Francesco Gatti
60 lines
1.3 KiB
C++
60 lines
1.3 KiB
C++
#include <iostream>
|
|
#include "NvInfer.h"
|
|
|
|
#include "tkdnn.h"
|
|
#include "Network.h"
|
|
#include "Layer.h"
|
|
|
|
namespace tkDNN {
|
|
|
|
Network::Network(dataDim_t input_dim) {
|
|
this->input_dim = input_dim;
|
|
|
|
float tk_ver = float(tkDNN::getVersion())/1000;
|
|
float cu_ver = float(cudnnGetVersion())/1000;
|
|
float rt_ver = float(NV_TENSORRT_MAJOR) + float(NV_TENSORRT_MINOR)/10 + float(NV_TENSORRT_PATCH)/100;
|
|
|
|
std::cout<<"New NETWORK (tkDNN v"<<tk_ver
|
|
<<", CUDNN v"<<cu_ver<<", TensorRT v"<<rt_ver<<")\n";
|
|
dataType = CUDNN_DATA_FLOAT;
|
|
tensorFormat = CUDNN_TENSOR_NCHW;
|
|
|
|
checkCUDNN( cudnnCreate(&cudnnHandle) );
|
|
checkERROR( cublasCreate(&cublasHandle) );
|
|
|
|
num_layers = 0;
|
|
}
|
|
|
|
Network::~Network() {
|
|
|
|
checkCUDNN( cudnnDestroy(cudnnHandle) );
|
|
checkERROR( cublasDestroy(cublasHandle) );
|
|
}
|
|
|
|
value_type* Network::infer(dataDim_t &dim, value_type* data) {
|
|
|
|
//do infer for every layer
|
|
for(int i=0; i<num_layers; i++)
|
|
data = layers[i]->infer(dim, data);
|
|
|
|
checkCuda(cudaDeviceSynchronize());
|
|
return data;
|
|
}
|
|
|
|
bool Network::addLayer(Layer *l) {
|
|
if(num_layers == MAX_LAYERS)
|
|
return false;
|
|
|
|
layers[num_layers++] = l;
|
|
return true;
|
|
}
|
|
|
|
dataDim_t Network::getOutputDim() {
|
|
|
|
if(num_layers == 0)
|
|
return input_dim;
|
|
else
|
|
return layers[num_layers-1]->output_dim;
|
|
}
|
|
|
|
} |