diff --git a/.gitignore b/.gitignore index 09a9677680..e5d951cc20 100644 --- a/.gitignore +++ b/.gitignore @@ -131,3 +131,7 @@ work_dirs/ # the generated header files /tests/test_csrc/test_define.h + +# +!docs/zh_cn/build +!docs/en/build diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml index 029a2dcd69..1540e469de 100644 --- a/.pre-commit-config.yaml +++ b/.pre-commit-config.yaml @@ -46,3 +46,9 @@ repos: hooks: - id: docformatter args: ["--in-place", "--wrap-descriptions", "79"] + + - repo: https://github.com/open-mmlab/pre-commit-hooks + rev: v0.2.0 + hooks: + - id: check-copyright + args: ["csrc", "mmdeploy", "tests", "demo", "tools"] diff --git a/CMakeLists.txt b/CMakeLists.txt index 2f8f8e4672..3ae98d745b 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -11,7 +11,11 @@ set(CMAKE_CXX_STANDARD 17) set(CMAKE_ARCHIVE_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib) -set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib) +if (MSVC) + set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin) +else () + set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib) +endif () set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin) # options @@ -39,12 +43,20 @@ endif () # notice that ubsan has linker issues for ubuntu < 18.04, see # https://stackoverflow.com/questions/50024731/ld-unrecognized-option-push-state-no-as-needed if (MMDEPLOY_UBSAN_ENABLE) - add_compile_options($<$:-fsanitize=undefined>) - add_link_options(-fsanitize=undefined) + add_compile_options($<$:-fsanitize=undefined>) + add_link_options(-fsanitize=undefined) +endif () + +if (MSVC) + add_compile_options($<$:/diagnostics:classic>) + add_compile_options($<$:/Zc:preprocessor>) # /experimental:preprocessor on VS2017 + add_compile_options($<$:/wd4251>) +else () + add_compile_options($<$:-fvisibility=hidden>) endif () -include(${CMAKE_SOURCE_DIR}/cmake/common.cmake) -# set INTERFACE target to gather linked modules +include(${CMAKE_SOURCE_DIR}/cmake/MMDeploy.cmake) + add_library(MMDeployStaticModules INTERFACE) add_library(MMDeployDynamicModules INTERFACE) add_library(MMDeployLibs INTERFACE) @@ -52,25 +64,10 @@ add_library(MMDeployLibs INTERFACE) add_subdirectory(csrc) if (MMDEPLOY_BUILD_SDK) - # get static modules and dynamic modules from ${MMDeployStaticModules} and ${MMDeployDynamicModules}, respectively - set(STATIC_MODULES) - get_target_property(STATIC_MODULES MMDeployStaticModules INTERFACE_LINK_LIBRARIES) - get_target_list("${STATIC_MODULES}" FILTERED_MODULES) - set(MMDEPLOY_STATIC_MODULES "${FILTERED_MODULES}" CACHE STRING "MMDeploy's static modules") - message(STATUS "MMDEPLOY_STATIC_MODULES: ${MMDEPLOY_STATIC_MODULES}") - - set(DYNAMIC_MODULES) - get_target_property(DYNAMIC_MODULES MMDeployDynamicModules INTERFACE_LINK_LIBRARIES) - get_target_list("${DYNAMIC_MODULES}" FILTERED_MODULES) - set(MMDEPLOY_DYNAMIC_MODULES "${FILTERED_MODULES}" CACHE STRING "MMDeploy's dynamic modules") - message(STATUS "MMDEPLOY_DYNAMIC_MODULES: ${MMDEPLOY_DYNAMIC_MODULES}") - - # get libs from ${MMDeployLibs} - set(LIBS) - get_target_property(LIBS MMDeployLibs INTERFACE_LINK_LIBRARIES) - get_target_list("${LIBS}" FILTERED_LIBS) - set(MMDEPLOY_LIBS "${FILTERED_LIBS}" CACHE STRING "MMDeploy's libs that can be linked directly by application") - message(STATUS "MMDEPLOY_LIBS: ${MMDEPLOY_LIBS}") + install(TARGETS MMDeployStaticModules + MMDeployDynamicModules + MMDeployLibs + EXPORT MMDeployTargets) if (MMDEPLOY_BUILD_TEST) add_subdirectory(tests/test_csrc) @@ -78,13 +75,11 @@ if (MMDEPLOY_BUILD_SDK) if (MMDEPLOY_BUILD_SDK_PYTHON_API) add_subdirectory(csrc/apis/python) - endif() + endif () # export MMDeploy package install(EXPORT MMDeployTargets - # NAMESPACE mmdeploy:: FILE MMDeployTargets.cmake - #EXPORT_LINK_INTERFACE_LIBRARIES DESTINATION lib/cmake/MMDeploy) include(CMakePackageConfigHelpers) @@ -105,6 +100,8 @@ if (MMDEPLOY_BUILD_SDK) install(FILES ${CMAKE_CURRENT_BINARY_DIR}/MMDeployConfig.cmake ${CMAKE_CURRENT_BINARY_DIR}/MMDeployConfigVersion.cmake + ${CMAKE_CURRENT_SOURCE_DIR}/cmake/MMDeploy.cmake + ${CMAKE_CURRENT_SOURCE_DIR}/cmake/loader.cpp.in DESTINATION lib/cmake/MMDeploy ) diff --git a/README.md b/README.md index ab05a73ad0..37a10deff3 100644 --- a/README.md +++ b/README.md @@ -48,6 +48,7 @@ a part of the [OpenMMLab](https://openmmlab.com/) project. - [x] MMSegmentation - [x] MMEditing - [x] MMOCR + - [x] MMPose - **Multiple inference backends are available** @@ -111,20 +112,21 @@ If you find this project useful in your research, please consider cite: ## Projects in OpenMMLab - [MMCV](https://github.com/open-mmlab/mmcv): OpenMMLab foundational library for computer vision. -- [MIM](https://github.com/open-mmlab/mim): MIM Installs OpenMMLab Packages. +- [MIM](https://github.com/open-mmlab/mim): MIM installs OpenMMLab packages. - [MMClassification](https://github.com/open-mmlab/mmclassification): OpenMMLab image classification toolbox and benchmark. - [MMDetection](https://github.com/open-mmlab/mmdetection): OpenMMLab detection toolbox and benchmark. - [MMDetection3D](https://github.com/open-mmlab/mmdetection3d): OpenMMLab's next-generation platform for general 3D object detection. +- [MMRotate](https://github.com/open-mmlab/mmrotate): OpenMMLab rotated object detection toolbox and benchmark. - [MMSegmentation](https://github.com/open-mmlab/mmsegmentation): OpenMMLab semantic segmentation toolbox and benchmark. +- [MMOCR](https://github.com/open-mmlab/mmocr): OpenMMLab text detection, recognition, and understanding toolbox. +- [MMPose](https://github.com/open-mmlab/mmpose): OpenMMLab pose estimation toolbox and benchmark. +- [MMHuman3D](https://github.com/open-mmlab/mmhuman3d): OpenMMLab 3D human parametric model toolbox and benchmark. +- [MMSelfSup](https://github.com/open-mmlab/mmselfsup): OpenMMLab self-supervised learning toolbox and benchmark. +- [MMRazor](https://github.com/open-mmlab/mmrazor): OpenMMLab model compression toolbox and benchmark. +- [MMFewShot](https://github.com/open-mmlab/mmfewshot): OpenMMLab fewshot learning toolbox and benchmark. - [MMAction2](https://github.com/open-mmlab/mmaction2): OpenMMLab's next-generation action understanding toolbox and benchmark. - [MMTracking](https://github.com/open-mmlab/mmtracking): OpenMMLab video perception toolbox and benchmark. -- [MMPose](https://github.com/open-mmlab/mmpose): OpenMMLab pose estimation toolbox and benchmark. +- [MMFlow](https://github.com/open-mmlab/mmflow): OpenMMLab optical flow toolbox and benchmark. - [MMEditing](https://github.com/open-mmlab/mmediting): OpenMMLab image and video editing toolbox. -- [MMOCR](https://github.com/open-mmlab/mmocr): A Comprehensive Toolbox for Text Detection, Recognition and Understanding. - [MMGeneration](https://github.com/open-mmlab/mmgeneration): OpenMMLab image and video generative models toolbox. -- [MMFlow](https://github.com/open-mmlab/mmflow): OpenMMLab optical flow toolbox and benchmark. -- [MMFewShot](https://github.com/open-mmlab/mmfewshot): OpenMMLab FewShot Learning Toolbox and Benchmark. -- [MMHuman3D](https://github.com/open-mmlab/mmhuman3d): OpenMMLab Human Pose and Shape Estimation Toolbox and Benchmark. -- [MMSelfSup](https://github.com/open-mmlab/mmselfsup): OpenMMLab self-supervised learning Toolbox and Benchmark. -- [MMRazor](https://github.com/open-mmlab/mmrazor): OpenMMLab Model Compression Toolbox and Benchmark. -- [MMDeploy](https://github.com/open-mmlab/mmdeploy): OpenMMLab Model Deployment Framework. +- [MMDeploy](https://github.com/open-mmlab/mmdeploy): OpenMMLab model deployment framework. diff --git a/README_zh-CN.md b/README_zh-CN.md index 3ccf14e8af..12a6c4c5e3 100644 --- a/README_zh-CN.md +++ b/README_zh-CN.md @@ -46,6 +46,7 @@ MMDeploy 是一个开源深度学习模型部署工具箱，它是 [OpenMMLab](h - [x] MMSegmentation - [x] MMEditing - [x] MMOCR + - [x] MMPose - **支持多种推理后端** @@ -114,18 +115,19 @@ MMDeploy 是一个开源深度学习模型部署工具箱，它是 [OpenMMLab](h - [MMClassification](https://github.com/open-mmlab/mmclassification): OpenMMLab 图像分类工具箱 - [MMDetection](https://github.com/open-mmlab/mmdetection): OpenMMLab 目标检测工具箱 - [MMDetection3D](https://github.com/open-mmlab/mmdetection3d): OpenMMLab 新一代通用 3D 目标检测平台 +- [MMRotate](https://github.com/open-mmlab/mmrotate): OpenMMLab 旋转框检测工具箱与测试基准 - [MMSegmentation](https://github.com/open-mmlab/mmsegmentation): OpenMMLab 语义分割工具箱 -- [MMAction2](https://github.com/open-mmlab/mmaction2): OpenMMLab 新一代视频理解工具箱 -- [MMTracking](https://github.com/open-mmlab/mmtracking): OpenMMLab 一体化视频目标感知平台 -- [MMPose](https://github.com/open-mmlab/mmpose): OpenMMLab 姿态估计工具箱 -- [MMEditing](https://github.com/open-mmlab/mmediting): OpenMMLab 图像视频编辑工具箱 - [MMOCR](https://github.com/open-mmlab/mmocr): OpenMMLab 全流程文字检测识别理解工具包 -- [MMGeneration](https://github.com/open-mmlab/mmgeneration): OpenMMLab 图片视频生成模型工具箱 -- [MMFlow](https://github.com/open-mmlab/mmflow): OpenMMLab 光流估计工具箱与测试基准 -- [MMFewShot](https://github.com/open-mmlab/mmfewshot): OpenMMLab 少样本学习工具箱与测试基准 +- [MMPose](https://github.com/open-mmlab/mmpose): OpenMMLab 姿态估计工具箱 - [MMHuman3D](https://github.com/open-mmlab/mmhuman3d): OpenMMLab 人体参数化模型工具箱与测试基准 - [MMSelfSup](https://github.com/open-mmlab/mmselfsup): OpenMMLab 自监督学习工具箱与测试基准 - [MMRazor](https://github.com/open-mmlab/mmrazor): OpenMMLab 模型压缩工具箱与测试基准 +- [MMFewShot](https://github.com/open-mmlab/mmfewshot): OpenMMLab 少样本学习工具箱与测试基准 +- [MMAction2](https://github.com/open-mmlab/mmaction2): OpenMMLab 新一代视频理解工具箱 +- [MMTracking](https://github.com/open-mmlab/mmtracking): OpenMMLab 一体化视频目标感知平台 +- [MMFlow](https://github.com/open-mmlab/mmflow): OpenMMLab 光流估计工具箱与测试基准 +- [MMEditing](https://github.com/open-mmlab/mmediting): OpenMMLab 图像视频编辑工具箱 +- [MMGeneration](https://github.com/open-mmlab/mmgeneration): OpenMMLab 图片视频生成模型工具箱 - [MMDeploy](https://github.com/open-mmlab/mmdeploy): OpenMMLab 模型部署框架 ## 欢迎加入 OpenMMLab 社区 diff --git a/cmake/MMDeploy.cmake b/cmake/MMDeploy.cmake new file mode 100644 index 0000000000..086b45681f --- /dev/null +++ b/cmake/MMDeploy.cmake @@ -0,0 +1,151 @@ +# Copyright (c) OpenMMLab. All rights reserved. + +function (mmdeploy_export NAME) + set(_LIB_DIR lib) + if (MSVC) + set(_LIB_DIR bin) + endif () + install(TARGETS ${NAME} + EXPORT MMDeployTargets + ARCHIVE DESTINATION lib + LIBRARY DESTINATION ${_LIB_DIR} + RUNTIME DESTINATION bin) +endfunction () + + +function (mmdeploy_add_library NAME) + cmake_parse_arguments(_MMDEPLOY "EXCLUDE" "" "" ${ARGN}) + add_library(${NAME} ${_MMDEPLOY_UNPARSED_ARGUMENTS}) + target_compile_definitions(${NAME} PRIVATE -DMMDEPLOY_API_EXPORTS=1) + get_target_property(_TYPE ${NAME} TYPE) + if (_TYPE STREQUAL STATIC_LIBRARY) + set_target_properties(${NAME} PROPERTIES POSITION_INDEPENDENT_CODE 1) + elseif (_TYPE STREQUAL SHARED_LIBRARY) + else () + message(FATAL_ERROR "unsupported type: ${_TYPE}") + endif () + if (NOT _MMDEPLOY_EXCLUDE) + target_link_libraries(MMDeployLibs INTERFACE ${NAME}) + mmdeploy_export(${NAME}) + endif () +endfunction () + + +function (mmdeploy_add_module NAME) + # EXCLUDE: exclude from registering & exporting as SDK module + # LIBRARY: the module is also a library (add_libray with SHARED instead of MODULE) + cmake_parse_arguments(_MMDEPLOY "EXCLUDE;LIBRARY" "" "" ${ARGN}) + # search for add_library keywords + cmake_parse_arguments(_KW "STATIC;SHARED;MODULE" "" "" ${_MMDEPLOY_UNPARSED_ARGUMENTS}) + + set(_MAYBE_MODULE) + # no library type specified + if (NOT (_KW_STATIC OR _KW_SHARED OR _KW_MODULE)) + # shared but not marked as a library, build module library so that no .lib dependency + # will be generated for MSVC + if (MSVC AND BUILD_SHARED_LIBS AND NOT _MMDEPLOY_LIBRARY) + set(_MAYBE_MODULE MODULE) + endif () + endif () + + add_library(${NAME} ${_MAYBE_MODULE} ${_MMDEPLOY_UNPARSED_ARGUMENTS}) + + # automatically link mmdeploy::core if exists + if (TARGET mmdeploy::core) + target_link_libraries(${NAME} PRIVATE mmdeploy::core) + endif () + + # export public symbols when marked as a library + if (_MMDEPLOY_LIBRARY) + target_compile_definitions(${NAME} PRIVATE -DMMDEPLOY_API_EXPORTS=1) + endif () + + get_target_property(_TYPE ${NAME} TYPE) + if (_TYPE STREQUAL STATIC_LIBRARY) + set_target_properties(${NAME} PROPERTIES POSITION_INDEPENDENT_CODE 1) + if (MSVC) + target_link_options(${NAME} INTERFACE "/WHOLEARCHIVE:${NAME}") + endif () + # register static modules + if (NOT _MMDEPLOY_EXCLUDE) + target_link_libraries(MMDeployStaticModules INTERFACE ${NAME}) + endif () + elseif (_TYPE STREQUAL SHARED_LIBRARY OR _TYPE STREQUAL MODULE_LIBRARY) + # register dynamic modules + if (NOT _MMDEPLOY_EXCLUDE) + target_link_libraries(MMDeployDynamicModules INTERFACE ${NAME}) + endif () + else () + message(FATAL_ERROR "unsupported type: ${_TYPE}") + endif () + if (NOT _MMDEPLOY_EXCLUDE) + mmdeploy_export(${NAME}) + endif () +endfunction () + + +function (_mmdeploy_flatten_modules RETVAL) + set(_RETVAL) + foreach (ARG IN LISTS ARGN) + get_target_property(TYPE ${ARG} TYPE) + if (TYPE STREQUAL "INTERFACE_LIBRARY") + get_target_property(LIBS ${ARG} INTERFACE_LINK_LIBRARIES) + if (LIBS) + # pattern for 3.17+ + list(FILTER LIBS EXCLUDE REGEX "^::@") + # pattern for 3.13-3.16 + list(TRANSFORM LIBS REPLACE "(.+)::@.*" "\\1") + list(APPEND _RETVAL ${LIBS}) + endif () + else () + list(APPEND _RETVAL ${ARG}) + endif () + endforeach () + set(${RETVAL} ${_RETVAL} PARENT_SCOPE) +endfunction () + + +function (mmdeploy_load_static NAME) + if (MSVC) + target_link_libraries(${NAME} PRIVATE ${ARGN}) + else () + _mmdeploy_flatten_modules(_MODULE_LIST ${ARGN}) + target_link_libraries(${NAME} PRIVATE + -Wl,--whole-archive + ${_MODULE_LIST} + -Wl,--no-whole-archive) + endif () +endfunction () + +function (mmdeploy_load_dynamic NAME) + _mmdeploy_flatten_modules(_MODULE_LIST ${ARGN}) + if (MSVC) + if (NOT _MODULE_LIST) + return () + endif () + # MSVC has nothing like "-Wl,--no-as-needed ... -Wl,--as-needed", as a + # workaround we build a static module which loads the dynamic modules + set(_MODULE_STR ${_MODULE_LIST}) + list(TRANSFORM _MODULE_STR REPLACE "(.+)" "\"\\1\"") + string(JOIN ",\n " _MODULE_STR ${_MODULE_STR}) + set(_MMDEPLOY_DYNAMIC_MODULES ${_MODULE_STR}) + + set(_LOADER_NAME ${NAME}_loader) + + add_dependencies(${NAME} ${_MODULE_LIST}) + + set(_LOADER_PATH ${CMAKE_BINARY_DIR}/${_LOADER_NAME}.cpp) + # ! CMAKE_CURRENT_FUNCTION_LIST_DIR requires cmake 3.17+ + configure_file( + ${CMAKE_CURRENT_FUNCTION_LIST_DIR}/loader.cpp.in + ${_LOADER_PATH}) + + mmdeploy_add_module(${_LOADER_NAME} STATIC EXCLUDE ${_LOADER_PATH}) + mmdeploy_load_static(${NAME} ${_LOADER_NAME}) + else () + target_link_libraries(${NAME} PRIVATE + -Wl,--no-as-needed + ${_MODULE_LIST} + -Wl,--as-needed) + endif () +endfunction () diff --git a/cmake/MMDeployConfig.cmake.in b/cmake/MMDeployConfig.cmake.in index 7a23a9e153..4bd05489e4 100644 --- a/cmake/MMDeployConfig.cmake.in +++ b/cmake/MMDeployConfig.cmake.in @@ -2,23 +2,26 @@ cmake_minimum_required(VERSION 3.14) -include ("${CMAKE_CURRENT_LIST_DIR}/MMDeployTargets.cmake") +include("${CMAKE_CURRENT_LIST_DIR}/MMDeployTargets.cmake") set(MMDEPLOY_CODEBASES @MMDEPLOY_CODEBASES@) set(MMDEPLOY_TARGET_DEVICES @MMDEPLOY_TARGET_DEVICES@) set(MMDEPLOY_TARGET_BACKENDS @MMDEPLOY_TARGET_BACKENDS@) set(MMDEPLOY_BUILD_TYPE @CMAKE_BUILD_TYPE@) -set(MMDEPLOY_STATIC_MODULES @MMDEPLOY_STATIC_MODULES@) -set(MMDEPLOY_DYNAMIC_MODULES @MMDEPLOY_DYNAMIC_MODULES@) set(MMDEPLOY_BUILD_SHARED @BUILD_SHARED_LIBS@) -set(MMDEPLOY_LIBS @MMDEPLOY_LIBS@) if (NOT MMDEPLOY_BUILD_SHARED) if ("cuda" IN_LIST MMDEPLOY_TARGET_DEVICES) + set(CMAKE_CUDA_RUNTIME_LIBRARY Shared) + enable_language(CUDA) find_package(pplcv REQUIRED) endif () endif () -set(MMDeploy_LIBS ${MMDEPLOY_LIBS} - -Wl,--no-as-needed ${MMDEPLOY_DYNAMIC_MODULES} -Wl,--as-needed - -Wl,--whole-archive ${MMDEPLOY_STATIC_MODULES} -Wl,--no-whole-archive) +find_package(spdlog REQUIRED) +find_package(OpenCV REQUIRED) + +set(THREADS_PREFER_PTHREAD_FLAG ON) +find_package(Threads REQUIRED) + +include("${CMAKE_CURRENT_LIST_DIR}/MMDeploy.cmake") diff --git a/cmake/common.cmake b/cmake/common.cmake deleted file mode 100644 index fae162bfe2..0000000000 --- a/cmake/common.cmake +++ /dev/null @@ -1,108 +0,0 @@ -# Copyright (c) OpenMMLab. All rights reserved. -function(set_targets PROJECT_NAME OBJ_TARGET STATIC_TARGET SHARED_TARGET) - set(${OBJ_TARGET} ${PROJECT_NAME}_obj PARENT_SCOPE) - set(${STATIC_TARGET} ${PROJECT_NAME}_static PARENT_SCOPE) - set(${SHARED_TARGET} ${PROJECT_NAME} PARENT_SCOPE) -endfunction() - -function(install_targets TARGET_NAMES) - foreach (TARGET_NAME ${TARGET_NAMES}) - install(TARGETS ${TARGET_NAME} - ARCHIVE DESTINATION lib - LIBRARY DESTINATION lib - RUNTIME DESTINATION bin - ) - endforeach () -endfunction() - -function(build_target TARGET_NAME TARGET_SRCS) - add_library(${TARGET_NAME} ${TARGET_SRCS}) - set_target_properties(${TARGET_NAME} PROPERTIES POSITION_INDEPENDENT_CODE 1) -endfunction() - -# When the object target ${TARGET_NAME} has more than one source file, -# "${SRCS_VARIABLE}" MUST be passed to ${TARGET_SRCS}. The quotation marks CANNOT be dismissed. -function(build_object_target TARGET_NAME TARGET_SRCS) - add_library(${TARGET_NAME} OBJECT) - target_sources(${TARGET_NAME} PRIVATE ${TARGET_SRCS}) - set_target_properties(${TARGET_NAME} PROPERTIES POSITION_INDEPENDENT_CODE 1) -endfunction() - -function(build_static_target TARGET_NAME OBJECT_TARGET LINK_TYPE) - add_library(${TARGET_NAME} STATIC $) - if (${LINK_TYPE} STREQUAL "PRIVATE") - target_link_libraries(${TARGET_NAME} PRIVATE ${OBJECT_TARGET}) - elseif (${LINK_TYPE} STREQUAL "PUBLIC") - target_link_libraries(${TARGET_NAME} PUBLIC ${OBJECT_TARGET}) - elseif (${LINK_TYPE} STREQUAL "INTERFACE") - target_link_libraries(${TARGET_NAME} INTERFACE ${OBJECT_TARGET}) - elseif (${LINK_TYPE} STREQUAL "") - target_link_libraries(${TARGET_NAME} ${OBJECT_TARGET}) - else () - message(FATAL_ERROR "Incorrect link type: ${LINK_TYPE}") - endif () -endfunction() - -function(build_shared_target TARGET_NAME OBJECT_TARGET LINK_TYPE) - add_library(${TARGET_NAME} SHARED $) - if (${LINK_TYPE} STREQUAL "PRIVATE") - target_link_libraries(${TARGET_NAME} PRIVATE ${OBJECT_TARGET}) - elseif (${LINK_TYPE} STREQUAL "PUBLIC") - target_link_libraries(${TARGET_NAME} PUBLIC ${OBJECT_TARGET}) - elseif (${LINK_TYPE} STREQUAL "INTERFACE") - target_link_libraries(${TARGET_NAME} INTERFACE ${OBJECT_TARGET}) - elseif (${LINK_TYPE} STREQUAL "") - target_link_libraries(${TARGET_NAME} ${OBJECT_TARGET}) - else () - message(FATAL_ERROR "Incorrect link type: ${LINK_TYPE}") - endif () -endfunction() - -function(build_module_target TARGET_NAME OBJECT_TARGET LINK_TYPE) - add_library(${TARGET_NAME} MODULE $) - if (${LINK_TYPE} STREQUAL "PRIVATE") - target_link_libraries(${TARGET_NAME} PRIVATE ${OBJECT_TARGET}) - elseif (${LINK_TYPE} STREQUAL "PUBLIC") - target_link_libraries(${TARGET_NAME} PUBLIC ${OBJECT_TARGET}) - elseif (${LINK_TYPE} STREQUAL "INTERFACE") - target_link_libraries(${TARGET_NAME} INTERFACE ${OBJECT_TARGET}) - elseif (${LINK_TYPE} STREQUAL "") - target_link_libraries(${TARGET_NAME} ${OBJECT_TARGET}) - else () - message(FATAL_ERROR "Incorrect link type: ${LINK_TYPE}") - endif () -endfunction() - - -function(export_target TARGET_NAME) - target_link_libraries(MMDeployLibs INTERFACE ${TARGET_NAME}) - install(TARGETS ${TARGET_NAME} - EXPORT MMDeployTargets - ARCHIVE DESTINATION lib - LIBRARY DESTINATION lib - ) -endfunction() - -function(export_module TARGET_NAME) - get_target_property(TARGET_TYPE ${TARGET_NAME} TYPE) - if (${TARGET_TYPE} STREQUAL "STATIC_LIBRARY") - target_link_libraries(MMDeployStaticModules INTERFACE ${TARGET_NAME}) - elseif (${TARGET_TYPE} STREQUAL "SHARED_LIBRARY") - target_link_libraries(MMDeployDynamicModules INTERFACE ${TARGET_NAME}) - endif () - install(TARGETS ${TARGET_NAME} - EXPORT MMDeployTargets - ARCHIVE DESTINATION lib - LIBRARY DESTINATION lib - ) -endfunction() - -function(get_target_list INPUT_TARGETS OUTPUT_TARGETS) - set(FILTERED_TARGETS) - foreach (INPUT_TARGET IN LISTS INPUT_TARGETS) - if (TARGET ${INPUT_TARGET}) - list(APPEND FILTERED_TARGETS ${INPUT_TARGET}) - endif() - endforeach () - set(${OUTPUT_TARGETS} "${FILTERED_TARGETS}" PARENT_SCOPE) -endfunction() diff --git a/cmake/cuda.cmake b/cmake/cuda.cmake index 158e542e16..9fe42596c4 100644 --- a/cmake/cuda.cmake +++ b/cmake/cuda.cmake @@ -23,35 +23,41 @@ else () set(CMAKE_CUDA_HOST_COMPILER ${CMAKE_CXX_COMPILER}) set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -Xcompiler=-fPIC,-Wall,-fvisibility=hidden") - set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -Xcompiler=-fno-gnu-unique") + if (CMAKE_CXX_COMPILER_ID MATCHES "GNU") + set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -Xcompiler=-fno-gnu-unique") + endif () endif () enable_language(CUDA) # set virtual compute architecture and real ones set(_NVCC_FLAGS) -set(_NVCC_FLAGS "${_NVCC_FLAGS} -gencode arch=compute_52,code=sm_52") -if (CUDA_VERSION_MAJOR VERSION_GREATER_EQUAL "8") - set(_NVCC_FLAGS "${_NVCC_FLAGS} -gencode arch=compute_60,code=sm_60") - set(_NVCC_FLAGS "${_NVCC_FLAGS} -gencode arch=compute_61,code=sm_61") -endif () -if (CUDA_VERSION_MAJOR VERSION_GREATER_EQUAL "9") - set(_NVCC_FLAGS "${_NVCC_FLAGS} -gencode arch=compute_70,code=sm_70") -endif () -if (CUDA_VERSION_MAJOR VERSION_GREATER_EQUAL "10") - set(_NVCC_FLAGS "${_NVCC_FLAGS} -gencode arch=compute_72,code=sm_72") - set(_NVCC_FLAGS "${_NVCC_FLAGS} -gencode arch=compute_75,code=sm_75") -endif () -if (CUDA_VERSION_MAJOR VERSION_GREATER_EQUAL "11") - set(_NVCC_FLAGS "${_NVCC_FLAGS} -gencode arch=compute_80,code=sm_80") - if (CUDA_VERSION_MINOR VERSION_GREATER_EQUAL "1") - # cuda doesn't support `sm_86` until version 11.1 - set(_NVCC_FLAGS "${_NVCC_FLAGS} -gencode arch=compute_86,code=sm_86") +if (NOT CMAKE_CUDA_ARCHITECTURES) + set(_NVCC_FLAGS "${_NVCC_FLAGS} -gencode arch=compute_52,code=sm_52") + if (CUDA_VERSION_MAJOR VERSION_GREATER_EQUAL "8") + set(_NVCC_FLAGS "${_NVCC_FLAGS} -gencode arch=compute_60,code=sm_60") + set(_NVCC_FLAGS "${_NVCC_FLAGS} -gencode arch=compute_61,code=sm_61") + endif () + if (CUDA_VERSION_MAJOR VERSION_GREATER_EQUAL "9") + set(_NVCC_FLAGS "${_NVCC_FLAGS} -gencode arch=compute_70,code=sm_70") + endif () + if (CUDA_VERSION_MAJOR VERSION_GREATER_EQUAL "10") + set(_NVCC_FLAGS "${_NVCC_FLAGS} -gencode arch=compute_72,code=sm_72") + set(_NVCC_FLAGS "${_NVCC_FLAGS} -gencode arch=compute_75,code=sm_75") + endif () + if (CUDA_VERSION_MAJOR VERSION_GREATER_EQUAL "11") + set(_NVCC_FLAGS "${_NVCC_FLAGS} -gencode arch=compute_80,code=sm_80") + if (CUDA_VERSION_MINOR VERSION_GREATER_EQUAL "1") + # cuda doesn't support `sm_86` until version 11.1 + set(_NVCC_FLAGS "${_NVCC_FLAGS} -gencode arch=compute_86,code=sm_86") + endif () endif () endif () set(CUDA_NVCC_FLAGS_DEBUG "-g -O0") set(CUDA_NVCC_FLAGS_RELEASE "-O3") set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS}") -set(CMAKE_CUDA_STANDARD 14) +if (NOT MSVC) + set(CMAKE_CUDA_STANDARD 14) +endif () set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} ${CUDA_NVCC_FLAGS} ${_NVCC_FLAGS}") diff --git a/cmake/loader.cpp.in b/cmake/loader.cpp.in new file mode 100644 index 0000000000..6627d6e2e7 --- /dev/null +++ b/cmake/loader.cpp.in @@ -0,0 +1,39 @@ +// Copyright (c) OpenMMLab. All rights reserved. + +#include + +#include + +namespace mmdeploy { +namespace { + +void* mmdeploy_load_library(const char* name) { + fprintf(stderr, "loading %s ...\n", name); + auto handle = LoadLibraryA(name); + if (!handle) { + fprintf(stderr, "failed to load library %s\n", name); + return nullptr; + } + return handle; +} + +// clang-format off + +class Loader { + public: + Loader() { + const char* modules[] = { + @_MMDEPLOY_DYNAMIC_MODULES@ + }; + for (const auto name : modules) { + mmdeploy_load_library(name); + } + } +}; + +// clang-format on + +static Loader loader; + +} // namespace +} // namespace mmdeploy diff --git a/configs/_base_/backends/torchscript.py b/configs/_base_/backends/torchscript.py new file mode 100644 index 0000000000..754fe488ac --- /dev/null +++ b/configs/_base_/backends/torchscript.py @@ -0,0 +1 @@ +backend_config = dict(type='torchscript') diff --git a/configs/_base_/torchscript_config.py b/configs/_base_/torchscript_config.py new file mode 100644 index 0000000000..b16a2e871d --- /dev/null +++ b/configs/_base_/torchscript_config.py @@ -0,0 +1,6 @@ +ir_config = dict( + type='torchscript', + save_file='end2end.pt', + input_names=['input'], + output_names=['output'], + input_shape=None) diff --git a/configs/mmcls/classification_torchscript.py b/configs/mmcls/classification_torchscript.py new file mode 100644 index 0000000000..559fd25c38 --- /dev/null +++ b/configs/mmcls/classification_torchscript.py @@ -0,0 +1,6 @@ +_base_ = [ + '../_base_/torchscript_config.py', '../_base_/backends/torchscript.py' +] + +ir_config = dict(input_shape=None) +codebase_config = dict(type='mmcls', task='Classification') diff --git a/configs/mmdet/_base_/base_instance-seg_torchscript.py b/configs/mmdet/_base_/base_instance-seg_torchscript.py new file mode 100644 index 0000000000..68eee07e72 --- /dev/null +++ b/configs/mmdet/_base_/base_instance-seg_torchscript.py @@ -0,0 +1,4 @@ +_base_ = ['./base_torchscript.py'] + +ir_config = dict(output_names=['dets', 'labels', 'masks']) +codebase_config = dict(post_processing=dict(export_postprocess_mask=False)) diff --git a/configs/mmdet/_base_/base_openvino_dynamic-300x300.py b/configs/mmdet/_base_/base_openvino_dynamic-300x300.py new file mode 100644 index 0000000000..ae1116df91 --- /dev/null +++ b/configs/mmdet/_base_/base_openvino_dynamic-300x300.py @@ -0,0 +1,6 @@ +_base_ = ['./base_dynamic.py', '../../_base_/backends/openvino.py'] + +onnx_config = dict(input_shape=None) + +backend_config = dict( + model_inputs=[dict(opt_shapes=dict(input=[1, 3, 300, 300]))]) diff --git a/configs/mmdet/_base_/base_torchscript.py b/configs/mmdet/_base_/base_torchscript.py new file mode 100644 index 0000000000..7e0ecc8ae5 --- /dev/null +++ b/configs/mmdet/_base_/base_torchscript.py @@ -0,0 +1,16 @@ +_base_ = ['../../_base_/torchscript_config.py'] + +ir_config = dict(output_names=['dets', 'labels']) +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, + )) diff --git a/configs/mmdet/detection/detection_openvino_dynamic-300x300.py b/configs/mmdet/detection/detection_openvino_dynamic-300x300.py new file mode 100644 index 0000000000..1df7d12114 --- /dev/null +++ b/configs/mmdet/detection/detection_openvino_dynamic-300x300.py @@ -0,0 +1 @@ +_base_ = ['../_base_/base_openvino_dynamic.py'] diff --git a/configs/mmdet/detection/detection_torchscript.py b/configs/mmdet/detection/detection_torchscript.py new file mode 100644 index 0000000000..69bfbd9c7f --- /dev/null +++ b/configs/mmdet/detection/detection_torchscript.py @@ -0,0 +1,3 @@ +_base_ = [ + '../_base_/base_torchscript.py', '../../_base_/backends/torchscript.py' +] diff --git a/configs/mmdet/instance-seg/instance-seg_torchscript.py b/configs/mmdet/instance-seg/instance-seg_torchscript.py new file mode 100644 index 0000000000..ba8ad7e041 --- /dev/null +++ b/configs/mmdet/instance-seg/instance-seg_torchscript.py @@ -0,0 +1,4 @@ +_base_ = [ + '../_base_/base_instance-seg_torchscript.py', + '../../_base_/backends/torchscript.py' +] diff --git a/configs/mmedit/super-resolution/super-resolution_torchscript.py b/configs/mmedit/super-resolution/super-resolution_torchscript.py new file mode 100644 index 0000000000..8ebef20e34 --- /dev/null +++ b/configs/mmedit/super-resolution/super-resolution_torchscript.py @@ -0,0 +1,7 @@ +_base_ = [ + '../../_base_/torchscript_config.py', + '../../_base_/backends/torchscript.py' +] + +ir_config = dict(input_shape=None) +codebase_config = dict(type='mmedit', task='SuperResolution') diff --git a/configs/mmocr/text-detection/text-detection_torchscript.py b/configs/mmocr/text-detection/text-detection_torchscript.py new file mode 100644 index 0000000000..48a27d44eb --- /dev/null +++ b/configs/mmocr/text-detection/text-detection_torchscript.py @@ -0,0 +1,7 @@ +_base_ = [ + '../../_base_/torchscript_config.py', + '../../_base_/backends/torchscript.py' +] + +ir_config = dict(input_shape=None) +codebase_config = dict(type='mmocr', task='TextDetection') diff --git a/configs/mmocr/text-recognition/text-recognition_torchscript.py b/configs/mmocr/text-recognition/text-recognition_torchscript.py new file mode 100644 index 0000000000..14e9112e49 --- /dev/null +++ b/configs/mmocr/text-recognition/text-recognition_torchscript.py @@ -0,0 +1,7 @@ +_base_ = [ + '../../_base_/torchscript_config.py', + '../../_base_/backends/torchscript.py' +] + +ir_config = dict(input_shape=None) +codebase_config = dict(type='mmocr', task='TextRecognition') diff --git a/configs/mmseg/segmentation_torchscript.py b/configs/mmseg/segmentation_torchscript.py new file mode 100644 index 0000000000..665f308ecc --- /dev/null +++ b/configs/mmseg/segmentation_torchscript.py @@ -0,0 +1,6 @@ +_base_ = [ + '../_base_/torchscript_config.py', '../_base_/backends/torchscript.py' +] + +ir_config = dict(input_shape=None) +codebase_config = dict(type='mmseg', task='Segmentation') diff --git a/csrc/CMakeLists.txt b/csrc/CMakeLists.txt index 889e54bb5d..b14c81c136 100644 --- a/csrc/CMakeLists.txt +++ b/csrc/CMakeLists.txt @@ -4,6 +4,7 @@ add_subdirectory(backend_ops) if (MMDEPLOY_BUILD_SDK) add_subdirectory(core) + add_subdirectory(utils) add_subdirectory(archive) add_subdirectory(device) add_subdirectory(graph) diff --git a/csrc/apis/c/CMakeLists.txt b/csrc/apis/c/CMakeLists.txt index 81da0a3fb1..f1809995bb 100644 --- a/csrc/apis/c/CMakeLists.txt +++ b/csrc/apis/c/CMakeLists.txt @@ -2,7 +2,7 @@ cmake_minimum_required(VERSION 3.14) project(capis) -include(${CMAKE_SOURCE_DIR}/cmake/common.cmake) +include(${CMAKE_SOURCE_DIR}/cmake/MMDeploy.cmake) if ("all" IN_LIST MMDEPLOY_CODEBASES) set(TASK_LIST "classifier;detector;segmentor;text_detector;text_recognizer;restorer;model") @@ -28,16 +28,13 @@ endif () foreach (TASK ${TASK_LIST}) set(TARGET_NAME mmdeploy_${TASK}) - build_target(${TARGET_NAME} ${TASK}.cpp) + mmdeploy_add_library(${TARGET_NAME} ${TASK}.cpp) target_link_libraries(${TARGET_NAME} PRIVATE mmdeploy::core) target_include_directories(${TARGET_NAME} PUBLIC $ $) - export_target(${TARGET_NAME}) - install(FILES ${CMAKE_CURRENT_SOURCE_DIR}/${TASK}.h DESTINATION include/c) - endforeach () install(FILES ${CMAKE_CURRENT_SOURCE_DIR}/common.h diff --git a/csrc/apis/c/classifier.cpp b/csrc/apis/c/classifier.cpp index 9236f5eae0..ecdfaafc87 100644 --- a/csrc/apis/c/classifier.cpp +++ b/csrc/apis/c/classifier.cpp @@ -55,28 +55,28 @@ int mmdeploy_classifier_create_impl(ModelType&& m, const char* device_name, int return MM_SUCCESS; } catch (const std::exception& e) { - ERROR("exception caught: {}", e.what()); + MMDEPLOY_ERROR("exception caught: {}", e.what()); } catch (...) { - ERROR("unknown exception caught"); + MMDEPLOY_ERROR("unknown exception caught"); } return MM_E_FAIL; } } // namespace -MM_SDK_API int mmdeploy_classifier_create(mm_model_t model, const char* device_name, int device_id, - mm_handle_t* handle) { +int mmdeploy_classifier_create(mm_model_t model, const char* device_name, int device_id, + mm_handle_t* handle) { return mmdeploy_classifier_create_impl(*static_cast(model), device_name, device_id, handle); } -MM_SDK_API int mmdeploy_classifier_create_by_path(const char* model_path, const char* device_name, - int device_id, mm_handle_t* handle) { +int mmdeploy_classifier_create_by_path(const char* model_path, const char* device_name, + int device_id, mm_handle_t* handle) { return mmdeploy_classifier_create_impl(model_path, device_name, device_id, handle); } -MM_SDK_API int mmdeploy_classifier_apply(mm_handle_t handle, const mm_mat_t* mats, int mat_count, - mm_class_t** results, int** result_count) { +int mmdeploy_classifier_apply(mm_handle_t handle, const mm_mat_t* mats, int mat_count, + mm_class_t** results, int** result_count) { if (handle == nullptr || mats == nullptr || mat_count == 0) { return MM_E_INVALID_ARG; } @@ -92,7 +92,7 @@ MM_SDK_API int mmdeploy_classifier_apply(mm_handle_t handle, const mm_mat_t* mat } auto output = classifier->Run(std::move(input)).value().front(); - DEBUG("output: {}", output); + MMDEPLOY_DEBUG("output: {}", output); auto classify_outputs = from_value>(output); @@ -124,20 +124,19 @@ MM_SDK_API int mmdeploy_classifier_apply(mm_handle_t handle, const mm_mat_t* mat return MM_SUCCESS; } catch (const std::exception& e) { - ERROR("exception caught: {}", e.what()); + MMDEPLOY_ERROR("exception caught: {}", e.what()); } catch (...) { - ERROR("unknown exception caught"); + MMDEPLOY_ERROR("unknown exception caught"); } return MM_E_FAIL; } -MM_SDK_API void mmdeploy_classifier_release_result(mm_class_t* results, const int* result_count, - int count) { +void mmdeploy_classifier_release_result(mm_class_t* results, const int* result_count, int count) { delete[] results; delete[] result_count; } -MM_SDK_API void mmdeploy_classifier_destroy(mm_handle_t handle) { +void mmdeploy_classifier_destroy(mm_handle_t handle) { if (handle != nullptr) { auto classifier = static_cast(handle); delete classifier; diff --git a/csrc/apis/c/classifier.h b/csrc/apis/c/classifier.h index 6834b8e401..a2209792ba 100644 --- a/csrc/apis/c/classifier.h +++ b/csrc/apis/c/classifier.h @@ -10,6 +10,10 @@ #include "common.h" +#ifdef __cplusplus +extern "C" { +#endif + typedef struct mm_class_t { int label_id; float score; @@ -25,8 +29,8 @@ typedef struct mm_class_t { * by \ref mmdeploy_classifier_destroy * @return status of creating classifier's handle */ -MM_SDK_API int mmdeploy_classifier_create(mm_model_t model, const char* device_name, int device_id, - mm_handle_t* handle); +MMDEPLOY_API int mmdeploy_classifier_create(mm_model_t model, const char* device_name, + int device_id, mm_handle_t* handle); /** * @brief Create classifier's handle @@ -37,8 +41,8 @@ MM_SDK_API int mmdeploy_classifier_create(mm_model_t model, const char* device_n * by \ref mmdeploy_classifier_destroy * @return status of creating classifier's handle */ -MM_SDK_API int mmdeploy_classifier_create_by_path(const char* model_path, const char* device_name, - int device_id, mm_handle_t* handle); +MMDEPLOY_API int mmdeploy_classifier_create_by_path(const char* model_path, const char* device_name, + int device_id, mm_handle_t* handle); /** * @brief Use classifier created by \ref mmdeploy_classifier_create_by_path to get label @@ -53,8 +57,8 @@ MM_SDK_API int mmdeploy_classifier_create_by_path(const char* model_path, const * mmdeploy_classifier_release_result * @return status of inference */ -MM_SDK_API int mmdeploy_classifier_apply(mm_handle_t handle, const mm_mat_t* mats, int mat_count, - mm_class_t** results, int** result_count); +MMDEPLOY_API int mmdeploy_classifier_apply(mm_handle_t handle, const mm_mat_t* mats, int mat_count, + mm_class_t** results, int** result_count); /** * @brief Release the inference result buffer created \ref mmdeploy_classifier_apply @@ -62,13 +66,17 @@ MM_SDK_API int mmdeploy_classifier_apply(mm_handle_t handle, const mm_mat_t* mat * @param[in] result_count \p results size buffer * @param[in] count length of \p result_count */ -MM_SDK_API void mmdeploy_classifier_release_result(mm_class_t* results, const int* result_count, - int count); +MMDEPLOY_API void mmdeploy_classifier_release_result(mm_class_t* results, const int* result_count, + int count); /** * @brief Destroy classifier's handle * @param[in] handle classifier's handle created by \ref mmdeploy_classifier_create_by_path */ -MM_SDK_API void mmdeploy_classifier_destroy(mm_handle_t handle); +MMDEPLOY_API void mmdeploy_classifier_destroy(mm_handle_t handle); + +#ifdef __cplusplus +} +#endif #endif // MMDEPLOY_CLASSIFIER_H diff --git a/csrc/apis/c/common.h b/csrc/apis/c/common.h index 1809f77727..dc82d44292 100644 --- a/csrc/apis/c/common.h +++ b/csrc/apis/c/common.h @@ -3,9 +3,23 @@ #ifndef MMDEPLOY_COMMON_H #define MMDEPLOY_COMMON_H -#include +#include -#define MM_SDK_API +#ifndef MMDEPLOY_EXPORT +#ifdef _MSC_VER +#define MMDEPLOY_EXPORT __declspec(dllexport) +#else +#define MMDEPLOY_EXPORT __attribute__((visibility("default"))) +#endif +#endif + +#ifndef MMDEPLOY_API +#ifdef MMDEPLOY_API_EXPORTS +#define MMDEPLOY_API MMDEPLOY_EXPORT +#else +#define MMDEPLOY_API +#endif +#endif // clang-format off diff --git a/csrc/apis/c/detector.cpp b/csrc/apis/c/detector.cpp index 190b8bf7d5..4dbb573f96 100644 --- a/csrc/apis/c/detector.cpp +++ b/csrc/apis/c/detector.cpp @@ -55,27 +55,27 @@ int mmdeploy_detector_create_impl(ModelType&& m, const char* device_name, int de return MM_SUCCESS; } catch (const std::exception& e) { - ERROR("exception caught: {}", e.what()); + MMDEPLOY_ERROR("exception caught: {}", e.what()); } catch (...) { - ERROR("unknown exception caught"); + MMDEPLOY_ERROR("unknown exception caught"); } return MM_E_FAIL; } } // namespace -MM_SDK_API int mmdeploy_detector_create(mm_model_t model, const char* device_name, int device_id, - mm_handle_t* handle) { +int mmdeploy_detector_create(mm_model_t model, const char* device_name, int device_id, + mm_handle_t* handle) { return mmdeploy_detector_create_impl(*static_cast(model), device_name, device_id, handle); } -MM_SDK_API int mmdeploy_detector_create_by_path(const char* model_path, const char* device_name, - int device_id, mm_handle_t* handle) { +int mmdeploy_detector_create_by_path(const char* model_path, const char* device_name, int device_id, + mm_handle_t* handle) { return mmdeploy_detector_create_impl(model_path, device_name, device_id, handle); } -MM_SDK_API int mmdeploy_detector_apply(mm_handle_t handle, const mm_mat_t* mats, int mat_count, - mm_detect_t** results, int** result_count) { +int mmdeploy_detector_apply(mm_handle_t handle, const mm_mat_t* mats, int mat_count, + mm_detect_t** results, int** result_count) { if (handle == nullptr || mats == nullptr || mat_count == 0) { return MM_E_INVALID_ARG; } @@ -91,7 +91,7 @@ MM_SDK_API int mmdeploy_detector_apply(mm_handle_t handle, const mm_mat_t* mats, } auto output = detector->Run(std::move(input)).value().front(); - DEBUG("output: {}", output); + MMDEPLOY_DEBUG("output: {}", output); auto detector_outputs = from_value>(output); @@ -142,15 +142,14 @@ MM_SDK_API int mmdeploy_detector_apply(mm_handle_t handle, const mm_mat_t* mats, return MM_SUCCESS; } catch (const std::exception& e) { - ERROR("exception caught: {}", e.what()); + MMDEPLOY_ERROR("exception caught: {}", e.what()); } catch (...) { - ERROR("unknown exception caught"); + MMDEPLOY_ERROR("unknown exception caught"); } return MM_E_FAIL; } -MM_SDK_API void mmdeploy_detector_release_result(mm_detect_t* results, const int* result_count, - int count) { +void mmdeploy_detector_release_result(mm_detect_t* results, const int* result_count, int count) { auto result_ptr = results; for (int i = 0; i < count; ++i) { for (int j = 0; j < result_count[i]; ++j, ++result_ptr) { @@ -164,7 +163,7 @@ MM_SDK_API void mmdeploy_detector_release_result(mm_detect_t* results, const int delete[] result_count; } -MM_SDK_API void mmdeploy_detector_destroy(mm_handle_t handle) { +void mmdeploy_detector_destroy(mm_handle_t handle) { if (handle != nullptr) { auto detector = static_cast(handle); delete detector; diff --git a/csrc/apis/c/detector.h b/csrc/apis/c/detector.h index 59689dd0de..bfcf0a8acb 100644 --- a/csrc/apis/c/detector.h +++ b/csrc/apis/c/detector.h @@ -10,6 +10,10 @@ #include "common.h" +#ifdef __cplusplus +extern "C" { +#endif + typedef struct mm_instance_mask_t { char* data; int height; @@ -32,8 +36,8 @@ typedef struct mm_detect_t { * @param[out] handle instance of a detector * @return status of creating detector's handle */ -MM_SDK_API int mmdeploy_detector_create(mm_model_t model, const char* device_name, int device_id, - mm_handle_t* handle); +MMDEPLOY_API int mmdeploy_detector_create(mm_model_t model, const char* device_name, int device_id, + mm_handle_t* handle); /** * @brief Create detector's handle @@ -43,8 +47,8 @@ MM_SDK_API int mmdeploy_detector_create(mm_model_t model, const char* device_nam * @param[out] handle instance of a detector * @return status of creating detector's handle */ -MM_SDK_API int mmdeploy_detector_create_by_path(const char* model_path, const char* device_name, - int device_id, mm_handle_t* handle); +MMDEPLOY_API int mmdeploy_detector_create_by_path(const char* model_path, const char* device_name, + int device_id, mm_handle_t* handle); /** * @brief Apply detector to batch images and get their inference results @@ -58,21 +62,25 @@ MM_SDK_API int mmdeploy_detector_create_by_path(const char* model_path, const ch * mmdeploy_detector_release_result * @return status of inference */ -MM_SDK_API int mmdeploy_detector_apply(mm_handle_t handle, const mm_mat_t* mats, int mat_count, - mm_detect_t** results, int** result_count); +MMDEPLOY_API int mmdeploy_detector_apply(mm_handle_t handle, const mm_mat_t* mats, int mat_count, + mm_detect_t** results, int** result_count); /** @brief Release the inference result buffer created by \ref mmdeploy_detector_apply * @param[in] results detection results buffer * @param[in] result_count \p results size buffer * @param[in] count length of \p result_count */ -MM_SDK_API void mmdeploy_detector_release_result(mm_detect_t* results, const int* result_count, - int count); +MMDEPLOY_API void mmdeploy_detector_release_result(mm_detect_t* results, const int* result_count, + int count); /** * @brief Destroy detector's handle * @param[in] handle detector's handle created by \ref mmdeploy_detector_create_by_path */ -MM_SDK_API void mmdeploy_detector_destroy(mm_handle_t handle); +MMDEPLOY_API void mmdeploy_detector_destroy(mm_handle_t handle); + +#ifdef __cplusplus +} +#endif #endif // MMDEPLOY_DETECTOR_H diff --git a/csrc/apis/c/handle.h b/csrc/apis/c/handle.h index 2af9c0da58..4223452995 100644 --- a/csrc/apis/c/handle.h +++ b/csrc/apis/c/handle.h @@ -20,12 +20,12 @@ class Handle { config["context"].update({{"device", device_}, {"stream", stream_}}); auto