mmr/tkDNN
Archived
1
0
Fork 0
Code Issues 46 Pull Requests 12 Releases 4 Activity

Adapt detection classes to use batches, adapt demos, update README

Browse Source 
Signed-off-by: Micaela Verucchi <micaelaverucchi@gmail.com>
This commit is contained in:
Micaela Verucchi
2020-05-14 16:41:51 +02:00
parent 2e3cb52cff
commit 40456592fc
10 changed files with 149 additions and 103 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/CenternetDetection.cpp
+14 -12
View File
@@ -3,10 +3,11 @@
namespace tk { namespace dnn {
bool CenternetDetection::init(const std::string& tensor_path, const int n_classes){
bool CenternetDetection::init(const std::string& tensor_path, const int n_classes, const int n_batches){
std::cout<<(tensor_path).c_str()<<"\n";
netRT = new tk::dnn::NetworkRT(NULL, (tensor_path).c_str() );
classes = n_classes;
nBatches = n_batches;
dim = netRT->input_dim;
@@ -41,7 +42,7 @@ bool CenternetDetection::init(const std::string& tensor_path, const int n_classe
trans = cv::Mat(cv::Size(3,2), CV_32F);
trans2 = cv::Mat(cv::Size(3,2), CV_32F);
checkCuda(cudaMalloc(&input_d, sizeof(dnnType)*netRT->input_dim.tot()));
checkCuda(cudaMalloc(&input_d, sizeof(dnnType)*netRT->input_dim.tot() * nBatches));
dim_hm = tk::dnn::dataDim_t(1, 80, 128, 128, 1);
dim_wh = tk::dnn::dataDim_t(1, 2, 128, 128, 1);
@@ -98,7 +99,7 @@ bool CenternetDetection::init(const std::string& tensor_path, const int n_classe
checkCuda(cudaMemcpy(mean_d, mean, 3*sizeof(float), cudaMemcpyHostToDevice));
checkCuda(cudaMemcpy(stddev_d, stddev, 3*sizeof(float), cudaMemcpyHostToDevice));
#else
checkCuda(cudaMallocHost(&input, sizeof(dnnType)*netRT->input_dim.tot()));
checkCuda(cudaMallocHost(&input, sizeof(dnnType)*netRT->input_dim.tot()* nBatches));
mean << 0.408, 0.447, 0.47;
stddev << 0.289, 0.274, 0.278;
#endif
@@ -120,7 +121,7 @@ bool CenternetDetection::init(const std::string& tensor_path, const int n_classe
}
void CenternetDetection::preprocess(cv::Mat &frame){
void CenternetDetection::preprocess(cv::Mat &frame, const int bi){
// -----------------------------------pre-process ------------------------------------------
// auto start_t = std::chrono::steady_clock::now();
@@ -212,7 +213,7 @@ void CenternetDetection::preprocess(cv::Mat &frame){
// std::cout << " TIME normalize: " << std::chrono::duration_cast<std::chrono:: microseconds>(end_t - step_t).count() << " us" << std::endl;
// step_t = end_t;
checkCuda(cudaMemcpy(input_d, d_ptrs, dim2.tot()*sizeof(dnnType), cudaMemcpyDeviceToDevice));
checkCuda(cudaMemcpy(input_d+ netRT->input_dim.tot()*bi, d_ptrs, dim2.tot()*sizeof(dnnType), cudaMemcpyDeviceToDevice));
// end_t = std::chrono::steady_clock::now();
// std::cout << " TIME Memcpy to input_d: " << std::chrono::duration_cast<std::chrono:: microseconds>(end_t - step_t).count() << " us" << std::endl;
@@ -254,18 +255,18 @@ void CenternetDetection::preprocess(cv::Mat &frame){
int idx = i*imageF.rows*imageF.cols;
