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

Merge branch 'master' into tensorrt8

Browse Source
This commit is contained in:
Francesco Gatti
2022-03-30 16:41:45 +02:00
parent 69bb7370a5 7c0620e391
commit df31375b67
16 changed files with 1719 additions and 98 deletions
Download Patch File Download Diff File Expand all files Collapse all files
CMakeLists.txt
+9
View File
@@ -5,6 +5,10 @@ set(CMAKE_CXX_STANDARD 14)
option(ENABLE_OPENCV_CUDA_CONTRIB "Enable OpenCV CUDA Contrib" OFF )
if(NOT CMAKE_BUILD_TYPE)
set(CMAKE_BUILD_TYPE "Release" CACHE STRING "default build" FORCE)
endif(NOT CMAKE_BUILD_TYPE)
find_package(CUDA 9.0 REQUIRED)
if (CUDA_FOUND)
set(OUTPUTFILE ${CMAKE_CURRENT_SOURCE_DIR}/cmake/cuda_script) # No suffix required
@@ -202,12 +206,17 @@ target_link_libraries(test_shelfnet_berkeley tkDNN)
add_executable(test_shelfnet_mapillary tests/shelfnet/shelfnet_mapillary.cpp)
target_link_libraries(test_shelfnet_mapillary tkDNN)
add_executable(test_shelfnet_coco tests/shelfnet/shelfnet_coco.cpp)
target_link_libraries(test_shelfnet_coco tkDNN)
# MONODEPTH2
add_executable(test_monodepth2_640 tests/monodepth2/monodepth2_640.cpp)
target_link_libraries(test_monodepth2_640 tkDNN)
add_executable(test_monodepth2_1024 tests/monodepth2/monodepth2_1024.cpp)
target_link_libraries(test_monodepth2_1024 tkDNN)
# DEMOS
add_executable(test_rtinference tests/test_rtinference/rtinference.cpp)
target_link_libraries(test_rtinference tkDNN)
README.md
+7 -9
View File
@@ -23,12 +23,11 @@ If you use tkDNN in your research, please cite the [following paper](https://iee
- [x] Support 2D/3D Object Detection and Tracking [README](docs/README_2d3dtracking.md)
#### 24 November 2021
- [x] Support to sematic segmentation on cuda 11
- [x] Support to TensorRT8 (tensort8 branch).
- [x] Support to TensorRT8. (thanks to [Harshvardhan Chandirasekar](https://github.com/perseusdg))
#### 30 March 2022
- [x] Support to monocular depth esitmation (tensort8 branch) [README](docs/README_depth.md)
- [x] Support to monocular depth esitmation [README](docs/README_depth.md) (thanks to [Harshvardhan Chandirasekar](https://github.com/perseusdg))
TensorRT8 (and therefore Jetpack 4.6) is currently supported only on the branch tensort8 due to [performance issue with TensorRT8](https://docs.nvidia.com/deeplearning/tensorrt/release-notes/tensorrt-8.html)). We will merge it to the master as soon as those issues are fixed (probably in future minor releases).
## FPS Results
Inference FPS of yolov4 with tkDNN, average of 1200 images with the same dimension as the input size, on
@@ -120,8 +119,8 @@ git clone https://github.com/ceccocats/tkDNN
cd tkDNN
mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release -G"Ninja" ..
ninja
cmake -DCMAKE_BUILD_TYPE=Release ..