creator = Registry::Get().GetCreator("Pipeline"); if (!creator) { - ERROR("failed to find Pipeline creator"); + MMDEPLOY_ERROR("failed to find Pipeline creator"); throw_exception(eEntryNotFound); } pipeline_ = creator->Create(config); if (!pipeline_) { - ERROR("create pipeline failed"); + MMDEPLOY_ERROR("create pipeline failed"); throw_exception(eFail); } pipeline_->Build(graph_); diff --git a/csrc/apis/c/model.cpp b/csrc/apis/c/model.cpp index 9834071c70..5101b92a47 100644 --- a/csrc/apis/c/model.cpp +++ b/csrc/apis/c/model.cpp @@ -1,11 +1,13 @@ // Copyright (c) OpenMMLab. All rights reserved. -#include "core/model.h" +// clang-format off +#include "model.h" #include #include "core/logger.h" -#include "model.h" +#include "core/model.h" +// clang-format on using namespace mmdeploy; @@ -15,9 +17,9 @@ int mmdeploy_model_create_by_path(const char *path, mm_model_t *model) { *model = ptr.release(); return MM_SUCCESS; } catch (const std::exception &e) { - ERROR("failed to create model: {}", e.what()); + MMDEPLOY_ERROR("failed to create model: {}", e.what()); } catch (...) { - ERROR("unknown exception caught"); + MMDEPLOY_ERROR("unknown exception caught"); } return MM_E_FAIL; } @@ -28,9 +30,9 @@ int mmdeploy_model_create(const void *buffer, int size, mm_model_t *model) { *model = ptr.release(); return MM_SUCCESS; } catch (const std::exception &e) { - ERROR("failed to create model: {}", e.what()); + MMDEPLOY_ERROR("failed to create model: {}", e.what()); } catch (...) { - ERROR("unknown exception caught"); + MMDEPLOY_ERROR("unknown exception caught"); } return MM_E_FAIL; } diff --git a/csrc/apis/c/model.h b/csrc/apis/c/model.h index 731bb0270b..6151ba43a5 100644 --- a/csrc/apis/c/model.h +++ b/csrc/apis/c/model.h @@ -10,13 +10,17 @@ #include "common.h" +#ifdef __cplusplus +extern "C" { +#endif + /** * @brief Create SDK Model instance from given model path * @param[in] path model path * @param[out] model sdk model instance that must be destroyed by \ref mmdeploy_model_destroy * @return status code of the operation */ -MM_SDK_API int mmdeploy_model_create_by_path(const char* path, mm_model_t* model); +MMDEPLOY_API int mmdeploy_model_create_by_path(const char* path, mm_model_t* model); /** * @brief Create SDK Model instance from memory @@ -25,13 +29,17 @@ MM_SDK_API int mmdeploy_model_create_by_path(const char* path, mm_model_t* model * @param[out] model sdk model instance that must be destroyed by \ref mmdeploy_model_destroy * @return status code of the operation */ -MM_SDK_API int mmdeploy_model_create(const void* buffer, int size, mm_model_t* model); +MMDEPLOY_API int mmdeploy_model_create(const void* buffer, int size, mm_model_t* model); /** * @brief Destroy model instance * @param[in] model sdk model instance created by \ref mmdeploy_model_create_by_path or \ref * mmdeploy_model_create */ -MM_SDK_API void mmdeploy_model_destroy(mm_model_t model); +MMDEPLOY_API void mmdeploy_model_destroy(mm_model_t model); + +#ifdef __cplusplus +} +#endif #endif // MMDEPLOY_SRC_APIS_C_MODEL_H_ diff --git a/csrc/apis/c/restorer.cpp b/csrc/apis/c/restorer.cpp index 0e12fa02c6..95e3679584 100644 --- a/csrc/apis/c/restorer.cpp +++ b/csrc/apis/c/restorer.cpp @@ -51,9 +51,9 @@ int mmdeploy_restorer_create_impl(ModelType &&m, const char *device_name, int de return MM_SUCCESS; } catch (const std::exception &e) { - ERROR("exception caught: {}", e.what()); + MMDEPLOY_ERROR("exception caught: {}", e.what()); } catch (...) { - ERROR("unknown exception caught"); + MMDEPLOY_ERROR("unknown exception caught"); } return MM_E_FAIL; } @@ -105,9 +105,9 @@ int mmdeploy_restorer_apply(mm_handle_t handle, const mm_mat_t *images, int coun *results = _results.release(); return MM_SUCCESS; } catch (const std::exception &e) { - ERROR("exception caught: {}", e.what()); + MMDEPLOY_ERROR("exception caught: {}", e.what()); } catch (...) { - ERROR("unknown exception caught"); + MMDEPLOY_ERROR("unknown exception caught"); } return MM_E_FAIL; } diff --git a/csrc/apis/c/restorer.h b/csrc/apis/c/restorer.h index 5d568cd418..4ae491a236 100644 --- a/csrc/apis/c/restorer.h +++ b/csrc/apis/c/restorer.h @@ -10,6 +10,10 @@ #include "common.h" +#ifdef __cplusplus +extern "C" { +#endif + /** * @brief Create a restorer instance * @param[in] model an instance of image restoration model created by @@ -20,8 +24,8 @@ * by \ref mmdeploy_restorer_destroy * @return status code of the operation */ -MM_SDK_API int mmdeploy_restorer_create(mm_model_t model, const char* device_name, int device_id, - mm_handle_t* handle); +MMDEPLOY_API int mmdeploy_restorer_create(mm_model_t model, const char* device_name, int device_id, + mm_handle_t* handle); /** * @brief Create a restorer instance @@ -32,8 +36,8 @@ MM_SDK_API int mmdeploy_restorer_create(mm_model_t model, const char* device_nam * by \ref mmdeploy_restorer_destroy * @return status code of the operation */ -MM_SDK_API int mmdeploy_restorer_create_by_path(const char* model_path, const char* device_name, - int device_id, mm_handle_t* handle); +MMDEPLOY_API int mmdeploy_restorer_create_by_path(const char* model_path, const char* device_name, + int device_id, mm_handle_t* handle); /** * @brief Apply restorer to a batch of images @@ -44,19 +48,23 @@ MM_SDK_API int mmdeploy_restorer_create_by_path(const char* model_path, const ch * by \ref mmdeploy_restorer_release_result * @return status code of the operation */ -MM_SDK_API int mmdeploy_restorer_apply(mm_handle_t handle, const mm_mat_t* images, int count, - mm_mat_t** results); +MMDEPLOY_API int mmdeploy_restorer_apply(mm_handle_t handle, const mm_mat_t* images, int count, + mm_mat_t** results); /** @brief Release result buffer returned by \ref mmdeploy_restorer_apply * @param[in] results result buffer by restorer * @param[in] count length of \p result */ -MM_SDK_API void mmdeploy_restorer_release_result(mm_mat_t* results, int count); +MMDEPLOY_API void mmdeploy_restorer_release_result(mm_mat_t* results, int count); /** * @brief destroy restorer * @param[in] handle handle of restorer created by \ref mmdeploy_restorer_create_by_path */ -MM_SDK_API void mmdeploy_restorer_destroy(mm_handle_t handle); +MMDEPLOY_API void mmdeploy_restorer_destroy(mm_handle_t handle); + +#ifdef __cplusplus +} +#endif #endif // MMDEPLOY_SRC_APIS_C_RESTORER_H_ diff --git a/csrc/apis/c/segmentor.cpp b/csrc/apis/c/segmentor.cpp index 2c578de321..bcdca722a7 100644 --- a/csrc/apis/c/segmentor.cpp +++ b/csrc/apis/c/segmentor.cpp @@ -53,28 +53,28 @@ int mmdeploy_segmentor_create_impl(ModelType&& m, const char* device_name, int d return MM_SUCCESS; } catch (const std::exception& e) { - ERROR("exception caught: {}", e.what()); + MMDEPLOY_ERROR("exception caught: {}", e.what()); } catch (...) { - ERROR("unknown exception caught"); + MMDEPLOY_ERROR("unknown exception caught"); } return MM_E_FAIL; } } // namespace -MM_SDK_API int mmdeploy_segmentor_create(mm_model_t model, const char* device_name, int device_id, - mm_handle_t* handle) { +int mmdeploy_segmentor_create(mm_model_t model, const char* device_name, int device_id, + mm_handle_t* handle) { return mmdeploy_segmentor_create_impl(*static_cast(model), device_name, device_id, handle); } -MM_SDK_API int mmdeploy_segmentor_create_by_path(const char* model_path, const char* device_name, - int device_id, mm_handle_t* handle) { +int mmdeploy_segmentor_create_by_path(const char* model_path, const char* device_name, + int device_id, mm_handle_t* handle) { return mmdeploy_segmentor_create_impl(model_path, device_name, device_id, handle); } -MM_SDK_API int mmdeploy_segmentor_apply(mm_handle_t handle, const mm_mat_t* mats, int mat_count, - mm_segment_t** results) { +int mmdeploy_segmentor_apply(mm_handle_t handle, const mm_mat_t* mats, int mat_count, + mm_segment_t** results) { if (handle == nullptr || mats == nullptr || mat_count == 0 || results == nullptr) { return MM_E_INVALID_ARG; } @@ -97,7 +97,7 @@ MM_SDK_API int mmdeploy_segmentor_apply(mm_handle_t handle, const mm_mat_t* mats auto results_ptr = _results.get(); for (auto i = 0; i < mat_count; ++i, ++results_ptr) { auto& output_item = output[i]; - DEBUG("the {}-th item in output: {}", i, output_item); + MMDEPLOY_DEBUG("the {}-th item in output: {}", i, output_item); auto segmentor_output = from_value(output_item); results_ptr->height = segmentor_output.height; results_ptr->width = segmentor_output.width; @@ -110,14 +110,14 @@ MM_SDK_API int mmdeploy_segmentor_apply(mm_handle_t handle, const mm_mat_t* mats return MM_SUCCESS; } catch (const std::exception& e) { - ERROR("exception caught: {}", e.what()); + MMDEPLOY_ERROR("exception caught: {}", e.what()); } catch (...) { - ERROR("unknown exception caught"); + MMDEPLOY_ERROR("unknown exception caught"); } return MM_E_FAIL; } -MM_SDK_API void mmdeploy_segmentor_release_result(mm_segment_t* results, int count) { +void mmdeploy_segmentor_release_result(mm_segment_t* results, int count) { if (results == nullptr) { return; } @@ -128,7 +128,7 @@ MM_SDK_API void mmdeploy_segmentor_release_result(mm_segment_t* results, int cou delete[] results; } -MM_SDK_API void mmdeploy_segmentor_destroy(mm_handle_t handle) { +void mmdeploy_segmentor_destroy(mm_handle_t handle) { if (handle != nullptr) { auto segmentor = static_cast(handle); delete segmentor; diff --git a/csrc/apis/c/segmentor.h b/csrc/apis/c/segmentor.h index 4abcd3cf68..741fbd9633 100644 --- a/csrc/apis/c/segmentor.h +++ b/csrc/apis/c/segmentor.h @@ -10,6 +10,10 @@ #include "common.h" +#ifdef __cplusplus +extern "C" { +#endif + typedef struct mm_segment_t { int height; ///< height of \p mask that equals to the input image's height int width; ///< width of \p mask that equals to the input image's width @@ -28,8 +32,8 @@ typedef struct mm_segment_t { * by \ref mmdeploy_segmentor_destroy * @return status of creating segmentor's handle */ -MM_SDK_API int mmdeploy_segmentor_create(mm_model_t model, const char* device_name, int device_id, - mm_handle_t* handle); +MMDEPLOY_API int mmdeploy_segmentor_create(mm_model_t model, const char* device_name, int device_id, + mm_handle_t* handle); /** * @brief Create segmentor's handle @@ -40,8 +44,8 @@ MM_SDK_API int mmdeploy_segmentor_create(mm_model_t model, const char* device_na * by \ref mmdeploy_segmentor_destroy * @return status of creating segmentor's handle */ -MM_SDK_API int mmdeploy_segmentor_create_by_path(const char* model_path, const char* device_name, - int device_id, mm_handle_t* handle); +MMDEPLOY_API int mmdeploy_segmentor_create_by_path(const char* model_path, const char* device_name, + int device_id, mm_handle_t* handle); /** * @brief Apply segmentor to batch images and get their inference results @@ -53,19 +57,23 @@ MM_SDK_API int mmdeploy_segmentor_create_by_path(const char* model_path, const c * image. It must be released by \ref mmdeploy_segmentor_release_result * @return status of inference */ -MM_SDK_API int mmdeploy_segmentor_apply(mm_handle_t handle, const mm_mat_t* mats, int mat_count, - mm_segment_t** results); +MMDEPLOY_API int mmdeploy_segmentor_apply(mm_handle_t handle, const mm_mat_t* mats, int mat_count, + mm_segment_t** results); /** @brief Release result buffer returned by \ref mmdeploy_segmentor_apply * @param[in] results result buffer * @param[in] count length of \p results */ -MM_SDK_API void mmdeploy_segmentor_release_result(mm_segment_t* results, int count); +MMDEPLOY_API void mmdeploy_segmentor_release_result(mm_segment_t* results, int count); /** * @brief Destroy segmentor's handle * @param[in] handle segmentor's handle created by \ref mmdeploy_segmentor_create_by_path */ -MM_SDK_API void mmdeploy_segmentor_destroy(mm_handle_t handle); +MMDEPLOY_API void mmdeploy_segmentor_destroy(mm_handle_t handle); + +#ifdef __cplusplus +} +#endif #endif // MMDEPLOY_SEGMENTOR_H diff --git a/csrc/apis/c/text_detector.cpp b/csrc/apis/c/text_detector.cpp index 1f0479f30d..6369044bfd 100644 --- a/csrc/apis/c/text_detector.cpp +++ b/csrc/apis/c/text_detector.cpp @@ -53,29 +53,28 @@ int mmdeploy_text_detector_create_impl(ModelType&& m, const char* device_name, i return MM_SUCCESS; } catch (const std::exception& e) { - ERROR("exception caught: {}", e.what()); + MMDEPLOY_ERROR("exception caught: {}", e.what()); } catch (...) { - ERROR("unknown exception caught"); + MMDEPLOY_ERROR("unknown exception caught"); } return MM_E_FAIL; } } // namespace -MM_SDK_API int mmdeploy_text_detector_create(mm_model_t model, const char* device_name, - int device_id, mm_handle_t* handle) { +int mmdeploy_text_detector_create(mm_model_t model, const char* device_name, int device_id, + mm_handle_t* handle) { return mmdeploy_text_detector_create_impl(*static_cast(model), device_name, device_id, handle); } -MM_SDK_API int mmdeploy_text_detector_create_by_path(const char* model_path, - const char* device_name, int device_id, - mm_handle_t* handle) { +int mmdeploy_text_detector_create_by_path(const char* model_path, const char* device_name, + int device_id, mm_handle_t* handle) { return mmdeploy_text_detector_create_impl(model_path, device_name, device_id, handle); } -MM_SDK_API int mmdeploy_text_detector_apply(mm_handle_t handle, const mm_mat_t* mats, int mat_count, - mm_text_detect_t** results, int** result_count) { +int mmdeploy_text_detector_apply(mm_handle_t handle, const mm_mat_t* mats, int mat_count, + mm_text_detect_t** results, int** result_count) { if (handle == nullptr || mats == nullptr || mat_count == 0) { return MM_E_INVALID_ARG; } @@ -91,7 +90,7 @@ MM_SDK_API int mmdeploy_text_detector_apply(mm_handle_t handle, const mm_mat_t* } auto output = text_detector->Run(std::move(input)).value().front(); - DEBUG("output: {}", output); + MMDEPLOY_DEBUG("output: {}", output); auto detector_outputs = from_value>(output); vector _result_count; @@ -125,20 +124,20 @@ MM_SDK_API int mmdeploy_text_detector_apply(mm_handle_t handle, const mm_mat_t* return MM_SUCCESS; } catch (const std::exception& e) { - ERROR("exception caught: {}", e.what()); + MMDEPLOY_ERROR("exception caught: {}", e.what()); } catch (...) { - ERROR("unknown exception caught"); + MMDEPLOY_ERROR("unknown exception caught"); } return MM_E_FAIL; } -MM_SDK_API void mmdeploy_text_detector_release_result(mm_text_detect_t* results, - const int* result_count, int count) { +void mmdeploy_text_detector_release_result(mm_text_detect_t* results, const int* result_count, + int count) { delete[] results; delete[] result_count; } -MM_SDK_API void mmdeploy_text_detector_destroy(mm_handle_t handle) { +void mmdeploy_text_detector_destroy(mm_handle_t handle) { if (handle != nullptr) { auto text_detector = static_cast(handle); delete text_detector; diff --git a/csrc/apis/c/text_detector.h b/csrc/apis/c/text_detector.h index 06cb78558c..0ca39b9003 100644 --- a/csrc/apis/c/text_detector.h +++ b/csrc/apis/c/text_detector.h @@ -10,6 +10,10 @@ #include "common.h" +#ifdef __cplusplus +extern "C" { +#endif + typedef struct mm_text_detect_t { mm_pointf_t bbox[4]; ///< a text bounding box of which the vertex are in clock-wise float score; @@ -25,8 +29,8 @@ typedef struct mm_text_detect_t { * by \ref mmdeploy_text_detector_destroy * @return status of creating text-detector's handle */ -MM_SDK_API int mmdeploy_text_detector_create(mm_model_t model, const char* device_name, - int device_id, mm_handle_t* handle); +MMDEPLOY_API int mmdeploy_text_detector_create(mm_model_t model, const char* device_name, + int device_id, mm_handle_t* handle); /** * @brief Create text-detector's handle @@ -37,9 +41,9 @@ MM_SDK_API int mmdeploy_text_detector_create(mm_model_t model, const char* devic * by \ref mmdeploy_text_detector_destroy * @return status of creating text-detector's handle */ -MM_SDK_API int mmdeploy_text_detector_create_by_path(const char* model_path, - const char* device_name, int device_id, - mm_handle_t* handle); +MMDEPLOY_API int mmdeploy_text_detector_create_by_path(const char* model_path, + const char* device_name, int device_id, + mm_handle_t* handle); /** * @brief Apply text-detector to batch images and get their inference results @@ -52,22 +56,27 @@ MM_SDK_API int mmdeploy_text_detector_create_by_path(const char* model_path, * results of each image. It must be released by \ref mmdeploy_detector_release_result * @return status of inference */ -MM_SDK_API int mmdeploy_text_detector_apply(mm_handle_t handle, const mm_mat_t* mats, int mat_count, - mm_text_detect_t** results, int** result_count); +MMDEPLOY_API int mmdeploy_text_detector_apply(mm_handle_t handle, const mm_mat_t* mats, + int mat_count, mm_text_detect_t** results, + int** result_count); /** @brief Release the inference result buffer returned by \ref mmdeploy_text_detector_apply * @param[in] results text detection result buffer * @param[in] result_count \p results size buffer * @param[in] count the length of buffer \p result_count */ -MM_SDK_API void mmdeploy_text_detector_release_result(mm_text_detect_t* results, - const int* result_count, int count); +MMDEPLOY_API void mmdeploy_text_detector_release_result(mm_text_detect_t* results, + const int* result_count, int count); /** * @brief Destroy text-detector's handle * @param[in] handle text-detector's handle created by \ref mmdeploy_text_detector_create_by_path or * \ref mmdeploy_text_detector_create */ -MM_SDK_API void mmdeploy_text_detector_destroy(mm_handle_t handle); +MMDEPLOY_API void mmdeploy_text_detector_destroy(mm_handle_t handle); + +#ifdef __cplusplus +} +#endif #endif // MMDEPLOY_TEXT_DETECTOR_H diff --git a/csrc/apis/c/text_recognizer.cpp b/csrc/apis/c/text_recognizer.cpp index 1ecf29e220..9458712b54 100644 --- a/csrc/apis/c/text_recognizer.cpp +++ b/csrc/apis/c/text_recognizer.cpp @@ -74,9 +74,9 @@ int mmdeploy_text_recognizer_create_impl(ModelType &&m, const char *device_name, return MM_SUCCESS; } catch (const std::exception &e) { - ERROR("exception caught: {}", e.what()); + MMDEPLOY_ERROR("exception caught: {}", e.what()); } catch (...) { - ERROR("unknown exception caught"); + MMDEPLOY_ERROR("unknown exception caught"); } return MM_E_FAIL; } @@ -178,9 +178,9 @@ int mmdeploy_text_recognizer_apply_bbox(mm_handle_t handle, const mm_mat_t *imag return MM_SUCCESS; } catch (const std::exception &e) { - ERROR("exception caught: {}", e.what()); + MMDEPLOY_ERROR("exception caught: {}", e.what()); } catch (...) { - ERROR("unknown exception caught"); + MMDEPLOY_ERROR("unknown exception caught"); } return MM_E_FAIL; } diff --git a/csrc/apis/c/text_recognizer.h b/csrc/apis/c/text_recognizer.h index d5bbd5e1ca..e257da5833 100644 --- a/csrc/apis/c/text_recognizer.h +++ b/csrc/apis/c/text_recognizer.h @@ -11,6 +11,10 @@ #include "common.h" #include "text_detector.h" +#ifdef __cplusplus +extern "C" { +#endif + typedef struct mm_text_recognize_t { char* text; float* score; @@ -27,8 +31,8 @@ typedef struct mm_text_recognize_t { * by \ref mmdeploy_text_recognizer_destroy * @return status code of the operation */ -MM_SDK_API int mmdeploy_text_recognizer_create(mm_model_t model, const char* device_name, - int device_id, mm_handle_t* handle); +MMDEPLOY_API int mmdeploy_text_recognizer_create(mm_model_t model, const char* device_name, + int device_id, mm_handle_t* handle); /** * @brief Create a text recognizer instance @@ -39,9 +43,9 @@ MM_SDK_API int mmdeploy_text_recognizer_create(mm_model_t model, const char* dev * by \ref mmdeploy_text_recognizer_destroy * @return status code of the operation */ -MM_SDK_API int mmdeploy_text_recognizer_create_by_path(const char* model_path, - const char* device_name, int device_id, - mm_handle_t* handle); +MMDEPLOY_API int mmdeploy_text_recognizer_create_by_path(const char* model_path, + const char* device_name, int device_id, + mm_handle_t* handle); /** * @brief Apply text recognizer to a batch of text images @@ -53,8 +57,8 @@ MM_SDK_API int mmdeploy_text_recognizer_create_by_path(const char* model_path, * by \ref mmdeploy_text_recognizer_release_result * @return status code of the operation */ -MM_SDK_API int mmdeploy_text_recognizer_apply(mm_handle_t handle, const mm_mat_t* images, int count, - mm_text_recognize_t** results); +MMDEPLOY_API int mmdeploy_text_recognizer_apply(mm_handle_t handle, const mm_mat_t* images, + int count, mm_text_recognize_t** results); /** * @brief Apply text recognizer to a batch of images supplied with text bboxes @@ -68,23 +72,28 @@ MM_SDK_API int mmdeploy_text_recognizer_apply(mm_handle_t handle, const mm_mat_t * bboxes, must be release by \ref mmdeploy_text_recognizer_release_result * @return status code of the operation */ -MM_SDK_API int mmdeploy_text_recognizer_apply_bbox(mm_handle_t handle, const mm_mat_t* images, - int image_count, const mm_text_detect_t* bboxes, - const int* bbox_count, - mm_text_recognize_t** results); +MMDEPLOY_API int mmdeploy_text_recognizer_apply_bbox(mm_handle_t handle, const mm_mat_t* images, + int image_count, + const mm_text_detect_t* bboxes, + const int* bbox_count, + mm_text_recognize_t** results); /** @brief Release result buffer returned by \ref mmdeploy_text_recognizer_apply or \ref * mmdeploy_text_recognizer_apply_bbox * @param[in] results result buffer by text recognizer * @param[in] count length of \p result */ -MM_SDK_API void mmdeploy_text_recognizer_release_result(mm_text_recognize_t* results, int count); +MMDEPLOY_API void mmdeploy_text_recognizer_release_result(mm_text_recognize_t* results, int count); /** * @brief destroy text recognizer * @param[in] handle handle of text recognizer created by \ref * mmdeploy_text_recognizer_create_by_path or \ref mmdeploy_text_recognizer_create */ -MM_SDK_API void mmdeploy_text_recognizer_destroy(mm_handle_t handle); +MMDEPLOY_API void mmdeploy_text_recognizer_destroy(mm_handle_t handle); + +#ifdef __cplusplus +} +#endif #endif // MMDEPLOY_SRC_APIS_C_TEXT_RECOGNIZER_H_ diff --git a/csrc/apis/python/CMakeLists.txt b/csrc/apis/python/CMakeLists.txt index 1d98d84d4b..0730268f07 100644 --- a/csrc/apis/python/CMakeLists.txt +++ b/csrc/apis/python/CMakeLists.txt @@ -23,11 +23,10 @@ mmdeploy_python_add_module(restorer) pybind11_add_module(${PROJECT_NAME} ${MMDEPLOY_PYTHON_SRCS}) -target_link_libraries(${PROJECT_NAME} PRIVATE - ${MMDEPLOY_LIBS} - -Wl,--whole-archive ${MMDEPLOY_STATIC_MODULES} -Wl,--no-whole-archive - -Wl,--no-as-needed ${MMDEPLOY_DYNAMIC_MODULES} -Wl,--as-need) +mmdeploy_load_static(${PROJECT_NAME} MMDeployStaticModules) +mmdeploy_load_dynamic(${PROJECT_NAME} MMDeployDynamicModules) +target_link_libraries(${PROJECT_NAME} PRIVATE MMDeployLibs) target_include_directories(${PROJECT_NAME} PRIVATE - ${CMAKE_CURRENT_SOURCE_DIR} - ${CMAKE_CURRENT_SOURCE_DIR}/../..) + ${CMAKE_CURRENT_SOURCE_DIR} + ${CMAKE_CURRENT_SOURCE_DIR}/../..) diff --git a/csrc/archive/CMakeLists.txt b/csrc/archive/CMakeLists.txt index 2b34ab1bb8..2645974152 100644 --- a/csrc/archive/CMakeLists.txt +++ b/csrc/archive/CMakeLists.txt @@ -6,7 +6,6 @@ add_library(${PROJECT_NAME} INTERFACE) target_link_libraries(${PROJECT_NAME} INTERFACE mmdeploy::core) add_library(mmdeploy::archive ALIAS mmdeploy_archive) -export_target(${PROJECT_NAME}) install(DIRECTORY ${CMAKE_SOURCE_DIR}/src/archive DESTINATION include/cpp FILES_MATCHING PATTERN "*.h") diff --git a/csrc/archive/json_archive.h b/csrc/archive/json_archive.h index 31666d89be..6f137b9a7c 100644 --- a/csrc/archive/json_archive.h +++ b/csrc/archive/json_archive.h @@ -133,7 +133,7 @@ inline Value json_to_value(const nlohmann::json& json) { return value; } default: - ERROR("unsupported json type: {}", json.type_name()); + MMDEPLOY_ERROR("unsupported json type: {}", json.type_name()); return {}; } } diff --git a/csrc/archive/value_archive.h b/csrc/archive/value_archive.h index 8500cbc424..f4115b3cb0 100644 --- a/csrc/archive/value_archive.h +++ b/csrc/archive/value_archive.h @@ -114,6 +114,8 @@ inline T from_value(const Value& value) { return x; } +namespace detail { + inline void load(ValueInputArchive& archive, Value& v) { archive.native(v); } template , Value>::value, bool> = true> @@ -121,6 +123,8 @@ inline void save(ValueOutputArchive& archive, T&& v) { archive.native(std::forward(v)); } +} // namespace detail + } // namespace mmdeploy #endif // MMDEPLOY_SRC_ARCHIVE_VALUE_ARCHIVE_H_ diff --git a/csrc/backend_ops/CMakeLists.txt b/csrc/backend_ops/CMakeLists.txt index 91117d2dbf..1537bd97fd 100644 --- a/csrc/backend_ops/CMakeLists.txt +++ b/csrc/backend_ops/CMakeLists.txt @@ -1,26 +1,28 @@ -set(CMAKE_CXX_STANDARD 14) +if (NOT MSVC) + set(CMAKE_CXX_STANDARD 14) +endif () set(CMAKE_CXX_FLAGS_RELEASE "-O3") # build ONNXRUNTIME ops if ("ort" IN_LIST MMDEPLOY_TARGET_BACKENDS) - if (NOT DEFINED ONNXRUNTIME_DIR) - set(ONNXRUNTIME_DIR $ENV{ONNXRUNTIME_DIR}) - endif () - if (NOT ONNXRUNTIME_DIR) - message(FATAL_ERROR " ONNXRUNTIME_DIR is not found.") - else () - message(STATUS "Build ONNXRUNTIME custom ops.") - add_subdirectory(onnxruntime) - endif () + if (NOT DEFINED ONNXRUNTIME_DIR) + set(ONNXRUNTIME_DIR $ENV{ONNXRUNTIME_DIR}) + endif () + if (NOT ONNXRUNTIME_DIR) + message(FATAL_ERROR " ONNXRUNTIME_DIR is not found.") + else () + message(STATUS "Build ONNXRUNTIME custom ops.") + add_subdirectory(onnxruntime) + endif () endif () # build TensorRT ops if ("trt" IN_LIST MMDEPLOY_TARGET_BACKENDS) - if (NOT DEFINED TENSORRT_DIR) - set(TENSORRT_DIR $ENV{TENSORRT_DIR}) - endif () - message(STATUS "Build TensorRT custom ops.") - add_subdirectory(tensorrt) + if (NOT DEFINED TENSORRT_DIR) + set(TENSORRT_DIR $ENV{TENSORRT_DIR}) + endif () + message(STATUS "Build TensorRT custom ops.") + add_subdirectory(tensorrt) endif () # build NCNN ops @@ -28,3 +30,9 @@ if ("ncnn" IN_LIST MMDEPLOY_TARGET_BACKENDS) message(STATUS "Build NCNN custom ops") add_subdirectory(ncnn) endif () + +# build TorchScript ops +if ("torchscript" IN_LIST MMDEPLOY_TARGET_BACKENDS) + message(STATUS "Build torchsciprt custom ops") + add_subdirectory(torchscript) +endif () diff --git a/csrc/backend_ops/common/modulated_deform_conv/common_cuda_helper.cuh b/csrc/backend_ops/common/modulated_deform_conv/common_cuda_helper.cuh new file mode 100644 index 0000000000..02c57c62e6 --- /dev/null +++ b/csrc/backend_ops/common/modulated_deform_conv/common_cuda_helper.cuh @@ -0,0 +1,94 @@ +// Copyright (c) OpenMMLab. All rights reserved. +#ifndef COMMON_CUDA_HELPER +#define COMMON_CUDA_HELPER + +#include +#include + +#include + +#define CUDA_1D_KERNEL_LOOP(i, n) \ + for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < (n); i += blockDim.x * gridDim.x) + +#define THREADS_PER_BLOCK 512 + +#define DIVUP(m, n) ((m) / (n) + ((m) % (n) > 0)) +inline int GET_BLOCKS(const int N) { + int optimal_block_num = DIVUP(N, THREADS_PER_BLOCK); + int max_block_num = 4096; + return std::min(optimal_block_num, max_block_num); +} + +#define cudaCheckError() \ + { \ + cudaError_t e = cudaGetLastError(); \ + if (e != cudaSuccess) { \ + printf("Cuda failure %s:%d: '%s'\n", __FILE__, __LINE__, cudaGetErrorString(e)); \ + exit(0); \ + } \ + } + +/** + * Returns a view of the original tensor with its dimensions permuted. + * + * @param[out] dst pointer to the destination tensor + * @param[in] src pointer to the source tensor + * @param[in] src_size shape of the src tensor + * @param[in] permute The desired ordering of dimensions + * @param[in] src_dim dim of src tensor + * @param[in] stream cuda stream handle + */ +template +void memcpyPermute(scalar_t* dst, const scalar_t* src, int* src_size, int* permute, int src_dim, + cudaStream_t stream = 0); + +template +cublasStatus_t cublasGemmWrap(cublasHandle_t handle, cublasOperation_t transa, + cublasOperation_t transb, int m, int n, int k, const scalar_t* alpha, + const scalar_t* A, int lda, const scalar_t* B, int ldb, + const scalar_t* beta, scalar_t* C, int ldc); + +template +__device__ scalar_t bilinear_interpolate(const scalar_t* input, const int height, const int width, + scalar_t y, scalar_t x) { + // deal with cases that inverse elements are out of feature map boundary + if (y < -1.0 || y > height || x < -1.0 || x > width) return 0; + + if (y <= 0) y = 0; + if (x <= 0) x = 0; + + int y_low = (int)y; + int x_low = (int)x; + int y_high; + int x_high; + + if (y_low >= height - 1) { + y_high = y_low = height - 1; + y = (scalar_t)y_low; + } else { + y_high = y_low + 1; + } + + if (x_low >= width - 1) { + x_high = x_low = width - 1; + x = (scalar_t)x_low; + } else { + x_high = x_low + 1; + } + + scalar_t ly = y - y_low; + scalar_t lx = x - x_low; + scalar_t hy = 1. - ly, hx = 1. - lx; + // do bilinear interpolation + scalar_t v1 = input[y_low * width + x_low]; + scalar_t v2 = input[y_low * width + x_high]; + scalar_t v3 = input[y_high * width + x_low]; + scalar_t v4 = input[y_high * width + x_high]; + scalar_t w1 = hy * hx, w2 = hy * lx, w3 = ly * hx, w4 = ly * lx; + + scalar_t val = (w1 * v1 + w2 * v2 + w3 * v3 + w4 * v4); + + return val; +} + +#endif // COMMON_CUDA_HELPER diff --git a/csrc/backend_ops/common/modulated_deform_conv/modulated_deform_conv_cpu.h b/csrc/backend_ops/common/modulated_deform_conv/modulated_deform_conv_cpu.h new file mode 100644 index 0000000000..a37e243109 --- /dev/null +++ b/csrc/backend_ops/common/modulated_deform_conv/modulated_deform_conv_cpu.h @@ -0,0 +1,82 @@ +#include +#include + +template +T bilinear_interpolate_2d(const T *src, const int64_t src_h, const int64_t src_w, const T h, + const T w) { + if (h <= -1 || src_h <= h || w <= -1 || src_w <= w) { + return 0; + } + + int64_t h_low = floor(h); + int64_t w_low = floor(w); + int64_t h_high = h_low + 1; + int64_t w_high = w_low + 1; + + T lh = h - h_low; + T lw = w - w_low; + T hh = 1 - lh; + T hw = 1 - lw; + + T v1 = 0; + if (h_low >= 0 && w_low >= 0) v1 = src[h_low * src_w + w_low]; + T v2 = 0; + if (h_low >= 0 && w_high <= src_w - 1) v2 = src[h_low * src_w + w_high]; + T v3 = 0; + if (h_high <= src_h - 1 && w_low >= 0) v3 = src[h_high * src_w + w_low]; + T v4 = 0; + if (h_high <= src_h - 1 && w_high <= src_w - 1) v4 = src[h_high * src_w + w_high]; + + T w1 = hh * hw, w2 = hh * lw, w3 = lh * hw, w4 = lh * lw; + + T val = (w1 * v1 + w2 * v2 + w3 * v3 + w4 * v4); + return val; +} + +// output: (channels * kernel_h * kernel_w, dst_h * dst_w) +template +void deformable_im2col_2d(const T *input, const T *offset, const T *mask, const int64_t src_h, + const int64_t src_w, const int64_t kernel_h, const int64_t kernel_w, + const int64_t pad_h, const int64_t pad_w, const int64_t stride_h, + const int64_t stride_w, const int64_t dilation_h, + const int64_t dilation_w, const int64_t channels, + const int64_t offset_groups, const int64_t dst_h, const int64_t dst_w, + const bool use_mask, T *columns) { + const int64_t workload = channels * dst_h * dst_w; + for (int64_t index = 0; index != workload; ++index) { + const int64_t ow = index % dst_w; + const int64_t oh = (index / dst_w) % dst_h; + const int64_t ic = index / (dst_w * dst_h); + const int64_t oc = ic * kernel_h * kernel_w; + + int64_t c_per_offset_grp = channels / offset_groups; + const int64_t grp_idx = ic / c_per_offset_grp; + + auto columns_ptr = columns + (oc * (dst_h * dst_w) + oh * dst_w + ow); + auto input_ptr = input + ic * (src_h * src_w); + auto offset_ptr = offset + grp_idx * 2 * kernel_h * kernel_w * dst_h * dst_w; + auto mask_ptr = mask; + if (use_mask) { + mask_ptr += grp_idx * kernel_h * kernel_w * dst_h * dst_w; + } + + for (int64_t kh = 0; kh < kernel_h; ++kh) { + for (int64_t kw = 0; kw < kernel_w; ++kw) { + const int64_t mask_idx = kh * kernel_w + kw; + const int64_t offset_idx = 2 * mask_idx; + + T mask_value = 1; + if (use_mask) { + mask_value = mask_ptr[mask_idx * (dst_h * dst_w) + oh * dst_w + ow]; + } + + const T offset_h = offset_ptr[offset_idx * (dst_h * dst_w) + oh * dst_w + ow]; + const T offset_w = offset_ptr[(offset_idx + 1) * (dst_h * dst_w) + oh * dst_w + ow]; + const T ih = (oh * stride_h - pad_h) + kh * dilation_h + offset_h; + const T iw = (ow * stride_w - pad_w) + kw * dilation_w + offset_w; + *columns_ptr = mask_value * bilinear_interpolate_2d(input_ptr, src_h, src_w, ih, iw); + columns_ptr += dst_h * dst_w; + } + } + } +} diff --git a/csrc/backend_ops/tensorrt/modulated_deform_conv/trt_modulated_deform_conv_kernel.hpp b/csrc/backend_ops/common/modulated_deform_conv/modulated_deform_conv_cuda.cuh similarity index 99% rename from csrc/backend_ops/tensorrt/modulated_deform_conv/trt_modulated_deform_conv_kernel.hpp rename to csrc/backend_ops/common/modulated_deform_conv/modulated_deform_conv_cuda.cuh index 2d78998a4d..3f4b2a55ce 100644 --- a/csrc/backend_ops/tensorrt/modulated_deform_conv/trt_modulated_deform_conv_kernel.hpp +++ b/csrc/backend_ops/common/modulated_deform_conv/modulated_deform_conv_cuda.cuh @@ -68,7 +68,7 @@ #include -#include "common_cuda_helper.hpp" +#include "common_cuda_helper.cuh" template __device__ T dmcn_im2col_bilinear(const T *input, const int data_width, const int height, diff --git a/csrc/backend_ops/ncnn/CMakeLists.txt b/csrc/backend_ops/ncnn/CMakeLists.txt index 6345448e6a..9580d3b96c 100755 --- a/csrc/backend_ops/ncnn/CMakeLists.txt +++ b/csrc/backend_ops/ncnn/CMakeLists.txt @@ -4,21 +4,20 @@ cmake_minimum_required(VERSION 3.14) # ncnn find_package(ncnn) -if(ncnn_FOUND) - message(STATUS "ncnn library found!") -else() - message(FATAL_ERROR "Could not locate ncnn") -endif() +if (ncnn_FOUND) + message(STATUS "ncnn library found!") +else () + message(FATAL_ERROR "Could not locate ncnn") +endif () -set_targets(mmdeploy_ncnn_ops NCNN_OPS_OBJ NCNN_OPS_STATIC NCNN_OPS_SHARED) -if(NOT ANDROID AND NOT IOS) - add_subdirectory(ops) - add_subdirectory(onnx2ncnn) - add_subdirectory(pyncnn_ext) -else() - # In case of embedded platform, like android, or ios, we only build custom ncnn - # ops, and leave the executable converter(onnx2ncnn, pyncnn_ext) built under - # the host platforms - add_subdirectory(ops) -endif() +if (NOT ANDROID AND NOT IOS) + add_subdirectory(ops) + add_subdirectory(onnx2ncnn) + add_subdirectory(pyncnn_ext) +else () + # In case of embedded platform, like android, or ios, we only build custom ncnn + # ops, and leave the executable converter(onnx2ncnn, pyncnn_ext) built under + # the host platforms + add_subdirectory(ops) +endif () diff --git a/csrc/backend_ops/ncnn/ops/CMakeLists.txt b/csrc/backend_ops/ncnn/ops/CMakeLists.txt index aa89729843..461301211a 100755 --- a/csrc/backend_ops/ncnn/ops/CMakeLists.txt +++ b/csrc/backend_ops/ncnn/ops/CMakeLists.txt @@ -2,18 +2,24 @@ cmake_minimum_required(VERSION 3.14) project(mmdeploy_ncnn_ops) -include(${CMAKE_SOURCE_DIR}/cmake/common.cmake) +include(${CMAKE_SOURCE_DIR}/cmake/MMDeploy.cmake) # add plugin source file(GLOB_RECURSE NCNN_OPS_SRCS *.cpp) -build_object_target(${NCNN_OPS_OBJ} "${NCNN_OPS_SRCS}") -target_link_libraries(${NCNN_OPS_OBJ} ncnn) +add_library(${PROJECT_NAME}_obj OBJECT "${NCNN_OPS_SRCS}") +target_compile_definitions(${PROJECT_NAME}_obj PRIVATE -DMMDEPLOY_API_EXPORTS=1) +set_target_properties(${PROJECT_NAME}_obj PROPERTIES POSITION_INDEPENDENT_CODE 1) +target_link_libraries(${PROJECT_NAME}_obj PRIVATE ncnn) +set(_COMMON_INCLUDE_DIRS + $ + $) +target_include_directories(${PROJECT_NAME}_obj + PUBLIC ${_COMMON_INCLUDE_DIRS}) +mmdeploy_export(${PROJECT_NAME}_obj) -build_shared_target(${NCNN_OPS_SHARED} ${NCNN_OPS_OBJ} "PRIVATE") -install_targets(${NCNN_OPS_SHARED}) +mmdeploy_add_library(${PROJECT_NAME} SHARED EXCLUDE "") +target_link_libraries(${PROJECT_NAME} PRIVATE ${PROJECT_NAME}_obj) +target_include_directories(${PROJECT_NAME} + PUBLIC ${_COMMON_INCLUDE_DIRS}) -if (MMDEPLOY_BUILD_SDK) - ## Build static library. SDK's uses it to build `ncnn_net` module - build_static_target(${NCNN_OPS_STATIC} ${NCNN_OPS_OBJ} "PRIVATE") - add_library(mmdeploy::ncnn_ops::static ALIAS ${NCNN_OPS_STATIC}) -endif () +add_library(mmdeploy::ncnn_ops ALIAS ${PROJECT_NAME}) diff --git a/csrc/backend_ops/ncnn/ops/ncnn_ops_register.h b/csrc/backend_ops/ncnn/ops/ncnn_ops_register.h index 2fb07d8b0f..333f174e79 100755 --- a/csrc/backend_ops/ncnn/ops/ncnn_ops_register.h +++ b/csrc/backend_ops/ncnn/ops/ncnn_ops_register.h @@ -5,13 +5,12 @@ #include #include +#include "core/macro.h" #include "net.h" -extern "C" { -std::map& get_mmdeploy_layer_creator(); -std::map& get_mmdeploy_layer_destroyer(); +MMDEPLOY_API std::map& get_mmdeploy_layer_creator(); +MMDEPLOY_API std::map& get_mmdeploy_layer_destroyer(); -int register_mmdeploy_custom_layers(ncnn::Net& net); -} +MMDEPLOY_API int register_mmdeploy_custom_layers(ncnn::Net& net); #endif diff --git a/csrc/backend_ops/ncnn/pyncnn_ext/CMakeLists.txt b/csrc/backend_ops/ncnn/pyncnn_ext/CMakeLists.txt index f0d4148963..c55acb488f 100755 --- a/csrc/backend_ops/ncnn/pyncnn_ext/CMakeLists.txt +++ b/csrc/backend_ops/ncnn/pyncnn_ext/CMakeLists.txt @@ -6,9 +6,12 @@ project(ncnn_ext) if (NOT TARGET pybind11) add_subdirectory(${CMAKE_SOURCE_DIR}/third_party/pybind11 pybind11) endif () -include_directories(${pybind11_INCLUDE_DIR} ${PYTHON_INCLUDE_DIRS}) + pybind11_add_module(ncnn_ext ncnn_ext.cpp) -target_link_libraries(ncnn_ext PUBLIC ncnn ${NCNN_OPS_SHARED}) -set_target_properties( - ncnn_ext PROPERTIES LIBRARY_OUTPUT_DIRECTORY - ${CMAKE_SOURCE_DIR}/mmdeploy/backend/ncnn) + +target_link_libraries(ncnn_ext PUBLIC mmdeploy_ncnn_ops ncnn) +set(_NCNN_EXT_DIR ${CMAKE_SOURCE_DIR}/mmdeploy/backend/ncnn) +set_target_properties(ncnn_ext PROPERTIES + LIBRARY_OUTPUT_DIRECTORY ${_NCNN_EXT_DIR} + LIBRARY_OUTPUT_DIRECTORY_DEBUG ${_NCNN_EXT_DIR} + LIBRARY_OUTPUT_DIRECTORY_RELEASE ${_NCNN_EXT_DIR}) diff --git a/csrc/backend_ops/ncnn/pyncnn_ext/ncnn_ext.cpp b/csrc/backend_ops/ncnn/pyncnn_ext/ncnn_ext.cpp index e9ff04eb41..ac158b9edb 100755 --- a/csrc/backend_ops/ncnn/pyncnn_ext/ncnn_ext.cpp +++ b/csrc/backend_ops/ncnn/pyncnn_ext/ncnn_ext.cpp @@ -1,7 +1,7 @@ // Copyright (c) OpenMMLab. All rights reserved. #include -#include "../ops/ncnn_ops_register.h" +#include "ncnn_ops_register.h" #include "net.h" PYBIND11_MODULE(ncnn_ext, m) { diff --git a/csrc/backend_ops/onnxruntime/CMakeLists.txt b/csrc/backend_ops/onnxruntime/CMakeLists.txt index f646bbc98a..b136781401 100644 --- a/csrc/backend_ops/onnxruntime/CMakeLists.txt +++ b/csrc/backend_ops/onnxruntime/CMakeLists.txt @@ -2,26 +2,24 @@ cmake_minimum_required(VERSION 3.14) project(mmdeploy_onnxruntime_ops) -include(${CMAKE_SOURCE_DIR}/cmake/common.cmake) -set_targets(${PROJECT_NAME} ORT_OPS_OBJ ORT_OPS_STATIC ORT_OPS_MODULE) +include(${CMAKE_SOURCE_DIR}/cmake/MMDeploy.cmake) # add plugin source file(GLOB_RECURSE ORT_OPS_SRCS *.cpp) -build_object_target(${ORT_OPS_OBJ} "${ORT_OPS_SRCS}") -target_include_directories(${ORT_OPS_OBJ} PUBLIC +add_library(${PROJECT_NAME}_obj OBJECT "${ORT_OPS_SRCS}") +target_compile_definitions(${PROJECT_NAME}_obj PRIVATE -DMMDEPLOY_API_EXPORTS=1) +set_target_properties(${PROJECT_NAME}_obj PROPERTIES POSITION_INDEPENDENT_CODE 1) +mmdeploy_export(${PROJECT_NAME}_obj) + +target_include_directories(${PROJECT_NAME}_obj PUBLIC $ - $) -target_link_directories(${ORT_OPS_OBJ} PUBLIC + $ + $ + $) +target_link_directories(${PROJECT_NAME}_obj PUBLIC ${ONNXRUNTIME_DIR}/lib) -target_link_libraries(${ORT_OPS_OBJ} PUBLIC onnxruntime) - -add_library(${ORT_OPS_MODULE} MODULE $) -target_link_libraries(${ORT_OPS_MODULE} PRIVATE ${ORT_OPS_OBJ}) -add_library(mmdeploy::onnxruntime::ops ALIAS ${ORT_OPS_MODULE}) -install_targets(${ORT_OPS_MODULE}) +target_link_libraries(${PROJECT_NAME}_obj PUBLIC onnxruntime) -if (MMDEPLOY_BUILD_SDK) - ## Build static library. SDK's uses it to build `ort_net` module - build_static_target(${ORT_OPS_STATIC} ${ORT_OPS_OBJ} "PRIVATE") - add_library(mmdeploy::onnxruntime::ops::static ALIAS ${ORT_OPS_STATIC}) -endif () +mmdeploy_add_library(${PROJECT_NAME} SHARED EXCLUDE "") +target_link_libraries(${PROJECT_NAME} PUBLIC ${PROJECT_NAME}_obj) +add_library(mmdeploy::onnxruntime::ops ALIAS ${PROJECT_NAME}) diff --git a/csrc/backend_ops/onnxruntime/common/onnxruntime_register.h b/csrc/backend_ops/onnxruntime/common/onnxruntime_register.h index 84318bc818..344031e791 100644 --- a/csrc/backend_ops/onnxruntime/common/onnxruntime_register.h +++ b/csrc/backend_ops/onnxruntime/common/onnxruntime_register.h @@ -3,11 +3,14 @@ #define ONNXRUNTIME_REGISTER_H #include +#include "core/macro.h" + #ifdef __cplusplus extern "C" { #endif -OrtStatus *ORT_API_CALL RegisterCustomOps(OrtSessionOptions *options, const OrtApiBase *api); +MMDEPLOY_API OrtStatus *ORT_API_CALL RegisterCustomOps(OrtSessionOptions *options, + const OrtApiBase *api); #ifdef __cplusplus } diff --git a/csrc/backend_ops/onnxruntime/modulated_deform_conv/modulated_deform_conv.cpp b/csrc/backend_ops/onnxruntime/modulated_deform_conv/modulated_deform_conv.cpp index 5561752cd6..3df1217a37 100644 --- a/csrc/backend_ops/onnxruntime/modulated_deform_conv/modulated_deform_conv.cpp +++ b/csrc/backend_ops/onnxruntime/modulated_deform_conv/modulated_deform_conv.cpp @@ -4,88 +4,11 @@ #include #include +#include "modulated_deform_conv/modulated_deform_conv_cpu.h" #include "ort_utils.h" namespace mmdeploy { -float bilinear_interpolate_2d(const float *src, const int64_t src_h, const int64_t src_w, - const float h, const float w) { - if (h <= -1 || src_h <= h || w <= -1 || src_w <= w) { - return 0; - } - - int64_t h_low = floor(h); - int64_t w_low = floor(w); - int64_t h_high = h_low + 1; - int64_t w_high = w_low + 1; - - float lh = h - h_low; - float lw = w - w_low; - float hh = 1 - lh; - float hw = 1 - lw; - - float v1 = 0; - if (h_low >= 0 && w_low >= 0) v1 = src[h_low * src_w + w_low]; - float v2 = 0; - if (h_low >= 0 && w_high <= src_w - 1) v2 = src[h_low * src_w + w_high]; - float v3 = 0; - if (h_high <= src_h - 1 && w_low >= 0) v3 = src[h_high * src_w + w_low]; - float v4 = 0; - if (h_high <= src_h - 1 && w_high <= src_w - 1) v4 = src[h_high * src_w + w_high]; - - float w1 = hh * hw, w2 = hh * lw, w3 = lh * hw, w4 = lh * lw; - - float val = (w1 * v1 + w2 * v2 + w3 * v3 + w4 * v4); - return val; -} - -// output: (channels * kernel_h * kernel_w, dst_h * dst_w) -void deformable_im2col_2d(const float *input, const float *offset, const float *mask, - const int64_t src_h, const int64_t src_w, const