Fix wrongly handled attribute in MVN and test generating scripts (onnx#1877)

* fix

* Add attribute

* update

* Resolve comments

Author: Raymond Yang

docs/Changelog.md

@@ -8761,8 +8761,8 @@ This version of the operator has been available since version 9 of the default O
 #### Attributes
 
 <dl>
-<dt><tt>axes</tt> : list of ints</dt>
-<dd>A list of integers, along which to reduce. The default is to reduce over all the dimensions of the input tensor. Use [0,2,3] (without C axis for N-D cases) for calculating means and variances along channels. Two variables with the same C-coordinate are associated with the same mean and variance.</dd>
+<dt><tt>axes</tt> : list of ints (default is ['0', '2', '3'])</dt>
+<dd>A list of integers, along which to reduce. The default is to caculate along axes [0,2,3] for calculating mean and variance along each channel. Two variables with the same C-coordinate are associated with the same mean and variance.</dd>
 </dl>
 
 #### Inputs

docs/Operators.md

@@ -6945,8 +6945,8 @@ This version of the operator has been available since version 9 of the default O
 #### Attributes
 
 <dl>
-<dt><tt>axes</tt> : list of ints</dt>
-<dd>A list of integers, along which to reduce. The default is to reduce over all the dimensions of the input tensor. Use [0,2,3] (without C axis for N-D cases) for calculating means and variances along channels. Two variables with the same C-coordinate are associated with the same mean and variance.</dd>
+<dt><tt>axes</tt> : list of ints (default is ['0', '2', '3'])</dt>
+<dd>A list of integers, along which to reduce. The default is to caculate along axes [0,2,3] for calculating mean and variance along each channel. Two variables with the same C-coordinate are associated with the same mean and variance.</dd>
 </dl>
 
 #### Inputs
@@ -6998,10 +6998,10 @@ input_data = np.array([[[[0.8439683], [0.5665144], [0.05836735]],
     [[0.69248444], [0.54119414], [0.07513223]]]], dtype=np.float32)
 
 # Calculate expected output data
-data_mean = np.mean(input_data, axis=(0, 1, 2, 3), keepdims=1)
+data_mean = np.mean(input_data, axis=(0, 2, 3), keepdims=1)
 data_mean_squared = np.power(data_mean, 2)
 data_squared = np.power(input_data, 2)
-data_squared_mean = np.mean(data_squared, axis=(0, 1, 2, 3), keepdims=1)
+data_squared_mean = np.mean(data_squared, axis=(0, 2, 3), keepdims=1)
 std = np.sqrt(data_squared_mean - data_mean_squared)
 expected_output = (input_data - data_mean) / (std + 1e-9)
 

docs/TestCoverage.md

@@ -3691,10 +3691,10 @@ input_data = np.array([[[[0.8439683], [0.5665144], [0.05836735]],
     [[0.69248444], [0.54119414], [0.07513223]]]], dtype=np.float32)
 
 # Calculate expected output data
-data_mean = np.mean(input_data, axis=(0, 1, 2, 3), keepdims=1)
+data_mean = np.mean(input_data, axis=(0, 2, 3), keepdims=1)
 data_mean_squared = np.power(data_mean, 2)
 data_squared = np.power(input_data, 2)
-data_squared_mean = np.mean(data_squared, axis=(0, 1, 2, 3), keepdims=1)
+data_squared_mean = np.mean(data_squared, axis=(0, 2, 3), keepdims=1)
 std = np.sqrt(data_squared_mean - data_mean_squared)
 expected_output = (input_data - data_mean) / (std + 1e-9)
 

onnx/backend/test/case/node/__init__.py

@@ -29,7 +29,7 @@ def function_expand_helper(node,  # type: NodeProto
     node_list = []
     input_names_map = dict()
     output_names_map = dict()
-    attribute_map = node.attribute
+    attribute_map = dict((a.name, a) for a in node.attribute)
 
     for idx in range(len(function_proto.input)):
         input_names_map[function_proto.input[idx]] = node.input[idx] \
@@ -77,6 +77,12 @@ def function_testcase_helper(node, name):  # type: (NodeProto, Text) -> List[Nod
         return []
     function_proto = schema.function_body  # type: ignore
 
