Co-DETR TensorRT Mo Xing Duan Dao Duan Jia Su Tui Li De C++Shi Xian
Xu Jing
0. Huan Jing Pei Zhi Shuo Ming
- Ubuntu16.04Xia An Zhuang mmdetection, mmdeploy, Qi Yi Lai mmcvHe mmengine
# mmdetection==3.3.0
git clone -b 3.3.0 https://github.com/open-mmlab/mmdetection
pip install -v -e .
# mmcv
pip install mmcv==2.0.0
# mmdeploy
# https://github.com/TommyZihao/MMDeploy_Tutorials
git clone -b 1.3.1 https://github.com/open-mmlab/mmdeploy --recursive
# Bian Yi Bing An Zhuang MMDeploy(Hao Shi Da Yue Shi Fen Zhong )
python tools/scripts/build_ubuntu_x64_ort.py
git clone -b 3.3.0 https://github.com/open-mmlab/mmdetection
pip install -v -e .
# mmcv
pip install mmcv==2.0.0
# mmdeploy
# https://github.com/TommyZihao/MMDeploy_Tutorials
git clone -b 1.3.1 https://github.com/open-mmlab/mmdeploy --recursive
# Bian Yi Bing An Zhuang MMDeploy(Hao Shi Da Yue Shi Fen Zhong )
python tools/scripts/build_ubuntu_x64_ort.py
- windows TensorRTDe Huan Jing
- TensorRT 8.5
- cuda 11.0, cudnn
- vs2017
- cmake version 3.22.1
- opencv
1.Co-DETR Zhuan ONNX
1.Xiu Gai Mo Xing Pei Zhi Wen Jian ,Guan Bi Ce Shi Guo Cheng Zhong De soft-nms(Hou Mian Yong EfficientNMS PluginDai Ti )
# mmdetection/projects/CO-DETR/configs/codino/co_dino_5scale_swin_l_16xb1_16e_o365tococo.py
#mmdetection/projects/CO-DETR/configs/codino/co_dino_5scale_r50_lsj_8xb2_1x_coco.py
test_cfg=[
# # Deferent from the DINO, we use the NMS.
dict(
max_per_img=300,
# NMS can improve the mAP by 0.2.
# nms=dict(type='soft_nms', iou_threshold=0.8)), # Guan Diao testGuo Cheng Zhong De soft nms
),
#mmdetection/projects/CO-DETR/configs/codino/co_dino_5scale_r50_lsj_8xb2_1x_coco.py
test_cfg=[
# # Deferent from the DINO, we use the NMS.
dict(
max_per_img=300,
# NMS can improve the mAP by 0.2.
# nms=dict(type='soft_nms', iou_threshold=0.8)), # Guan Diao testGuo Cheng Zhong De soft nms
),
2.Xiu Gai mmdeployZhong Guan Yu onnxDe Dao Chu Pei Zhi
# mmdeploy/configs/_base_/onnx_config.py
onnx_config = dict(
type='onnx',
export_params=True,
keep_initializers_as_inputs=False,
opset_version=11, # opset Ban Ben
save_file='end2end.onnx', #Zhuan Chu onnxDe Bao Cun Ming Zi
input_names=['input'], # inputDe Ming Zi
output_names=['output'], # outputDe Ming Zi
input_shape=None,
optimize=True)
# mmdeploy/configs/mmdet/_base_/base_static.py
_base_ = ['../../_base_/onnx_config.py']
onnx_config = dict(output_names=['dets', 'labels'], input_shape=[640,640]) # static inputDe Da Xiao She Zhi Wei 640x640
codebase_config = dict(
type='mmdet',
task='ObjectDetection',
model_type='end2end',
post_processing=dict(
score_threshold=0.05,
confidence_threshold=0.005, # for YOLOv3
iou_threshold=0.5,
max_output_boxes_per_class=200,
pre_top_k=5000,
keep_top_k=100,
background_label_id=-1,
))
