Multiclass Model with interactions leads to invalid model

[pschleiter]

• ebm2onnx version: 3.3.0
• python version: 3.9
• Operating System: Windows

Description

I like to convert an ExplainableBoostingClassifier for a multiclass problem (with interactions set to a value greater than 0) in to an onnx model. The created onnx_model can not be instancated in a onnxruntime, because some reshape nodes have a wrong shape tensor.

What I Did

Here is a minimal example demonstrating the issue:

# /// script
# requires-python = ">=3.9"
# dependencies = [
#     "ebm2onnx==3.3.0",
#     "interpret==0.6.14",
#     "onnx==1.18.0",
#     "onnxruntime==1.19.2",
#     "pandas==2.3.0",
#     "skl2onnx==1.19.1",
# ]
# ///

if __name__ == "__main__":
    import pandas as pd
    from interpret.glassbox import ExplainableBoostingClassifier
    import ebm2onnx

    # Small dataset
    X = pd.DataFrame(
        {
            "feature1": [0, 0, 1, 1] * 8,
            "feature2": [0] * 16 + [1] * 16,
        }
    )
    y = pd.Series([0] * 24 + [1] * 4 + [2] * 4)

    # Train model
    model = ExplainableBoostingClassifier(interactions=2)
    model.fit(X=X, y=y)

    # Convert mode to onnx
    onnx_model = ebm2onnx.to_onnx(
        model,
        ebm2onnx.get_dtype_from_pandas(X),
        predict_proba=True,
    )
    onnx_model.ir_version = 10  # Issue #22

    # Score with the onnx model
    import onnxruntime as rt

    session = rt.InferenceSession(onnx_model.SerializeToString())
    # Creating the session fails with the following error
    #   onnxruntime.capi.onnxruntime_pybind11_state.Fail: [ONNXRuntimeError] : 1 : FAIL : Node (Concat_1) Op (Concat) [ShapeInferenceError] Can't merge shape info. Both inferred and declared dimension have values but they differ. Inferred=1 Declared=3 Dimension=2

    print(
        list(
            filter(
                lambda i: i.name.startswith("score_reshape"),
                onnx_model.graph.initializer,
            )
        )
    )
    # Provides the following output:
    #   [
    #   "dims: 3\ndata_type: 7\nint64_data: -1\nint64_data: 1\nint64_data: 3\nname: \"score_reshape_0\"\n",
    #   "dims: 3\ndata_type: 7\nint64_data: -1\nint64_data: 1\nint64_data: 3\nname: \"score_reshape_1\"\n",
    #   "dims: 3\ndata_type: 7\nint64_data: -1\nint64_data: 1\nint64_data: 1\nname: \"score_reshape_2\"\n"
    #   ]
    # The tensor "score_reshape_2" is faulty if changing to [-1, 1, 3] manually the session can be create and scoring works