int64_t kernel_h, - const int64_t kernel_w, const int64_t pad_h, const int64_t pad_w, - const int64_t stride_h, const int64_t stride_w, const int64_t dilation_h, - const int64_t dilation_w, const int64_t channels, - const int64_t offset_groups, const int64_t dst_h, const int64_t dst_w, - const bool use_mask, float *columns) { - const int64_t workload = channels * dst_h * dst_w; - for (int64_t index = 0; index != workload; ++index) { - const int64_t ow = index % dst_w; - const int64_t oh = (index / dst_w) % dst_h; - const int64_t ic = index / (dst_w * dst_h); - const int64_t oc = ic * kernel_h * kernel_w; - - int64_t c_per_offset_grp = channels / offset_groups; - const int64_t grp_idx = ic / c_per_offset_grp; - - auto columns_ptr = columns + (oc * (dst_h * dst_w) + oh * dst_w + ow); - auto input_ptr = input + ic * (src_h * src_w); - auto offset_ptr = offset + grp_idx * 2 * kernel_h * kernel_w * dst_h * dst_w; - auto mask_ptr = mask; - if (use_mask) { - mask_ptr += grp_idx * kernel_h * kernel_w * dst_h * dst_w; - } - - for (int64_t kh = 0; kh < kernel_h; ++kh) { - for (int64_t kw = 0; kw < kernel_w; ++kw) { - const int64_t mask_idx = kh * kernel_w + kw; - const int64_t offset_idx = 2 * mask_idx; - - float mask_value = 1; - if (use_mask) { - mask_value = mask_ptr[mask_idx * (dst_h * dst_w) + oh * dst_w + ow]; - } - - const float offset_h = offset_ptr[offset_idx * (dst_h * dst_w) + oh * dst_w + ow]; - const float offset_w = offset_ptr[(offset_idx + 1) * (dst_h * dst_w) + oh * dst_w + ow]; - const float ih = (oh * stride_h - pad_h) + kh * dilation_h + offset_h; - const float iw = (ow * stride_w - pad_w) + kw * dilation_w + offset_w; - *columns_ptr = mask_value * bilinear_interpolate_2d(input_ptr, src_h, src_w, ih, iw); - columns_ptr += dst_h * dst_w; - } - } - } -} - void gemm_ref_fp32(const float *A, const float *B, const float *V, const float *H, const int32_t trans_A, const int32_t trans_B, const int32_t M, const int32_t N, const int32_t K, const float alpha, const float beta, float *Y) { @@ -162,12 +85,12 @@ void deformable_conv2d_ref_fp32(const float *src, const float *offset, const flo for (int64_t b = 0; b < batch; ++b) { for (int64_t g = 0; g < group; ++g) { - deformable_im2col_2d(src + b * src_c * src_h * src_w + g * ic_per_gp * src_h * src_w, - offset + b * offset_group * 2 * kernel_h * kernel_w * dst_h * dst_w, - mask + b * offset_group * kernel_h * kernel_w * dst_h * dst_w, src_h, - src_w, kernel_h, kernel_w, pad_h, pad_w, stride_h, stride_w, dilation_h, - dilation_w, ic_per_gp, offset_group, dst_h, dst_w, mask != nullptr, - columns); + deformable_im2col_2d( + src + b * src_c * src_h * src_w + g * ic_per_gp * src_h * src_w, + offset + b * offset_group * 2 * kernel_h * kernel_w * dst_h * dst_w, + mask + b * offset_group * kernel_h * kernel_w * dst_h * dst_w, src_h, src_w, kernel_h, + kernel_w, pad_h, pad_w, stride_h, stride_w, dilation_h, dilation_w, ic_per_gp, + offset_group, dst_h, dst_w, mask != nullptr, columns); float *dst_ptr = dst + b * dst_c * dst_h * dst_w + g * oc_per_gp * dst_h * dst_w; if (bias != nullptr) { const float *bias_ptr = bias + g * oc_per_gp; diff --git a/csrc/backend_ops/onnxruntime/onnxruntime_register.cpp b/csrc/backend_ops/onnxruntime/onnxruntime_register.cpp index 9f2ce2cc0f..f7b9cedff8 100644 --- a/csrc/backend_ops/onnxruntime/onnxruntime_register.cpp +++ b/csrc/backend_ops/onnxruntime/onnxruntime_register.cpp @@ -7,7 +7,6 @@ const char *c_MMDeployOpDomain = "mmdeploy"; OrtStatus *ORT_API_CALL RegisterCustomOps(OrtSessionOptions *options, const OrtApiBase *api) { const OrtApi *kOrtApi = api->GetApi(ORT_API_VERSION); - OrtStatus *status = nullptr; for (auto &_op_list_pair : mmdeploy::get_mmdeploy_custom_ops()) { OrtCustomOpDomain *domain = nullptr; diff --git a/csrc/backend_ops/tensorrt/CMakeLists.txt b/csrc/backend_ops/tensorrt/CMakeLists.txt index 88a0176df6..626fb6c92a 100644 --- a/csrc/backend_ops/tensorrt/CMakeLists.txt +++ b/csrc/backend_ops/tensorrt/CMakeLists.txt @@ -3,41 +3,37 @@ cmake_minimum_required(VERSION 3.14) include(${CMAKE_SOURCE_DIR}/cmake/cuda.cmake NO_POLICY_SCOPE) project(mmdeploy_tensorrt_ops CUDA CXX) -include(${CMAKE_SOURCE_DIR}/cmake/common.cmake) +include(${CMAKE_SOURCE_DIR}/cmake/MMDeploy.cmake) include(${CMAKE_SOURCE_DIR}/cmake/cuda.cmake NO_POLICY_SCOPE) include(${CMAKE_SOURCE_DIR}/cmake/tensorrt.cmake NO_POLICY_SCOPE) -set_targets(${PROJECT_NAME} BACKEND_OPS_OBJ BACKEND_OPS_STATIC BACKEND_OPS_MODULE) - # cub if (NOT DEFINED CUB_ROOT_DIR) if (CUDA_VERSION VERSION_LESS 11.0) set(CUB_ROOT_DIR "${CMAKE_SOURCE_DIR}/third_party/cub") - endif() + endif () endif () file(GLOB_RECURSE BACKEND_OPS_SRCS *.cpp *.cu) -build_object_target(${BACKEND_OPS_OBJ} "${BACKEND_OPS_SRCS}") -target_compile_definitions(${BACKEND_OPS_OBJ} +add_library(${PROJECT_NAME}_obj OBJECT "${BACKEND_OPS_SRCS}") +set_target_properties(${PROJECT_NAME}_obj PROPERTIES POSITION_INDEPENDENT_CODE 1) +target_compile_definitions(${PROJECT_NAME}_obj PRIVATE -DTHRUST_IGNORE_DEPRECATED_CPP_DIALECT=1) -target_include_directories(${BACKEND_OPS_OBJ} +target_include_directories(${PROJECT_NAME}_obj + PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/../common) +target_include_directories(${PROJECT_NAME}_obj PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/common) -target_include_directories(${BACKEND_OPS_OBJ} +target_include_directories(${PROJECT_NAME}_obj PRIVATE ${CUDA_TOOLKIT_ROOT_DIR}/include) -target_include_directories(${BACKEND_OPS_OBJ} PRIVATE ${TENSORRT_INCLUDE_DIR}) -target_include_directories(${BACKEND_OPS_OBJ} PRIVATE ${CUDNN_DIR}/include) -target_include_directories(${BACKEND_OPS_OBJ} PRIVATE ${CUB_ROOT_DIR}) -target_link_directories(${BACKEND_OPS_OBJ} PUBLIC ${CUDNN_DIR}/lib64) -target_link_libraries(${BACKEND_OPS_OBJ} - PRIVATE ${TENSORRT_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} cudnn) +target_include_directories(${PROJECT_NAME}_obj PRIVATE ${TENSORRT_INCLUDE_DIR}) +target_include_directories(${PROJECT_NAME}_obj PRIVATE ${CUDNN_DIR}/include) +target_include_directories(${PROJECT_NAME}_obj PRIVATE ${CUB_ROOT_DIR}) +target_link_directories(${PROJECT_NAME}_obj PUBLIC ${CUDNN_DIR}/lib64 ${CUDNN_DIR}/lib/x64) +target_link_libraries(${PROJECT_NAME}_obj + PUBLIC ${TENSORRT_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} cudnn) +mmdeploy_export(${PROJECT_NAME}_obj) # Build module library. It is used to convert onnx model to tensorrt engine -build_module_target(${BACKEND_OPS_MODULE} ${BACKEND_OPS_OBJ} "PRIVATE") -add_library(mmdeploy::tensorrt_ops ALIAS ${BACKEND_OPS_MODULE}) -install_targets(${BACKEND_OPS_MODULE}) - -if (MMDEPLOY_BUILD_SDK) - ## Build static library. SDK's uses it to build `trt_net` module - build_static_target(${BACKEND_OPS_STATIC} ${BACKEND_OPS_OBJ} "PRIVATE") - add_library(mmdeploy::tensorrt_ops::static ALIAS ${BACKEND_OPS_STATIC}) -endif () +mmdeploy_add_module(${PROJECT_NAME} MODULE EXCLUDE "") +target_link_libraries(${PROJECT_NAME} PRIVATE ${PROJECT_NAME}_obj) +add_library(mmdeploy::tensorrt_ops ALIAS ${PROJECT_NAME}) diff --git a/csrc/backend_ops/tensorrt/modulated_deform_conv/trt_modulated_deform_conv_kernel.cu b/csrc/backend_ops/tensorrt/modulated_deform_conv/trt_modulated_deform_conv_kernel.cu index a5940b5e40..ed284e7809 100644 --- a/csrc/backend_ops/tensorrt/modulated_deform_conv/trt_modulated_deform_conv_kernel.cu +++ b/csrc/backend_ops/tensorrt/modulated_deform_conv/trt_modulated_deform_conv_kernel.cu @@ -3,7 +3,7 @@ #include #include "common_cuda_helper.hpp" -#include "trt_modulated_deform_conv_kernel.hpp" +#include "modulated_deform_conv/modulated_deform_conv_cuda.cuh" #include "trt_plugin_helper.hpp" template diff --git a/csrc/backend_ops/torchscript/CMakeLists.txt b/csrc/backend_ops/torchscript/CMakeLists.txt new file mode 100644 index 0000000000..e383129992 --- /dev/null +++ b/csrc/backend_ops/torchscript/CMakeLists.txt @@ -0,0 +1,5 @@ +# Copyright (c) OpenMMLab. All rights reserved. +cmake_minimum_required(VERSION 3.14) + +add_subdirectory(ops) +add_subdirectory(optimizer) diff --git a/csrc/backend_ops/torchscript/bind.cpp b/csrc/backend_ops/torchscript/bind.cpp new file mode 100644 index 0000000000..cfc08d1489 --- /dev/null +++ b/csrc/backend_ops/torchscript/bind.cpp @@ -0,0 +1,10 @@ +// Copyright (c) OpenMMLab. All rights reserved. +#include "torch/script.h" + +TORCH_LIBRARY(mmdeploy, m) { + m.def( + "modulated_deform_conv(Tensor input, Tensor weight, Tensor bias, Tensor offset, Tensor " + "mask, " + "int kernel_h, int kernel_w, int stride_h, int stride_w, int pad_h, int pad_w, int " + "dilation_h,int dilation_w, int groups, int deform_groups, bool with_bias) -> Tensor"); +} diff --git a/csrc/backend_ops/torchscript/ops/CMakeLists.txt b/csrc/backend_ops/torchscript/ops/CMakeLists.txt new file mode 100644 index 0000000000..71c7256cd4 --- /dev/null +++ b/csrc/backend_ops/torchscript/ops/CMakeLists.txt @@ -0,0 +1,36 @@ +# Copyright (c) OpenMMLab. All rights reserved. +cmake_minimum_required(VERSION 3.14) + +if("cuda" IN_LIST MMDEPLOY_TARGET_DEVICES) + project(mmdeploy_torchscript_ops CUDA CXX) + include(${CMAKE_SOURCE_DIR}/cmake/cuda.cmake NO_POLICY_SCOPE) + file(GLOB_RECURSE BACKEND_OPS_SRCS *.cpp *.cu) +else() + project(mmdeploy_torchscript_ops CXX) + file(GLOB_RECURSE BACKEND_OPS_SRCS *.cpp) +endif() + +include(${CMAKE_SOURCE_DIR}/cmake/MMDeploy.cmake) +find_package(Torch REQUIRED) + + +add_library(${PROJECT_NAME}_obj OBJECT "${BACKEND_OPS_SRCS}") +set_target_properties(${PROJECT_NAME}_obj PROPERTIES POSITION_INDEPENDENT_CODE 1) +target_compile_definitions(${PROJECT_NAME}_obj + PRIVATE -DTHRUST_IGNORE_DEPRECATED_CPP_DIALECT=1) +target_include_directories(${PROJECT_NAME}_obj + PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/../../common) +target_include_directories(${PROJECT_NAME}_obj + PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/common) + +if("cuda" IN_LIST MMDEPLOY_TARGET_DEVICES) + target_include_directories(${PROJECT_NAME}_obj + PRIVATE ${CUDA_TOOLKIT_ROOT_DIR}/include) +endif() +target_link_libraries(${PROJECT_NAME}_obj PRIVATE ${TORCH_LIBRARIES}) +mmdeploy_export(${PROJECT_NAME}_obj) + +# Build module library. It is used to inference with torchscript +mmdeploy_add_module(${PROJECT_NAME} MODULE EXCLUDE "") +target_link_libraries(${PROJECT_NAME} PUBLIC ${PROJECT_NAME}_obj) +add_library(mmdeploy::torchscript_ops ALIAS ${PROJECT_NAME}) diff --git a/csrc/backend_ops/torchscript/ops/modulated_deform_conv/modulated_deform_conv_cpu.cpp b/csrc/backend_ops/torchscript/ops/modulated_deform_conv/modulated_deform_conv_cpu.cpp new file mode 100644 index 0000000000..c6d980919f --- /dev/null +++ b/csrc/backend_ops/torchscript/ops/modulated_deform_conv/modulated_deform_conv_cpu.cpp @@ -0,0 +1,94 @@ +// Copyright (c) OpenMMLab. All rights reserved. +#include "modulated_deform_conv/modulated_deform_conv_cpu.h" + +#include "torch/script.h" + +namespace mmdeploy { + +void modulated_deformable_im2col_cpu( + const at::Tensor data_im, const at::Tensor data_offset, const at::Tensor data_mask, + const int64_t batch_size, const int64_t channels, const int64_t height_im, + const int64_t width_im, const int64_t height_col, const int64_t width_col, + const int64_t kernel_h, const int64_t kernel_w, const int64_t pad_h, const int64_t pad_w, + const int64_t stride_h, const int64_t stride_w, const int64_t dilation_h, + const int64_t dilation_w, int64_t deformable_group, at::Tensor data_col) { + // num_axes should be smaller than block size + + AT_DISPATCH_FLOATING_TYPES_AND_HALF( + data_im.scalar_type(), "modulated_deformable_im2col_cpu", ([&] { + const scalar_t *data_im_ = data_im.data_ptr(); + const scalar_t *data_offset_ = data_offset.data_ptr(); + const scalar_t *data_mask_ = data_mask.data_ptr(); + scalar_t *data_col_ = data_col.data_ptr(); + + deformable_im2col_2d(data_im_, data_offset_, data_mask_, height_im, width_im, + kernel_h, kernel_w, pad_h, pad_w, stride_h, stride_w, + dilation_h, dilation_w, channels, deformable_group, + height_col, width_col, data_mask_ != nullptr, data_col_); + })); +} + +at::Tensor modulated_deform_conv_forward_cpu(at::Tensor input, at::Tensor weight, at::Tensor bias, + at::Tensor offset, at::Tensor mask, int64_t kernel_h, + int64_t kernel_w, int64_t stride_h, int64_t stride_w, + int64_t pad_h, int64_t pad_w, int64_t dilation_h, + int64_t dilation_w, int64_t group, + int64_t deformable_group, bool with_bias) { + at::DeviceGuard guard(input.device()); + + const int batch = input.size(0); + const int channels = input.size(1); + const int height = input.size(2); + const int width = input.size(3); + + const int channels_out = weight.size(0); + const int channels_kernel = weight.size(1); + const int kernel_h_ = weight.size(2); + const int kernel_w_ = weight.size(3); + + if (kernel_h_ != kernel_h || kernel_w_ != kernel_w) + AT_ERROR("Input shape and kernel shape won't match: (%d x %d vs %d x %d).", kernel_h_, kernel_w, + kernel_h_, kernel_w_); + if (channels != channels_kernel * group) + AT_ERROR("Input shape and kernel channels won't match: (%d vs %d).", channels, + channels_kernel * group); + + const int height_out = (height + 2 * pad_h - (dilation_h * (kernel_h - 1) + 1)) / stride_h + 1; + const int width_out = (width + 2 * pad_w - (dilation_w * (kernel_w - 1) + 1)) / stride_w + 1; + + // resize output + at::Tensor output = + at::zeros({batch, group, channels_out / group, height_out, width_out}, input.options()); + // resize temporary columns + at::Tensor columns = at::zeros( + {group, channels * kernel_h * kernel_w / group, 1 * height_out * width_out}, input.options()); + + // divide into group + weight = + weight.view({group, weight.size(0) / group, weight.size(1), weight.size(2), weight.size(3)}); + for (int b = 0; b < batch; b++) { + modulated_deformable_im2col_cpu(input[b], offset[b], mask[b], 1, channels, height, width, + height_out, width_out, kernel_h, kernel_w, pad_h, pad_w, + stride_h, stride_w, dilation_h, dilation_w, deformable_group, + columns); + + for (int g = 0; g < group; g++) { + output[b][g] = + output[b][g].flatten(1).addmm_(weight[g].flatten(1), columns[g]).view_as(output[b][g]); + } + } + + output = output.view( + {output.size(0), output.size(1) * output.size(2), output.size(3), output.size(4)}); + + if (with_bias) { + output += bias.view({1, bias.size(0), 1, 1}); + } + + return output; +} + +TORCH_LIBRARY_IMPL(mmdeploy, CPU, m) { + m.impl("modulated_deform_conv", modulated_deform_conv_forward_cpu); +} +} // namespace mmdeploy diff --git a/csrc/backend_ops/torchscript/ops/modulated_deform_conv/modulated_deform_conv_cuda.cu b/csrc/backend_ops/torchscript/ops/modulated_deform_conv/modulated_deform_conv_cuda.cu new file mode 100644 index 0000000000..3f9b6aef08 --- /dev/null +++ b/csrc/backend_ops/torchscript/ops/modulated_deform_conv/modulated_deform_conv_cuda.cu @@ -0,0 +1,97 @@ +// Copyright (c) OpenMMLab. All rights reserved. +#include "c10/cuda/CUDAStream.h" +#include "modulated_deform_conv/modulated_deform_conv_cuda.cuh" +#include "torch/script.h" + +namespace mmdeploy { + +void modulated_deformable_im2col_cuda( + const at::Tensor data_im, const at::Tensor data_offset, const at::Tensor data_mask, + const int64_t batch_size, const int64_t channels, const int64_t height_im, + const int64_t width_im, const int64_t height_col, const int64_t width_col, + const int64_t kernel_h, const int64_t kernel_w, const int64_t pad_h, const int64_t pad_w, + const int64_t stride_h, const int64_t stride_w, const int64_t dilation_h, + const int64_t dilation_w, const int64_t deformable_group, at::Tensor data_col) { + // num_axes should be smaller than block size + const int channel_per_deformable_group = channels / deformable_group; + const int num_kernels = channels * batch_size * height_col * width_col; + + AT_DISPATCH_FLOATING_TYPES_AND_HALF( + data_im.scalar_type(), "modulated_deformable_im2col_cuda", ([&] { + const scalar_t *data_im_ = data_im.data_ptr(); + const scalar_t *data_offset_ = data_offset.data_ptr(); + const scalar_t *data_mask_ = data_mask.data_ptr(); + scalar_t *data_col_ = data_col.data_ptr(); + modulated_deformable_im2col_gpu_kernel + <<>>( + num_kernels, data_im_, data_offset_, data_mask_, height_im, width_im, kernel_h, + kernel_w, pad_h, pad_w, stride_h, stride_w, dilation_h, dilation_w, + channel_per_deformable_group, batch_size, channels, deformable_group, height_col, + width_col, data_col_); + })); +} + +at::Tensor modulated_deform_conv_forward_cuda(at::Tensor input, at::Tensor weight, at::Tensor bias, + at::Tensor offset, at::Tensor mask, int64_t kernel_h, + int64_t kernel_w, int64_t stride_h, int64_t stride_w, + int64_t pad_h, int64_t pad_w, int64_t dilation_h, + int64_t dilation_w, int64_t group, + int64_t deformable_group, bool with_bias) { + at::DeviceGuard guard(input.device()); + + const int batch = input.size(0); + const int channels = input.size(1); + const int height = input.size(2); + const int width = input.size(3); + + const int channels_out = weight.size(0); + const int channels_kernel = weight.size(1); + const int kernel_h_ = weight.size(2); + const int kernel_w_ = weight.size(3); + + if (kernel_h_ != kernel_h || kernel_w_ != kernel_w) + AT_ERROR("Input shape and kernel shape won't match: (%d x %d vs %d x %d).", kernel_h_, kernel_w, + kernel_h_, kernel_w_); + if (channels != channels_kernel * group) + AT_ERROR("Input shape and kernel channels won't match: (%d vs %d).", channels, + channels_kernel * group); + + const int height_out = (height + 2 * pad_h - (dilation_h * (kernel_h - 1) + 1)) / stride_h + 1; + const int width_out = (width + 2 * pad_w - (dilation_w * (kernel_w - 1) + 1)) / stride_w + 1; + + // resize output + at::Tensor output = + at::zeros({batch, group, channels_out / group, height_out, width_out}, input.options()); + // resize temporary columns + at::Tensor columns = at::zeros( + {group, channels * kernel_h * kernel_w / group, 1 * height_out * width_out}, input.options()); + + // divide into group + weight = + weight.view({group, weight.size(0) / group, weight.size(1), weight.size(2), weight.size(3)}); + for (int b = 0; b < batch; b++) { + modulated_deformable_im2col_cuda(input[b], offset[b], mask[b], 1, channels, height, width, + height_out, width_out, kernel_h, kernel_w, pad_h, pad_w, + stride_h, stride_w, dilation_h, dilation_w, deformable_group, + columns); + + for (int g = 0; g < group; g++) { + output[b][g] = + output[b][g].flatten(1).addmm_(weight[g].flatten(1), columns[g]).view_as(output[b][g]); + } + } + + output = output.view( + {output.size(0), output.size(1) * output.size(2), output.size(3), output.size(4)}); + + if (with_bias) { + output += bias.view({1, bias.size(0), 1, 1}); + } + + return output; +} + +TORCH_LIBRARY_IMPL(mmdeploy, CUDA, m) { + m.impl("modulated_deform_conv", modulated_deform_conv_forward_cuda); +} +} // namespace mmdeploy diff --git a/csrc/backend_ops/torchscript/optimizer/CMakeLists.txt b/csrc/backend_ops/torchscript/optimizer/CMakeLists.txt new file mode 100644 index 0000000000..8f3cb46d71 --- /dev/null +++ b/csrc/backend_ops/torchscript/optimizer/CMakeLists.txt @@ -0,0 +1,17 @@ +# Copyright (c) OpenMMLab. All rights reserved. +cmake_minimum_required(VERSION 3.14) +project(ts_optimizer) + +find_package(Torch REQUIRED) +if (NOT TARGET pybind11) + add_subdirectory(${CMAKE_SOURCE_DIR}/third_party/pybind11 pybind11) +endif () + +file(GLOB_RECURSE OPTIMIZER_SRCS *.cpp) + +pybind11_add_module(${PROJECT_NAME} ${OPTIMIZER_SRCS}) +target_link_libraries(${PROJECT_NAME} PRIVATE ${TORCH_LIBRARIES}) +target_link_directories(${PROJECT_NAME} PRIVATE mmdeploy::torchscript_ops) +set_target_properties( + ${PROJECT_NAME} PROPERTIES LIBRARY_OUTPUT_DIRECTORY + ${CMAKE_SOURCE_DIR}/mmdeploy/backend/torchscript) diff --git a/csrc/backend_ops/torchscript/optimizer/bind.cpp b/csrc/backend_ops/torchscript/optimizer/bind.cpp new file mode 100644 index 0000000000..73594776a3 --- /dev/null +++ b/csrc/backend_ops/torchscript/optimizer/bind.cpp @@ -0,0 +1,26 @@ +// Copyright (c) OpenMMLab. All rights reserved. +#include + +#include + +#include "optimizer.h" + +void optimize_for_backend(torch::jit::Module& model, const std::string& ir = "torchscript", + const std::string& backend = "torchscript") { + if (ir == "torchscript") { + model = mmdeploy::optimize_for_torchscript(model); + } else if (ir == "onnx") { + model = mmdeploy::optimize_for_onnx(model); + } else { + fprintf(stderr, "No optimize for combination ir: %s backend: %s\n", ir.c_str(), + backend.c_str()); + exit(-1); + } +} + +PYBIND11_MODULE(ts_optimizer, m) { + namespace py = pybind11; + m.def("optimize_for_backend", optimize_for_backend, py::arg("module"), + py::arg("ir") = std::string("torchscript"), + py::arg("backend") = std::string("torchscript")); +} diff --git a/csrc/backend_ops/torchscript/optimizer/optimizer.cpp b/csrc/backend_ops/torchscript/optimizer/optimizer.cpp new file mode 100644 index 0000000000..05ef9d54cd --- /dev/null +++ b/csrc/backend_ops/torchscript/optimizer/optimizer.cpp @@ -0,0 +1,70 @@ +// Copyright (c) OpenMMLab. All rights reserved. +#include "optimizer.h" + +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#if TORCH_VERSION_MINOR >= 9 +#include +#include +#include +#endif + +namespace mmdeploy { + +using torch::jit::Graph; +const std::shared_ptr& required_passes(const std::shared_ptr& graph) { + RemoveExpands(graph); + CanonicalizeOps(graph); + EliminateDeadCode(graph); + return graph; +} + +Module optimize_for_torchscript(const Module& model) { + auto frozen_model = freeze_module(model); + auto graph = frozen_model.get_method("forward").graph(); + OptimizeFrozenGraph(graph, true); + +#if TORCH_VERSION_MINOR >= 9 + FuseFrozenConvAddRelu(graph); + ConvertFrozenOpsToMKLDNN(graph); + FrozenLinearTranspose(graph); +#endif + + graph = required_passes(graph); + EliminateCommonSubexpression(graph); + PeepholeOptimize(graph); + ConstantPropagation(graph); + ConstantPooling(graph); + + // TODO: add more custom passes + + return frozen_model; +} + +Module optimize_for_onnx(const Module& model) { + auto frozen_model = freeze_module(model, {"training"}); + auto graph = frozen_model.get_method("forward").graph(); + OptimizeFrozenGraph(graph, true); + +#if TORCH_VERSION_MINOR >= 9 + FuseFrozenConvAddRelu(graph); + ConvertFrozenOpsToMKLDNN(graph); + FrozenLinearTranspose(graph); +#endif + + // TODO: add more custom passes + + return frozen_model; +} + +// TODO: add optimizer for other backend/onnx + +} // namespace mmdeploy diff --git a/csrc/backend_ops/torchscript/optimizer/optimizer.h b/csrc/backend_ops/torchscript/optimizer/optimizer.h new file mode 100644 index 0000000000..d0d91c627d --- /dev/null +++ b/csrc/backend_ops/torchscript/optimizer/optimizer.h @@ -0,0 +1,10 @@ +// Copyright (c) OpenMMLab. All rights reserved. +#include + +namespace mmdeploy { +using torch::jit::script::Module; + +Module optimize_for_torchscript(const Module &model); + +Module optimize_for_onnx(const Module &model); +} // namespace mmdeploy diff --git a/csrc/codebase/CMakeLists.txt b/csrc/codebase/CMakeLists.txt index 023be7668d..9ef6490a8c 100644 --- a/csrc/codebase/CMakeLists.txt +++ b/csrc/codebase/CMakeLists.txt @@ -9,8 +9,8 @@ if ("all" IN_LIST MMDEPLOY_CODEBASES) list(APPEND CODEBASES "mmseg") list(APPEND CODEBASES "mmocr") list(APPEND CODEBASES "mmedit") -else() - set (CODEBASES ${MMDEPLOY_CODEBASES}) +else () + set(CODEBASES ${MMDEPLOY_CODEBASES}) endif () foreach (codebase IN LISTS CODEBASES) diff --git a/csrc/codebase/common.h b/csrc/codebase/common.h index b0b164ddb7..c815aa11f7 100644 --- a/csrc/codebase/common.h +++ b/csrc/codebase/common.h @@ -14,7 +14,7 @@ namespace mmdeploy { class Context { public: explicit Context(const Value& config) { - DEBUG("config: {}", config); + MMDEPLOY_DEBUG("config: {}", config); device_ = config["context"]["device"].get(); stream_ = config["context"]["stream"].get(); } @@ -35,17 +35,17 @@ class CodebaseCreator : public Creator { std::unique_ptr Create(const Value& cfg) override { constexpr auto key{"component"}; if (!cfg.contains(key)) { - ERROR("no key '{}' in config {}", key, cfg); + MMDEPLOY_ERROR("no key '{}' in config {}", key, cfg); throw_exception(eInvalidArgument); } if (!cfg[key].is_string()) { - ERROR("key '{}' is not a string", key); + MMDEPLOY_ERROR("key '{}' is not a string", key); throw_exception(eInvalidArgument); } auto postprocess_type = cfg[key].get(); auto creator = Registry::Get().GetCreator(postprocess_type); if (creator == nullptr) { - ERROR("could not found entry '{}' in {}", postprocess_type, Tag::name); + MMDEPLOY_ERROR("could not found entry '{}' in {}", postprocess_type, Tag::name); throw_exception(eEntryNotFound); } return creator->Create(cfg); diff --git a/csrc/codebase/mmcls/CMakeLists.txt b/csrc/codebase/mmcls/CMakeLists.txt index c2b254149c..259b653cd1 100644 --- a/csrc/codebase/mmcls/CMakeLists.txt +++ b/csrc/codebase/mmcls/CMakeLists.txt @@ -2,10 +2,8 @@ cmake_minimum_required(VERSION 3.14) project(mmdeploy_mmcls) -include(${CMAKE_SOURCE_DIR}/cmake/common.cmake) +include(${CMAKE_SOURCE_DIR}/cmake/MMDeploy.cmake) file(GLOB_RECURSE SRCS ${CMAKE_CURRENT_SOURCE_DIR} "*.cpp") -build_target(${PROJECT_NAME} "${SRCS}") -target_link_libraries(${PROJECT_NAME} PRIVATE mmdeploy::core) +mmdeploy_add_module(${PROJECT_NAME} "${SRCS}") add_library(mmdeploy::mmcls ALIAS ${PROJECT_NAME}) -export_module(${PROJECT_NAME}) diff --git a/csrc/codebase/mmcls/linear_cls.cpp b/csrc/codebase/mmcls/linear_cls.cpp index 07704cd088..8b14f4e926 100644 --- a/csrc/codebase/mmcls/linear_cls.cpp +++ b/csrc/codebase/mmcls/linear_cls.cpp @@ -1,5 +1,6 @@ // Copyright (c) OpenMMLab. All rights reserved. +#include #include #include "codebase/mmcls/mmcls.h" @@ -18,19 +19,19 @@ class LinearClsHead : public MMClassification { if (cfg.contains("params")) { topk_ = cfg["params"].value("topk", 1); if (topk_ <= 0) { - ERROR("'topk' should be greater than 0, but got '{}'", topk_); + MMDEPLOY_ERROR("'topk' should be greater than 0, but got '{}'", topk_); throw_exception(eInvalidArgument); } } } Result operator()(const Value& infer_res) { - DEBUG("infer_res: {}", infer_res); + MMDEPLOY_DEBUG("infer_res: {}", infer_res); auto output = infer_res["output"].get(); if (!(output.shape().size() >= 2 && output.data_type() == DataType::kFLOAT)) { - ERROR("unsupported `output` tensor, shape: {}, dtype: {}", output.shape(), - (int)output.data_type()); + MMDEPLOY_ERROR("unsupported `output` tensor, shape: {}, dtype: {}", output.shape(), + (int)output.data_type()); return Status(eNotSupported); } @@ -53,7 +54,7 @@ class LinearClsHead : public MMClassification { [&](int i, int j) { return scores_data[i] > scores_data[j]; }); for (int i = 0; i < topk_; ++i) { auto label = ClassifyOutput::Label{idx[i], scores_data[idx[i]]}; - DEBUG("label_id: {}, score: {}", label.label_id, label.score); + MMDEPLOY_DEBUG("label_id: {}, score: {}", label.label_id, label.score); output.labels.push_back(label); } return to_value(std::move(output)); diff --git a/csrc/codebase/mmcls/mmcls.cpp b/csrc/codebase/mmcls/mmcls.cpp index cd59907244..973a4c6d80 100644 --- a/csrc/codebase/mmcls/mmcls.cpp +++ b/csrc/codebase/mmcls/mmcls.cpp @@ -2,8 +2,12 @@ #include "codebase/mmcls/mmcls.h" -namespace mmdeploy::mmcls { +namespace mmdeploy { +namespace mmcls { REGISTER_CODEBASE(MMClassification); -} // namespace mmdeploy::mmcls +} + +MMDEPLOY_DEFINE_REGISTRY(mmcls::MMClassification); +} // namespace mmdeploy diff --git a/csrc/codebase/mmcls/mmcls.h b/csrc/codebase/mmcls/mmcls.h index 5cddfd197d..2b87b2d538 100644 --- a/csrc/codebase/mmcls/mmcls.h +++ b/csrc/codebase/mmcls/mmcls.h @@ -8,7 +8,8 @@ #include "core/module.h" #include "core/serialization.h" -namespace mmdeploy::mmcls { +namespace mmdeploy { +namespace mmcls { struct ClassifyOutput { struct Label { @@ -21,7 +22,9 @@ struct ClassifyOutput { }; DECLARE_CODEBASE(MMClassification, mmcls); +} // namespace mmcls -} // namespace mmdeploy::mmcls +MMDEPLOY_DECLARE_REGISTRY(mmcls::MMClassification); +} // namespace mmdeploy #endif // MMDEPLOY_SRC_CODEBASE_MMCLS_MMCLS_H_ diff --git a/csrc/codebase/mmdet/CMakeLists.txt b/csrc/codebase/mmdet/CMakeLists.txt index 31a00813bb..75ae6c4d4b 100644 --- a/csrc/codebase/mmdet/CMakeLists.txt +++ b/csrc/codebase/mmdet/CMakeLists.txt @@ -3,10 +3,11 @@ cmake_minimum_required(VERSION 3.14) project(mmdeploy_mmdet) include(${CMAKE_SOURCE_DIR}/cmake/opencv.cmake) -include(${CMAKE_SOURCE_DIR}/cmake/common.cmake) +include(${CMAKE_SOURCE_DIR}/cmake/MMDeploy.cmake) file(GLOB_RECURSE SRCS ${CMAKE_CURRENT_SOURCE_DIR} "*.cpp") -build_target(${PROJECT_NAME} "${SRCS}") -target_link_libraries(${PROJECT_NAME} PRIVATE mmdeploy::core opencv_core) +mmdeploy_add_module(${PROJECT_NAME} "${SRCS}") +target_link_libraries(${PROJECT_NAME} + PRIVATE mmdeploy_opencv_utils) + add_library(mmdeploy::mmdet ALIAS ${PROJECT_NAME}) -export_module(${PROJECT_NAME}) diff --git a/csrc/codebase/mmdet/instance_segmentation.cpp b/csrc/codebase/mmdet/instance_segmentation.cpp index 481e0b1e23..638ce0b80a 100644 --- a/csrc/codebase/mmdet/instance_segmentation.cpp +++ b/csrc/codebase/mmdet/instance_segmentation.cpp @@ -5,7 +5,7 @@ #include "experimental/module_adapter.h" #include "object_detection.h" #include "opencv2/imgproc/imgproc.hpp" -#include "preprocess/cpu/opencv_utils.h" +#include "opencv_utils.h" namespace mmdeploy::mmdet { @@ -19,35 +19,35 @@ class ResizeInstanceMask : public ResizeBBox { // TODO: remove duplication Result operator()(const Value& prep_res, const Value& infer_res) { - DEBUG("prep_res: {}\ninfer_res: {}", prep_res, infer_res); + MMDEPLOY_DEBUG("prep_res: {}\ninfer_res: {}", prep_res, infer_res); try { auto dets = infer_res["dets"].get(); auto labels = infer_res["labels"].get(); auto masks = infer_res["masks"].get(); - DEBUG("dets.shape: {}", dets.shape()); - DEBUG("labels.shape: {}", labels.shape()); - DEBUG("masks.shape: {}", masks.shape()); + MMDEPLOY_DEBUG("dets.shape: {}", dets.shape()); + MMDEPLOY_DEBUG("labels.shape: {}", labels.shape()); + MMDEPLOY_DEBUG("masks.shape: {}", masks.shape()); // `dets` is supposed to have 3 dims. They are 'batch', 'bboxes_number' // and 'channels' respectively if (!(dets.shape().size() == 3 && dets.data_type() == DataType::kFLOAT)) { - ERROR("unsupported `dets` tensor, shape: {}, dtype: {}", dets.shape(), - (int)dets.data_type()); + MMDEPLOY_ERROR("unsupported `dets` tensor, shape: {}, dtype: {}", dets.shape(), + (int)dets.data_type()); return Status(eNotSupported); } // `labels` is supposed to have 2 dims, which are 'batch' and // 'bboxes_number' if (labels.shape().size() != 2) { - ERROR("unsupported `labels`, tensor, shape: {}, dtype: {}", labels.shape(), - (int)labels.data_type()); + MMDEPLOY_ERROR("unsupported `labels`, tensor, shape: {}, dtype: {}", labels.shape(), + (int)labels.data_type()); return Status(eNotSupported); } if (!(masks.shape().size() == 4 && masks.data_type() == DataType::kFLOAT)) { - ERROR("unsupported `mask` tensor, shape: {}, dtype: {}", masks.shape(), - (int)masks.data_type()); + MMDEPLOY_ERROR("unsupported `mask` tensor, shape: {}, dtype: {}", masks.shape(), + (int)masks.data_type()); return Status(eNotSupported); } @@ -65,7 +65,7 @@ class ResizeInstanceMask : public ResizeBBox { return to_value(result); } catch (const std::exception& e) { - ERROR("{}", e.what()); + MMDEPLOY_ERROR("{}", e.what()); return Status(eFail); } } @@ -74,10 +74,10 @@ class ResizeInstanceMask : public ResizeBBox { void ProcessMasks(DetectorOutput& result, Tensor cpu_masks, int img_w, int img_h) const { auto shape = TensorShape{cpu_masks.shape(1), cpu_masks.shape(2), cpu_masks.shape(3)}; cpu_masks.Reshape(shape); - + MMDEPLOY_DEBUG("{}, {}", cpu_masks.shape(), cpu_masks.data_type()); for (auto& det : result.detections) { auto mask = cpu_masks.Slice(det.index); - cv::Mat mask_mat(mask.shape(1), mask.shape(2), CV_32F, mask.data()); + cv::Mat mask_mat((int)mask.shape(1), (int)mask.shape(2), CV_32F, mask.data()); cv::Mat warped_mask; auto& bbox = det.bbox; // same as mmdet with skip_empty = True @@ -97,7 +97,9 @@ class ResizeInstanceMask : public ResizeBBox { cv::warpAffine(mask_mat, warped_mask, m, cv::Size{width, height}, cv::INTER_LINEAR | cv::WARP_INVERSE_MAP); warped_mask = warped_mask > mask_thr_binary_; - det.mask = cpu::CVMat2Mat(warped_mask, PixelFormat::kGRAYSCALE); + + det.mask = Mat(height, width, PixelFormat::kGRAYSCALE, DataType::kINT8, + std::shared_ptr(warped_mask.data, [mat = warped_mask](void*) {})); } } diff --git a/csrc/codebase/mmdet/mmdet.cpp b/csrc/codebase/mmdet/mmdet.cpp index 218b73c502..45fe21f9b0 100644 --- a/csrc/codebase/mmdet/mmdet.cpp +++ b/csrc/codebase/mmdet/mmdet.cpp @@ -2,8 +2,12 @@ #include "codebase/mmdet/mmdet.h" -namespace mmdeploy::mmdet { +namespace mmdeploy { +namespace mmdet { REGISTER_CODEBASE(MMDetection); -} // namespace mmdeploy::mmdet +} + +MMDEPLOY_DEFINE_REGISTRY(mmdet::MMDetection); +} // namespace mmdeploy diff --git a/csrc/codebase/mmdet/mmdet.h b/csrc/codebase/mmdet/mmdet.h index a663004038..dcb1e18194 100644 --- a/csrc/codebase/mmdet/mmdet.h +++ b/csrc/codebase/mmdet/mmdet.h @@ -3,13 +3,17 @@ #ifndef MMDEPLOY_SRC_CODEBASE_MMDET_MMDET_H_ #define MMDEPLOY_SRC_CODEBASE_MMDET_MMDET_H_ +#include + #include "codebase/common.h" #include "core/device.h" #include "core/mat.h" #include "core/module.h" +#include "core/registry.h" #include "core/serialization.h" -namespace mmdeploy::mmdet { +namespace mmdeploy { +namespace mmdet { struct DetectorOutput { struct Detection { @@ -25,7 +29,9 @@ struct DetectorOutput { }; DECLARE_CODEBASE(MMDetection, mmdet); +} // namespace mmdet -} // namespace mmdeploy::mmdet +MMDEPLOY_DECLARE_REGISTRY(mmdet::MMDetection); +} // namespace mmdeploy #endif // MMDEPLOY_SRC_CODEBASE_MMDET_MMDET_H_ diff --git a/csrc/codebase/mmdet/object_detection.cpp b/csrc/codebase/mmdet/object_detection.cpp index 6a7c6d6a37..62a5c21017 100644 --- a/csrc/codebase/mmdet/object_detection.cpp +++ b/csrc/codebase/mmdet/object_detection.cpp @@ -17,26 +17,27 @@ ResizeBBox::ResizeBBox(const Value& cfg) : MMDetection(cfg) { } } Result ResizeBBox::operator()(const Value& prep_res, const Value& infer_res) { - DEBUG("prep_res: {}\ninfer_res: {}", prep_res, infer_res); + MMDEPLOY_DEBUG("prep_res: {}\ninfer_res: {}", prep_res, infer_res); try { auto dets = infer_res["dets"].get(); auto labels = infer_res["labels"].get(); - DEBUG("dets.shape: {}", dets.shape()); - DEBUG("labels.shape: {}", labels.shape()); + MMDEPLOY_DEBUG("dets.shape: {}", dets.shape()); + MMDEPLOY_DEBUG("labels.shape: {}", labels.shape()); // `dets` is supposed to have 3 dims. They are 'batch', 'bboxes_number' // and 'channels' respectively if (!(dets.shape().size() == 3 && dets.data_type() == DataType::kFLOAT)) { - ERROR("unsupported `dets` tensor, shape: {}, dtype: {}", dets.shape(), (int)dets.data_type()); + MMDEPLOY_ERROR("unsupported `dets` tensor, shape: {}, dtype: {}", dets.shape(), + (int)dets.data_type()); return Status(eNotSupported); } // `labels` is supposed to have 2 dims, which are 'batch' and // 'bboxes_number' if (labels.shape().size() != 2) { - ERROR("unsupported `labels`, tensor, shape: {}, dtype: {}", labels.shape(), - (int)labels.data_type()); + MMDEPLOY_ERROR("unsupported `labels`, tensor, shape: {}, dtype: {}", labels.shape(), + (int)labels.data_type()); return Status(eNotSupported); } @@ -98,16 +99,17 @@ Result ResizeBBox::GetBBoxes(const Value& prep_res, const Tensor auto right = dets_ptr[2]; auto bottom = dets_ptr[3]; - DEBUG("ori left {}, top {}, right {}, bottom {}, label {}", left, top, right, bottom, - *labels_ptr); + MMDEPLOY_DEBUG("ori left {}, top {}, right {}, bottom {}, label {}", left, top, right, bottom, + *labels_ptr); auto rect = MapToOriginImage(left, top, right, bottom, scale_factor.data(), w_offset, h_offset, ori_width, ori_height); if (rect[2] - rect[0] < min_bbox_size_ || rect[3] - rect[1] < min_bbox_size_) { - DEBUG("ignore small bbox with width '{}' and height '{}", rect[2] - rect[0], - rect[3] - rect[1]); + MMDEPLOY_DEBUG("ignore small bbox with width '{}' and height '{}", rect[2] - rect[0], + rect[3] - rect[1]); continue; } - DEBUG("remap left {}, top {}, right {}, bottom {}", rect[0], rect[1], rect[2], rect[3]); + MMDEPLOY_DEBUG("remap left {}, top {}, right {}, bottom {}", rect[0], rect[1], rect[2], + rect[3]); DetectorOutput::Detection det{}; det.index = i; det.label_id = static_cast(*labels_ptr); diff --git a/csrc/codebase/mmedit/CMakeLists.txt b/csrc/codebase/mmedit/CMakeLists.txt index a546642551..59646d0f6a 100644 --- a/csrc/codebase/mmedit/CMakeLists.txt +++ b/csrc/codebase/mmedit/CMakeLists.txt @@ -3,10 +3,9 @@ cmake_minimum_required(VERSION 3.14) project(mmdeploy_mmedit) include(${CMAKE_SOURCE_DIR}/cmake/opencv.cmake) -include(${CMAKE_SOURCE_DIR}/cmake/common.cmake) +include(${CMAKE_SOURCE_DIR}/cmake/MMDeploy.cmake) file(GLOB_RECURSE SRCS ${CMAKE_CURRENT_SOURCE_DIR} "*.cpp") -build_target(${PROJECT_NAME} "${SRCS}") -target_link_libraries(${PROJECT_NAME} PRIVATE mmdeploy::core opencv_core) +mmdeploy_add_module(${PROJECT_NAME} "${SRCS}") +target_link_libraries(${PROJECT_NAME} PRIVATE opencv_core) add_library(mmdeploy::mmedit ALIAS ${PROJECT_NAME}) -export_module(${PROJECT_NAME}) diff --git a/csrc/codebase/mmedit/mmedit.cpp b/csrc/codebase/mmedit/mmedit.cpp index 28a8dfa1d2..b10c084b24 100644 --- a/csrc/codebase/mmedit/mmedit.cpp +++ b/csrc/codebase/mmedit/mmedit.cpp @@ -4,8 +4,12 @@ #include "core/registry.h" -namespace mmdeploy::mmedit { +namespace mmdeploy { +namespace mmedit { REGISTER_CODEBASE(MMEdit); -} // namespace mmdeploy::mmedit +} // namespace mmedit + +MMDEPLOY_DEFINE_REGISTRY(mmedit::MMEdit); +} // namespace mmdeploy diff --git a/csrc/codebase/mmedit/mmedit.h b/csrc/codebase/mmedit/mmedit.h index ed9c3cdc88..e7c4285134 100644 --- a/csrc/codebase/mmedit/mmedit.h +++ b/csrc/codebase/mmedit/mmedit.h @@ -9,12 +9,16 @@ #include "core/module.h" #include "core/serialization.h" -namespace mmdeploy::mmedit { +namespace mmdeploy { +namespace mmedit { using RestorerOutput = Mat; DECLARE_CODEBASE(MMEdit, mmedit); -} // namespace mmdeploy::mmedit +} // namespace mmedit + +MMDEPLOY_DECLARE_REGISTRY(mmedit::MMEdit); +} // namespace mmdeploy #endif // MMDEPLOY_SRC_CODEBASE_MMEDIT_MMEDIT_H_ diff --git a/csrc/codebase/mmedit/restorer.cpp b/csrc/codebase/mmedit/restorer.cpp index da06075a4d..84626d15be 100644 --- a/csrc/codebase/mmedit/restorer.cpp +++ b/csrc/codebase/mmedit/restorer.cpp @@ -32,8 +32,8 @@ class TensorToImg : public MMEdit { mat_hwc.convertTo(rescale_uint8, CV_8UC(channels), 255.f); return mat; } else { - ERROR("unsupported `output` tensor, shape: {}, dtype: {}", upscale.shape(), - (int)upscale.data_type()); + MMDEPLOY_ERROR("unsupported `output` tensor, shape: {}, dtype: {}", upscale.shape(), + (int)upscale.data_type()); return Status(eNotSupported); } } diff --git a/csrc/codebase/mmocr/CMakeLists.txt b/csrc/codebase/mmocr/CMakeLists.txt index 42e63a0dff..60ac5c6ff4 100644 --- a/csrc/codebase/mmocr/CMakeLists.txt +++ b/csrc/codebase/mmocr/CMakeLists.txt @@ -3,13 +3,13 @@ cmake_minimum_required(VERSION 3.14) project(mmdeploy_mmocr) include(${CMAKE_SOURCE_DIR}/cmake/opencv.cmake) -include(${CMAKE_SOURCE_DIR}/cmake/common.cmake) +include(${CMAKE_SOURCE_DIR}/cmake/MMDeploy.cmake) aux_source_directory(${CMAKE_CURRENT_SOURCE_DIR} OCR_SRCS) aux_source_directory(${CMAKE_SOURCE_DIR}/third_party/clipper CLIPPER_SRCS) set(SRCS ${OCR_SRCS} ${CLIPPER_SRCS}) -build_target(${PROJECT_NAME} "${SRCS}") +mmdeploy_add_module(${PROJECT_NAME} "${SRCS}") target_include_directories(${PROJECT_NAME} PRIVATE ${CMAKE_SOURCE_DIR}/third_party/clipper) -target_link_libraries(${PROJECT_NAME} PRIVATE mmdeploy::core opencv_core opencv_imgproc) +target_link_libraries(${PROJECT_NAME} + PRIVATE mmdeploy::transform mmdeploy_opencv_utils) add_library(mmdeploy::mmocr ALIAS ${PROJECT_NAME}) -export_module(${PROJECT_NAME}) diff --git a/csrc/codebase/mmocr/crnn.cpp b/csrc/codebase/mmocr/crnn.cpp index bd6c4a6179..fc611e8588 100644 --- a/csrc/codebase/mmocr/crnn.cpp +++ b/csrc/codebase/mmocr/crnn.cpp @@ -1,5 +1,6 @@ // Copyright (c) OpenMMLab. All rights reserved. +#include #include #include "core/device.h" @@ -22,7 +23,7 @@ class CTCConvertor : public MMOCR { explicit CTCConvertor(const Value& cfg) : MMOCR(cfg) { auto model = cfg["context"]["model"].get(); if (!cfg.contains("params")) { - ERROR("'params' is required, but it's not in the config"); + MMDEPLOY_ERROR("'params' is required, but it's not in the config"); throw_exception(eInvalidArgument); } // BaseConverter @@ -40,11 +41,11 @@ class CTCConvertor : public MMOCR { } else if (dict_type == "DICT90") { idx2char_ = SplitChars(DICT90); } else { - ERROR("unknown dict_type: {}", dict_type); + MMDEPLOY_ERROR("unknown dict_type: {}", dict_type); throw_exception(eInvalidArgument); } } else { - ERROR("either dict_file, dict_list or dict_type must be specified"); + MMDEPLOY_ERROR("either dict_file, dict_list or dict_type must be specified"); throw_exception(eInvalidArgument); } // CTCConverter @@ -62,8 +63,8 @@ class CTCConvertor : public MMOCR { auto d_conf = _prob["output"].get(); if (!(d_conf.shape().size() == 3 && d_conf.data_type() == DataType::kFLOAT)) { - ERROR("unsupported `output` tensor, shape: {}, dtype: {}", d_conf.shape(), - (int)d_conf.data_type()); + MMDEPLOY_ERROR("unsupported `output` tensor, shape: {}, dtype: {}", d_conf.shape(), + (int)d_conf.data_type()); return Status(eNotSupported); } @@ -80,7 +81,7 @@ class CTCConvertor : public MMOCR { auto [indexes, scores] = Tensor2Idx(data, w, c, valid_ratio); auto text = Idx2Str(indexes); - DEBUG("text: {}", text); + MMDEPLOY_DEBUG("text: {}", text); TextRecognizerOutput output{text, scores}; diff --git a/csrc/codebase/mmocr/dbnet.cpp b/csrc/codebase/mmocr/dbnet.cpp index 93a3d0400b..bd90dca3f3 100644 --- a/csrc/codebase/mmocr/dbnet.cpp +++ b/csrc/codebase/mmocr/dbnet.cpp @@ -13,7 +13,7 @@ #include "core/value.h" #include "experimental/module_adapter.h" #include "mmocr.h" -#include "preprocess/cpu/opencv_utils.h" +#include "opencv_utils.h" namespace mmdeploy::mmocr { @@ -37,21 +37,21 @@ class DBHead : public MMOCR { } Result operator()(const Value& _data, const Value& _prob) { - DEBUG("preprocess_result: {}", _data); - DEBUG("inference_result: {}", _prob); + MMDEPLOY_DEBUG("preprocess_result: {}", _data); + MMDEPLOY_DEBUG("inference_result: {}", _prob); auto img = _data["img"].get(); - DEBUG("img shape: {}", img.shape()); + MMDEPLOY_DEBUG("img shape: {}", img.shape()); Device cpu_device{"cpu"}; OUTCOME_TRY(auto conf, MakeAvailableOnDevice(_prob["output"].get(), cpu_device, stream_)); OUTCOME_TRY(stream_.Wait()); - DEBUG("shape: {}", conf.shape()); + MMDEPLOY_DEBUG("shape: {}", conf.shape()); if (!