Merge branch 'master' into master

.azure-pipelines/linux-conda-CI.yml

@@ -26,7 +26,7 @@ jobs:
       Py39-Onnx1101-Rt190-Skl0242:
         do.bench: '0'
         python.version: '3.9'
-        numpy.version: '>=1.18.1'
+        numpy.version: '<1.21.0'
         onnx.version: '==1.10.1'  # 'test'
         onnx.target_opset: ''
         onnxrt.version: 'onnxruntime==1.9.0'  # '-i https://test.pypi.org/simple/ ort-nightly'
@@ -36,17 +36,17 @@ jobs:
       Py39-Onnx1101-Rt181-Skl0242:
         do.bench: '0'
         python.version: '3.9'
-        numpy.version: '>=1.18.1'
+        numpy.version: '<1.21.0'
         onnx.version: '==1.10.1'  # 'test'
         onnx.target_opset: ''
         onnxrt.version: 'onnxruntime==1.8.1'  # '-i https://test.pypi.org/simple/ ort-nightly'
-        sklearn.version: '>=0.24.2'
+        sklearn.version: '>=0.24.2 lightgbm==3.2.0'
         onnxcc.version: '>=1.8.1'  # git
         run.example: '0'
       Py39-Onnx190-Rt180-Skl0242:
         do.bench: '0'
         python.version: '3.9'
-        numpy.version: '>=1.18.1'
+        numpy.version: '<1.21.0'
         onnx.version: '==1.9.0'
         onnx.target_opset: ''
         onnxrt.version: 'onnxruntime>=1.8.1'  # '-i https://test.pypi.org/simple/ ort-nightly'
@@ -56,7 +56,7 @@ jobs:
       Py39-Onnx190-Rt170-Skl0242:
         do.bench: '0'
         python.version: '3.9'
-        numpy.version: '>=1.18.1'
+        numpy.version: '<1.21.0'
         onnx.version: '==1.9.0'
         onnx.target_opset: ''
         onnxrt.version: 'onnxruntime==1.7.0'  # '-i https://test.pypi.org/simple/ ort-nightly'
@@ -66,7 +66,7 @@ jobs:
       Py38-Onnx181-Rt170-Skl0242:
         do.bench: '0'
         python.version: '3.8'
-        numpy.version: '>=1.18.1'
+        numpy.version: '<1.21.0'
         onnx.version: '==1.8.1'
         onnx.target_opset: ''
         onnxrt.version: 'onnxruntime==1.7.0'  # '-i https://test.pypi.org/simple/ ort-nightly'
@@ -76,7 +76,7 @@ jobs:
       Py38-Onnx181-Rt170-Skl0241:
         do.bench: '0'
         python.version: '3.8'
-        numpy.version: '>=1.18.1'
+        numpy.version: '<1.21.0'
         onnx.version: '==1.8.1'
         onnx.target_opset: ''
         onnxrt.version: 'onnxruntime==1.7.0'  # '-i https://test.pypi.org/simple/ ort-nightly'
@@ -86,7 +86,7 @@ jobs:
       Py38-Onnx181-Rt160-Skl0240:
         do.bench: '0'
         python.version: '3.8'
-        numpy.version: '>=1.18.1'
+        numpy.version: '<1.21.0'
         onnx.version: '==1.8.1'
         onnx.target_opset: ''
         onnxrt.version: 'onnxruntime==1.6.0'
@@ -96,7 +96,7 @@ jobs:
       Py38-Onnx170-Rt160-Skl0240:
         do.bench: '0'
         python.version: '3.8'
-        numpy.version: '>=1.18.1'
+        numpy.version: '<1.21.0'
         onnx.version: '==1.7.0'
         onnx.target_opset: ''
         onnxrt.version: 'onnxruntime==1.6.0'
@@ -106,7 +106,7 @@ jobs:
       Py37-Onnx170-Rt151-Skl0232:
         do.bench: '0'
         python.version: '3.7'
-        numpy.version: '>=1.18.1'
+        numpy.version: '<1.21.0'
         onnx.version: '==1.7.0'
         onnx.target_opset: ''
         onnxrt.version: 'onnxruntime==1.5.1'
@@ -116,7 +116,7 @@ jobs:
       Py37-Onnx170-Rt140-Skl0232:
         do.bench: '0'
         python.version: '3.7'
-        numpy.version: '>=1.18.1'
+        numpy.version: '<1.21.0'
         onnx.version: '==1.7.0'
         onnx.target_opset: ''
         onnxrt.version: 'onnxruntime==1.4.0'  # '-i https://test.pypi.org/simple/ ort-nightly'
@@ -126,7 +126,7 @@ jobs:
       Py37-Onnx170-Rt130-Skl0231:
         do.bench: '0'
         python.version: '3.7'
-        numpy.version: '>=1.18.1'
+        numpy.version: '<1.21.0'
         onnx.version: '==1.7.0'
         onnx.target_opset: ''
         onnxrt.version: 'onnxruntime==1.3.0'
@@ -136,7 +136,7 @@ jobs:
       Py37-Onnx170-Rt120-Skl0230-Op11:
         do.bench: '0'
         python.version: '3.7'
-        numpy.version: '>=1.18.1'
+        numpy.version: '<1.21.0'
         onnx.version: '==1.7.0'
         onnx.target_opset: '11'
         onnxrt.version: 'onnxruntime==1.2.0'
@@ -146,7 +146,7 @@ jobs:
       Py37-Onnx160-Rt120-Skl0230:
         do.bench: '0'
         python.version: '3.7'
-        numpy.version: '>=1.18.1'
+        numpy.version: '<1.21.0'
         onnx.version: '==1.6.0'
         onnx.target_opset: ''
         onnxcc.version: '==1.7.0'
@@ -156,7 +156,7 @@ jobs:
       Py37-Onnx160-Rt111-Skl0222:
         do.bench: '0'
         python.version: '3.7'
-        numpy.version: '>=1.18.1'
+        numpy.version: '<1.21.0'
         onnx.version: '==1.6.0'
         onnx.target_opset: ''
         onnxcc.version: '==1.7.0'
@@ -166,7 +166,7 @@ jobs:
       Py37-Onnx160-Rt100-Skl0213:
         do.bench: '0'
         python.version: '3.7'
-        numpy.version: '==1.18.1'
+        numpy.version: '<1.21.0'
         onnx.version: '==1.6.0'
         onnx.target_opset: ''
         onnxcc.version: '==1.7.0'
@@ -257,6 +257,8 @@ jobs:
       echo "---------------"
       pip show numpy
       echo "---------------"
+      pip show scipy
+      echo "---------------"
       pip show pandas
       echo "---------------"
       pip show onnx

