Custom trained .pth model for labeling images on Azure ML

Question

Hello,

Can i use a locally trained custom model.pth/ model.onnx file for labeling my images on Azure ML? Is it possible? If yes, How?

Answer 1

Azure Machine Learning’s built-in data labeling tool does not support plugging in a custom .pth or .onnx model directly as the labeling engine.

Supported image labeling capabilities in Azure Machine Learning are:

• Human-in-the-loop labeling
• ML-assisted labeling using the service’s own mechanisms

These projects operate over images stored in an Azure Blob Storage–backed datastore or an MLTable data asset, and then export labels as an Azure Machine Learning dataset (for example, COCO) that can be used later to train or evaluate custom models. The documentation describes creating and managing labeling projects, but does not provide a way to upload or attach a custom PyTorch/ONNX model to drive the labeling UI.

To use a custom .pth/.onnx model in the broader workflow, the supported pattern is:

1. Use Azure Machine Learning data labeling to create and export labeled datasets.
2. Train or fine-tune custom models (including PyTorch models exported to ONNX) in Azure Machine Learning or elsewhere.
3. Use those models for inference or deployment (for example, ONNX with ONNX Runtime, or Windows ML), but not as pluggable labelers inside the labeling tool itself.

---

References:

• Set up an image labeling project
• Build and deploy custom document processing models on Azure
• ONNX and Azure Machine Learning