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

Yolo3_tiny CUDNN works, TensorRT doesn't. Add n_masks to Yolo layer.

Browse Source 
Signed-off-by: Micaela Verucchi <micaelaverucchi@gmail.com>
This commit is contained in:
Micaela Verucchi
2019-12-05 20:43:48 +00:00
parent a2db98670a
commit 44b2bce3ff
5 changed files with 74 additions and 39 deletions
Download Patch File Download Diff File Expand all files Collapse all files
include/tkDNN/Layer.h
+2 -2
View File
@@ -379,11 +379,11 @@ public:
int sort_class;
};
Yolo(Network *net, int classes, int num, std::string fname_weights);
Yolo(Network *net, int classes, int num, std::string fname_weights, int n_masks=3);
virtual ~Yolo();
virtual layerType_t getLayerType() { return LAYER_YOLO; };
int classes, num;
int classes, num, n_masks;
dnnType *mask_h, *mask_d; //anchors
dnnType *bias_h, *bias_d; //anchors
std::vector<std::string> classesNames;
include/tkDNN/pluginsRT/YoloRT.h
+12 -10
View File
@@ -8,16 +8,17 @@ class YoloRT : public IPlugin {
public:
YoloRT(int classes, int num, tk::dnn::Yolo *yolo = nullptr) {
YoloRT(int classes, int num, tk::dnn::Yolo *yolo = nullptr, int n_masks=3) {
this->classes = classes;
this->num = num;
this->n_masks = n_masks;
mask = new dnnType[num];
bias = new dnnType[num*3*2];
mask = new dnnType[n_masks];
bias = new dnnType[num*n_masks*2];
if(yolo != nullptr) {
memcpy(mask, yolo->mask_h, sizeof(dnnType)*num);
memcpy(bias, yolo->bias_h, sizeof(dnnType)*num*3*2);
memcpy(mask, yolo->mask_h, sizeof(dnnType)*n_masks);
memcpy(bias, yolo->bias_h, sizeof(dnnType)*num*n_masks*2);
classesNames = yolo->classesNames;
}
}
@@ -60,7 +61,7 @@ public:
checkCuda( cudaMemcpyAsync(dstData, srcData, batchSize*c*h*w*sizeof(dnnType), cudaMemcpyDeviceToDevice, stream));
for (int b = 0; b < batchSize; ++b){
for(int n = 0; n < num; ++n){
for(int n = 0; n < n_masks; ++n){
int index = entry_index(b, n*w*h, 0, batchSize);
activationLOGISTICForward(srcData + index, dstData + index, 2*w*h, stream);
@@ -75,19 +76,20 @@ public:
virtual size_t getSerializationSize() override {
return 5*sizeof(int) + num*sizeof(dnnType) + num*3*2*sizeof(dnnType) + YOLORT_CLASSNAME_W*classes*sizeof(char);
return 6*sizeof(int) + n_masks*sizeof(dnnType) + num*n_masks*2*sizeof(dnnType) + YOLORT_CLASSNAME_W*classes*sizeof(char);
}
virtual void serialize(void* buffer) override {
char *buf = reinterpret_cast<char*>(buffer);
tk::dnn::writeBUF(buf, classes);
tk::dnn::writeBUF(buf, num);
tk::dnn::writeBUF(buf, n_masks);
tk::dnn::writeBUF(buf, c);
tk::dnn::writeBUF(buf, h);
tk::dnn::writeBUF(buf, w);
for(int i=0; i<num; i++)
for(int i=0; i<n_masks; i++)
tk::dnn::writeBUF(buf, mask[i]);
for(int i=0; i<3*2*num; i++)
for(int i=0; i<n_masks*2*num; i++)
tk::dnn::writeBUF(buf, bias[i]);
// save classes names
@@ -101,7 +103,7 @@ public:
}
int c, h, w;
int classes, num;
int classes, num, n_masks;
std::vector<std::string> classesNames;
dnnType *mask;
src/NetworkRT.cpp
+41 -9
View File
@@ -15,9 +15,9 @@ using namespace nvinfer1;
