Problem

Today in ML.NET, for scenarios where a transform isn't implemented, you can build a custom transform either using:

• Expression transform
• CustomMapping transform

However, neither is exportable to ONNX making it difficult to share and reuse code for ONNX models.

Existing solutions

In Python, this is currently possible via onnxruntime-extensions.

Proposed solution

Enable .NET developers to write export code that is embeddable as part of the exported ONNX model making any pre/post-processing code as well as any other custom code written to be used as part of an ML.NET pipeline is exportable to ONNX.