CMakeLists.txt

@@ -103,6 +103,9 @@ target_link_libraries(test_yolo3_flir tkDNN)
 add_executable(test_mobilenetv2ssd tests/mobilenetv2ssd/mobilenetv2ssd.cpp)
 target_link_libraries(test_mobilenetv2ssd tkDNN)
 
+add_executable(test_mobilenetv2ssd512 tests/mobilenetv2ssd512/mobilenetv2ssd512.cpp)
+target_link_libraries(test_mobilenetv2ssd512 tkDNN)
+
 add_executable(test_resnet101 tests/resnet101/resnet101.cpp)
 target_link_libraries(test_resnet101 tkDNN)
 

demo/demo/demo.cpp

@@ -44,7 +44,7 @@ int main(int argc, char *argv[]) {
             cnet.init(net);
             break;
         case 'm':
-            mbnet.init(net);
+            mbnet.init(net, 512, 81);
             break;
         default:
         FatalError("Network type not allowed (3rd parameter)\n");

include/tkDNN/MobilenetDetection.h

@@ -15,31 +15,31 @@
 
 struct SSDSpec
 {
-    int feature_size = 0;
+    int featureSize = 0;
     int shrinkage = 0;
-    int box_width = 0;
-    int box_height = 0;
+    int boxWidth = 0;
+    int boxHeight = 0;
     int ratio1 = 0;
     int ratio2 = 0;
 
     SSDSpec() {}
 
-    SSDSpec(int feature_size, int shrinkage, int box_width, int box_height, int ratio1, int ratio2) : feature_size(feature_size), shrinkage(shrinkage), box_width(box_width), box_height(box_height),
+    SSDSpec(int feature_size, int shrinkage, int box_width, int box_height, int ratio1, int ratio2) : featureSize(feature_size), shrinkage(shrinkage), boxWidth(box_width), boxHeight(box_height),
                                                                                                       ratio1(ratio1), ratio2(ratio2) {}
 
     void setAll(int feature_size, int shrinkage, int box_width, int box_height, int ratio1, int ratio2)
     {
-        this->feature_size = feature_size;
+        this->featureSize = feature_size;
         this->shrinkage = shrinkage;
-        this->box_width = box_width;
-        this->box_height = box_height;
+        this->boxWidth = box_width;
+        this->boxHeight = box_height;
         this->ratio1 = ratio1;
         this->ratio2 = ratio2;
     }
 
     void print()
     {
-        std::cout << "fsize: " << feature_size << "\tshrinkage: " << shrinkage << "\t box W:" << box_width << "\tbox H: " << box_height << "\t x ratio:" << ratio1 << "\t y ratio:" << ratio2 << std::endl;
+        std::cout << "fsize: " << featureSize << "\tshrinkage: " << shrinkage << "\t box W:" << boxWidth << "\tbox H: " << boxHeight << "\t x ratio:" << ratio1 << "\t y ratio:" << ratio2 << std::endl;
     }
 };
 
@@ -54,14 +54,12 @@ class MobilenetDetection
 private:
     tk::dnn::NetworkRT *netRT = nullptr;
 
-    int classes = 21;
-    float iou_threshold = 0.45;
-    float center_variance = 0.1;
-    float size_variance = 0.2;
-    float conf_thresh = 0.4;
-    int input_h = 300;
-    int input_w = 300;
-    int image_size = 300;
+    int classes;
+    float iouThreshold = 0.45;
+    float centerVariance = 0.1;
+    float sizeVariance = 0.2;
+    float confThresh = 0.4;
+    int imageSize;
 
     float *priors = nullptr;
     int n_priors = 0;
@@ -91,7 +89,7 @@ private:
     float get_color2(int c, int x, int max);
 
     cv::Scalar colors[256];
-    std::vector<std::string> voc_class_name;
+    std::vector<std::string> classesNames;
 
 public:
     // keep track of inference times (ms)
@@ -101,7 +99,7 @@ public:
     MobilenetDetection() {}
     ~MobilenetDetection() {}
 
-    void init(std::string tensor_path);
+    void init(std::string tensor_path, int input_size, int n_classes);
     cv::Mat draw();
     void update(cv::Mat &img);
 };

src/MobilenetDetection.cpp

@@ -15,7 +15,7 @@ void MobilenetDetection::generate_ssd_priors(const SSDSpec *specs, const int n_s
     n_priors = 0;
     for (int i = 0; i < n_specs; i++)
     {
-        n_priors += specs[i].feature_size * specs[i].feature_size * 6;
+        n_priors += specs[i].featureSize * specs[i].featureSize * 6;
     }
 
     // std::cout<<"n priors: "<<n_priors<<std::endl;
@@ -28,18 +28,18 @@ void MobilenetDetection::generate_ssd_priors(const SSDSpec *specs, const int n_s
     int min, max;
     for (int i = 0; i < n_specs; i++)
     {
-        scale = (float)image_size / (float)specs[i].shrinkage;
-        min = specs[i].box_height > specs[i].box_width ? specs[i].box_width : specs[i].box_height;
-        max = specs[i].box_height < specs[i].box_width ? specs[i].box_width : specs[i].box_height;
-        for (int j = 0; j < specs[i].feature_size; j++)
+        scale = (float)imageSize / (float)specs[i].shrinkage;
+        min = specs[i].boxHeight > specs[i].boxWidth ? specs[i].boxWidth : specs[i].boxHeight;
+        max = specs[i].boxHeight < specs[i].boxWidth ? specs[i].boxWidth : specs[i].boxHeight;
+        for (int j = 0; j < specs[i].featureSize; j++)
         {
-            for (int k = 0; k < specs[i].feature_size; k++)
+            for (int k = 0; k < specs[i].featureSize; k++)
             {
                 //small sized square box
                 size = min;
                 x_center = (k + 0.5f) / scale;
                 y_center = (j + 0.5f) / scale;
-                h = w = (float)size / (float)image_size;
+                h = w = (float)size / (float)imageSize;
 
                 priors[i_prio * N_COORDS + 0] = x_center;
                 priors[i_prio * N_COORDS + 1] = y_center;
@@ -49,7 +49,7 @@ void MobilenetDetection::generate_ssd_priors(const SSDSpec *specs, const int n_s
 
