Update a script to generate text graphs for Faster-RCNN networks from TensorFlow

modules/dnn/src/dnn.cpp

@@ -1794,44 +1794,46 @@ struct Net::Impl
                 }
 
                 // fuse convolution layer followed by eltwise + relu
-                if ( IS_DNN_OPENCL_TARGET(preferableTarget) )
+                if ( IS_DNN_OPENCL_TARGET(preferableTarget) && ld.layerInstance->type == "Convolution" )
                 {
                     Ptr<EltwiseLayer> nextEltwiseLayer;
                     if( nextData )
                         nextEltwiseLayer = nextData->layerInstance.dynamicCast<EltwiseLayer>();
 
-                    if( !nextEltwiseLayer.empty() && pinsToKeep.count(lpNext) == 0 )
+                    if( !nextEltwiseLayer.empty() && pinsToKeep.count(lpNext) == 0 &&
+                        nextData->inputBlobsId.size() == 2 )
                     {
                         LayerData *eltwiseData = nextData;
-                        // go down from the second input and find the first non-skipped layer.
-                        LayerData *downLayerData = &layers[eltwiseData->inputBlobsId[1].lid];
-                        CV_Assert(downLayerData);
-                        while (downLayerData->skip)
-                        {
-                            downLayerData = &layers[downLayerData->inputBlobsId[0].lid];
-                        }
-                        CV_Assert(downLayerData);
 
-                        // second input layer is current layer.
+                        // Eltwise layer has two inputs. We need to determine which
-                        if ( ld.id == downLayerData->id )
+                        // is a base convolution layer and which could be used as it's bias.
+                        LayerData* biasLayerData = 0;
+                        for (int i = 0; i < 2; ++i)
                         {
-                            // go down from the first input and find the first non-skipped layer
+                            LayerData *downLayerData = &layers[eltwiseData->inputBlobsId[i].lid];
-                            downLayerData = &layers[eltwiseData->inputBlobsId[0].lid];
+                            CV_Assert(downLayerData);
                             while (downLayerData->skip)
                             {
-                                if ( !downLayerData->type.compare("Eltwise") )
+                                if (downLayerData->inputBlobsId.size() == 1)
-                                    downLayerData = &layers[downLayerData->inputBlobsId[1].lid];
-                                else
                                     downLayerData = &layers[downLayerData->inputBlobsId[0].lid];
+                                else
+                                {
+                                    downLayerData = 0;
+                                    break;
+                                }
                             }
-
+                            if (downLayerData && ld.id == downLayerData->id)
-                            Ptr<ConvolutionLayer> convLayer = downLayerData->layerInstance.dynamicCast<ConvolutionLayer>();
+                            {
-
+                                biasLayerData = &layers[eltwiseData->inputBlobsId[1 - i].lid];
-                            //  first input layer is convolution layer
+                                break;
-                            if( !convLayer.empty() && eltwiseData->consumers.size() == 1 )
+                            }
+                        }
+                        CV_Assert(biasLayerData);
+                        {
+                            if( eltwiseData->consumers.size() == 1 )
                             {
                                 // fuse eltwise + activation layer
-                                LayerData *firstConvLayerData = downLayerData;
+                                if (biasLayerData->id < ld.id)
                                 {
                                     nextData = &layers[eltwiseData->consumers[0].lid];
                                     lpNext = LayerPin(eltwiseData->consumers[0].lid, 0);
@@ -1845,8 +1847,8 @@ struct Net::Impl
                                              !nextData->type.compare("Power")) &&
                                             currLayer->setActivation(nextActivLayer) )
                                     {
-                                        CV_Assert(firstConvLayerData->outputBlobsWrappers.size() == 1 && ld.inputBlobsWrappers.size() == 1);
+                                        CV_Assert_N(biasLayerData->outputBlobsWrappers.size() == 1, ld.inputBlobsWrappers.size() == 1);
-                                        ld.inputBlobsWrappers.push_back(firstConvLayerData->outputBlobsWrappers[0]);
+                                        ld.inputBlobsWrappers.push_back(biasLayerData->outputBlobsWrappers[0]);
                                         printf_(("\tfused with %s\n", nextEltwiseLayer->name.c_str()));
                                         printf_(("\tfused with %s\n", nextActivLayer->name.c_str()));
                                         eltwiseData->skip = true;
@@ -1897,9 +1899,6 @@ struct Net::Impl
                 }
             }
 
