Add dilation attribute to MaxPool (onnx#1864)

* add dialation attribute to MaxPool

* add shape inference test for dilated MaxPool

* regen docs

* fix wrong formatting

* update test coverage

* fix dilation parameter for maxpool

* add dilation case for MaxPool op

* fix formatting

* regen docs

Author: karljang

docs/Changelog.md

@@ -9574,6 +9574,8 @@ This version of the operator has been available since version 10 of the default
 <dd>auto_pad must be either NOTSET, SAME_UPPER, SAME_LOWER or VALID. Where default value is NOTSET, which means explicit padding is used. SAME_UPPER or SAME_LOWER mean pad the input so that the output size match the input.In case of odd number add the extra padding at the end for SAME_UPPER and at the beginning for SAME_LOWER. VALID mean no padding. DEPRECATION NOTE: auto_pad is only intended to support legacy uses, and for framework authors, one is explicitly encouraged to use explicit padding specified in the pads attribute.</dd>
 <dt><tt>ceil_mode</tt> : int (default is 0)</dt>
 <dd>Wether to use ceil or floor (default) to compute the output shape.</dd>
+<dt><tt>dilations</tt> : list of ints</dt>
+<dd>Dilation value along each axis of filter.</dd>
 <dt><tt>kernel_shape</tt> : list of ints (required)</dt>
 <dd>The size of the kernel along each axis.</dd>
 <dt><tt>pads</tt> : list of ints</dt>

docs/Operators.md

@@ -6199,6 +6199,8 @@ Other versions of this operator: <a href="Changelog.md#MaxPool-1">MaxPool-1</a>,
 <dd>auto_pad must be either NOTSET, SAME_UPPER, SAME_LOWER or VALID. Where default value is NOTSET, which means explicit padding is used. SAME_UPPER or SAME_LOWER mean pad the input so that the output size match the input.In case of odd number add the extra padding at the end for SAME_UPPER and at the beginning for SAME_LOWER. VALID mean no padding. DEPRECATION NOTE: auto_pad is only intended to support legacy uses, and for framework authors, one is explicitly encouraged to use explicit padding specified in the pads attribute.</dd>
 <dt><tt>ceil_mode</tt> : int (default is 0)</dt>
 <dd>Wether to use ceil or floor (default) to compute the output shape.</dd>
+<dt><tt>dilations</tt> : list of ints</dt>
+<dd>Dilation value along each axis of filter.</dd>
 <dt><tt>kernel_shape</tt> : list of ints (required)</dt>
 <dd>The size of the kernel along each axis.</dd>
 <dt><tt>pads</tt> : list of ints</dt>
@@ -6325,6 +6327,38 @@ expect(node, inputs=[x], outputs=[y], name='test_maxpool_2d_default')
 </details>
 
 
+<details>
+<summary>maxpool_2d_dilations</summary>
+
+```python
+"""
+input_shape: [1, 1, 4, 4]
+output_shape: [1, 1, 2, 2]
+"""
+node = onnx.helper.make_node(
+    'MaxPool',
+    inputs=['x'],
+    outputs=['y'],
+    kernel_shape=[2, 2],
+    strides=[1, 1],
+    dilations=[2, 2]
+)
+x = np.array([[[
+    [1, 2, 3, 4],
+    [5, 6, 7, 8],
+    [9, 10, 11, 12],
+    [13, 14, 15, 16],
+]]]).astype(np.float32)
+y = np.array([[[
+    [11, 12],
+    [15, 16]]]]).astype(np.float32)
+
+expect(node, inputs=[x], outputs=[y], name='test_maxpool_2d_dilations')
+```
+
+</details>
+
+
 <details>
 <summary>maxpool_2d_pads</summary>
 

docs/TestCoverage.md

@@ -3142,7 +3142,7 @@ expect(node, inputs=[data_0, data_1], outputs=[result],
 
 
 ### MaxPool
-There are 13 test cases, listed as following:
+There are 14 test cases, listed as following:
 <details>
 <summary>maxpool_1d_default</summary>
 