make
```
## Workflow
@@ -179,11 +178,10 @@ For specific details on how to run tkDNN on Windows 10/11 see [HERE](./docs/wind
| yolo4_320 | Yolov4 <sup>8</sup> | [COCO 2017](http://cocodataset.org/) | 80 | 320x320 | [weights](https://cloud.hipert.unimore.it/s/d97CFzYqCPCp5Hg/download) |
| yolo4_512 | Yolov4 <sup>8</sup> | [COCO 2017](http://cocodataset.org/) | 80 | 512x512 | [weights](https://cloud.hipert.unimore.it/s/d97CFzYqCPCp5Hg/download) |
| yolo4_608 | Yolov4 <sup>8</sup> | [COCO 2017](http://cocodataset.org/) | 80 | 608x608 | [weights](https://cloud.hipert.unimore.it/s/d97CFzYqCPCp5Hg/download) |
| yolo4_berkeley | Yolov4 <sup>8</sup> | [BDD100K ](https://bair.berkeley.edu/blog/2018/05/30/bdd/) | 10 | 540x320 | [weights](https://cloud.hipert.unimore.it/s/nkWFa5fgb4NTdnB/download) |
| yolo4_berkeley | Yolov4 <sup>8</sup> | [BDD100K ](https://bair.berkeley.edu/blog/2018/05/30/bdd/) | 10 | 544x320 | [weights](https://cloud.hipert.unimore.it/s/nkWFa5fgb4NTdnB/download) |
| yolo4tiny | Yolov4 tiny <sup>9</sup> | [COCO 2017](http://cocodataset.org/) | 80 | 416x416 | [weights](https://cloud.hipert.unimore.it/s/iRnc4pSqmx78gJs/download) |
| yolo4x | Yolov4x-mish <sup>9</sup> | [COCO 2017](http://cocodataset.org/) |
| yolo4x | Yolov4x-mish <sup>9</sup> | [COCO 2017](http://cocodataset.org/) | 80 | 640x640 | [weights](https://cloud.hipert.unimore.it/s/5MFjtNtgbDGdJEo/download) |
| yolo4tiny_512 | Yolov4 tiny <sup>9</sup> | [COCO 2017](http://cocodataset.org/) | 80 | 512x512 | [weights](https://cloud.hipert.unimore.it/s/iRnc4pSqmx78gJs/download) |
80 | 640x640 | [weights](https://cloud.hipert.unimore.it/s/5MFjtNtgbDGdJEo/download) |
| yolo4x-cps | Scaled Yolov4 <sup>10</sup> | [COCO 2017](http://cocodataset.org/) | 80 | 512x512 | [weights](https://cloud.hipert.unimore.it/s/AfzHE4BfTeEm2gH/download) |
| shelfnet | ShelfNet18_realtime<sup>11</sup> | [Cityscapes](https://www.cityscapes-dataset.com/) | 19 | 1024x1024 | [weights](https://cloud.hipert.unimore.it/s/mEDZMRJaGCFWSJF/download) |
| shelfnet_berkeley | ShelfNet18_realtime<sup>11</sup> | [DeepDrive](https://bdd-data.berkeley.edu/) | 20 | 1024x1024 | [weights](https://cloud.hipert.unimore.it/s/m92e7QdD9gYMF7f/download) |
demo/demo/demo.cpp
+21 -24
View File
@@ -9,7 +9,6 @@
#include "Yolo3Detection.h"
bool gRun;
bool SAVE_RESULT = false;
void sig_handler(int signo) {
std::cout<<"request gateway stop\n";
@@ -20,36 +19,33 @@ int main(int argc, char *argv[]) {
signal(SIGINT, sig_handler);
#ifdef __linux__
// get config file path and read it
#ifdef __linux__
std::string config_file = "../demo/demoConfig.yaml";
#elif _WIN32
#elif _WIN32
std::string config_file = "..\\..\\..\\demo\\demoConfig.yaml";
#endif
if(argc > 1){
#endif
if(argc > 1)
config_file = argv[1];
}
YAML::Node conf = YAMLloadConf(config_file);
if(!conf){
if(!conf)
FatalError("Problem with config file");
}
std::string net = YAMLgetConf<std::string>(conf,"net","yolo4tiny_fp32.rt");
if(!fileExist(net.c_str())) {
// read settings from config file
std::string net = YAMLgetConf<std::string>(conf, "net", "yolo4tiny_fp32.rt");
if(!fileExist(net.c_str()))
FatalError("The given network does not exist. Create the rt first.");
}
#ifdef __linux__
#ifdef __linux__
std::string input = YAMLgetConf<std::string>(conf, "input", "../demo/yolo_test.mp4");