skl2onnx/__init__.py

@@ -3,7 +3,7 @@
 """
 Main entry point to the converter from the *scikit-learn* to *onnx*.
 """
-__version__ = "1.10.0"
+__version__ = "1.10.1"
 __author__ = "Microsoft"
 __producer__ = "skl2onnx"
 __producer_version__ = __version__

skl2onnx/_parse.py

@@ -413,17 +413,9 @@ def _parse_sklearn_random_trees_embedding(scope, model, inputs,
     return res
 
 
-def _parse_sklearn_classifier(scope, model, inputs, custom_parsers=None):
-    options = scope.get_options(model, dict(zipmap=True))
-    no_zipmap = (
-        (isinstance(options['zipmap'], bool) and not options['zipmap']) or
-        (model.__class__ in [NuSVC, SVC] and not model.probability))
-    probability_tensor = _parse_sklearn_simple_model(
-        scope, model, inputs, custom_parsers=custom_parsers)
-    if no_zipmap:
-        return probability_tensor
-
-    if options['zipmap'] == 'columns':
+def _apply_zipmap(zipmap_options, scope, model, input_type,
+                  probability_tensor):
+    if zipmap_options == 'columns':
         zipmap_operator = scope.declare_local_operator('SklearnZipMapColumns')
         classes = get_label_classes(scope, model)
         classes_names = get_label_classes(scope, model, node_names=True)
@@ -456,10 +448,11 @@ def _parse_sklearn_classifier(scope, model, inputs, custom_parsers=None):
         label_type = StringTensorType([None])
 
     zip_label = scope.declare_local_variable('output_label', label_type)
-    zipmap_operator.outputs.append(zip_label)
+    if len(probability_tensor) == 2:
+        zipmap_operator.outputs.append(zip_label)
 
