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

yoloRT load anchors

Browse Source
This commit is contained in:
Francesco Gatti
2019-02-18 21:39:14 +01:00
parent bdd8e0bc26
commit 87fe342ca2
18 changed files with 164 additions and 207 deletions
Download Patch File Download Diff File Expand all files Collapse all files
demo/demo/demo.cpp
+1 -1
View File
@@ -24,7 +24,7 @@ int main(int argc, char *argv[]) {
signal(SIGINT, sig_handler);
tk::dnn::Yolo3Detection yolo;
yolo.init("yolo3_berkeley");
yolo.init("yolo3_berkeley.rt");
gRun = true;
include/Layer.h
+2 -2
View File
@@ -343,7 +343,7 @@ public:
float x, y, w, h;
};
typedef struct detection{
struct detection{
Yolo::box bbox;
int classes;
float *prob;
@@ -361,7 +361,7 @@ public:
dnnType *bias_h, *bias_d; //anchors
virtual dnnType* infer(dataDim_t &dim, dnnType* srcData);
int computeDetections(Yolo::detection *dets, int &ndets, int w, int h, int netw, int neth, float thresh);
int computeDetections(Yolo::detection *dets, int &ndets, int netw, int neth, float thresh);
dnnType *predictions;
include/Network.h
+1 -1
View File
@@ -15,7 +15,7 @@ namespace tk { namespace dnn {
*/
struct dataDim_t {
int n, c, h, w, l;
int n = 0, c = 0, h = 0, w = 0, l = 0;
dataDim_t() : n(1), c(1), h(1), w(1), l(1) {};
include/NetworkRT.h
+37 -14
View File
@@ -1,6 +1,7 @@
#ifndef NETWORKRT_H
#define NETWORKRT_H
#include <string.h> // memcpy
#include "utils.h"
#include "Network.h"
#include "Layer.h"
@@ -8,6 +9,40 @@
namespace tk { namespace dnn {
template<typename T> void writeBUF(char*& buffer, const T& val)
{
*reinterpret_cast<T*>(buffer) = val;
buffer += sizeof(T);
}
template<typename T> T readBUF(const char*& buffer)
{
T val = *reinterpret_cast<const T*>(buffer);
buffer += sizeof(T);
return val;
}
using namespace nvinfer1;
#include "pluginsRT/ActivationLeakyRT.h"
#include "pluginsRT/ReorgRT.h"
#include "pluginsRT/RegionRT.h"
//#include "pluginsRT/RouteRT.h"
#include "pluginsRT/ShortcutRT.h"
#include "pluginsRT/YoloRT.h"
#include "pluginsRT/UpsampleRT.h"
//#include "pluginsRT/Int8Calibrator.h"
class PluginFactory : IPluginFactory
{
public:
YoloRT *yolos[16];
int n_yolos;
virtual IPlugin* createPlugin(const char* layerName, const void* serialData, size_t serialLength);
};
class NetworkRT {
public:
@@ -27,6 +62,8 @@ public:
dnnType *output;
cudaStream_t stream;
PluginFactory *pluginFactory;
NetworkRT(Network *net, const char *name);
virtual ~NetworkRT();
@@ -53,19 +90,5 @@ public:
bool deserialize(const char *filename);
};
template<typename T> void writeBUF(char*& buffer, const T& val)
{
*reinterpret_cast<T*>(buffer) = val;
buffer += sizeof(T);
}
template<typename T> T readBUF(const char*& buffer)
{
T val = *reinterpret_cast<const T*>(buffer);
buffer += sizeof(T);
return val;
}
}}
#endif //NETWORKRT_H
include/Yolo3Detection.h
+2 -2
View File
