Added possibility of getting any intermediate blob with thrifty memory management

Author: arrybn

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

@@ -337,19 +337,35 @@ namespace dnn //! This namespace is used for dnn module functionlaity.
          * In fact, this layer provides the only way to pass user data into the network.
          * As any other layer, this layer can label its outputs and this function provides an easy way to do this.
          */
-        CV_WRAP void setNetInputs(const std::vector<String> &inputBlobNames);
+        CV_WRAP void setInputsNames(const std::vector<String> &inputBlobNames);
 
-        /** @brief Initializes and allocates all layers. */
-        CV_WRAP void allocate();
-
-        /** @brief Runs forward pass to compute output of layer @p toLayer.
+        /** @brief Runs forward pass to compute output of layer with name @p outputName.
+         *  @param outputName name for layer which output is needed to get
+         *  @return blob for first output of specified layer.
           * @details By default runs forward pass for the whole network.
           */
-        CV_WRAP void forward(LayerId toLayer = String());
-        /** @brief Runs forward pass to compute output of layer @p toLayer, but computations start from @p startLayer */
-        void forward(LayerId startLayer, LayerId toLayer);
-        /** @overload */
-        void forward(const std::vector<LayerId> &startLayers, const std::vector<LayerId> &toLayers);
+        CV_WRAP Mat forward(const String& outputName = String());
+
+        /** @brief Runs forward pass to compute output of layer with name @p outputName.
+         *  @param outputBlobs contains all output blobs for specified layer.
+         *  @param outputName name for layer which output is needed to get
+          * @details If @p outputName is empty, runs forward pass for the whole network.
+          */
+        CV_WRAP void forward(std::vector<Mat>& outputBlobs, const String& outputName = String());
+
+        /** @brief Runs forward pass to compute outputs of layers listed in @p outBlobNames.
+         *  @param outputBlobs contains blobs for first outputs of specified layers.
+         *  @param outBlobNames names for layers which outputs are needed to get
+          */
+        CV_WRAP void forward(std::vector<Mat>& outputBlobs,
+                             const std::vector<String>& outBlobNames);
+
+        /** @brief Runs forward pass to compute outputs of layers listed in @p outBlobNames.
+         *  @param outputBlobs contains all output blobs for each layer specified in @p outBlobNames.
+         *  @param outBlobNames names for layers which outputs are needed to get
+          */
+        CV_WRAP void forward(std::vector<std::vector<Mat> >& outputBlobs,
+                             const std::vector<String>& outBlobNames);
 
         //TODO:
         /** @brief Optimized forward.
@@ -369,7 +385,7 @@ namespace dnn //! This namespace is used for dnn module functionlaity.
          * specific target. For layers that not represented in scheduling file
          * or if no manual scheduling used at all, automatic scheduling will be applied.
          */
-        void compileHalide(const std::string& scheduler = "");
+        void setHalideScheduler(const String& scheduler);
 
         /**
          * @brief Ask network to use specific computation backend where it supported.
@@ -379,19 +395,13 @@ namespace dnn //! This namespace is used for dnn module functionlaity.
         void setPreferableBackend(int backendId);
 
         /** @brief Sets the new value for the layer output blob
-         *  @param outputName descriptor of the updating layer output blob.
+         *  @param name descriptor of the updating layer output blob.
          *  @param blob new blob.
          *  @see connect(String, String) to know format of the descriptor.
          *  @note If updating blob is not empty then @p blob must have the same shape,
          *  because network reshaping is not implemented yet.
          */
-        CV_WRAP void setBlob(String outputName, const Mat &blob);
-
-        /** @brief Returns the layer output blob.
-         *  @param outputName the descriptor of the returning layer output blob.
-         *  @see connect(String, String)
-         */
-        CV_WRAP Mat getBlob(String outputName);
+        CV_WRAP void setInput(const Mat &blob, const String& name = "");
 
         /** @brief Sets the new value for the learned param of the layer.
          *  @param layer name or id of the layer.

modules/dnn/misc/python/pyopencv_dnn.hpp

@@ -3,6 +3,7 @@ typedef dnn::DictValue LayerId;
 typedef std::vector<dnn::MatShape> vector_MatShape;
 typedef std::vector<std::vector<dnn::MatShape> > vector_vector_MatShape;
 typedef std::vector<size_t> vector_size_t;
+typedef std::vector<std::vector<Mat> > vector_vector_Mat;
 
 template<>
 bool pyopencv_to(PyObject *o, dnn::DictValue &dv, const char *name)

modules/dnn/samples/caffe_googlenet.cpp

@@ -119,16 +119,14 @@ int main(int argc, char **argv)
     //! [Prepare blob]
 
     //! [Set input blob]
-    net.setBlob(".data", inputBlob);        //set the network input
+    net.setInput(inputBlob, "data");        //set the network input
     //! [Set input blob]
 
     //! [Make forward pass]
-    net.forward();                          //compute output
+    Mat prob = net.forward("prob");                          //compute output
     //! [Make forward pass]
 
     //! [Gather output]
-    Mat prob = net.getBlob("prob");   //gather output of "prob" layer
-
     int classId;
     double classProb;
     getMaxClass(prob, &classId, &classProb);//find the best class

modules/dnn/samples/fcn_semsegm.cpp

@@ -134,19 +134,16 @@ int main(int argc, char **argv)
     //! [Prepare blob]
 
     //! [Set input blob]
-    net.setBlob(".data", inputBlob);        //set the network input
+    net.setInput(inputBlob, "data");        //set the network input
     //! [Set input blob]
 