-    if options['zipmap'] == 'columns':
-        prob_type = probability_tensor[1].type
+    if zipmap_options == 'columns':
+        prob_type = probability_tensor[-1].type
         for cl in classes_names:
             output_cl = scope.declare_local_variable(cl, prob_type.__class__())
             zipmap_operator.outputs.append(output_cl)
@@ -468,24 +461,84 @@ def _parse_sklearn_classifier(scope, model, inputs, custom_parsers=None):
             'output_probability',
             SequenceType(
                 DictionaryType(
-                    label_type, guess_tensor_type(inputs[0].type))))
+                    label_type, guess_tensor_type(input_type))))
         zipmap_operator.outputs.append(zip_probability)
 
     zipmap_operator.init_status(is_evaluated=True)
     return zipmap_operator.outputs
 
 
+def _parse_sklearn_classifier(scope, model, inputs, custom_parsers=None):
+    options = scope.get_options(model, dict(zipmap=True))
+    no_zipmap = (
+        (isinstance(options['zipmap'], bool) and not options['zipmap']) or
+        (model.__class__ in [NuSVC, SVC] and not model.probability))
+    probability_tensor = _parse_sklearn_simple_model(
+        scope, model, inputs, custom_parsers=custom_parsers)
+    if no_zipmap:
+        return probability_tensor
+    return _apply_zipmap(
+        options['zipmap'], scope, model, inputs[0].type, probability_tensor)
+
+
 def _parse_sklearn_multi_output_classifier(scope, model, inputs,
                                            custom_parsers=None):
     alias = _get_sklearn_operator_name(type(model))
     this_operator = scope.declare_local_operator(alias, model)
     this_operator.inputs = inputs
-    label = scope.declare_local_variable("label", Int64TensorType())
-    proba = scope.declare_local_variable(
-        "probabilities", SequenceType(guess_tensor_type(inputs[0].type)))
+    guessed_output_type = guess_tensor_type(inputs[0].type)
+    label_type = Int64TensorType()
+    label = scope.declare_local_variable("label", label_type)
+
+    options = scope.get_options(model, dict(zipmap=True))
+    no_zipmap = isinstance(options['zipmap'], bool) and not options['zipmap']
+
+    if no_zipmap:
+        proba = scope.declare_local_variable(
+            "probabilities", SequenceType(guessed_output_type))
+        this_operator.outputs.append(label)
+        this_operator.outputs.append(proba)
+        return this_operator.outputs
+
+    warnings.warn(
+        "The current converter for class %r "
+        "creates a sequence of maps as an output."
+        "This is not allowed in standards ONNX specifications."
+        "" % model.__class__.__name__,
+        RuntimeWarning)
+
     this_operator.outputs.append(label)
+    proba = scope.declare_local_variable(
+        "output_probabilities", SequenceType(guessed_output_type))
     this_operator.outputs.append(proba)
-    return this_operator.outputs
+
+    zipmap = options.get('zipmap', False)
+    if zipmap == 'columns':
+        raise RuntimeError(
+            "Unable to convert model %r with option zipmap=%r." % (
+                model.__class__.__name__, zipmap))
+
+    # zipmap is True
+    zipped = []
+    for i in range(len(model.estimators_)):
+        output_name = scope.declare_local_variable(
+            "multi_output_%d" % i, guessed_output_type)
+        sequence_at = scope.declare_local_operator('SklearnSequenceAt')
+        sequence_at.inputs.append(proba)
+        sequence_at.outputs.append(output_name)
+        sequence_at.index = i
+        zipped.extend(
+            _apply_zipmap(
+                zipmap, scope, model.estimators_[i],
+                guessed_output_type, sequence_at.outputs))
+
+    sequence_build = scope.declare_local_operator('SklearnSequenceConstruct')
+    sequence_build.inputs.extend(zipped)
+    proba_zipped = scope.declare_local_variable(
+        "probabilities",
+        SequenceType(DictionaryType(label_type, guessed_output_type)))
+    sequence_build.outputs.append(proba_zipped)
+    return sequence_build.outputs
 
 
 def _parse_sklearn_gaussian_process(scope, model, inputs, custom_parsers=None):

skl2onnx/operator_converters/__init__.py

@@ -50,6 +50,7 @@
 from . import ransac_regressor
 from . import replace_op
 from . import scaler_op
+from . import sequence
 from . import sgd_classifier
 from . import stacking
 from . import support_vector_machines
@@ -108,6 +109,7 @@
     ransac_regressor,
     replace_op,
     scaler_op,
+    sequence,
     sgd_classifier,
     stacking,
     support_vector_machines,