std::string cfgPath = YAMLgetConf<std::string>(conf,"cfg_input", "../tests/darknet/cfg/yolo4tiny.cfg");
std::string namePath = YAMLgetConf<std::string>(conf,"name_input","../tests/darknet/names/coco.names");
#elif _WIN32
#elif _WIN32
std::string input = YAMLgetConf<std::string>(conf, "win_input", "..\\..\\..\\demo\\yolo_test.mp4");
std::string cfgPath = YAMLgetConf<std::string>(conf,"cfg_win_input","..\\..\\..\\tests\\darknet\\cfg\\yolo4tiny.cfg");
std::string namePath = YAMLgetConf<std::string>(conf,"name_win_input","..\\..\\..\\tests\\darknet\\names\\coco.names");
#endif
#endif
if(!fileExist(input.c_str()))
FatalError("The given input video does not exist.");
@@ -71,6 +67,7 @@ int main(int argc, char *argv[]) {
<<", show: "<< show
<<", save: "<< save<<"\n\n";
// create detection network
tk::dnn::Yolo3Detection yolo;
tk::dnn::CenternetDetection cnet;
tk::dnn::MobilenetDetection mbnet;
@@ -100,8 +97,7 @@ int main(int argc, char *argv[]) {
detNN->init(net,cfgPath,namePath,n_classes,n_batch,conf_thresh);
gRun = true;
// open video stream
cv::VideoCapture cap(input);
if(!cap.isOpened())
gRun = false;
@@ -115,13 +111,15 @@ int main(int argc, char *argv[]) {
resultVideo.open("result.mp4", cv::VideoWriter::fourcc('M','P','4','V'), 30, cv::Size(w, h));
}
cv::Mat frame;
if(show)
cv::namedWindow("detection", cv::WINDOW_NORMAL);
cv::Mat frame;
std::vector<cv::Mat> batch_frame;
std::vector<cv::Mat> batch_dnn_input;
// start detection loop
gRun = true;
while(gRun) {
batch_dnn_input.clear();
batch_frame.clear();
@@ -154,14 +152,13 @@ int main(int argc, char *argv[]) {
}
std::cout<<"detection end\n";
double mean = 0;
std::cout<<COL_GREENB<<"\n\nTime stats:\n";
std::cout<<"Min: "<<*std::min_element(detNN->stats.begin(), detNN->stats.end())/n_batch<<" ms\n";
std::cout<<"Max: "<<*std::max_element(detNN->stats.begin(), detNN->stats.end())/n_batch<<" ms\n";
std::cout<<"Min: "<<*std::min_element(detNN->stats.begin(), detNN->stats.end())<<" ms\n";
std::cout<<"Max: "<<*std::max_element(detNN->stats.begin(), detNN->stats.end())<<" ms\n";
for(int i=0; i<detNN->stats.size(); i++) mean += detNN->stats[i]; mean /= detNN->stats.size();
std::cout<<"Avg: "<<mean/n_batch<<" ms\t"<<1000/(mean/n_batch)<<" FPS\n"<<COL_END;
std::cout<<"Avg: "<<mean<<" ms\t"<<1000/(mean)<<" FPS\n"<<COL_END;
return 0;
}
include/tkDNN/SegmentationNN.h
+1
View File
@@ -185,6 +185,7 @@ class SegmentationNN {
checkCuda(cudaMemcpyAsync(stddev_d, stddev.data(), stddev.size() * sizeof(float), cudaMemcpyHostToDevice, netRT->stream));
return true;
return true;
}
/**
include/tkDNN/utils.h
+1 -1
View File
@@ -23,6 +23,7 @@
#include <ios>
#include <chrono>
#include <yaml-cpp/yaml.h>
@@ -176,5 +177,4 @@ inline T YAMLgetConf(YAML::Node conf, std::string key, T defaultVal) {
return val;
}
#endif //UTILS_H
scripts/test_all_tests.sh
+43 -39
View File
@@ -1,6 +1,6 @@
#!/bin/bash
cd build
#cd build
RED='\033[1;31m'
GREEN='\033[1;32m'
@@ -29,24 +29,28 @@ function print_output {
}
out_dir=results
out_file=results.log
rm $out_file
rm -rf $out_dir/
mkdir -p $out_dir
function test_net {
./test_$1 &>> $out_file
./test_$1 &> $out_dir/$1_${TKDNN_MODE}_build_$out_file
print_output $? $1
./test_rtinference $1*.rt $TKDNN_BATCHSIZE &>> $out_file
./test_rtinference $1*.rt 1 &> $out_dir/$1_${TKDNN_MODE}_inference_batch1_$out_file
print_output $? "infer $1"
