前言
因为最近在研究PaddleX和PaddleSeg在win10下的部署,主要是奔向C++去的,所以这篇其实大部分记录工程搭建的工作;这篇以今年刚出的STDC为例。
我的环境:
- Visual Studio 2019
- CUDA 11.5,cudnn 8.2
- CMake 3.17.1
一.模型解读
可以参考STDC语义分割网络,模型关键的地方都有做介绍。需要具体研究的建议去看一下github项目。
二.模型训练
可参考官方教程,如果要训练自定义的数据(voc语义
分割格式),需要修改配置文件。
-
配置文件的参数修改:
configs\_base_\pascal_voc12aug.yml(Line:4) (mode: train) configs\_base_\pascal_voc12.yml(Line:2、Line:27) Line:2 (iters: 80000) Line:27 (target_size: [你的图像尺寸宽度, 你的图像尺寸长度]) paddleseg\datasets\voc.py(Line:39) (NUM_CLASSES = 数据集总类别数量) -
训练:
#开始训练(若没准备数据集它会自己下载): python train.py --config configs/stdcseg/stdc2_seg_voc12aug_512x512_40k.yml --save_dir output_voc2012 #恢复训练: python train.py --config configs/stdcseg/stdc2_seg_voc12aug_512x512_40k.yml --save_dir output_voc2012 --resume_model output_voc2012/iter_54000 --do_eval --use_vdl
-
预测:
python predict.py --config configs/stdcseg/stdc2_seg_voc12aug_512x512_40k.yml --model_path output_voc2012/best_model/model.pdparams --image_path data/test_imgs/ADE_test_00003102.jpg --save_dir output_voc2012/result --custom_color 0 0 0 255 255 255
-
导出预测模型
python export.py --config configs/stdcseg/stdc2_seg_voc12aug_512x512_40k.yml --model_path output_voc2012/best_model/model.pdparams --save_dir output_voc2012_shared
三.VS2019运行C++预测
Step0: 搭建环境
进入到paddleseg/deploy/cpp目录,新建一个build文件夹;将cmakelists.txt替换成我修改后的cmakelists.txt,注意cuda(第77行)和yaml(第193~197行)设置为自己对应的路径。
cmake_minimum_required(VERSION 3.0)
project(cpp_inference_demo CXX C)
option(WITH_MKL "Compile demo with MKL/OpenBlas support, default use MKL." ON)
option(WITH_GPU "Compile demo with GPU/CPU, default use CPU." OFF)
option(WITH_STATIC_LIB "Compile demo with static/shared library, default use static." ON)
option(USE_TENSORRT "Compile demo with TensorRT." OFF)
option(WITH_ROCM "Compile demo with rocm." OFF)
if(NOT WITH_STATIC_LIB)
add_definitions("-DPADDLE_WITH_SHARED_LIB")
else()