                 //big sized square box
                 size = sqrt(max * min);
-                h = w = (float)size / (float)image_size;
+                h = w = (float)size / (float)imageSize;
 
                 priors[i_prio * N_COORDS + 0] = x_center;
                 priors[i_prio * N_COORDS + 1] = y_center;
@@ -59,7 +59,7 @@ void MobilenetDetection::generate_ssd_priors(const SSDSpec *specs, const int n_s
 
                 //change h/w ratio of the small sized box
                 size = min;
-                h = w = size / (float)image_size;
+                h = w = size / (float)imageSize;
                 ratio = sqrt(specs[i].ratio1);
                 priors[i_prio * N_COORDS + 0] = x_center;
                 priors[i_prio * N_COORDS + 1] = y_center;
@@ -103,15 +103,15 @@ void MobilenetDetection::generate_ssd_priors(const SSDSpec *specs, const int n_s
     }
 }
 
-void MobilenetDetection::convert_locatios_to_boxes_and_center(float *priors, const int n_priors, float *locations, const float center_variance, const float size_variance)
+void MobilenetDetection::convert_locatios_to_boxes_and_center(float *priors, const int n_priors, float *locations, const float centerVariance, const float sizeVariance)
 {
     float cur_x, cur_y;
     for (int i = 0; i < n_priors; i++)
     {
-        locations[i * N_COORDS + 0] = locations[i * N_COORDS + 0] * center_variance * priors[i * N_COORDS + 2] + priors[i * N_COORDS + 0];
-        locations[i * N_COORDS + 1] = locations[i * N_COORDS + 1] * center_variance * priors[i * N_COORDS + 3] + priors[i * N_COORDS + 1];
-        locations[i * N_COORDS + 2] = exp(locations[i * N_COORDS + 2] * size_variance) * priors[i * N_COORDS + 2];
-        locations[i * N_COORDS + 3] = exp(locations[i * N_COORDS + 3] * size_variance) * priors[i * N_COORDS + 3];
+        locations[i * N_COORDS + 0] = locations[i * N_COORDS + 0] * centerVariance * priors[i * N_COORDS + 2] + priors[i * N_COORDS + 0];
+        locations[i * N_COORDS + 1] = locations[i * N_COORDS + 1] * centerVariance * priors[i * N_COORDS + 3] + priors[i * N_COORDS + 1];
+        locations[i * N_COORDS + 2] = exp(locations[i * N_COORDS + 2] * sizeVariance) * priors[i * N_COORDS + 2];
+        locations[i * N_COORDS + 3] = exp(locations[i * N_COORDS + 3] * sizeVariance) * priors[i * N_COORDS + 3];
 
         cur_x = locations[i * N_COORDS + 0];
         cur_y = locations[i * N_COORDS + 1];
@@ -221,16 +221,38 @@ float MobilenetDetection::get_color2(int c, int x, int max)
     return r;
 }
 
-void MobilenetDetection::init(std::string tensor_path)
+void MobilenetDetection::init(std::string tensor_path, int input_size, int n_classes)
 {
+    this->imageSize = input_size;
+    this->classes = n_classes;
+
     const int n_SSDSpec = 6;
     SSDSpec specs[6];
-    specs[0].setAll(19, 16, 60, 105, 2, 3);
-    specs[1].setAll(10, 32, 105, 150, 2, 3);
-    specs[2].setAll(5, 64, 150, 195, 2, 3);
-    specs[3].setAll(3, 100, 195, 240, 2, 3);
-    specs[4].setAll(2, 150, 240, 285, 2, 3);
-    specs[5].setAll(1, 300, 285, 330, 2, 3);
+
+
+    if(input_size == 300)
+    {
+        specs[0].setAll(19, 16, 60, 105, 2, 3);
+        specs[1].setAll(10, 32, 105, 150, 2, 3);
+        specs[2].setAll(5, 64, 150, 195, 2, 3);
+        specs[3].setAll(3, 100, 195, 240, 2, 3);
+        specs[4].setAll(2, 150, 240, 285, 2, 3);
+        specs[5].setAll(1, 300, 285, 330, 2, 3);
+    }
+    else if(input_size == 512)
+    {
+        specs[0].setAll(32, 16, 60, 105, 2, 3);
+        specs[1].setAll(16, 32, 105, 150, 2, 3);
+        specs[2].setAll(8, 64, 150, 195, 2, 3);
+        specs[3].setAll(4, 100, 195, 240, 2, 3);
+        specs[4].setAll(2, 150, 240, 285, 2, 3);
+        specs[5].setAll(1, 300, 285, 330, 2, 3);
+    }  
+    else
+    {
+        FatalError("Input size for mobilenet not supported");
+    }
+     
 
     generate_ssd_priors(specs, n_SSDSpec);
 
@@ -242,7 +264,7 @@ void MobilenetDetection::init(std::string tensor_path)
     locations_h = (float *)malloc(N_COORDS * n_priors * sizeof(float));
     confidences_h = (float *)malloc(n_priors * classes * sizeof(float));
 
-    dim = tk::dnn::dataDim_t(1, 3, input_w, input_h, 1);
+    dim = tk::dnn::dataDim_t(1, 3, imageSize, imageSize, 1);
 
     for (int c = 0; c < classes; c++)
     {
@@ -253,11 +275,39 @@ void MobilenetDetection::init(std::string tensor_path)
         colors[c] = cv::Scalar(int(255.0 * b), int(255.0 * g), int(255.0 * r));
     }
 