./test_rtinference $1*.rt $TKDNN_BATCHSIZE &> $out_dir/$1_${TKDNN_MODE}_inference_batch${TKDNN_BATCHSIZE}_$out_file
print_output $? "batched $1"
}
modes=( 1 ) # only FP32
# modes=( 1 2 ) # FP32 and FP16
# modes=( 1 ) # only FP32
modes=( 1 2 ) # FP32 and FP16
# modes=( 1 2 3 ) # FP32, FP16 and INT8
for i in "${modes[@]}"
do
rm *rt
rm -f *rt
if [ $i -eq 1 ]
then
export TKDNN_MODE=FP32
@@ -73,37 +77,37 @@ do
# print_output $? imuodom
test_net yolo4
test_net yolo4_320
test_net yolo4_320_coco2
test_net yolo4_512
test_net yolo4_608
test_net yolo4-csp
test_net yolo4x
test_net yolo4_berkeley
test_net yolo4_berkeley_f1
test_net yolo4tiny
test_net yolo4tiny_512
test_net yolo3
test_net yolo3_berkeley
test_net yolo3_coco4
test_net yolo3_flir
test_net yolo3_512
test_net yolo3tiny
test_net yolo3tiny_512
test_net yolo2
test_net yolo2_voc
#test_net yolo2tiny
test_net csresnext50-panet-spp
#test_net csresnext50-panet-spp_berkeley
test_net resnet101_cnet
test_net dla34_cnet
test_net dla34_cnet3d
test_net mobilenetv2ssd
test_net mobilenetv2ssd512
test_net bdd-mobilenetv2ssd
test_net dla34_ctrack
test_net shelfnet
test_net shelfnet_berkeley
# test_net yolo4_320
# test_net yolo4_320_coco2
# test_net yolo4_512
# test_net yolo4_608
# test_net yolo4-csp
# test_net yolo4x
# test_net yolo4_berkeley
# test_net yolo4_berkeley_f1
# test_net yolo4tiny
# test_net yolo4tiny_512
# test_net yolo3
# test_net yolo3_berkeley
# test_net yolo3_coco4
# test_net yolo3_flir
# test_net yolo3_512
# test_net yolo3tiny
# test_net yolo3tiny_512
# test_net yolo2
# test_net yolo2_voc
# test_net yolo2tiny
# test_net csresnext50-panet-spp
# test_net csresnext50-panet-spp_berkeley
# test_net resnet101_cnet
# test_net dla34_cnet
# test_net dla34_cnet3d
# test_net mobilenetv2ssd
# test_net mobilenetv2ssd512
# test_net bdd-mobilenetv2ssd
# test_net dla34_ctrack
# test_net shelfnet
# test_net shelfnet_berkeley
done
echo "If errors occured, check logfile $out_file"
echo "If errors occured, check logfiles in directory: $out_dir"
src/CenternetDetection.cpp
+1
View File
@@ -120,6 +120,7 @@ bool CenternetDetection::init(const std::string& tensor_path, const std::string&
dst2.at<float>(2,1)=dst2.at<float>(1,1) + (dst2.at<float>(0,0)-dst2.at<float>(1,0) );
return true;
return true;
}
src/MobilenetDetection.cpp
+1 -1
View File
@@ -207,7 +207,7 @@ bool MobilenetDetection::init(const std::string& tensor_path, const std::string&
else{
FatalError("Number of classes not supported for mobilenet");
}
return 1;
return true;
}
void MobilenetDetection::preprocess(cv::Mat &frame, const int bi){
src/Yolo.cpp
+3 -2
View File
@@ -278,6 +278,7 @@ void Yolo::mergeDetections(Yolo::detection *dets, int ndets, int classes, double
}
total = k+1;
float thresh = 0.45f;
for(k = 0; k < classes; ++k){
for(i = 0; i < total; ++i){
dets[i].sort_class = k;
@@ -288,9 +289,9 @@ void Yolo::mergeDetections(Yolo::detection *dets, int ndets, int classes, double
box a = dets[i].bbox;
for(j = i+1; j < total; ++j){
box b = dets[j].bbox;
if (nsm_kind == GREEDY_NMS && yolo_box_iou(a, b) > nms_thresh)
if (nsm_kind == GREEDY_NMS && yolo_box_iou(a, b) > thresh)
dets[j].prob[k] = 0;
else if (nsm_kind == DIOU_NMS && yolo_box_diou(a, b, nms_thresh) > nms_thresh)
else if (nsm_kind == DIOU_NMS && yolo_box_diou(a, b, nms_thresh) > thresh)
dets[j].prob[k] = 0;
}
}
tests/darknet/cfg/yolo4-csp_crowd.cfg