# PD_INFER_DECL is mainly used to set the dllimport/dllexport attribute in dynamic library mode.
# Set it to empty in static library mode to avoid compilation issues.
add_definitions("/DPD_INFER_DECL=")
endif()
macro(safe_set_static_flag)
foreach(flag_var
CMAKE_CXX_FLAGS CMAKE_CXX_FLAGS_DEBUG CMAKE_CXX_FLAGS_RELEASE
CMAKE_CXX_FLAGS_MINSIZEREL CMAKE_CXX_FLAGS_RELWITHDEBINFO)
if(${flag_var} MATCHES "/MD")
string(REGEX REPLACE "/MD" "/MT" ${flag_var} "${${flag_var}}")
endif(${flag_var} MATCHES "/MD")
endforeach(flag_var)
endmacro()
if(NOT DEFINED PADDLE_LIB)
message(FATAL_ERROR "please set PADDLE_LIB with -DPADDLE_LIB=/path/paddle/lib")
endif()
if(NOT DEFINED DEMO_NAME)
message(FATAL_ERROR "please set DEMO_NAME with -DDEMO_NAME=demo_name")
endif()
include_directories("${PADDLE_LIB}/")
set(PADDLE_LIB_THIRD_PARTY_PATH "${PADDLE_LIB}/third_party/install/")
include_directories("${PADDLE_LIB_THIRD_PARTY_PATH}protobuf/include")
include_directories("${PADDLE_LIB_THIRD_PARTY_PATH}glog/include")
include_directories("${PADDLE_LIB_THIRD_PARTY_PATH}gflags/include")
include_directories("${PADDLE_LIB_THIRD_PARTY_PATH}xxhash/include")
include_directories("${PADDLE_LIB_THIRD_PARTY_PATH}cryptopp/include")
link_directories("${PADDLE_LIB_THIRD_PARTY_PATH}protobuf/lib")
link_directories("${PADDLE_LIB_THIRD_PARTY_PATH}glog/lib")
link_directories("${PADDLE_LIB_THIRD_PARTY_PATH}gflags/lib")
link_directories("${PADDLE_LIB_THIRD_PARTY_PATH}xxhash/lib")
link_directories("${PADDLE_LIB_THIRD_PARTY_PATH}cryptopp/lib")
link_directories("${PADDLE_LIB}/paddle/lib")
if (WIN32)
add_definitions("/DGOOGLE_GLOG_DLL_DECL=")
option(MSVC_STATIC_CRT "use static C Runtime library by default" ON)
if (MSVC_STATIC_CRT)
if (WITH_MKL)
set(FLAG_OPENMP "/openmp")
endif()
set(CMAKE_C_FLAGS_DEBUG "${CMAKE_C_FLAGS_DEBUG} /bigobj /MTd ${FLAG_OPENMP}")
set(CMAKE_C_FLAGS_RELEASE "${CMAKE_C_FLAGS_RELEASE} /bigobj /MT ${FLAG_OPENMP}")
set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} /bigobj /MTd ${FLAG_OPENMP}")
set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /bigobj /MT ${FLAG_OPENMP}")
safe_set_static_flag()
if (WITH_STATIC_LIB)
add_definitions(-DSTATIC_LIB)
endif()
endif()
else()
if(WITH_MKL)
set(FLAG_OPENMP "-fopenmp")
endif()
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11 ${FLAG_OPENMP}")
endif()
if(WITH_GPU)
set(CUDA_LIB "C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v11.5\\lib\\x64")
#if(NOT WIN32)
#set(CUDA_LIB "/usr/local/cuda/lib64/" CACHE STRING "CUDA Library")
#else()
#if(CUDA_LIB STREQUAL "")
#set(CUDA_LIB "C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v11.5\\lib\\x64")
#endif()
#endif(NOT WIN32)
endif()
if (USE_TENSORRT AND WITH_GPU)
set(TENSORRT_ROOT "D:\\lbq\\TensorRT-7.2.3.4")
if("${TENSORRT_ROOT}" STREQUAL "")
message(FATAL_ERROR "The TENSORRT_ROOT is empty, you must assign it a value with CMake command. Such as: -DTENSORRT_ROOT=TENSORRT_ROOT_PATH ")
endif()
set(TENSORRT_INCLUDE_DIR ${TENSORRT_ROOT}/include)
set(TENSORRT_LIB_DIR ${TENSORRT_ROOT}/lib)
file(READ ${TENSORRT_INCLUDE_DIR}/NvInfer.h TENSORRT_VERSION_FILE_CONTENTS)
string(REGEX MATCH "define NV_TENSORRT_MAJOR +([0-9]+)" TENSORRT_MAJOR_VERSION
"${TENSORRT_VERSION_FILE_CONTENTS}")
if("${TENSORRT_MAJOR_VERSION}" STREQUAL "")
file(READ ${TENSORRT_INCLUDE_DIR}/NvInferVersion.h TENSORRT_VERSION_FILE_CONTENTS)
string(REGEX MATCH "define NV_TENSORRT_MAJOR +([0-9]+)" TENSORRT_MAJOR_VERSION
"${TENSORRT_VERSION_FILE_CONTENTS}")
endif()
if("${TENSORRT_MAJOR_VERSION}" STREQUAL "")
message(SEND_ERROR "Failed to detect TensorRT version.")
