Merge pull request #22199 from zihaomu:bug_fix_22195

DNN: Reduce Layer (add dynamic batch and ReduceSum support)

modules/dnn/include/opencv2/dnn/all_layers.hpp

@@ -334,7 +334,8 @@ CV__DNN_INLINE_NS_BEGIN
     {
     public:
         int reduceType;
-        std::vector<size_t> reduceDims;
+        // reduceDims contains the dimensions that need to be reduced, targetDims is the target output dimension.
+        std::vector<size_t> reduceDims, targetDims;
         static Ptr<ReduceLayer> create(const LayerParams& params);
     };
 

modules/dnn/src/int8layers/reduce_layer.cpp

@@ -38,6 +38,15 @@ public:
         {
             reduceDims[i] = tempDims.get<int>(i);
         }
+
+        CV_Assert(params.has("target_dims"));
+        tempDims = params.get("target_dims");
+        n = tempDims.size();
+        targetDims.resize(n);
+        for (i = 0; i < n; i++)
+        {
+            targetDims[i] = tempDims.get<int>(i);
+        }
     }
 
     virtual bool supportBackend(int backendId) CV_OVERRIDE
@@ -161,18 +170,30 @@ public:
                          std::vector<MatShape> &internals) const CV_OVERRIDE
     {
         CV_Assert(inputs.size() > 0);
-        CV_Assert(reduceDims.size() != 0 && inputs[0].size() >= reduceDims.size());
+        CV_Assert( reduceDims.size() !=0 && targetDims.size() != 0 && inputs[0].size() >= reduceDims.size());
 
-        std::vector<int> outShape;
+        // outShapeTmp can save the right number of `total(outShapeTmp)`. And the outShape is used as the final output shape.
+        std::vector<int> outShapeTmp, outShape;
+        outShape.assign(targetDims.begin(), targetDims.end());
         if (inputs[0].size() == reduceDims.size())
-            outShape.push_back(1);
+            outShapeTmp.push_back(1);
         else
         {
             for (int i = 0; i < inputs[0].size() - reduceDims.size(); i++)
             {
-                outShape.push_back(inputs[0][i]);
+                outShapeTmp.push_back(inputs[0][i]);
             }
         }
+
+        // Support dynamic shape of Batch size.
+        // Note that: when there are multiple dynamic inputs, we will give an error.
+        if (total(outShape) != total(outShapeTmp))
+        {
+            if (outShape[0] != outShapeTmp[0])
+                outShape[0] = outShapeTmp[0];
+        }
+
+        CV_Assert(total(outShape) == total(outShapeTmp));
         outputs.assign(1, outShape);
 
         return false;

modules/dnn/src/layers/reduce_layer.cpp

@@ -61,6 +61,15 @@ public:
         {
             reduceDims[i] = tempDims.get<int>(i);
         }
+
+        CV_Assert(params.has("target_dims"));
+        tempDims = params.get("target_dims");
+        n = tempDims.size();
+        targetDims.resize(n);
+        for (i = 0; i < n; i++)
+        {
+            targetDims[i] = tempDims.get<int>(i);
+        }
     }
 
     virtual bool supportBackend(int backendId) CV_OVERRIDE
@@ -325,18 +334,29 @@ public:
                          std::vector<MatShape> &internals) const CV_OVERRIDE
     {
         CV_Assert(inputs.size() > 0);
-        CV_Assert(reduceDims.size() != 0 && inputs[0].size() >= reduceDims.size());
+        CV_Assert( reduceDims.size() !=0 && targetDims.size() != 0 && inputs[0].size() >= reduceDims.size());
 
-        std::vector<int> outShape;
+        // outShapeTmp can save the right number of `total(outShapeTmp)`. And the outShape is used as the final output shape.
+        std::vector<int> outShapeTmp, outShape;
+        outShape.assign(targetDims.begin(), targetDims.end());
         if (inputs[0].size() == reduceDims.size())
-            outShape.push_back(1);
+            outShapeTmp.push_back(1);
         else
         {
             for (int i = 0; i < inputs[0].size() - reduceDims.size(); i++)
             {
-                outShape.push_back(inputs[0][i]);
+                outShapeTmp.push_back(inputs[0][i]);
             }
         }
+
+        // Support dynamic shape of Batch size.
+        // Note that: when there are multiple dynamic inputs, we will give an error.
+        if (total(outShape) != total(outShapeTmp) && outShape[0] != outShapeTmp[0])
+        {
+                outShape[0] = outShapeTmp[0];
+        }
+
+        CV_Assert(total(outShape) == total(outShapeTmp));
         outputs.assign(1, outShape);
 
         return false;

modules/dnn/src/onnx/onnx_importer.cpp

@@ -1203,36 +1203,47 @@ void ONNXImporter::parseReduce(LayerParams& layerParams, const opencv_onnx::Node
     layerParams.set("reduce", reduceType);
     bool keepdims = layerParams.get<int>("keepdims", 1) == 1;
 
