diff --git a/README.md b/README.md index 93a7574..3b16ec3 100644 --- a/README.md +++ b/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 -* CUDNN 7.105 -* TENSORRT 4.02 +* CUDA 10.0 +* CUDNN 7.603 +* TENSORRT 6.01 +* OPENCV 4.1 ## Workflow The recommended workflow follow these step: diff --git a/include/tkDNN/Layer.h b/include/tkDNN/Layer.h index c6bee42..ee73109 100644 --- a/include/tkDNN/Layer.h +++ b/include/tkDNN/Layer.h @@ -24,6 +24,8 @@ enum layerType_t { LAYER_YOLO }; +#define TKDNN_BN_MIN_EPSILON 1e-5 + /** Simple layer Father class */ diff --git a/src/Conv2d.cpp b/src/Conv2d.cpp index 3dfdce6..b275198 100644 --- a/src/Conv2d.cpp +++ b/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; diff --git a/src/LayerWgs.cpp b/src/LayerWgs.cpp index f27da0f..7851164 100644 --- a/src/LayerWgs.cpp +++ b/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