endif()
string(REGEX REPLACE "define NV_TENSORRT_MAJOR +([0-9]+)" "\\1"
TENSORRT_MAJOR_VERSION "${TENSORRT_MAJOR_VERSION}")
message(STATUS "Current TensorRT header is ${TENSORRT_INCLUDE_DIR}/NvInfer.h. "
"Current TensorRT version is v${TENSORRT_MAJOR_VERSION}. ")
include_directories("${TENSORRT_INCLUDE_DIR}")
link_directories("${TENSORRT_LIB_DIR}")
endif()
if(WITH_MKL)
set(MATH_LIB_PATH "${PADDLE_LIB_THIRD_PARTY_PATH}mklml")
include_directories("${MATH_LIB_PATH}/include")
if(WIN32)
set(MATH_LIB ${MATH_LIB_PATH}/lib/mklml${CMAKE_STATIC_LIBRARY_SUFFIX}
${MATH_LIB_PATH}/lib/libiomp5md${CMAKE_STATIC_LIBRARY_SUFFIX})
else()
set(MATH_LIB ${MATH_LIB_PATH}/lib/libmklml_intel${CMAKE_SHARED_LIBRARY_SUFFIX}
${MATH_LIB_PATH}/lib/libiomp5${CMAKE_SHARED_LIBRARY_SUFFIX})
endif()
set(MKLDNN_PATH "${PADDLE_LIB_THIRD_PARTY_PATH}mkldnn")
if(EXISTS ${MKLDNN_PATH})
include_directories("${MKLDNN_PATH}/include")
if(WIN32)
set(MKLDNN_LIB ${MKLDNN_PATH}/lib/mkldnn.lib)
else(WIN32)
set(MKLDNN_LIB ${MKLDNN_PATH}/lib/libmkldnn.so.0)
endif(WIN32)
endif()
else()
set(OPENBLAS_LIB_PATH "${PADDLE_LIB_THIRD_PARTY_PATH}openblas")
include_directories("${OPENBLAS_LIB_PATH}/include/openblas")
if(WIN32)
set(MATH_LIB ${OPENBLAS_LIB_PATH}/lib/openblas${CMAKE_STATIC_LIBRARY_SUFFIX})
else()
set(MATH_LIB ${OPENBLAS_LIB_PATH}/lib/libopenblas${CMAKE_STATIC_LIBRARY_SUFFIX})
endif()
endif()
if(WITH_STATIC_LIB)
set(DEPS ${PADDLE_LIB}/paddle/lib/libpaddle_inference${CMAKE_STATIC_LIBRARY_SUFFIX})
else()
if(WIN32)
set(DEPS ${PADDLE_LIB}/paddle/lib/paddle_inference${CMAKE_STATIC_LIBRARY_SUFFIX})
else()
set(DEPS ${PADDLE_LIB}/paddle/lib/libpaddle_inference${CMAKE_SHARED_LIBRARY_SUFFIX})
endif()
endif()
if (NOT WIN32)
set(EXTERNAL_LIB "-lrt -ldl -lpthread")
set(DEPS ${DEPS}
${MATH_LIB} ${MKLDNN_LIB}
glog gflags protobuf xxhash cryptopp
${EXTERNAL_LIB})
else()
set(DEPS ${DEPS}
${MATH_LIB} ${MKLDNN_LIB}
glog gflags_static libprotobuf xxhash cryptopp-static ${EXTERNAL_LIB})
set(DEPS ${DEPS} shlwapi.lib)
endif(NOT WIN32)
if(WITH_GPU)
if(NOT WIN32)
if (USE_TENSORRT)
