Changes for SKL1.2

[ksaur]

For #660

[ksaur]

Now we also have 2 other SKL1.2 bugs

=========================== short test summary info ============================
FAILED tests/test_sklearn_decomposition.py::TestSklearnMatrixDecomposition::test_pca_converter_mle_whiten - sklearn.utils._param_validation.InvalidParameterError: The 'whiten' parameter of PCA must be an instance of 'bool', an instance of 'numpy.bool_' or an instance of 'int'. Got 'arbitrary-variance' instead.
FAILED tests/test_sklearn_normalizer_converter.py::TestSklearnNormalizer::test_normalizer_converter_raises_wrong_type - sklearn.utils._param_validation.InvalidParameterError: The 'norm' parameter of Normalizer must be a str among {'l1', 'max', 'l2'}. Got 'invalid' instead.
===== 2 failed, 356 passed, 272 skipped, 71 warnings in 184.05s (0:03:04) ======

[ksaur]

Fixed those two.

And now:

=========================== short test summary info ===========================
FAILED tests/test_backends.py::TestBackends::test_onnx_no_test_data_double - AttributeError: 'DecisionTreeRegressor' object has no attribute 'n_features_'
FAILED tests/test_backends.py::TestBackends::test_onnx_no_test_data_float - AttributeError: 'DecisionTreeRegressor' object has no attribute 'n_features_'
FAILED tests/test_extra_conf.py::TestExtraConf::test_onnx_deafault_n_threads - AttributeError: 'DecisionTreeRegressor' object has no attribute 'n_features_'
FAILED tests/test_extra_conf.py::TestExtraConf::test_pandas_batch_onnxml - AttributeError: 'DecisionTreeRegressor' object has no attribute 'n_features_'
===== 4 failed, 522 passed, 104 skipped, 94 warnings in 287.73s (0:04:47) =====

[ksaur]

@interesaaat this last one seems to be a conflict with sklearn-onnx and SKL1.2. Double check and we'll open an issue?

The error 'DecisionTreeRegressor' object has no attribute 'n_features_' is in sklearn-onnx convert_sklearn_gradient_boosting_classifier

And from skl:
n_features_ is deprecated in 1.0 and will be removed in 1.2.

[ksaur]

Tagging onnx/sklearn-onnx#951 onnx/sklearn-onnx#952

[interesaaat]

@interesaaat this last one seems to be a conflict with sklearn-onnx and SKL1.2. Double check and we'll open an issue?

The error 'DecisionTreeRegressor' object has no attribute 'n_features_' is in sklearn-onnx convert_sklearn_gradient_boosting_classifier

And from skl: n_features_ is deprecated in 1.0 and will be removed in 1.2.

Oh I thought that we already fixed this. Let me check.

[interesaaat]

Yea we fixed in our code but probably sklern-onnx has not upgraded yet. Good job finding this!

[ksaur]

Yes thanks to @mshr-h in #610 we had n_features_ already deprecated on our side!

[mshr-h]

I've checked load_bostom() API docs and it says there're two alternatives for that, fetch_california_housing() and fetch_openml(name="house_prices", as_frame=True) which is called Ames housing dataset.
The former has 20,640 samples in total. But the latter has 506 which is identical to the Boston dataset.
I guess the Ames housing dataset is a better alternative for us because we want to make sure the output of the HB model is close enough to the output of the SKL model.

[ksaur]

@mshr-h I tried the Ames dataset but it has various types, objects:

>>> X_train.dtypes
Id             float64
MSSubClass     float64
MSZoning        object
LotFrontage    float64
                ...   
MiscFeature     object

and i want to keep it simple. rather than encoding, etc. Boston was all float64 and so is Cali.

In the meantime, my onnx/sklearn-onnx#952 was merged (but not released yet).
@interesaaat should we try to change the pipeline to pull from head to get the change?

[ksaur]

I thought these all passed fine (works on my machine!) but it seems with the new skl2onnx we have one more:

TestONNXSVC.test_logistic_regression_onnxml_multi_torch

        for name in names:
>           new_names.append(map[name])
E           KeyError: 'probabilities'`