+    for attr in schema.attributes:
+        if attr in [a.name for a in node.attribute]:
+            continue
+        if schema.attributes[attr].default_value:
+            node.attribute.extend([schema.attributes[attr].default_value])
+
     # function_proto.attributes
     node_list = function_expand_helper(node, function_proto, op_prefix)
     return node_list

onnx/backend/test/case/node/meanvariancenormalization.py

@@ -31,10 +31,10 @@ def export():  # type: () -> None
             [[0.69248444], [0.54119414], [0.07513223]]]], dtype=np.float32)
 
         # Calculate expected output data
-        data_mean = np.mean(input_data, axis=(0, 1, 2, 3), keepdims=1)
+        data_mean = np.mean(input_data, axis=(0, 2, 3), keepdims=1)
         data_mean_squared = np.power(data_mean, 2)
         data_squared = np.power(input_data, 2)
-        data_squared_mean = np.mean(data_squared, axis=(0, 1, 2, 3), keepdims=1)
+        data_squared_mean = np.mean(data_squared, axis=(0, 2, 3), keepdims=1)
         std = np.sqrt(data_squared_mean - data_mean_squared)
         expected_output = (input_data - data_mean) / (std + 1e-9)
 

onnx/backend/test/data/node/test_mvn/model.onnx

@@ -1,4 +1,4 @@
-backend-test:g
+backend-test:g
 !
 
X
Y"MeanVarianceNormalizationtest_mvnZ
 
X
@@ -12,4 +12,4 @@
 
 
 
-
B	
+
B

onnx/backend/test/data/node/test_mvn_expanded/model.onnx

@@ -1 +1 @@
-

B
YJlY�~?��>E�������W�����"��X?�u�?uª?[m�2��EG�Y�>�Ŀ�å>�ƞ?&��?T���+Q��.h�?�h>�hk?H�Q?�-����?S*�=���
+

B
YJl�d�?�;�>:�ſr����d���>D��>6nQ?��[?:���#rc�vyH�E3U?>	����,?�UD?�Pi?�ҿ�o����?��>�2�?�ϗ?��m����=msþ���

onnx/backend/test/data/node/test_mvn_expanded/test_data_set_0/output_0.pb

@@ -1776,6 +1776,8 @@ static const char* mvn_ver9_doc = R"DOC(
       on the input tensor X using formula: <br/> ``` (X-EX)/sqrt(E(X-EX)^2) ```
 )DOC";
 
+static std::vector<int64_t> mvn_default_axes = {0, 2, 3};
+
 ONNX_OPERATOR_SET_SCHEMA(
     MeanVarianceNormalization,
     9,
@@ -1785,13 +1787,12 @@ ONNX_OPERATOR_SET_SCHEMA(
         .Output(0, "Y", "Output tensor", "T")
         .Attr(
             "axes",
-            "A list of integers, along which to reduce. The default is to reduce over "
-            "all the dimensions of the input tensor. Use [0,2,3] (without C axis for "
-            "N-D cases) for calculating means and variances along channels. Two "
-            "variables with the same C-coordinate are associated "
-            "with the same mean and variance.",
+            "A list of integers, along which to reduce. The default is to "
+            "caculate along axes [0,2,3] for calculating mean and variance "
+            "along each channel. Two variables with the same C-coordinate "
+            "are associated with the same mean and variance.",
             AttributeProto::INTS,
-            OPTIONAL)
+            mvn_default_axes)
         .TypeConstraint(
             "T",
             {"tensor(float16)", "tensor(float)", "tensor(double)"},