+1279
View File
File diff suppressed because it is too large Load Diff
tests/darknet/names/crowdhuman.names
+2
View File
@@ -0,0 +1,2 @@
person
head
tests/darknet/yolo4-csp_crowd.cpp
+34
View File
@@ -0,0 +1,34 @@
#include<iostream>
#include<vector>
#include "tkdnn.h"
#include "test.h"
#include "DarknetParser.h"
int main() {
std::string bin_path = "yolo4-csp_crowd";
std::vector<std::string> input_bins = {
bin_path + "/layers/input.bin"
};
std::vector<std::string> output_bins = {
bin_path + "/debug/layer144_out.bin",
bin_path + "/debug/layer159_out.bin",
bin_path + "/debug/layer174_out.bin"
};
std::string wgs_path = bin_path + "/layers";
std::string cfg_path = std::string(TKDNN_PATH) + "/tests/darknet/cfg/yolo4-csp_crowd.cfg";
std::string name_path = std::string(TKDNN_PATH) + "/tests/darknet/names/crowdhuman.names";
downloadWeightsifDoNotExist(input_bins[0], bin_path, "https://cloud.hipert.unimore.it/s/RKWfWNmWXfJigsK/download");
// parse darknet network
tk::dnn::Network *net = tk::dnn::darknetParser(cfg_path, wgs_path, name_path);
net->print();
//convert network to tensorRT
tk::dnn::NetworkRT *netRT = new tk::dnn::NetworkRT(net, net->getNetworkRTName(bin_path.c_str()));
int ret = testInference(input_bins, output_bins, net, netRT);
net->releaseLayers();
delete net;
delete netRT;
return ret;
}
tests/darknet/yolo4_berkeley_f1.cpp
+1 -1
View File
@@ -17,7 +17,7 @@ int main() {
std::string wgs_path = bin_path + "/layers";
std::string cfg_path = std::string(TKDNN_PATH) + "/tests/darknet/cfg/yolo4_berkeley.cfg";
std::string name_path = std::string(TKDNN_PATH) + "/tests/darknet/names/berkeley.names";
downloadWeightsifDoNotExist(input_bins[0], bin_path, "https://cloud.hipert.unimore.it/s/M7WJdGoGDaDACnN/download");
downloadWeightsifDoNotExist(input_bins[0], bin_path, "https://cloud.hipert.unimore.it/s/q9dwoqQ5YQqEi7s/download");
// parse darknet network
tk::dnn::Network *net = tk::dnn::darknetParser(cfg_path, wgs_path, name_path);
tests/shelfnet/shelfnet_coco.cpp
+295
View File
@@ -0,0 +1,295 @@
#include <iostream>
#include <opencv2/highgui/highgui.hpp>
#include <opencv2/imgproc/imgproc.hpp>
#include "tkdnn.h"
#include "NetworkViz.h"
const char *input_bin = "shelfnet_coco/debug/input.bin";
const char *backbone[] = {
"shelfnet_coco/layers/backbone-conv1.bin",
"shelfnet_coco/layers/backbone-layer1-0-conv1.bin",
"shelfnet_coco/layers/backbone-layer1-0-conv2.bin",
"shelfnet_coco/layers/backbone-layer1-1-conv1.bin",
"shelfnet_coco/layers/backbone-layer1-1-conv2.bin",
"shelfnet_coco/layers/backbone-layer2-0-conv1.bin",
"shelfnet_coco/layers/backbone-layer2-0-conv2.bin",
"shelfnet_coco/layers/backbone-layer2-0-downsample-0.bin",
"shelfnet_coco/layers/backbone-layer2-1-conv1.bin",
"shelfnet_coco/layers/backbone-layer2-1-conv2.bin",
"shelfnet_coco/layers/backbone-layer3-0-conv1.bin",
"shelfnet_coco/layers/backbone-layer3-0-conv2.bin",
"shelfnet_coco/layers/backbone-layer3-0-downsample-0.bin",
"shelfnet_coco/layers/backbone-layer3-1-conv1.bin",
"shelfnet_coco/layers/backbone-layer3-1-conv2.bin",
"shelfnet_coco/layers/backbone-layer4-0-conv1.bin",
"shelfnet_coco/layers/backbone-layer4-0-conv2.bin",
"shelfnet_coco/layers/backbone-layer4-0-downsample-0.bin",
"shelfnet_coco/layers/backbone-layer4-1-conv1.bin",
"shelfnet_coco/layers/backbone-layer4-1-conv2.bin"};