-    const char *voc_class_name_[] = {
+    if(classes == 21)
+    {
+        const char *classes_names_[] = {
         "BACKGROUND", "aeroplane", "bicycle", "bird", "boat", "bottle", "bus",
         "car", "cat", "chair", "cow", "diningtable", "dog", "horse", "motorbike",
         "person", "pottedplant", "sheep", "sofa", "train", "tvmonitor"};
-    voc_class_name = std::vector<std::string>(voc_class_name_, std::end(voc_class_name_));
+        classesNames = std::vector<std::string>(classes_names_, std::end(classes_names_));
+
+    }
+    else if (classes == 81)
+    {
+        const char *classes_names_[] = {
+        "BACKGROUND", "person" , "bicycle" , "car" , "motorbike" , "aeroplane" , "bus" ,
+        "train" , "truck" , "boat" , "traffic light" , "fire hydrant" , "stop sign" , 
+        "parking meter" , "bench" , "bird" , "cat" , "dog" , "horse" , "sheep" , "cow" , 
+        "elephant" , "bear" , "zebra" , "giraffe" , "backpack" , "umbrella" , "handbag" , 
+        "tie" , "suitcase" , "frisbee" , "skis" , "snowboard" , "sports ball" , "kite" , 
+        "baseball bat" , "baseball glove" , "skateboard" , "surfboard" , "tennis racket" , 
+        "bottle" , "wine glass" , "cup" , "fork" , "knife" , "spoon" , "bowl" , "banana" , 
+        "apple" , "sandwich" , "orange" , "broccoli" , "carrot" , "hot dog" , "pizza" , 
+        "donut" , "cake" , "chair" , "sofa" , "pottedplant" , "bed" , "diningtable" , 
+        "toilet" , "tvmonitor" , "laptop" , "mouse" , "remote" , "keyboard" , 
+        "cell phone" , "microwave" , "oven" , "toaster" , "sink" , "refrigerator" , 
+        "book" , "clock" , "vase" , "scissors" , "teddy bear" , "hair drier" , "toothbrush"};
+        classesNames = std::vector<std::string>(classes_names_, std::end(classes_names_));
+
+    }
+    else
+    {
+        FatalError("Number of classes not supported for mobilenet");
+    }
+    
+    
 }
 
 cv::Mat MobilenetDetection::draw()
@@ -266,7 +316,7 @@ cv::Mat MobilenetDetection::draw()
     for (size_t i = 0; i < detected.size(); i++)
     {
         b = detected[i];
-        std::string det_class = voc_class_name[b.cl];
+        std::string det_class = classesNames[b.cl];
         cv::rectangle(origImg, cv::Point(b.x, b.y), cv::Point(b.w, b.h), colors[b.cl], 2);
         // draw label
         cv::Size textSize = getTextSize(det_class, cv::FONT_HERSHEY_SIMPLEX, fontScale, thickness, &baseline);
@@ -319,8 +369,8 @@ void MobilenetDetection::update(cv::Mat &img)
     checkCuda(cudaMemcpy(locations_h, loc, N_COORDS * n_priors * sizeof(float), cudaMemcpyDeviceToHost));
 
     //postprocess
-    convert_locatios_to_boxes_and_center(priors, n_priors, locations_h, center_variance, size_variance);
-    detected = postprocess(locations_h, confidences_h, n_priors, conf_thresh, classes, iou_threshold, sz.width, sz.height);
+    convert_locatios_to_boxes_and_center(priors, n_priors, locations_h, centerVariance, sizeVariance);
+    detected = postprocess(locations_h, confidences_h, n_priors, confThresh, classes, iouThreshold, sz.width, sz.height);
 
     TIMER_STOP
     stats.push_back(t_ns);