// Logger for info/warning/errors
class Logger : public ILogger {
void log(Severity severity, const char* msg) override {
#ifdef DEBUG
// #ifdef DEBUG
std::cout <<"TENSORRT LOG: "<< msg << std::endl;
#endif
// #endif
}
} loggerRT;
@@ -253,14 +253,18 @@ ILayer* NetworkRT::convert_layer(ITensor *input, Conv2d *l) {
lRTconv->setStride(DimsHW{l->strideH, l->strideW});
lRTconv->setPadding(DimsHW{l->paddingH, l->paddingW});
lRT = (ILayer*) lRTconv;
}
checkNULL(lRT);
if(l->batchnorm) {
Weights power{dtRT, power_b, l->outputs};
Weights shift{dtRT, mean_b, l->outputs};
Weights scale{dtRT, variance_b, l->outputs};
std::cout<<lRT->getNbOutputs()<<std::endl;
IScaleLayer *lRT2 = networkRT->addScale(*lRT->getOutput(0), ScaleMode::kCHANNEL,
shift, scale, power);
checkNULL(lRT2);
Weights shift2{dtRT, bias_b, l->outputs};
@@ -277,7 +281,7 @@ ILayer* NetworkRT::convert_layer(ITensor *input, Conv2d *l) {
ILayer* NetworkRT::convert_layer(ITensor *input, Pooling *l) {
std::cout<<"convert Pooling\n";
// printf("%d %d\n", l->winW, l->winH);
printf("%d %d %d %d %d %d %d %d %d %d %d %d (layer)\n", l->input_dim.h, l->input_dim.w, l->output_dim.h, l->output_dim.w, l->winW, l->winH, l->strideH, l->strideW, l->paddingH, l->paddingW, l->pool_mode, tkdnnPoolingMode_t::POOLING_MAX) ;
PoolingType ptype;
if(l->pool_mode == tkdnnPoolingMode_t::POOLING_MAX) ptype = PoolingType::kMAX;
@@ -287,8 +291,27 @@ ILayer* NetworkRT::convert_layer(ITensor *input, Pooling *l) {
IPoolingLayer *lRT = networkRT->addPooling(*input,
ptype, DimsHW{l->winH, l->winW});
checkNULL(lRT);
lRT->setStride(DimsHW{l->strideH, l->strideW});
lRT->setPadding(DimsHW{l->paddingH, l->paddingW});
// if (l->input_dim.h == 13 && l->output_dim.h == 13)
// {
// lRT->setPadding(DimsHW{7, 7});
// lRT->setStride(DimsHW{2, 2});
// }
// else
// {
lRT->setPadding(DimsHW{l->paddingH, l->paddingW});
lRT->setStride(DimsHW{l->strideH, l->strideW});
// }
// IResizeLayer *lRT = networkRT->addResize(*lRT->getOutput(0));
// checkNULL(lRT);
// lRT->setOutputDimensions(l->output_dim);
ITensor *t = lRT->getOutput(0);
for(int j=0; j<t->getDimensions().nbDims; j++) {
std::cout<<t->getDimensions().d[j]<<" ";
}
std::cout<<" (TensorRT)\n";
return lRT;
}
@@ -324,12 +347,19 @@ ILayer* NetworkRT::convert_layer(ITensor *input, Softmax *l) {
}
ILayer* NetworkRT::convert_layer(ITensor *input, Route *l) {
//std::cout<<"convert route\n";
std::cout<<"convert route\n";
ITensor **tens = new ITensor*[l->layers_n];
for(int i=0; i<l->layers_n; i++) {
tens[i] = tensors[l->layers[i]];
for(int j=0; j<tens[i]->getDimensions().nbDims; j++) {
std::cout<<tens[i]->getDimensions().d[j]<<" ";
}
std::cout<<"\n";
}
IConcatenationLayer *lRT = networkRT->addConcatenation(tens, l->layers_n);
//IPlugin *plugin = new RouteRT();
//IPluginLayer *lRT = networkRT->addPlugin(tens, l->layers_n, *plugin);
@@ -474,13 +504,15 @@ IPlugin* PluginFactory::createPlugin(const char* layerName, const void* serialDa
if(name.find("Yolo") == 0) {
YoloRT *r = new YoloRT(readBUF<int>(buf), //classes