int ch = dim2.c-3 +i;
// std::cout<<"i: "<<i<<", idx: "<<idx<<", ch: "<<ch<<std::endl;
memcpy((void*)&input[idx], (void*)bgr[ch].data, imageF.rows*imageF.cols*sizeof(dnnType));
memcpy((void*)&input[idx+ netRT->input_dim.tot()*bi], (void*)bgr[ch].data, imageF.rows*imageF.cols*sizeof(dnnType));
}
checkCuda(cudaMemcpyAsync(input_d, input, dim2.tot()*sizeof(dnnType), cudaMemcpyHostToDevice));
checkCuda(cudaMemcpyAsync(input_d+ netRT->input_dim.tot()*bi, input+ netRT->input_dim.tot()*bi, dim2.tot()*sizeof(dnnType), cudaMemcpyHostToDevice));
#endif
}
void CenternetDetection::postprocess(const bool mAP){
void CenternetDetection::postprocess(const int bi, const bool mAP){
dnnType *rt_out[4];
rt_out[0] = (dnnType *)netRT->buffersRT[1];
rt_out[1] = (dnnType *)netRT->buffersRT[2];
rt_out[2] = (dnnType *)netRT->buffersRT[3];
rt_out[3] = (dnnType *)netRT->buffersRT[4];
rt_out[0] = (dnnType *)netRT->buffersRT[1]+ netRT->buffersDIM[0].tot()*bi;
rt_out[1] = (dnnType *)netRT->buffersRT[2]+ netRT->buffersDIM[1].tot()*bi;
rt_out[2] = (dnnType *)netRT->buffersRT[3]+ netRT->buffersDIM[2].tot()*bi;
rt_out[3] = (dnnType *)netRT->buffersRT[4]+ netRT->buffersDIM[3].tot()*bi;
// auto start_t = std::chrono::steady_clock::now();
// auto step_t = std::chrono::steady_clock::now();
@@ -389,6 +390,7 @@ void CenternetDetection::postprocess(const bool mAP){
}
}
batchDetected.push_back(detected);
// end_t = std::chrono::steady_clock::now();
// std::cout << " TIME detections: " << std::chrono::duration_cast<std::chrono:: microseconds>(end_t - step_t).count() << " us" << std::endl;
// step_t = end_t;
src/MobilenetDetection.cpp
+12 -10
View File
@@ -126,11 +126,12 @@ float MobilenetDetection::iou(const tk::dnn::box &a, const tk::dnn::box &b){
return iou;
}
bool MobilenetDetection::init(const std::string& tensor_path, const int n_classes){
bool MobilenetDetection::init(const std::string& tensor_path, const int n_classes, const int n_batches){
std::cout<<(tensor_path).c_str()<<"\n";
netRT = new tk::dnn::NetworkRT(NULL, (tensor_path).c_str());
imageSize = netRT->input_dim.h;
classes = n_classes;
nBatches = n_batches;
SSDSpec specs[N_SSDSPEC];
@@ -157,9 +158,9 @@ bool MobilenetDetection::init(const std::string& tensor_path, const int n_classe
generate_ssd_priors(specs, N_SSDSPEC);
#ifndef OPENCV_CUDACONTRIB
checkCuda(cudaMallocHost(&input, sizeof(dnnType) * netRT->input_dim.tot()));
checkCuda(cudaMallocHost(&input, sizeof(dnnType) * netRT->input_dim.tot() * nBatches));
#endif
checkCuda(cudaMalloc(&input_d, sizeof(dnnType) * netRT->input_dim.tot()));
checkCuda(cudaMalloc(&input_d, sizeof(dnnType) * netRT->input_dim.tot() * nBatches));
locations_h = (float *)malloc(N_COORDS * nPriors * sizeof(float));
confidences_h = (float *)malloc(nPriors * classes * sizeof(float));
@@ -208,7 +209,7 @@ bool MobilenetDetection::init(const std::string& tensor_path, const int n_classe
return 1;
}
void MobilenetDetection::preprocess(cv::Mat &frame){