-            if (preferableBackend != DNN_BACKEND_OPENCV)
-                continue;  // Go to the next layer.
-
             // the optimization #2. if there is no layer that takes max pooling layer's computed
             // max indices (and only some semantical segmentation networks might need this;
             // many others only take the maximum values), then we switch the max pooling

modules/dnn/src/layers/pooling_layer.cpp

@@ -95,7 +95,6 @@ public:
         else if (params.has("pooled_w") || params.has("pooled_h"))
         {
             type = ROI;
-            computeMaxIdx = false;
             pooledSize.width = params.get<uint32_t>("pooled_w", 1);
             pooledSize.height = params.get<uint32_t>("pooled_h", 1);
         }
@@ -141,6 +140,7 @@ public:
 #ifdef HAVE_OPENCL
         poolOp.release();
 #endif
+        computeMaxIdx = type == MAX;
     }
 
     virtual bool supportBackend(int backendId) CV_OVERRIDE
@@ -190,19 +190,14 @@ public:
             poolOp = Ptr<OCL4DNNPool<float> >(new OCL4DNNPool<float>(config));
         }
 
-        for (size_t ii = 0; ii < inputs.size(); ii++)
+        CV_Assert_N(inputs.size() == 1, !outputs.empty(), !computeMaxIdx || outputs.size() == 2);
-        {
+        UMat& inpMat = inputs[0];
-            UMat& inpMat = inputs[ii];
+        UMat& outMat = outputs[0];
-            int out_index = (type == MAX) ? 2 : 1;
+        UMat maskMat = computeMaxIdx ? outputs[1] : UMat();
-            UMat& outMat = outputs[out_index * ii];
-            UMat maskMat = (type == MAX) ? outputs[2 * ii + 1] : UMat();
 
-            CV_Assert(inpMat.offset == 0 && outMat.offset == 0);
+        CV_Assert(inpMat.offset == 0 && outMat.offset == 0);
 
-            if (!poolOp->Forward(inpMat, outMat, maskMat))
+        return poolOp->Forward(inpMat, outMat, maskMat);
-                return false;
-        }
-        return true;
     }
 #endif
 
@@ -229,9 +224,12 @@ public:
         switch (type)
         {
             case MAX:
-                CV_Assert_N(inputs.size() == 1, outputs.size() == 2);
+            {
-                maxPooling(inputs[0], outputs[0], outputs[1]);
+                CV_Assert_N(inputs.size() == 1, !computeMaxIdx || outputs.size() == 2);
+                Mat mask = computeMaxIdx ? outputs[1] : Mat();
+                maxPooling(inputs[0], outputs[0], mask);
                 break;
+            }
             case AVE:
                 CV_Assert_N(inputs.size() == 1, outputs.size() == 1);
                 avePooling(inputs[0], outputs[0]);
@@ -912,7 +910,10 @@ public:
             dims[0] = inputs[1][0];  // Number of proposals;
             dims[1] = psRoiOutChannels;
         }
-        outputs.assign(type == MAX ? 2 : 1, shape(dims, 4));
+
+        int numOutputs = requiredOutputs ? requiredOutputs : (type == MAX ? 2 : 1);
+        CV_Assert(numOutputs == 1 || (numOutputs == 2 && type == MAX));
+        outputs.assign(numOutputs, shape(dims, 4));
 
         return false;
     }

modules/dnn/test/test_tf_importer.cpp

@@ -358,7 +358,7 @@ TEST_P(Test_TensorFlow_nets, Faster_RCNN)
         (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16))
         throw SkipTestException("");
 