tests/mobilenetv2ssd512/mobilenetv2ssd512.cpp

@@ -0,0 +1,540 @@
+#include <iostream>
+#include "tkdnn.h"
+
+
+const char *output_bin1 = "../tests/mobilenetv2ssd512/debug/classification_headers-5.bin";
+const char *output_bin2 = "../tests/mobilenetv2ssd512/debug/regression_headers-5.bin";
+const char *input_bin = "../tests/mobilenetv2ssd512/debug/input.bin";
+
+const char *conv0_bin = "../tests/mobilenetv2ssd512/layers/base_net-0-0.bin";
+const char *inverted_residual1[] = {
+    "../tests/mobilenetv2ssd512/layers/base_net-1-conv-0.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-1-conv-3.bin"};
+const char *inverted_residual2[] = {
+    "../tests/mobilenetv2ssd512/layers/base_net-2-conv-0.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-2-conv-3.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-2-conv-6.bin"};
+const char *inverted_residual3[] = {
+    "../tests/mobilenetv2ssd512/layers/base_net-3-conv-0.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-3-conv-3.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-3-conv-6.bin"};
+const char *inverted_residual4[] = {
+    "../tests/mobilenetv2ssd512/layers/base_net-4-conv-0.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-4-conv-3.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-4-conv-6.bin"};
+const char *inverted_residual5[] = {
+    "../tests/mobilenetv2ssd512/layers/base_net-5-conv-0.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-5-conv-3.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-5-conv-6.bin"};
+const char *inverted_residual6[] = {
+    "../tests/mobilenetv2ssd512/layers/base_net-6-conv-0.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-6-conv-3.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-6-conv-6.bin"};
+const char *inverted_residual7[] = {
+    "../tests/mobilenetv2ssd512/layers/base_net-7-conv-0.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-7-conv-3.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-7-conv-6.bin"};
+const char *inverted_residual8[] = {
+    "../tests/mobilenetv2ssd512/layers/base_net-8-conv-0.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-8-conv-3.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-8-conv-6.bin"};
+const char *inverted_residual9[] = {
+    "../tests/mobilenetv2ssd512/layers/base_net-9-conv-0.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-9-conv-3.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-9-conv-6.bin"};
+const char *inverted_residual10[] = {
+    "../tests/mobilenetv2ssd512/layers/base_net-10-conv-0.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-10-conv-3.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-10-conv-6.bin"};
+const char *inverted_residual11[] = {
+    "../tests/mobilenetv2ssd512/layers/base_net-11-conv-0.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-11-conv-3.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-11-conv-6.bin"};
+const char *inverted_residual12[] = {
+    "../tests/mobilenetv2ssd512/layers/base_net-12-conv-0.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-12-conv-3.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-12-conv-6.bin"};
+const char *inverted_residual13[] = {
+    "../tests/mobilenetv2ssd512/layers/base_net-13-conv-0.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-13-conv-3.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-13-conv-6.bin"};
+const char *inverted_residual14[] = {
+    "../tests/mobilenetv2ssd512/layers/base_net-14-conv-0.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-14-conv-3.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-14-conv-6.bin"};
+const char *inverted_residual15[] = {
+    "../tests/mobilenetv2ssd512/layers/base_net-15-conv-0.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-15-conv-3.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-15-conv-6.bin"};
+const char *inverted_residual16[] = {
+    "../tests/mobilenetv2ssd512/layers/base_net-16-conv-0.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-16-conv-3.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-16-conv-6.bin"};
+const char *inverted_residual17[] = {
+    "../tests/mobilenetv2ssd512/layers/base_net-17-conv-0.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-17-conv-3.bin",
+    "../tests/mobilenetv2ssd512/layers/base_net-17-conv-6.bin"};
+
+const char *conv18 = "../tests/mobilenetv2ssd512/layers/base_net-18-0.bin";
+
+const char *extras0[] = {
+    "../tests/mobilenetv2ssd512/layers/extras-0-conv-0.bin",
+    "../tests/mobilenetv2ssd512/layers/extras-0-conv-3.bin",
+    "../tests/mobilenetv2ssd512/layers/extras-0-conv-6.bin"};
+const char *extras1[] = {
+    "../tests/mobilenetv2ssd512/layers/extras-1-conv-0.bin",
+    "../tests/mobilenetv2ssd512/layers/extras-1-conv-3.bin",
+    "../tests/mobilenetv2ssd512/layers/extras-1-conv-6.bin"};
+const char *extras2[] = {
+    "../tests/mobilenetv2ssd512/layers/extras-2-conv-0.bin",
+    "../tests/mobilenetv2ssd512/layers/extras-2-conv-3.bin",
+    "../tests/mobilenetv2ssd512/layers/extras-2-conv-6.bin"};
+const char *extras3[] = {
+    "../tests/mobilenetv2ssd512/layers/extras-3-conv-0.bin",
+    "../tests/mobilenetv2ssd512/layers/extras-3-conv-3.bin",
+    "../tests/mobilenetv2ssd512/layers/extras-3-conv-6.bin"};
+
+const char *classification_header0[] = {
+    "../tests/mobilenetv2ssd512/layers/classification_headers-0-0.bin",
+    "../tests/mobilenetv2ssd512/layers/classification_headers-0-3.bin"};
+const char *classification_header1[] = {
+    "../tests/mobilenetv2ssd512/layers/classification_headers-1-0.bin",