set(DEPS ${DEPS} ${TENSORRT_LIB_DIR}/libnvinfer${CMAKE_SHARED_LIBRARY_SUFFIX})
set(DEPS ${DEPS} ${TENSORRT_LIB_DIR}/libnvinfer_plugin${CMAKE_SHARED_LIBRARY_SUFFIX})
endif()
set(DEPS ${DEPS} ${CUDA_LIB}/libcudart${CMAKE_SHARED_LIBRARY_SUFFIX})
else()
if(USE_TENSORRT)
set(DEPS ${DEPS} ${TENSORRT_LIB_DIR}/nvinfer${CMAKE_STATIC_LIBRARY_SUFFIX})
set(DEPS ${DEPS} ${TENSORRT_LIB_DIR}/nvinfer_plugin${CMAKE_STATIC_LIBRARY_SUFFIX})
if(${TENSORRT_MAJOR_VERSION} GREATER_EQUAL 7)
set(DEPS ${DEPS} ${TENSORRT_LIB_DIR}/myelin64_1${CMAKE_STATIC_LIBRARY_SUFFIX})
endif()
endif()
set(DEPS ${DEPS} ${CUDA_LIB}/cudart${CMAKE_STATIC_LIBRARY_SUFFIX} )
set(DEPS ${DEPS} ${CUDA_LIB}/cublas${CMAKE_STATIC_LIBRARY_SUFFIX} )
set(DEPS ${DEPS} ${CUDA_LIB}/cudnn${CMAKE_STATIC_LIBRARY_SUFFIX} )
endif()
endif()
if(WITH_ROCM)
if(NOT WIN32)
set(DEPS ${DEPS} ${ROCM_LIB}/libamdhip64${CMAKE_SHARED_LIBRARY_SUFFIX})
endif()
endif()
include_directories(/usr/local/include)
link_directories(/usr/local/lib)
#find_package(yaml-cpp REQUIRED)
set(YAML_CPP "D:\\lbq\\yaml-cpp-master") #7
set(YAML_CPP_INCLUDE_DIRS ${YAML_CPP}\\include) #8
set(YAML_CPP_LIBRARIES ${YAML_CPP}\\lib_mine) #9
include_directories(${YAML_CPP_INCLUDE_DIRS})
link_directories(${YAML_CPP_LIBRARIES})
#set(DEPS ${DEPS} "-lyaml-cpp")
find_package(OpenCV REQUIRED)
include_directories(${OpenCV_INCLUDE_DIRS})
set(DEPS ${DEPS} ${OpenCV_LIBS})
add_executable(${DEMO_NAME} src/${DEMO_NAME}.cc)
target_link_libraries(${DEMO_NAME} ${DEPS})
if(WIN32)
if(USE_TENSORRT)
add_custom_command(TARGET ${DEMO_NAME} POST_BUILD
COMMAND ${CMAKE_COMMAND} -E copy ${TENSORRT_LIB_DIR}/nvinfer${CMAKE_SHARED_LIBRARY_SUFFIX}
${CMAKE_BINARY_DIR}/${CMAKE_BUILD_TYPE}
COMMAND ${CMAKE_COMMAND} -E copy ${TENSORRT_LIB_DIR}/nvinfer_plugin${CMAKE_SHARED_LIBRARY_SUFFIX}
${CMAKE_BINARY_DIR}/${CMAKE_BUILD_TYPE}
)
if(${TENSORRT_MAJOR_VERSION} GREATER_EQUAL 7)
add_custom_command(TARGET ${DEMO_NAME} POST_BUILD
COMMAND ${CMAKE_COMMAND} -E copy ${TENSORRT_LIB_DIR}/myelin64_1${CMAKE_SHARED_LIBRARY_SUFFIX}
${CMAKE_BINARY_DIR}/${CMAKE_BUILD_TYPE})
endif()
endif()
if(WITH_MKL)