@@ -31,8 +31,8 @@ class Yolo3Detection {
cv::Mat bgr[3];
public:
int classes = 10;
int num = 3;
int classes = 0;
int num = 0;
float thresh = 0.3;
cv::Scalar colors[256];
src/pluginsRT/ActivationLeakyRT.cpp → include/pluginsRT/ActivationLeakyRT.h
View File
src/pluginsRT/Int8Calibrator.cpp → include/pluginsRT/Int8Calibrator.h
View File
src/pluginsRT/RegionRT.cpp → include/pluginsRT/RegionRT.h
View File
src/pluginsRT/ReorgRT.cpp → include/pluginsRT/ReorgRT.h
View File
src/pluginsRT/RouteRT.cpp → include/pluginsRT/RouteRT.h
View File
src/pluginsRT/ShortcutRT.cpp → include/pluginsRT/ShortcutRT.h
View File
src/pluginsRT/UpsampleRT.cpp → include/pluginsRT/UpsampleRT.h
View File
src/pluginsRT/YoloRT.cpp → include/pluginsRT/YoloRT.h
View File
src/NetworkRT.cpp
+80 -100
View File
@@ -11,14 +11,6 @@
#include "NetworkRT.h"
using namespace nvinfer1;
#include "pluginsRT/ActivationLeakyRT.cpp"
#include "pluginsRT/ReorgRT.cpp"
#include "pluginsRT/RegionRT.cpp"
//#include "pluginsRT/RouteRT.cpp"
#include "pluginsRT/ShortcutRT.cpp"
#include "pluginsRT/YoloRT.cpp"
#include "pluginsRT/UpsampleRT.cpp"
#include "pluginsRT/Int8Calibrator.cpp"
// Logger for info/warning/errors
class Logger : public ILogger {
@@ -54,14 +46,7 @@ NetworkRT::NetworkRT(Network *net, const char *name) {
builderRT->setMaxBatchSize(1);
builderRT->setMaxWorkspaceSize(1 << 30);
//change datatype based on system specs
if(builderRT->platformHasFastInt8()) {
BatchStream bstream({32,dim.c, dim.h, dim.w}, 32, 1);
Int8EntropyCalibrator calib(bstream, 0, false);
builderRT->setInt8Mode(true);
builderRT->setInt8Calibrator(&calib);
} else if(net->fp16 && builderRT->platformHasFastFp16()) {
if(net->fp16 && builderRT->platformHasFastFp16()) {
dtRT = DataType::kHALF;
builderRT->setHalf2Mode(true);
}
@@ -393,87 +378,6 @@ bool NetworkRT::serialize(const char *filename) {
return true;
}
class PluginFactory : IPluginFactory
{
public:
virtual IPlugin* createPlugin(const char* layerName, const void* serialData, size_t serialLength) {
const char * buf = reinterpret_cast<const char*>(serialData);
std::string name(layerName);
if(name.find("Activation") == 0) {
ActivationLeakyRT *a = new ActivationLeakyRT();
a->size = readBUF<int>(buf);
return a;
}
if(name.find("Region") == 0) {
RegionRT *r = new RegionRT(readBUF<int>(buf), //classes
readBUF<int>(buf), //coords
readBUF<int>(buf)); //num
r->c = readBUF<int>(buf);
r->h = readBUF<int>(buf);
r->w = readBUF<int>(buf);
return r;
}
if(name.find("Reorg") == 0) {
ReorgRT *r = new ReorgRT(readBUF<int>(buf)); //stride
r->c = readBUF<int>(buf);
r->h = readBUF<int>(buf);
r->w = readBUF<int>(buf);
return r;
}
if(name.find("Shortcut") == 0) {
ShortcutRT *r = new ShortcutRT();
r->c = readBUF<int>(buf);
r->h = readBUF<int>(buf);
r->w = readBUF<int>(buf);
return r;
}
if(name.find("Yolo") == 0) {
YoloRT *r = new YoloRT(readBUF<int>(buf), //classes