@@ -3224,6 +3224,36 @@ y = pool(padded, x_shape, kernel_shape, strides, out_shape, (0, 0), 'MAX')
 expect(node, inputs=[x], outputs=[y], name='test_maxpool_2d_default')
 ```
 
+</details>
+<details>
+<summary>maxpool_2d_dilations</summary>
+
+```python
+"""
+input_shape: [1, 1, 4, 4]
+output_shape: [1, 1, 2, 2]
+"""
+node = onnx.helper.make_node(
+    'MaxPool',
+    inputs=['x'],
+    outputs=['y'],
+    kernel_shape=[2, 2],
+    strides=[1, 1],
+    dilations=[2, 2]
+)
+x = np.array([[[
+    [1, 2, 3, 4],
+    [5, 6, 7, 8],
+    [9, 10, 11, 12],
+    [13, 14, 15, 16],
+]]]).astype(np.float32)
+y = np.array([[[
+    [11, 12],
+    [15, 16]]]]).astype(np.float32)
+
+expect(node, inputs=[x], outputs=[y], name='test_maxpool_2d_dilations')
+```
+
 </details>
 <details>
 <summary>maxpool_2d_pads</summary>
@@ -7128,10 +7158,11 @@ bias: 1
 size: 1
 </details>
 <details>
-<summary>MaxPool: 3 out of 6 attributes covered</summary>
+<summary>MaxPool: 3 out of 7 attributes covered</summary>
 
 auto_pad: 0
 ceil_mode: 0
+dilations: 0
 kernel_shape: 1
 pads: 2
 storage_order: 0
@@ -7201,10 +7232,11 @@ bias: 1
 size: 1
 </details>
 <details>
-<summary>MaxPool: 3 out of 6 attributes covered</summary>
+<summary>MaxPool: 3 out of 7 attributes covered</summary>
 
 auto_pad: 0
 ceil_mode: 0
+dilations: 0
 kernel_shape: 1
 pads: 3
 storage_order: 0
@@ -7279,10 +7311,11 @@ bias: 1
 size: 1
 </details>
 <details>
-<summary>MaxPool: 3 out of 6 attributes covered</summary>
+<summary>MaxPool: 3 out of 7 attributes covered</summary>
 
 auto_pad: 0
 ceil_mode: 0
+dilations: 0
 kernel_shape: 1
 pads: 3
 storage_order: 0
@@ -7357,10 +7390,11 @@ bias: 1
 size: 1
 </details>
 <details>
-<summary>MaxPool: 3 out of 6 attributes covered</summary>
+<summary>MaxPool: 3 out of 7 attributes covered</summary>
 
 auto_pad: 0
 ceil_mode: 0
+dilations: 0
 kernel_shape: 1
 pads: 3
 storage_order: 0
@@ -7435,10 +7469,11 @@ bias: 1
 size: 1
 </details>
 <details>
-<summary>MaxPool: 3 out of 6 attributes covered</summary>
+<summary>MaxPool: 3 out of 7 attributes covered</summary>
 
 auto_pad: 0
 ceil_mode: 0
+dilations: 0
 kernel_shape: 1
 pads: 3
 storage_order: 0
@@ -7513,10 +7548,11 @@ bias: 1
 size: 1
 </details>
 <details>
-<summary>MaxPool: 3 out of 6 attributes covered</summary>
+<summary>MaxPool: 3 out of 7 attributes covered</summary>
 
 auto_pad: 0
 ceil_mode: 0
+dilations: 0
 kernel_shape: 1
 pads: 3
 storage_order: 0
@@ -7596,10 +7632,11 @@ bias: 1
 size: 1
 </details>
 <details>
-<summary>MaxPool: 3 out of 6 attributes covered</summary>
+<summary>MaxPool: 3 out of 7 attributes covered</summary>
 
 auto_pad: 0
 ceil_mode: 0
+dilations: 0
 kernel_shape: 1
 pads: 3
 storage_order: 0
@@ -7679,10 +7716,11 @@ bias: 1
 size: 1
 </details>
 <details>
-<summary>MaxPool: 3 out of 6 attributes covered</summary>
+<summary>MaxPool: 3 out of 7 attributes covered</summary>
 
 auto_pad: 0
 ceil_mode: 0
+dilations: 0
 kernel_shape: 2
 pads: 3
 storage_order: 0
@@ -7762,10 +7800,11 @@ bias: 2
 size: 1
 </details>
 <details>
-<summary>MaxPool: 3 out of 6 attributes covered</summary>
+<summary>MaxPool: 3 out of 7 attributes covered</summary>
 
 auto_pad: 0
 ceil_mode: 0
+dilations: 0
 kernel_shape: 2
 pads: 3
 storage_order: 0

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

@@ -362,3 +362,29 @@ def export_maxpool_2d_ceil():  # type: () -> None
             [15, 16]]]]).astype(np.float32)
 
         expect(node, inputs=[x], outputs=[y], name='export_maxpool_2d_ceil')
+
+    @staticmethod
+    def export_maxpool_2d_dilations():  # type: () -> None
+        """
+        input_shape: [1, 1, 4, 4]
+        output_shape: [1, 1, 2, 2]
+        """
+        node = onnx.helper.make_node(
+            'MaxPool',
+            inputs=['x'],
+            outputs=['y'],
+            kernel_shape=[2, 2],
+            strides=[1, 1],