skl2onnx/operator_converters/calibrated_classifier_cv.py

@@ -12,6 +12,7 @@
     guess_numpy_type, Int64TensorType, guess_proto_type)
 from ..common._registration import register_converter
 from .._supported_operators import sklearn_operator_name_map
+from sklearn.ensemble import RandomForestClassifier
 
 
 def _handle_zeros(scope, container, concatenated_prob_name,
@@ -256,6 +257,8 @@ def convert_calibrated_classifier_base_estimator(scope, operator, container,
     #                           |                        |
     #                           V                        |
     #                        class_prob_tensor [M, C] <--'
+    model_proba = {RandomForestClassifier}
+
     if scope.get_options(operator.raw_operator, dict(nocl=False))['nocl']:
         raise RuntimeError(
             "Option 'nocl' is not implemented for operator '{}'.".format(
@@ -275,7 +278,8 @@ def convert_calibrated_classifier_base_estimator(scope, operator, container,
     prob_name = [None] * n_classes
 
     this_operator = scope.declare_local_operator(op_type, base_model)
-    if container.has_options(base_model, 'raw_scores'):
+    if (container.has_options(base_model, 'raw_scores') and
+            not type(base_model) in model_proba):
         container.add_options(id(base_model), {'raw_scores': True})
         scope.add_options(id(base_model), {'raw_scores': True})
     this_operator.inputs = operator.inputs

skl2onnx/operator_converters/multioutput.py

@@ -61,13 +61,12 @@ def convert_multi_output_classifier_converter(
                                    op_version=op_version)
                   for y in y_list]
 
-    label = OnnxConcat(*label_list, axis=1, op_version=op_version,
-                       output_names=[operator.outputs[0]])
-    label.add_to(scope=scope, container=container)
-
     # probabilities
     proba_list = [OnnxIdentity(y[1], op_version=op_version)
                   for y in y_list]
+    label = OnnxConcat(*label_list, axis=1, op_version=op_version,
+                       output_names=[operator.outputs[0]])
+    label.add_to(scope=scope, container=container)
 
     proba = OnnxSequenceConstruct(
         *proba_list, op_version=op_version,
@@ -79,4 +78,5 @@ def convert_multi_output_classifier_converter(
                    convert_multi_output_regressor_converter)
 register_converter('SklearnMultiOutputClassifier',
                    convert_multi_output_classifier_converter,
-                   options={'nocl': [False, True]})
+                   options={'nocl': [False, True],
+                            'zipmap': [False, True]})

skl2onnx/operator_converters/random_forest.py

@@ -191,7 +191,7 @@ def convert_sklearn_random_forest_classifier(
         elif use_raw_scores:
             raise RuntimeError(
                 "The converter cannot implement decision_function for "
-                "'{}'.".format(type(op)))
+                "'{}' and loss '{}'.".format(type(op), loss))
 
         input_name = operator.input_full_names
         if type(operator.inputs[0].type) == BooleanTensorType:

skl2onnx/operator_converters/sequence.py

@@ -0,0 +1,31 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from ..proto import onnx_proto
+from ..common._registration import register_converter
+from ..common._topology import Scope, Operator
+from ..common._container import ModelComponentContainer
+
+
+def convert_sklearn_sequence_at(scope: Scope, operator: Operator,
+                                container: ModelComponentContainer):
+    i_index = operator.index
+    index_name = scope.get_unique_variable_name("seq_at%d" % i_index)
+    container.add_initializer(
+        index_name, onnx_proto.TensorProto.INT64, [], [i_index])
+    container.add_node(
+        'SequenceAt', [operator.inputs[0].full_name, index_name],
+        operator.outputs[0].full_name,
+        name=scope.get_unique_operator_name('SequenceAt%d' % i_index))
+
+
+def convert_sklearn_sequence_construct(scope: Scope, operator: Operator,
+                                       container: ModelComponentContainer):
+    container.add_node(
+        'SequenceConstruct', [i.full_name for i in operator.inputs],
+        operator.outputs[0].full_name,
+        name=scope.get_unique_operator_name('SequenceConstruct'))
+
+
+register_converter('SklearnSequenceAt', convert_sklearn_sequence_at)
+register_converter(
+    'SklearnSequenceConstruct', convert_sklearn_sequence_construct)