Fixed Linear classifier converter to resolve RS5 test error in LinearSVC (#194)

Author: prabhat00155

onnxmltools/convert/common/shape_calculator.py

@@ -33,31 +33,30 @@ def calculate_linear_classifier_output_shapes(operator):
     N = operator.inputs[0].type.shape[0]
 
     class_labels = operator.raw_operator.classes_
-    number_of_classes = len(class_labels)
     if all(isinstance(i, np.ndarray) for i in class_labels):
         class_labels = np.concatenate(class_labels)
     if all(isinstance(i, (six.string_types, six.text_type)) for i in class_labels):
         operator.outputs[0].type = StringTensorType(shape=[N])
-        if number_of_classes > 2 or operator.type != 'SklearnLinearSVC':
+        if len(class_labels) > 2 or operator.type != 'SklearnLinearSVC':
             # For multi-class classifier, we produce a map for encoding the probabilities of all classes
             if operator.target_opset < 7:
                 operator.outputs[1].type = DictionaryType(StringTensorType([1]), FloatTensorType([1]))
             else:
                 operator.outputs[1].type = SequenceType(DictionaryType(StringTensorType([]), FloatTensorType([])), N)
         else:
-            # For binary LinearSVC, we produce the probability tensor
-            operator.outputs[1].type = FloatTensorType(shape=[N, number_of_classes])
+            # For binary LinearSVC, we produce probability of the positive class
+            operator.outputs[1].type = FloatTensorType(shape=[N, 1])
     elif all(isinstance(i, (numbers.Real, bool, np.bool_)) for i in class_labels):
         operator.outputs[0].type = Int64TensorType(shape=[N])
-        if number_of_classes > 2 or operator.type != 'SklearnLinearSVC':
+        if len(class_labels) > 2 or operator.type != 'SklearnLinearSVC':
             # For multi-class classifier, we produce a map for encoding the probabilities of all classes
             if operator.target_opset < 7:
                 operator.outputs[1].type = DictionaryType(Int64TensorType([1]), FloatTensorType([1]))
             else:
                 operator.outputs[1].type = SequenceType(DictionaryType(Int64TensorType([]), FloatTensorType([])), N)
         else:
-            # For binary LinearSVC, we produce the probability tensor
-            operator.outputs[1].type = FloatTensorType(shape=[N, number_of_classes])
+            # For binary LinearSVC, we produce probability of the positive class
+            operator.outputs[1].type = FloatTensorType(shape=[N, 1])
     else:
         raise ValueError('Unsupported or mixed label types')
 

onnxmltools/convert/sklearn/operator_converters/LinearClassifier.py

@@ -7,6 +7,7 @@
 import numpy as np
 import six, numbers
 from ...common._registration import register_converter
+from ....proto import onnx_proto
 
 
 def convert_sklearn_linear_classifier(scope, operator, container):
@@ -67,9 +68,18 @@ def convert_sklearn_linear_classifier(scope, operator, container):
     probability_tensor_name = scope.get_unique_variable_name('probability_tensor')
 
     if op.__class__.__name__ == 'LinearSVC' and op.classes_.shape[0] <= 2:
+        raw_scores_tensor_name = scope.get_unique_variable_name('raw_scores_tensor')
+        positive_class_index_name = scope.get_unique_variable_name('positive_class_index')
+
+        container.add_initializer(positive_class_index_name, onnx_proto.TensorProto.INT64,
+                              [], [1])
+
         container.add_node(classifier_type, operator.inputs[0].full_name,
-                           [label_name, operator.outputs[1].full_name],
+                           [label_name, raw_scores_tensor_name],
                            op_domain='ai.onnx.ml', **classifier_attrs)
+        container.add_node('ArrayFeatureExtractor', [raw_scores_tensor_name, positive_class_index_name],
+                           operator.outputs[1].full_name, name=scope.get_unique_operator_name('ArrayFeatureExtractor'),
+                           op_domain='ai.onnx.ml')
     else:
         container.add_node(classifier_type, operator.inputs[0].full_name,
                            [label_name, probability_tensor_name],