readBUF<int>(buf)); //num
readBUF<int>(buf), //num
nullptr,
readBUF<int>(buf)); //n_masks
r->c = readBUF<int>(buf);
r->h = readBUF<int>(buf);
r->w = readBUF<int>(buf);
for(int i=0; i<r->num; i++)
for(int i=0; i<r->n_masks; i++)
r->mask[i] = readBUF<dnnType>(buf);
for(int i=0; i<3*2*r->num; i++)
for(int i=0; i<r->n_masks*2*r->num; i++)
r->bias[i] = readBUF<dnnType>(buf);
// save classes names
src/Yolo.cpp
+8 -7
View File
@@ -11,19 +11,20 @@
namespace tk { namespace dnn {
Yolo::Yolo(Network *net, int classes, int num, std::string fname_weights) :
Yolo::Yolo(Network *net, int classes, int num, std::string fname_weights, int n_masks) :
Layer(net) {
this->classes = classes;
this->num = 3;
this->num = num;
this->n_masks = n_masks;
// load anchors
if(fname_weights != "") {
int seek = 0;
readBinaryFile(fname_weights, 3, &mask_h, &mask_d, seek);
seek += 3;
readBinaryFile(fname_weights, 3*num*2, &bias_h, &bias_d, seek);
for(int i=0; i<3*num*2; i++)
readBinaryFile(fname_weights, n_masks, &mask_h, &mask_d, seek);
seek += n_masks;
readBinaryFile(fname_weights, n_masks*num*2, &bias_h, &bias_d, seek);
for(int i=0; i<n_masks*num*2; i++)
printf("%f\n", bias_h[i]);
}
@@ -70,7 +71,7 @@ dnnType* Yolo::infer(dataDim_t &dim, dnnType* srcData) {
checkCuda( cudaMemcpy(dstData, srcData, dim.tot()*sizeof(dnnType), cudaMemcpyDeviceToDevice));
for (int b = 0; b < dim.n; ++b){
for(int n = 0; n < num; ++n){
for(int n = 0; n < n_masks; ++n){
int index = entry_index(b, n*dim.w*dim.h, 0, classes, input_dim, output_dim);
activationLOGISTICForward(srcData + index, dstData + index, 2*dim.w*dim.h);
tests/yolo3_tiny/yolo3_tiny.cpp
+11 -11
View File
@@ -52,7 +52,7 @@ int main() {
tk::dnn::Conv2d c10(&net, 512, 3, 3, 1, 1, 1, 1, c10_bin, true);
tk::dnn::Activation a10(&net, tk::dnn::ACTIVATION_LEAKY);
tk::dnn::Pooling p11(&net, 2, 2, 1, 1, tk::dnn::POOLING_MAX);
tk::dnn::Pooling p11(&net, 2, 2, 1, 1,0,0, tk::dnn::POOLING_MAX);
tk::dnn::Conv2d c12(&net, 1024, 3, 3, 1, 1, 1, 1, c12_bin, true);
tk::dnn::Activation a12(&net, tk::dnn::ACTIVATION_LEAKY);
@@ -102,21 +102,21 @@ int main() {
dim1.print();
}
// tk::dnn::dataDim_t dim2 = dim;
// printCenteredTitle(" TENSORRT inference ", '=', 30); {
// dim2.print();
// TIMER_START
// out_data2 = netRT.infer(dim2, data);
// TIMER_STOP
// dim2.print();
// }
tk::dnn::dataDim_t dim2 = dim;
printCenteredTitle(" TENSORRT inference ", '=', 30); {
dim2.print();
TIMER_START
out_data2 = netRT.infer(dim2, data);
TIMER_STOP
dim2.print();
}
printCenteredTitle(" CHECK RESULTS ", '=', 30);
dnnType *out, *out_h;
int out_dim = net.getOutputDim().tot();
readBinaryFile(output_bin, out_dim, &out_h, &out);
std::cout<<"CUDNN vs correct"; checkResult(out_dim, out_data, out);
// std::cout<<"TRT vs correct"; checkResult(out_dim, out_data2, out);
// std::cout<<"CUDNN vs TRT "; checkResult(out_dim, out_data, out_data2);
std::cout<<"TRT vs correct"; checkResult(out_dim, out_data2, out);
std::cout<<"CUDNN vs TRT "; checkResult(out_dim, out_data, out_data2);
return 0;
}