     //! [Make forward pass]
     double t = (double)cv::getTickCount();
-    net.forward();                          //compute output
+    Mat score = net.forward("score");                          //compute output
     t = (double)cv::getTickCount() - t;
     printf("processing time: %.1fms\n", t*1000./getTickFrequency());
     //! [Make forward pass]
 
-    //! [Gather output]
-    Mat score = net.getBlob("score");
-
     Mat colorize;
     colorizeSegmentation(score, colors, colorize);
     Mat show;

modules/dnn/samples/squeezenet_halide.cpp

@@ -93,23 +93,23 @@ int main(int argc, char **argv)
     //! [Prepare blob]
 
     //! [Set input blob]
-    net.setBlob("", inputBlob);                      // Set the network input.
+    net.setInput(inputBlob);                      // Set the network input.
     //! [Set input blob]
 
     //! [Enable Halide backend]
     net.setPreferableBackend(DNN_BACKEND_HALIDE);    // Tell engine to use Halide where it possible.
     //! [Enable Halide backend]
 
     //! [Compile Halide pipeline]
-    net.compileHalide();                             // Compile Halide pipeline.
+//    net.compileHalide();                             // Compile Halide pipeline.
     //! [Compile Halide pipeline]
 
     //! [Make forward pass]
-    net.forward();                                   // Compute output.
+    Mat prob = net.forward("prob");                                   // Compute output.
     //! [Make forward pass]
 
     //! [Gather output]
-    Mat prob = net.getBlob("prob");                  // Gather output of "prob" layer.
+//    net.getBlob();                  // Gather output of "prob" layer.
 
     int classId;
     double classProb;

modules/dnn/samples/ssd_object_detection.cpp

@@ -108,15 +108,13 @@ int main(int argc, char** argv)
     //! [Prepare blob]
 
     //! [Set input blob]
-    net.setBlob(".data", inputBlob);                //set the network input
+    net.setInput(inputBlob, "data");                //set the network input
     //! [Set input blob]
 
     //! [Make forward pass]
-    net.forward();                                  //compute output
+    Mat detection = net.forward("detection_out");                                  //compute output
     //! [Make forward pass]
 
-    //! [Gather output]
-    Mat detection = net.getBlob("detection_out");
     Mat detectionMat(detection.size[2], detection.size[3], CV_32F, detection.ptr<float>());
 
     float confidenceThreshold = parser.get<float>("min_confidence");

modules/dnn/samples/tf_inception.cpp

@@ -26,7 +26,7 @@ const String keys =
                        "https://storage.googleapis.com/download.tensorflow.org/models/inception5h.zip }"
         "{model m   |tensorflow_inception_graph.pb| path to TensorFlow .pb model file }"
         "{image i   || path to image file }"
-        "{i_blob    | .input | input blob name) }"
+        "{i_blob    | input | input blob name) }"
         "{o_blob    | softmax2 | output blob name) }"
         "{c_names c | imagenet_comp_graph_label_strings.txt | path to file with classnames for class id }"
         "{result r  || path to save output blob (optional, binary format, NCHW order) }"
@@ -101,21 +101,18 @@ int main(int argc, char **argv)
     //! [Prepare blob]
     inputBlob -= 117.0;
     //! [Set input blob]
-    net.setBlob(inBlobName, inputBlob);        //set the network input
+    net.setInput(inputBlob, inBlobName);        //set the network input
     //! [Set input blob]
 
     cv::TickMeter tm;
     tm.start();
 
     //! [Make forward pass]
-    net.forward();                          //compute output
+    Mat result = net.forward(outBlobName);                          //compute output
     //! [Make forward pass]
 
     tm.stop();
 
-    //! [Gather output]
-    Mat result = net.getBlob(outBlobName);   //gather output of "prob" layer
-
     if (!resultFile.empty()) {
         CV_Assert(result.isContinuous());
 

modules/dnn/samples/torch_enet.cpp

@@ -73,32 +73,19 @@ int main(int argc, char **argv)
     //! [Prepare blob]
 
     //! [Set input blob]
-    net.setBlob("", inputBlob);        //set the network input
+    net.setInput(inputBlob, "");        //set the network input
     //! [Set input blob]
 
-    const int N = 3;
     TickMeter tm;
 
-    //! [Make forward pass]
-    for( int i = 0; i < N; i++ )
-    {
-        TickMeter tm_;
-        tm_.start();
-        net.forward();                          //compute output
-        tm_.stop();
-        if( i == 0 || tm_.getTimeTicks() < tm.getTimeTicks() )
-            tm = tm_;
-    }
-
-    //! [Gather output]
-
     String oBlob = net.getLayerNames().back();
     if (!parser.get<String>("o_blob").empty())
     {
         oBlob = parser.get<String>("o_blob");
     }
 
-    Mat result = net.getBlob(oBlob);   //gather output of "prob" layer
+    //! [Make forward pass]
+    Mat result = net.forward(oBlob);
 
     if (!resultFile.empty()) {
         CV_Assert(result.isContinuous());

modules/dnn/src/caffe/caffe_importer.cpp

@@ -277,7 +277,7 @@ class CaffeImporter : public Importer
                 addedBlobs.push_back(BlobNote(net.input(inNum), 0, inNum));
                 netInputs[inNum] = net.input(inNum);
             }
-            dstNet.setNetInputs(netInputs);
+            dstNet.setInputsNames(netInputs);
         }
 
         for (int li = 0; li < layersSize; li++)