+    "../tests/mobilenetv2ssd512/layers/classification_headers-1-3.bin"};
+const char *classification_header2[] = {
+    "../tests/mobilenetv2ssd512/layers/classification_headers-2-0.bin",
+    "../tests/mobilenetv2ssd512/layers/classification_headers-2-3.bin"};
+const char *classification_header3[] = {
+    "../tests/mobilenetv2ssd512/layers/classification_headers-3-0.bin",
+    "../tests/mobilenetv2ssd512/layers/classification_headers-3-3.bin"};
+const char *classification_header4[] = {
+    "../tests/mobilenetv2ssd512/layers/classification_headers-4-0.bin",
+    "../tests/mobilenetv2ssd512/layers/classification_headers-4-3.bin"};
+
+const char *classification_header5 = "../tests/mobilenetv2ssd512/layers/classification_headers-5.bin";
+
+const char *regression_header0[] = {
+    "../tests/mobilenetv2ssd512/layers/regression_headers-0-0.bin",
+    "../tests/mobilenetv2ssd512/layers/regression_headers-0-3.bin"};
+const char *regression_header1[] = {
+    "../tests/mobilenetv2ssd512/layers/regression_headers-1-0.bin",
+    "../tests/mobilenetv2ssd512/layers/regression_headers-1-3.bin"};
+const char *regression_header2[] = {
+    "../tests/mobilenetv2ssd512/layers/regression_headers-2-0.bin",
+    "../tests/mobilenetv2ssd512/layers/regression_headers-2-3.bin"};
+const char *regression_header3[] = {
+    "../tests/mobilenetv2ssd512/layers/regression_headers-3-0.bin",
+    "../tests/mobilenetv2ssd512/layers/regression_headers-3-3.bin"};
+const char *regression_header4[] = {
+    "../tests/mobilenetv2ssd512/layers/regression_headers-4-0.bin",
+    "../tests/mobilenetv2ssd512/layers/regression_headers-4-3.bin"};
+
+const char *regression_header5 = "../tests/mobilenetv2ssd512/layers/regression_headers-5.bin";
+
+
+int main()
+{
+
+    // downloadWeightsifDoNotExist(input_bin, "./tests/mobilenetv2ssd512");
+
+    int classes = 81;
+
+    // Network layout
+    tk::dnn::dataDim_t dim(1, 3, 512, 512, 1);
+    tk::dnn::Network net(dim);
+
+    tk::dnn::Conv2d conv1(&net, 32, 3, 3, 2, 2, 1, 1, conv0_bin, true);
+    tk::dnn::Activation relu3(&net, CUDNN_ACTIVATION_RELU);
+
+    //Inverted Residual 1
+
+    tk::dnn::Conv2d conv2(&net, 32, 3, 3, 1, 1, 1, 1, inverted_residual1[0], true, false, false, 32);
+    tk::dnn::Activation relu5(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d conv3(&net, 16, 1, 1, 1, 1, 0, 0, inverted_residual1[1], true);
+
+    //Inverted Residual 2
+    tk::dnn::Conv2d ir_2_conv1(&net, 96, 1, 1, 1, 1, 0, 0, inverted_residual2[0], true);
+    tk::dnn::Activation relu_2_1(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_2_conv2(&net, 96, 3, 3, 2, 2, 1, 1, inverted_residual2[1], true, false, false, 96);
+    tk::dnn::Activation relu_2_2(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_2_conv3(&net, 24, 1, 1, 1, 1, 0, 0, inverted_residual2[2], true);
+
+    //Inverted Residual 3
+    tk::dnn::Layer *last = &ir_2_conv3;
+    tk::dnn::Conv2d ir_3_conv1(&net, 144, 1, 1, 1, 1, 0, 0, inverted_residual3[0], true);
+    tk::dnn::Activation relu_3_1(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_3_conv2(&net, 144, 3, 3, 1, 1, 1, 1, inverted_residual3[1], true, false, false, 144);
+    tk::dnn::Activation relu_3_2(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_3_conv3(&net, 24, 1, 1, 1, 1, 0, 0, inverted_residual3[2], true);
+
+    tk::dnn::Shortcut s3_0(&net, last);
+    // //Inverted Residual 4
+    tk::dnn::Conv2d ir_4_conv1(&net, 144, 1, 1, 1, 1, 0, 0, inverted_residual4[0], true);
+    tk::dnn::Activation relu_4_1(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_4_conv2(&net, 144, 3, 3, 2, 2, 1, 1, inverted_residual4[1], true, false, false, 144);
+    tk::dnn::Activation relu_4_2(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_4_conv3(&net, 32, 1, 1, 1, 1, 0, 0, inverted_residual4[2], true);
+
+    // // //Inverted Residual 5
+    last = &ir_4_conv3;
+    tk::dnn::Conv2d ir_5_conv1(&net, 192, 1, 1, 1, 1, 0, 0, inverted_residual5[0], true);
+    tk::dnn::Activation relu_5_1(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_5_conv2(&net, 192, 3, 3, 1, 1, 1, 1, inverted_residual5[1], true, false, false, 192);
+    tk::dnn::Activation relu_5_2(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_5_conv3(&net, 32, 1, 1, 1, 1, 0, 0, inverted_residual5[2], true);
+
+    tk::dnn::Shortcut s5_0(&net, last);
+    // // // //Inverted Residual 6
+    last = &s5_0;
+    tk::dnn::Conv2d ir_6_conv1(&net, 192, 1, 1, 1, 1, 0, 0, inverted_residual6[0], true);
+    tk::dnn::Activation relu_6_1(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_6_conv2(&net, 192, 3, 3, 1, 1, 1, 1, inverted_residual6[1], true, false, false, 192);
+    tk::dnn::Activation relu_6_2(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_6_conv3(&net, 32, 1, 1, 1, 1, 0, 0, inverted_residual6[2], true);
+
+    tk::dnn::Shortcut s6_0(&net, last);
+    //Inverted Residual 7
+    tk::dnn::Conv2d ir_7_conv1(&net, 192, 1, 1, 1, 1, 0, 0, inverted_residual7[0], true);
+    tk::dnn::Activation relu_7_1(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_7_conv2(&net, 192, 3, 3, 2, 2, 1, 1, inverted_residual7[1], true, false, false, 192);
+    tk::dnn::Activation relu_7_2(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_7_conv3(&net, 64, 1, 1, 1, 1, 0, 0, inverted_residual7[2], true);
+
+    // //Inverted Residual 8
+    last = &ir_7_conv3;
+    tk::dnn::Conv2d ir_8_conv1(&net, 384, 1, 1, 1, 1, 0, 0, inverted_residual8[0], true);
+    tk::dnn::Activation relu_8_1(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_8_conv2(&net, 384, 3, 3, 1, 1, 1, 1, inverted_residual8[1], true, false, false, 384);
