Pipeline convert

Hello! Following pipeline doesn’t convert for me:

classifier = LogisticRegression(C=0.01, 
                                random_state=SEED,
                                class_weight=dict(zip([False, True], c_w)),
                                n_jobs=40,
                                max_iter=500
                                solver='lbfgs',
                                verbose = 10,
                                tol=1e-3)

numeric_transformer = Pipeline(steps=[
    ('imputer', SimpleImputer(strategy='median')),
    ('scaler', StandardScaler())
])

categorical_transformer = Pipeline(steps=[
    ('imputer', SimpleImputer(strategy='constant', fill_value='nan')),
    ('onehot', OneHotEncoder(sparse=True, handle_unknown='ignore')),
    ('tsvd', TruncatedSVD(n_components=60,
                          algorithm='arpack',
                          random_state=SEED,
                          tol=1e-4))
])

preprocessor = ColumnTransformer(
    transformers=[
        ('num', numeric_transformer, numeric_features),
        ('cat', categorical_transformer, categorical_features)
    ])

model = Pipeline(steps=[
    ('precprocessor', preprocessor),
    ('classifier', classifier)
])

model.fit(X_train, y_train)

initial_type = [('s1', StringTensorType([1,15])), ('float1', FloatTensorType([1,21])),
                ('s2', StringTensorType([1,27])), ('float2', FloatTensorType([1,30]))]

onnx = convert_sklearn(model, initial_types=initial_type)

This strange error occurs:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-134-d0cc22f876ed> in <module>
----> 1 onnx = convert_sklearn(model, initial_types=initial_type)

/opt/conda/lib/python3.6/site-packages/skl2onnx/convert.py in convert_sklearn(model, name, initial_types, doc_string, target_opset, custom_conversion_functions, custom_shape_calculators)
     62     target_opset = target_opset if target_opset else get_opset_number_from_onnx()
     63     # Parse scikit-learn model as our internal data structure (i.e., Topology)
---> 64     topology = parse_sklearn(model, initial_types, target_opset, custom_conversion_functions, custom_shape_calculators)
     65 
     66     # Infer variable shapes

/opt/conda/lib/python3.6/site-packages/skl2onnx/_parse.py in parse_sklearn(model, initial_types, target_opset, custom_conversion_functions, custom_shape_calculators)
    294 
    295     # Parse the input scikit-learn model as a Topology object.
--> 296     outputs = _parse_sklearn(scope, model, inputs)
    297 
    298     # THe object raw_model_container is a part of the topology we're going to return. We use it to store the outputs of

/opt/conda/lib/python3.6/site-packages/skl2onnx/_parse.py in _parse_sklearn(scope, model, inputs)
    235     '''
    236     if isinstance(model, pipeline.Pipeline):
--> 237         return _parse_sklearn_pipeline(scope, model, inputs)
    238     elif isinstance(model, pipeline.FeatureUnion):
    239         return _parse_sklearn_feature_union(scope, model, inputs)

/opt/conda/lib/python3.6/site-packages/skl2onnx/_parse.py in _parse_sklearn_pipeline(scope, model, inputs)
    160     '''
    161     for step in model.steps:
--> 162         inputs = _parse_sklearn(scope, step[1], inputs)
    163     return inputs
    164 

/opt/conda/lib/python3.6/site-packages/skl2onnx/_parse.py in _parse_sklearn(scope, model, inputs)
    239         return _parse_sklearn_feature_union(scope, model, inputs)
    240     elif isinstance(model, ColumnTransformer):
--> 241         return _parse_sklearn_column_transformer(scope, model, inputs)
    242     elif type(model) in sklearn_classifier_list and type(model) not in [LinearSVC, SVC, NuSVC]:
    243         probability_tensor = _parse_sklearn_simple_model(scope, model, inputs)

/opt/conda/lib/python3.6/site-packages/skl2onnx/_parse.py in _parse_sklearn_column_transformer(scope, model, inputs)
    213         transform_inputs = _fetch_input_slice(scope, inputs, column_indices)
    214         transformed_result_names.append(_parse_sklearn_simple_model(scope, model.named_transformers_[name],
--> 215                                                                     transform_inputs)[0])
    216     # Create a Concat ONNX node
    217     concat_operator = scope.declare_local_operator('SklearnConcat')

/opt/conda/lib/python3.6/site-packages/skl2onnx/_parse.py in _parse_sklearn_simple_model(scope, model, inputs)
    131     :return: A list of output variables which will be passed to next stage
    132     '''
--> 133     this_operator = scope.declare_local_operator(_get_sklearn_operator_name(type(model)), model)
    134     this_operator.inputs = inputs
    135 

/opt/conda/lib/python3.6/site-packages/skl2onnx/_parse.py in _get_sklearn_operator_name(model_type)
    118     '''
    119     if model_type not in sklearn_operator_name_map:
--> 120         raise ValueError("No proper operator name found for '%s'" % model_type)
    121     return sklearn_operator_name_map[model_type]
    122 

ValueError: No proper operator name found for '<class 'sklearn.pipeline.Pipeline'>'

[xadupre]

Working on it right now.

[xadupre]

It should be fixed except one operator I could not convert but you can look the following unit test: https://github.com/xadupre/sklearn-onnx/blob/i26/tests/test_SklearnPipeline.py. ONNX does not implement an imputer on strings, see https://github.com/onnx/onnx/blob/master/docs/Operators-ml.md#ai.onnx.ml.Imputer.