-    for (int i = 1; i < 2; ++i)
+    for (int i = 0; i < 2; ++i)
     {
         std::string proto = findDataFile("dnn/" + names[i] + ".pbtxt", false);
         std::string model = findDataFile("dnn/" + names[i] + ".pb", false);

samples/dnn/tf_text_graph_faster_rcnn.py

@@ -32,6 +32,8 @@ def createFasterRCNNGraph(modelPath, configPath, outputPath):
     width_stride = float(grid_anchor_generator['width_stride'][0])
     height_stride = float(grid_anchor_generator['height_stride'][0])
     features_stride = float(config['feature_extractor'][0]['first_stage_features_stride'][0])
+    first_stage_nms_iou_threshold = float(config['first_stage_nms_iou_threshold'][0])
+    first_stage_max_proposals = int(config['first_stage_max_proposals'][0])
 
     print('Number of classes: %d' % num_classes)
     print('Scales:            %s' % str(scales))
@@ -47,7 +49,8 @@ def createFasterRCNNGraph(modelPath, configPath, outputPath):
     removeIdentity(graph_def)
 
     def to_remove(name, op):
-        return name.startswith(scopesToIgnore) or not name.startswith(scopesToKeep)
+        return name.startswith(scopesToIgnore) or not name.startswith(scopesToKeep) or \
+               (name.startswith('CropAndResize') and op != 'CropAndResize')
 
     removeUnusedNodesAndAttrs(to_remove, graph_def)
 
@@ -114,10 +117,10 @@ def createFasterRCNNGraph(modelPath, configPath, outputPath):
     detectionOut.addAttr('num_classes', 2)
     detectionOut.addAttr('share_location', True)
     detectionOut.addAttr('background_label_id', 0)
-    detectionOut.addAttr('nms_threshold', 0.7)
+    detectionOut.addAttr('nms_threshold', first_stage_nms_iou_threshold)
     detectionOut.addAttr('top_k', 6000)
     detectionOut.addAttr('code_type', "CENTER_SIZE")
-    detectionOut.addAttr('keep_top_k', 100)
+    detectionOut.addAttr('keep_top_k', first_stage_max_proposals)
     detectionOut.addAttr('clip', False)
 
     graph_def.node.extend([detectionOut])
@@ -147,9 +150,11 @@ def createFasterRCNNGraph(modelPath, configPath, outputPath):
               'SecondStageBoxPredictor/Reshape_1/Reshape', [1, -1], graph_def)
 
     # Replace Flatten subgraph onto a single node.
+    cropAndResizeNodeName = ''
     for i in reversed(range(len(graph_def.node))):
         if graph_def.node[i].op == 'CropAndResize':
             graph_def.node[i].input.insert(1, 'detection_out/clip_by_value')
+            cropAndResizeNodeName = graph_def.node[i].name
 
         if graph_def.node[i].name == 'SecondStageBoxPredictor/Reshape':
             addConstNode('SecondStageBoxPredictor/Reshape/shape2', [1, -1, 4], graph_def)
@@ -159,11 +164,15 @@ def createFasterRCNNGraph(modelPath, configPath, outputPath):
 
         if graph_def.node[i].name in ['SecondStageBoxPredictor/Flatten/flatten/Shape',
                                       'SecondStageBoxPredictor/Flatten/flatten/strided_slice',
-                                      'SecondStageBoxPredictor/Flatten/flatten/Reshape/shape']:
+                                      'SecondStageBoxPredictor/Flatten/flatten/Reshape/shape',
+                                      'SecondStageBoxPredictor/Flatten_1/flatten/Shape',
+                                      'SecondStageBoxPredictor/Flatten_1/flatten/strided_slice',
+                                      'SecondStageBoxPredictor/Flatten_1/flatten/Reshape/shape']:
             del graph_def.node[i]
 
     for node in graph_def.node:
-        if node.name == 'SecondStageBoxPredictor/Flatten/flatten/Reshape':
+        if node.name == 'SecondStageBoxPredictor/Flatten/flatten/Reshape' or \
+           node.name == 'SecondStageBoxPredictor/Flatten_1/flatten/Reshape':
             node.op = 'Flatten'
             node.input.pop()
 
@@ -171,6 +180,11 @@ def createFasterRCNNGraph(modelPath, configPath, outputPath):
                          'SecondStageBoxPredictor/BoxEncodingPredictor/MatMul']:
             node.addAttr('loc_pred_transposed', True)
 
+        if node.name.startswith('MaxPool2D'):
+            assert(node.op == 'MaxPool')
+            assert(cropAndResizeNodeName)
+            node.input = [cropAndResizeNodeName]
+
     ################################################################################
     ### Postprocessing
     ################################################################################