-    if (layer_type == "ReduceSum" && node_proto.input_size() == 2)
-    {
-        // TODO support the opset 13 of ReduceSum.
-        //  in opset 13, the ReduceSum has two input, it takes axes as input instead of attribute
-        //  details:https://github.com/onnx/onnx/issues/3420#issuecomment-844295687
-        CV_Error(Error::StsNotImplemented, "Unsupported " + layer_type + " operation of opset 13, please try to "
-                                                                         "re-export the onnx model with opset 11.");
-    }
-
     MatShape inpShape = outShapes[node_proto.input(0)];
     std::vector<bool> shouldDelete(inpShape.size(), false);
 
-    if (layerParams.has("axes"))
+    if (layer_type == "ReduceSum" && node_proto.input_size() == 2)
     {
-        DictValue axes = layerParams.get("axes");
+        if (constBlobs.find(node_proto.input(1)) != constBlobs.end())
-        for (int i = 0; i < axes.size(); i++)
         {
-            int axis = normalize_axis(axes.get<int>(i), inpShape.size());
+            Mat axesMat = getBlob(node_proto, 1);
-            shouldDelete[axis] = true;
+            int axesNum = axesMat.total();
+            for (int i = 0; i < axesNum; i++)
+            {
+                int axis = normalize_axis(axesMat.at<int>(i), inpShape.size());
+                shouldDelete[axis] = true;
+            }
         }
+        else
+            //  in opset 13, the ReduceSum has two input, it takes axes as input instead of attribute
+            //  details:https://github.com/onnx/onnx/issues/3420#issuecomment-844295687
+            CV_Error(Error::StsNotImplemented, "Non-constant axis values in ReduceSum are not supported.");
     }
     else
     {
-        for (int i = 0; i < inpShape.size(); i++)
+        if (layerParams.has("axes"))
         {
-            shouldDelete[i] = true;
+            DictValue axes = layerParams.get("axes");
+            for (int i = 0; i < axes.size(); i++)
+            {
+                int axis = normalize_axis(axes.get<int>(i), inpShape.size());
+                shouldDelete[axis] = true;
+            }
+        }
+        else
+        {
+            for (int i = 0; i < inpShape.size(); i++)
+            {
+                shouldDelete[i] = true;
+            }
         }
     }
 
-    MatShape targetShape;
+    std::vector<int> targetShape;
     for (int i = 0; i < inpShape.size(); ++i)
     {
         if (!shouldDelete[i])
@@ -1302,19 +1313,10 @@ void ONNXImporter::parseReduce(LayerParams& layerParams, const opencv_onnx::Node
         }
     }
 
-    LayerParams reduceLp = layerParams;
+    layerParams.set("deleted_dims", DictValue::arrayInt(&deletedDims[0], deletedDims.size()));
-    reduceLp.name = layerParams.name + "/reduce";
+    layerParams.set("target_dims", DictValue::arrayInt(&targetShape[0], targetShape.size()));
-    CV_Assert(layer_id.find(reduceLp.name) == layer_id.end());
-    reduceLp.set("deleted_dims", DictValue::arrayInt(&deletedDims[0], deletedDims.size()));
 
     node_proto.set_input(0, inputString);
-    node_proto.set_output(0, reduceLp.name);
-    addLayer(reduceLp, node_proto);
-
-    layerParams.type = (depth == CV_8S) ? "ReshapeInt8" : "Reshape";
-    layerParams.set("dim", DictValue::arrayInt(&targetShape[0], targetShape.size()));
-
-    node_proto.set_input(0, node_proto.output(0));
     node_proto.set_output(0, output_name);
 
     addLayer(layerParams, node_proto);

modules/dnn/test/test_onnx_importer.cpp

@@ -411,6 +411,7 @@ TEST_P(Test_ONNX_layers, ReduceMean)
 TEST_P(Test_ONNX_layers, ReduceSum)
 {
     testONNXModels("reduce_sum");
+    testONNXModels("reduce_sum_axis_dynamic_batch");
 }
 
 TEST_P(Test_ONNX_layers, ReduceMax)