+    tk::dnn::Activation relu_8_2(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_8_conv3(&net, 64, 1, 1, 1, 1, 0, 0, inverted_residual8[2], true);
+
+    tk::dnn::Shortcut s8_0(&net, last);
+    //Inverted Residual 9
+    last = &s8_0;
+    tk::dnn::Conv2d ir_9_conv1(&net, 384, 1, 1, 1, 1, 0, 0, inverted_residual9[0], true);
+    tk::dnn::Activation relu_9_1(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_9_conv2(&net, 384, 3, 3, 1, 1, 1, 1, inverted_residual9[1], true, false, false, 384);
+    tk::dnn::Activation relu_9_2(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_9_conv3(&net, 64, 1, 1, 1, 1, 0, 0, inverted_residual9[2], true);
+
+    tk::dnn::Shortcut s9_0(&net, last);
+    //Inverted Residual 10
+    last = &s9_0;
+    tk::dnn::Conv2d ir_10_conv1(&net, 384, 1, 1, 1, 1, 0, 0, inverted_residual10[0], true);
+    tk::dnn::Activation relu_10_1(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_10_conv2(&net, 384, 3, 3, 1, 1, 1, 1, inverted_residual10[1], true, false, false, 384);
+    tk::dnn::Activation relu_10_2(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_10_conv3(&net, 64, 1, 1, 1, 1, 0, 0, inverted_residual10[2], true);
+
+    tk::dnn::Shortcut s10_0(&net, last);
+    //Inverted Residual 11
+    tk::dnn::Conv2d ir_11_conv1(&net, 384, 1, 1, 1, 1, 0, 0, inverted_residual11[0], true);
+    tk::dnn::Activation relu_11_1(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_11_conv2(&net, 384, 3, 3, 1, 1, 1, 1, inverted_residual11[1], true, false, false, 384);
+    tk::dnn::Activation relu_11_2(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_11_conv3(&net, 96, 1, 1, 1, 1, 0, 0, inverted_residual11[2], true);
+
+    last = &ir_11_conv3;
+    //Inverted Residual 12
+    tk::dnn::Conv2d ir_12_conv1(&net, 576, 1, 1, 1, 1, 0, 0, inverted_residual12[0], true);
+    tk::dnn::Activation relu_12_1(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_12_conv2(&net, 576, 3, 3, 1, 1, 1, 1, inverted_residual12[1], true, false, false, 576);
+    tk::dnn::Activation relu_12_2(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_12_conv3(&net, 96, 1, 1, 1, 1, 0, 0, inverted_residual12[2], true);
+
+    tk::dnn::Shortcut s12_0(&net, last);
+    last = &s12_0;
+    //Inverted Residual 13
+    tk::dnn::Conv2d ir_13_conv1(&net, 576, 1, 1, 1, 1, 0, 0, inverted_residual13[0], true);
+    tk::dnn::Activation relu_13_1(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_13_conv2(&net, 576, 3, 3, 1, 1, 1, 1, inverted_residual13[1], true, false, false, 576);
+    tk::dnn::Activation relu_13_2(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_13_conv3(&net, 96, 1, 1, 1, 1, 0, 0, inverted_residual13[2], true);
+
+    tk::dnn::Shortcut s13_0(&net, last);
+    // //Inverted Residual 14
+    tk::dnn::Conv2d ir_14_conv1(&net, 576, 1, 1, 1, 1, 0, 0, inverted_residual14[0], true);
+    tk::dnn::Activation relu_14_1(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_14_conv2(&net, 576, 3, 3, 2, 2, 1, 1, inverted_residual14[1], true, false, false, 576);
+    tk::dnn::Activation relu_14_2(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_14_conv3(&net, 160, 1, 1, 1, 1, 0, 0, inverted_residual14[2], true);
+
+    // //Inverted Residual 15
+    last = &ir_14_conv3;
+    tk::dnn::Conv2d ir_15_conv1(&net, 960, 1, 1, 1, 1, 0, 0, inverted_residual15[0], true);
+    tk::dnn::Activation relu_15_1(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_15_conv2(&net, 960, 3, 3, 1, 1, 1, 1, inverted_residual15[1], true, false, false, 960);
+    tk::dnn::Activation relu_15_2(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_15_conv3(&net, 160, 1, 1, 1, 1, 0, 0, inverted_residual15[2], true);
+
+    tk::dnn::Shortcut s15_0(&net, last);
+    //Inverted Residual 16
+    last = &s15_0;
+    tk::dnn::Conv2d ir_16_conv1(&net, 960, 1, 1, 1, 1, 0, 0, inverted_residual16[0], true);
+    tk::dnn::Activation relu_16_1(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_16_conv2(&net, 960, 3, 3, 1, 1, 1, 1, inverted_residual16[1], true, false, false, 960);
+    tk::dnn::Activation relu_16_2(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_16_conv3(&net, 160, 1, 1, 1, 1, 0, 0, inverted_residual16[2], true);
+
+    tk::dnn::Shortcut s16_0(&net, last);
+    //Inverted Residual 17
+    tk::dnn::Conv2d ir_17_conv1(&net, 960, 1, 1, 1, 1, 0, 0, inverted_residual17[0], true);
+    tk::dnn::Activation relu_17_1(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_17_conv2(&net, 960, 3, 3, 1, 1, 1, 1, inverted_residual17[1], true, false, false, 960);
+    tk::dnn::Activation relu_17_2(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d ir_17_conv3(&net, 320, 1, 1, 1, 1, 0, 0, inverted_residual17[2], true);
+
+    //Conv 18
+    tk::dnn::Conv2d ir_18_conv1(&net, 1280, 1, 1, 1, 1, 0, 0, conv18, true);
+    tk::dnn::Activation relu_18_1(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Layer *header_1[1] = {&relu_18_1};
+
+    // //extras Inverted Residual 0
+    tk::dnn::Conv2d e_0_conv1(&net, 256, 1, 1, 1, 1, 0, 0, extras0[0], true);
+    tk::dnn::Activation e_relu_0_1(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d e_0_conv2(&net, 256, 3, 3, 2, 2, 1, 1, extras0[1], true, false, false, 256);
+    tk::dnn::Activation e_relu_0_2(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d e_0_conv3(&net, 512, 1, 1, 1, 1, 0, 0, extras0[2], true);
+    tk::dnn::Layer *header_2[1] = {&e_0_conv3};
+
+    // //extras Inverted Residual 1
+    tk::dnn::Conv2d e_1_conv1(&net, 128, 1, 1, 1, 1, 0, 0, extras1[0], true);
