What it does ONNXMLTools is a unified Python package that converts trained models from eight different ML toolkits—scikit-learn, TensorFlow, XGBoost, LightGBM, CatBoost, Apple Core ML, Spark ML, and H2O—into the ONNX format. For several of these, it’s essentially a thin wrapper around dedicated converters like tf2onnx and skl2onnx, with the value being that you only need to remember one API.

The interesting bit The opset negotiation logic is more subtle than it looks. You can request a target_opset, but the converter may return a model with a lower version if that’s the newest opset shared by all operators in your graph. It’s a pragmatic backwards-compatibility choice that prevents silent breakage.

Key highlights

• Supports Keras, scikit-learn, Core ML, LightGBM, XGBoost, CatBoost, H2O, libsvm, and Spark ML (experimental)
• TensorFlow and scikit-learn conversions delegate to tf2onnx and skl2onnx respectively
• PyTorch is explicitly out of scope—use PyTorch’s built-in exporter instead
• Requires Python 3.7+; needs the source toolkit installed at conversion time
• Unit tests verify against onnxruntime automatically

Caveats

• Core ML support is pinned to CoreMLTools 3.1 or lower, which is increasingly dated
• Spark ML support is marked experimental
• Installing from source requires setting ONNX_ML=1 before building the onnx dependency, an easy step to miss

Verdict Worth a look if you’re maintaining a polyglot ML pipeline and want one import to handle most of your conversion chores. Skip it if you’re all-in on PyTorch or need bleeding-edge Core ML support.