(conf.shape().size() == 4 && conf.data_type() == DataType::kFLOAT)) { - ERROR("unsupported `output` tensor, shape: {}, dtype: {}", conf.shape(), - (int)conf.data_type()); + MMDEPLOY_ERROR("unsupported `output` tensor, shape: {}, dtype: {}", conf.shape(), + (int)conf.data_type()); return Status(eNotSupported); } @@ -103,7 +103,7 @@ class DBHead : public MMOCR { } else { assert(0); } - DEBUG("score: {}", score); + MMDEPLOY_DEBUG("score: {}", score); // cv::drawContours(score_map, vector>{approx}, -1, 1); vector scaled(begin(approx), end(approx)); @@ -133,7 +133,7 @@ class DBHead : public MMOCR { cv::Mat mask(rect.size(), CV_8U, cv::Scalar(0)); - cv::fillPoly(mask, std::vector{box}, 1, cv::LINE_8, 0, -rect.tl()); + cv::fillPoly(mask, std::vector>{box}, 1, cv::LINE_8, 0, -rect.tl()); auto mean = cv::mean(bitmap(rect), mask)[0]; return static_cast(mean); } diff --git a/csrc/codebase/mmocr/mmocr.cpp b/csrc/codebase/mmocr/mmocr.cpp index 2935f03b5a..f34f918afa 100644 --- a/csrc/codebase/mmocr/mmocr.cpp +++ b/csrc/codebase/mmocr/mmocr.cpp @@ -5,8 +5,12 @@ #include "core/registry.h" #include "core/utils/formatter.h" -namespace mmdeploy::mmocr { +namespace mmdeploy { +namespace mmocr { REGISTER_CODEBASE(MMOCR); -} // namespace mmdeploy::mmocr +} // namespace mmocr + +MMDEPLOY_DEFINE_REGISTRY(mmocr::MMOCR); +} // namespace mmdeploy diff --git a/csrc/codebase/mmocr/mmocr.h b/csrc/codebase/mmocr/mmocr.h index 42098af746..1871b6755c 100644 --- a/csrc/codebase/mmocr/mmocr.h +++ b/csrc/codebase/mmocr/mmocr.h @@ -7,7 +7,8 @@ #include "core/device.h" #include "core/module.h" -namespace mmdeploy::mmocr { +namespace mmdeploy { +namespace mmocr { struct TextDetectorOutput { std::vector> boxes; @@ -23,6 +24,9 @@ struct TextRecognizerOutput { DECLARE_CODEBASE(MMOCR, mmocr); -} // namespace mmdeploy::mmocr +} // namespace mmocr + +MMDEPLOY_DECLARE_REGISTRY(mmocr::MMOCR); +} // namespace mmdeploy #endif // MMDEPLOY_MMOCR_H diff --git a/csrc/codebase/mmocr/resize_ocr.cpp b/csrc/codebase/mmocr/resize_ocr.cpp index 2da9bac7a7..69d7602d18 100644 --- a/csrc/codebase/mmocr/resize_ocr.cpp +++ b/csrc/codebase/mmocr/resize_ocr.cpp @@ -4,12 +4,14 @@ #include "archive/json_archive.h" #include "archive/value_archive.h" +#include "core/registry.h" #include "core/tensor.h" #include "core/utils/device_utils.h" #include "core/utils/formatter.h" #include "opencv2/imgproc.hpp" -#include "preprocess/cpu/opencv_utils.h" +#include "opencv_utils.h" #include "preprocess/transform/resize.h" +#include "preprocess/transform/transform.h" using namespace std; @@ -37,7 +39,7 @@ class ResizeOCRImpl : public Module { ~ResizeOCRImpl() override = default; Result Process(const Value& input) override { - DEBUG("input: {}", input); + MMDEPLOY_DEBUG("input: {}", input); auto dst_height = height_; auto dst_min_width = min_width_; auto dst_max_width = max_width_; @@ -84,7 +86,7 @@ class ResizeOCRImpl : public Module { output["resize_shape"] = to_value(img_resize.desc().shape); output["pad_shape"] = output["resize_shape"]; output["valid_ratio"] = valid_ratio; - DEBUG("output: {}", to_json(output).dump(2)); + MMDEPLOY_DEBUG("output: {}", to_json(output).dump(2)); return output; } @@ -95,7 +97,7 @@ class ResizeOCRImpl : public Module { int h = desc.shape[1]; int w = desc.shape[2]; int c = desc.shape[3]; - assert(c == 3 or c == 1); + assert(c == 3 || c == 1); cv::Mat src_mat, dst_mat; if (3 == c) { // rgb src_mat = cv::Mat(h, w, CV_8UC3, const_cast(img.data())); @@ -135,6 +137,8 @@ class ResizeOCRImplCreator : public Creator { ReturnType Create(const Value& args) override { return std::make_unique(args); } }; +MMDEPLOY_DEFINE_REGISTRY(ResizeOCRImpl); + REGISTER_MODULE(ResizeOCRImpl, ResizeOCRImplCreator); class ResizeOCR : public Transform { diff --git a/csrc/codebase/mmocr/warp.cpp b/csrc/codebase/mmocr/warp.cpp index b54e6a7897..56566f6d12 100644 --- a/csrc/codebase/mmocr/warp.cpp +++ b/csrc/codebase/mmocr/warp.cpp @@ -8,7 +8,7 @@ #include "core/utils/formatter.h" #include "core/value.h" #include "experimental/module_adapter.h" -#include "preprocess/cpu/opencv_utils.h" +#include "opencv_utils.h" namespace mmdeploy { diff --git a/csrc/codebase/mmseg/CMakeLists.txt b/csrc/codebase/mmseg/CMakeLists.txt index 55bb7e9d08..89a15bb32f 100644 --- a/csrc/codebase/mmseg/CMakeLists.txt +++ b/csrc/codebase/mmseg/CMakeLists.txt @@ -3,10 +3,9 @@ cmake_minimum_required(VERSION 3.14) project(mmdeploy_mmseg) include(${CMAKE_SOURCE_DIR}/cmake/opencv.cmake) -include(${CMAKE_SOURCE_DIR}/cmake/common.cmake) +include(${CMAKE_SOURCE_DIR}/cmake/MMDeploy.cmake) file(GLOB_RECURSE SRCS ${CMAKE_CURRENT_SOURCE_DIR} "*.cpp") -build_target(${PROJECT_NAME} "${SRCS}") -target_link_libraries(${PROJECT_NAME} PRIVATE mmdeploy::core opencv_core) +mmdeploy_add_module(${PROJECT_NAME} "${SRCS}") +target_link_libraries(${PROJECT_NAME} PRIVATE mmdeploy_opencv_utils) add_library(mmdeploy::mmseg ALIAS ${PROJECT_NAME}) -export_module(${PROJECT_NAME}) diff --git a/csrc/codebase/mmseg/mmseg.cpp b/csrc/codebase/mmseg/mmseg.cpp index 78e1d490e5..6f080b8221 100644 --- a/csrc/codebase/mmseg/mmseg.cpp +++ b/csrc/codebase/mmseg/mmseg.cpp @@ -4,8 +4,12 @@ using namespace std; -namespace mmdeploy::mmseg { +namespace mmdeploy { +namespace mmseg { REGISTER_CODEBASE(MMSegmentation); -} // namespace mmdeploy::mmseg +} + +MMDEPLOY_DEFINE_REGISTRY(mmseg::MMSegmentation); +} // namespace mmdeploy diff --git a/csrc/codebase/mmseg/mmseg.h b/csrc/codebase/mmseg/mmseg.h index 3685c35bf2..9122047836 100644 --- a/csrc/codebase/mmseg/mmseg.h +++ b/csrc/codebase/mmseg/mmseg.h @@ -8,7 +8,8 @@ #include "core/module.h" #include "core/tensor.h" -namespace mmdeploy::mmseg { +namespace mmdeploy { +namespace mmseg { struct SegmentorOutput { Tensor mask; @@ -20,6 +21,9 @@ struct SegmentorOutput { DECLARE_CODEBASE(MMSegmentation, mmseg); -} // namespace mmdeploy::mmseg +} // namespace mmseg + +MMDEPLOY_DECLARE_REGISTRY(mmseg::MMSegmentation); +} // namespace mmdeploy #endif // MMDEPLOY_MMSEG_H diff --git a/csrc/codebase/mmseg/segment.cpp b/csrc/codebase/mmseg/segment.cpp index 48afa9b572..8d5aeef08e 100644 --- a/csrc/codebase/mmseg/segment.cpp +++ b/csrc/codebase/mmseg/segment.cpp @@ -4,7 +4,7 @@ #include "core/tensor.h" #include "core/utils/device_utils.h" #include "core/utils/formatter.h" -#include "preprocess/cpu/opencv_utils.h" +#include "opencv_utils.h" #include "preprocess/transform/transform.h" namespace mmdeploy::mmseg { @@ -15,19 +15,19 @@ class ResizeMask : public MMSegmentation { try { classes_ = cfg["params"]["num_classes"].get(); } catch (const std::exception &e) { - ERROR("no ['params']['num_classes'] is specified in cfg: {}", cfg); + MMDEPLOY_ERROR("no ['params']['num_classes'] is specified in cfg: {}", cfg); throw_exception(eInvalidArgument); } } Result operator()(const Value &preprocess_result, const Value &inference_result) { - DEBUG("preprocess: {}\ninference: {}", preprocess_result, inference_result); + MMDEPLOY_DEBUG("preprocess: {}\ninference: {}", preprocess_result, inference_result); auto mask = inference_result["output"].get(); - DEBUG("tensor.name: {}, tensor.shape: {}, tensor.data_type: {}", mask.name(), mask.shape(), - mask.data_type()); + MMDEPLOY_DEBUG("tensor.name: {}, tensor.shape: {}, tensor.data_type: {}", mask.name(), + mask.shape(), mask.data_type()); if (!(mask.shape().size() == 4 && mask.shape(0) == 1 && mask.shape(1) == 1)) { - ERROR("unsupported `output` tensor, shape: {}", mask.shape()); + MMDEPLOY_ERROR("unsupported `output` tensor, shape: {}", mask.shape()); return Status(eNotSupported); } @@ -40,16 +40,14 @@ class ResizeMask : public MMSegmentation { OUTCOME_TRY(stream_.Wait()); if (mask.data_type() == DataType::kINT64) { // change kINT64 to 2 INT32 - TensorDesc desc{.device = host_tensor.device(), - .data_type = DataType::kINT32, - .shape = {1, 2, height, width}, - .name = host_tensor.name()}; + TensorDesc desc{ + host_tensor.device(), DataType::kINT32, {1, 2, height, width}, host_tensor.name()}; Tensor _host_tensor(desc, mask.buffer()); return MaskResize(_host_tensor, input_height, input_width); } else if (mask.data_type() == DataType::kINT32) { return MaskResize(host_tensor, input_height, input_width); } else { - ERROR("unsupported `output` tensor, dtype: {}", (int)mask.data_type()); + MMDEPLOY_ERROR("unsupported `output` tensor, dtype: {}", (int)mask.data_type()); return Status(eNotSupported); } } diff --git a/csrc/core/CMakeLists.txt b/csrc/core/CMakeLists.txt index ef539a05b2..19b9a64a97 100644 --- a/csrc/core/CMakeLists.txt +++ b/csrc/core/CMakeLists.txt @@ -11,15 +11,14 @@ reliably on all generators, or if a new generator is added in the future that ca projects using it will be stuck. Even if CONFIGURE_DEPENDS works reliably, there is still a cost to perform the check on every rebuild. #]==] -# file(GLOB_RECURSE CORE_SRCS *.cpp) set(SPDLOG_LIB) find_package(spdlog QUIET) if (spdlog_FOUND) message(STATUS "spdlog is found") - set(SPDLOG_LIB $) + set(SPDLOG_LIB spdlog::spdlog) endif () -include(${CMAKE_SOURCE_DIR}/cmake/common.cmake) +include(${CMAKE_SOURCE_DIR}/cmake/MMDeploy.cmake) set(SRCS device_impl.cpp @@ -32,11 +31,14 @@ set(SRCS operator.cpp status_code.cpp tensor.cpp + registry.cpp utils/device_utils.cpp utils/formatter.cpp utils/stacktrace.cpp) -build_target(${PROJECT_NAME} "${SRCS}") + +mmdeploy_add_library(${PROJECT_NAME} ${SRCS}) target_compile_definitions(${PROJECT_NAME} PUBLIC -DMMDEPLOY_STATUS_USE_SOURCE_LOCATION=1) + target_include_directories(${PROJECT_NAME} PUBLIC $ @@ -45,7 +47,10 @@ target_include_directories(${PROJECT_NAME} $ $ ) -target_link_libraries(${PROJECT_NAME} PUBLIC ${SPDLOG_LIB} stdc++fs) +target_link_libraries(${PROJECT_NAME} PUBLIC ${SPDLOG_LIB}) +if (NOT MSVC) + target_link_libraries(${PROJECT_NAME} PUBLIC stdc++fs) +endif () add_library(mmdeploy::core ALIAS ${PROJECT_NAME}) install(DIRECTORY ${CMAKE_SOURCE_DIR}/csrc/core @@ -57,5 +62,3 @@ install(FILES ${CMAKE_SOURCE_DIR}/third_party/outcome/outcome-experimental.hpp install(DIRECTORY ${CMAKE_SOURCE_DIR}/csrc/experimental DESTINATION include/cpp FILES_MATCHING PATTERN "*.h") - -export_target(${PROJECT_NAME}) diff --git a/csrc/core/device.h b/csrc/core/device.h index 6695ca3807..f337ad4dcd 100644 --- a/csrc/core/device.h +++ b/csrc/core/device.h @@ -10,6 +10,7 @@ #include #include +#include "core/macro.h" #include "core/status_code.h" namespace mmdeploy { @@ -67,7 +68,7 @@ class Device { constexpr explicit Device(int platform_id, int device_id = 0) : platform_id_(platform_id), device_id_(device_id) {} - explicit Device(const char *platform_name, int device_id = 0); + MMDEPLOY_API explicit Device(const char *platform_name, int device_id = 0); constexpr int device_id() const noexcept { return device_id_; } @@ -100,7 +101,7 @@ class Device { enum class MemcpyKind : int { HtoD, DtoH, DtoD }; -class Platform { +class MMDEPLOY_API Platform { public: // throws if not found explicit Platform(const char *platform_name); @@ -133,7 +134,7 @@ Platform GetPlatform(int platform_id); Platform GetPlatform(const char *platform_name); -class Stream { +class MMDEPLOY_API Stream { public: Stream() = default; @@ -187,7 +188,7 @@ T GetNative(Stream &stream, ErrorCode *ec = nullptr) { return reinterpret_cast(stream.GetNative(ec)); } -class Event { +class MMDEPLOY_API Event { public: Event() = default; @@ -226,7 +227,7 @@ T GetNative(Event &event, ErrorCode *ec = nullptr) { return reinterpret_cast(event.GetNative(ec)); } -class Kernel { +class MMDEPLOY_API Kernel { public: Kernel() = default; explicit Kernel(std::shared_ptr impl) : impl_(std::move(impl)) {} @@ -246,7 +247,7 @@ T GetNative(Kernel &kernel, ErrorCode *ec = nullptr) { return reinterpret_cast(kernel.GetNative(ec)); } -class Allocator { +class MMDEPLOY_API Allocator { friend class Access; public: @@ -259,7 +260,7 @@ class Allocator { std::shared_ptr impl_; }; -class Buffer { +class MMDEPLOY_API Buffer { public: Buffer() = default; @@ -304,7 +305,7 @@ T GetNative(const Buffer &buffer, ErrorCode *ec = nullptr) { return reinterpret_cast(buffer.GetNative(ec)); } -class PlatformRegistry { +class MMDEPLOY_API PlatformRegistry { public: using Creator = std::function()>; @@ -332,6 +333,6 @@ class PlatformRegistry { std::vector entries_; }; -PlatformRegistry &gPlatformRegistry(); +MMDEPLOY_API PlatformRegistry &gPlatformRegistry(); } // namespace mmdeploy diff --git a/csrc/core/device_impl.cpp b/csrc/core/device_impl.cpp index 04b772314e..32ed9e104f 100644 --- a/csrc/core/device_impl.cpp +++ b/csrc/core/device_impl.cpp @@ -144,7 +144,7 @@ Stream::Stream(Device device, uint64_t flags) { r.error().throw_exception(); } } else { - ERROR("{}, {}", device.device_id(), device.platform_id()); + MMDEPLOY_ERROR("{}, {}", device.device_id(), device.platform_id()); throw_exception(eInvalidArgument); } } diff --git a/csrc/core/graph.cpp b/csrc/core/graph.cpp index a824630059..524542928e 100644 --- a/csrc/core/graph.cpp +++ b/csrc/core/graph.cpp @@ -3,8 +3,10 @@ #include "core/graph.h" #include "archive/value_archive.h" +#include "core/registry.h" -namespace mmdeploy::graph { +namespace mmdeploy { +namespace graph { TaskGraph::Handle* TaskGraph::Add(TaskFunction fn) { function_.push_back(std::move(fn)); @@ -14,7 +16,8 @@ TaskGraph::Handle* TaskGraph::Add(TaskFunction fn) { TaskGraph::~TaskGraph() { for (int i = 0; i < time_.size(); ++i) { - INFO("node {} ({}): {} ms", i, handle_[i]->name(), static_cast(time_[i]) / count_); + MMDEPLOY_INFO("node {} ({}): {} ms", i, handle_[i]->name(), + static_cast(time_[i]) / count_); } } @@ -75,4 +78,8 @@ std::vector> Context::Execute(Span()>> return graph_->Execute(tasks); } -} // namespace mmdeploy::graph +} // namespace graph + +MMDEPLOY_DEFINE_REGISTRY(graph::Node); + +} // namespace mmdeploy diff --git a/csrc/core/graph.h b/csrc/core/graph.h index 37c1aa189f..d55afe70fa 100644 --- a/csrc/core/graph.h +++ b/csrc/core/graph.h @@ -14,7 +14,9 @@ #include "taskflow/taskflow.hpp" #endif -namespace mmdeploy::graph { +namespace mmdeploy { + +namespace graph { using std::pair; using std::string; @@ -24,7 +26,7 @@ using std::vector; class TaskGraph; class Node; -class Context { +class MMDEPLOY_API Context { public: explicit Context(TaskGraph* graph) : graph_(graph) {} @@ -48,7 +50,7 @@ class Context { TaskGraph* graph_; }; -class TaskGraph { +class MMDEPLOY_API TaskGraph { friend class Context; public: @@ -65,6 +67,10 @@ class TaskGraph { ~TaskGraph(); + TaskGraph() = default; + TaskGraph(const TaskGraph&) = delete; + TaskGraph& operator=(const TaskGraph&) = delete; + Handle* Add(TaskFunction fn); Result Run(Value inputs); @@ -82,7 +88,7 @@ class TaskGraph { int64_t count_{}; }; -class Node { +class MMDEPLOY_API Node { public: virtual ~Node() = default; virtual void Build(TaskGraph& graph) = 0; @@ -96,6 +102,10 @@ class Node { std::vector outputs_; }; -} // namespace mmdeploy::graph +} // namespace graph + +MMDEPLOY_DECLARE_REGISTRY(graph::Node); + +} // namespace mmdeploy #endif // MMDEPLOY_SRC_EXPERIMENTAL_PIPELINE_IR_H_ diff --git a/csrc/core/logger.cpp b/csrc/core/logger.cpp index b858ce785c..a1499cc072 100644 --- a/csrc/core/logger.cpp +++ b/csrc/core/logger.cpp @@ -9,6 +9,9 @@ #include #else #include +#if defined(_MSC_VER) +#include +#endif #endif #endif diff --git a/csrc/core/logger.h b/csrc/core/logger.h index 4f7c2ebb84..ff326c4511 100644 --- a/csrc/core/logger.h +++ b/csrc/core/logger.h @@ -5,11 +5,13 @@ #include +#include "core/macro.h" + namespace mmdeploy { -spdlog::logger *GetLogger(); +MMDEPLOY_API spdlog::logger *GetLogger(); -void SetLogger(spdlog::logger *logger); +MMDEPLOY_API void SetLogger(spdlog::logger *logger); } // namespace mmdeploy @@ -86,18 +88,4 @@ void SetLogger(spdlog::logger *logger); #define MMDEPLOY_CRITICAL(...) (void)0; #endif -#undef CRITICAL -#undef ERROR -#undef WARN -#undef INFO -#undef DEBUG -#undef TRACE - -#define CRITICAL MMDEPLOY_CRITICAL -#define ERROR MMDEPLOY_ERROR -#define WARN MMDEPLOY_WARN -#define INFO MMDEPLOY_INFO -#define DEBUG MMDEPLOY_DEBUG -#define TRACE MMDEPLOY_TRACE - #endif // !CORE_LOG_H diff --git a/csrc/core/macro.h b/csrc/core/macro.h index f9822094e9..6f52f0b5fa 100644 --- a/csrc/core/macro.h +++ b/csrc/core/macro.h @@ -3,34 +3,119 @@ #ifndef MMDEPLOY_SRC_CORE_MARCO_H_ #define MMDEPLOY_SRC_CORE_MARCO_H_ +#ifndef MMDEPLOY_EXPORT #ifdef _MSC_VER -#ifdef SDK_EXPORTS -#define MM_SDK_API __declspec(dllexport) +#define MMDEPLOY_EXPORT __declspec(dllexport) #else -#define MM_SDK_API -#endif -#else /* _MSC_VER */ -#ifdef SDK_EXPORTS -#define MM_SDK_API __attribute__((visibility("default"))) -#else -#define MM_SDK_API +#define MMDEPLOY_EXPORT __attribute__((visibility("default"))) #endif #endif -#ifdef __cplusplus -#define CV_SDK_API extern "C" MM_SDK_API +#ifndef MMDEPLOY_API +#ifdef MMDEPLOY_API_EXPORTS +#define MMDEPLOY_API MMDEPLOY_EXPORT #else -#define CV_SDK_API MM_SDK_API +#define MMDEPLOY_API +#endif #endif -#define MMDEPLOY_CONCATENATE_IMPL(s1, s2) s1##s2 -#define MMDEPLOY_CONCATENATE(s1, s2) MMDEPLOY_CONCATENATE_IMPL(s1, s2) +#define _MMDEPLOY_PP_CONCAT_IMPL(s1, s2) s1##s2 +#define MMDEPLOY_PP_CONCAT(s1, s2) _MMDEPLOY_PP_CONCAT_IMPL(s1, s2) // ! Be aware of ODR violation when using __COUNTER__ #ifdef __COUNTER__ -#define MMDEPLOY_ANONYMOUS_VARIABLE(str) MMDEPLOY_CONCATENATE(str, __COUNTER__) +#define MMDEPLOY_ANONYMOUS_VARIABLE(str) MMDEPLOY_PP_CONCAT(str, __COUNTER__) #else -#define MMDEPLOY_ANONYMOUS_VARIABLE(str) MMDEPLOY_CONCATENATE(str, __LINE__) +#define MMDEPLOY_ANONYMOUS_VARIABLE(str) MMDEPLOY_PP_CONCAT(str, __LINE__) #endif +#define MMDEPLOY_PP_NARG(...) _MMDEPLOY_PP_NARG(__VA_ARGS__, _MMDEPLOY_PP_RESQ_N()) + +#define _MMDEPLOY_PP_NARG(...) _MMDEPLOY_PP_ARG_N(__VA_ARGS__) + +#define _MMDEPLOY_PP_ARG_N(_1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, \ + _17, _18, _19, _20, _21, _22, _23, _24, _25, _26, _27, _28, _29, _30, \ + _31, _32, _33, _34, _35, _36, _37, _38, _39, _40, _41, _42, _43, _44, \ + _45, _46, _47, _48, _49, _50, _51, _52, _53, _54, _55, _56, _57, _58, \ + _59, _60, _61, _62, _63, N, ...) \ + N + +#define _MMDEPLOY_PP_RESQ_N() \ + 63, 62, 61, 60, 59, 58, 57, 56, 55, 54, 53, 52, 51, 50, 49, 48, 47, 46, 45, 44, 43, 42, 41, 40, \ + 39, 38, 37, 36, 35, 34, 33, 32, 31, 30, 29, 28, 27, 26, 25, 24, 23, 22, 21, 20, 19, 18, 17, \ + 16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0 + +#define MMDEPLOY_PP_MAP_1(f, x) f(x) +#define MMDEPLOY_PP_MAP_2(f, x, ...) f(x), MMDEPLOY_PP_MAP_1(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_3(f, x, ...) f(x), MMDEPLOY_PP_MAP_2(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_4(f, x, ...) f(x), MMDEPLOY_PP_MAP_3(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_5(f, x, ...) f(x), MMDEPLOY_PP_MAP_4(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_6(f, x, ...) f(x), MMDEPLOY_PP_MAP_5(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_7(f, x, ...) f(x), MMDEPLOY_PP_MAP_6(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_8(f, x, ...) f(x), MMDEPLOY_PP_MAP_7(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_9(f, x, ...) f(x), MMDEPLOY_PP_MAP_8(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_10(f, x, ...) f(x), MMDEPLOY_PP_MAP_9(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_11(f, x, ...) f(x), MMDEPLOY_PP_MAP_10(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_12(f, x, ...) f(x), MMDEPLOY_PP_MAP_11(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_13(f, x, ...) f(x), MMDEPLOY_PP_MAP_12(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_14(f, x, ...) f(x), MMDEPLOY_PP_MAP_13(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_15(f, x, ...) f(x), MMDEPLOY_PP_MAP_14(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_16(f, x, ...) f(x), MMDEPLOY_PP_MAP_15(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_17(f, x, ...) f(x), MMDEPLOY_PP_MAP_16(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_18(f, x, ...) f(x), MMDEPLOY_PP_MAP_17(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_19(f, x, ...) f(x), MMDEPLOY_PP_MAP_18(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_20(f, x, ...) f(x), MMDEPLOY_PP_MAP_19(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_21(f, x, ...) f(x), MMDEPLOY_PP_MAP_20(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_22(f, x, ...) f(x), MMDEPLOY_PP_MAP_21(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_23(f, x, ...) f(x), MMDEPLOY_PP_MAP_22(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_24(f, x, ...) f(x), MMDEPLOY_PP_MAP_23(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_25(f, x, ...) f(x), MMDEPLOY_PP_MAP_24(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_26(f, x, ...) f(x), MMDEPLOY_PP_MAP_25(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_27(f, x, ...) f(x), MMDEPLOY_PP_MAP_26(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_28(f, x, ...) f(x), MMDEPLOY_PP_MAP_27(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_29(f, x, ...) f(x), MMDEPLOY_PP_MAP_28(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_30(f, x, ...) f(x), MMDEPLOY_PP_MAP_29(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_31(f, x, ...) f(x), MMDEPLOY_PP_MAP_30(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_32(f, x, ...) f(x), MMDEPLOY_PP_MAP_31(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_33(f, x, ...) f(x), MMDEPLOY_PP_MAP_32(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_34(f, x, ...) f(x), MMDEPLOY_PP_MAP_33(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_35(f, x, ...) f(x), MMDEPLOY_PP_MAP_34(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_36(f, x, ...) f(x), MMDEPLOY_PP_MAP_35(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_37(f, x, ...) f(x), MMDEPLOY_PP_MAP_36(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_38(f, x, ...) f(x), MMDEPLOY_PP_MAP_37(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_39(f, x, ...) f(x), MMDEPLOY_PP_MAP_38(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_40(f, x, ...) f(x), MMDEPLOY_PP_MAP_39(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_41(f, x, ...) f(x), MMDEPLOY_PP_MAP_40(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_42(f, x, ...) f(x), MMDEPLOY_PP_MAP_41(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_43(f, x, ...) f(x), MMDEPLOY_PP_MAP_42(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_44(f, x, ...) f(x), MMDEPLOY_PP_MAP_43(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_45(f, x, ...) f(x), MMDEPLOY_PP_MAP_44(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_46(f, x, ...) f(x), MMDEPLOY_PP_MAP_45(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_47(f, x, ...) f(x), MMDEPLOY_PP_MAP_46(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_48(f, x, ...) f(x), MMDEPLOY_PP_MAP_47(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_49(f, x, ...) f(x), MMDEPLOY_PP_MAP_48(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_50(f, x, ...) f(x), MMDEPLOY_PP_MAP_49(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_51(f, x, ...) f(x), MMDEPLOY_PP_MAP_50(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_52(f, x, ...) f(x), MMDEPLOY_PP_MAP_51(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_53(f, x, ...) f(x), MMDEPLOY_PP_MAP_52(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_54(f, x, ...) f(x), MMDEPLOY_PP_MAP_53(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_55(f, x, ...) f(x), MMDEPLOY_PP_MAP_54(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_56(f, x, ...) f(x), MMDEPLOY_PP_MAP_55(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_57(f, x, ...) f(x), MMDEPLOY_PP_MAP_56(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_58(f, x, ...) f(x), MMDEPLOY_PP_MAP_57(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_59(f, x, ...) f(x), MMDEPLOY_PP_MAP_58(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_60(f, x, ...) f(x), MMDEPLOY_PP_MAP_59(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_61(f, x, ...) f(x), MMDEPLOY_PP_MAP_60(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_62(f, x, ...) f(x), MMDEPLOY_PP_MAP_61(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_63(f, x, ...) f(x), MMDEPLOY_PP_MAP_62(f, __VA_ARGS__) +#define MMDEPLOY_PP_MAP_64(f, x, ...) f(x), MMDEPLOY_PP_MAP_63(f, __VA_ARGS__) + +#define MMDEPLOY_PP_MAP(f, ...) \ + _MMDEPLOY_PP_MAP_IMPL1(f, MMDEPLOY_PP_NARG(__VA_ARGS__), __VA_ARGS__) + +#define _MMDEPLOY_PP_MAP_IMPL1(f, n, ...) \ + _MMDEPLOY_PP_MAP_IMPL2(f, MMDEPLOY_PP_CONCAT(MMDEPLOY_PP_MAP_, n), __VA_ARGS__) + +#define _MMDEPLOY_PP_MAP_IMPL2(f, M_, ...) M_(f, __VA_ARGS__) + #endif // MMDEPLOY_SRC_CORE_MARCO_H_ diff --git a/csrc/core/mat.h b/csrc/core/mat.h index d6ec4ba206..4702df7e9e 100644 --- a/csrc/core/mat.h +++ b/csrc/core/mat.h @@ -11,7 +11,7 @@ namespace mmdeploy { -class Mat final { +class MMDEPLOY_API Mat final { public: Mat() = default; diff --git a/csrc/core/model.cpp b/csrc/core/model.cpp index 080504a2a6..d4b6361a91 100644 --- a/csrc/core/model.cpp +++ b/csrc/core/model.cpp @@ -4,14 +4,7 @@ #include "core/logger.h" #include "core/model_impl.h" - -#if __GNUC__ >= 8 -#include -namespace fs = std::filesystem; -#else -#include -namespace fs = std::experimental::filesystem; -#endif +#include "core/utils/filesystem.h" using namespace std; @@ -19,7 +12,7 @@ namespace mmdeploy { Model::Model(const std::string& model_path) { if (auto r = Model::Init(model_path); !r) { - ERROR("load model failed. Its file path is '{}'", model_path); + MMDEPLOY_ERROR("load model failed. Its file path is '{}'", model_path); r.error().throw_exception(); } } @@ -28,7 +21,7 @@ Model::Model(const void* buffer, size_t size) { Init(buffer, size).value(); } Result Model::Init(const std::string& model_path) { if (!fs::exists(model_path)) { - ERROR("'{}' doesn't exist", model_path); + MMDEPLOY_ERROR("'{}' doesn't exist", model_path); return Status(eFileNotExist); } @@ -42,13 +35,13 @@ Result Model::Init(const std::string& model_path) { } OUTCOME_TRY(auto meta, impl->ReadMeta()); - INFO("{} successfully load sdk model {}", entry.name, model_path); + MMDEPLOY_INFO("{} successfully load sdk model {}", entry.name, model_path); impl_ = std::move(impl); meta_ = std::move(meta); return success(); } - ERROR("no ModelImpl can read sdk_model {}", model_path); + MMDEPLOY_ERROR("no ModelImpl can read sdk_model {}", model_path); return Status(eNotSupported); } @@ -63,13 +56,13 @@ Result Model::Init(const void* buffer, size_t size) { } OUTCOME_TRY(auto meta, impl->ReadMeta()); - INFO("{} successfully load sdk model {}", entry.name); + MMDEPLOY_INFO("{} successfully load sdk model {}", entry.name); impl_ = std::move(impl); meta_ = std::move(meta); return success(); } - ERROR("no ModelImpl can parse buffer"); + MMDEPLOY_ERROR("no ModelImpl can parse buffer"); return Status(eNotSupported); } @@ -79,7 +72,7 @@ Result Model::GetModelConfig(const std::string& name) const { return info; } } - ERROR("cannot find model '{}' in meta file", name); + MMDEPLOY_ERROR("cannot find model '{}' in meta file", name); return Status(eEntryNotFound); } @@ -87,14 +80,19 @@ Result Model::ReadFile(const std::string& file_path) noexcept { return impl_->ReadFile(file_path); } +ModelRegistry& ModelRegistry::Get() { + static ModelRegistry inst; + return inst; +} + Result ModelRegistry::Register(const std::string& name, Creator creator) { for (auto& entry : entries_) { if (entry.name == name) { - ERROR("{} is already registered", name); + MMDEPLOY_ERROR("{} is already registered", name); return Status(eFail); } } - INFO("Register '{}'", name); + MMDEPLOY_INFO("Register '{}'", name); entries_.push_back({name, std::move(creator)}); return success(); } diff --git a/csrc/core/model.h b/csrc/core/model.h index a9ce11eff3..5193128c31 100644 --- a/csrc/core/model.h +++ b/csrc/core/model.h @@ -39,7 +39,7 @@ class ModelImpl; * in case of faster-rcnn model, it splits into two models, one is rpn and the * other is cnn for roi classification. */ -class Model { +class MMDEPLOY_API Model { public: Model() = default; @@ -115,7 +115,7 @@ class Model { * }; * ANewModelImplRegister a_new_model_impl_register; */ -class ModelRegistry { +class MMDEPLOY_API ModelRegistry { public: using Creator = std::function()>; struct Entry { @@ -126,10 +126,7 @@ class ModelRegistry { /** * @brief Return global instance of `ModelRegistry` */ - static ModelRegistry& Get() { - static ModelRegistry inst; - return inst; - } + static ModelRegistry& Get(); /** * @brief Register an sdk model format denoted by an specified `ModelImpl` diff --git a/csrc/core/module.cpp b/csrc/core/module.cpp index d21ea31bdc..28857f4206 100644 --- a/csrc/core/module.cpp +++ b/csrc/core/module.cpp @@ -6,7 +6,9 @@ namespace mmdeploy { -template class Registry; -template class Creator; +// template class Registry; +// template class Creator; + +MMDEPLOY_DEFINE_REGISTRY(Module); } // namespace mmdeploy diff --git a/csrc/core/module.h b/csrc/core/module.h index 96d0c5cffa..6debc6a6dd 100644 --- a/csrc/core/module.h +++ b/csrc/core/module.h @@ -4,17 +4,20 @@ #define MMDEPLOY_SRC_CORE_MODULE_H_ #include "core/macro.h" +#include "core/registry.h" #include "core/status_code.h" #include "core/value.h" namespace mmdeploy { -class MM_SDK_API Module { +class MMDEPLOY_API Module { public: virtual ~Module() = default; virtual Result Process(const Value& args) = 0; }; +MMDEPLOY_DECLARE_REGISTRY(Module); + } // namespace mmdeploy #endif // MMDEPLOY_SRC_CORE_MODULE_H_ diff --git a/csrc/core/net.cpp b/csrc/core/net.cpp index f548042507..9f057dc88e 100644 --- a/csrc/core/net.cpp +++ b/csrc/core/net.cpp @@ -6,7 +6,9 @@ namespace mmdeploy { -template class Registry; -template class Creator; +// template class Registry; +// template class Creator; + +MMDEPLOY_DEFINE_REGISTRY(Net); } // namespace mmdeploy diff --git a/csrc/core/net.h b/csrc/core/net.h index b96551cf90..c49a7ceeb7 100644 --- a/csrc/core/net.h +++ b/csrc/core/net.h @@ -22,6 +22,8 @@ class Net { virtual Result ForwardAsync(Event* event) = 0; }; +MMDEPLOY_DECLARE_REGISTRY(Net); + } // namespace mmdeploy #endif // MMDEPLOY_SRC_CORE_NET_H_ diff --git a/csrc/core/operator.cpp b/csrc/core/operator.cpp index c40bfee508..e4e536b2c5 100644 --- a/csrc/core/operator.cpp +++ b/csrc/core/operator.cpp @@ -2,6 +2,8 @@ #include "operator.h" +#include + namespace mmdeploy::graph { Result Gather(const Value::Array& array, const vector& idxs, Value::Array& output) { diff --git a/csrc/core/operator.h b/csrc/core/operator.h index 0936fbfaf3..c71c62f051 100644 --- a/csrc/core/operator.h +++ b/csrc/core/operator.h @@ -11,13 +11,18 @@ using std::string; using std::tuple; using std::vector; -Result Gather(const Value::Array& array, const vector& idxs, Value::Array& output); -Result Gather(Value::Array&& array, const vector& idxs, Value::Array& output); -Result Gather(const Value::Object& object, const vector& keys, - Value::Array& output); -Result Gather(Value::Object&& object, const vector& keys, Value::Array& output); -Result Scatter(Value::Array array, const vector& idxs, Value::Array& output); -Result Scatter(Value::Array array, const vector& keys, Value::Object& output); +MMDEPLOY_API Result Gather(const Value::Array& array, const vector& idxs, + Value::Array& output); +MMDEPLOY_API Result Gather(Value::Array&& array, const vector& idxs, + Value::Array& output); +MMDEPLOY_API Result Gather(const Value::Object& object, const vector& keys, + Value::Array& output); +MMDEPLOY_API Result Gather(Value::Object&& object, const vector& keys, + Value::Array& output); +MMDEPLOY_API Result Scatter(Value::Array array, const vector& idxs, + Value::Array& output); +MMDEPLOY_API Result Scatter(Value::Array array, const vector& keys, + Value::Object& output); inline Result Gather(const Value::Array& array, const vector& idxs) { Value::Array output; @@ -95,13 +100,13 @@ Result Unflatten(V&& input, const vector& idxs) { } // object of arrays -> array of objects, all arrays must be of same length -Result DistribOA(const Value& oa); +MMDEPLOY_API Result DistribOA(const Value& oa); // array of objects -> object of arrays, all objects must be isomorphic -Result DistribAO(const Value& ao); +MMDEPLOY_API Result DistribAO(const Value& ao); // array of arrays -> array of arrays, this is equivalent to transpose -Result DistribAA(const Value& a); +MMDEPLOY_API Result DistribAA(const Value& a); } // namespace mmdeploy::graph diff --git a/csrc/core/registry.cpp b/csrc/core/registry.cpp new file mode 100644 index 0000000000..d0d543ee5d --- /dev/null +++ b/csrc/core/registry.cpp @@ -0,0 +1,46 @@ +// Copyright (c) OpenMMLab. All rights reserved. + +#include "core/registry.h" + +namespace mmdeploy { + +Registry::Registry() = default; + +Registry::~Registry() = default; + +bool Registry::AddCreator(Creator &creator) { + MMDEPLOY_DEBUG("Adding creator: {}", creator.GetName()); + auto key = creator.GetName(); + if (entries_.find(key) == entries_.end()) { + entries_.insert(std::make_pair(key, &creator)); + return true; + } + + for (auto iter = entries_.lower_bound(key); iter != entries_.upper_bound(key); ++iter) { + if (iter->second->GetVersion() == creator.GetVersion()) { + return false; + } + } + + entries_.insert(std::make_pair(key, &creator)); + return true; +} + +Creator *Registry::GetCreator(const std::string &type, int version) { + auto iter = entries_.find(type); + if (iter == entries_.end()) { + return nullptr; + } + if (0 == version) { + return iter->second; + } + + for (auto iter = entries_.lower_bound(type); iter != entries_.upper_bound(type); ++iter) { + if (iter->second->GetVersion() == version) { + return iter->second; + } + } + return nullptr; +} + +} // namespace mmdeploy diff --git a/csrc/core/registry.h b/csrc/core/registry.h index 03d5e3f234..bde878a35c 100644 --- a/csrc/core/registry.h +++ b/csrc/core/registry.h @@ -9,6 +9,7 @@ #include #include +#include "macro.h" #include "value.h" namespace mmdeploy { @@ -30,73 +31,58 @@ using get_return_type_t = typename get_return_type::type; } // namespace detail +template +class Creator; + +template <> +class Creator { + public: + virtual ~Creator() = default; + virtual const char *GetName() const = 0; + virtual int GetVersion() const { return 0; } +}; + template -class Creator { +class Creator : public Creator { public: using ReturnType = detail::get_return_type_t; public: - virtual ~Creator() = default; - virtual const char *GetName() const = 0; - virtual int GetVersion() const = 0; virtual ReturnType Create(const Value &args) = 0; }; -template -class Registry { +template +class Registry; + +template <> +class MMDEPLOY_API Registry { public: - static Registry &Get() { - static Registry registry; - return registry; - } + Registry(); - bool AddCreator(Creator &creator) { - auto key = creator.GetName(); - if (entries_.find(key) == entries_.end()) { - entries_.insert(std::make_pair(key, &creator)); - return true; - } - - for (auto iter = entries_.lower_bound(key); iter != entries_.upper_bound(key); ++iter) { - if (iter->second->GetVersion() == creator.GetVersion()) { - return false; - } - } - - entries_.insert(std::make_pair(key, &creator)); - return true; - } + ~Registry(); + + bool AddCreator(Creator &creator); + + Creator *GetCreator(const std::string &type, int version = 0); + + private: + std::multimap *> entries_; +}; + +template +class Registry : public Registry { + public: + bool AddCreator(Creator &creator) { return Registry::AddCreator(creator); } Creator *GetCreator(const std::string &type, int version = 0) { - auto iter = entries_.find(type); - if (iter == entries_.end()) { - return nullptr; - } - if (0 == version) { - return iter->second; - } - - for (auto iter = entries_.lower_bound(type); iter != entries_.upper_bound(type); ++iter) { - if (iter->second->GetVersion() == version) { - return iter->second; - } - } - return nullptr; + auto creator = Registry::GetCreator(type, version); + return static_cast *>(creator); } - std::vector ListCreators() { - std::vector keys; - for (const auto &[key, _] : entries_) { - keys.push_back(key); - } - return keys; - } + static Registry &Get(); private: Registry() = default; - - private: - std::multimap *> entries_; }; template @@ -110,6 +96,17 @@ class Registerer { } // namespace mmdeploy +#define MMDEPLOY_DECLARE_REGISTRY(EntryType) \ + template <> \ + Registry &Registry::Get(); + +#define MMDEPLOY_DEFINE_REGISTRY(EntryType) \ + template <> \ + MMDEPLOY_EXPORT Registry &Registry::Get() { \ + static Registry v; \ + return v; \ + } + #define REGISTER_MODULE(EntryType, CreatorType) \ static ::mmdeploy::Registerer g_register_##EntryType##_##CreatorType{}; diff --git a/csrc/core/serialization.h b/csrc/core/serialization.h index aeea43bb50..6a37d8c3ff 100644 --- a/csrc/core/serialization.h +++ b/csrc/core/serialization.h @@ -8,47 +8,14 @@ #include #include +#include "core/macro.h" #include "core/status_code.h" #include "mpl/detected.h" #include "mpl/type_traits.h" namespace mmdeploy { -#define _MMDEPLOY_NTH_ARG(_1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, \ - N, ...) \ - N - -#define _MMDEPLOY_ARCHIVE_1(x) MMDEPLOY_NVP(x) -#define _MMDEPLOY_ARCHIVE_2(x, ...) MMDEPLOY_NVP(x), _MMDEPLOY_ARCHIVE_1(__VA_ARGS__) -#define _MMDEPLOY_ARCHIVE_3(x, ...) MMDEPLOY_NVP(x), _MMDEPLOY_ARCHIVE_2(__VA_ARGS__) -#define _MMDEPLOY_ARCHIVE_4(x, ...) MMDEPLOY_NVP(x), _MMDEPLOY_ARCHIVE_3(__VA_ARGS__) -#define _MMDEPLOY_ARCHIVE_5(x, ...) MMDEPLOY_NVP(x), _MMDEPLOY_ARCHIVE_4(__VA_ARGS__) -#define _MMDEPLOY_ARCHIVE_6(x, ...) MMDEPLOY_NVP(x), _MMDEPLOY_ARCHIVE_5(__VA_ARGS__) -#define _MMDEPLOY_ARCHIVE_7(x, ...) MMDEPLOY_NVP(x), _MMDEPLOY_ARCHIVE_6(__VA_ARGS__) -#define _MMDEPLOY_ARCHIVE_8(x, ...) MMDEPLOY_NVP(x), _MMDEPLOY_ARCHIVE_7(__VA_ARGS__) -#define _MMDEPLOY_ARCHIVE_9(x, ...) MMDEPLOY_NVP(x), _MMDEPLOY_ARCHIVE_8(__VA_ARGS__) -#define _MMDEPLOY_ARCHIVE_10(x, ...) MMDEPLOY_NVP(x), _MMDEPLOY_ARCHIVE_9(__VA_ARGS__) -#define _MMDEPLOY_ARCHIVE_11(x, ...) MMDEPLOY_NVP(x), _MMDEPLOY_ARCHIVE_10(__VA_ARGS__) -#define _MMDEPLOY_ARCHIVE_12(x, ...) MMDEPLOY_NVP(x), _MMDEPLOY_ARCHIVE_11(__VA_ARGS__) -#define _MMDEPLOY_ARCHIVE_13(x, ...) MMDEPLOY_NVP(x), _MMDEPLOY_ARCHIVE_12(__VA_ARGS__) -#define _MMDEPLOY_ARCHIVE_14(x, ...) MMDEPLOY_NVP(x), _MMDEPLOY_ARCHIVE_13(__VA_ARGS__) -#define _MMDEPLOY_ARCHIVE_15(x, ...) MMDEPLOY_NVP(x), _MMDEPLOY_ARCHIVE_14(__VA_ARGS__) -#define _MMDEPLOY_ARCHIVE_16(x, ...) MMDEPLOY_NVP(x), _MMDEPLOY_ARCHIVE_15(__VA_ARGS__) - -#define _MMDEPLOY_ARCHIVE_DISPATCH(...) \ - _MMDEPLOY_NTH_ARG(__VA_ARGS__, _MMDEPLOY_ARCHIVE_16(__VA_ARGS__), \ - _MMDEPLOY_ARCHIVE_15(__VA_ARGS__), _MMDEPLOY_ARCHIVE_14(__VA_ARGS__), \ - _MMDEPLOY_ARCHIVE_13(__VA_ARGS__), _MMDEPLOY_ARCHIVE_12(__VA_ARGS__), \ - _MMDEPLOY_ARCHIVE_11(__VA_ARGS__), _MMDEPLOY_ARCHIVE_10(__VA_ARGS__), \ - _MMDEPLOY_ARCHIVE_9(__VA_ARGS__), _MMDEPLOY_ARCHIVE_8(__VA_ARGS__), \ - _MMDEPLOY_ARCHIVE_7(__VA_ARGS__), _MMDEPLOY_ARCHIVE_6(__VA_ARGS__), \ - _MMDEPLOY_ARCHIVE_5(__VA_ARGS__), _MMDEPLOY_ARCHIVE_4(__VA_ARGS__), \ - _MMDEPLOY_ARCHIVE_3(__VA_ARGS__), _MMDEPLOY_ARCHIVE_2(__VA_ARGS__), \ - _MMDEPLOY_ARCHIVE_1(__VA_ARGS__)) - -#define MMDEPLOY_ARCHIVE_NVP(archive, ...) archive(_MMDEPLOY_ARCHIVE_DISPATCH(__VA_ARGS__)) - -#define MMDEPLOY_ARCHIVE(archive, ...) archive(__VA_ARGS__) +#define MMDEPLOY_ARCHIVE_NVP(archive, ...) archive(MMDEPLOY_PP_MAP(MMDEPLOY_NVP, __VA_ARGS__)) #define MMDEPLOY_ARCHIVE_MEMBERS(...) \ template \ @@ -273,32 +240,26 @@ void load(Archive &&archive, T &&object) { } } -template -using save_t = decltype(save(std::declval(), std::declval())); - struct save_fn { template - auto operator()(Archive &&a, T &&v) const -> save_t { + auto operator()(Archive &&a, T &&v) const + -> decltype(save(std::forward(a), std::forward(v))) { return save(std::forward(a), std::forward(v)); } }; -template -using load_t = decltype(load(std::declval(), std::declval())); - struct load_fn { template - auto operator()(Archive &&a, T &&v) const -> load_t { + auto operator()(Archive &&a, T &&v) const + -> decltype(load(std::forward(a), std::forward(v))) { return load(std::forward(a), std::forward(v)); } }; -template -using serialize_t = decltype(serialize(std::declval(), std::declval())); - struct serialize_fn { template - auto operator()(Archive &&a, T &&v) const -> serialize_t { + auto operator()(Archive &&a, T &&v) const + -> decltype(serialize(std::forward(a), std::forward(v))) { return serialize(std::forward(a), std::forward(v)); } }; @@ -319,15 +280,18 @@ struct adl_serializer; template struct adl_serializer { template - static auto save(Archive &&a, T &&v) -> detail::save_t { + static auto save(Archive &&a, T &&v) + -> decltype(::mmdeploy::save(std::forward(a), std::forward(v))) { ::mmdeploy::save(std::forward(a), std::forward(v)); } template - static auto load(Archive &&a, T &&v) -> detail::load_t { + static auto load(Archive &&a, T &&v) + -> decltype(::mmdeploy::load(std::forward(a), std::forward(v))) { ::mmdeploy::load(std::forward(a), std::forward(v)); } template - static auto serialize(Archive &&a, T &&v) -> detail::serialize_t { + static auto serialize(Archive &&a, T &&v) + -> decltype(::mmdeploy::serialize(std::forward(a), std::forward(v))) { ::mmdeploy::serialize(std::forward(a), std::forward(v)); } }; diff --git a/csrc/core/status_code.h b/csrc/core/status_code.h index 3b719e677a..f84ce8a886 100644 --- a/csrc/core/status_code.h +++ b/csrc/core/status_code.h @@ -5,6 +5,7 @@ #include +#include "core/macro.h" #include "outcome-experimental.hpp" #if MMDEPLOY_STATUS_USE_SOURCE_LOCATION #include "utils/source_location.h" @@ -71,7 +72,7 @@ inline const char *to_string(ErrorCode code) { } } -struct Status { +struct MMDEPLOY_API Status { ErrorCode ec{}; Status() = default; SYSTEM_ERROR2_NAMESPACE::status_code_domain::string_ref message() const; @@ -94,7 +95,7 @@ class StatusDomain; using StatusCode = SYSTEM_ERROR2_NAMESPACE::status_code; -class StatusDomain : public SYSTEM_ERROR2_NAMESPACE::status_code_domain { +class MMDEPLOY_API StatusDomain : public SYSTEM_ERROR2_NAMESPACE::status_code_domain { using _base = status_code_domain; public: diff --git a/csrc/core/tensor.cpp b/csrc/core/tensor.cpp index aed5d6c3e0..6a040ce046 100644 --- a/csrc/core/tensor.cpp +++ b/csrc/core/tensor.cpp @@ -87,16 +87,16 @@ void Tensor::Reshape(const TensorShape& shape) { Result Tensor::CopyFrom(const Tensor& tensor, Stream stream) { if (desc_.shape.empty() || tensor.desc().shape.empty()) { - ERROR("uninitialized tensor"); + MMDEPLOY_ERROR("uninitialized tensor"); return Status(eInvalidArgument); } if (!