+    tk::dnn::Activation e_relu_1_1(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d e_1_conv2(&net, 128, 3, 3, 2, 2, 1, 1, extras1[1], true, false, false, 128);
+    tk::dnn::Activation e_relu_1_2(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d e_1_conv3(&net, 256, 1, 1, 1, 1, 0, 0, extras1[2], true);
+    tk::dnn::Layer *header_3[1] = {&e_1_conv3};
+
+    //extras Inverted Residual 2
+    tk::dnn::Conv2d e_2_conv1(&net, 128, 1, 1, 1, 1, 0, 0, extras2[0], true);
+    tk::dnn::Activation e_relu_2_1(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d e_2_conv2(&net, 128, 3, 3, 2, 2, 1, 1, extras2[1], true, false, false, 128);
+    tk::dnn::Activation e_relu_2_2(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d e_2_conv3(&net, 256, 1, 1, 1, 1, 0, 0, extras2[2], true);
+    tk::dnn::Layer *header_4[1] = {&e_2_conv3};
+
+    //extras Inverted Residual 3
+    tk::dnn::Conv2d e_3_conv1(&net, 64, 1, 1, 1, 1, 0, 0, extras3[0], true);
+    tk::dnn::Activation e_relu_3_1(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d e_3_conv2(&net, 64, 3, 3, 2, 2, 1, 1, extras3[1], true, false, false, 64);
+    tk::dnn::Activation e_relu_3_2(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d e_3_conv3(&net, 64, 1, 1, 1, 1, 0, 0, extras3[2], true);
+    tk::dnn::Layer *header_5[1] = {&e_3_conv3};
+
+    // classification header 0
+    tk::dnn::Layer *header_0[1] = {&relu_14_1};
+    tk::dnn::Route rout_ch_0(&net, header_0, 1);
+    tk::dnn::Conv2d ch_0_conv1(&net, 576, 3, 3, 1, 1, 1, 1, classification_header0[0], true, false, false, 576, true);
+    tk::dnn::Activation ch_relu_0_1(&net, CUDNN_ACTIVATION_CLIPPED_RELU, 6);
+    tk::dnn::Conv2d ch_0_conv2(&net, 486, 1, 1, 1, 1, 0, 0, classification_header0[1], false);
+    tk::dnn::Layer *conf0[1] = {&ch_0_conv2};
+
+    // // classification header 1
+    tk::dnn::Route rout_ch_1(&net, header_1, 1);
+    tk::dnn::Conv2d ch_1_conv1(&net, 1280, 3, 3, 1, 1, 1, 1, classification_header1[0], true, false, false, 1280, true);
+    tk::dnn::Activation ch_relu_1_1(&net, CUDNN_ACTIVATION_CLIPPED_RELU, 6);
+    tk::dnn::Conv2d ch_1_conv2(&net, 486, 1, 1, 1, 1, 0, 0, classification_header1[1], false);
+    tk::dnn::Layer *conf1[1] = {&ch_1_conv2};
+
+    // //classification header 2
+    tk::dnn::Route rout_ch_2(&net, header_2, 1);
+    tk::dnn::Conv2d ch_2_conv1(&net, 512, 3, 3, 1, 1, 1, 1, classification_header2[0], true, false, false, 512, true);
+    tk::dnn::Activation ch_relu_2_1(&net, CUDNN_ACTIVATION_CLIPPED_RELU, 6);
+    tk::dnn::Conv2d ch_2_conv2(&net, 486, 1, 1, 1, 1, 0, 0, classification_header2[1], false);
+    tk::dnn::Layer *conf2[1] = {&ch_2_conv2};
+
+    // //classification header 3
+    tk::dnn::Route rout_ch_3(&net, header_3, 1);
+    tk::dnn::Conv2d ch_3_conv1(&net, 256, 3, 3, 1, 1, 1, 1, classification_header3[0], true, false, false, 256, true);
+    tk::dnn::Activation ch_relu_3_1(&net, CUDNN_ACTIVATION_CLIPPED_RELU, 6);
+    tk::dnn::Conv2d ch_3_conv2(&net, 486, 1, 1, 1, 1, 0, 0, classification_header3[1], false);
+    tk::dnn::Layer *conf3[1] = {&ch_3_conv2};
+
+    // //classification header 4
+    tk::dnn::Route rout_ch_4(&net, header_4, 1);
+    tk::dnn::Conv2d ch_4_conv1(&net, 256, 3, 3, 1, 1, 1, 1, classification_header4[0], true, false, false, 256, true);
+    tk::dnn::Activation ch_relu_4_1(&net, CUDNN_ACTIVATION_CLIPPED_RELU, 6);
+    tk::dnn::Conv2d ch_4_conv2(&net, 486, 1, 1, 1, 1, 0, 0, classification_header4[1], false);
+    tk::dnn::Layer *conf4[1] = {&ch_4_conv2};
+
+    // //classification header 5
+    tk::dnn::Route rout_ch_5(&net, header_5, 1);
+    tk::dnn::Conv2d ch_5_conv(&net, 486, 1, 1, 1, 1, 0, 0, classification_header5, false, false, true);
+    tk::dnn::Layer *conf5[1] = {&ch_5_conv};
+
+    //regression header 0
+    tk::dnn::Route rout_rh_0(&net, header_0, 1);
+    tk::dnn::Conv2d rh_0_conv1(&net, 576, 3, 3, 1, 1, 1, 1, regression_header0[0], true, false, false, 576, true);
+    tk::dnn::Activation rh_relu_0_1(&net, CUDNN_ACTIVATION_CLIPPED_RELU, 6);
+    tk::dnn::Conv2d rh_0_conv2(&net, 24, 1, 1, 1, 1, 0, 0, regression_header0[1], false);
+    tk::dnn::Layer *loc0[1] = {&rh_0_conv2};
+
+    // //regression header 1
+    tk::dnn::Route rout_rh_1(&net, header_1, 1);
+    tk::dnn::Conv2d rh_1_conv1(&net, 1280, 3, 3, 1, 1, 1, 1, regression_header1[0], true, false, false, 1280, true);
+    tk::dnn::Activation rh_relu_1_1(&net, CUDNN_ACTIVATION_CLIPPED_RELU, 6);
+    tk::dnn::Conv2d rh_1_conv2(&net, 24, 1, 1, 1, 1, 0, 0, regression_header1[1], false);
+    tk::dnn::Layer *loc1[1] = {&rh_1_conv2};
+
+    //regression header 2
+    tk::dnn::Route rout_rh_2(&net, header_2, 1);
+    tk::dnn::Conv2d rh_2_conv1(&net, 512, 3, 3, 1, 1, 1, 1, regression_header2[0], true, false, false, 512, true);
+    tk::dnn::Activation rh_relu_2_1(&net, CUDNN_ACTIVATION_CLIPPED_RELU, 6);
+    tk::dnn::Conv2d rh_2_conv2(&net, 24, 1, 1, 1, 1, 0, 0, regression_header2[1], false);
+    tk::dnn::Layer *loc2[1] = {&rh_2_conv2};
+
+    //regression header 3
+    tk::dnn::Route rout_rh_3(&net, header_3, 1);
+    tk::dnn::Conv2d rh_3_conv1(&net, 256, 3, 3, 1, 1, 1, 1, regression_header3[0], true, false, false, 256, true);
+    tk::dnn::Activation rh_relu_3_1(&net, CUDNN_ACTIVATION_CLIPPED_RELU, 6);
+    tk::dnn::Conv2d rh_3_conv2(&net, 24, 1, 1, 1, 1, 0, 0, regression_header3[1], false);
+    tk::dnn::Layer *loc3[1] = {&rh_3_conv2};
+
+    //regression header 4
+
+    tk::dnn::Route rout_rh_4(&net, header_4, 1);
+    tk::dnn::Conv2d rh_4_conv1(&net, 256, 3, 3, 1, 1, 1, 1, regression_header4[0], true, false, false, 256, true);