const char *conv_out[] = {
"shelfnet_coco/layers/conv_out-conv-conv.bin",
"shelfnet_coco/layers/conv_out-conv_out.bin",
"shelfnet_coco/layers/conv_out16-conv-conv.bin",
"shelfnet_coco/layers/conv_out16-conv_out.bin",
"shelfnet_coco/layers/conv_out32-conv-conv.bin",
"shelfnet_coco/layers/conv_out32-conv_out.bin"
};
const char *decoder[] = {
"shelfnet_coco/layers/decoder-bottom-conv1.bin",
"shelfnet_coco/layers/decoder-bottom-conv12.bin",
"shelfnet_coco/layers/decoder-up_conv_list-0-conv-conv.bin",
"shelfnet_coco/layers/decoder-up_conv_list-0-conv_atten.bin",
"shelfnet_coco/layers/decoder-up_dense_list-0-conv.bin",
"shelfnet_coco/layers/decoder-up_conv_list-1-conv-conv.bin",
"shelfnet_coco/layers/decoder-up_conv_list-1-conv_atten.bin",
"shelfnet_coco/layers/decoder-up_dense_list-1-conv.bin"
};
const char *ladder[] = {
"shelfnet_coco/layers/ladder-inconv-conv1.bin",
"shelfnet_coco/layers/ladder-inconv-conv12.bin",
"shelfnet_coco/layers/ladder-down_module_list-0-conv1.bin",
"shelfnet_coco/layers/ladder-down_module_list-0-conv12.bin",
"shelfnet_coco/layers/ladder-down_conv_list-0.bin",
"shelfnet_coco/layers/ladder-down_module_list-1-conv1.bin",
"shelfnet_coco/layers/ladder-down_module_list-1-conv12.bin",
"shelfnet_coco/layers/ladder-down_conv_list-1.bin",
"shelfnet_coco/layers/ladder-bottom-conv1.bin",
"shelfnet_coco/layers/ladder-bottom-conv12.bin",
"shelfnet_coco/layers/ladder-up_conv_list-0-conv-conv.bin",
"shelfnet_coco/layers/ladder-up_conv_list-0-conv_atten.bin",
"shelfnet_coco/layers/ladder-up_dense_list-0-conv.bin",
"shelfnet_coco/layers/ladder-up_conv_list-1-conv-conv.bin",
"shelfnet_coco/layers/ladder-up_conv_list-1-conv_atten.bin",
"shelfnet_coco/layers/ladder-up_dense_list-1-conv.bin"};
const char *trans[] = {
"shelfnet_coco/layers/trans1-conv.bin",
"shelfnet_coco/layers/trans2-conv.bin",
"shelfnet_coco/layers/trans3-conv.bin"};
int main()
{
downloadWeightsifDoNotExist(input_bin, "shelfnet_coco", "https://cloud.hipert.unimore.it/s/KfQ9fGJQsgzNbiW/download");
int classes = 183;
// Network layout
tk::dnn::dataDim_t dim(1, 3, 1024, 1024, 1);
tk::dnn::Network net(dim);
int bi = 0, di = 0, li = 0, ci = 0;
new tk::dnn::Conv2d(&net, 64, 7, 7, 2, 2, 3, 3, backbone[bi++], true);
new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
tk::dnn::Layer* last = new tk::dnn::Pooling (&net, 3, 3, 2, 2, 1, 1, tk::dnn::POOLING_MAX);
for(int i=0; i<2; ++i){
new tk::dnn::Conv2d (&net, 64, 3, 3, 1, 1, 1, 1, backbone[bi++], true);
new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
new tk::dnn::Conv2d (&net, 64, 3, 3, 1, 1, 1, 1, backbone[bi++], true);
new tk::dnn::Shortcut(&net, last);
last = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
}
std::vector<tk::dnn::Layer*> features;
for(int i=0;i<3;++i){
int out_channel = pow(2,7+i);
std::cout<<out_channel<<std::endl;
new tk::dnn::Conv2d (&net, out_channel, 3, 3, 2, 2, 1, 1, backbone[bi++], true);
new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
tk::dnn::Layer* bn2 = new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, backbone[bi++], true);
new tk::dnn::Route(&net, &last, 1);
new tk::dnn::Conv2d (&net, out_channel, 1, 1, 2, 2, 0, 0, backbone[bi++], true);
new tk::dnn::Shortcut(&net, bn2);
last = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, backbone[bi++], true);
new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, backbone[bi++], true);
