exportDirectory

Optional. Path used to store exported ONNX models and other training results.
If no path is specified, the results are stored in the directory containing the training data specified in trainingData.

finetuneDataRoot

Path to the data used for fine-tuning. This is typically the same path as validationData.

classes

List of classes or labels that the network should recognize. The class names must match the labels used in the annotation data.

nonRotables

List of classes without a defined orientation, for example a disk or a scratch.
Each entry must exactly match a class name defined in classes.

Arrow, Disk

Examples of class names.

imageScale

Specifies the image scaling factor relative to the original image size. Supported values are 0.5 and 1.0.

networkBase

Path to an existing trained network that should be used as the starting point for training.
Set the value to null to start training without an existing network.
Parameter Description

networkType

Specifies the network type. The available values depend on the target application and whether the object’s orientation or only its rotation needs to be determined.

For LumiScan VGR:

  • ORTML, ORTMM, ORTMS, ORTMXS: Oriented models. Use these when the object has a meaningful orientation, for example a distinguishable top side, that must be identified in addition to its position.

  • RRTML, RRTMM, RRTMS, RRTMXS: Rotated models. Use these when only the rotation of the object needs to be determined and no specific side or orientation needs to be distinguished.

  • LumiScanOOD (deprecated): Previous model type for oriented object detection.

  • LumiScanROD (deprecated): Previous model type for rotated object detection.

For AI Object Matching:

  • OOD: For objects with a meaningful orientation that must be identified.

  • ROD: For objects where only the rotation needs to be determined.

The final letters of the ORTM…​ and RRTM…​ values specify the network size:

  • L: large

  • M: medium

  • S: small

  • XS: extra small

Larger networks can provide higher accuracy but require more memory and generally take longer to train.

Enter the value exactly as listed, including the size suffix. ORTM and RRTM without a size suffix are not valid values.

schedule

Specifies the training schedule. Supported values are Full and Finetune.

  • Full: Performs a new training run. Use this option for initial training.

  • Finetune: Continues training using an existing network and the data specified in finetuneDataRoot. Use this option for subsequent refinement steps.

trainingProfile

Specifies the training profile and influences GPU memory usage. Currently, only one GPU is supported.

Available values:

  • B8x1: Default value.

  • B4x1: Use this option if training fails with B8x1 because of insufficient GPU memory or if training is using shared GPU memory.

Shared GPU memory usage can be checked in the Windows Task Manager. An unusually long training time can also indicate insufficient GPU memory.

deploymentConfiguration

Specifies the target deployment configuration. For LumiScan VGR, the value is LumiScanVGR and should not be changed.

tmpFiles

Path used to store temporary files during training. The application must have write permission for this location. Temporary files are deleted after training.