+    tk::dnn::Activation rh_relu_4_1(&net, CUDNN_ACTIVATION_CLIPPED_RELU, 6);
+    tk::dnn::Conv2d rh_4_conv2(&net, 24, 1, 1, 1, 1, 0, 0, regression_header4[1], false);
+    tk::dnn::Layer *loc4[1] = {&rh_4_conv2};
+
+    //regression header 5
+    tk::dnn::Route rout_rh_5(&net, header_5, 1);
+    tk::dnn::Conv2d rh_5_conv(&net, 24, 1, 1, 1, 1, 0, 0, regression_header5, false, false, true);
+    tk::dnn::Layer *loc5[1] = {&rh_5_conv};
+
+    last = &rh_5_conv;
+
+    //flatten all confidence
+    tk::dnn::Route r_conf_0(&net, conf0, 1);
+    tk::dnn::Flatten fl_c_0(&net);
+    tk::dnn::Route r_conf_1(&net, conf1, 1);
+    tk::dnn::Flatten fl_c_1(&net);
+    tk::dnn::Route r_conf_2(&net, conf2, 1);
+    tk::dnn::Flatten fl_c_2(&net);
+    tk::dnn::Route r_conf_3(&net, conf3, 1);
+    tk::dnn::Flatten fl_c_3(&net);
+    tk::dnn::Route r_conf_4(&net, conf4, 1);
+    tk::dnn::Flatten fl_c_4(&net);
+    tk::dnn::Route r_conf_5(&net, conf5, 1);
+    tk::dnn::Flatten fl_c_5(&net);
+
+    // //flatten all locations
+    tk::dnn::Route r_loc_0(&net, loc0, 1);
+    tk::dnn::Flatten fl_l_0(&net);
+    tk::dnn::Route r_loc_1(&net, loc1, 1);
+    tk::dnn::Flatten fl_l_1(&net);
+    tk::dnn::Route r_loc_2(&net, loc2, 1);
+    tk::dnn::Flatten fl_l_2(&net);
+    tk::dnn::Route r_loc_3(&net, loc3, 1);
+    tk::dnn::Flatten fl_l_3(&net);
+    tk::dnn::Route r_loc_4(&net, loc4, 1);
+    tk::dnn::Flatten fl_l_4(&net);
+    tk::dnn::Route r_loc_5(&net, loc5, 1);
+    tk::dnn::Flatten fl_l_5(&net);
+
+    // //concat confidence + softmax
+    tk::dnn::Layer *confidences[6] = {&fl_c_0, &fl_c_1, &fl_c_2, &fl_c_3, &fl_c_4, &fl_c_5};
+    tk::dnn::Route rout_conf(&net, confidences, 6);
+    tk::dnn::dataDim_t olddim_c = net.layers[net.num_layers - 1]->output_dim;
+    tk::dnn::dataDim_t dim_resh(1, olddim_c.c * olddim_c.h * olddim_c.w / classes, classes, 1, 1);
+
+    tk::dnn::Reshape reshape_conf1(&net, dim_resh);
+    tk::dnn::Flatten fl_l_6(&net);
+    tk::dnn::dataDim_t newdim_c(1, classes, olddim_c.c * olddim_c.h * olddim_c.w / classes, 1, 1);
+
+    tk::dnn::Reshape reshape_conf2(&net, newdim_c);
+
+    tk::dnn::Softmax sm_1(&net, &newdim_c, true);
+    // tk::dnn::Flatten fl_l_7(&net);
+    // tk::dnn::Reshape reshape_conf3(&net,dim_resh, true);
+    tk::dnn::Layer *conf = &sm_1;
+
+    //concat locations
+    tk::dnn::Layer *locations[6] = {&fl_l_0, &fl_l_1, &fl_l_2, &fl_l_3, &fl_l_4, &fl_l_5};
+    tk::dnn::Route rout_loc(&net, locations, 6);
+    tk::dnn::dataDim_t olddim_l = net.layers[net.num_layers - 1]->output_dim;
+    tk::dnn::dataDim_t newdim_l(1, olddim_l.c * olddim_l.h * olddim_l.w / 4, 1, 4, 1);
+    tk::dnn::Reshape reshape_loc(&net, newdim_l, true);
+    tk::dnn::Layer *loc = &reshape_loc;
+
+    // Load input
+    dnnType *data;
+    dnnType *input_h;
+    readBinaryFile(input_bin, dim.tot(), &input_h, &data);
+    //printDeviceVector(64, data, true);
+
+    //print network model
+    net.print();
+
+    // convert network to tensorRT
+    tk::dnn::NetworkRT netRT(&net, "mobilenetv2ssd512.rt");
+
+    tk::dnn::dataDim_t dim1 = dim; //input dim
+    printCenteredTitle(" CUDNN inference ", '=', 30);
+    {
+        dim1.print();
+        TIMER_START
+        net.infer(dim1, data);
+        TIMER_STOP
+        dim1.print();
+    }
+
+    dnnType *cudnn_out1 = conf5[0]->dstData;
+    tk::dnn::dataDim_t out_dim1 = conf5[0]->output_dim;
+    dnnType *cudnn_out2 = loc5[0]->dstData;
+    tk::dnn::dataDim_t out_dim2 = loc5[0]->output_dim;
+
+    tk::dnn::dataDim_t dim2 = dim;
+    printCenteredTitle(" TENSORRT inference ", '=', 30);
+    {
+        dim2.print();
+        TIMER_START
+        netRT.infer(dim2, data);
+        TIMER_STOP
+        dim2.print();
+    }
+
+    dnnType *rt_out1 = (dnnType *)netRT.buffersRT[1];
+    dnnType *rt_out2 = (dnnType *)netRT.buffersRT[2];
+    dnnType *rt_out3 = (dnnType *)netRT.buffersRT[3];
+    dnnType *rt_out4 = (dnnType *)netRT.buffersRT[4];
+
+    printCenteredTitle(std::string(" RESNET CHECK RESULTS ").c_str(), '=', 30);
+    dnnType *out1, *out1_h;
+    int odim1 = out_dim1.tot();
+    readBinaryFile(output_bin1, odim1, &out1_h, &out1);
+
+    dnnType *out2, *out2_h;
+    int odim2 = out_dim2.tot();
+    readBinaryFile(output_bin2, odim2, &out2_h, &out2);
+    std::cout << "CUDNN vs correct" << std::endl;
+    checkResult(odim1, cudnn_out1, out1);
+    checkResult(odim2, cudnn_out2, out2);
+
+    std::cout << "TRT   vs correct" << std::endl;
+    checkResult(odim1, rt_out1, out1);
+    checkResult(odim2, rt_out2, out2);
+
+    std::cout << "CUDNN vs TRT    " << std::endl;
+    checkResult(odim1, cudnn_out1, rt_out1);
+    checkResult(odim2, cudnn_out2, rt_out2);
+
+    std::cout << "---------------------------------------------------" << std::endl;
+    std::cout << "Confidence CUDNN" << std::endl;
+    printDeviceVector(64, conf->dstData, true);
+    std::cout << "Locations CUDNN" << std::endl;
+    printDeviceVector(64, loc->dstData, true);
+    std::cout << "---------------------------------------------------" << std::endl;
+
+    std::cout << "Confidence tensorRT" << std::endl;
+    printDeviceVector(64, rt_out3, true);
+    std::cout << "Locations tensorRT" << std::endl;
+    printDeviceVector(64, rt_out4, true);
+    std::cout << "---------------------------------------------------" << std::endl;
+
+    std::cout << "CUDNN vs TRT    " << std::endl;
+    checkResult(conf->output_dim.tot(), conf->dstData, rt_out3);
+    checkResult(loc->output_dim.tot(), loc->dstData, rt_out4);
+
+    return 0;
+}