# co-dinoShi Yong Liao Duo Chi Du Xun Lian ,Zhe Li Wo Men Jiang test inputDe Chi Du She Wei 640x640,Jian Shao Ji Suan Liang
onnx_config = dict(
type='onnx',
export_params=True,
keep_initializers_as_inputs=False,
opset_version=11, # opset Ban Ben
save_file='end2end.onnx', #Zhuan Chu onnxDe Bao Cun Ming Zi
input_names=['input'], # inputDe Ming Zi
output_names=['output'], # outputDe Ming Zi
input_shape=None,
optimize=True)
# mmdeploy/configs/mmdet/_base_/base_static.py
_base_ = ['../../_base_/onnx_config.py']
onnx_config = dict(output_names=['dets', 'labels'], input_shape=[640,640]) # static inputDe Da Xiao She Zhi Wei 640x640
codebase_config = dict(
type='mmdet',
task='ObjectDetection',
model_type='end2end',
post_processing=dict(
score_threshold=0.05,
confidence_threshold=0.005, # for YOLOv3
iou_threshold=0.5,
max_output_boxes_per_class=200,
pre_top_k=5000,
keep_top_k=100,
background_label_id=-1,
))
# co-dinoShi Yong Liao Duo Chi Du Xun Lian ,Zhe Li Wo Men Jiang test inputDe Chi Du She Wei 640x640,Jian Shao Ji Suan Liang
3.mmdeployZhuan onnx
python mmdeploy/tools/deploy.py \
mmdeploy/configs/mmdet/detection/detection_onnxruntime_stati c.py \
mmdetection/projects/CO-DETR/configs/codino/co_dino_5scale_s win_l_16xb1_16e_o365tococo.py \
mmdetection/checkpoints/co_dino_5scale_swin_large_16e_o365to coco-614254c9.pth \
mmdetection/demo/demo.jpg \
--work-dir mmdetection/checkpoints \
--device cpu
# Zhe Ge Guo Cheng Sheng Cheng Liao end2end.onnxDe ,Dan Shi onnxruntimeDe Shi Hou Huo Bao Cuo ,Bao Cuo De Yuan Yin Shi grid_samplerSuan Zi onnxruntimeHe tensorrtJun Bu Zhi Chi ,Shao Hou Hui Bian Yi tensorrt pluginJie Jue Gai Wei Wen Ti
mmdeploy/configs/mmdet/detection/detection_onnxruntime_stati c.py \
mmdetection/projects/CO-DETR/configs/codino/co_dino_5scale_s win_l_16xb1_16e_o365tococo.py \
mmdetection/checkpoints/co_dino_5scale_swin_large_16e_o365to coco-614254c9.pth \
mmdetection/demo/demo.jpg \
--work-dir mmdetection/checkpoints \
--device cpu
# Zhe Ge Guo Cheng Sheng Cheng Liao end2end.onnxDe ,Dan Shi onnxruntimeDe Shi Hou Huo Bao Cuo ,Bao Cuo De Yuan Yin Shi grid_samplerSuan Zi onnxruntimeHe tensorrtJun Bu Zhi Chi ,Shao Hou Hui Bian Yi tensorrt pluginJie Jue Gai Wei Wen Ti
4.Dui onnxJin Xing onnxsimHe fold constants
polygraphy surgeon sanitize end2end.onnx --fold-constants -o end2end_folded.onnx
python -m onnxsim end2end_folded.onnx end2end_folded_sim.onnx
python -m onnxsim end2end_folded.onnx end2end_folded_sim.onnx
Zhu Yi :
# Chang Liang Zhe Die He simplifierShe Ji Dao De Ku De Ban Ben
polygraphy==0.49.0
onnxruntime-gpu==1.19.2
onnx-simplifier=0.4.36
2. Windows Xia Dan Du Bian Yi mmdeployZhong Jin She Ji Co-DETRDe TensorRT Plugin
cmake_minimum_required(VERSION 2.6)
project(mmdeploy_plugins)
add_definitions(-std=c++11)
add_definitions(-DAPI_EXPORTS)