void MobilenetDetection::preprocess(cv::Mat &frame, const int bi){
#ifdef OPENCV_CUDACONTRIB
//move original image on GPU
cv::cuda::GpuMat orig_img, frame_nomean;
@@ -224,7 +225,7 @@ void MobilenetDetection::preprocess(cv::Mat &frame){
for(int i=0; i < netRT->input_dim.c; i++){
int idx = i * imagePreproc.rows * imagePreproc.cols;
checkCuda( cudaMemcpy((void *)&input_d[idx], (void *)bgr[i].data, imagePreproc.rows * imagePreproc.cols* sizeof(float), cudaMemcpyDeviceToDevice) );
checkCuda( cudaMemcpy((void *)&input_d[idx + netRT->input_dim.tot()*bi], (void *)bgr[i].data, imagePreproc.rows * imagePreproc.cols* sizeof(float), cudaMemcpyDeviceToDevice) );
}
#else
//resize image, remove mean, divide by std
@@ -237,17 +238,17 @@ void MobilenetDetection::preprocess(cv::Mat &frame){
cv::split(imagePreproc, bgr);
for (int i = 0; i < netRT->input_dim.c; i++){
int idx = i * imagePreproc.rows * imagePreproc.cols;
memcpy((void *)&input[idx], (void *)bgr[i].data, imagePreproc.rows * imagePreproc.cols * sizeof(dnnType));
memcpy((void *)&input[idx + netRT->input_dim.tot()*bi], (void *)bgr[i].data, imagePreproc.rows * imagePreproc.cols * sizeof(dnnType));
}
checkCuda(cudaMemcpyAsync(input_d, input, netRT->input_dim.tot() * sizeof(dnnType), cudaMemcpyHostToDevice, netRT->stream));
checkCuda(cudaMemcpyAsync(input_d+ netRT->input_dim.tot()*bi, input + netRT->input_dim.tot()*bi, netRT->input_dim.tot() * sizeof(dnnType), cudaMemcpyHostToDevice, netRT->stream));
#endif
}
void MobilenetDetection::postprocess(const bool mAP){
void MobilenetDetection::postprocess(const int bi, const bool mAP){
//get confidences and locations_h
dnnType *rt_out[2];
rt_out[0] = (dnnType *)netRT->buffersRT[3];
rt_out[1] = (dnnType *)netRT->buffersRT[4];
rt_out[0] = (dnnType *)netRT->buffersRT[3]+ netRT->buffersDIM[3].tot()*bi;
rt_out[1] = (dnnType *)netRT->buffersRT[4]+ netRT->buffersDIM[4].tot()*bi;
detected.clear();
@@ -302,6 +303,7 @@ void MobilenetDetection::postprocess(const bool mAP){
boxes = remaining;
}
}
batchDetected.push_back(detected);
}
src/Yolo3Detection.cpp
+17 -12
View File
@@ -3,12 +3,16 @@
namespace tk { namespace dnn {
bool Yolo3Detection::init(const std::string& tensor_path, const int n_classes) {
bool Yolo3Detection::init(const std::string& tensor_path, const int n_classes, const int n_batches) {
//convert network to tensorRT
std::cout<<(tensor_path).c_str()<<"\n";
netRT = new tk::dnn::NetworkRT(NULL, (tensor_path).c_str() );
nBatches = n_batches;
tk::dnn::dataDim_t idim = netRT->input_dim;
idim.n = nBatches;
if(netRT->pluginFactory->n_yolos < 2 ) {
FatalError("this is not yolo3");
}
@@ -19,7 +23,7 @@ bool Yolo3Detection::init(const std::string& tensor_path, const int n_classes) {
num = yRT->num;
nMasks = yRT->n_masks;
// make a yolo layer for interpret predictions
// make a yolo layer to interpret predictions
yolo[i] = new tk::dnn::Yolo(nullptr, classes, nMasks, ""); // yolo without input and bias