readBUF<int>(buf)); //num
r->c = readBUF<int>(buf);
r->h = readBUF<int>(buf);
r->w = readBUF<int>(buf);
for(int i=0; i<r->num; i++)
r->mask[i] = readBUF<dnnType>(buf);
for(int i=0; i<3*2*r->num; i++)
r->bias[i] = readBUF<dnnType>(buf);
std::cout<<"YOLO: "<<r->c<<" "<<r->h<<" "<<r->w<<"\n";
return r;
}
if(name.find("Upsample") == 0) {
UpsampleRT *r = new UpsampleRT(readBUF<int>(buf)); //stride
r->c = readBUF<int>(buf);
r->h = readBUF<int>(buf);
r->w = readBUF<int>(buf);
return r;
}
/*
if(name.find("Route") == 0) {
RouteRT *r = new RouteRT();
r->in = readBUF<int>(buf);
for(int i=0; i<RouteRT::MAX_INPUTS; i++)
r->c_in[i] = readBUF<int>(buf);
r->c = readBUF<int>(buf);
r->h = readBUF<int>(buf);
r->w = readBUF<int>(buf);
return r;
}
*/
FatalError("Cant deserialize Plugin");
return NULL;
}
};
bool NetworkRT::deserialize(const char *filename) {
char *gieModelStream{nullptr};
@@ -488,13 +392,89 @@ bool NetworkRT::deserialize(const char *filename) {
file.close();
}
PluginFactory plfact;
runtimeRT = createInferRuntime(loggerRT);
engineRT = runtimeRT->deserializeCudaEngine(gieModelStream, size, (IPluginFactory *) &plfact);
engineRT = runtimeRT->deserializeCudaEngine(gieModelStream, size, (IPluginFactory *) pluginFactory);
//if (gieModelStream) delete [] gieModelStream;
return true;
}
IPlugin* PluginFactory::createPlugin(const char* layerName, const void* serialData, size_t serialLength) {
const char * buf = reinterpret_cast<const char*>(serialData);
std::string name(layerName);
if(name.find("Activation") == 0) {
ActivationLeakyRT *a = new ActivationLeakyRT();
a->size = readBUF<int>(buf);
return a;
}
if(name.find("Region") == 0) {
RegionRT *r = new RegionRT(readBUF<int>(buf), //classes
readBUF<int>(buf), //coords
readBUF<int>(buf)); //num
r->c = readBUF<int>(buf);
r->h = readBUF<int>(buf);
r->w = readBUF<int>(buf);
return r;
}
if(name.find("Reorg") == 0) {
ReorgRT *r = new ReorgRT(readBUF<int>(buf)); //stride
r->c = readBUF<int>(buf);
r->h = readBUF<int>(buf);
r->w = readBUF<int>(buf);
return r;
}
if(name.find("Shortcut") == 0) {
ShortcutRT *r = new ShortcutRT();
r->c = readBUF<int>(buf);
r->h = readBUF<int>(buf);
r->w = readBUF<int>(buf);
return r;
}
if(name.find("Yolo") == 0) {
YoloRT *r = new YoloRT(readBUF<int>(buf), //classes
readBUF<int>(buf)); //num
r->c = readBUF<int>(buf);
r->h = readBUF<int>(buf);
r->w = readBUF<int>(buf);
for(int i=0; i<r->num; i++)
r->mask[i] = readBUF<dnnType>(buf);
for(int i=0; i<3*2*r->num; i++)
r->bias[i] = readBUF<dnnType>(buf);
yolos[n_yolos++] = r;
return r;
}
if(name.find("Upsample") == 0) {
UpsampleRT *r = new UpsampleRT(readBUF<int>(buf)); //stride
r->c = readBUF<int>(buf);
r->h = readBUF<int>(buf);
r->w = readBUF<int>(buf);
return r;
}
/*
if(name.find("Route") == 0) {
RouteRT *r = new RouteRT();
r->in = readBUF<int>(buf);
