add DLA, plugin for shortcut and leaky. new verison 0.4

Author: mive93

Diff for: demo/demo/demo.cpp

@@ -97,6 +97,13 @@ int main(int argc, char *argv[]) {
     }
 
     std::cout<<"detection end\n";   
+
+
+    std::cout<<COL_GREENB<<"\n\nTime stats:\n";
+    std::cout<<"Min: "<<*std::min_element(yolo.stats.begin(), yolo.stats.end())<<" ms\n";    
+    std::cout<<"Max: "<<*std::max_element(yolo.stats.begin(), yolo.stats.end())<<" ms\n";    
+    double mean = 0; for(int i=0; i<yolo.stats.size(); i++) mean += yolo.stats[i]; mean /= yolo.stats.size();
+    std::cout<<"Avg: "<<mean<<" ms\n"<<COL_END;    
     return 0;
 }
 

Diff for: include/tkDNN/Yolo3Detection.h

@@ -39,6 +39,9 @@ class Yolo3Detection {
         // this is filled with results
         std::vector<tk::dnn::box> detected;
 
+        // keep track of inference times (ms)
+        std::vector<double> stats;
+
         Yolo3Detection() {}
 
         virtual ~Yolo3Detection() {}

Diff for: include/tkDNN/tkdnn.h

@@ -5,4 +5,4 @@
 #include "Layer.h"
 #include "NetworkRT.h"
 
-#define TKDNN_VERSION 300
+#define TKDNN_VERSION 400

Diff for: src/NetworkRT.cpp

@@ -35,7 +35,7 @@ NetworkRT::NetworkRT(Network *net, const char *name) {
     builderRT = createInferBuilder(loggerRT);
     std::cout<<"Float16 support: "<<builderRT->platformHasFastFp16()<<"\n";
     std::cout<<"Int8 support: "<<builderRT->platformHasFastInt8()<<"\n";
-    //std::cout<<"DLAs: "<<builderRT->getNbDLACores()<<"\n";
+    std::cout<<"DLAs: "<<builderRT->getNbDLACores()<<"\n";
     networkRT = builderRT->createNetwork();
 
     if(!fileExist(name)) {
@@ -51,15 +51,13 @@ NetworkRT::NetworkRT(Network *net, const char *name) {
             dtRT = DataType::kHALF;
             builderRT->setHalf2Mode(true);
         }
-        /*
         if(net->dla && builderRT->getNbDLACores() > 0) {
             dtRT = DataType::kHALF;
             builderRT->setFp16Mode(true);
             builderRT->allowGPUFallback(true);
             builderRT->setDefaultDeviceType(DeviceType::kDLA);
             builderRT->setDLACore(0);
         }
-        */
 
         //add input layer
         ITensor *input = networkRT->addInput("data", DataType::kFLOAT, 
@@ -276,10 +274,19 @@ ILayer* NetworkRT::convert_layer(ITensor *input, Activation *l) {
 
     if(l->act_mode == ACTIVATION_LEAKY) {
         //std::cout<<"New plugin LEAKY\n";
+        
+        /*
+        // plugin version
         IPlugin *plugin = new ActivationLeakyRT();
         IPluginLayer *lRT = networkRT->addPlugin(&input, 1, *plugin);
         checkNULL(lRT);
         return lRT;
+        */
+
+        IActivationLayer *lRT = networkRT->addActivation(*input, ActivationType::kLEAKY_RELU);
+        lRT->setAlpha(0.1);
+        checkNULL(lRT);
+        return lRT;
 
     } else if(l->act_mode == CUDNN_ACTIVATION_RELU) {
         IActivationLayer *lRT = networkRT->addActivation(*input, ActivationType::kRELU);
@@ -340,14 +347,21 @@ ILayer* NetworkRT::convert_layer(ITensor *input, Shortcut *l) {
     //std::cout<<"convert Shortcut\n";
 
     //std::cout<<"New plugin Shortcut\n";
+    
     ITensor *back_tens = tensors[l->backLayer];
+    /*
+    // plugin version
     IPlugin *plugin = new ShortcutRT();
-
     ITensor **inputs = new ITensor*[2];
     inputs[0] = input;
     inputs[1] = back_tens; 
     IPluginLayer *lRT = networkRT->addPlugin(inputs, 2, *plugin);
     checkNULL(lRT);
+    */
+
+    IElementWiseLayer *lRT = networkRT->addElementWise(*input, *back_tens, ElementWiseOperation::kSUM);
+    checkNULL(lRT);
+    
     return lRT;
 }
 

Diff for: src/Yolo3Detection.cpp

@@ -94,6 +94,8 @@ void Yolo3Detection::update(cv::Mat &imageORIG) {
         netRT->infer(dim, input_d);
         TIMER_STOP
         dim.print();
+    
+        stats.push_back(t_ns);
     }
 
     TIMER_START