option(CUDA_USE_STATIC_CUDA_RUNTIME OFF)
set(CMAKE_CXX_STANDARD 11)
set(CMAKE_BUILD_TYPE Release)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /O2")
add_compile_definitions(WIN32_LEAN_AND_MEAN NOMINMAX)
find_package(CUDA REQUIRED)
#if(WIN32)
#enable_language(CUDA)
#endif(WIN32)
# cuda
set(cuda_inc "C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v11.0/include")
set(cuda_lib "C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v11.0/lib/x64")
include_directories(${cuda_inc})
link_directories(${cuda_lib})
#cub
set(CUB_ROOT_DIR "C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v11.0/include/cub")
include_directories(${CUB_ROOT_DIR})
# tensorrt
set(tensorrt_inc "D:/trt_install/TensorRT-8.5.1.7/include")
set(tensorrt_lib "D:/trt_install/TensorRT-8.5.1.7/lib")
include_directories(${tensorrt_inc})
link_directories(${tensorrt_lib})
# opencv
#include_directories("${PROJECT_SOURCE_DIR}/third_party/CV460_64/include")
#set(opencv_lib "${PROJECT_SOURCE_DIR}/third_party/CV460_64/lib/opencv_world460.lib")
# common files,Lai Yuan Yu mmdeploy
include_directories(common)
file(GLOB grid_sampler_src ${PROJECT_SOURCE_DIR}/grid_sampler/*.cpp ${PROJECT_SOURCE_DIR}/grid_sampler/*.cu)
cuda_add_library(trtgrid_sampler SHARED ${grid_sampler_src})
#cuda_add_library(trtgrid_sampler STATIC ${grid_sampler_src})
target_link_libraries(trtgrid_sampler nvinfer cudart)
file(GLOB topk_src ${PROJECT_SOURCE_DIR}/gather_topk/*.cpp ${PROJECT_SOURCE_DIR}/gather_topk/*.cu)
cuda_add_library(trtgather_topk SHARED ${topk_src})
#cuda_add_library(trtgather_topk STATIC ${topk_src})
target_link_libraries(trtgather_topk nvinfer cudart)
if(UNIX)
add_definitions(-O2 -pthread)
endif(UNIX)
- Da Kai vs studio 2017De Zhong Duan
x64 Native Tools Command ...,cdDao Xiang Mu De Mu Lu Jin Xing Bian Yi
2.windowsXia Bian Yi TensorRT Plugin
mkdir build && cd build
cmake -G "NMake Makefiles" ..
nmake
cmake -G "NMake Makefiles" ..
nmake
Zai buildWen Jian Jia Xia Sheng Cheng Liao trtgrid_sampler.dllHe trtgather_topk.dll,Xia Mian Wo Men Hui Shi Yong trtgrid_sampler.dllDe plugin .
3.Co-DETR ONNX GraphXiu Gai He Bian Ji
Yuan Shi Dao Chu De Bu Bao Han nmsDe graph
Zhi Xing Bian Ji onnx graphDe Jiao Ben :
python co_detr_add_nms.py
Mo Xing Jie Gou Bian Wei :
4.WindowsXia Xu Lie Hua Co-DETR TensorRT engine
trtexec --onnx=end2end_foled_sim_nms.onnx --saveEngine=test_1.plan --workspace=60000 --verbose --plugins=./trtgrid_sampler.dll
5.mmdetectionZhong Co-DETRMo Xing Qian Chu Li Shi Xian De Fen Xi He C++Zhong Xie
mmdetection 3.3.0 co-dinoDe Qian Chu Li :
-
opencvDu Ru BGRTu Xiang
-
Deng Bi Li Suo Fang ,Chang Bian Suo Fang Dao 640,Suo Fang Fang Fa bilinear
-
normalize:
mean=[123.675, 116.28, 103.53], # RGB
std=[58.395, 57.12, 57.375], #RGB -
BGR2RGB