for(int i=0; i<RouteRT::MAX_INPUTS; i++)
r->c_in[i] = readBUF<int>(buf);
r->c = readBUF<int>(buf);
r->h = readBUF<int>(buf);
r->w = readBUF<int>(buf);
return r;
}
*/
FatalError("Cant deserialize Plugin");
return NULL;
}
}}
src/Yolo.cpp
+7 -44
View File
@@ -18,13 +18,12 @@ Yolo::Yolo(Network *net, int classes, int num, const char* fname_weights) :
this->num = num;
// load anchors
int seek = 0;
readBinaryFile(fname_weights, num, &mask_h, &mask_d, seek);
seek += num;
readBinaryFile(fname_weights, 3*num*2, &bias_h, &bias_d, seek);
printDeviceVector(num, mask_h, false);
printDeviceVector(3*num*2, bias_h, false);
if(fname_weights != nullptr) {
int seek = 0;
readBinaryFile(fname_weights, num, &mask_h, &mask_d, seek);
seek += num;
readBinaryFile(fname_weights, 3*num*2, &bias_h, &bias_d, seek);
}
// same
output_dim.n = input_dim.n;
@@ -33,10 +32,6 @@ Yolo::Yolo(Network *net, int classes, int num, const char* fname_weights) :
output_dim.w = input_dim.w;
output_dim.l = input_dim.l;
std::cout<<"YOLO INPUT: ";
input_dim.print();
std::cout<<"\n";
checkCuda( cudaMalloc(&dstData, output_dim.tot()*sizeof(dnnType)) );
predictions = nullptr;
}
@@ -62,35 +57,6 @@ Yolo::box get_yolo_box(float *x, float *biases, int n, int index, int i, int j,
return b;
}
void correct_yolo_boxes(Yolo::detection *dets, int n, int w, int h, int netw, int neth, int relative)
{
int i;
int new_w=0;
int new_h=0;
if (((float)netw/w) < ((float)neth/h)) {
new_w = netw;
new_h = (h * netw)/w;
} else {
new_h = neth;
new_w = (w * neth)/h;
}
for (i = 0; i < n; ++i){
Yolo::box b = dets[i].bbox;
b.x = (b.x - (netw - new_w)/2./netw) / ((float)new_w/netw);
b.y = (b.y - (neth - new_h)/2./neth) / ((float)new_h/neth);
b.w *= (float)netw/new_w;
b.h *= (float)neth/new_h;
if(!relative){
b.x *= w;
b.w *= w;
b.y *= h;
b.h *= h;
}
dets[i].bbox = b;
}
}
dnnType* Yolo::infer(dataDim_t &dim, dnnType* srcData) {
checkCuda( cudaMemcpy(dstData, srcData, dim.tot()*sizeof(dnnType), cudaMemcpyDeviceToDevice));
@@ -109,14 +75,12 @@ dnnType* Yolo::infer(dataDim_t &dim, dnnType* srcData) {
return dstData;
}
int Yolo::computeDetections(Yolo::detection *dets, int &ndets, int w, int h, int netw, int neth, float thresh) {
int Yolo::computeDetections(Yolo::detection *dets, int &ndets, int netw, int neth, float thresh) {
if(predictions == nullptr)
predictions = new dnnType[output_dim.tot()];
checkCuda( cudaMemcpy(predictions, dstData, output_dim.tot()*sizeof(dnnType), cudaMemcpyDeviceToHost));
int relative = 0;
int lw = output_dim.w;
int lh = output_dim.h;
@@ -150,7 +114,6 @@ int Yolo::computeDetections(Yolo::detection *dets, int &ndets, int w, int h, int
}
}
correct_yolo_boxes(dets + ndets, count, w, h, netw, neth, relative);
ndets = count;
return count;
}
src/Yolo3Detection.cpp
+28 -19
View File
@@ -2,10 +2,36 @@
namespace tk { namespace dnn {