new tk::dnn::Shortcut(&net, last);
last = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
features.push_back(last);
}
for(int i=0; i<features.size(); ++i){
new tk::dnn::Route(&net, &features[i], 1);
int out_channel = pow(2,6+i);
new tk::dnn::Conv2d (&net, out_channel, 1, 1, 1, 1, 0, 0, trans[i], true);
features[i] = new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
}
//DECODER
last = features[2];
std::vector<tk::dnn::Layer*> up_out;
//bottom
new tk::dnn::Conv2d (&net, 256, 3, 3, 1, 1, 1, 1, decoder[di++], true, false, 1, true);
new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
new tk::dnn::Conv2d (&net, 256, 3, 3, 1, 1, 1, 1, decoder[di++], true, false, 1, true);
new tk::dnn::Shortcut(&net, last);
last = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
up_out.push_back(last);
for(int i=0; i<2; ++i){
int out_channel = pow(2,7-i);
//up-conv
std::cout<<out_channel<<std::endl;
new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, decoder[di++], true);
last = new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
new tk::dnn::Pooling(&net, last->output_dim.w, last->output_dim.h, last->output_dim.w, last->output_dim.h, 0, 0, tk::dnn::POOLING_AVERAGE);
new tk::dnn::Conv2d (&net, out_channel, 1, 1, 1, 1, 0, 0, decoder[di++], true);
tk::dnn::Layer* act = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_SIGMOID);
new tk::dnn::Route(&net, &last, 1);
new tk::dnn::Shortcut(&net, act, true);
//interpolate
new tk::dnn::Resize(&net, 1,2,2);
new tk::dnn::Shortcut(&net, features[1-i]);
//up-dense
new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, decoder[di++], true);
last = new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
up_out.push_back(last);
}
//LADDER
std::vector<tk::dnn::Layer*> down_out;
new tk::dnn::Conv2d (&net, 64, 3, 3, 1, 1, 1, 1, ladder[li++], true, false, 1, true);
new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
new tk::dnn::Conv2d (&net, 64, 3, 3, 1, 1, 1, 1, ladder[li++], true, false, 1, true);
new tk::dnn::Shortcut(&net, last);
new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
for(int i=0; i<2;++i){
int out_channel = pow(2,6+i);
tk::dnn::Layer* l_last = new tk::dnn::Shortcut(&net, up_out[2-i]);
new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, ladder[li++], true, false, 1, true);
new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, ladder[li++], true, false, 1, true);
new tk::dnn::Shortcut(&net, l_last);
l_last = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
down_out.push_back(l_last);
new tk::dnn::Conv2d (&net, out_channel*2, 3, 3, 2, 2, 1, 1, ladder[li++], false);
last = new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.0f); //should be ReLU
}
new tk::dnn::Conv2d (&net, 256, 3, 3, 1, 1, 1, 1, ladder[li++], true, false, 1, true);
new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
new tk::dnn::Conv2d (&net, 256, 3, 3, 1, 1, 1, 1, ladder[li++], true, false, 1, true);
new tk::dnn::Shortcut(&net, last);
last = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
up_out.clear();
up_out.push_back(last);
for(int i=0; i<2; ++i){
int out_channel = pow(2,7-i);
//up-conv
new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, ladder[li++], true);
last = new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