yolo[i]->mask_h = new dnnType[nMasks];
yolo[i]->bias_h = new dnnType[num*nMasks*2];
@@ -31,9 +35,9 @@ bool Yolo3Detection::init(const std::string& tensor_path, const int n_classes) {
dets = tk::dnn::Yolo::allocateDetections(tk::dnn::Yolo::MAX_DETECTIONS, classes);
#ifndef OPENCV_CUDACONTRIB
checkCuda(cudaMallocHost(&input, sizeof(dnnType)*netRT->input_dim.tot()));
checkCuda(cudaMallocHost(&input, sizeof(dnnType)*idim.tot()));
#endif
checkCuda(cudaMalloc(&input_d, sizeof(dnnType)*netRT->input_dim.tot()));
checkCuda(cudaMalloc(&input_d, sizeof(dnnType)*idim.tot()));
// class colors precompute
for(int c=0; c<classes; c++) {
@@ -48,7 +52,7 @@ bool Yolo3Detection::init(const std::string& tensor_path, const int n_classes) {
return true;
}
void Yolo3Detection::preprocess(cv::Mat &frame){
void Yolo3Detection::preprocess(cv::Mat &frame, const int bi){
#ifdef OPENCV_CUDACONTRIB
cv::cuda::GpuMat orig_img, img_resized;
orig_img = cv::cuda::GpuMat(frame);
@@ -64,7 +68,7 @@ void Yolo3Detection::preprocess(cv::Mat &frame){
int size = imagePreproc.rows * imagePreproc.cols;
int ch = netRT->input_dim.c-1 -i;
bgr[ch].download(bgr_h); //TODO: don't copy back on CPU
checkCuda( cudaMemcpy(input_d + i*size, (float*)bgr_h.data, size*sizeof(dnnType), cudaMemcpyHostToDevice));
checkCuda( cudaMemcpy(input_d + i*size + netRT->input_dim.tot()*bi, (float*)bgr_h.data, size*sizeof(dnnType), cudaMemcpyHostToDevice));
}
#else
cv::resize(frame, frame, cv::Size(netRT->input_dim.w, netRT->input_dim.h));
@@ -77,18 +81,18 @@ void Yolo3Detection::preprocess(cv::Mat &frame){
for(int i=0; i<netRT->input_dim.c; i++) {
int idx = i*imagePreproc.rows*imagePreproc.cols;
int ch = netRT->input_dim.c-1 -i;
memcpy((void*)&input[idx], (void*)bgr[ch].data, imagePreproc.rows*imagePreproc.cols*sizeof(dnnType));
memcpy((void*)&input[idx + netRT->input_dim.tot()*bi], (void*)bgr[ch].data, imagePreproc.rows*imagePreproc.cols*sizeof(dnnType));
}
checkCuda(cudaMemcpyAsync(input_d, input, netRT->input_dim.tot()*sizeof(dnnType), cudaMemcpyHostToDevice, netRT->stream));
checkCuda(cudaMemcpyAsync(input_d + netRT->input_dim.tot()*bi, input + netRT->input_dim.tot()*bi, netRT->input_dim.tot()*sizeof(dnnType), cudaMemcpyHostToDevice, netRT->stream));
#endif
}
void Yolo3Detection::postprocess(const bool mAP){
void Yolo3Detection::postprocess(const int bi, const bool mAP){
//get yolo outputs
dnnType *rt_out[netRT->pluginFactory->n_yolos];
for(int i=0; i<netRT->pluginFactory->n_yolos; i++) {
rt_out[i] = (dnnType*)netRT->buffersRT[i+1];
}
for(int i=0; i<netRT->pluginFactory->n_yolos; i++)
rt_out[i] = (dnnType*)netRT->buffersRT[i+1] + netRT->buffersDIM[i+1].tot()*bi;
float x_ratio = float(originalSize.width) / float(netRT->input_dim.w);
float y_ratio = float(originalSize.height) / float(netRT->input_dim.h);
@@ -138,6 +142,7 @@ void Yolo3Detection::postprocess(const bool mAP){
detected.push_back(res);
}
}
batchDetected.push_back(detected);
}