bool Yolo3Detection::init(std::string tensor_folder) {
bool Yolo3Detection::init(std::string tensor_path) {
//const char *tensor_path = "../data/yolo3/yolo3_berkeley.rt";
//convert network to tensorRT
std::cout<<(tensor_path).c_str()<<"\n";
netRT = new tk::dnn::NetworkRT(NULL, (tensor_path).c_str() );
if(netRT->pluginFactory->n_yolos != 3) {
FatalError("this is not yolo3");
}
for(int i=0; i<netRT->pluginFactory->n_yolos; i++) {
YoloRT *yRT = netRT->pluginFactory->yolos[i];
classes = yRT->classes;
num = yRT->num;
// make a yolo layer for interpret predictions
yolo[i] = new tk::dnn::Yolo(nullptr, classes, num, nullptr); // yolo without input and bias
memcpy(yolo[i]->mask_h, yRT->mask, sizeof(dnnType)*num);
memcpy(yolo[i]->bias_h, yRT->bias, sizeof(dnnType)*num*3*2);
yolo[i]->input_dim = yolo[i]->output_dim = tk::dnn::dataDim_t(1, yRT->c, yRT->h, yRT->w);
}
dets = tk::dnn::Yolo::allocateDetections(tk::dnn::Yolo::MAX_DETECTIONS, classes);
checkCuda(cudaMallocHost(&input, sizeof(dnnType)*netRT->input_dim.tot()));
checkCuda(cudaMalloc(&input_d, sizeof(dnnType)*netRT->input_dim.tot()));
// class colors precompute
for(int c=0; c<classes; c++) {
int cc = c+1;
@@ -19,23 +45,6 @@ bool Yolo3Detection::init(std::string tensor_folder) {
//std::cout<<r<<" "<<g<<" "<<b<<"\n";
colors[c] = cv::Scalar(int(255.0*b), int(255.0*g), int(255.0*r));
}
//convert network to tensorRT
std::cout<<(tensor_folder + ".rt").c_str()<<"\n";
netRT = new tk::dnn::NetworkRT(NULL, (tensor_folder + ".rt").c_str() );
yolo[0] = new tk::dnn::Yolo(nullptr, classes, num, (tensor_folder + "_0.bin").c_str() ); // yolo without input and bias
yolo[0]->input_dim = yolo[0]->output_dim = tk::dnn::dataDim_t(1, 45, 10, 17);
yolo[1] = new tk::dnn::Yolo(nullptr, classes, num, (tensor_folder + "_1.bin").c_str() ); // yolo without input and bias
yolo[1]->input_dim = yolo[1]->output_dim = tk::dnn::dataDim_t(1, 45, 20, 34);
yolo[2] = new tk::dnn::Yolo(nullptr, classes, num, (tensor_folder + "_2.bin").c_str() ); // yolo without input and bias
yolo[2]->input_dim = yolo[2]->output_dim = tk::dnn::dataDim_t(1, 45, 40, 68);
dets = tk::dnn::Yolo::allocateDetections(tk::dnn::Yolo::MAX_DETECTIONS, classes);
checkCuda(cudaMallocHost(&input, sizeof(dnnType)*netRT->input_dim.tot()));
checkCuda(cudaMalloc(&input_d, sizeof(dnnType)*netRT->input_dim.tot()));
return true;
}
@@ -82,7 +91,7 @@ void Yolo3Detection::update(cv::Mat &imageORIG) {
for(int i=0; i<3; i++) {
rt_out[i] = (dnnType*)netRT->buffersRT[i+1];
yolo[i]->dstData = rt_out[i];
yolo[i]->computeDetections(dets, ndets, netRT->input_dim.w, netRT->input_dim.h, netRT->input_dim.w, netRT->input_dim.h, thresh);
yolo[i]->computeDetections(dets, ndets, netRT->input_dim.w, netRT->input_dim.h, thresh);
}
tk::dnn::Yolo::mergeDetections(dets, ndets, classes);
TIMER_STOP
tests/yolo3_berkeley/yolo3_berkeley.cpp
+3 -12
View File
