Issue using OneHotEncoder with drop="if_binary"

[dahlbaek]

We've hit what appears to be an interesting bug that triggers when one uses OneHotEncoder with drop="if_binary". It can be replicated by dropping the following test in https://github.com/onnx/sklearn-onnx/blob/main/tests/test_sklearn_one_hot_encoder_converter.py

    @unittest.skipIf(
        pv.Version(ort_version) <= pv.Version("0.4.0"), reason="issues with shapes"
    )
    @unittest.skipIf(
        not one_hot_encoder_supports_drop(),
        reason="OneHotEncoder does not support drop in scikit versions < 0.21",
    )
    def test_one_hot_encoder_drop_if_binary(self):
        data = [
            ["c0.4", "c0.2", 0],
            ["c1.4", "c1.2", 0],
            ["c0.2", "c2.2", 1],
            ["c0.2", "c2.2", 1],
            ["c0.2", "c2.2", 1],
            ["c0.2", "c2.2", 1],
        ]
        test = [["c0.2", "c2.2", 1]]
        model = OneHotEncoder(categories="auto", drop="if_binary")
        model.fit(data)
        inputs = [
            ("input1", StringTensorType([None, 2])),
            ("input2", Int64TensorType([None, 1])),
        ]
        model_onnx = convert_sklearn(
            model, "one-hot encoder", inputs, target_opset=TARGET_OPSET
        )
        self.assertTrue(model_onnx is not None)
        dump_data_and_model(
            test,
            model,
            model_onnx,
            verbose=False,
            basename="SklearnOneHotEncoderMixedStringIntDrop",
        )

Investigating the stack trace, one winds up here

sklearn-onnx/skl2onnx/operator_converters/one_hot_encoder.py

Lines 161 to 163 in 895c3a7

	indices_to_keep = np.delete(
	np.arange(len(categories)), ohe_op.drop_idx_[index]
	)

with ohe_op.drop_idx_ = [None, None, 0], and therefore np.delete(np.arange(3), None).

[dahlbaek]

The test suite passes if one changes

sklearn-onnx/skl2onnx/operator_converters/one_hot_encoder.py

Lines 161 to 163 in 895c3a7

	indices_to_keep = np.delete(
	np.arange(len(categories)), ohe_op.drop_idx_[index]
	)

to

            if ohe_op.drop_idx_[index] is not None:
                indices_to_keep = np.delete(
                    np.arange(len(categories)), ohe_op.drop_idx_[index]
                )
            else:
                indices_to_keep = np.arange(len(categories))

[xadupre]

Closing this issue as the fix was merged into main branch.