+            dilations=[2, 2]
+        )
+        x = np.array([[[
+            [1, 2, 3, 4],
+            [5, 6, 7, 8],
+            [9, 10, 11, 12],
+            [13, 14, 15, 16],
+        ]]]).astype(np.float32)
+        y = np.array([[[
+            [11, 12],
+            [15, 16]]]]).astype(np.float32)
+
+        expect(node, inputs=[x], outputs=[y], name='test_maxpool_2d_dilations')

onnx/backend/test/data/node/test_maxpool_2d_dilations/model.onnx

@@ -0,0 +1,18 @@
+backend-test:�

+L
+
x
y"MaxPool*
+	dilations@@�
*
+kernel_shape@@�
*
+strides@
@
�
test_maxpool_2d_dilationsZ
+
x
+

+

+

+
+b
+
y
+

+

+

+
+B

onnx/backend/test/data/node/test_maxpool_2d_dilations/test_data_set_0/input_0.pb

@@ -250,7 +250,8 @@ std::function<void(OpSchema&)> PoolOpSchemaGenerator_9(
 std::function<void(OpSchema&)> PoolOpSchemaGenerator(
     const char* name,
     const char* opName,
-    const char* additionalDescription) {
+    const char* additionalDescription,
+    bool use_dilation) {
   return [=](OpSchema& schema) {
     std::string doc = R"DOC(
  {name} consumes an input tensor X and applies {opName} pooling across
@@ -330,8 +331,8 @@ std::function<void(OpSchema&)> PoolOpSchemaGenerator(
         "T",
         {"tensor(float16)", "tensor(float)", "tensor(double)"},
         "Constrain input and output types to float tensors.");
-    schema.TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
-      convPoolTypeAndShapeInference(ctx, false, true);
+    schema.TypeAndShapeInferenceFunction([use_dilation](InferenceContext& ctx) {
+      convPoolTypeAndShapeInference(ctx, use_dilation, true);
     });
   };
 } // namespace ONNX_NAMESPACE
@@ -365,7 +366,8 @@ ONNX_OPERATOR_SET_SCHEMA(
         .FillUsing(PoolOpSchemaGenerator(
             "AveragePool",
             "average",
-            "The output of each pooling window is divided by the number of elements (exclude pad when attribute count_include_pad is zero)."))
+            "The output of each pooling window is divided by the number of elements (exclude pad when attribute count_include_pad is zero).",
+            false))
         .Attr(
             "count_include_pad",
             "Whether include pad pixels when calculating values for the edges. Default is 0, doesn't count include pad.",
@@ -416,12 +418,18 @@ ONNX_OPERATOR_SET_SCHEMA(
         .FillUsing(PoolOpSchemaGenerator(
             "MaxPool",
             "max",
-            "The output of each pooling window is maximum number of elements exclude pad."))
+            "The output of each pooling window is maximum number of elements exclude pad.",
+            true))
         .Attr(
             "storage_order",
             "The storage order of the tensor. 0 is row major, and 1 is column major.",
             AttributeProto::INT,
             static_cast<int64_t>(0))
+        .Attr(
+            "dilations",
+            "Dilation value along each axis of filter.",
+            AttributeProto::INTS,
+            OPTIONAL)
         .Output(
             1,
             "Indices",

onnx/backend/test/data/node/test_maxpool_2d_dilations/test_data_set_0/output_0.pb

@@ -958,6 +958,13 @@ def test_maxpool_ceil(self):  # type: () -> None
             [])
         self._assert_inferred(graph, [make_tensor_value_info("Y", TensorProto.FLOAT, (1, 1, 2, 2))])
 
+    def test_maxpool_with_dilations(self):  # type: () -> None
+        graph = self._make_graph(
+            [("X", TensorProto.FLOAT, (5, 3, 4, 4))],
+            [make_node("MaxPool", ["X"], ["Y"], kernel_shape=[2, 2], dilations=[2, 2])],
+            [])
+        self._assert_inferred(graph, [make_tensor_value_info("Y", TensorProto.FLOAT, (5, 3, 2, 2))])
+
     def test_averagepool(self):  # type: () -> None
         graph = self._make_graph(
             [("X", TensorProto.FLOAT, (5, 3, 4, 4))],