@@ -326,9 +326,9 @@ int main() {
int ndets = 0;
int classes = yolo0.classes;
tk::dnn::Yolo::detection *dets = tk::dnn::Yolo::allocateDetections(tk::dnn::Yolo::MAX_DETECTIONS, classes);
yolo0.computeDetections(dets, ndets, net.input_dim.w, net.input_dim.h, net.input_dim.w, net.input_dim.h, 0.5);
yolo1.computeDetections(dets, ndets, net.input_dim.w, net.input_dim.h, net.input_dim.w, net.input_dim.h, 0.5);
yolo2.computeDetections(dets, ndets, net.input_dim.w, net.input_dim.h, net.input_dim.w, net.input_dim.h, 0.5);
yolo0.computeDetections(dets, ndets, net.input_dim.w, net.input_dim.h, 0.5);
yolo1.computeDetections(dets, ndets, net.input_dim.w, net.input_dim.h, 0.5);
yolo2.computeDetections(dets, ndets, net.input_dim.w, net.input_dim.h, 0.5);
tk::dnn::Yolo::mergeDetections(dets, ndets, classes);
for(int j=0; j<ndets; j++) {
@@ -369,14 +369,5 @@ int main() {
std::cout<<"TRT vs correct"; checkResult(odim, rt_out[i], out);
std::cout<<"CUDNN vs TRT "; checkResult(odim, cudnn_out[i], rt_out[i]);
}
std::cout<<"copyng layer config to this folder\n";
std::string cmd;
cmd = "cp " + std::string(g82_bin) + " yolo3_berkeley_0.bin";
std::cout<<cmd<<"\n"; system(cmd.c_str());
cmd = "cp " + std::string(g94_bin) + " yolo3_berkeley_1.bin";
std::cout<<cmd<<"\n"; system(cmd.c_str());
cmd = "cp " + std::string(g106_bin) + " yolo3_berkeley_2.bin";
std::cout<<cmd<<"\n"; system(cmd.c_str());
return 0;
}
tests/yolo3_voc/yolo3_voc.cpp
+3 -12
View File
@@ -326,9 +326,9 @@ int main() {
int ndets = 0;
int classes = yolo0.classes;
tk::dnn::Yolo::detection *dets = tk::dnn::Yolo::allocateDetections(tk::dnn::Yolo::MAX_DETECTIONS, classes);
yolo0.computeDetections(dets, ndets, net.input_dim.w, net.input_dim.h, net.input_dim.w, net.input_dim.h, 0.5);
yolo1.computeDetections(dets, ndets, net.input_dim.w, net.input_dim.h, net.input_dim.w, net.input_dim.h, 0.5);
yolo2.computeDetections(dets, ndets, net.input_dim.w, net.input_dim.h, net.input_dim.w, net.input_dim.h, 0.5);
yolo0.computeDetections(dets, ndets, net.input_dim.w, net.input_dim.h, 0.5);
yolo1.computeDetections(dets, ndets, net.input_dim.w, net.input_dim.h, 0.5);
yolo2.computeDetections(dets, ndets, net.input_dim.w, net.input_dim.h, 0.5);
tk::dnn::Yolo::mergeDetections(dets, ndets, classes);
for(int j=0; j<ndets; j++) {
@@ -369,14 +369,5 @@ int main() {
std::cout<<"TRT vs correct"; checkResult(odim, rt_out[i], out);
std::cout<<"CUDNN vs TRT "; checkResult(odim, cudnn_out[i], rt_out[i]);
}
std::cout<<"copyng layer config to this folder\n";
std::string cmd;
cmd = "cp " + std::string(g82_bin) + " yolo3_voc_0.bin";
std::cout<<cmd<<"\n"; system(cmd.c_str());
cmd = "cp " + std::string(g94_bin) + " yolo3_voc_1.bin";
std::cout<<cmd<<"\n"; system(cmd.c_str());
cmd = "cp " + std::string(g106_bin) + " yolo3_voc_2.bin";
std::cout<<cmd<<"\n"; system(cmd.c_str());
return 0;
}