add_custom_command(TARGET ${DEMO_NAME} POST_BUILD
COMMAND ${CMAKE_COMMAND} -E copy ${MATH_LIB_PATH}/lib/mklml.dll ${CMAKE_BINARY_DIR}/Release
COMMAND ${CMAKE_COMMAND} -E copy ${MATH_LIB_PATH}/lib/libiomp5md.dll ${CMAKE_BINARY_DIR}/Release
COMMAND ${CMAKE_COMMAND} -E copy ${MKLDNN_PATH}/lib/mkldnn.dll ${CMAKE_BINARY_DIR}/Release
)
else()
add_custom_command(TARGET ${DEMO_NAME} POST_BUILD
COMMAND ${CMAKE_COMMAND} -E copy ${OPENBLAS_LIB_PATH}/lib/openblas.dll ${CMAKE_BINARY_DIR}/Release
)
endif()
if(NOT WITH_STATIC_LIB)
add_custom_command(TARGET ${DEMO_NAME} POST_BUILD
COMMAND ${CMAKE_COMMAND} -E copy "${PADDLE_LIB}/paddle/lib/paddle_inference.dll" ${CMAKE_BINARY_DIR}/${CMAKE_BUILD_TYPE}
)
endif()
endif()
message(STATUS " libraries: ${CUDA_LIB}")
message(STATUS " MATH_LIB_path: ${MATH_LIB}")
此处附上yaml在windows下的静态编译方式:
# 1.克隆并进入项目
git clone https://github.com/jbeder/yaml-cpp.git
cd yaml-cpp
# 2.启动 cmake-gui
cmake-gui
# 3.输入源码路径D:\lbq\yaml-cpp-master,和构建路径D:\lbq\yaml-cpp-master\mt_build,然后点击Configure进行配置
# 4.在弹窗里选择Visual Studio 16 2019
# 5.只勾选这2个即可,别的都不要勾选
YAML_CPP_BUILD_CONTRIB YAML_CPP_BUILD_TOOLS
# 6.再次点击Configure,再点击Generate,项目文件生成完成后,点击YAML_CPP.sln,然后熟悉的vs就出来了;在vs里选择Release和x64编译生成,便会在build/Release下出现yaml-cpp.lib,这个正是cmakelists.txt中第195行对应的库文件。
Step1: 下载PaddlePaddle C++ 预测库 fluid_inference
Step2:编译
配置完成之后选择以下三种方式编译,我选的是gpu方式编译,其中需要设置的是-DPADDLE_LIB,这个是在PaddleX提到过的PaddlePaddle C++ 预测库,请提前下载。
1)cpu
cmake .. -DDEMO_NAME=test_seg -DWITH_MKL=ON -DWITH_GPU=OFF -DUSE_TENSORRT=OFF -DWITH_STATIC_LIB=OFF -DPADDLE_LIB=D:\lbq\code\4_ocr\paddle_c_plus
2)gpu
cmake .. -DDEMO_NAME=test_seg -DWITH_MKL=ON -DWITH_GPU=ON -DUSE_TENSORRT=OFF -DWITH_STATIC_LIB=OFF -DPADDLE_LIB=D:\lbq\code\4_ocr\paddle_c_plus
3)tensorrt
cmake .. -DDEMO_NAME=test_seg -DWITH_MKL=ON -DWITH_GPU=ON -DUSE_TENSORRT=ON -DWITH_STATIC_LIB=OFF -DPADDLE_LIB=D:\lbq\code\4_ocr\paddle_c_plus
编译成功后,打开cpp_inference_demo.sln,在test_seg项目的属性->输入->附加依赖项 中添加一行’ yaml-cpp.lib’
然后便可生成test_seg.exe 。最后将python导出的模型软链接到Release目录下,并测试一张图片。
PaddleX和PaddleSeg都有tensorrt的解决方案,但是我调试的时候都出现了问题,目前找到了原因,正在进一步修复中。