Merge branch 'master' of https://github.com/ceccocats/tkDNN

README.md

@@ -3,9 +3,10 @@ tkDNN is a Deep Neural Network library built with cuDNN primitives specifically
 The main scope is to do high performance inference on already trained models.
 
 this branch actually work on every NVIDIA GPU that support the dependencies:
-* CUDA 9
+* CUDA 10.0
-* CUDNN 7.105
+* CUDNN 7.603
-* TENSORRT 4.02
+* TENSORRT 6.01
+* OPENCV 4.1
 
 ## Workflow
 The recommended workflow follow these step:

demo/demo/demo.cpp

@@ -7,6 +7,7 @@
 
 #include <opencv2/core/core.hpp>
 #include <opencv2/highgui/highgui.hpp>
+#include <opencv2/videoio.hpp>
 #include <opencv2/imgproc/imgproc.hpp>
 
 #include "Yolo3Detection.h"
@@ -46,9 +47,9 @@ int main(int argc, char *argv[]) {
 
     cv::VideoWriter resultVideo;
     if(SAVE_RESULT) {
-        int w = cap.get(CV_CAP_PROP_FRAME_WIDTH);
+        int w = cap.get(cv::CAP_PROP_FRAME_WIDTH);
-        int h = cap.get(CV_CAP_PROP_FRAME_HEIGHT);
+        int h = cap.get(cv::CAP_PROP_FRAME_HEIGHT);
-        resultVideo.open("result.mp4", CV_FOURCC('M','P','4','V'), 30, cv::Size(w, h));
+        resultVideo.open("result.mp4", cv::VideoWriter::fourcc('M','P','4','V'), 30, cv::Size(w, h));
     }
 
     cv::Mat frame;

include/tkDNN/Layer.h

@@ -24,6 +24,8 @@ enum layerType_t {
     LAYER_YOLO
 };
 
+#define TKDNN_BN_MIN_EPSILON 1e-5
+
 /** 
     Simple layer Father class
 */

src/Conv2d.cpp

@@ -117,7 +117,7 @@ dnnType* Conv2d::infer(dataDim_t &dim, dnnType* srcData) {
             dstTensorDesc, dstData, dstTensorDesc, 
             dstData, biasTensorDesc, //same tensor descriptor as bias 
             scales_d, bias_d, mean_d, variance_d, 
-            CUDNN_BN_MIN_EPSILON);
+            TKDNN_BN_MIN_EPSILON);
     }
     //update data dimensions    
     dim = output_dim;

src/LayerWgs.cpp

@@ -29,7 +29,7 @@ LayerWgs::LayerWgs(Network *net, int inputs, int outputs,
         seek += outputs;
         readBinaryFile(weights_path.c_str(), outputs, &variance_h, &variance_d, seek);
 
-        float eps = CUDNN_BN_MIN_EPSILON;
+        float eps = TKDNN_BN_MIN_EPSILON;
 
         power_h = new dnnType[outputs];
         for(int i=0; i<outputs; i++) power_h[i] = 1.0f;

src/Network.cpp

@@ -17,10 +17,6 @@ Network::Network(dataDim_t input_dim) {
              <<", CUDNN v"<<cu_ver<<")\n";
     dataType = CUDNN_DATA_FLOAT;
     tensorFormat = CUDNN_TENSOR_NCHW;
-
-    checkCUDNN( cudnnCreate(&cudnnHandle) );
-    checkERROR( cublasCreate(&cublasHandle) );
-
     num_layers = 0;
 
     fp16 = false;
@@ -38,6 +34,11 @@ Network::Network(dataDim_t input_dim) {
         std::cout<<COL_REDB<<"!! FP16 INERENCE ENABLED !!"<<COL_END<<"\n";
     if(dla)
         std::cout<<COL_GREENB<<"!! DLA INERENCE ENABLED !!"<<COL_END<<"\n";
+
+
+    checkCUDNN( cudnnCreate(&cudnnHandle) );
+    checkERROR( cublasCreate(&cublasHandle) );
+
 }
 
 Network::~Network() {