TreeEnsembleRegressor/Classifier have been deprecated in favour of TreeEnsemble

[khoover]

The decision tree converters currently default to using the deprecated operator types, even in op versions where they've both been long-deprecated (didn't realize it was under the ml ops, not the regular ones). I believe the fix would be to use a default of op_type=None, plus if op_type is None: op_type = "TreeEnsemble" if op_version >= 5 else "TreeEnsemble...". There might be more that needs to happen for the node attributes, I haven't looked at the differences between the old op types and the new one too close.

[khoover]

Related to #1051

[xadupre]

Would you like to contribute?

[khoover]

I could give it a shot, sure. Do we have any examples in the repo of converters that have different outputs based on the op_version? Also, how are converters meant to determine the opset version? I see the node container has a target opset, but then the existing converter for e.g. GradientBoosting requires the opset to use be passed in as a separate parameter?

[xadupre]

This file may be interesting: https://github.com/onnx/sklearn-onnx/blob/main/skl2onnx/operator_converters/label_encoder.py#L25.

[khoover]

@xadupre is there a CLA for this project? Legal would want to review it before I submit any PRs.

[xadupre]

That should be covered by the license (https://github.com/onnx/sklearn-onnx/blob/main/LICENSE). Did I miss something?

[khoover]

Some projects have a separate one from the project license (e.g. onnxruntime), just making sure that isn't the case here.

[xadupre]

sklearn-onnx follows the same rules as onnx: https://lfaidata.foundation/.