Fix bug in import of Conv op in ONNX.

Source/CNTKv2LibraryDll/proto/onnx/ONNXToCNTK.cpp

@@ -3176,7 +3176,7 @@ FunctionPtr ONNXToCNTKHelper::CreateCNTKConvTransposeNode(const Node *node, cons
 FunctionPtr ONNXToCNTKHelper::CreateCNTKConvNode(const Node *node, const std::vector<Variable> &inputs)
 {
     Variable convolutionMap = inputs[1];
-    size_t numSpatialDim = convolutionMap.Shape().Rank() - 1; // This is conv op dimension, i.e. 2 for 2D conv, 3 for 3D conv.
+    size_t numSpatialDim = convolutionMap.Shape().Rank() - 2; // This is conv op dimension, i.e. 2 for 2D conv, 3 for 3D conv.
 
     NDShape strides = GetNamedAttributeAsShape(node, "strides", false, NDShape(std::vector<size_t>(numSpatialDim, 1u)));
     NDShape dilation = GetNamedAttributeAsShape(node, "dilations", false, NDShape(std::vector<size_t>(numSpatialDim, 1u)));
@@ -3189,6 +3189,14 @@ FunctionPtr ONNXToCNTKHelper::CreateCNTKConvNode(const Node *node, const std::ve
     std::vector<bool> cntkConvAutoPadding;
     auto convOperand = GetNodeOperandWithPaddingResolved(/*output arg first*/ cntkConvAutoPadding, strides, node, inputs[0]);
 
+    // At this point length of vectors strides, dilation, and cntkConvAutoPadding must be equal to
+    // number of spatial dimensions (2 for 2D conv, 3 for 3D conv). 
+    // In order to match the expected input for CNTK Convolution API we will append one more element
+    // in each for the "channel" axis. 
+    strides = strides.AppendShape({ 1 });
+    dilation = dilation.AppendShape({ 1 });
+    cntkConvAutoPadding.push_back(false);
+
     auto operandPlaceholder = PlaceholderVariable(convOperand.Shape(), L"operand", {});
     auto convmapPlaceholder = PlaceholderVariable(convolutionMap.Shape(), L"convolutionMap", {});
     FunctionPtr operandWithBatchAxis = ToBatch(operandPlaceholder);

bindings/python/cntk/tests/onnx_op_test.py

@@ -425,6 +425,24 @@ def test_Concat(tmpdir, dtype):
 
         verify_one_input(model, data1, tmpdir, 'Concat_1')
 
+@pytest.mark.parametrize("dtype", DType_Config)
+def test_Conv(tmpdir, dtype, device_id):
+    if device_id == -1 and dtype == np.float16:
+        pytest.skip('Test is skipped on CPU with float16 data')
+    device = cntk_device(device_id)
+    with C.default_options(dtype=dtype):
+        input_shape = (3, 20, 32) 
+        img = np.reshape(np.arange(np.prod(input_shape), dtype = dtype), input_shape) 
+
+        x = C.input_variable(input_shape)
+
+        kernel_shape = (64, 3, 3, 3) # For convolution the shape is (O x I x W x H)
+        kernel = C.constant(value = np.ones(shape=(kernel_shape), dtype = dtype))
+
+        conv_model = C.convolution(kernel, x, auto_padding = [False, True, True])
+
+        verify_one_input(conv_model, img, tmpdir, 'Conv_0', device)
+        
 @pytest.mark.parametrize("dtype", DType_Config)
 def test_ConvTranspose(tmpdir, dtype, device_id):
     if device_id == -1 and dtype == np.float16: