[SPEC] Fix output_shape input in (Group)ConvolutionBackpropData ops (#…

…4005)

src/ngraph/op/convolution.cpp

@@ -205,13 +205,28 @@ op::v1::ConvolutionBackpropData::ConvolutionBackpropData(const Output<Node>& dat
 
 const PartialShape op::v1::ConvolutionBackpropData::get_output_shape() const
 {
-    PartialShape shape{PartialShape::dynamic()};
+    PartialShape shape{vector<Dimension>(m_strides.size() + 2)};
+    auto data_pshape = get_input_partial_shape(0);
+    if (data_pshape.rank().is_static())
+    {
+        shape[0] = data_pshape[0]; // N
+    }
+    auto filters_pshape = get_input_partial_shape(1);
+    if (filters_pshape.rank().is_static())
+    {
+        shape[1] = filters_pshape[1]; // C
+    }
     bool is_output_shape_present = get_inputs().size() == 3;
     if (is_output_shape_present)
     {
         if (auto const_op = as_type<op::Constant>(input_value(2).get_node()))
         {
-            shape = const_op->get_shape_val();
+            auto output_shape = const_op->get_shape_val();
+            // Populate spatials
+            for (int i = 0; i < output_shape.size(); ++i)
+            {
+                shape[i + 2] = output_shape[i];
+            }
         }
     }
     return shape;
@@ -270,13 +285,6 @@ void op::v1::ConvolutionBackpropData::validate_and_infer_types()
     if (is_output_shape_present)
     {
         set_input_is_relevant_to_shape(2);
-        if (output_pshape.is_static() && data_pshape.is_static())
-        {
-            auto data_shape = data_pshape.to_shape();
-            auto output_shape = output_pshape.to_shape();
-            output_shape.insert(output_shape.begin(), data_shape.begin(), data_shape.begin() + 1);
-            output_pshape = output_shape;
-        }
     }
     else
     {
@@ -295,12 +303,13 @@ void op::v1::ConvolutionBackpropData::validate_and_infer_types()
             for (size_t i = 0; i < data_spatial_rank; ++i)
             {
                 size_t tmp = m_strides[i] * (data_shape[i + 2] - 1) +
-                             ((filters_shape[i] + 2 - 1) * m_dilations[i] + 1) - m_pads_begin[i] -
+                             ((filters_shape[i + 2] - 1) * m_dilations[i] + 1) - m_pads_begin[i] -
                              m_pads_end[i] + output_padding[i];
                 output_shape.push_back(tmp);
-                output_pshape = output_shape;
             }
-            output_shape.insert(output_shape.begin(), data_shape.begin(), data_shape.begin() + 1);
+            output_shape.insert(output_shape.begin(), filters_shape.at(1));
+            output_shape.insert(output_shape.begin(), data_shape.at(0));
+            output_pshape = output_shape;
         }
     }
 

src/ngraph/op/fused/group_conv.cpp

@@ -192,13 +192,28 @@ op::v1::GroupConvolutionBackpropData::GroupConvolutionBackpropData(
 
 const PartialShape op::v1::GroupConvolutionBackpropData::get_output_shape() const
 {
-    PartialShape shape{PartialShape::dynamic()};
+    PartialShape shape{vector<Dimension>(m_strides.size() + 2)};
+    auto data_pshape = get_input_partial_shape(0);
+    if (data_pshape.rank().is_static())
+    {
+        shape[0] = data_pshape[0]; // N
+    }
+    auto filters_pshape = get_input_partial_shape(1);
+    if (filters_pshape.rank().is_static())
+    {
+        shape[1] = filters_pshape[1]; // C
+    }
     bool is_output_shape_present = get_inputs().size() == 3;
     if (is_output_shape_present)
     {
         if (auto const_op = as_type<op::Constant>(input_value(2).get_node()))
         {
-            shape = const_op->get_shape_val();
+            auto output_shape = const_op->get_shape_val();
+            // Populate spatials
+            for (int i = 0; i < output_shape.size(); ++i)
+            {
+                shape[i + 2] = output_shape[i];
+            }
         }
     }
     return shape;
@@ -257,40 +272,32 @@ void op::v1::GroupConvolutionBackpropData::validate_and_infer_types()
     if (is_output_shape_present)
     {
         set_input_is_relevant_to_shape(2);
-        if (output_pshape.is_static() && data_pshape.is_static())
-        {
-            auto data_shape = data_pshape.to_shape();
-            auto output_shape = output_pshape.to_shape();
-            output_shape.insert(output_shape.begin(), data_shape.begin(), data_shape.begin() + 1);
-            output_pshape = output_shape;
-        }
     }
     else
     {
         if (filters_pshape.is_static() && data_pshape.is_static())
         {
             auto filters_shape = filters_pshape.to_shape();
-            filters_shape.erase(filters_shape.begin(),
-                                filters_shape.begin() + 3); // remove {G, O, I}
             auto data_shape = data_pshape.to_shape();
-            data_shape.erase(data_shape.begin(), data_shape.begin() + 2); // remove {N, C}
 
             Shape output_shape;
-            auto data_spatial_rank = data_shape.size();
+            auto data_spatial_rank = data_shape.size() - 2;
             auto output_padding = get_output_padding();
             if (output_padding.size() == 0)
             {
                 output_padding.insert(output_padding.begin(), data_spatial_rank, 0);
             }
             for (size_t i = 0; i < data_spatial_rank; ++i)
             {
-                size_t tmp = m_strides[i] * (data_shape[i] - 1) +
-                             ((filters_shape[i] - 1) * m_dilations[i] + 1) - m_pads_begin[i] -
+                size_t tmp = m_strides[i] * (data_shape[i + 2] - 1) +
+                             ((filters_shape[i + 3] - 1) * m_dilations[i] + 1) - m_pads_begin[i] -
                              m_pads_end[i] + output_padding[i];
                 output_shape.push_back(tmp);
-                output_pshape = output_shape;
             }
-            output_shape.insert(output_shape.begin(), data_shape.begin(), data_shape.begin() + 1);
+            output_shape.insert(output_shape.begin(),
+                                filters_shape.at(0) * filters_shape.at(2)); // GROUP * C_OUTPUT
+            output_shape.insert(output_shape.begin(), data_shape.at(0));
+            output_pshape = output_shape;
         }
     }
 

src/ngraph/pass/opset0_downgrade.cpp

@@ -135,11 +135,12 @@ namespace
 
     bool op_cast(shared_ptr<op::v1::ConvolutionBackpropData> node)
     {
-        auto output_shape = as_type_ptr<op::Constant>(node->input_value(2).get_node_shared_ptr());
+        auto output_shape_node =
+            as_type_ptr<op::Constant>(node->input_value(2).get_node_shared_ptr());
         const auto data_arg = node->input(0).get_source_output();
         const auto filters_arg = node->input(1).get_source_output();
         const auto strides = node->get_strides();
-        NGRAPH_CHECK(output_shape,
+        NGRAPH_CHECK(output_shape_node,
                      "Unable to convert ConvolutionBackpropData:v1 to ConvolutionBackpropData:v0 "
                      "if output_shape is not constant. Node: ",
                      *node);
@@ -155,8 +156,22 @@ namespace
                      "with output padding other than `0`. Node: ",
                      *node);
 
+        auto data_pshape = data_arg.get_partial_shape();
+        auto filters_pshape = filters_arg.get_partial_shape();
+
+        NGRAPH_CHECK(data_pshape.rank().is_static() && data_pshape[0].is_static() &&
+                         filters_pshape.rank().is_static() && filters_pshape[1].is_static(),
+                     "Unable to convert ConvolutionBackpropData:v1 to ConvolutionBackpropData:v0 "
+                     "if data shape N and filters shape C dimensions are not static. Node: ",
+                     *node);
+
+        // Add N and C dimenstions to output_shape
+        auto output_shape = output_shape_node->get_shape_val();
+        output_shape.insert(output_shape.begin(), static_cast<size_t>(filters_pshape[1]));
+        output_shape.insert(output_shape.begin(), static_cast<size_t>(data_pshape[0]));
+
         auto replacement_node =
-            make_shared<op::v0::ConvolutionBackpropData>(output_shape->get_shape_val(),
+            make_shared<op::v0::ConvolutionBackpropData>(output_shape,
                                                          filters_arg,
                                                          data_arg,
                                                          node->get_strides(),

src/ngraph/pass/opset1_upgrade.cpp

@@ -179,7 +179,10 @@ namespace
         auto replacement_node = make_shared<op::v1::ConvolutionBackpropData>(
             node->input_value(1), // data
             node->input_value(0), // filters
-            op::Constant::create(element::i64, Shape{data_batch_shape.size()}, data_batch_shape),
+            op::Constant::create(
+                element::i64,
+                Shape{data_batch_shape.size() - 2},
+                vector<size_t>(data_batch_shape.begin() + 2, data_batch_shape.end())),
             strides,
             pads_begin,
             pads_end,

test/backend/convolution.in.cpp

@@ -175,7 +175,7 @@ NGRAPH_TEST(${BACKEND_NAME}, dyn_convolution_backprop_data)
     for (int i = 0; i < 2 * 3 * 5 * 5; i++)
         expected_result.emplace_back(i);
 
-    vector<int64_t> shapes = {2, 3, 5, 5};
+    vector<int64_t> shapes = {5, 5};
 
     // Create some tensors for input/output
     auto a = backend->create_tensor(element::f32, shape_delta);

test/opset_pass/convolution_opset_pass.cpp

@@ -78,7 +78,7 @@ TEST(opset_transform, opset1_convolution_downgrade_pass)
 
 TEST(opset_transform, opset1_convolution_backprop_data_downgrade_pass)
 {
-    auto data_batch_shape = op::Constant::create<int64_t>(element::i64, Shape{3}, {64, 3, 100});
+    auto data_batch_shape = op::Constant::create<int64_t>(element::i64, Shape{1}, {100});
     auto filters = make_shared<op::Parameter>(element::f32, Shape{128, 3, 10});
     auto delta = make_shared<op::Parameter>(element::f32, Shape{64, 128, 96});
     auto strides = Strides{1};