-
Duan Bian You Xia Jiao Tian Chong Wei 0
C++Shi Xian Ru Xia :
//mmdetection3.3.0 co-detrQian Chu Li
void codetr::preprocess(cv::Mat &img, float data[]) {
int w, h, x, y;
float r_w = INPUT_W / (img.cols*1.0);
float r_h = INPUT_H / (img.rows*1.0);
if (r_h > r_w) {
w = INPUT_W;
h = r_w * img.rows;
}
else {
w = r_h * img.cols;
h = INPUT_H;
}
cv::Mat re(h, w, CV_8UC3);
cv::resize(img, re, re.size(), 0, 0, cv::INTER_LINEAR);
cv::Mat out(INPUT_H, INPUT_W, CV_8UC3, cv::Scalar(103, 116, 123)); //(0,0,0)Xiang Su Tian Chong
re.copyTo(out(cv::Rect(0, 0, re.cols, re.rows))); //You Xia Jiao
int i = 0;
for (int row = 0; row < INPUT_H; ++row) {
uchar* uc_pixel = out.data + row * out.step;
for (int col = 0; col < INPUT_W; ++col) {
data[i] = ((float)uc_pixel[2] - 123.675)/58.395; //R
data[i + INPUT_H * INPUT_W] = ((float)uc_pixel[1] - 116.28) / 57.12; //G
data[i + 2 * INPUT_H * INPUT_W] = ((float)uc_pixel[0] - 103.53)/ 57.375; //B
uc_pixel += 3;
++i;
}
}
}
void codetr::preprocess(cv::Mat &img, float data[]) {
int w, h, x, y;
float r_w = INPUT_W / (img.cols*1.0);
float r_h = INPUT_H / (img.rows*1.0);
if (r_h > r_w) {
w = INPUT_W;
h = r_w * img.rows;
}
else {
w = r_h * img.cols;
h = INPUT_H;
}
cv::Mat re(h, w, CV_8UC3);
cv::resize(img, re, re.size(), 0, 0, cv::INTER_LINEAR);
cv::Mat out(INPUT_H, INPUT_W, CV_8UC3, cv::Scalar(103, 116, 123)); //(0,0,0)Xiang Su Tian Chong
re.copyTo(out(cv::Rect(0, 0, re.cols, re.rows))); //You Xia Jiao
int i = 0;
for (int row = 0; row < INPUT_H; ++row) {
uchar* uc_pixel = out.data + row * out.step;
for (int col = 0; col < INPUT_W; ++col) {
data[i] = ((float)uc_pixel[2] - 123.675)/58.395; //R
data[i + INPUT_H * INPUT_W] = ((float)uc_pixel[1] - 116.28) / 57.12; //G
data[i + 2 * INPUT_H * INPUT_W] = ((float)uc_pixel[0] - 103.53)/ 57.375; //B
uc_pixel += 3;
++i;
}
}
}
6.Co-DETR TensorRT C++Shi Xian He Ce Shi
Zhu Yi C++Jia Zai Zi Ji Ding Yi De Plugin
bool didInitPlugins = initLibNvInferPlugins(nullptr, "");
void* handle_grid_sampler = LoadLibrary(L"trtgrid_sampler.dll");
void* handle_grid_sampler = LoadLibrary(L"trtgrid_sampler.dll");
TensorRT C++De Tui Li Demo:
| bus.jpg | zidane.jpg |
|---|---|
7. LinuxRu He Bian Yi Gai Cheng Xu
- Wo Ti Gong Liao Zai LinuxXia Bian Yi Co-DETRJin Xing Duan Dao Duan Tui Li De Dai Ma ,Qing Can Kao
linux_cc/,Qi ZhongpluginWei grid_sampler pluginDe Bian Yi ,co_detrWei Co-DETRDe TensorRTDiao Yong .
[!NOTE]\
- Co-DETR TensorRTDe Shi Xian ,Keng Que Shi Bi Jiao Duo ,Can Kao De Wang Luo Zi Yuan Ji Ben Mei You
- Wo Men Jiang soft-nmsSuan Zi Shan Chu ,Ti Huan Wei TensorRT EfficientNMS Plugin
- Wo Men Zai windowsXia Bian Yi Liao TensorRT Plugin grid_sampler
Zui Zhong Cheng Gong Shi Xian Liao Co-DETRDe Duan Dao Duan De TensorRT Mo Xing Tui Li Yi Gou Ji Suan Jia Su Tui Li !