new tk::dnn::Pooling(&net, last->output_dim.w, last->output_dim.h, last->output_dim.w, last->output_dim.h, 0, 0, tk::dnn::POOLING_AVERAGE);
new tk::dnn::Conv2d (&net, out_channel, 1, 1, 1, 1, 0, 0, ladder[li++], true);
tk::dnn::Layer* act = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_SIGMOID);
new tk::dnn::Route(&net, &last, 1);
new tk::dnn::Shortcut(&net, act, true);
//interpolate
new tk::dnn::Resize(&net, 1,2,2);
new tk::dnn::Shortcut(&net, down_out[1-i]);
// //up-dense
new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, ladder[li++], true);
last = new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
up_out.push_back(last);
}
// for(int i=2;i>=0;--i){
// new tk::dnn::Route(&net, &up_out[i], 1);
new tk::dnn::Conv2d (&net, 64, 3, 3, 1, 1, 1, 1, conv_out[ci++], true);
new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
new tk::dnn::Conv2d (&net, classes, 3, 3, 1, 1, 1, 1, conv_out[ci++], false);
/*up_out[i] =*/ new tk::dnn::Resize(&net, classes, net.input_dim.h, net.input_dim.w, true, tk::dnn::ResizeMode_t::LINEAR);
// // }
new tk::dnn::Softmax(&net);
const char *output_bin = "shelfnet_coco/debug/softmax.bin";
// Load input
dnnType *data;
dnnType *input_h;
readBinaryFile(input_bin, dim.tot(), &input_h, &data);
std::cout<<"Input:"<<std::endl;
//print network model
net.print();
// // convert network to tensorRT
tk::dnn::NetworkRT netRT(&net, net.getNetworkRTName("shelfnet_coco"));
tk::dnn::dataDim_t dim1 = dim; //input dim
dnnType *cudnn_out = nullptr;
printCenteredTitle(" CUDNN inference ", '=', 30);
{
dim1.print();
TKDNN_TSTART
cudnn_out = net.infer(dim1, data);
TKDNN_TSTOP
dim1.print();
}
tk::dnn::dataDim_t dim2 = dim;
printCenteredTitle(" TENSORRT inference ", '=', 30);
{
dim2.print();
TKDNN_TSTART
netRT.infer(dim2, data);
TKDNN_TSTOP
dim2.print();
}
dnnType *rt_out1 = (dnnType *)netRT.buffersRT[1];
printCenteredTitle(std::string(" CHECK RESULTS ").c_str(), '=', 30);
dnnType *out1, *out1_h;
int odim1 = dim1.tot();
readBinaryFile(output_bin, odim1, &out1_h, &out1);
int ret_cudnn = 0, ret_tensorrt = 0, ret_cudnn_tensorrt = 0;
// std::cout << "CUDNN vs correct" << std::endl;
// ret_cudnn |= checkResult(odim1, cudnn_out, out1, true, 20) == 0 ? 0 : ERROR_CUDNN;
std::cout << "TRT vs correct" << std::endl;
ret_tensorrt |=checkResult(odim1, rt_out1, out1) == 0 ? 0 : ERROR_TENSORRT;
std::cout << "CUDNN vs TRT " << std::endl;
ret_cudnn_tensorrt |= checkResult(odim1, cudnn_out, rt_out1) == 0 ? 0 : ERROR_CUDNNvsTENSORRT;
cv::Mat viz = vizLayer2Mat(&net, net.num_layers-1);
cv::imwrite("test.png", viz);
return ret_cudnn | ret_tensorrt | ret_cudnn_tensorrt;
}
tests/test_rtinference/rtinference.cpp
+4 -4
View File
@@ -1,6 +1,6 @@
#include<iostream>
#include<algorithm>
#include "tkdnn.h"
#include "tkDNN/tkdnn.h"
#include <stdlib.h> /* srand, rand */
@@ -66,11 +66,11 @@ int main(int argc, char *argv[]) {
}
}
double min = *std::min_element(stats.begin(), stats.end())/BATCH_SIZE;
double max = *std::max_element(stats.begin(), stats.end())/BATCH_SIZE;
double min = *std::min_element(stats.begin(), stats.end()); ///BATCH_SIZE;
double max = *std::max_element(stats.begin(), stats.end()); ///BATCH_SIZE;
double mean =0;
for(int i=0; i<stats.size(); i++) mean += stats[i]; mean /= stats.size();
mean /=BATCH_SIZE;
//mean /=BATCH_SIZE;
std::cout<<"Min: "<<min<<" ms\n";
std::cout<<"Max: "<<max<<" ms\n";