(desc_.shape == tensor.desc().shape)) { - ERROR("mismatched shape {} vs {}", shape_string(desc_.shape), - shape_string(tensor.desc().shape)); + MMDEPLOY_ERROR("mismatched shape {} vs {}", shape_string(desc_.shape), + shape_string(tensor.desc().shape)); return Status(eShapeMismatch); } if (desc_.data_type != tensor.desc().data_type) { - ERROR("mismatched data type {} vs {}", desc_.data_type, tensor.desc().data_type); + MMDEPLOY_ERROR("mismatched data type {} vs {}", desc_.data_type, tensor.desc().data_type); return Status(eShapeMismatch); } Allocate(); @@ -112,17 +112,17 @@ Result Tensor::CopyFrom(const Tensor& tensor, Stream stream) { Result Tensor::CopyTo(Tensor& tensor, Stream stream) const { if (desc_.shape.empty() || tensor.desc().shape.empty()) { - ERROR("uninitialized tensor"); + MMDEPLOY_ERROR("uninitialized tensor"); return Status(eInvalidArgument); } if (!(desc_.shape == tensor.desc().shape)) { - ERROR("mismatched shape {} vs {}", shape_string(desc_.shape), - shape_string(tensor.desc().shape)); + MMDEPLOY_ERROR("mismatched shape {} vs {}", shape_string(desc_.shape), + shape_string(tensor.desc().shape)); return Status(eShapeMismatch); } if (desc_.data_type != tensor.desc().data_type) { - ERROR("mismatched data type {} vs {}", desc_.data_type, tensor.desc().data_type); + MMDEPLOY_ERROR("mismatched data type {} vs {}", desc_.data_type, tensor.desc().data_type); return Status(eShapeMismatch); } tensor.Allocate(); @@ -140,7 +140,7 @@ Result Tensor::CopyFrom(void* host_ptr, Stream stream) { return Status(eInvalidArgument); } if (desc_.shape.empty()) { - ERROR("uninitialized tensor"); + MMDEPLOY_ERROR("uninitialized tensor"); return Status(eInvalidArgument); } Allocate(); @@ -157,7 +157,7 @@ Result Tensor::CopyTo(void* host_ptr, Stream stream) const { return Status(eInvalidArgument); } if (desc_.shape.empty()) { - ERROR("uninitialized tensor"); + MMDEPLOY_ERROR("uninitialized tensor"); return Status(eInvalidArgument); } if (!stream) { diff --git a/csrc/core/tensor.h b/csrc/core/tensor.h index 264c6d7b84..78be82ad39 100644 --- a/csrc/core/tensor.h +++ b/csrc/core/tensor.h @@ -19,7 +19,7 @@ struct TensorDesc { std::string name; }; -class Tensor { +class MMDEPLOY_API Tensor { public: Tensor() = default; Tensor(const Tensor&) = default; diff --git a/csrc/core/utils/device_utils.h b/csrc/core/utils/device_utils.h index 81621efecc..65422664e8 100644 --- a/csrc/core/utils/device_utils.h +++ b/csrc/core/utils/device_utils.h @@ -14,7 +14,8 @@ namespace mmdeploy { * @param stream * @return */ -Result MakeAvailableOnDevice(const Mat& src, const Device& device, Stream& stream); +MMDEPLOY_API Result MakeAvailableOnDevice(const Mat& src, const Device& device, + Stream& stream); /** * @@ -23,7 +24,8 @@ Result MakeAvailableOnDevice(const Mat& src, const Device& device, Stream& * @param stream * @return */ -Result MakeAvailableOnDevice(const Tensor& src, const Device& device, Stream& stream); +MMDEPLOY_API Result MakeAvailableOnDevice(const Tensor& src, const Device& device, + Stream& stream); } // namespace mmdeploy #endif // MMDEPLOY_TRANSFORM_UTILS_H diff --git a/csrc/core/utils/filesystem.h b/csrc/core/utils/filesystem.h new file mode 100644 index 0000000000..7aca6a8d8e --- /dev/null +++ b/csrc/core/utils/filesystem.h @@ -0,0 +1,15 @@ +// Copyright (c) OpenMMLab. All rights reserved. + +#ifndef MMDEPLOY_CSRC_CORE_UTILS_FILESYSTEM_H_ +#define MMDEPLOY_CSRC_CORE_UTILS_FILESYSTEM_H_ + +// TODO: what about clang? +#if __GNUC__ >= 8 || _MSC_VER +#include +namespace fs = std::filesystem; +#else +#include +namespace fs = std::experimental::filesystem; +#endif + +#endif // MMDEPLOY_CSRC_CORE_UTILS_FILESYSTEM_H_ diff --git a/csrc/core/utils/formatter.h b/csrc/core/utils/formatter.h index 14075b3866..af28f8c9c0 100644 --- a/csrc/core/utils/formatter.h +++ b/csrc/core/utils/formatter.h @@ -13,7 +13,7 @@ namespace mmdeploy { class Value; -std::string format_value(const Value& value); +MMDEPLOY_API std::string format_value(const Value& value); } // namespace mmdeploy diff --git a/csrc/core/utils/source_location.h b/csrc/core/utils/source_location.h index b7362cc12b..f0d579b76b 100644 --- a/csrc/core/utils/source_location.h +++ b/csrc/core/utils/source_location.h @@ -3,7 +3,7 @@ #ifndef MMDEPLOY_SRC_UTILS_SOURCE_LOCATION_H_ #define MMDEPLOY_SRC_UTILS_SOURCE_LOCATION_H_ -#if __has_include() +#if __has_include() && !_MSC_VER #include namespace mmdeploy { using SourceLocation = std::source_location; diff --git a/csrc/core/utils/stacktrace.h b/csrc/core/utils/stacktrace.h index 1c1cacbb7d..53b1a44b99 100644 --- a/csrc/core/utils/stacktrace.h +++ b/csrc/core/utils/stacktrace.h @@ -4,6 +4,7 @@ #define MMDEPLOY_SRC_CORE_STACKTRACE_H_ #include +#include namespace mmdeploy { diff --git a/csrc/core/value.h b/csrc/core/value.h index b73fba4302..3241330565 100644 --- a/csrc/core/value.h +++ b/csrc/core/value.h @@ -650,22 +650,22 @@ class Value { template bool contains(Key&& key) const { - return _unwrap().template _contains(std::forward(key)); + return _unwrap()._contains(std::forward(key)); } template iterator find(Key&& key) { - return _unwrap().template _find(std::forward(key)); + return _unwrap()._find(std::forward(key)); } template const_iterator find(Key&& key) const { - return _unwrap().template _find(std::forward(key)); + return _unwrap()._find(std::forward(key)); } template T value(const typename Object::key_type& key, const T& default_value) const { - return _unwrap().template _value(key, default_value); + return _unwrap()._value(key, default_value); } iterator begin() { return _unwrap()._begin(); } diff --git a/csrc/device/cpu/CMakeLists.txt b/csrc/device/cpu/CMakeLists.txt index 226d0894bf..f7e7c46117 100644 --- a/csrc/device/cpu/CMakeLists.txt +++ b/csrc/device/cpu/CMakeLists.txt @@ -2,10 +2,14 @@ cmake_minimum_required(VERSION 3.14) project(mmdeploy_cpu_device) -include(${CMAKE_SOURCE_DIR}/cmake/common.cmake) +include(${CMAKE_SOURCE_DIR}/cmake/MMDeploy.cmake) file(GLOB_RECURSE SRCS "*.cpp") -build_target(${PROJECT_NAME} "${SRCS}") -target_link_libraries(${PROJECT_NAME} PUBLIC pthread PRIVATE mmdeploy::core) + +mmdeploy_add_module(${PROJECT_NAME} "${SRCS}") + +set(THREADS_PREFER_PTHREAD_FLAG ON) +find_package(Threads REQUIRED) +target_link_libraries(${PROJECT_NAME} PRIVATE Threads::Threads) + add_library(mmdeploy::device::cpu ALIAS ${PROJECT_NAME}) -export_module(${PROJECT_NAME}) diff --git a/csrc/device/cuda/CMakeLists.txt b/csrc/device/cuda/CMakeLists.txt index 6a36e513bd..1ac67bb8b8 100644 --- a/csrc/device/cuda/CMakeLists.txt +++ b/csrc/device/cuda/CMakeLists.txt @@ -9,17 +9,13 @@ if (${CMAKE_VERSION} VERSION_GREATER_EQUAL "3.18.0") cmake_policy(SET CMP0104 OLD) endif () -include(${CMAKE_SOURCE_DIR}/cmake/common.cmake) -set_targets(${PROJECT_NAME} CUDA_DEVICE_OBJ CUDA_DEVICE_STATIC CUDA_DEVICE_SHARED) +include(${CMAKE_SOURCE_DIR}/cmake/MMDeploy.cmake) set(SRCS cuda_device.cpp cuda_builtin_kernels.cu) -build_target(${PROJECT_NAME} "${SRCS}") +mmdeploy_add_module(${PROJECT_NAME} "${SRCS}") target_include_directories(${PROJECT_NAME} PUBLIC ${CUDA_INCLUDE_DIRS}) target_link_directories(${PROJECT_NAME} PUBLIC ${CUDA_TOOLKIT_ROOT_DIR}/lib64) -target_link_libraries(${PROJECT_NAME} - PRIVATE mmdeploy::core - PUBLIC cudart cuda) +target_link_libraries(${PROJECT_NAME} PRIVATE cudart cuda) add_library(mmdeploy::device::cuda ALIAS ${PROJECT_NAME}) -export_module(${PROJECT_NAME}) diff --git a/csrc/device/cuda/buddy_allocator.h b/csrc/device/cuda/buddy_allocator.h index f94ec1c046..3e26a0edff 100644 --- a/csrc/device/cuda/buddy_allocator.h +++ b/csrc/device/cuda/buddy_allocator.h @@ -25,7 +25,7 @@ class BuddyAllocator { block_count_ = size / block_size_; if (!IsPowerOfTwo(block_count_)) { block_count_ = RoundToPowerOfTwo(block_count_); - WARN("Rounding up block_count to next power of 2 {}", block_count_); + MMDEPLOY_WARN("Rounding up block_count to next power of 2 {}", block_count_); } base_ = LogPowerOfTwo(block_count_); size_ = block_size_ * block_count_; @@ -34,17 +34,18 @@ class BuddyAllocator { free_.resize(base_ + 1); Build(1, 0); Add(1, 0); - ERROR("size = {}, block_size = {}, block_count = {}", size_, block_size_, block_count_); + MMDEPLOY_ERROR("size = {}, block_size = {}, block_count = {}", size_, block_size_, + block_count_); size = size_; for (int i = 0; i <= base_; ++i) { - ERROR("level {}, size = {}", i, size); + MMDEPLOY_ERROR("level {}, size = {}", i, size); size /= 2; } } ~BuddyAllocator() { for (int i = 0; i < free_.size(); ++i) { - ERROR("free_[{}].size(): {}", i, free_[i].size()); + MMDEPLOY_ERROR("free_[{}].size(): {}", i, free_[i].size()); } gDefaultAllocator().Deallocate(memory_, size_); } @@ -62,7 +63,7 @@ class BuddyAllocator { } } if (level < 0) { - WARN("failed to allocate memory size = {} bytes", n); + MMDEPLOY_WARN("failed to allocate memory size = {} bytes", n); return nullptr; } for (; level < n_level; ++level) { @@ -80,7 +81,7 @@ class BuddyAllocator { std::lock_guard lock{mutex_}; auto offset = static_cast(p) - static_cast(memory_); if (offset < 0 || offset % block_size_) { - ERROR("invalid address: {}", p); + MMDEPLOY_ERROR("invalid address: {}", p); } offset /= static_cast(block_size_); auto level = GetLevel(n); diff --git a/csrc/device/cuda/cuda_builtin_kernels.cu b/csrc/device/cuda/cuda_builtin_kernels.cu index c2cf5460cf..463da81369 100644 --- a/csrc/device/cuda/cuda_builtin_kernels.cu +++ b/csrc/device/cuda/cuda_builtin_kernels.cu @@ -1,5 +1,7 @@ // Copyright (c) OpenMMLab. All rights reserved. +#include + namespace mmdeploy { namespace cuda { @@ -17,8 +19,8 @@ __global__ void FillKernel(void* dst, size_t dst_size, const void* pattern, size int Fill(void* dst, size_t dst_size, const void* pattern, size_t pattern_size, cudaStream_t stream) { - const uint n_threads = 256; - const uint n_blocks = (dst_size + n_threads - 1) / n_threads; + const unsigned int n_threads = 256; + const unsigned int n_blocks = (dst_size + n_threads - 1) / n_threads; FillKernel<<>>(dst, dst_size, pattern, pattern_size); diff --git a/csrc/device/cuda/cuda_device.cpp b/csrc/device/cuda/cuda_device.cpp index 768a267d90..9825182e9e 100644 --- a/csrc/device/cuda/cuda_device.cpp +++ b/csrc/device/cuda/cuda_device.cpp @@ -79,9 +79,9 @@ Allocator CreateDefaultAllocator() { using namespace device_allocator; AllocatorImplPtr allocator = std::make_shared(); allocator = std::make_shared(allocator, "cudaMalloc"); - allocator = std::make_shared(allocator, -1, .0); + allocator = std::make_shared(allocator, -1, .5); allocator = std::make_shared(allocator, "Tree"); - INFO("Default CUDA allocator initialized"); + MMDEPLOY_INFO("Default CUDA allocator initialized"); return Access::create(allocator); } @@ -265,7 +265,7 @@ void CudaPlatformImpl::PerDeviceData::init() { CudaPlatformImpl::CudaPlatformImpl() { int count{}; if (auto err = cudaGetDeviceCount(&count); err != cudaSuccess) { - ERROR("error getting device count: {}", cudaGetErrorString(err)); + MMDEPLOY_ERROR("error getting device count: {}", cudaGetErrorString(err)); throw_exception(eFail); } per_device_data_storage_.reserve(count); diff --git a/csrc/device/cuda/cuda_device.h b/csrc/device/cuda/cuda_device.h index 71623d42b6..d4588d2fd9 100644 --- a/csrc/device/cuda/cuda_device.h +++ b/csrc/device/cuda/cuda_device.h @@ -16,6 +16,16 @@ class CudaPlatformImpl : public PlatformImpl { public: CudaPlatformImpl(); + ~CudaPlatformImpl() override { + // The CUDA driver may have already shutdown before the platform dtor is called. + // As a workaround, simply leak per device resources and let the driver handle it + // FIXME: maybe a pair of global mmdeploy_init/deinit function would be a + // better solution + for (auto& data : per_device_data_storage_) { + data.release(); + } + } + const char* GetPlatformName() const noexcept override { return "cuda"; } shared_ptr CreateBuffer(Device device) override; diff --git a/csrc/device/cuda/default_allocator.h b/csrc/device/cuda/default_allocator.h index ca4d794e9d..a8b2177ccc 100644 --- a/csrc/device/cuda/default_allocator.h +++ b/csrc/device/cuda/default_allocator.h @@ -16,11 +16,11 @@ class DefaultAllocator { public: DefaultAllocator() = default; ~DefaultAllocator() { - ERROR("=== CUDA Default Allocator ==="); - ERROR(" Allocation: count={}, size={}MB, time={}ms", alloc_count_, - alloc_size_ / (1024 * 1024.f), alloc_time_ / 1000000.f); - ERROR("Deallocation: count={}, size={}MB, time={}ms", dealloc_count_, - dealloc_size_ / (1024 * 1024.f), dealloc_time_ / 1000000.f); + MMDEPLOY_ERROR("=== CUDA Default Allocator ==="); + MMDEPLOY_ERROR(" Allocation: count={}, size={}MB, time={}ms", alloc_count_, + alloc_size_ / (1024 * 1024.f), alloc_time_ / 1000000.f); + MMDEPLOY_ERROR("Deallocation: count={}, size={}MB, time={}ms", dealloc_count_, + dealloc_size_ / (1024 * 1024.f), dealloc_time_ / 1000000.f); } [[nodiscard]] void* Allocate(std::size_t n) { void* p{}; @@ -29,7 +29,7 @@ class DefaultAllocator { auto t1 = std::chrono::high_resolution_clock::now(); alloc_time_ += (int64_t)std::chrono::duration(t1 - t0).count(); if (ret != cudaSuccess) { - ERROR("error allocating cuda memory: {}", cudaGetErrorString(ret)); + MMDEPLOY_ERROR("error allocating cuda memory: {}", cudaGetErrorString(ret)); return nullptr; } alloc_count_ += 1; @@ -43,7 +43,7 @@ class DefaultAllocator { auto t1 = std::chrono::high_resolution_clock::now(); dealloc_time_ += (int64_t)std::chrono::duration(t1 - t0).count(); if (ret != cudaSuccess) { - ERROR("error deallocating cuda memory: {}", cudaGetErrorString(ret)); + MMDEPLOY_ERROR("error deallocating cuda memory: {}", cudaGetErrorString(ret)); return; } dealloc_count_ += 1; diff --git a/csrc/device/cuda/linear_allocator.h b/csrc/device/cuda/linear_allocator.h index 15be01bc36..59133e9332 100644 --- a/csrc/device/cuda/linear_allocator.h +++ b/csrc/device/cuda/linear_allocator.h @@ -25,11 +25,11 @@ class LinearAllocator { std::size_t space = base_ + size_ - ptr_; if (std::align(16, n, ptr, space)) { - ERROR("success n={}, total={}, count={}", n, total_, count_); + MMDEPLOY_ERROR("success n={}, total={}, count={}", n, total_, count_); ptr_ = static_cast(ptr) + n; return ptr; } - ERROR("fallback {}, total={}, count={}", n, total_, count_); + MMDEPLOY_ERROR("fallback {}, total={}, count={}", n, total_, count_); return gDefaultAllocator().Allocate(n); } void Deallocate(void* _p, std::size_t n) { @@ -43,7 +43,7 @@ class LinearAllocator { } total_ -= n; --count_; - ERROR("deallocate total={}, count={}", total_, count_); + MMDEPLOY_ERROR("deallocate total={}, count={}", total_, count_); if (total_ == 0) { assert(count_ == 0); ptr_ = base_; diff --git a/csrc/device/device_allocator.h b/csrc/device/device_allocator.h index 4539e12d4e..06bb5730df 100644 --- a/csrc/device/device_allocator.h +++ b/csrc/device/device_allocator.h @@ -162,12 +162,14 @@ class Stats : public AllocatorImpl { : allocator_(std::move(allocator)), name_(std::move(name)) {} ~Stats() override { - INFO("=== {} ===", name_); - INFO(" Allocation: count={}, size={}MB, time={}ms", data_.allocation_count, - data_.allocated_bytes / (1024 * 1024.f), static_cast(data_.allocation_time)); - INFO("Deallocation: count={}, size={}MB, time={}ms", data_.deallocation_count, - data_.deallocated_bytes / (1024 * 1024.f), static_cast(data_.deallocation_time)); - INFO("Peak memory usage: size={}MB", data_.peak / (1024 * 1024.f)); + MMDEPLOY_INFO("=== {} ===", name_); + MMDEPLOY_INFO(" Allocation: count={}, size={}MB, time={}ms", data_.allocation_count, + data_.allocated_bytes / (1024 * 1024.f), + static_cast(data_.allocation_time)); + MMDEPLOY_INFO("Deallocation: count={}, size={}MB, time={}ms", data_.deallocation_count, + data_.deallocated_bytes / (1024 * 1024.f), + static_cast(data_.deallocation_time)); + MMDEPLOY_INFO("Peak memory usage: size={}MB", data_.peak / (1024 * 1024.f)); } Block Allocate(size_t size) noexcept override { @@ -281,10 +283,10 @@ class Bucketizer : public AllocatorImpl { Bucketizer(const AllocatorCreator& creator, size_t min_size, size_t max_size, size_t step_size) : min_size_(min_size), max_size_(max_size), step_size_(step_size) { for (auto base = min_size_; base < max_size_; base += step_size_) { - // ERROR("{}, {}", base, base + step_size - 1); + // MMDEPLOY_ERROR("{}, {}", base, base + step_size - 1); allocator_.push_back(creator(base, base + step_size - 1)); } - // ERROR("{}", allocator_.size()); + // MMDEPLOY_ERROR("{}", allocator_.size()); } Block Allocate(size_t size) noexcept override { diff --git a/csrc/experimental/collection.h b/csrc/experimental/collection.h index e26be6da99..a65de2ec65 100644 --- a/csrc/experimental/collection.h +++ b/csrc/experimental/collection.h @@ -1,93 +1,93 @@ -// Copyright (c) OpenMMLab. All rights reserved. - -#ifndef MMDEPLOY_SRC_EXPERIMENTAL_COLLECTION_H_ -#define MMDEPLOY_SRC_EXPERIMENTAL_COLLECTION_H_ - -#include "token.h" - -namespace mmdeploy { - -class Collection { - public: - template - friend Collection& operator<<(Collection& c, const Token& value) { - c.put(value); - return c; - } - - template - friend const Collection& operator>>(const Collection& c, Token& value) { - c.get(value); - return c; - } - - template - Result maybe() const { - T token; - if (get(token)) { - return token; - } - return Status(eFail); - } - - private: - std::vector keys_; - std::vector> values_; - - template - void put(const Token& value) { - keys_.push_back(Token::key()); - values_.push_back(std::make_shared>(value)); - } - - template - bool get(Token& value) const { - for (int i = 0; i < keys_.size(); ++i) { - if (keys_[i] == Token::key()) { - value = *static_cast*>(values_[i].get()); - return true; - } - } - return false; - } -}; - -namespace detail { - -template -struct function_traits { - template - static std::tuple get_args(std::function); - - template - static R get_ret(std::function); - - using args_t = decltype(get_args(std::function{std::declval()})); - using ret_t = decltype(get_ret(std::function{std::declval()})); -}; - -// TODO: obtain first error -// TODO: combine all errors -template > -Result Apply(F&& f, const Result&... args) { - if ((... && args)) { - return std::invoke(std::forward(f), args.value()...); - } - return Status(eFail); -} - -template > -Result ApplyImpl(F&& f, const Collection& c, std::tuple*) { - return Apply(std::forward(f), c.maybe()...); -} - -} // namespace detail - -template ::args_t> -decltype(auto) Apply(F&& f, const Collection& c) { - return detail::ApplyImpl(std::forward(f), c, std::add_pointer_t{}); -} - -} // namespace mmdeploy - -#endif // MMDEPLOY_SRC_EXPERIMENTAL_COLLECTION_H_ +//// Copyright (c) OpenMMLab. All rights reserved. +// +//#ifndef MMDEPLOY_SRC_EXPERIMENTAL_COLLECTION_H_ +//#define MMDEPLOY_SRC_EXPERIMENTAL_COLLECTION_H_ +// +//#include "token.h" +// +// namespace mmdeploy { +// +// class Collection { +// public: +// template +// friend Collection& operator<<(Collection& c, const Token& value) { +// c.put(value); +// return c; +// } +// +// template +// friend const Collection& operator>>(const Collection& c, Token& value) { +// c.get(value); +// return c; +// } +// +// template +// Result maybe() const { +// T token; +// if (get(token)) { +// return token; +// } +// return Status(eFail); +// } +// +// private: +// std::vector keys_; +// std::vector> values_; +// +// template +// void put(const Token& value) { +// keys_.push_back(Token::key()); +// values_.push_back(std::make_shared>(value)); +// } +// +// template +// bool get(Token& value) const { +// for (int i = 0; i < keys_.size(); ++i) { +// if (keys_[i] == Token::key()) { +// value = *static_cast*>(values_[i].get()); +// return true; +// } +// } +// return false; +// } +//}; +// +// namespace detail { +// +// template +// struct function_traits { +// template +// static std::tuple get_args(std::function); +// +// template +// static R get_ret(std::function); +// +// using args_t = decltype(get_args(std::function{std::declval()})); +// using ret_t = decltype(get_ret(std::function{std::declval()})); +//}; +// +//// TODO: obtain first error +//// TODO: combine all errors +// template > +// Result Apply(F&& f, const Result&... args) { +// if ((... && args)) { +// return std::invoke(std::forward(f), args.value()...); +// } +// return Status(eFail); +// } +// +// template > +// Result ApplyImpl(F&& f, const Collection& c, std::tuple*) { +// return Apply(std::forward(f), c.maybe()...); +// } +// +// } // namespace detail +// +// template ::args_t> +// decltype(auto) Apply(F&& f, const Collection& c) { +// return detail::ApplyImpl(std::forward(f), c, std::add_pointer_t{}); +// } +// +// } // namespace mmdeploy +// +//#endif // MMDEPLOY_SRC_EXPERIMENTAL_COLLECTION_H_ diff --git a/csrc/experimental/module_adapter.h b/csrc/experimental/module_adapter.h index 8d652cfb4f..581e7f2bea 100644 --- a/csrc/experimental/module_adapter.h +++ b/csrc/experimental/module_adapter.h @@ -31,7 +31,7 @@ struct InvokeImpl { std::forward(ts)...); return make_ret_val(std::move(ret)); } catch (const std::exception& e) { - ERROR("unhandled exception: {}", e.what()); + MMDEPLOY_ERROR("unhandled exception: {}", e.what()); return Status(eFail); } catch (...) { return Status(eFail); diff --git a/csrc/experimental/token.h b/csrc/experimental/token.h index e1c951352e..6d6ae7f884 100644 --- a/csrc/experimental/token.h +++ b/csrc/experimental/token.h @@ -1,72 +1,72 @@ -// Copyright (c) OpenMMLab. All rights reserved. - -#ifndef MMDEPLOY_SRC_TOKEN_TOKEN_H_ -#define MMDEPLOY_SRC_TOKEN_TOKEN_H_ - -#include -#include -#include -#include -#include -#include - -#include "core/status_code.h" - -namespace mmdeploy { - -namespace token { - -template -using String = std::integer_sequence; - -// this is a GCC only extension -template -constexpr String operator""_ts() { - return {}; -} - -template -const char* c_str(String) { - static constexpr const char str[sizeof...(cs) + 1] = {cs..., '\0'}; - return str; -} - -} // namespace token - -// template -// static void* signature() { -// static char id = 0; -// return &id; +//// Copyright (c) OpenMMLab. All rights reserved. +// +//#ifndef MMDEPLOY_SRC_TOKEN_TOKEN_H_ +//#define MMDEPLOY_SRC_TOKEN_TOKEN_H_ +// +//#include +//#include +//#include +//#include +//#include +//#include +// +//#include "core/status_code.h" +// +// namespace mmdeploy { +// +// namespace token { +// +// template +// using String = std::integer_sequence; +// +//// this is a GCC only extension +// template +// constexpr String operator""_ts() { +// return {}; // } // -// using signature_t = decltype(signature()); - -template -struct Token { - using signature_t = void*; - using value_type = T; - - Token(T value = {}) : value_(value) {} // NOLINT - - operator T() const { return value_; } // NOLINT - static const char* key() { return token::c_str(Key{}); } - - T& operator*() { return value_; } - T* operator->() { return &value_; } - - private: - T value_; -}; - -template -class Identifier { - public: - constexpr explicit Identifier(const char* key) : key_(key) {} - const char* key_; -}; - -constexpr inline Identifier batch_size{"batch_size"}; - -} // namespace mmdeploy - -#endif // MMDEPLOY_SRC_TOKEN_TOKEN_H_ +// template +// const char* c_str(String) { +// static constexpr const char str[sizeof...(cs) + 1] = {cs..., '\0'}; +// return str; +// } +// +// } // namespace token +// +//// template +//// static void* signature() { +//// static char id = 0; +//// return &id; +//// } +//// +//// using signature_t = decltype(signature()); +// +// template +// struct Token { +// using signature_t = void*; +// using value_type = T; +// +// Token(T value = {}) : value_(value) {} // NOLINT +// +// operator T() const { return value_; } // NOLINT +// static const char* key() { return token::c_str(Key{}); } +// +// T& operator*() { return value_; } +// T* operator->() { return &value_; } +// +// private: +// T value_; +//}; +// +// template +// class Identifier { +// public: +// constexpr explicit Identifier(const char* key) : key_(key) {} +// const char* key_; +//}; +// +// constexpr inline Identifier batch_size{"batch_size"}; +// +//} // namespace mmdeploy +// +//#endif // MMDEPLOY_SRC_TOKEN_TOKEN_H_ diff --git a/csrc/graph/CMakeLists.txt b/csrc/graph/CMakeLists.txt index a0c09946e1..e39fbcf5c2 100644 --- a/csrc/graph/CMakeLists.txt +++ b/csrc/graph/CMakeLists.txt @@ -2,7 +2,7 @@ cmake_minimum_required(VERSION 3.14) project(mmdeploy_graph) -include(${CMAKE_SOURCE_DIR}/cmake/common.cmake) +include(${CMAKE_SOURCE_DIR}/cmake/MMDeploy.cmake) set(SRCS common.cpp inference.cpp @@ -10,7 +10,5 @@ set(SRCS task.cpp flatten.cpp unflatten.cpp) -build_target(${PROJECT_NAME} "${SRCS}") -target_link_libraries(${PROJECT_NAME} PRIVATE mmdeploy::core) +mmdeploy_add_module(${PROJECT_NAME} "${SRCS}") add_library(mmdeploy::graph ALIAS ${PROJECT_NAME}) -export_module(${PROJECT_NAME}) diff --git a/csrc/graph/common.cpp b/csrc/graph/common.cpp index 7881b91f81..fba4b23619 100644 --- a/csrc/graph/common.cpp +++ b/csrc/graph/common.cpp @@ -10,7 +10,7 @@ mmdeploy::graph::BaseNode::BaseNode(const mmdeploy::Value& cfg) { from_value(cfg["output"], outputs_); name_ = cfg.value("name", ""); } catch (...) { - ERROR("error parsing config: {}", cfg); + MMDEPLOY_ERROR("error parsing config: {}", cfg); throw; } } diff --git a/csrc/graph/common.h b/csrc/graph/common.h index 95e8c2d08c..a94d8c5fff 100644 --- a/csrc/graph/common.h +++ b/csrc/graph/common.h @@ -12,15 +12,16 @@ namespace mmdeploy::graph { template ::ReturnType> inline Result CreateFromRegistry(const Value& config, const char* key = "type") { - INFO("config: {}", config); + MMDEPLOY_INFO("config: {}", config); auto type = config[key].get(); auto creator = Registry::Get().GetCreator(type); if (!creator) { + MMDEPLOY_ERROR("failed to find module creator: {}", type); return Status(eEntryNotFound); } auto inst = creator->Create(config); if (!inst) { - ERROR("failed to create module: {}", type); + MMDEPLOY_ERROR("failed to create module: {}", type); return Status(eFail); } return std::move(inst); diff --git a/csrc/graph/flatten.cpp b/csrc/graph/flatten.cpp index 3153e586b5..d180470eb6 100644 --- a/csrc/graph/flatten.cpp +++ b/csrc/graph/flatten.cpp @@ -20,7 +20,7 @@ void FlattenNode::Build(TaskGraph& graph) { if (idxs.empty()) { idxs = std::move(idx); } else if (idx != idxs) { - ERROR("args does not have same structure"); + MMDEPLOY_ERROR("args does not have same structure"); return Status(eInvalidArgument); } rets.push_back(std::move(ret)); diff --git a/csrc/graph/inference.cpp b/csrc/graph/inference.cpp index 34d9c8a374..15c11d581a 100644 --- a/csrc/graph/inference.cpp +++ b/csrc/graph/inference.cpp @@ -17,7 +17,7 @@ Inference::Inference(const Value& cfg) : BaseNode(cfg) { auto model_path = model_value.get(); model_ = Model(model_path); } else { - ERROR("unsupported model specification"); + MMDEPLOY_ERROR("unsupported model specification"); throw_exception(eInvalidArgument); } @@ -31,7 +31,7 @@ Inference::Inference(const Value& cfg) : BaseNode(cfg) { value["context"] = context; pipeline_ = std::make_unique(value); if (!pipeline_) { - ERROR("failed to create pipeline"); + MMDEPLOY_ERROR("failed to create pipeline"); throw_exception(eFail); } } diff --git a/csrc/graph/pipeline.cpp b/csrc/graph/pipeline.cpp index ef77a242be..4d668c249a 100644 --- a/csrc/graph/pipeline.cpp +++ b/csrc/graph/pipeline.cpp @@ -21,7 +21,7 @@ Pipeline::Pipeline(const Value& cfg) : BaseNode(cfg["pipeline"]) { node_input_idx_.push_back(UpdateBindings(nodes_.back()->inputs(), kRead)); node_output_idx_.push_back(UpdateBindings(nodes_.back()->outputs(), kWrite)); } else { - ERROR("could not create {}:{}", name, type); + MMDEPLOY_ERROR("could not create {}:{}", name, type); throw_exception(eFail); } } @@ -57,7 +57,7 @@ std::vector Pipeline::UpdateBindings(const vector& names, Bind auto it = binding_name_to_idx_.lower_bound(name); if (it == binding_name_to_idx_.end() || it->first != name) { if (type == kRead) { - ERROR("unknown binding name: {}", name); + MMDEPLOY_ERROR("unknown binding name: {}", name); throw_exception(eEntryNotFound); } else { auto index = static_cast(binding_name_to_idx_.size()); diff --git a/csrc/graph/task.cpp b/csrc/graph/task.cpp index 0791a84a04..ee1f0ebf84 100644 --- a/csrc/graph/task.cpp +++ b/csrc/graph/task.cpp @@ -36,7 +36,7 @@ static size_t GetBatchSize(const Value& args) { Task::Task(const Value& cfg) : BaseNode(cfg) { auto module = CreateFromRegistry(cfg, "module"); if (!module) { - ERROR("failed to create task: {}", cfg); + MMDEPLOY_ERROR("failed to create task: {}", cfg); throw_exception(eFail); } module_ = std::move(module).value(); @@ -50,7 +50,8 @@ void Task::Build(TaskGraph& graph) { auto args = ctx.pop().array(); auto rets = Value::Array{}; auto batch_size = GetBatchSize(args); - // ERROR("name: {}, is_batched: {}, INPUT batch_size: {}", name_, is_batched_, batch_size); + // MMDEPLOY_ERROR("name: {}, is_batched: {}, INPUT batch_size: {}", name_, is_batched_, + // batch_size); if (!is_batched_ && batch_size) { rets.resize(outputs_.size(), Value::kArray); if (!is_thread_safe_) { @@ -86,7 +87,7 @@ void Task::Build(TaskGraph& graph) { rets = std::move(tmp).array(); } ctx.push(std::move(rets)); - // ERROR("name: {}, is_batched: {}, OUTPUT batch_size: {}", name_, is_batched_, + // MMDEPLOY_ERROR("name: {}, is_batched: {}, OUTPUT batch_size: {}", name_, is_batched_, // GetBatchSize(rets)); return success(); }); diff --git a/csrc/model/CMakeLists.txt b/csrc/model/CMakeLists.txt index fbeaaf6050..ebfbf2167e 100644 --- a/csrc/model/CMakeLists.txt +++ b/csrc/model/CMakeLists.txt @@ -2,7 +2,7 @@ cmake_minimum_required(VERSION 3.14) project(model) -include(${CMAKE_SOURCE_DIR}/cmake/common.cmake) +include(${CMAKE_SOURCE_DIR}/cmake/MMDeploy.cmake) set(MODEL_NAMES "directory_model") if (${MMDEPLOY_ZIP_MODEL}) @@ -11,18 +11,17 @@ endif () foreach (MODEL_NAME ${MODEL_NAMES}) set(TARGET_MODEL_NAME mmdeploy_${MODEL_NAME}) - build_target(${TARGET_MODEL_NAME} ${MODEL_NAME}_impl.cpp) - target_link_libraries(${TARGET_MODEL_NAME} - PRIVATE mmdeploy::core - PUBLIC stdc++fs) + mmdeploy_add_module(${TARGET_MODEL_NAME} ${MODEL_NAME}_impl.cpp) + if (NOT MSVC) + target_link_libraries(${TARGET_MODEL_NAME} PUBLIC stdc++fs) + endif () if (${MODEL_NAME} STREQUAL "zip_model") find_package(libzip QUIET) if (libzip_FOUND) target_link_libraries(${TARGET_MODEL_NAME} PUBLIC libzip::zip) - else() + else () target_link_libraries(${TARGET_MODEL_NAME} PUBLIC zip) - endif() + endif () endif () add_library(mmdeploy::${MODEL_NAME} ALIAS ${TARGET_MODEL_NAME}) - export_module(${TARGET_MODEL_NAME}) endforeach () diff --git a/csrc/model/directory_model_impl.cpp b/csrc/model/directory_model_impl.cpp index 2de9d5ec23..202eafb12f 100644 --- a/csrc/model/directory_model_impl.cpp +++ b/csrc/model/directory_model_impl.cpp @@ -5,17 +5,10 @@ #include "archive/json_archive.h" #include "core/model.h" #include "core/model_impl.h" +#include "core/utils/filesystem.h" using nlohmann::json; -#if __GNUC__ >= 8 -#include -namespace fs = std::filesystem; -#else -#include -namespace fs = std::experimental::filesystem; -#endif - namespace mmdeploy { class DirectoryModelImpl : public ModelImpl { @@ -52,7 +45,7 @@ class DirectoryModelImpl : public ModelImpl { from_json(json::parse(deploy_json), meta); return meta; } catch (std::exception& e) { - ERROR("exception happened: {}", e.what()); + MMDEPLOY_ERROR("exception happened: {}", e.what()); return Status(eFail); } } diff --git a/csrc/model/zip_model_impl.cpp b/csrc/model/zip_model_impl.cpp index 0f1479f64c..54545860f8 100644 --- a/csrc/model/zip_model_impl.cpp +++ b/csrc/model/zip_model_impl.cpp @@ -7,14 +7,8 @@ #include "core/logger.h" #include "core/model.h" #include "core/model_impl.h" +#include "core/utils/filesystem.h" #include "zip.h" -#if __GNUC__ >= 8 -#include -namespace fs = std::filesystem; -#else -#include -namespace fs = std::experimental::filesystem; -#endif using nlohmann::json; @@ -40,10 +34,10 @@ class ZipModelImpl : public ModelImpl { int ret = 0; zip_ = zip_open(model_path.c_str(), 0, &ret); if (ret != 0) { - INFO("open zip file {} failed, ret {}", model_path.c_str(), ret); + MMDEPLOY_INFO("open zip file {} failed, ret {}", model_path.c_str(), ret); return Status(eInvalidArgument); } - INFO("open sdk model file {} successfully", model_path.c_str()); + MMDEPLOY_INFO("open sdk model file {} successfully", model_path.c_str()); return InitZip(); } @@ -70,24 +64,25 @@ class ZipModelImpl : public ModelImpl { auto iter = file_index_.find(file_path); if (iter == file_index_.end()) { - ERROR("cannot find file {} under dir {}", file_path.c_str(), root_dir_.c_str()); + MMDEPLOY_ERROR("cannot find file {} under dir {}", file_path.c_str(), root_dir_.c_str()); return Status(eFail); } index = iter->second; struct zip_file* pzip = zip_fopen_index(zip_, index, 0); if (nullptr == pzip) { - ERROR("read file {} in zip file failed, whose index is {}", file_path.c_str(), index); + MMDEPLOY_ERROR("read file {} in zip file failed, whose index is {}", file_path.c_str(), + index); return Status(eFail); } struct zip_stat stat {}; if ((ret = zip_stat_index(zip_, index, 0, &stat)) < 0) { - ERROR("get stat of file {} error, ret {}", file_path.c_str(), ret); + MMDEPLOY_ERROR("get stat of file {} error, ret {}", file_path.c_str(), ret); return Status(eFail); } - DEBUG("file size {}", (int)stat.size); + MMDEPLOY_DEBUG("file size {}", (int)stat.size); std::vector buf(stat.size); if ((ret = zip_fread(pzip, buf.data(), stat.size)) < 0) { - ERROR("read data of file {} error, ret {}", file_path.c_str(), ret); + MMDEPLOY_ERROR("read data of file {} error, ret {}", file_path.c_str(), ret); return Status(eFail); } return std::string(buf.begin(), buf.end()); @@ -100,7 +95,7 @@ class ZipModelImpl : public ModelImpl { from_json(json::parse(deploy_json), meta); return meta; } catch (std::exception& e) { - ERROR("exception happened: {}", e.what()); + MMDEPLOY_ERROR("exception happened: {}", e.what()); return Status(eFail); } } @@ -108,7 +103,7 @@ class ZipModelImpl : public ModelImpl { private: Result InitZip() { int files = zip_get_num_files(zip_); - INFO("there are {} files in sdk model file", files); + MMDEPLOY_INFO("there are {} files in sdk model file", files); if (files == 0) { return Status(eFail); } @@ -119,9 +114,9 @@ class ZipModelImpl : public ModelImpl { fs::path path(stat.name); auto file_name = path.filename().string(); if (file_name == ".") { - DEBUG("{}-th file name is: {}， which is a directory", i, stat.name); + MMDEPLOY_DEBUG("{}-th file name is: {}， which is a directory", i, stat.name); } else { - DEBUG("{}-th file name is: {}， which is a file", i, stat.name); + MMDEPLOY_DEBUG("{}-th file name is: {}， which is a file", i, stat.name); file_index_[file_name] = i; } } diff --git a/csrc/net/CMakeLists.txt b/csrc/net/CMakeLists.txt index f339801c04..f411abe5e4 100644 --- a/csrc/net/CMakeLists.txt +++ b/csrc/net/CMakeLists.txt @@ -2,7 +2,7 @@ cmake_minimum_required(VERSION 3.14) project(mmdeploy_net_module) -include(${CMAKE_SOURCE_DIR}/cmake/common.cmake) +include(${CMAKE_SOURCE_DIR}/cmake/MMDeploy.cmake) if ("trt" IN_LIST MMDEPLOY_TARGET_BACKENDS) add_subdirectory(trt) @@ -24,7 +24,5 @@ if ("openvino" IN_LIST MMDEPLOY_TARGET_BACKENDS) add_subdirectory(openvino) endif () -build_target(${PROJECT_NAME} net_module.cpp) -target_link_libraries(${PROJECT_NAME} PRIVATE mmdeploy::core) +mmdeploy_add_module(${PROJECT_NAME} net_module.cpp) add_library(mmdeploy::net_module ALIAS ${PROJECT_NAME}) -export_module(${PROJECT_NAME}) diff --git a/csrc/net/ncnn/CMakeLists.txt b/csrc/net/ncnn/CMakeLists.txt index 5e83abbf53..3c0e1ff6e7 100644 --- a/csrc/net/ncnn/CMakeLists.txt +++ b/csrc/net/ncnn/CMakeLists.txt @@ -2,19 +2,17 @@ cmake_minimum_required(VERSION 3.14) project(mmdeploy_ncnn_net) -if("cpu" IN_LIST MMDEPLOY_TARGET_DEVICES) - include(${CMAKE_SOURCE_DIR}/cmake/common.cmake) +if ("cpu" IN_LIST MMDEPLOY_TARGET_DEVICES) + include(${CMAKE_SOURCE_DIR}/cmake/MMDeploy.cmake) - find_package(ncnn REQUIRED) - add_library(${PROJECT_NAME} SHARED ncnn_net.cpp) - target_link_libraries(${PROJECT_NAME} PRIVATE mmdeploy::core ncnn) - target_link_libraries( - ${PROJECT_NAME} PRIVATE -Wl,--whole-archive mmdeploy::ncnn_ops::static - -Wl,--no-whole-archive) - add_library(mmdeploy::ncnn_net ALIAS ${PROJECT_NAME}) - export_module(${PROJECT_NAME}) -else() - message( - ERROR - "'ncnn_net' is NOT supported in target devices: ${MMDEPLOY_TARGET_DEVICES}") -endif() + find_package(ncnn REQUIRED) + + mmdeploy_add_module(${PROJECT_NAME} ncnn_net.cpp) + target_link_libraries(${PROJECT_NAME} PRIVATE mmdeploy_ncnn_ops_obj) + target_link_libraries(${PROJECT_NAME} PRIVATE ncnn) + add_library(mmdeploy::ncnn_net ALIAS ${PROJECT_NAME}) +else () + message( + ERROR + "'ncnn_net' is NOT supported in target devices: ${MMDEPLOY_TARGET_DEVICES}") +endif () diff --git a/csrc/net/ncnn/ncnn_net.cpp b/csrc/net/ncnn/ncnn_net.cpp index 29b37242e3..1a958ccfe8 100644 --- a/csrc/net/ncnn/ncnn_net.cpp +++ b/csrc/net/ncnn/ncnn_net.cpp @@ -5,6 +5,7 @@ #include "core/logger.h" #include "core/model.h" #include "core/utils/formatter.h" +#include "ncnn_ops_register.h" namespace mmdeploy { @@ -33,27 +34,27 @@ Result NCNNNet::Init(const Value& args) { OUTCOME_TRY(params_, model.ReadFile(config.net)); OUTCOME_TRY(weights_, model.ReadFile(config.weights)); + register_mmdeploy_custom_layers(net_); + OUTCOME_TRY(ncnn_status(net_.load_param_mem(params_.c_str()))); net_.load_model(reinterpret_cast(weights_.data())); input_indices_ = net_.input_indexes(); for (const auto& x : net_.input_names()) { - // input_names_.emplace_back(x); input_tensors_.emplace_back(TensorDesc{ - .device = Device("cpu"), - .data_type = DataType::kFLOAT, - .shape = {}, - .name = x, + Device("cpu"), + DataType::kFLOAT, + {}, + x, }); } output_indices_ = net_.output_indexes(); for (const auto& x : net_.output_names()) { - // output_names_.emplace_back(x); output_tensors_.emplace_back(TensorDesc{ - .device = Device("cpu"), - .data_type = DataType::kFLOAT, - .shape = {}, - .name = x, + Device("cpu"), + DataType::kFLOAT, + {}, + x, }); } @@ -107,7 +108,7 @@ class NCNNNetCreator : public Creator { if (auto r = p->Init(args)) { return p; } else { - ERROR("error creating NCNNNet: {}", r.error().message().c_str()); + MMDEPLOY_ERROR("error creating NCNNNet: {}", r.error().message().c_str()); return nullptr; } } diff --git a/csrc/net/net_module.cpp b/csrc/net/net_module.cpp index 7ebd78df94..1216d4d7ee 100644 --- a/csrc/net/net_module.cpp +++ b/csrc/net/net_module.cpp @@ -24,7 +24,7 @@ struct NetModule::Impl { using Output = std::map; explicit Impl(const Value& args) { - DEBUG("Net Module cfg: {}", args); + MMDEPLOY_DEBUG("Net Module cfg: {}", args); auto init = [&]() -> Result { auto name = args["name"].get(); auto& context = args["context"]; @@ -34,7 +34,7 @@ struct NetModule::Impl { stream_ = context.value("stream", Stream::GetDefault(device_)); auto creator = Registry::Get().GetCreator(config.backend); if (!creator) { - ERROR("Net backend not found: {}", config.backend); + MMDEPLOY_ERROR("Net backend not found: {}", config.backend); return Status(eEntryNotFound); } auto net_cfg = args; @@ -82,13 +82,13 @@ struct NetModule::Impl { return shape; } if (shape[0] != 1) { - ERROR("unsupported shape for batch assemble: {}", shape); + MMDEPLOY_ERROR("unsupported shape for batch assemble: {}", shape); return Status(eNotSupported); } for (int i = 1; i < input.size(); ++i) { auto& sample = input[i]; if (sample.shape() != shape) { - ERROR("shapes are not consistent across the batch"); + MMDEPLOY_ERROR("shapes are not consistent across the batch"); return Status(eNotSupported); } } @@ -122,7 +122,7 @@ struct NetModule::Impl { if (auto it = sample.find(name); it != sample.end()) { tmp.push_back(it->second); } else { - ERROR("sample {} missing key {}", i, name); + MMDEPLOY_ERROR("sample {} missing key {}", i, name); return Status(eInvalidArgument); } } @@ -140,7 +140,7 @@ struct NetModule::Impl { auto& src = input_samples[i]; auto& dst = inputs_[i]; if (dst.shape() != input_shapes[i]) { - ERROR("inconsistent input shape, expect {}, got {}", input_shapes[i], dst.shape()); + MMDEPLOY_ERROR("inconsistent input shape, expect {}, got {}", input_shapes[i], dst.shape()); return Status(eFail); } if (src.size() > 1) { @@ -165,7 +165,7 @@ struct NetModule::Impl { if (tmp.size()) { OUTCOME_TRY(t.CopyTo(tmp, stream_)); } else { - WARN("copy skipped due to zero sized tensor"); + MMDEPLOY_WARN("copy skipped due to zero sized tensor"); } if (output.size() > 1) { for (int i = 0; i < output.size(); ++i) { diff --git a/csrc/net/openvino/CMakeLists.txt b/csrc/net/openvino/CMakeLists.txt index 14542aa94c..6963739e21 100644 --- a/csrc/net/openvino/CMakeLists.txt +++ b/csrc/net/openvino/CMakeLists.txt @@ -3,15 +3,13 @@ cmake_minimum_required(VERSION 3.14) project(mmdeploy_openvino_net) if ("cpu" IN_LIST MMDEPLOY_TARGET_DEVICES) - include(${CMAKE_SOURCE_DIR}/cmake/common.cmake) + include(${CMAKE_SOURCE_DIR}/cmake/MMDeploy.cmake) find_package(InferenceEngine REQUIRED) - add_library(${PROJECT_NAME} SHARED openvino_net.cpp) + mmdeploy_add_module(${PROJECT_NAME} openvino_net.cpp) target_link_libraries(${PROJECT_NAME} PRIVATE - mmdeploy::core ${InferenceEngine_LIBRARIES}) add_library(mmdeploy::openvino_net ALIAS ${PROJECT_NAME}) - export_module(${PROJECT_NAME}) else () message(ERROR "'openvino_net' is NOT supported in target devices: ${MMDEPLOY_TARGET_DEVICES}") endif () diff --git a/csrc/net/openvino/openvino_net.cpp b/csrc/net/openvino/openvino_net.cpp index af20899a6b..1176967c9a 100644 --- a/csrc/net/openvino/openvino_net.cpp +++ b/csrc/net/openvino/openvino_net.cpp @@ -3,17 +3,11 @@ #include -#if __GNUC__ >= 8 -#include -namespace fs = std::filesystem; -#else -#include -namespace fs = std::experimental::filesystem; -#endif #include #include "core/logger.h" #include "core/model.h" +#include "core/utils/filesystem.h" #include "core/utils/formatter.h" namespace mmdeploy { @@ -40,7 +34,7 @@ static Result ConvertElementType(InferenceEngine::Precision prec) { case InferenceEngine::Precision::ePrecision::I64: return DataType::kINT64; default: - ERROR("unsupported InferenceEngine Precision: {}", static_cast(type)); + MMDEPLOY_ERROR("unsupported InferenceEngine Precision: {}", static_cast(type)); return Status(eNotSupported); } } @@ -58,7 +52,7 @@ static Result ConvertPrecision(DataType case DataType::kINT64: return InferenceEngine::Precision::ePrecision::I64; default: - ERROR("unsupported DataType: {}", static_cast(type)); + MMDEPLOY_ERROR("unsupported DataType: {}", static_cast(type)); return Status(eNotSupported); } } @@ -99,7 +93,7 @@ Result OpenVINONet::Init(const Value& args) { bin_out << raw_bin; bin_out.close(); } catch (const std::exception& e) { - ERROR("unhandled exception when creating tmp xml/bin: {}", e.what()); + MMDEPLOY_ERROR("unhandled exception when creating tmp xml/bin: {}", e.what()); return Status(eFail); } @@ -116,8 +110,7 @@ Result OpenVINONet::Init(const Value& args) { OUTCOME_TRY(auto data_type, ConvertElementType(input_data->getPrecision())); const auto& size_vector = input_data->getTensorDesc().getDims(); TensorShape shape{size_vector.begin(), size_vector.end()}; - input_tensors_.emplace_back(TensorDesc{ - .device = device_, .data_type = data_type, .shape = shape, .name = input_name}); + input_tensors_.emplace_back(TensorDesc{device_, data_type, shape, input_name}); } // set output tensor @@ -128,8 +121,7 @@ Result OpenVINONet::Init(const Value& args) { OUTCOME_TRY(auto data_type, ConvertElementType(output_data->getPrecision())); const auto& size_vector = output_data->getDims(); TensorShape shape{size_vector.begin(), size_vector.end()}; - output_tensors_.emplace_back(TensorDesc{ - .device = device_, .data_type = data_type, .shape = shape, .name = output_name}); + output_tensors_.emplace_back(TensorDesc{device_, data_type, shape, output_name}); } // create request @@ -141,7 +133,7 @@ Result OpenVINONet::Init(const Value& args) { request_ = executable_network.CreateInferRequest(); } catch (const std::exception& e) { - ERROR("unhandled exception when creating OpenVINO: {}", e.what()); + MMDEPLOY_ERROR("unhandled exception when creating OpenVINO: {}", e.what()); return Status(eFail); } return success(); @@ -190,7 +182,7 @@ static Result SetBlob(InferenceEngine::InferRequest& request, Tensor& tens InferenceEngine::make_shared_blob(ie_desc, tensor.data())); break; default: - ERROR("unsupported DataType: {}", static_cast(desc.data_type)); + MMDEPLOY_ERROR("unsupported DataType: {}", static_cast(desc.data_type)); return Status(eNotSupported); } return success(); @@ -211,9 +203,7 @@ static Result GetBlob(InferenceEngine::InferRequest& request, Tensor& tens auto moutputHolder = moutput->rmap(); std::shared_ptr data(const_cast(moutputHolder.as()), [](void*) {}); - Tensor blob_tensor = { - TensorDesc{.device = device, .data_type = data_type, .shape = shape, .name = output_name}, - data}; + Tensor blob_tensor = {TensorDesc{device, data_type, shape, output_name}, data}; if (!std::equal(blob_tensor.shape().begin(), blob_tensor.shape().end(), tensor.shape().begin())) tensor.Reshape(shape); OUTCOME_TRY(tensor.CopyFrom(blob_tensor, stream)); @@ -272,11 +262,11 @@ class OpenVINONetCreator : public Creator { if (auto r = p->Init(args)) { return p; } else { - ERROR("error creating OpenVINONet: {}", r.error().message().c_str()); + MMDEPLOY_ERROR("error creating OpenVINONet: {}", r.error().message().c_str()); return nullptr; } } catch (const std::exception& e) { - ERROR("unhandled exception when creating OpenVINONet: {}", e.what()); + MMDEPLOY_ERROR("unhandled exception when creating OpenVINONet: {}", e.what()); return nullptr; } } diff --git a/csrc/net/ort/CMakeLists.txt b/csrc/net/ort/CMakeLists.txt index 4b7af7aa52..b4b78eff47 100644 --- a/csrc/net/ort/CMakeLists.txt +++ b/csrc/net/ort/CMakeLists.txt @@ -3,18 +3,12 @@ cmake_minimum_required(VERSION 3.14) project(mmdeploy_ort_net) if ("cpu" IN_LIST MMDEPLOY_TARGET_DEVICES) - include(${CMAKE_SOURCE_DIR}/cmake/common.cmake) - add_library(${PROJECT_NAME} SHARED ort_net.cpp) - target_include_directories(${PROJECT_NAME} PUBLIC ${ONNXRUNTIME_DIR}/include) + include(${CMAKE_SOURCE_DIR}/cmake/MMDeploy.cmake) + mmdeploy_add_module(${PROJECT_NAME} ort_net.cpp) + target_include_directories(${PROJECT_NAME} PRIVATE ${ONNXRUNTIME_DIR}/include) target_link_directories(${PROJECT_NAME} PUBLIC ${ONNXRUNTIME_DIR}/lib) - target_link_libraries(${PROJECT_NAME} PRIVATE mmdeploy::core - PUBLIC onnxruntime) - target_link_libraries(${PROJECT_NAME} PRIVATE - -Wl,--whole-archive - mmdeploy::onnxruntime::ops::static - -Wl,--no-whole-archive) + target_link_libraries(${PROJECT_NAME} PRIVATE mmdeploy_onnxruntime_ops_obj) add_library(mmdeploy::ort_net ALIAS ${PROJECT_NAME}) - export_module(${PROJECT_NAME}) else () message(ERROR "'ort_net' is NOT supported in target devices: ${MMDEPLOY_TARGET_DEVICES}") endif () diff --git a/csrc/net/ort/ort_net.cpp b/csrc/net/ort/ort_net.cpp index 5a2fb2e637..10ab9f6e1f 100644 --- a/csrc/net/ort/ort_net.cpp +++ b/csrc/net/ort/ort_net.cpp @@ -1,9 +1,13 @@ // Copyright (c) OpenMMLab. All rights reserved. + #include "ort_net.h" +#include + #include "core/logger.h" #include "core/model.h" #include "core/utils/formatter.h" +#include "onnxruntime_register.h" namespace mmdeploy { @@ -25,7 +29,7 @@ static Result ConvertElementType(ONNXTensorElementDataType type) { case ONNX_TENSOR_ELEMENT_DATA_TYPE_INT64: return DataType::kINT64; default: - ERROR("unsupported ONNXTensorElementDataType: {}", static_cast(type)); + MMDEPLOY_ERROR("unsupported ONNXTensorElementDataType: {}", static_cast(type)); return Status(eNotSupported); } } @@ -45,6 +49,9 @@ Result OrtNet::Init(const Value& args) { Ort::SessionOptions options; options.SetLogSeverityLevel(3); + + RegisterCustomOps(options, OrtGetApiBase()); + if (device_.is_device()) { OrtCUDAProviderOptions cuda_options{}; cuda_options.device_id = device_.device_id(); @@ -69,12 +76,11 @@ Result OrtNet::Init(const Value& args) { auto input_name = session_.GetInputName(i, allocator); auto type_info = session_.GetInputTypeInfo(i); auto shape = to_shape(type_info); - INFO("input {}, shape = {}", i, shape); + MMDEPLOY_INFO("input {}, shape = {}", i, shape); filter_shape(shape); OUTCOME_TRY(auto data_type, ConvertElementType(type_info.GetTensorTypeAndShapeInfo().GetElementType())); - input_tensors_.emplace_back( - TensorDesc{.device = device_, .data_type = data_type, .shape = shape, .name = input_name}); + input_tensors_.emplace_back(TensorDesc{device_, data_type, shape, input_name}); allocator.Free(input_name); } @@ -84,12 +90,11 @@ Result OrtNet::Init(const Value& args) { auto output_name = session_.GetOutputName(i, allocator); auto type_info = session_.GetOutputTypeInfo(i); auto shape = to_shape(type_info); - INFO("output {}, shape = {}", i, shape); + MMDEPLOY_INFO("output {}, shape = {}", i, shape); filter_shape(shape); OUTCOME_TRY(auto data_type, ConvertElementType(type_info.GetTensorTypeAndShapeInfo().GetElementType())); - output_tensors_.emplace_back( - TensorDesc{.device = device_, .data_type = data_type, .shape = shape, .name = output_name}); + output_tensors_.emplace_back(TensorDesc{device_, data_type, shape, output_name}); allocator.Free(output_name); } @@ -166,7 +171,7 @@ Result OrtNet::Forward() { OUTCOME_TRY(stream_.Wait()); } catch (const std::exception& e) { - ERROR(e.what()); + MMDEPLOY_ERROR(e.what()); return Status(eFail); } return success(); @@ -182,11 +187,11 @@ class OrtNetCreator : public Creator { if (auto r = p->Init(args)) { return p; } else { - ERROR("error creating OrtNet: {}", r.error().message().c_str()); + MMDEPLOY_ERROR("error creating OrtNet: {}", r.error().message().c_str()); return nullptr; } } catch (const std::exception& e) { - ERROR("unhandled exception when creating ORTNet: {}", e.what()); + MMDEPLOY_ERROR("unhandled exception when creating ORTNet: {}", e.what()); return nullptr; } } diff --git a/csrc/net/ppl/CMakeLists.txt b/csrc/net/ppl/CMakeLists.txt index cb6c0fb31c..dd859f8e29 100644 --- a/csrc/net/ppl/CMakeLists.txt +++ b/csrc/net/ppl/CMakeLists.txt @@ -2,10 +2,10 @@ cmake_minimum_required(VERSION 3.14) project(mmdeploy_pplnn_net) -include(${CMAKE_SOURCE_DIR}/cmake/common.cmake) +include(${CMAKE_SOURCE_DIR}/cmake/MMDeploy.cmake) find_package(pplnn REQUIRED) -add_library(${PROJECT_NAME} SHARED ppl_net.cpp) +mmdeploy_add_module(${PROJECT_NAME} ppl_net.cpp) target_include_directories(${PROJECT_NAME} PUBLIC $) if ("cpu" IN_LIST MMDEPLOY_TARGET_DEVICES) @@ -17,7 +17,5 @@ if ("cuda" IN_LIST MMDEPLOY_TARGET_DEVICES) target_link_directories(${PROJECT_NAME} PUBLIC ${CUDA_TOOLKIT_ROOT_DIR}/lib64) endif () target_link_libraries(${PROJECT_NAME} - PRIVATE mmdeploy::core ${PPLNN_LIBRARIES} - PUBLIC nvrtc) + PRIVATE ${PPLNN_LIBRARIES} nvrtc) add_library(mmdeploy::pplnn_net ALIAS ${PROJECT_NAME}) -export_module(${PROJECT_NAME}) diff --git a/csrc/net/ppl/ppl_net.cpp b/csrc/net/ppl/ppl_net.cpp index 72e8f99ac6..f0d4b16030 100644 --- a/csrc/net/ppl/ppl_net.cpp +++ b/csrc/net/ppl/ppl_net.cpp @@ -22,7 +22,7 @@ Result ppl_try(int code) { if (code == 0) { return success(); } - ERROR("ppl error: {}", ppl::common::GetRetCodeStr(code)); + MMDEPLOY_ERROR("ppl error: {}", ppl::common::GetRetCodeStr(code)); return Status(eFail); } @@ -86,9 +86,9 @@ Result PPLNet::Init(const Value& args) { /// debug only auto& desc = inputs_internal_[i]->GetShape(); std::vector shape_(desc.GetDims(), desc.GetDims() + desc.GetDimCount()); - DEBUG("input {}: datatype = {}, dataformat = {}, shape = {}", i, - ppl::common::GetDataTypeStr(desc.GetDataType()), - ppl::common::GetDataFormatStr(desc.GetDataFormat()), shape_); + MMDEPLOY_DEBUG("input {}: datatype = {}, dataformat = {}, shape = {}", i, + ppl::common::GetDataTypeStr(desc.GetDataType()), + ppl::common::GetDataFormatStr(desc.GetDataFormat()), shape_); } for (int i = 0; i < runtime->GetOutputCount(); ++i) { @@ -98,9 +98,9 @@ Result PPLNet::Init(const Value& args) { auto desc = outputs_internal_[i]->GetShape(); std::vector shape_(desc.GetDims(), desc.GetDims() + desc.GetDimCount()); - DEBUG("output {}: datatype = {}, dataformat = {}, shape = {}", i, - ppl::common::GetDataTypeStr(desc.GetDataType()), - ppl::common::GetDataFormatStr(desc.GetDataFormat()), shape_); + MMDEPLOY_DEBUG("output {}: datatype = {}, dataformat = {}, shape = {}", i, + ppl::common::GetDataTypeStr(desc.GetDataType()), + ppl::common::GetDataFormatStr(desc.GetDataFormat()), shape_); TensorShape shape(desc.GetDims(), desc.GetDims() + desc.GetDimCount()); } @@ -176,8 +176,8 @@ Result PPLNet::Forward() { auto& internal = *outputs_internal_[i]; auto format = internal.GetShape().GetDataFormat(); if (format != ppl::common::DATAFORMAT_NDARRAY) { - ERROR("output {}'s format is {}, only NDARRAY is currently supported", i, - ppl::common::GetDataFormatStr(format)); + MMDEPLOY_ERROR("output {}'s format is {}, only NDARRAY is currently supported", i, + ppl::common::GetDataFormatStr(format)); return Status(eNotSupported); } auto& external = outputs_external_[i]; @@ -200,7 +200,8 @@ Result PPLNet::Forward() { if (external.size() > 0) { OUTCOME_TRY(Tensor(external.desc(), data).CopyTo(external, stream_)); } else { - WARN("copy skipped due to zero sized tensor: {} {}", external.name(), external.shape()); + MMDEPLOY_WARN("copy skipped due to zero sized tensor: {} {}", external.name(), + external.shape()); } } } @@ -235,7 +236,7 @@ Result PPLNet::Reshape(Span input_shapes) { if (can_infer_output_shapes_) { OUTCOME_TRY(auto output_shapes, InferOutputShapes(input_shapes, prev_in_shapes, prev_out_shapes)); - // ERROR("inferred output shapes: {}", output_shapes); + // MMDEPLOY_ERROR("inferred output shapes: {}", output_shapes); for (int i = 0; i < outputs_external_.size(); ++i) { auto& output = outputs_external_[i]; output.Reshape(output_shapes[i]); @@ -304,7 +305,7 @@ class PPLNetCreator : public Creator { if (auto r = p->Init(args)) { return p; } else { - ERROR("error creating PPLNet: {}", r.error().message().c_str()); + MMDEPLOY_ERROR("error creating PPLNet: {}", r.error().message().c_str()); return nullptr; } } diff --git a/csrc/net/trt/CMakeLists.txt b/csrc/net/trt/CMakeLists.txt index 1368e93352..94f08070b0 100644 --- a/csrc/net/trt/CMakeLists.txt +++ b/csrc/net/trt/CMakeLists.txt @@ -2,24 +2,16 @@ cmake_minimum_required(VERSION 3.14) project(mmdeploy_trt_net) -include(${CMAKE_SOURCE_DIR}/cmake/common.cmake) +include(${CMAKE_SOURCE_DIR}/cmake/MMDeploy.cmake) include(${CMAKE_SOURCE_DIR}/cmake/tensorrt.cmake) -add_library(${PROJECT_NAME} SHARED trt_net.cpp) +mmdeploy_add_module(${PROJECT_NAME} trt_net.cpp) target_include_directories(${PROJECT_NAME} PRIVATE ${TENSORRT_INCLUDE_DIR}) target_include_directories(${PROJECT_NAME} PRIVATE ${CUDNN_DIR}/include) target_include_directories(${PROJECT_NAME} PRIVATE ${CUDA_TOOLKIT_ROOT_DIR}/include) -target_link_directories(${PROJECT_NAME} PUBLIC ${CUDNN_DIR}/lib64) +target_link_directories(${PROJECT_NAME} PUBLIC ${CUDNN_DIR}/lib64 ${CUDNN_DIR}/lib/x64) +target_link_libraries(${PROJECT_NAME} PRIVATE mmdeploy_tensorrt_ops_obj) target_link_libraries(${PROJECT_NAME} PUBLIC ${TENSORRT_LIBRARY} cudnn) -target_link_libraries(${PROJECT_NAME} - PRIVATE mmdeploy::core - ) -target_link_libraries(${PROJECT_NAME} - PRIVATE -Wl,--whole-archive - mmdeploy::tensorrt_ops::static - -Wl,--no-whole-archive - ) add_library(mmdeploy::trt_net ALIAS ${PROJECT_NAME}) -export_module(${PROJECT_NAME}) diff --git a/csrc/net/trt/trt_net.cpp b/csrc/net/trt/trt_net.cpp index 6f4cb940a1..9300aad10e 100644 --- a/csrc/net/trt/trt_net.cpp +++ b/csrc/net/trt/trt_net.cpp @@ -18,14 +18,14 @@ class TRTLogger : public nvinfer1::ILogger { void log(Severity severity, const char* msg) noexcept override { switch (severity) { case Severity::kINFO: - // INFO("TRTNet: {}", msg); + // MMDEPLOY_INFO("TRTNet: {}", msg); break; case Severity::kWARNING: - WARN("TRTNet: {}", msg); + MMDEPLOY_WARN("TRTNet: {}", msg); break; case Severity::kERROR: case Severity::kINTERNAL_ERROR: - ERROR("TRTNet: {}", msg); + MMDEPLOY_ERROR("TRTNet: {}", msg); break; default: break; @@ -72,7 +72,7 @@ static inline Result trt_try(bool code, const char* msg = nullptr, Status return success(); } if (msg) { - ERROR("{}", msg); + MMDEPLOY_ERROR("{}", msg); } return e; } @@ -102,7 +102,7 @@ Result TRTNet::Init(const Value& args) { auto& context = args["context"]; device_ = context["device"].get(); if (device_.is_host()) { - ERROR("TRTNet: device must be a GPU!"); + MMDEPLOY_ERROR("TRTNet: device must be a GPU!"); return Status(eNotSupported); } stream_ = context["stream"].get(); @@ -129,19 +129,18 @@ Result TRTNet::Init(const Value& args) { auto binding_name = engine_->getBindingName(i); auto dims = engine_->getBindingDimensions(i); if (engine_->isShapeBinding(i)) { - ERROR("shape binding is not supported."); + MMDEPLOY_ERROR("shape binding is not supported."); return Status(eNotSupported); } OUTCOME_TRY(auto dtype, MapDataType(engine_->getBindingDataType(i))); - TensorDesc desc{ - .device = device_, .data_type = dtype, .shape = to_shape(dims), .name = binding_name}; + TensorDesc desc{device_, dtype, to_shape(dims), binding_name}; if (engine_->bindingIsInput(i)) { - DEBUG("input binding {} {} {}", i, binding_name, to_string(dims)); + MMDEPLOY_DEBUG("input binding {} {} {}", i, binding_name, to_string(dims)); input_ids_.push_back(i); input_names_.emplace_back(binding_name); input_tensors_.emplace_back(desc, Buffer()); } else { - DEBUG("output binding {} {} {}", i, binding_name, to_string(dims)); + MMDEPLOY_DEBUG("output binding {} {} {}", i, binding_name, to_string(dims)); output_ids_.push_back(i); output_names_.emplace_back(binding_name); output_tensors_.emplace_back(desc, Buffer()); @@ -169,17 +168,17 @@ Result TRTNet::Reshape(Span input_shapes) { } for (int i = 0; i < input_tensors_.size(); ++i) { auto dims = to_dims(input_shapes[i]); - // ERROR("input shape: {}", to_string(dims)); + // MMDEPLOY_ERROR("input shape: {}", to_string(dims)); TRT_TRY(context_->setBindingDimensions(input_ids_[i], dims)); input_tensors_[i].Reshape(input_shapes[i]); } if (!context_->allInputDimensionsSpecified()) { - ERROR("not all input dimensions specified"); + MMDEPLOY_ERROR("not all input dimensions specified"); return Status(eFail); } for (int i = 0; i < output_tensors_.size(); ++i) { auto dims = context_->getBindingDimensions(output_ids_[i]); - // ERROR("output shape: {}", to_string(dims)); + // MMDEPLOY_ERROR("output shape: {}", to_string(dims)); output_tensors_[i].Reshape(to_shape(dims)); } return success(); diff --git a/csrc/preprocess/CMakeLists.txt b/csrc/preprocess/CMakeLists.txt index ab1084198e..503ead8f8d 100644 --- a/csrc/preprocess/CMakeLists.txt +++ b/csrc/preprocess/CMakeLists.txt @@ -8,8 +8,7 @@ if ("cuda" IN_LIST MMDEPLOY_TARGET_DEVICES) add_subdirectory(cuda) endif () -include(${CMAKE_SOURCE_DIR}/cmake/common.cmake) -build_target(${PROJECT_NAME} transform_module.cpp) -target_link_libraries(${PROJECT_NAME} PRIVATE mmdeploy::core) +include(${CMAKE_SOURCE_DIR}/cmake/MMDeploy.cmake) +mmdeploy_add_module(${PROJECT_NAME} transform_module.cpp) +target_link_libraries(${PROJECT_NAME} PRIVATE mmdeploy::transform) add_library(mmdeploy::transform_module ALIAS ${PROJECT_NAME}) -export_module(${PROJECT_NAME}) diff --git a/csrc/preprocess/cpu/CMakeLists.txt b/csrc/preprocess/cpu/CMakeLists.txt index 00c87cd7ee..d2a75b10e8 100644 --- a/csrc/preprocess/cpu/CMakeLists.txt +++ b/csrc/preprocess/cpu/CMakeLists.txt @@ -3,7 +3,7 @@ cmake_minimum_required(VERSION 3.14) project(mmdeploy_cpu_transform_impl) include(${CMAKE_SOURCE_DIR}/cmake/opencv.cmake) -include(${CMAKE_SOURCE_DIR}/cmake/common.cmake) +include(${CMAKE_SOURCE_DIR}/cmake/MMDeploy.cmake) set(SRCS collect_impl.cpp @@ -11,13 +11,10 @@ set(SRCS image2tensor_impl.cpp load_impl.cpp normalize_impl.cpp - opencv_utils.cpp - opencv_utils.h pad_impl.cpp resize_impl.cpp) -build_target(${PROJECT_NAME} "${SRCS}") +mmdeploy_add_module(${PROJECT_NAME} "${SRCS}") target_link_libraries(${PROJECT_NAME} - PUBLIC opencv_imgproc opencv_core - PRIVATE mmdeploy::core) + PRIVATE mmdeploy::transform + mmdeploy_opencv_utils) add_library(mmdeploy::transform_impl::cpu ALIAS ${PROJECT_NAME}) -export_module(${PROJECT_NAME}) diff --git a/csrc/preprocess/cpu/pad_impl.cpp b/csrc/preprocess/cpu/pad_impl.cpp index 3a139d70ad..c75ba4139b 100644 --- a/csrc/preprocess/cpu/pad_impl.cpp +++ b/csrc/preprocess/cpu/pad_impl.cpp @@ -17,7 +17,7 @@ class PadImpl : public ::mmdeploy::PadImpl { {"reflect", cv::BORDER_REFLECT_101}, {"symmetric", cv::BORDER_REFLECT}}; if (border_map.find(arg_.padding_mode) == border_map.end()) { - ERROR("unsupported padding_mode '{}'", arg_.padding_mode); + MMDEPLOY_ERROR("unsupported padding_mode '{}'", arg_.padding_mode); throw std::invalid_argument("unsupported padding_mode"); } border_type_ = border_map[arg_.padding_mode]; diff --git a/csrc/preprocess/cuda/CMakeLists.txt b/csrc/preprocess/cuda/CMakeLists.txt index ac4def77d3..76caeb214b 100644 --- a/csrc/preprocess/cuda/CMakeLists.txt +++ b/csrc/preprocess/cuda/CMakeLists.txt @@ -2,29 +2,29 @@ cmake_minimum_required(VERSION 3.14) project(mmdeploy_cuda_transform_impl CUDA CXX) -if(${CMAKE_VERSION} VERSION_GREATER_EQUAL "3.18.0") - # suppress 'CMAKE_CUDA_ARCHITECTURES' warning - cmake_policy(SET CMP0104 OLD) -endif() +if (${CMAKE_VERSION} VERSION_GREATER_EQUAL "3.18.0") + # suppress 'CMAKE_CUDA_ARCHITECTURES' warning + cmake_policy(SET CMP0104 OLD) +endif () find_package(pplcv REQUIRED) -include(${CMAKE_SOURCE_DIR}/cmake/common.cmake) +include(${CMAKE_SOURCE_DIR}/cmake/MMDeploy.cmake) set(SRCS - crop_impl.cpp - image2tensor_impl.cpp - load_impl.cpp - normalize_impl.cpp - pad_impl.cpp - resize_impl.cpp - cast.cu - crop.cu - normalize.cu - transpose.cu) -build_target(${PROJECT_NAME} "${SRCS}") + crop_impl.cpp + image2tensor_impl.cpp + load_impl.cpp + normalize_impl.cpp + pad_impl.cpp + resize_impl.cpp + cast.cu + crop.cu + normalize.cu + transpose.cu) +mmdeploy_add_module(${PROJECT_NAME} "${SRCS}") +target_link_libraries(${PROJECT_NAME} PRIVATE + mmdeploy::transform ${PPLCV_LIBRARIES}) target_include_directories(${PROJECT_NAME} - PUBLIC ${CUDA_TOOLKIT_ROOT_DIR}/include) -target_link_libraries(${PROJECT_NAME} PRIVATE ${PPLCV_LIBRARIES} mmdeploy::core) + PUBLIC ${CUDA_TOOLKIT_ROOT_DIR}/include ${PPLCV_INCLUDE_DIRS}) add_library(mmdeploy::transform_impl::cuda ALIAS ${PROJECT_NAME}) -export_module(${PROJECT_NAME}) diff --git a/csrc/preprocess/cuda/crop_impl.cpp b/csrc/preprocess/cuda/crop_impl.cpp index 0808e8c733..eb6f64f835 100644 --- a/csrc/preprocess/cuda/crop_impl.cpp +++ b/csrc/preprocess/cuda/crop_impl.cpp @@ -43,7 +43,7 @@ class CenterCropImpl : public ::mmdeploy::CenterCropImpl { } else if (1 == c) { Crop(input, desc.shape[2], output, h, w, top, left, stream); } else { - ERROR("unsupported channels {}", c); + MMDEPLOY_ERROR("unsupported channels {}", c); return Status(eNotSupported); } } else if (DataType::kFLOAT == type) { @@ -54,11 +54,11 @@ class CenterCropImpl : public ::mmdeploy::CenterCropImpl { } else if (1 == c) { Crop(input, desc.shape[2], output, h, w, top, left, stream); } else { - ERROR("unsupported channels {}", c); + MMDEPLOY_ERROR("unsupported channels {}", c); return Status(eNotSupported); } } else { - ERROR("unsupported channels {}", c); + MMDEPLOY_ERROR("unsupported channels {}", c); return Status(eNotSupported); } return dst_tensor; diff --git a/csrc/preprocess/cuda/load_impl.cpp b/csrc/preprocess/cuda/load_impl.cpp index 2d8df26b87..e7ffe506d2 100644 --- a/csrc/preprocess/cuda/load_impl.cpp +++ b/csrc/preprocess/cuda/load_impl.cpp @@ -70,11 +70,11 @@ class PrepareImageImpl : public ::mmdeploy::PrepareImageImpl { BGRA2BGR(stream, src_h, src_w, src_stride, src_ptr, dst_stride, dst_ptr); break; default: - ERROR("src type: unknown type {}", img.pixel_format()); + MMDEPLOY_ERROR("src type: unknown type {}", img.pixel_format()); return Status(eNotSupported); } if (ret != 0) { - ERROR("color transfer from {} to BGR failed, ret {}", img.pixel_format(), ret); + MMDEPLOY_ERROR("color transfer from {} to BGR failed, ret {}", img.pixel_format(), ret); return Status(eFail); } if (arg_.to_float32) { @@ -140,11 +140,11 @@ class PrepareImageImpl : public ::mmdeploy::PrepareImageImpl { BGRA2GRAY(stream, src_h, src_w, src_stride, src_ptr, dst_stride, dst_ptr); break; default: - ERROR("src type: unknown type {}", img.pixel_format()); + MMDEPLOY_ERROR("src type: unknown type {}", img.pixel_format()); throw Status(eNotSupported); } if (ret != 0) { - ERROR("color transfer from {} to Gray failed", img.pixel_format()); + MMDEPLOY_ERROR("color transfer from {} to Gray failed", img.pixel_format()); throw Status(eFail); } if (arg_.to_float32) { diff --git a/csrc/preprocess/cuda/normalize.cu b/csrc/preprocess/cuda/normalize.cu index 696abcc7d3..9536ecd054 100644 --- a/csrc/preprocess/cuda/normalize.cu +++ b/csrc/preprocess/cuda/normalize.cu @@ -1,5 +1,7 @@ // Copyright (c) OpenMMLab. All rights reserved. +#include + #include #include @@ -12,7 +14,7 @@ __global__ void normalize(const T* src, int height, int width, int stride, float int x = (int)(blockIdx.x * blockDim.x + threadIdx.x); int y = (int)(blockIdx.y * blockDim.y + threadIdx.y); - if (x >= width or y >= height) { + if (x >= width || y >= height) { return; } diff --git a/csrc/preprocess/cuda/normalize_impl.cpp b/csrc/preprocess/cuda/normalize_impl.cpp index 639f31aa98..48e6647990 100644 --- a/csrc/preprocess/cuda/normalize_impl.cpp +++ b/csrc/preprocess/cuda/normalize_impl.cpp @@ -41,7 +41,7 @@ class NormalizeImpl : public ::mmdeploy::NormalizeImpl { Normalize(input, h, w, stride, output, arg_.mean.data(), arg_.std.data(), arg_.to_rgb, stream); } else { - ERROR("unsupported channels {}", c); + MMDEPLOY_ERROR("unsupported channels {}", c); return Status(eNotSupported); } } else if (DataType::kFLOAT == src_desc.data_type) { @@ -53,11 +53,11 @@ class NormalizeImpl : public ::mmdeploy::NormalizeImpl { Normalize(input, h, w, stride, output, arg_.mean.data(), arg_.std.data(), arg_.to_rgb, stream); } else { - ERROR("unsupported channels {}", c); + MMDEPLOY_ERROR("unsupported channels {}", c); return Status(eNotSupported); } } else { - ERROR("unsupported data type {}", src_desc.data_type); + MMDEPLOY_ERROR("unsupported data type {}", src_desc.data_type); assert(0); return Status(eNotSupported); } diff --git a/csrc/preprocess/cuda/pad_impl.cpp b/csrc/preprocess/cuda/pad_impl.cpp index ae567cb092..77781c0485 100644 --- a/csrc/preprocess/cuda/pad_impl.cpp +++ b/csrc/preprocess/cuda/pad_impl.cpp @@ -14,12 +14,20 @@ namespace cuda { class PadImpl : public ::mmdeploy::PadImpl { public: explicit PadImpl(const Value& args) : ::mmdeploy::PadImpl(args) { +#if PPLCV_VERSION_MAJOR >= 0 && PPLCV_VERSION_MINOR >= 6 && PPLCV_VERSION_PATCH >= 2 + map border_map{{"constant", ppl::cv::BORDER_CONSTANT}, + {"edge", ppl::cv::BORDER_REPLICATE}, + {"reflect", ppl::cv::BORDER_REFLECT_101}, + { "symmetric", + ppl::cv::BORDER_REFLECT }}; +#else map border_map{{"constant", ppl::cv::BORDER_TYPE_CONSTANT}, {"edge", ppl::cv::BORDER_TYPE_REPLICATE}, {"reflect", ppl::cv::BORDER_TYPE_REFLECT_101}, {"symmetric", ppl::cv::BORDER_TYPE_REFLECT}}; +#endif if (border_map.find(arg_.padding_mode) == border_map.end()) { - ERROR("unsupported padding_mode '{}'", arg_.padding_mode); + MMDEPLOY_ERROR("unsupported padding_mode '{}'", arg_.padding_mode); throw_exception(eNotSupported); } padding_mode_ = border_map[arg_.padding_mode]; @@ -55,7 +63,7 @@ class PadImpl : public ::mmdeploy::PadImpl { dst_buffer, padding[1], padding[3], padding[0], padding[2], padding_mode_, arg_.pad_val); } else { - ERROR("unsupported channels {}", c); + MMDEPLOY_ERROR("unsupported channels {}", c); assert(0); return Status(eNotSupported); } @@ -71,17 +79,17 @@ class PadImpl : public ::mmdeploy::PadImpl { stream, height, width, width * c, src_buffer, dst_width * c, dst_buffer, padding[1], padding[3], padding[0], padding[2], padding_mode_, (ppl::cv::uchar)arg_.pad_val); } else { - ERROR("unsupported channels {}", c); + MMDEPLOY_ERROR("unsupported channels {}", c); assert(0); return Status(eNotSupported); } } else { - ERROR("unsupported data type {}", desc.data_type); + MMDEPLOY_ERROR("unsupported data type {}", desc.data_type); assert(0); return Status(eNotSupported); } if (ret != 0) { - ERROR("unexpected exception happened"); + MMDEPLOY_ERROR("unexpected exception happened"); assert(0); return Status(eNotSupported); } diff --git a/csrc/preprocess/cuda/resize_impl.cpp b/csrc/preprocess/cuda/resize_impl.cpp index ce0a891c3c..8a37664801 100644 --- a/csrc/preprocess/cuda/resize_impl.cpp +++ b/csrc/preprocess/cuda/resize_impl.cpp @@ -14,7 +14,7 @@ class ResizeImpl final : public ::mmdeploy::ResizeImpl { public: explicit ResizeImpl(const Value& args) : ::mmdeploy::ResizeImpl(args) { if (arg_.interpolation != "bilinear" && arg_.interpolation != "nearest") { - ERROR("{} interpolation is not supported", arg_.interpolation); + MMDEPLOY_ERROR("{} interpolation is not supported", arg_.interpolation); throw_exception(eNotSupported); } } @@ -33,7 +33,7 @@ class ResizeImpl final : public ::mmdeploy::ResizeImpl { } else if (tensor.data_type() == DataType::kFLOAT) { OUTCOME_TRY(ResizeDispatch(src_tensor, dst_tensor, stream)); } else { - ERROR("unsupported data type {}", tensor.data_type()); + MMDEPLOY_ERROR("unsupported data type {}", tensor.data_type()); return Status(eNotSupported); } return dst_tensor; @@ -42,23 +42,23 @@ class ResizeImpl final : public ::mmdeploy::ResizeImpl { private: template ppl::common::RetCode DispatchImpl(Args&&... args) { -#ifdef PPLCV_VERSION_MAJOR +#if PPLCV_VERSION_MAJOR >= 0 && PPLCV_VERSION_MINOR >= 6 && PPLCV_VERSION_PATCH >= 2 if (arg_.interpolation == "bilinear") { return ppl::cv::cuda::Resize(std::forward(args)..., - ppl::cv::INTERPOLATION_TYPE_LINEAR); + ppl::cv::INTERPOLATION_LINEAR); } if (arg_.interpolation == "nearest") { return ppl::cv::cuda::Resize(std::forward(args)..., - ppl::cv::INTERPOLATION_TYPE_NEAREST_POINT); + ppl::cv::INTERPOLATION_NEAREST_POINT); } - #else -#warning "support for ppl.cv < 0.6 is deprecated and will be dropped in the future" if (arg_.interpolation == "bilinear") { - return ppl::cv::cuda::ResizeLinear(std::forward(args)...); + return ppl::cv::cuda::Resize(std::forward(args)..., + ppl::cv::INTERPOLATION_TYPE_LINEAR); } if (arg_.interpolation == "nearest") { - return ppl::cv::cuda::ResizeNearestPoint(std::forward(args)...); + return ppl::cv::cuda::Resize(std::forward(args)..., + ppl::cv::INTERPOLATION_TYPE_NEAREST_POINT); } #endif return ppl::common::RC_UNSUPPORTED; @@ -82,7 +82,7 @@ class ResizeImpl final : public ::mmdeploy::ResizeImpl { } else if (4 == c) { ret = DispatchImpl(stream, h, w, w * c, input, dst_h, dst_w, dst_w * c, output); } else { - ERROR("unsupported channels {}", c); + MMDEPLOY_ERROR("unsupported channels {}", c); return Status(eNotSupported); } return ret == 0 ? success() : Result(Status(eFail)); diff --git a/csrc/preprocess/transform/CMakeLists.txt b/csrc/preprocess/transform/CMakeLists.txt index e9a9c14026..8e13a67ae1 100644 --- a/csrc/preprocess/transform/CMakeLists.txt +++ b/csrc/preprocess/transform/CMakeLists.txt @@ -2,21 +2,19 @@ cmake_minimum_required(VERSION 3.14) project(mmdeploy_transform) -include(${CMAKE_SOURCE_DIR}/cmake/common.cmake) +include(${CMAKE_SOURCE_DIR}/cmake/MMDeploy.cmake) set(SRCS - collect.cpp - compose.cpp - crop.cpp - image2tensor.cpp - load.cpp - normalize.cpp - pad.cpp - resize.cpp - transform.cpp) -build_target(${PROJECT_NAME} "${SRCS}") + collect.cpp + compose.cpp + crop.cpp + image2tensor.cpp + load.cpp + normalize.cpp + pad.cpp + resize.cpp + transform.cpp) +mmdeploy_add_module(${PROJECT_NAME} LIBRARY "${SRCS}") target_include_directories( - ${PROJECT_NAME} PUBLIC $) -target_link_libraries(${PROJECT_NAME} PRIVATE mmdeploy::core) + ${PROJECT_NAME} PUBLIC $) add_library(mmdeploy::transform ALIAS ${PROJECT_NAME}) -export_module(${PROJECT_NAME}) diff --git a/csrc/preprocess/transform/collect.cpp b/csrc/preprocess/transform/collect.cpp index 673514c780..d01d1cf4b3 100644 --- a/csrc/preprocess/transform/collect.cpp +++ b/csrc/preprocess/transform/collect.cpp @@ -26,7 +26,7 @@ CollectImpl::CollectImpl(const Value &args) { } Result CollectImpl::Process(const Value &input) { - DEBUG("input: {}", to_json(input).dump(2)); + MMDEPLOY_DEBUG("input: {}", to_json(input).dump(2)); Value output; // collect 'ori_img' and 'attribute' from `input`, because those two fields @@ -45,7 +45,7 @@ Result CollectImpl::Process(const Value &input) { } for (auto &key : arg_.keys) { if (!input.contains(key)) { - ERROR("missed key '{}' in input", key); + MMDEPLOY_ERROR("missed key '{}' in input", key); // return Status(eInvalidArgument); return Status(eInvalidArgument); } else { @@ -53,7 +53,7 @@ Result CollectImpl::Process(const Value &input) { } } - DEBUG("output: {}", to_json(output).dump(2)); + MMDEPLOY_DEBUG("output: {}", to_json(output).dump(2)); return output; } @@ -77,4 +77,6 @@ class CollectCreator : public Creator { REGISTER_MODULE(Transform, CollectCreator); +MMDEPLOY_DEFINE_REGISTRY(CollectImpl); + } // namespace mmdeploy diff --git a/csrc/preprocess/transform/collect.h b/csrc/preprocess/transform/collect.h index 92439120f1..327c5191e4 100644 --- a/csrc/preprocess/transform/collect.h +++ b/csrc/preprocess/transform/collect.h @@ -6,7 +6,7 @@ #include "transform.h" namespace mmdeploy { -class CollectImpl : public Module { +class MMDEPLOY_API CollectImpl : public Module { public: explicit CollectImpl(const Value& args); ~CollectImpl() = default; @@ -24,7 +24,7 @@ class CollectImpl : public Module { ArgType arg_; }; -class Collect : public Transform { +class MMDEPLOY_API Collect : public Transform { public: explicit Collect(const Value& args, int version = 0); ~Collect() = default; @@ -35,6 +35,8 @@ class Collect : public Transform { std::unique_ptr impl_; }; +MMDEPLOY_DECLARE_REGISTRY(CollectImpl); + } // namespace mmdeploy #endif // MMDEPLOY_COLLECT_H diff --git a/csrc/preprocess/transform/compose.cpp b/csrc/preprocess/transform/compose.cpp index 57417eb480..a52b6848a1 100644 --- a/csrc/preprocess/transform/compose.cpp +++ b/csrc/preprocess/transform/compose.cpp @@ -17,15 +17,15 @@ Compose::Compose(const Value& args, int version) : Transform(args) { for (auto cfg : args["transforms"]) { cfg["context"] = context; auto type = cfg.value("type", std::string{}); - DEBUG("creating transform: {} with cfg: {}", type, mmdeploy::to_json(cfg).dump(2)); + MMDEPLOY_DEBUG("creating transform: {} with cfg: {}", type, mmdeploy::to_json(cfg).dump(2)); auto creator = Registry::Get().GetCreator(type, version); if (!creator) { - ERROR("unable to find creator: {}", type); + MMDEPLOY_ERROR("unable to find creator: {}", type); throw std::invalid_argument("unable to find creator"); } auto transform = creator->Create(cfg); if (!transform) { - ERROR("failed to create transform: {}", type); + MMDEPLOY_ERROR("failed to create transform: {}", type); throw std::invalid_argument("failed to create transform"); } transforms_.push_back(std::move(transform)); diff --git a/csrc/preprocess/transform/compose.h b/csrc/preprocess/transform/compose.h index 3472d3e209..41f170371c 100644 --- a/csrc/preprocess/transform/compose.h +++ b/csrc/preprocess/transform/compose.h @@ -7,7 +7,7 @@ namespace mmdeploy { -class Compose : public Transform { +class MMDEPLOY_API Compose : public Transform { public: explicit Compose(const Value& args, int version = 0); ~Compose() override = default; diff --git a/csrc/preprocess/transform/crop.cpp b/csrc/preprocess/transform/crop.cpp index d2b9977dc4..1ea8867cab 100644 --- a/csrc/preprocess/transform/crop.cpp +++ b/csrc/preprocess/transform/crop.cpp @@ -24,7 +24,7 @@ CenterCropImpl::CenterCropImpl(const Value& args) : TransformImpl(args) { } Result CenterCropImpl::Process(const Value& input) { - DEBUG("input: {}", to_json(input).dump(2)); + MMDEPLOY_DEBUG("input: {}", to_json(input).dump(2)); auto img_fields = GetImageFields(input); // copy input data, and update its properties @@ -63,14 +63,14 @@ Result CenterCropImpl::Process(const Value& input) { } } - DEBUG("output: {}", to_json(output).dump(2)); + MMDEPLOY_DEBUG("output: {}", to_json(output).dump(2)); return output; } CenterCrop::CenterCrop(const Value& args, int version) : Transform(args) { auto impl_creator = Registry::Get().GetCreator(specified_platform_, version); if (nullptr == impl_creator) { - ERROR("'CenterCrop' is not supported on '{}' platform", specified_platform_); + MMDEPLOY_ERROR("'CenterCrop' is not supported on '{}' platform", specified_platform_); throw std::domain_error("'Resize' is not supported on specified platform"); } impl_ = impl_creator->Create(args); @@ -87,4 +87,5 @@ class CenterCropCreator : public Creator { }; REGISTER_MODULE(Transform, CenterCropCreator); +MMDEPLOY_DEFINE_REGISTRY(CenterCropImpl); } // namespace mmdeploy diff --git a/csrc/preprocess/transform/crop.h b/csrc/preprocess/transform/crop.h index 46bd50737b..76c567271e 100644 --- a/csrc/preprocess/transform/crop.h +++ b/csrc/preprocess/transform/crop.h @@ -3,12 +3,14 @@ #ifndef MMDEPLOY_CROP_H #define MMDEPLOY_CROP_H +#include + #include "core/tensor.h" #include "transform.h" namespace mmdeploy { -class CenterCropImpl : public TransformImpl { +class MMDEPLOY_API CenterCropImpl : public TransformImpl { public: explicit CenterCropImpl(const Value& args); ~CenterCropImpl() = default; @@ -29,7 +31,7 @@ class CenterCropImpl : public TransformImpl { ArgType arg_; }; -class CenterCrop : public Transform { +class MMDEPLOY_API CenterCrop : public Transform { public: explicit CenterCrop(const Value& args, int version = 0); ~CenterCrop() = default; @@ -40,6 +42,8 @@ class CenterCrop : public Transform { std::unique_ptr impl_; }; +MMDEPLOY_DECLARE_REGISTRY(CenterCropImpl); + } // namespace mmdeploy #endif // MMDEPLOY_CROP_H diff --git a/csrc/preprocess/transform/image2tensor.cpp b/csrc/preprocess/transform/image2tensor.cpp index 2adf959904..e2ccd3bb5d 100644 --- a/csrc/preprocess/transform/image2tensor.cpp +++ b/csrc/preprocess/transform/image2tensor.cpp @@ -16,7 +16,7 @@ ImageToTensorImpl::ImageToTensorImpl(const Value& args) : TransformImpl(args) { } Result ImageToTensorImpl::Process(const Value& input) { - DEBUG("input: {}", to_json(input).dump(2)); + MMDEPLOY_DEBUG("input: {}", to_json(input).dump(2)); Value output = input; for (auto& key : arg_.keys) { assert(input.contains(key)); @@ -28,14 +28,14 @@ Result ImageToTensorImpl::Process(const Value& input) { OUTCOME_TRY(output[key], HWC2CHW(src_tensor)); } // for key - DEBUG("output: {}", to_json(output).dump(2)); + MMDEPLOY_DEBUG("output: {}", to_json(output).dump(2)); return output; } ImageToTensor::ImageToTensor(const Value& args, int version) : Transform(args) { auto impl_creator = Registry::Get().GetCreator(specified_platform_, version); if (nullptr == impl_creator) { - ERROR("'ImageToTensor' is not supported on '{}' platform", specified_platform_); + MMDEPLOY_ERROR("'ImageToTensor' is not supported on '{}' platform", specified_platform_); throw std::domain_error("'ImageToTensor' is not supported on specified platform"); } impl_ = impl_creator->Create(args); @@ -53,4 +53,5 @@ class ImageToTensorCreator : public Creator { int version_{1}; }; REGISTER_MODULE(Transform, ImageToTensorCreator); +MMDEPLOY_DEFINE_REGISTRY(ImageToTensorImpl); } // namespace mmdeploy diff --git a/csrc/preprocess/transform/image2tensor.h b/csrc/preprocess/transform/image2tensor.h index cca2c5db7a..49eefd9f47 100644 --- a/csrc/preprocess/transform/image2tensor.h +++ b/csrc/preprocess/transform/image2tensor.h @@ -14,7 +14,7 @@ namespace mmdeploy { * it to (1, C, H, W). * */ -class ImageToTensorImpl : public TransformImpl { +class MMDEPLOY_API ImageToTensorImpl : public TransformImpl { public: ImageToTensorImpl(const Value& args); ~ImageToTensorImpl() = default; @@ -34,7 +34,7 @@ class ImageToTensorImpl : public TransformImpl { ArgType arg_; }; -class ImageToTensor : public Transform { +class MMDEPLOY_API ImageToTensor : public Transform { public: explicit ImageToTensor(const Value& args, int version = 0); ~ImageToTensor() = default; @@ -45,6 +45,8 @@ class ImageToTensor : public Transform { std::unique_ptr impl_; }; +MMDEPLOY_DECLARE_REGISTRY(ImageToTensorImpl); + } // namespace mmdeploy #endif // MMDEPLOY_IMAGE2TENSOR_H diff --git a/csrc/preprocess/transform/load.cpp b/csrc/preprocess/transform/load.cpp index 671948f2d8..462c70a837 100644 --- a/csrc/preprocess/transform/load.cpp +++ b/csrc/preprocess/transform/load.cpp @@ -31,7 +31,7 @@ PrepareImageImpl::PrepareImageImpl(const Value& args) : TransformImpl(args) { */ Result PrepareImageImpl::Process(const Value& input) { - DEBUG("input: {}", to_json(input).dump(2)); + MMDEPLOY_DEBUG("input: {}", to_json(input).dump(2)); assert(input.contains("ori_img")); // copy input data, and update its properties later @@ -50,7 +50,7 @@ Result PrepareImageImpl::Process(const Value& input) { } output["ori_shape"] = {1, src_mat.height(), src_mat.width(), src_mat.channel()}; output["img_fields"].push_back("img"); - DEBUG("output: {}", to_json(output).dump(2)); + MMDEPLOY_DEBUG("output: {}", to_json(output).dump(2)); return output; } @@ -58,7 +58,7 @@ Result PrepareImageImpl::Process(const Value& input) { PrepareImage::PrepareImage(const Value& args, int version) : Transform(args) { auto impl_creator = Registry::Get().GetCreator(specified_platform_, version); if (nullptr == impl_creator) { - ERROR("'PrepareImage' is not supported on '{}' platform", specified_platform_); + MMDEPLOY_ERROR("'PrepareImage' is not supported on '{}' platform", specified_platform_); throw std::domain_error("'PrepareImage' is not supported on specified platform"); } impl_ = impl_creator->Create(args); @@ -80,4 +80,7 @@ class PrepareImageCreator : public Creator { }; REGISTER_MODULE(Transform, PrepareImageCreator); + +MMDEPLOY_DEFINE_REGISTRY(PrepareImageImpl); + } // namespace mmdeploy diff --git a/csrc/preprocess/transform/load.h b/csrc/preprocess/transform/load.h index 32f0bdfdbd..a05d4c136f 100644 --- a/csrc/preprocess/transform/load.h +++ b/csrc/preprocess/transform/load.h @@ -8,7 +8,7 @@ #include "transform.h" namespace mmdeploy { -class PrepareImageImpl : public TransformImpl { +class MMDEPLOY_API PrepareImageImpl : public TransformImpl { public: explicit PrepareImageImpl(const Value& args); ~PrepareImageImpl() = default; @@ -29,7 +29,7 @@ class PrepareImageImpl : public TransformImpl { ArgType arg_; }; -class PrepareImage : public Transform { +class MMDEPLOY_API PrepareImage : public Transform { public: explicit PrepareImage(const Value& args, int version = 0); ~PrepareImage() = default; @@ -40,6 +40,8 @@ class PrepareImage : public Transform { std::unique_ptr impl_; }; +MMDEPLOY_DECLARE_REGISTRY(PrepareImageImpl); + } // namespace mmdeploy #endif // MMDEPLOY_LOAD_H diff --git a/csrc/preprocess/transform/normalize.cpp b/csrc/preprocess/transform/normalize.cpp index a0bc5f7ba2..7fc9c2ad31 100644 --- a/csrc/preprocess/transform/normalize.cpp +++ b/csrc/preprocess/transform/normalize.cpp @@ -10,9 +10,11 @@ using namespace std; namespace mmdeploy { +// MMDEPLOY_DEFINE_REGISTRY(NormalizeImpl); + NormalizeImpl::NormalizeImpl(const Value& args) : TransformImpl(args) { - if (!args.contains("mean") or !args.contains("std")) { - ERROR("no 'mean' or 'std' is configured"); + if (!args.contains("mean") || !args.contains("std")) { + MMDEPLOY_ERROR("no 'mean' or 'std' is configured"); throw std::invalid_argument("no 'mean' or 'std' is configured"); } for (auto& v : args["mean"]) { @@ -50,7 +52,7 @@ NormalizeImpl::NormalizeImpl(const Value& args) : TransformImpl(args) { */ Result NormalizeImpl::Process(const Value& input) { - DEBUG("input: {}", to_json(input).dump(2)); + MMDEPLOY_DEBUG("input: {}", to_json(input).dump(2)); // copy input data, and update its properties later Value output = input; @@ -73,14 +75,14 @@ Result NormalizeImpl::Process(const Value& input) { } output["img_norm_cfg"]["to_rgb"] = arg_.to_rgb; } - DEBUG("output: {}", to_json(output).dump(2)); + MMDEPLOY_DEBUG("output: {}", to_json(output).dump(2)); return output; } Normalize::Normalize(const Value& args, int version) : Transform(args) { auto impl_creator = Registry::Get().GetCreator(specified_platform_, version); if (nullptr == impl_creator) { - ERROR("'Normalize' is not supported on '{}' platform", specified_platform_); + MMDEPLOY_ERROR("'Normalize' is not supported on '{}' platform", specified_platform_); throw std::domain_error("'Normalize' is not supported on specified platform"); } impl_ = impl_creator->Create(args); @@ -98,4 +100,6 @@ class NormalizeCreator : public Creator { REGISTER_MODULE(Transform, NormalizeCreator); +MMDEPLOY_DEFINE_REGISTRY(NormalizeImpl); + } // namespace mmdeploy diff --git a/csrc/preprocess/transform/normalize.h b/csrc/preprocess/transform/normalize.h index 14a4edb43e..fef8fd17c6 100644 --- a/csrc/preprocess/transform/normalize.h +++ b/csrc/preprocess/transform/normalize.h @@ -8,7 +8,7 @@ namespace mmdeploy { -class NormalizeImpl : public TransformImpl { +class MMDEPLOY_API NormalizeImpl : public TransformImpl { public: explicit NormalizeImpl(const Value& args); ~NormalizeImpl() = default; @@ -28,7 +28,7 @@ class NormalizeImpl : public TransformImpl { ArgType arg_; }; -class Normalize : public Transform { +class MMDEPLOY_API Normalize : public Transform { public: explicit Normalize(const Value& args, int version = 0); ~Normalize() = default; @@ -39,5 +39,7 @@ class Normalize : public Transform { std::unique_ptr impl_; }; +MMDEPLOY_DECLARE_REGISTRY(NormalizeImpl); + } // namespace mmdeploy #endif // MMDEPLOY_NORMALIZE_H diff --git a/csrc/preprocess/transform/pad.cpp b/csrc/preprocess/transform/pad.cpp index 4d9c6c69a2..9eb60748f0 100644 --- a/csrc/preprocess/transform/pad.cpp +++ b/csrc/preprocess/transform/pad.cpp @@ -28,7 +28,7 @@ PadImpl::PadImpl(const Value& args) : TransformImpl(args) { } Result PadImpl::Process(const Value& input) { - DEBUG("input: {}", to_json(input).dump(2)); + MMDEPLOY_DEBUG("input: {}", to_json(input).dump(2)); Value output = input; auto img_fields = GetImageFields(input); @@ -38,7 +38,7 @@ Result PadImpl::Process(const Value& input) { assert(tensor.desc().shape.size() == 4); assert(tensor.desc().shape[0] == 1); - assert(tensor.desc().shape[3] == 3 or tensor.desc().shape[3] == 1); + assert(tensor.desc().shape[3] == 3 || tensor.desc().shape[3] == 1); int height = tensor.desc().shape[1]; int width = tensor.desc().shape[2]; @@ -75,14 +75,14 @@ Result PadImpl::Process(const Value& input) { } } - DEBUG("output: {}", to_json(output).dump(2)); + MMDEPLOY_DEBUG("output: {}", to_json(output).dump(2)); return output; } Pad::Pad(const Value& args, int version) : Transform(args) { auto impl_creator = Registry::Get().GetCreator(specified_platform_, version); if (nullptr == impl_creator) { - ERROR("'Pad' is not supported on '{}' platform", specified_platform_); + MMDEPLOY_ERROR("'Pad' is not supported on '{}' platform", specified_platform_); throw std::domain_error("'Pad' is not supported on specified platform"); } impl_ = impl_creator->Create(args); @@ -100,4 +100,6 @@ class PadCreator : public Creator { REGISTER_MODULE(Transform, PadCreator); +MMDEPLOY_DEFINE_REGISTRY(PadImpl); + } // namespace mmdeploy diff --git a/csrc/preprocess/transform/pad.h b/csrc/preprocess/transform/pad.h index e684791a5c..1b5ccbcd5f 100644 --- a/csrc/preprocess/transform/pad.h +++ b/csrc/preprocess/transform/pad.h @@ -3,12 +3,14 @@ #ifndef MMDEPLOY_PAD_H #define MMDEPLOY_PAD_H +#include + #include "core/tensor.h" #include "transform.h" namespace mmdeploy { -class PadImpl : public TransformImpl { +class MMDEPLOY_API PadImpl : public TransformImpl { public: explicit PadImpl(const Value& args); ~PadImpl() override = default; @@ -33,7 +35,7 @@ class PadImpl : public TransformImpl { ArgType arg_; }; -class Pad : public Transform { +class MMDEPLOY_API Pad : public Transform { public: explicit Pad(const Value& args, int version = 0); ~Pad() override = default; @@ -43,6 +45,9 @@ class Pad : public Transform { protected: std::unique_ptr impl_; }; + +MMDEPLOY_DECLARE_REGISTRY(PadImpl); + } // namespace mmdeploy #endif // MMDEPLOY_PAD_H diff --git a/csrc/preprocess/transform/resize.cpp b/csrc/preprocess/transform/resize.cpp index 604103e6c6..98398e3dce 100644 --- a/csrc/preprocess/transform/resize.cpp +++ b/csrc/preprocess/transform/resize.cpp @@ -19,14 +19,14 @@ ResizeImpl::ResizeImpl(const Value& args) : TransformImpl(args) { arg_.img_scale = {size, size}; } else if (args["size"].is_array()) { if (args["size"].size() != 2) { - ERROR("'size' expects an array of size 2, but got {}", args["size"].size()); + MMDEPLOY_ERROR("'size' expects an array of size 2, but got {}", args["size"].size()); throw std::length_error("'size' expects an array of size 2"); } auto height = args["size"][0].get(); auto width = args["size"][1].get(); arg_.img_scale = {height, width}; } else { - ERROR("'size' is expected to be an integer or and array of size 2"); + MMDEPLOY_ERROR("'size' is expected to be an integer or and array of size 2"); throw std::domain_error("'size' is expected to be an integer or and array of size 2"); } } @@ -35,13 +35,13 @@ ResizeImpl::ResizeImpl(const Value& args) : TransformImpl(args) { vector interpolations{"nearest", "bilinear", "bicubic", "area", "lanczos"}; if (std::find(interpolations.begin(), interpolations.end(), arg_.interpolation) == interpolations.end()) { - ERROR("'{}' interpolation is not supported", arg_.interpolation); + MMDEPLOY_ERROR("'{}' interpolation is not supported", arg_.interpolation); throw std::invalid_argument("unexpected interpolation"); } } Result ResizeImpl::Process(const Value& input) { - DEBUG("input: {}", to_json(input).dump(2)); + MMDEPLOY_DEBUG("input: {}", to_json(input).dump(2)); Value output = input; auto img_fields = GetImageFields(input); @@ -66,7 +66,7 @@ Result ResizeImpl::Process(const Value& input) { dst_h = int(h * scale_factor + 0.5); dst_w = int(w * scale_factor + 0.5); } else if (!arg_.img_scale.empty()) { - DEBUG( + MMDEPLOY_WARN( "neither 'scale' or 'scale_factor' is provided in input value. " "'img_scale' will be used"); if (-1 == arg_.img_scale[1]) { @@ -82,7 +82,7 @@ Result ResizeImpl::Process(const Value& input) { dst_w = arg_.img_scale[1]; } } else { - ERROR("no resize related parameter is provided"); + MMDEPLOY_ERROR("no resize related parameter is provided"); return Status(eInvalidArgument); } if (arg_.keep_ratio) { @@ -111,14 +111,14 @@ Result ResizeImpl::Process(const Value& input) { output[key] = dst_img; } - DEBUG("output: {}", to_json(output).dump(2)); + MMDEPLOY_DEBUG("output: {}", to_json(output).dump(2)); return output; } Resize::Resize(const Value& args, int version) : Transform(args) { auto impl_creator = Registry::Get().GetCreator(specified_platform_, version); if (nullptr == impl_creator) { - ERROR("'Resize' is not supported on '{}' platform", specified_platform_); + MMDEPLOY_ERROR("'Resize' is not supported on '{}' platform", specified_platform_); throw std::domain_error("'Resize' is not supported on specified platform"); } impl_ = impl_creator->Create(args); @@ -136,4 +136,6 @@ class ResizeCreator : public Creator { REGISTER_MODULE(Transform, ResizeCreator); +MMDEPLOY_DEFINE_REGISTRY(ResizeImpl); + } // namespace mmdeploy diff --git a/csrc/preprocess/transform/resize.h b/csrc/preprocess/transform/resize.h index 1b9c10034d..54947bee48 100644 --- a/csrc/preprocess/transform/resize.h +++ b/csrc/preprocess/transform/resize.h @@ -3,11 +3,13 @@ #ifndef MMDEPLOY_RESIZE_H #define MMDEPLOY_RESIZE_H +#include + #include "core/tensor.h" #include "transform.h" namespace mmdeploy { -class ResizeImpl : public TransformImpl { +class MMDEPLOY_API ResizeImpl : public TransformImpl { public: explicit ResizeImpl(const Value& args); ~ResizeImpl() override = default; @@ -29,7 +31,7 @@ class ResizeImpl : public TransformImpl { ArgType arg_; }; -class Resize : public Transform { +class MMDEPLOY_API Resize : public Transform { public: explicit Resize(const Value& args, int version = 0); ~Resize() override = default; @@ -40,5 +42,8 @@ class Resize : public Transform { std::unique_ptr impl_; static const std::string name_; }; + +MMDEPLOY_DECLARE_REGISTRY(ResizeImpl); + } // namespace mmdeploy #endif // MMDEPLOY_RESIZE_H diff --git a/csrc/preprocess/transform/transform.cpp b/csrc/preprocess/transform/transform.cpp index d57cfa393d..f2be7519c4 100644 --- a/csrc/preprocess/transform/transform.cpp +++ b/csrc/preprocess/transform/transform.cpp @@ -48,4 +48,6 @@ Transform::Transform(const Value &args) { } } +MMDEPLOY_DEFINE_REGISTRY(Transform); + } // namespace mmdeploy diff --git a/csrc/preprocess/transform/transform.h b/csrc/preprocess/transform/transform.h index fa1a700ea2..ba96e91a14 100644 --- a/csrc/preprocess/transform/transform.h +++ b/csrc/preprocess/transform/transform.h @@ -9,7 +9,7 @@ namespace mmdeploy { -class TransformImpl : public Module { +class MMDEPLOY_API TransformImpl : public Module { public: TransformImpl() = default; explicit TransformImpl(const Value& args); @@ -23,41 +23,36 @@ class TransformImpl : public Module { Stream stream_; }; -class Transform : public Module { +class MMDEPLOY_API Transform : public Module { public: + ~Transform() override = default; + Transform() = default; explicit Transform(const Value& args); - ~Transform() override = default; + Transform(const Transform&) = delete; + Transform& operator=(const Transform&) = delete; const std::string& RuntimePlatform() const { return runtime_platform_; } protected: template - [[deprecated]] - /* - * We cannot LOG the error message, because WARN/INFO/ERROR causes - * redefinition when building UTs "catch2.hpp" used in UTs has the same LOG - * declaration - */ - std::unique_ptr - Instantiate(const char* transform_type, const Value& args, int version = 0) { + [[deprecated]] std::unique_ptr Instantiate(const char* transform_type, const Value& args, + int version = 0) { std::unique_ptr impl(nullptr); auto impl_creator = Registry::Get().GetCreator(specified_platform_, version); if (nullptr == impl_creator) { - // WARN("cannot find {} implementation on specific platform {} ", - // transform_type, specified_platform_); + MMDEPLOY_WARN("cannot find {} implementation on specific platform {} ", transform_type, + specified_platform_); for (auto& name : candidate_platforms_) { impl_creator = Registry::Get().GetCreator(name); if (impl_creator) { - // INFO("fallback {} implementation to platform {}", transform_type, - // name); + MMDEPLOY_INFO("fallback {} implementation to platform {}", transform_type, name); break; } } } if (nullptr == impl_creator) { - // ERROR("cannot find {} implementation on any registered platform ", - // transform_type); + MMDEPLOY_ERROR("cannot find {} implementation on any registered platform ", transform_type); return nullptr; } else { return impl_creator->Create(args); @@ -70,6 +65,8 @@ class Transform : public Module { std::vector candidate_platforms_; }; +MMDEPLOY_DECLARE_REGISTRY(Transform); + } // namespace mmdeploy #endif // MMDEPLOY_TRANSFORM_H diff --git a/csrc/preprocess/transform_module.cpp b/csrc/preprocess/transform_module.cpp index b769878212..9b7b2f01fd 100644 --- a/csrc/preprocess/transform_module.cpp +++ b/csrc/preprocess/transform_module.cpp @@ -3,6 +3,7 @@ #include "transform_module.h" #include "archive/value_archive.h" +#include "core/module.h" #include "core/utils/formatter.h" #include "experimental/module_adapter.h" #include "preprocess/transform/transform.h" @@ -15,12 +16,12 @@ TransformModule::TransformModule(const Value& args) { const auto type = "Compose"; auto creator = Registry::Get().GetCreator(type, 1); if (!creator) { - ERROR("unable to find creator: {}", type); + MMDEPLOY_ERROR("unable to find creator: {}", type); throw_exception(eEntryNotFound); } auto cfg = args; if (cfg.contains("device")) { - WARN("force using device: {}", cfg["device"].get()); + MMDEPLOY_WARN("force using device: {}", cfg["device"].get()); auto device = Device(cfg["device"].get()); cfg["context"]["device"] = device; cfg["context"]["stream"] = Stream::GetDefault(device); @@ -31,7 +32,7 @@ TransformModule::TransformModule(const Value& args) { Result TransformModule::operator()(const Value& input) { auto output = transform_->Process(input); if (!output) { - ERROR("error: {}", output.error().message().c_str()); + MMDEPLOY_ERROR("error: {}", output.error().message().c_str()); } auto& ret = output.value(); if (ret.is_object()) { @@ -39,13 +40,13 @@ Result TransformModule::operator()(const Value& input) { } else if (ret.is_array() && ret.size() == 1 && ret[0].is_object()) { ret = ret[0]; } else { - ERROR("unsupported return value: {}", ret); + MMDEPLOY_ERROR("unsupported return value: {}", ret); return Status(eNotSupported); } return ret; } -class TransformModuleCreator : public Creator { +class MMDEPLOY_API TransformModuleCreator : public Creator { public: const char* GetName() const override { return "Transform"; } int GetVersion() const override { return 0; } diff --git a/csrc/utils/CMakeLists.txt b/csrc/utils/CMakeLists.txt new file mode 100644 index 0000000000..1eef35940a --- /dev/null +++ b/csrc/utils/CMakeLists.txt @@ -0,0 +1,3 @@ +# Copyright (c) OpenMMLab. All rights reserved. + +add_subdirectory(opencv) diff --git a/csrc/utils/opencv/CMakeLists.txt b/csrc/utils/opencv/CMakeLists.txt new file mode 100644 index 0000000000..6eb8bd2e91 --- /dev/null +++ b/csrc/utils/opencv/CMakeLists.txt @@ -0,0 +1,17 @@ +# Copyright (c) OpenMMLab. All rights reserved. +cmake_minimum_required(VERSION 3.14) +project(mmdeploy_opencv_utils) + +include(${CMAKE_SOURCE_DIR}/cmake/opencv.cmake) +include(${CMAKE_SOURCE_DIR}/cmake/MMDeploy.cmake) + +mmdeploy_add_library(${PROJECT_NAME} opencv_utils.cpp) + +target_link_libraries(${PROJECT_NAME} + PRIVATE mmdeploy::core + PUBLIC ${OpenCV_LIBS}) + +target_include_directories(${PROJECT_NAME} + INTERFACE $) + +#export_module(${PROJECT_NAME}) diff --git a/csrc/preprocess/cpu/opencv_utils.cpp b/csrc/utils/opencv/opencv_utils.cpp similarity index 91% rename from csrc/preprocess/cpu/opencv_utils.cpp rename to csrc/utils/opencv/opencv_utils.cpp index d02d5571cd..ef05dbf2eb 100644 --- a/csrc/preprocess/cpu/opencv_utils.cpp +++ b/csrc/utils/opencv/opencv_utils.cpp @@ -42,7 +42,7 @@ cv::Mat Mat2CVMat(const Mat& mat) { {DataType::kINT32, CV_32S}}; auto type = CV_MAKETYPE(type_mapper[mat.type()], mat.channel()); auto format = mat.pixel_format(); - if (PixelFormat::kBGR == format or PixelFormat::kRGB == format) { + if (PixelFormat::kBGR == format || PixelFormat::kRGB == format) { return cv::Mat(mat.height(), mat.width(), type, mat.data()); } else if (PixelFormat::kGRAYSCALE == format) { return cv::Mat(mat.height(), mat.width(), type, mat.data()); @@ -59,7 +59,7 @@ cv::Mat Mat2CVMat(const Mat& mat) { } else if (PixelFormat::kBGRA == format) { return cv::Mat(mat.height(), mat.width(), type, mat.data()); } else { - ERROR("unsupported mat format {}", format); + MMDEPLOY_ERROR("unsupported mat format {}", format); return {}; } } @@ -78,7 +78,7 @@ cv::Mat Tensor2CVMat(const Tensor& tensor) { return {h, w, CV_32SC(c), const_cast(tensor.data())}; } else { assert(0); - ERROR("unsupported type: {}", desc.data_type); + MMDEPLOY_ERROR("unsupported type: {}", desc.data_type); return {}; } } @@ -95,7 +95,7 @@ Tensor CVMat2Tensor(const cv::Mat& mat) { shape = {1, mat.rows, mat.cols, mat.channels()}; data_type = DataType::kINT32; } else { - ERROR("unsupported mat dat type {}", mat.type()); + MMDEPLOY_ERROR("unsupported mat dat type {}", mat.type()); assert(0); return {}; } @@ -118,7 +118,7 @@ cv::Mat Resize(const cv::Mat& src, int dst_height, int dst_width, } else if (interpolation == "lanczos") { cv::resize(src, dst, dst.size(), 0, 0, cv::INTER_LANCZOS4); } else { - ERROR("{} interpolation is not supported", interpolation); + MMDEPLOY_ERROR("{} interpolation is not supported", interpolation); assert(0); } return dst; @@ -189,7 +189,7 @@ cv::Mat ColorTransfer(const cv::Mat& src, PixelFormat src_format, PixelFormat ds cv::cvtColor(src, dst, cv::COLOR_BGRA2BGR); break; default: - ERROR("unsupported src mat's element type {}", src_format); + MMDEPLOY_ERROR("unsupported src mat's element type {}", src_format); assert(0); return {}; } @@ -214,7 +214,7 @@ cv::Mat ColorTransfer(const cv::Mat& src, PixelFormat src_format, PixelFormat ds cv::cvtColor(src, dst, cv::COLOR_BGRA2RGB); break; default: - ERROR("unsupported src mat's element type {}", src_format); + MMDEPLOY_ERROR("unsupported src mat's element type {}", src_format); assert(0); return {}; } @@ -239,12 +239,12 @@ cv::Mat ColorTransfer(const cv::Mat& src, PixelFormat src_format, PixelFormat ds cv::cvtColor(src, dst, cv::COLOR_BGRA2GRAY); break; default: - ERROR("unsupported src mat's element type {}", src_format); + MMDEPLOY_ERROR("unsupported src mat's element type {}", src_format); assert(0); return {}; } } else { - ERROR("unsupported target mat's element type {}", dst_format); + MMDEPLOY_ERROR("unsupported target mat's element type {}", dst_format); assert(0); return {}; } @@ -267,7 +267,7 @@ bool Compare(const cv::Mat& src1, const cv::Mat& src2) { cv::subtract(_src1, _src2, diff); diff = cv::abs(diff); auto sum = cv::sum(cv::sum(diff)); - DEBUG("sum: {}, average: {}", sum[0], sum[0] * 1.0 / (src1.rows * src1.cols)); + MMDEPLOY_DEBUG("sum: {}, average: {}", sum[0], sum[0] * 1.0 / (src1.rows * src1.cols)); return sum[0] / (src1.rows * src1.cols) < 0.5f; } diff --git a/csrc/preprocess/cpu/opencv_utils.h b/csrc/utils/opencv/opencv_utils.h similarity index 69% rename from csrc/preprocess/cpu/opencv_utils.h rename to csrc/utils/opencv/opencv_utils.h index 45aa360ff3..05f8405eb4 100644 --- a/csrc/preprocess/cpu/opencv_utils.h +++ b/csrc/utils/opencv/opencv_utils.h @@ -1,7 +1,7 @@ // Copyright (c) OpenMMLab. All rights reserved. -#ifndef MMDEPLOY_OPENCV_UTILS_H -#define MMDEPLOY_OPENCV_UTILS_H +#ifndef MMDEPLOY_CSRC_UTILS_OPENCV_OPENCV_UTILS_H_ +#define MMDEPLOY_CSRC_UTILS_OPENCV_OPENCV_UTILS_H_ #include "core/mat.h" #include "core/mpl/type_traits.h" @@ -12,11 +12,11 @@ namespace mmdeploy { namespace cpu { -cv::Mat Mat2CVMat(const Mat& mat); -cv::Mat Tensor2CVMat(const Tensor& tensor); +MMDEPLOY_API cv::Mat Mat2CVMat(const Mat& mat); +MMDEPLOY_API cv::Mat Tensor2CVMat(const Tensor& tensor); -Mat CVMat2Mat(const cv::Mat& mat, PixelFormat format); -Tensor CVMat2Tensor(const cv::Mat& mat); +MMDEPLOY_API Mat CVMat2Mat(const cv::Mat& mat, PixelFormat format); +MMDEPLOY_API Tensor CVMat2Tensor(const cv::Mat& mat); /** * @brief resize an image to specified size @@ -26,7 +26,8 @@ Tensor CVMat2Tensor(const cv::Mat& mat); * @param dst_width output image's width * @return output image if success, error code otherwise */ -cv::Mat Resize(const cv::Mat& src, int dst_height, int dst_width, const std::string& interpolation); +MMDEPLOY_API cv::Mat Resize(const cv::Mat& src, int dst_height, int dst_width, + const std::string& interpolation); /** * @brief crop an image @@ -38,7 +39,7 @@ cv::Mat Resize(const cv::Mat& src, int dst_height, int dst_width, const std::str * @param right * @return cv::Mat */ -cv::Mat Crop(const cv::Mat& src, int top, int left, int bottom, int right); +MMDEPLOY_API cv::Mat Crop(const cv::Mat& src, int top, int left, int bottom, int right); /** * @brief Do normalization to an image @@ -50,8 +51,8 @@ cv::Mat Crop(const cv::Mat& src, int top, int left, int bottom, int right); * @param inplace * @return cv::Mat */ -cv::Mat Normalize(cv::Mat& src, const std::vector& mean, const std::vector& std, - bool to_rgb, bool inplace = true); +MMDEPLOY_API cv::Mat Normalize(cv::Mat& src, const std::vector& mean, + const std::vector& std, bool to_rgb, bool inplace = true); /** * @brief tranpose an image, from {h, w, c} to {c, h, w} @@ -59,7 +60,7 @@ cv::Mat Normalize(cv::Mat& src, const std::vector& mean, const std::vecto * @param src input image * @return */ -cv::Mat Transpose(const cv::Mat& src); +MMDEPLOY_API cv::Mat Transpose(const cv::Mat& src); /** * @brief convert an image to another color space @@ -69,7 +70,8 @@ cv::Mat Transpose(const cv::Mat& src); * @param dst_format * @return */ -cv::Mat ColorTransfer(const cv::Mat& src, PixelFormat src_format, PixelFormat dst_format); +MMDEPLOY_API cv::Mat ColorTransfer(const cv::Mat& src, PixelFormat src_format, + PixelFormat dst_format); /** * @@ -82,8 +84,8 @@ cv::Mat ColorTransfer(const cv::Mat& src, PixelFormat src_format, PixelFormat ds * @param val * @return */ -cv::Mat Pad(const cv::Mat& src, int top, int left, int bottom, int right, int border_type, - float val); +MMDEPLOY_API cv::Mat Pad(const cv::Mat& src, int top, int left, int bottom, int right, + int border_type, float val); /** * @brief compare two images @@ -92,7 +94,7 @@ cv::Mat Pad(const cv::Mat& src, int top, int left, int bottom, int right, int bo * @param src2 the other input image * @return bool true means the images are the same */ -bool Compare(const cv::Mat& src1, const cv::Mat& src2); +MMDEPLOY_API bool Compare(const cv::Mat& src1, const cv::Mat& src2); } // namespace cpu @@ -104,8 +106,6 @@ struct IsCvPoint : std::false_type {}; template struct IsCvPoint<::cv::Point_> : std::true_type {}; -} // namespace detail - template >::value, int> = 0> void serialize(Archive&& archive, T&& p) { @@ -146,6 +146,8 @@ void load(Archive& archive, std::vector& v) { } } +} // namespace detail + } // namespace mmdeploy -#endif // MMDEPLOY_OPENCV_UTILS_H +#endif // MMDEPLOY_CSRC_UTILS_OPENCV_OPENCV_UTILS_H_ diff --git a/demo/csrc/CMakeLists.txt b/demo/csrc/CMakeLists.txt index 3da3b04b79..3e1bdcc6fb 100644 --- a/demo/csrc/CMakeLists.txt +++ b/demo/csrc/CMakeLists.txt @@ -2,13 +2,18 @@ cmake_minimum_required(VERSION 3.14) project(mmdeploy-example) -find_package(OpenCV REQUIRED) find_package(MMDeploy REQUIRED) function(add_example name) - add_executable(${name} ${name}.cpp) - target_link_libraries(${name} ${MMDeploy_LIBS} -Wl,--disable-new-dtags - opencv_imgcodecs opencv_imgproc opencv_core) + file(GLOB _SRCS ${name}.c*) + add_executable(${name} ${_SRCS}) + if (NOT MSVC) + # disable new dtags so that executables can run even without LD_LIBRARY_PATH set + target_link_libraries(${name} PRIVATE -Wl,--disable-new-dtags) + endif () + mmdeploy_load_static(${name} MMDeployStaticModules) + mmdeploy_load_dynamic(${name} MMDeployDynamicModules) + target_link_libraries(${name} PRIVATE MMDeployLibs ${OpenCV_LIBS}) endfunction() add_example(image_classification) diff --git a/demo/csrc/image_classification.cpp b/demo/csrc/image_classification.cpp index be618659be..18d1e0793e 100644 --- a/demo/csrc/image_classification.cpp +++ b/demo/csrc/image_classification.cpp @@ -1,5 +1,5 @@ #include -#include +#include #include #include "classifier.h" diff --git a/demo/csrc/image_restorer.cpp b/demo/csrc/image_restorer.cpp index 4e462ce083..3984c88a8a 100644 --- a/demo/csrc/image_restorer.cpp +++ b/demo/csrc/image_restorer.cpp @@ -1,7 +1,7 @@ // Copyright (c) OpenMMLab. All rights reserved. #include -#include +#include #include #include diff --git a/demo/csrc/image_segmentation.cpp b/demo/csrc/image_segmentation.cpp index 71b3108152..8502ecec0c 100644 --- a/demo/csrc/image_segmentation.cpp +++ b/demo/csrc/image_segmentation.cpp @@ -1,7 +1,7 @@ // Copyright (c) OpenMMLab. All rights reserved. #include -#include +#include #include #include #include @@ -13,7 +13,7 @@ using namespace std; vector gen_palette(int num_classes) { std::mt19937 gen; - std::uniform_int_distribution uniform_dist(0, 255); + std::uniform_int_distribution uniform_dist(0, 255); vector palette; palette.reserve(num_classes); diff --git a/demo/csrc/object_detection.cpp b/demo/csrc/object_detection.cpp index a57b4f41f6..1843407532 100644 --- a/demo/csrc/object_detection.cpp +++ b/demo/csrc/object_detection.cpp @@ -1,5 +1,5 @@ #include -#include +#include #include #include diff --git a/demo/csrc/ocr.cpp b/demo/csrc/ocr.cpp index f82d1eca67..1bb8d43ef2 100644 --- a/demo/csrc/ocr.cpp +++ b/demo/csrc/ocr.cpp @@ -1,5 +1,5 @@ #include -#include +#include #include #include diff --git a/demo/demo_rewrite.py b/demo/demo_rewrite.py index 5adc581f30..a624c26eba 100644 --- a/demo/demo_rewrite.py +++ b/demo/demo_rewrite.py @@ -1,3 +1,4 @@ +# Copyright (c) OpenMMLab. All rights reserved. import asyncio import os import shutil diff --git a/docs/en/backends/openvino.md b/docs/en/backends/openvino.md index d043c578ac..12a6686d36 100644 --- a/docs/en/backends/openvino.md +++ b/docs/en/backends/openvino.md @@ -27,7 +27,7 @@ sudo apt-get install libpython3.7 Example: ```bash python tools/deploy.py \ - configs/mmdet/detection/detection_openvino_dynamic.py \ + configs/mmdet/detection/detection_openvino_static-300x300.py \ /mmdetection_dir/mmdetection/configs/ssd/ssd300_coco.py \ /tmp/snapshots/ssd300_coco_20210803_015428-d231a06e.pth \ tests/data/tiger.jpeg \ diff --git a/docs/en/backends/torchscript.md b/docs/en/backends/torchscript.md new file mode 100644 index 0000000000..30449444a1 --- /dev/null +++ b/docs/en/backends/torchscript.md @@ -0,0 +1,54 @@ +## TorchScript support + +### Introduction of TorchScript + +**TorchScript** a way to create serializable and optimizable models from PyTorch code. Any TorchScript program can be saved from a Python process and loaded in a process where there is no Python dependency. Check the [Introduction to TorchScript](https://pytorch.org/tutorials/beginner/Intro_to_TorchScript_tutorial.html) for more details. + +### Build custom ops + +#### Prerequisite + +- Download libtorch from the official website [here](https://pytorch.org/get-started/locally/). + +*Please note that only **Pre-cxx11 ABI** and **version 1.8.1+** on Linux platform are supported by now.* + +For previous versions of libtorch, users can find through the [issue comment](https://github.com/pytorch/pytorch/issues/40961#issuecomment-1017317786). Libtorch1.8.1+cu111 as an example, extract it, expose `Torch_DIR` and add the lib path to `LD_LIBRARY_PATH` as below: + +```bash +wget https://download.pytorch.org/libtorch/cu111/libtorch-shared-with-deps-1.8.1%2Bcu111.zip + +unzip libtorch-shared-with-deps-1.8.1+cu111.zip +cd libtorch +export Torch_DIR=$(pwd) +export LD_LIBRARY_PATH=$Torch_DIR/lib:$LD_LIBRARY_PATH +``` + +Note: + +- If you want to save libtorch env variables to bashrc, you could run + + ```bash + echo '# set env for libtorch' >> ~/.bashrc + echo "export Torch_DIR=${Torch_DIR}" >> ~/.bashrc + echo 'export LD_LIBRARY_PATH=$Torch_DIR/lib:$LD_LIBRARY_PATH' >> ~/.bashrc + source ~/.bashrc + ``` + +#### Build on Linux + +```bash +cd ${MMDEPLOY_DIR} # To MMDeploy root directory +mkdir -p build && cd build +cmake -DMMDEPLOY_TARGET_BACKENDS=torchscript -DTorch_DIR=${Torch_DIR} .. +make -j$(nproc) +``` + +### How to convert a model + +- You could follow the instructions of tutorial [How to convert model](../tutorials/how_to_convert_model.md) + +### FAQs + +- Error: `projects/thirdparty/libtorch/share/cmake/Caffe2/Caffe2Config.cmake:96 (message):Your installed Caffe2 version uses cuDNN but I cannot find the cuDNN libraries. Please set the proper cuDNN prefixes and / or install cuDNN.` + + May export CUDNN_ROOT=/root/path/to/cudnn to resolve the build error. diff --git a/docs/en/benchmark.md b/docs/en/benchmark.md index 31b26e2742..60331d7bf4 100644 --- a/docs/en/benchmark.md +++ b/docs/en/benchmark.md @@ -690,6 +690,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut MMCls PyTorch + TorchScript ONNX Runtime TensorRT PPLNN @@ -702,6 +703,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut Task Metrics fp32 + seresnet fp32 fp32 fp16 @@ -714,6 +716,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut Classification top-1 69.90 + 69.90 69.88 69.88 69.86 @@ -724,6 +727,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut top-5 89.43 + 89.43 89.34 89.34 89.33 @@ -737,6 +741,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut 77.90 77.90 77.90 + 77.90 - 77.78 77.89 @@ -747,6 +752,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut 93.66 93.66 93.66 + 93.66 - 93.64 93.65 @@ -758,6 +764,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut 77.74 77.74 77.74 + 77.74 77.75 77.63 77.73 @@ -768,6 +775,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut 93.84 93.84 93.84 + 93.84 93.83 93.72 93.84 @@ -780,6 +788,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut 68.13 68.13 68.13 + 68.13 67.71 68.11 $MMCLS_DIR/configs/shufflenet_v1/shufflenet_v1_1x_b64x16_linearlr_bn_nowd_imagenet.py @@ -790,6 +799,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut 87.81 87.81 87.81 + 87.81 87.58 87.80 @@ -800,6 +810,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut 69.55 69.55 69.55 + 69.55 69.54 69.10 69.54 @@ -810,6 +821,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut 88.92 88.92 88.92 + 88.92 88.91 88.58 88.92 @@ -821,6 +833,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut 71.86 71.86 71.86 + 71.86 71.87 70.91 71.84 @@ -831,6 +844,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut 90.42 90.42 90.42 + 90.42 90.40 89.85 90.41 @@ -848,6 +862,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut MMDet Pytorch + TorchScript ONNXRuntime TensorRT PPLNN @@ -864,6 +879,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut fp32 fp32 fp32 + fp32 fp16 int8 fp16 @@ -876,6 +892,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut COCO2017 box AP 33.7 + 33.7 - 33.5 33.5 @@ -890,6 +907,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut COCO2017 box AP 25.5 + 25.5 - 25.5 25.5 @@ -904,6 +922,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut COCO2017 box AP 36.5 + 36.4 - 36.4 36.4 @@ -919,6 +938,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut box AP 36.6 - + - 36.6 36.5 - @@ -932,6 +952,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut COCO2017 box AP 37.4 + 37.4 - 37.4 37.4 @@ -946,6 +967,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut COCO2017 box AP 40.5 + 40.3 - 40.3 40.3 @@ -960,6 +982,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut COCO2017 box AP 37.4 + 37.3 - 37.3 37.3 @@ -975,6 +998,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut box AP 39.4 - + - 39.4 39.4 - @@ -989,6 +1013,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut box AP 40.4 - + - 40.4 40.4 - @@ -996,12 +1021,27 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut - $MMDET_DIR/configs/cascade_rcnn/cascade_rcnn_r50_caffe_fpn_1x_coco.py + + GFL + Object Detection + COCO2017 + box AP + 40.2 + - + 40.2 + 40.2 + 40.0 + - + - + $MMDET_DIR/configs/gfl/gfl_r50_fpn_1x_coco.py + Mask R-CNN Instance Segmentation COCO2017 box AP 38.2 + 38.1 - 38.1 38.1 @@ -1013,6 +1053,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut mask AP 34.7 + 34.7 - 33.7 33.7 @@ -1033,6 +1074,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut MMEdit Pytorch + TorchScript ONNX Runtime TensorRT PPLNN @@ -1040,7 +1082,6 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut - Model Task @@ -1049,6 +1090,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut fp32 fp32 fp32 + fp32 fp16 int8 fp16 @@ -1060,6 +1102,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut Set5 PSNR 28.4316 + 28.4120 28.4323 28.4323 28.4286 @@ -1070,6 +1113,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut SSIM 0.8099 + 0.8106 0.8097 0.8097 0.8096 @@ -1082,6 +1126,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut Set5 PSNR 28.2700 + 28.2619 28.2592 28.2592 - @@ -1092,6 +1137,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut SSIM 0.7778 + 0.7784 0.7764 0.7774 - @@ -1104,6 +1150,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut Set5 PSNR 30.6428 + 30.6306 30.6444 30.6430 - @@ -1112,8 +1159,9 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut $MMEDIT_DIR/configs/restorers/esrgan/esrgan_psnr_x4c64b23g32_g1_1000k_div2k.py - + SSIM 0.8559 + 0.8565 0.8558 0.8558 - @@ -1126,16 +1174,18 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut Set5 PSNR 27.9499 + 27.9252 27.9408 27.9408 - - 27.9388 - $MMEDIT_DIR/configs/restorers/srresnet_srgan/srgan_x4c64b16_g1_1000k_div2k.pyy + $MMEDIT_DIR/configs/restorers/srresnet_srgan/srgan_x4c64b16_g1_1000k_div2k.py SSIM 0.7846 + 0.7851 0.7839 0.7839 - @@ -1148,6 +1198,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut Set5 PSNR 30.2252 + 30.2069 30.2300 30.2300 - @@ -1156,8 +1207,9 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut $MMEDIT_DIR/configs/restorers/srresnet_srgan/msrresnet_x4c64b16_g1_1000k_div2k.py - + SSIM 0.8491 + 0.8497 0.8488 0.8488 - @@ -1170,6 +1222,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut Set5 PSNR 28.0297 + - 27.7016 27.7016 - @@ -1180,6 +1233,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut SSIM 0.8236 + - 0.8122 0.8122 - @@ -1192,6 +1246,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut Set5 PSNR 30.2223 + 30.2192 30.2214 30.2214 30.2211 @@ -1202,6 +1257,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut SSIM 0.8500 + 0.8507 0.8497 0.8497 0.8497 @@ -1221,6 +1277,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut MMOCR Pytorch + TorchScript ONNXRuntime TensorRT PPLNN @@ -1237,6 +1294,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut fp32 fp32 fp32 + fp32 fp16 int8 fp16 @@ -1249,6 +1307,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut ICDAR2015 recall 0.7310 + 0.7308 0.7304 0.7198 0.7179 @@ -1261,6 +1320,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut precision 0.8714 0.8718 + 0.8714 0.8677 0.8674 0.8688 @@ -1271,6 +1331,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut hmean 0.7950 0.7949 + 0.7950 0.7868 0.7856 0.7821 @@ -1285,6 +1346,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut 0.8067 0.8067 0.8067 + 0.8067 0.8063 0.8067 0.8067 @@ -1297,6 +1359,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut IIIT5K acc 0.9517 + - 0.9287 - - @@ -1318,6 +1381,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut MMSeg Pytorch + TorchScript ONNXRuntime TensorRT PPLNN @@ -1332,6 +1396,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut fp32 fp32 fp32 + fp32 fp16 int8 fp16 @@ -1342,6 +1407,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut Cityscapes mIoU 72.25 + 72.36 - 72.36 72.35 @@ -1354,6 +1420,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut Cityscapes mIoU 78.55 + 78.66 - 78.26 78.24 @@ -1366,6 +1433,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut Cityscapes mIoU 79.09 + 79.12 - 79.12 79.12 @@ -1378,6 +1446,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut Cityscapes mIoU 79.61 + 79.60 - 79.60 79.60 @@ -1390,6 +1459,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut Cityscapes mIoU 70.96 + 70.96 - 70.93 70.92 @@ -1403,6 +1473,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](tut mIoU 69.10 - + - 69.10 69.10 68.95 diff --git a/docs/en/build.md b/docs/en/build.md index 005950d1ab..075c5204e7 100644 --- a/docs/en/build.md +++ b/docs/en/build.md @@ -89,6 +89,7 @@ Build the inference engine extension libraries you need. - [ncnn](backends/ncnn.md) - [pplnn](backends/pplnn.md) - [OpenVINO](backends/openvino.md) +- [TorchScript](backends/torchscript.md) ### Install mmdeploy diff --git a/docs/en/build/linux.md b/docs/en/build/linux.md new file mode 100644 index 0000000000..1333ed77b7 --- /dev/null +++ b/docs/en/build/linux.md @@ -0,0 +1 @@ +TODO diff --git a/docs/en/build/windows.md b/docs/en/build/windows.md new file mode 100644 index 0000000000..1333ed77b7 --- /dev/null +++ b/docs/en/build/windows.md @@ -0,0 +1 @@ +TODO diff --git a/docs/en/codebases/mmdet.md b/docs/en/codebases/mmdet.md index e5b4f5409d..f03bf7c60f 100644 --- a/docs/en/codebases/mmdet.md +++ b/docs/en/codebases/mmdet.md @@ -22,6 +22,7 @@ Please refer to [get_started.md](https://github.com/open-mmlab/mmdetection/blob/ | Cascade R-CNN | ObjectDetection | Y | Y | N | Y | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/cascade_rcnn) | | Faster R-CNN | ObjectDetection | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/faster_rcnn) | | Faster R-CNN + DCN | ObjectDetection | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/faster_rcnn) | +| GFL | ObjectDetection | Y | Y | N | ? | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/gfl) | | Cascade Mask R-CNN | InstanceSegmentation | Y | N | N | N | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/cascade_rcnn) | | Mask R-CNN | InstanceSegmentation | Y | Y | N | N | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/mask_rcnn) | diff --git a/docs/en/get_started.md b/docs/en/get_started.md index 9336a06c67..a151c2bb01 100644 --- a/docs/en/get_started.md +++ b/docs/en/get_started.md @@ -49,7 +49,7 @@ Now you can do model inference with the APIs provided by the backend. But what i ```python from mmdeploy.apis import inference_model -result = inference_model(model_cfg, deploy_cfg, backend_models, img=img, device=device) +result = inference_model(model_cfg, deploy_cfg, backend_files, img=img, device=device) ``` The `inference_model` will create a wrapper module and do the inference for you. The result has the same format as the original OpenMMLab repo. diff --git a/docs/en/index.rst b/docs/en/index.rst index bc7e10d91f..717011adb0 100644 --- a/docs/en/index.rst +++ b/docs/en/index.rst @@ -46,6 +46,7 @@ You can switch between Chinese and English documents in the lower-left corner of backends/openvino.md backends/ncnn.md backends/pplnn.md + backends/torchscript.md .. toctree:: :maxdepth: 1 diff --git a/docs/en/supported_models.md b/docs/en/supported_models.md index 32a44b1da8..fa7cf4f4ea 100644 --- a/docs/en/supported_models.md +++ b/docs/en/supported_models.md @@ -2,64 +2,65 @@ The table below lists the models that are guaranteed to be exportable to other backends. -| Model | Codebase | OnnxRuntime | TensorRT | NCNN | PPLNN | OpenVINO | Model config | -|:--------------------------|:-----------------|:-----------:|:--------:|:----:|:-----:|:--------:|:----------------------------------------------------------------------------------------------:| -| RetinaNet | MMDetection | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/retinanet) | -| Faster R-CNN | MMDetection | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/faster_rcnn) | -| YOLOv3 | MMDetection | Y | Y | Y | N | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/yolo) | -| YOLOX | MMDetection | Y | Y | Y | N | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/yolox) | -| FCOS | MMDetection | Y | Y | Y | N | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/fcos) | -| FSAF | MMDetection | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/fsaf) | -| Mask R-CNN | MMDetection | Y | Y | N | N | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/mask_rcnn) | -| SSD[*](#note) | MMDetection | Y | Y | Y | N | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/ssd) | -| FoveaBox | MMDetection | Y | N | N | N | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/foveabox) | -| ATSS | MMDetection | Y | Y | N | N | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/atss) | -| Cascade R-CNN | MMDetection | Y | Y | N | Y | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/cascade_rcnn) | -| Cascade Mask R-CNN | MMDetection | Y | N | N | N | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/cascade_rcnn) | -| VFNet | MMDetection | N | N | N | N | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/vfnet) | -| ResNet | MMClassification | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmclassification/tree/master/configs/resnet) | -| ResNeXt | MMClassification | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmclassification/tree/master/configs/resnext) | -| SE-ResNet | MMClassification | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmclassification/tree/master/configs/seresnet) | -| MobileNetV2 | MMClassification | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmclassification/tree/master/configs/mobilenet_v2) | -| ShuffleNetV1 | MMClassification | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmclassification/tree/master/configs/shufflenet_v1) | -| ShuffleNetV2 | MMClassification | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmclassification/tree/master/configs/shufflenet_v2) | -| FCN | MMSegmentation | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/fcn) | -| PSPNet[*static](#note) | MMSegmentation | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/pspnet) | -| DeepLabV3 | MMSegmentation | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/deeplabv3) | -| DeepLabV3+ | MMSegmentation | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/deeplabv3plus) | -| Fast-SCNN[*static](#note) | MMSegmentation | Y | Y | N | Y | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/fastscnn) | -| UNet[*static](#note) | MMSegmentation | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/unet) | -| ANN[*](#note) | MMSegmentation | Y | Y | N | N | N | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/ann) | -| APCNet | MMSegmentation | Y | Y | Y | N | N | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/apcnet) | -| BiSeNetV1 | MMSegmentation | Y | Y | Y | N | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/bisenetv1) | -| BiSeNetV2 | MMSegmentation | Y | Y | Y | N | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/bisenetv2) | -| CGNet | MMSegmentation | Y | Y | Y | N | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/cgnet) | -| DMNet | MMSegmentation | Y | N | N | N | N | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/dmnet) | -| DNLNet | MMSegmentation | Y | Y | Y | N | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/dnlnet) | -| EMANet | MMSegmentation | Y | Y | N | N | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/emanet) | -| EncNet | MMSegmentation | Y | Y | N | N | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/encnet) | -| ERFNet | MMSegmentation | Y | Y | Y | N | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/erfnet) | -| FastFCN | MMSegmentation | Y | Y | Y | N | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/fastfcn) | -| GCNet | MMSegmentation | Y | Y | N | N | N | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/gcnet) | -| ICNet[*](#note) | MMSegmentation | Y | Y | N | N | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/icnet) | -| ISANet | MMSegmentation | Y | Y | N | N | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/isanet) | -| NonLocal Net | MMSegmentation | Y | Y | Y | N | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/nonlocal_net) | -| OCRNet | MMSegmentation | Y | Y | Y | N | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/ocrnet) | -| PointRend | MMSegmentation | Y | Y | N | N | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/point_rend) | -| Semantic FPN | MMSegmentation | Y | Y | Y | N | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/sem_fpn) | -| STDC | MMSegmentation | Y | Y | Y | N | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/stdc) | -| UPerNet[*](#note) | MMSegmentation | Y | Y | N | N | N | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/upernet) | -| DANet | MMSegmentation | Y | Y | N | N | N | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/danet) | -| SRCNN | MMEditing | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmediting/tree/master/configs/restorers/srcnn) | -| ESRGAN | MMEditing | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmediting/tree/master/configs/restorers/esrgan) | -| SRGAN | MMEditing | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmediting/tree/master/configs/restorers/srresnet_srgan) | -| SRResNet | MMEditing | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmediting/tree/master/configs/restorers/srresnet_srgan) | -| Real-ESRGAN | MMEditing | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmediting/tree/master/configs/restorers/real_esrgan) | -| EDSR | MMEditing | Y | Y | Y | N | Y | [config](https://github.com/open-mmlab/mmediting/tree/master/configs/restorers/edsr) | -| RDN | MMEditing | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmediting/tree/master/configs/restorers/rdn) | -| DBNet | MMOCR | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmocr/tree/main/configs/textdet/dbnet) | -| CRNN | MMOCR | Y | Y | Y | Y | N | [config](https://github.com/open-mmlab/mmocr/tree/main/configs/textrecog/crnn) | -| SAR | MMOCR | Y | N | N | N | N | [config](https://github.com/open-mmlab/mmocr/tree/main/configs/textrecog/sar) | +| Model | Codebase | TorchScript | OnnxRuntime | TensorRT | NCNN | PPLNN | OpenVINO | Model config | +|---------------------------|------------------|:-----------:|:-----------:|:--------:|:----:|:-----:|:--------:|------------------------------------------------------------------------------------------------| +| RetinaNet | MMDetection | Y | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/retinanet) | +| Faster R-CNN | MMDetection | Y | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/faster_rcnn) | +| YOLOv3 | MMDetection | Y | Y | Y | Y | N | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/yolo) | +| YOLOX | MMDetection | Y | Y | Y | Y | N | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/yolox) | +| FCOS | MMDetection | Y | Y | Y | Y | N | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/fcos) | +| FSAF | MMDetection | Y | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/fsaf) | +| Mask R-CNN | MMDetection | Y | Y | Y | N | N | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/mask_rcnn) | +| SSD[*](#note) | MMDetection | Y | Y | Y | Y | N | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/ssd) | +| FoveaBox | MMDetection | Y | Y | N | N | N | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/foveabox) | +| ATSS | MMDetection | N | Y | Y | N | N | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/atss) | +| GFL | MMDetection | N | Y | Y | N | ? | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/gfl) | +| Cascade R-CNN | MMDetection | N | Y | Y | N | Y | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/cascade_rcnn) | +| Cascade Mask R-CNN | MMDetection | N | Y | N | N | N | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/cascade_rcnn) | +| VFNet | MMDetection | N | N | N | N | N | Y | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/vfnet) | +| ResNet | MMClassification | Y | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmclassification/tree/master/configs/resnet) | +| ResNeXt | MMClassification | Y | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmclassification/tree/master/configs/resnext) | +| SE-ResNet | MMClassification | Y | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmclassification/tree/master/configs/seresnet) | +| MobileNetV2 | MMClassification | Y | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmclassification/tree/master/configs/mobilenet_v2) | +| ShuffleNetV1 | MMClassification | Y | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmclassification/tree/master/configs/shufflenet_v1) | +| ShuffleNetV2 | MMClassification | Y | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmclassification/tree/master/configs/shufflenet_v2) | +| FCN | MMSegmentation | Y | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/fcn) | +| PSPNet[*static](#note) | MMSegmentation | Y | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/pspnet) | +| DeepLabV3 | MMSegmentation | Y | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/deeplabv3) | +| DeepLabV3+ | MMSegmentation | Y | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/deeplabv3plus) | +| Fast-SCNN[*static](#note) | MMSegmentation | Y | Y | Y | N | Y | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/fastscnn) | +| UNet[*static](#note) | MMSegmentation | Y | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/unet) | +| ANN[*](#note) | MMSegmentation | ? | Y | Y | N | N | N | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/ann) | +| APCNet | MMSegmentation | ? | Y | Y | Y | N | N | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/apcnet) | +| BiSeNetV1 | MMSegmentation | ? | Y | Y | Y | N | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/bisenetv1) | +| BiSeNetV2 | MMSegmentation | ? | Y | Y | Y | N | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/bisenetv2) | +| CGNet | MMSegmentation | ? | Y | Y | Y | N | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/cgnet) | +| DMNet | MMSegmentation | ? | Y | N | N | N | N | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/dmnet) | +| DNLNet | MMSegmentation | ? | Y | Y | Y | N | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/dnlnet) | +| EMANet | MMSegmentation | ? | Y | Y | N | N | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/emanet) | +| EncNet | MMSegmentation | ? | Y | Y | N | N | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/encnet) | +| ERFNet | MMSegmentation | ? | Y | Y | Y | N | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/erfnet) | +| FastFCN | MMSegmentation | ? | Y | Y | Y | N | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/fastfcn) | +| GCNet | MMSegmentation | ? | Y | Y | N | N | N | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/gcnet) | +| ICNet[*](#note) | MMSegmentation | ? | Y | Y | N | N | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/icnet) | +| ISANet | MMSegmentation | ? | Y | Y | N | N | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/isanet) | +| NonLocal Net | MMSegmentation | ? | Y | Y | Y | N | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/nonlocal_net) | +| OCRNet | MMSegmentation | ? | Y | Y | Y | N | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/ocrnet) | +| PointRend | MMSegmentation | ? | Y | Y | N | N | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/point_rend) | +| Semantic FPN | MMSegmentation | ? | Y | Y | Y | N | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/sem_fpn) | +| STDC | MMSegmentation | ? | Y | Y | Y | N | Y | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/stdc) | +| UPerNet[*](#note) | MMSegmentation | ? | Y | Y | N | N | N | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/upernet) | +| DANet | MMSegmentation | ? | Y | Y | N | N | N | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/danet) | +| SRCNN | MMEditing | Y | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmediting/tree/master/configs/restorers/srcnn) | +| ESRGAN | MMEditing | Y | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmediting/tree/master/configs/restorers/esrgan) | +| SRGAN | MMEditing | Y | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmediting/tree/master/configs/restorers/srresnet_srgan) | +| SRResNet | MMEditing | Y | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmediting/tree/master/configs/restorers/srresnet_srgan) | +| Real-ESRGAN | MMEditing | Y | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmediting/tree/master/configs/restorers/real_esrgan) | +| EDSR | MMEditing | Y | Y | Y | Y | N | Y | [config](https://github.com/open-mmlab/mmediting/tree/master/configs/restorers/edsr) | +| RDN | MMEditing | Y | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmediting/tree/master/configs/restorers/rdn) | +| DBNet | MMOCR | Y | Y | Y | Y | Y | Y | [config](https://github.com/open-mmlab/mmocr/tree/main/configs/textdet/dbnet) | +| CRNN | MMOCR | Y | Y | Y | Y | Y | N | [config](https://github.com/open-mmlab/mmocr/tree/main/configs/textrecog/crnn) | +| SAR | MMOCR | N | Y | N | N | N | N | [config](https://github.com/open-mmlab/mmocr/tree/main/configs/textrecog/sar) | ### Note diff --git a/docs/en/tutorials/how_to_support_new_backends.md b/docs/en/tutorials/how_to_support_new_backends.md index 85319b8d91..c18cd86148 100644 --- a/docs/en/tutorials/how_to_support_new_backends.md +++ b/docs/en/tutorials/how_to_support_new_backends.md @@ -218,6 +218,7 @@ Although the backend engines are usually implemented in C/C++, it is convenient def _build_wrapper(backend: Backend, backend_files: Sequence[str], device: str, + input_names: Optional[Sequence[str]] = None, output_names: Optional[Sequence[str]] = None): if backend == Backend.ONNXRUNTIME: from mmdeploy.backend.onnxruntime import ORTWrapper diff --git a/docs/zh_cn/benchmark.md b/docs/zh_cn/benchmark.md index 6ba3ababc0..52afc8d1b8 100644 --- a/docs/zh_cn/benchmark.md +++ b/docs/zh_cn/benchmark.md @@ -682,6 +682,7 @@ GPU: ncnn, TensorRT, PPLNN 用户可以直接通过[如何测试性能](tutorials/how_to_evaluate_a_model.md)获得想要的性能测试结果。下面是我们环境中的测试结果： +
MMCls
@@ -691,6 +692,7 @@ GPU: ncnn, TensorRT, PPLNN MMCls PyTorch + TorchScript ONNX Runtime TensorRT PPLNN @@ -703,6 +705,7 @@ GPU: ncnn, TensorRT, PPLNN Task Metrics fp32 + seresnet fp32 fp32 fp16 @@ -715,6 +718,7 @@ GPU: ncnn, TensorRT, PPLNN Classification top-1 69.90 + 69.90 69.88 69.88 69.86 @@ -725,6 +729,7 @@ GPU: ncnn, TensorRT, PPLNN top-5 89.43 + 89.43 89.34 89.34 89.33 @@ -738,6 +743,7 @@ GPU: ncnn, TensorRT, PPLNN 77.90 77.90 77.90 + 77.90 - 77.78 77.89 @@ -748,6 +754,7 @@ GPU: ncnn, TensorRT, PPLNN 93.66 93.66 93.66 + 93.66 - 93.64 93.65 @@ -759,6 +766,7 @@ GPU: ncnn, TensorRT, PPLNN 77.74 77.74 77.74 + 77.74 77.75 77.63 77.73 @@ -769,6 +777,7 @@ GPU: ncnn, TensorRT, PPLNN 93.84 93.84 93.84 + 93.84 93.83 93.72 93.84 @@ -781,6 +790,7 @@ GPU: ncnn, TensorRT, PPLNN 68.13 68.13 68.13 + 68.13 67.71 68.11 $MMCLS_DIR/configs/shufflenet_v1/shufflenet_v1_1x_b64x16_linearlr_bn_nowd_imagenet.py @@ -791,6 +801,7 @@ GPU: ncnn, TensorRT, PPLNN 87.81 87.81 87.81 + 87.81 87.58 87.80 @@ -801,6 +812,7 @@ GPU: ncnn, TensorRT, PPLNN 69.55 69.55 69.55 + 69.55 69.54 69.10 69.54 @@ -811,6 +823,7 @@ GPU: ncnn, TensorRT, PPLNN 88.92 88.92 88.92 + 88.92 88.91 88.58 88.92 @@ -822,6 +835,7 @@ GPU: ncnn, TensorRT, PPLNN 71.86 71.86 71.86 + 71.86 71.87 70.91 71.84 @@ -832,6 +846,7 @@ GPU: ncnn, TensorRT, PPLNN 90.42 90.42 90.42 + 90.42 90.40 89.85 90.41 @@ -849,6 +864,7 @@ GPU: ncnn, TensorRT, PPLNN MMDet Pytorch + TorchScript ONNXRuntime TensorRT PPLNN @@ -865,6 +881,7 @@ GPU: ncnn, TensorRT, PPLNN fp32 fp32 fp32 + fp32 fp16 int8 fp16 @@ -877,6 +894,7 @@ GPU: ncnn, TensorRT, PPLNN COCO2017 box AP 33.7 + 33.7 - 33.5 33.5 @@ -891,6 +909,7 @@ GPU: ncnn, TensorRT, PPLNN COCO2017 box AP 25.5 + 25.5 - 25.5 25.5 @@ -905,6 +924,7 @@ GPU: ncnn, TensorRT, PPLNN COCO2017 box AP 36.5 + 36.4 - 36.4 36.4 @@ -920,6 +940,7 @@ GPU: ncnn, TensorRT, PPLNN box AP 36.6 - + - 36.6 36.5 - @@ -933,6 +954,7 @@ GPU: ncnn, TensorRT, PPLNN COCO2017 box AP 37.4 + 37.4 - 37.4 37.4 @@ -947,6 +969,7 @@ GPU: ncnn, TensorRT, PPLNN COCO2017 box AP 40.5 + 40.3 - 40.3 40.3 @@ -961,6 +984,7 @@ GPU: ncnn, TensorRT, PPLNN COCO2017 box AP 37.4 + 37.3 - 37.3 37.3 @@ -976,6 +1000,7 @@ GPU: ncnn, TensorRT, PPLNN box AP 39.4 - + - 39.4 39.4 - @@ -990,6 +1015,7 @@ GPU: ncnn, TensorRT, PPLNN box AP 40.4 - + - 40.4 40.4 - @@ -1003,6 +1029,7 @@ GPU: ncnn, TensorRT, PPLNN COCO2017 box AP 38.2 + 38.1 - 38.1 38.1 @@ -1014,6 +1041,7 @@ GPU: ncnn, TensorRT, PPLNN mask AP 34.7 + 34.7 - 33.7 33.7 @@ -1034,6 +1062,7 @@ GPU: ncnn, TensorRT, PPLNN MMEdit Pytorch + TorchScript ONNX Runtime TensorRT PPLNN @@ -1041,7 +1070,6 @@ GPU: ncnn, TensorRT, PPLNN - Model Task @@ -1050,6 +1078,7 @@ GPU: ncnn, TensorRT, PPLNN fp32 fp32 fp32 + fp32 fp16 int8 fp16 @@ -1061,6 +1090,7 @@ GPU: ncnn, TensorRT, PPLNN Set5 PSNR 28.4316 + 28.4120 28.4323 28.4323 28.4286 @@ -1071,6 +1101,7 @@ GPU: ncnn, TensorRT, PPLNN SSIM 0.8099 + 0.8106 0.8097 0.8097 0.8096 @@ -1083,6 +1114,7 @@ GPU: ncnn, TensorRT, PPLNN Set5 PSNR 28.2700 + 28.2619 28.2592 28.2592 - @@ -1093,6 +1125,7 @@ GPU: ncnn, TensorRT, PPLNN SSIM 0.7778 + 0.7784 0.7764 0.7774 - @@ -1105,6 +1138,7 @@ GPU: ncnn, TensorRT, PPLNN Set5 PSNR 30.6428 + 30.6306 30.6444 30.6430 - @@ -1113,8 +1147,9 @@ GPU: ncnn, TensorRT, PPLNN $MMEDIT_DIR/configs/restorers/esrgan/esrgan_psnr_x4c64b23g32_g1_1000k_div2k.py - + SSIM 0.8559 + 0.8565 0.8558 0.8558 - @@ -1127,16 +1162,18 @@ GPU: ncnn, TensorRT, PPLNN Set5 PSNR 27.9499 + 27.9252 27.9408 27.9408 - - 27.9388 - $MMEDIT_DIR/configs/restorers/srresnet_srgan/srgan_x4c64b16_g1_1000k_div2k.pyy + $MMEDIT_DIR/configs/restorers/srresnet_srgan/srgan_x4c64b16_g1_1000k_div2k.py SSIM 0.7846 + 0.7851 0.7839 0.7839 - @@ -1149,6 +1186,7 @@ GPU: ncnn, TensorRT, PPLNN Set5 PSNR 30.2252 + 30.2069 30.2300 30.2300 - @@ -1157,8 +1195,9 @@ GPU: ncnn, TensorRT, PPLNN $MMEDIT_DIR/configs/restorers/srresnet_srgan/msrresnet_x4c64b16_g1_1000k_div2k.py - + SSIM 0.8491 + 0.8497 0.8488 0.8488 - @@ -1171,6 +1210,7 @@ GPU: ncnn, TensorRT, PPLNN Set5 PSNR 28.0297 + - 27.7016 27.7016 - @@ -1181,6 +1221,7 @@ GPU: ncnn, TensorRT, PPLNN SSIM 0.8236 + - 0.8122 0.8122 - @@ -1193,6 +1234,7 @@ GPU: ncnn, TensorRT, PPLNN Set5 PSNR 30.2223 + 30.2192 30.2214 30.2214 30.2211 @@ -1203,6 +1245,7 @@ GPU: ncnn, TensorRT, PPLNN SSIM 0.8500 + 0.8507 0.8497 0.8497 0.8497 @@ -1222,6 +1265,7 @@ GPU: ncnn, TensorRT, PPLNN MMOCR Pytorch + TorchScript ONNXRuntime TensorRT PPLNN @@ -1238,6 +1282,7 @@ GPU: ncnn, TensorRT, PPLNN fp32 fp32 fp32 + fp32 fp16 int8 fp16 @@ -1250,6 +1295,7 @@ GPU: ncnn, TensorRT, PPLNN ICDAR2015 recall 0.7310 + 0.7308 0.7304 0.7198 0.7179 @@ -1262,6 +1308,7 @@ GPU: ncnn, TensorRT, PPLNN precision 0.8714 0.8718 + 0.8714 0.8677 0.8674 0.8688 @@ -1272,6 +1319,7 @@ GPU: ncnn, TensorRT, PPLNN hmean 0.7950 0.7949 + 0.7950 0.7868 0.7856 0.7821 @@ -1286,6 +1334,7 @@ GPU: ncnn, TensorRT, PPLNN 0.8067 0.8067 0.8067 + 0.8067 0.8063 0.8067 0.8067 @@ -1298,6 +1347,7 @@ GPU: ncnn, TensorRT, PPLNN IIIT5K acc 0.9517 + - 0.9287 - - @@ -1319,6 +1369,7 @@ GPU: ncnn, TensorRT, PPLNN MMSeg Pytorch + TorchScript ONNXRuntime TensorRT PPLNN @@ -1333,6 +1384,7 @@ GPU: ncnn, TensorRT, PPLNN fp32 fp32 fp32 + fp32 fp16 int8 fp16 @@ -1343,6 +1395,7 @@ GPU: ncnn, TensorRT, PPLNN Cityscapes mIoU 72.25 + 72.36 - 72.36 72.35 @@ -1355,6 +1408,7 @@ GPU: ncnn, TensorRT, PPLNN Cityscapes mIoU 78.55 + 78.66 - 78.26 78.24 @@ -1367,6 +1421,7 @@ GPU: ncnn, TensorRT, PPLNN Cityscapes mIoU 79.09 + 79.12 - 79.12 79.12 @@ -1379,6 +1434,7 @@ GPU: ncnn, TensorRT, PPLNN Cityscapes mIoU 79.61 + 79.60 - 79.60 79.60 @@ -1391,6 +1447,7 @@ GPU: ncnn, TensorRT, PPLNN Cityscapes mIoU 70.96 + 70.96 - 70.93 70.92 @@ -1404,6 +1461,7 @@ GPU: ncnn, TensorRT, PPLNN mIoU 69.10 - + - 69.10 69.10 68.95 diff --git a/docs/zh_cn/build.md b/docs/zh_cn/build.md index 3e0717ddc8..8093cbb0b0 100644 --- a/docs/zh_cn/build.md +++ b/docs/zh_cn/build.md @@ -87,6 +87,7 @@ - [ncnn](https://mmdeploy.readthedocs.io/en/latest/backends/ncnn.html) - [PPLNN](https://mmdeploy.readthedocs.io/en/latest/backends/pplnn.html) - [OpenVINO](https://mmdeploy.readthedocs.io/en/latest/backends/openvino.html) +- [TorchScript](https://mmdeploy.readthedocs.io/en/latest/backends/torchscript.md) ### 安装 MMDeploy diff --git a/docs/zh_cn/build/linux.md b/docs/zh_cn/build/linux.md new file mode 100644 index 0000000000..1333ed77b7 --- /dev/null +++ b/docs/zh_cn/build/linux.md @@ -0,0 +1 @@ +TODO diff --git a/docs/zh_cn/build/windows.md b/docs/zh_cn/build/windows.md new file mode 100644 index 0000000000..e0bcdfbc37 --- /dev/null +++ b/docs/zh_cn/build/windows.md @@ -0,0 +1,336 @@ +- [Windows 下构建方式](#windows-下构建方式) + - [源码安装](#源码安装) + - [安装构建和编译工具链](#安装构建和编译工具链) + - [安装依赖包](#安装依赖包) + - [安装 MMDeploy Converter 依赖](#安装-mmdeploy-converter-依赖) + - [安装 MMDeploy SDK 依赖](#安装-mmdeploy-sdk-依赖) + - [安装推理引擎](#安装推理引擎) + - [编译 MMDeploy](#编译-mmdeploy) + - [编译安装 Model Converter](#编译安装-model-converter) + - [编译自定义算子](#编译自定义算子) + - [安装 Model Converter](#安装-model-converter) + - [编译 SDK](#编译-sdk) + - [编译选项说明](#编译选项说明) + - [编译样例](#编译样例) + - [编译 SDK Demo](#编译-sdk-demo) + - [注意事项](#注意事项) + +--- +# Windows 下构建方式 + +目前，MMDeploy 在 Windows 平台下仅提供源码编译安装方式。未来会提供预编译包方式。 + +## 源码安装 +下述安装方式，均是在 **Windows 10** 下进行 +### 安装构建和编译工具链 +1. 下载并安装 [Visual Studio 2019](https://visualstudio.microsoft.com) 。安装时请勾选 "使用C++的桌面开发, "Windows 10 SDK
+2. 把 cmake 路径加入到环境变量 PATH 中, "C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\Common7\IDE\CommonExtensions\Microsoft\CMake\CMake\bin"
+3. 如果系统中配置了 NVIDIA 显卡，根据官网教程，下载并安装 cuda toolkit。
+### 安装依赖包 + +#### 安装 MMDeploy Converter 依赖 + + + + + + + + + + + + + + + + + + + + + +
名称 安装方法
conda 强烈建议安装conda，或者miniconda。比如，
https://repo.anaconda.com/miniconda/Miniconda3-py37_4.11.0-Windows-x86_64.exe
安装完毕后，打开系统开始菜单，输入prompt，选择并打开 anaconda powershell prompt。
下文中的安装命令均是在 anaconda powershell 中测试验证的。
pytorch
(>=1.8.0) 		+ 参考pytorch官网，根据系统环境, 选择合适的预编译包进行安装。比如,
+ 

+    pip install torch==1.8.1+cu111 torchvision==0.9.1+cu111 torchaudio==0.8.1 -f https://download.pytorch.org/whl/torch_stable.html
+
+
mmcv-full 参考mmcv官网，根据系统环境，选择预编译包进行安装。比如，
+ 

+    $env:cu_version="cu111"
+    $env:torch_version="torch1.8.0"
+    pip install mmcv-full==1.4.0 -f https://download.openmmlab.com/mmcv/dist/$env:cu_version/$env:torch_version/index.html
+
+
+ + +#### 安装 MMDeploy SDK 依赖 + + + + + + + + + + + + + + + + + + + + + + +
名称 安装方法
spdlog spdlog是一个精巧的日志管理库。请参考如下命令安装：
+ 1. 下载 https://github.com/gabime/spdlog/archive/refs/tags/v1.9.2.zip
+ 2. 解压后，进入到文件夹 spdlog-v1.9.2
+ 3. 执行编译安装命令
+ 

+    mkdir build
+    cd build
+    cmake .. -G "Visual Studio 16 2019" -A x64 -T v142 -DCMAKE_POSITION_INDEPENDENT_CODE=ON -DCMAKE_INSTALL_PREFIX=install -DCMAKE_BUILD_TYPE=Release
+    cmake --build . --target install -j --config Release
+
+
OpenCV + 1. 下载并安装 OpenCV 在 windows 下的预编译包: https://github.com/opencv/opencv/releases/download/4.5.5/opencv-4.5.5-vc14_vc15.exe
+ 2. 把 OpenCV 库的路径加入到环境变量 PATH 中
pplcv pplcv 是在x86和cuda平台下的高性能图像处理库。 此依赖项为可选项，只有在cuda平台下，才需安装。而且，目前必须使用v0.6.2，且需要使用git clone的方式下载源码并编译安装
+ 

+    git clone --recursive git@github.com:openppl-public/ppl.cv.git
+    cd ppl.cv
+    git checkout tags/v0.6.2 -b v0.6.2
+    ./build.bat -G "Visual Studio 16 2019" -T v142 -A x64 -DHPCC_USE_CUDA=ON -DHPCC_MSVC_MD=ON
+
+
+ + + +#### 安装推理引擎 +目前，在 Windows 平台下，MMDeploy 支持 ONNXRuntime 和 TensorRT 两种推理引擎。其他推理引擎尚未进行验证，或者验证未通过。后续将陆续予以支持 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
推理引擎 依赖包安装方法
ONNXRuntimeonnxruntime + 1. 下载二进制包：https://github.com/microsoft/onnxruntime/releases/download/v1.8.0/onnxruntime-win-x64-1.8.0.zip
+ 2. 解压到目标路径。我们使用 onnxruntime_dir 代表此路径
+ 3. 在 PATH 中增加 onnxruntime libs 路径, + 

+    $env:path = "{onnxruntime_dir}/lib;" + $env:path
+
+
TensorRT
TensorRT
+ 1. 从NVIDIA官网下载二进制包, 比如，
+ https://developer.nvidia.com/compute/machine-learning/tensorrt/secure/8.2.3.0/zip/TensorRT-8.2.3.0.Windows10.x86_64.cuda-11.4.cudnn8.2.zip
+ 1. 解压二进制包到目标路径。我们使用 tensorrt_dir 代表此路径
+ 2. 安装 tensorrt 的 python package
+ 3. 在 PATH 中增加 tensorrt libs 路径 + 

+   pip install {tensorrt_dir}/python/tensorrt-8.2.3.0-cp37-none-win_amd64.whl
+   $env:path = "{tensorrt_dir}/lib;" + $env:path
+
+
cudnn + 1. 从NVIDIA官网下载二进制包, 比如,
+ https://developer.nvidia.com/compute/machine-learning/cudnn/secure/8.2.1.32/11.3_06072021/cudnn-11.3-windows-x64-v8.2.1.32.zip
+ 1. 解压二进制包到目标路径。我们使用 cudnn_dir 代表此路径
+ 2. 在 PATH 中增加 cudnn libs 路径 + 

+   $env:path = "{cudnn_dir}/bin;" + $env:path
+   
+
PPL.NNppl.nn TODO
OpenVINOopenvino TODO
ncnn ncnn TODO
+ +### 编译 MMDeploy + +#### 编译安装 Model Converter +##### 编译自定义算子 +- **ONNXRuntime** 自定义算子 +```powershell +mkdir build +cd build +cmake .. -G "Visual Studio 16 2019" -A x64 -T v142 -DMMDEPLOY_TARGET_BACKENDS="ort" -DONNXRUNTIME_DIR={onnxruntime_dir} +cmake --build . --config Release -- /maxcpucount:4 +``` + +- **TensorRT** 自定义算子 + +```powershell +mkdir build +cd build +cmake .. -G "Visual Studio 16 2019" -A x64 -T v142 -DMMDEPLOY_TARGET_BACKENDS="trt" -DTENSORRT_DIR={tensorrt_dir} -DCUDNN_DIR={cudnn_dir} +cmake --build . --config Release -- /maxcpucount:4 +``` + +- **ncnn** 自定义算子 + + TODO + +##### 安装 Model Converter +```powershell +cd root/path/of/MMDeploy +pip install -e . +``` +#### 编译 SDK +##### 编译选项说明 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
编译选项取值范围缺省值说明
MMDEPLOY_BUILD_SDK{ON, OFF}OFFMMDeploy SDK 编译开关
MMDEPLOY_BUILD_SDK_PYTHON_API{ON, OFF}OFFMMDeploy SDK python package的编译开关
MMDEPLOY_BUILD_TEST{ON, OFF}OFFMMDeploy SDK的测试程序编译开关
MMDEPLOY_TARGET_DEVICES{"cpu", "cuda"}cpu设置目标设备。当有多个设备时，设备名称之间使用分号隔开。 比如，-DMMDEPLOY_TARGET_DEVICES="cpu;cuda"
MMDEPLOY_TARGET_BACKENDS{"trt", "ort", "pplnn", "ncnn", "openvino"}N/A 默认情况下，SDK不设置任何后端, 因为它与应用场景高度相关。 当选择多个后端时， 中间使用分号隔开。比如，
-DMMDEPLOY_TARGET_BACKENDS="trt;ort;pplnn;ncnn;openvino"
+ 构建时，几乎每个后端，都需设置一些环境变量，用来查找依赖包。
+ 1. trt: 表示 TensorRT, 需要设置 TENSORRT_DIR 和 CUDNN_DIR。类似， 
-DTENSORRT_DIR={tensorrt_dir}
-DCUDNN_DIR={cudnn_dir}
+ 2. ort: 表示 ONNXRuntime，需要设置 ONNXRUNTIME_DIR。类似， 
-DONNXRUNTIME_DIR={onnxruntime_dir}
+ 3. pplnn: 表示 PPL.NN，需要设置 pplnn_DIR。当前版本尚未验证
+ 4. ncnn：需要设置 ncnn_DIR。当前版本尚未验证
+ 5. openvino: 表示 OpenVINO，需要设置 InferenceEngine_DIR。当前版本尚未验证通过 +
MMDEPLOY_CODEBASES{"mmcls", "mmdet", "mmseg", "mmedit", "mmocr", "all"}N/A用来设置SDK后处理组件，加载OpenMMLab算法仓库的后处理功能。已支持的算法仓库有'mmcls'，'mmdet'，'mmedit'，'mmseg'和'mmocr'。如果选择多个codebase，中间使用分号隔开。比如，-DMMDEPLOY_CODEBASES="mmcls;mmdet"。也可以通过 -DMMDEPLOY_CODEBASES=all 方式，加载所有codebase。
BUILD_SHARED_LIBS{ON, OFF}ON动态库的编译开关。设置OFF时，编译静态库
+ + +##### 编译样例 + +下文展示2个构建SDK的样例，分别用于不同的运行环境。 + +- cpu + ONNXRuntime + + ```PowerShell + mkdir build + cd build + cmake .. -G "Visual Studio 16 2019" -A x64 -T v142 ` + -DMMDEPLOY_BUILD_SDK=ON ` + -DMMDEPLOY_TARGET_DEVICES="cpu" ` + -DMMDEPLOY_TARGET_BACKENDS="ort" ` + -DMMDEPLOY_CODEBASES="all" ` + -DONNXRUNTIME_DIR={onnxruntime_dir} ` + -Dspdlog_DIR={spdlog_dir}/build/install/lib/cmake/spdlog ` + -DOpenCV_DIR={opencv_dir}/build + cmake --build . --config Release -- /maxcpucount:4 + cmake --install . --config Release + ``` + +- cuda + TensorRT + + ```PowerShell + mkdir build + cd build + cmake .. -G "Visual Studio 16 2019" -A x64 -T v142 ` + -DMMDEPLOY_BUILD_SDK=ON ` + -DMMDEPLOY_TARGET_DEVICES="cuda" ` + -DMMDEPLOY_TARGET_BACKENDS="trt" ` + -DMMDEPLOY_CODEBASES="all" ` + -Dpplcv_DIR={pplcv_dir}/pplcv-build/install/lib/cmake/ppl ` + -DTENSORRT_DIR={tensorrt_dir} ` + -DCUDNN_DIR={cudnn_dir} ` + -Dspdlog_DIR={spdlog_dir}/build/install/lib/cmake/spdlog ` + -DOpenCV_DIR={opencv_dir}/build + cmake --build . --config Release -- /maxcpucount:4 + cmake --install . --config Release + ``` +- 其他 + + 请参考上述两个示例，以及前述 SDK 的编译选项，在其他运行环境下编译 SDK + +##### 编译 SDK Demo + +```PowerShell +cd install/example +mkdir build +cd build +cmake .. -G "Visual Studio 16 2019" -A x64 -T v142 ` + -DMMDeploy_DIR={mmdeploy_dir}/build/install/lib/cmake/MMDeploy ` + -Dspdlog_DIR={spdlog_dir}/build/install/lib/cmake/spdlog ` + -DOpenCV_DIR={opencv_dir}/build +cmake --build . --config Release -- /maxcpucount:4 + +$env:path = "${mmdeploy_dir}/build/install/bin;" + $env:path + +``` + +### 注意事项 + 1. Release / Debug 库不能混用。MMDeploy要是编译Debug版本，所有第三方依赖都要是Debug版本。 diff --git a/docs/zh_cn/get_started.md b/docs/zh_cn/get_started.md index 414bc481d3..f817128296 100644 --- a/docs/zh_cn/get_started.md +++ b/docs/zh_cn/get_started.md @@ -49,7 +49,7 @@ python ${MMDEPLOY_DIR}/tools/deploy.py \ ```python from mmdeploy.apis import inference_model -result = inference_model(model_cfg, deploy_cfg, backend_models, img=img, device=device) +result = inference_model(model_cfg, deploy_cfg, backend_files, img=img, device=device) ``` `inference_model`会创建一个对后端模型的封装，通过该封装进行推理。推理的结果会保持与OpenMMLab中原模型同样的格式。 diff --git a/mmdeploy/apis/__init__.py b/mmdeploy/apis/__init__.py index 70e6e479d5..48b1339d15 100644 --- a/mmdeploy/apis/__init__.py +++ b/mmdeploy/apis/__init__.py @@ -3,11 +3,12 @@ from .extract_model import extract_model from .inference import inference_model from .pytorch2onnx import torch2onnx, torch2onnx_impl +from .pytorch2torchscript import torch2torchscript, torch2torchscript_impl from .utils import build_task_processor, get_predefined_partition_cfg from .visualize import visualize_model __all__ = [ 'create_calib_table', 'extract_model', 'inference_model', 'torch2onnx', - 'torch2onnx_impl', 'build_task_processor', 'get_predefined_partition_cfg', - 'visualize_model' + 'torch2onnx_impl', 'torch2torchscript', 'torch2torchscript_impl', + 'build_task_processor', 'get_predefined_partition_cfg', 'visualize_model' ] diff --git a/mmdeploy/apis/pytorch2torchscript.py b/mmdeploy/apis/pytorch2torchscript.py new file mode 100644 index 0000000000..c984892360 --- /dev/null +++ b/mmdeploy/apis/pytorch2torchscript.py @@ -0,0 +1,111 @@ +# Copyright (c) OpenMMLab. All rights reserved. +import os.path as osp +from typing import Any, Optional, Sequence, Union + +import mmcv +import torch +from packaging.version import parse as version_parse + +from mmdeploy.backend.torchscript import get_ops_path +from mmdeploy.core import RewriterContext, patch_model +from mmdeploy.utils import (IR, get_backend, get_input_shape, get_root_logger, + load_config) + + +def torch2torchscript_impl(model: torch.nn.Module, + inputs: Union[torch.Tensor, Sequence[torch.Tensor]], + deploy_cfg: Union[str, + mmcv.Config], output_file: str): + """Converting torch model to torchscript. + + Args: + model (torch.nn.Module): Input pytorch model. + inputs (torch.Tensor | Sequence[torch.Tensor]): Input tensors used to + convert model. + deploy_cfg (str | mmcv.Config): Deployment config file or + Config object. + output_file (str): Output file to save torchscript model. + """ + # load custom ops if exist + custom_ops_path = get_ops_path() + if osp.exists(custom_ops_path): + torch.ops.load_library(custom_ops_path) + + deploy_cfg = load_config(deploy_cfg)[0] + + backend = get_backend(deploy_cfg).value + + patched_model = patch_model(model, cfg=deploy_cfg, backend=backend) + + with RewriterContext( + cfg=deploy_cfg, backend=backend, + ir=IR.TORCHSCRIPT), torch.no_grad(), torch.jit.optimized_execution( + True): + # for exporting models with weight that depends on inputs + patched_model( + *inputs) if isinstance(inputs, Sequence) else patched_model(inputs) + ts_model = torch.jit.trace(patched_model, inputs) + + # perform optimize, note that optimizing models may trigger errors when + # loading the saved .pt file, as described in + # https://github.com/pytorch/pytorch/issues/62706 + logger = get_root_logger() + logger.info('perform torchscript optimizer.') + try: + # custom optimizer + from mmdeploy.backend.torchscript import ts_optimizer + logger = get_root_logger() + ts_optimizer.optimize_for_backend( + ts_model._c, ir=IR.TORCHSCRIPT.value, backend=backend) + except Exception: + # use pytorch builtin optimizer + ts_model = torch.jit.freeze(ts_model) + torch_version = version_parse(torch.__version__) + if torch_version.minor >= 9: + ts_model = torch.jit.optimize_for_inference(ts_model) + + # save model + torch.jit.save(ts_model, output_file) + + +def torch2torchscript(img: Any, + work_dir: str, + save_file: str, + deploy_cfg: Union[str, mmcv.Config], + model_cfg: Union[str, mmcv.Config], + model_checkpoint: Optional[str] = None, + device: str = 'cuda:0'): + """Convert PyTorch model to torchscript model. + + Args: + img (str | np.ndarray | torch.Tensor): Input image used to assist + converting model. + work_dir (str): A working directory to save files. + save_file (str): Filename to save torchscript model. + deploy_cfg (str | mmcv.Config): Deployment config file or + Config object. + model_cfg (str | mmcv.Config): Model config file or Config object. + model_checkpoint (str): A checkpoint path of PyTorch model, + defaults to `None`. + device (str): A string specifying device type, defaults to 'cuda:0'. + """ + # load deploy_cfg if necessary + deploy_cfg, model_cfg = load_config(deploy_cfg, model_cfg) + mmcv.mkdir_or_exist(osp.abspath(work_dir)) + output_file = osp.join(work_dir, save_file) + + input_shape = get_input_shape(deploy_cfg) + + from mmdeploy.apis import build_task_processor + task_processor = build_task_processor(model_cfg, deploy_cfg, device) + + torch_model = task_processor.init_pytorch_model(model_checkpoint) + _, model_inputs = task_processor.create_input(img, input_shape) + if not isinstance(model_inputs, torch.Tensor): + model_inputs = model_inputs[0] + + torch2torchscript_impl( + torch_model, + model_inputs, + deploy_cfg=deploy_cfg, + output_file=output_file) diff --git a/mmdeploy/backend/ncnn/init_plugins.py b/mmdeploy/backend/ncnn/init_plugins.py index 97667defd4..10f9ca2655 100644 --- a/mmdeploy/backend/ncnn/init_plugins.py +++ b/mmdeploy/backend/ncnn/init_plugins.py @@ -1,7 +1,8 @@ # Copyright (c) OpenMMLab. All rights reserved. -import glob import os +from mmdeploy.utils import get_file_path + def get_ops_path() -> str: """Get NCNN custom ops library path. @@ -9,14 +10,11 @@ def get_ops_path() -> str: Returns: str: The library path of NCNN custom ops. """ - wildcard = os.path.abspath( - os.path.join( - os.path.dirname(__file__), - '../../../build/lib/libmmdeploy_ncnn_ops.so')) - - paths = glob.glob(wildcard) - lib_path = paths[0] if len(paths) > 0 else '' - return lib_path + candidates = [ + '../../../build/lib/libmmdeploy_ncnn_ops.so', + '../../../build/bin/*/mmdeploy_ncnn_ops.dll' + ] + return get_file_path(os.path.dirname(__file__), candidates) def get_onnx2ncnn_path() -> str: @@ -25,10 +23,7 @@ def get_onnx2ncnn_path() -> str: Returns: str: A path of onnx2ncnn tool. """ - wildcard = os.path.abspath( - os.path.join( - os.path.dirname(__file__), '../../../build/bin/onnx2ncnn')) - - paths = glob.glob(wildcard) - lib_path = paths[0] if len(paths) > 0 else '' - return lib_path + candidates = [ + '../../../build/bin/onnx2ncnn', '../../../build/bin/*/onnx2ncnn.exe' + ] + return get_file_path(os.path.dirname(__file__), candidates) diff --git a/mmdeploy/backend/onnxruntime/init_plugins.py b/mmdeploy/backend/onnxruntime/init_plugins.py index 06cd001502..e8622eedf3 100644 --- a/mmdeploy/backend/onnxruntime/init_plugins.py +++ b/mmdeploy/backend/onnxruntime/init_plugins.py @@ -1,7 +1,8 @@ # Copyright (c) OpenMMLab. All rights reserved. -import glob import os +from mmdeploy.utils import get_file_path + def get_ops_path() -> str: """Get the library path of onnxruntime custom ops. @@ -9,11 +10,8 @@ def get_ops_path() -> str: Returns: str: The library path to onnxruntime custom ops. """ - wildcard = os.path.abspath( - os.path.join( - os.path.dirname(__file__), - '../../../build/lib/libmmdeploy_onnxruntime_ops.so')) - - paths = glob.glob(wildcard) - lib_path = paths[0] if len(paths) > 0 else '' - return lib_path + candidates = [ + '../../../build/lib/libmmdeploy_onnxruntime_ops.so', + '../../../build/bin/*/mmdeploy_onnxruntime_ops.dll', + ] + return get_file_path(os.path.dirname(__file__), candidates) diff --git a/mmdeploy/backend/sdk/__init__.py b/mmdeploy/backend/sdk/__init__.py index 95a9548505..ef648c4d5b 100644 --- a/mmdeploy/backend/sdk/__init__.py +++ b/mmdeploy/backend/sdk/__init__.py @@ -3,14 +3,24 @@ import os import sys -lib_dir = os.path.abspath( - os.path.join(os.path.dirname(__file__), '../../../build/lib')) - -sys.path.insert(0, lib_dir) +from mmdeploy.utils import get_file_path _is_available = False -if importlib.util.find_spec('mmdeploy_python') is not None: +module_name = 'mmdeploy_python' + +candidates = [ + f'../../../build/lib/{module_name}.*.so', + f'../../../build/bin/*/{module_name}.*.pyd' +] + +lib_path = get_file_path(os.path.dirname(__file__), candidates) + +if lib_path: + lib_dir = os.path.dirname(lib_path) + sys.path.insert(0, lib_dir) + +if importlib.util.find_spec(module_name) is not None: from .wrapper import SDKWrapper __all__ = ['SDKWrapper'] _is_available = True diff --git a/mmdeploy/backend/tensorrt/init_plugins.py b/mmdeploy/backend/tensorrt/init_plugins.py index 9bb0da7f43..80c6eea4d7 100644 --- a/mmdeploy/backend/tensorrt/init_plugins.py +++ b/mmdeploy/backend/tensorrt/init_plugins.py @@ -1,9 +1,8 @@ # Copyright (c) OpenMMLab. All rights reserved. import ctypes -import glob import os -from mmdeploy.utils import get_root_logger +from mmdeploy.utils import get_file_path, get_root_logger def get_ops_path() -> str: @@ -12,14 +11,11 @@ def get_ops_path() -> str: Returns: str: A path of the TensorRT plugin library. """ - wildcard = os.path.abspath( - os.path.join( - os.path.dirname(__file__), - '../../../build/lib/libmmdeploy_tensorrt_ops.so')) - - paths = glob.glob(wildcard) - lib_path = paths[0] if len(paths) > 0 else '' - return lib_path + candidates = [ + '../../../build/lib/libmmdeploy_tensorrt_ops.so', + '../../../build/bin/*/mmdeploy_tensorrt_ops.dll' + ] + return get_file_path(os.path.dirname(__file__), candidates) def load_tensorrt_plugin() -> bool: diff --git a/mmdeploy/backend/torchscript/__init__.py b/mmdeploy/backend/torchscript/__init__.py new file mode 100644 index 0000000000..9179ef3da6 --- /dev/null +++ b/mmdeploy/backend/torchscript/__init__.py @@ -0,0 +1,20 @@ +# Copyright (c) OpenMMLab. All rights reserved. +# flake8: noqa +from .init_plugins import get_ops_path, ops_available + + +def is_available(): + """Torchscript available. + + Returns: + bool: Always True. + """ + return True + + +__all__ = ['get_ops_path', 'ops_available'] + +if is_available(): + from .wrapper import TorchscriptWrapper + + __all__ += ['TorchscriptWrapper'] diff --git a/mmdeploy/backend/torchscript/init_plugins.py b/mmdeploy/backend/torchscript/init_plugins.py new file mode 100644 index 0000000000..ec0371b59a --- /dev/null +++ b/mmdeploy/backend/torchscript/init_plugins.py @@ -0,0 +1,28 @@ +# Copyright (c) OpenMMLab. All rights reserved. +import glob +import os.path as osp + + +def get_ops_path() -> str: + """Get path of the torchscript extension library. + + Returns: + str: A path of the torchscript extension library. + """ + wildcard = osp.abspath( + osp.join( + osp.dirname(__file__), + '../../../build/lib/libmmdeploy_torchscript_ops.so')) + + paths = glob.glob(wildcard) + lib_path = paths[0] if len(paths) > 0 else '' + return lib_path + + +def ops_available() -> bool: + """Return whether ops are available. + + Returns: + bool: Whether ops are available. + """ + return osp.exists(get_ops_path()) diff --git a/mmdeploy/backend/torchscript/wrapper.py b/mmdeploy/backend/torchscript/wrapper.py new file mode 100644 index 0000000000..668ab23aa0 --- /dev/null +++ b/mmdeploy/backend/torchscript/wrapper.py @@ -0,0 +1,108 @@ +# Copyright (c) OpenMMLab. All rights reserved. +import os.path as osp +from typing import Dict, Optional, Sequence, Union + +import torch + +from mmdeploy.utils import Backend +from mmdeploy.utils.timer import TimeCounter +from ..base import BACKEND_WRAPPER, BaseWrapper +from .init_plugins import get_ops_path + + +@BACKEND_WRAPPER.register_module(Backend.TORCHSCRIPT.value) +class TorchscriptWrapper(BaseWrapper): + """Torchscript engine wrapper for inference. + + Args: + model (torch.jit.RecursiveScriptModule): torchscript engine to wrap. + input_names (Sequence[str] | None): Names of model inputs in order. + Defaults to `None` and the wrapper will accept list or Tensor. + output_names (Sequence[str] | None): Names of model outputs in order. + Defaults to `None` and the wrapper will return list or Tensor. + + Note: + If the engine is converted from onnx model. The input_names and + output_names should be the same as onnx model. + + Examples: + >>> from mmdeploy.backend.torchscript import TorchscriptWrapper + >>> engine_file = 'resnet.engine' + >>> model = TorchscriptWrapper(engine_file, input_names=['input'], \ + >>> output_names=['output']) + >>> inputs = dict(input=torch.randn(1, 3, 224, 224)) + >>> outputs = model(inputs) + >>> print(outputs) + """ + + def __init__(self, + model: Union[str, torch.jit.RecursiveScriptModule], + input_names: Optional[Sequence[str]] = None, + output_names: Optional[Sequence[str]] = None): + # load custom ops if exist + custom_ops_path = get_ops_path() + if osp.exists(custom_ops_path): + torch.ops.load_library(custom_ops_path) + super().__init__(output_names) + self.ts_model = model + if isinstance(self.ts_model, str): + self.ts_model = torch.jit.load(self.ts_model) + + assert isinstance(self.ts_model, torch.jit.RecursiveScriptModule + ), 'failed to load torchscript model.' + + self._input_names = input_names + self._output_names = output_names + + def forward( + self, inputs: Union[torch.Tensor, Sequence[torch.Tensor], + Dict[str, torch.Tensor]] + ) -> Union[torch.Tensor, Sequence[torch.Tensor], Dict[str, torch.Tensor]]: + """Run forward inference. + + Args: + inputs (torch.Tensor | Sequence[torch.Tensor] | Dict[str, + torch.Tensor]): The input tensor, or tensor sequence, or pairs + of input names and tensors. + + Return: + outputs (torch.Tensor | Sequence[torch.Tensor] | Dict[str, + torch.Tensor]): The input tensor, or tensor sequence, or pairs + of input names and tensors. + """ + + is_dict_inputs = isinstance(inputs, Dict) + if is_dict_inputs: + # inputs to dict + assert self._input_names is not None, \ + 'input names have not been given.' + inputs = [inputs[input_name] for input_name in self._input_names] + elif isinstance(inputs, torch.Tensor): + inputs = [inputs] + + outputs = self.__torchscript_execute(inputs) + + if self._output_names is not None and is_dict_inputs: + # output to dict + if isinstance(outputs, torch.Tensor): + outputs = [outputs] + outputs = dict(zip(self._output_names, outputs)) + + if isinstance(outputs, tuple) and self._output_names is not None: + assert len(outputs) == len(self._output_names) + outputs = dict(zip(self._output_names, outputs)) + return outputs + + @TimeCounter.count_time() + def __torchscript_execute( + self, inputs: Sequence[torch.Tensor]) -> Sequence[torch.Tensor]: + """Run inference with TorchScript. + + Args: + inputs (Sequence[torch.Tensor]): A list of integer binding the + input/output. + Returns: + torch.Tensor | Sequence[torch.Tensor]: The inference outputs from + TorchScript. + """ + return self.ts_model(*inputs) diff --git a/mmdeploy/codebase/base/backend_model.py b/mmdeploy/codebase/base/backend_model.py index 6a21440693..93dc2fe74f 100644 --- a/mmdeploy/codebase/base/backend_model.py +++ b/mmdeploy/codebase/base/backend_model.py @@ -35,8 +35,11 @@ def __init__(self, def _build_wrapper(backend: Backend, backend_files: Sequence[str], device: str, + input_names: Optional[Sequence[str]] = None, output_names: Optional[Sequence[str]] = None, - deploy_cfg: Optional[mmcv.Config] = None): + deploy_cfg: Optional[mmcv.Config] = None, + *args, + **kwargs): """The default methods to build backend wrappers. Args: @@ -44,6 +47,8 @@ def _build_wrapper(backend: Backend, beckend_files (Sequence[str]): Paths to all required backend files( e.g. '.onnx' for ONNX Runtime, '.param' and '.bin' for ncnn). device (str): A string specifying device type. + input_names (Sequence[str] | None): Names of model inputs in + order. Defaults to `None`. output_names (Sequence[str] | None): Names of model outputs in order. Defaults to `None` and the wrapper will load the output names from the model. @@ -85,6 +90,12 @@ def _build_wrapper(backend: Backend, model_file=backend_files[0], task_name=task_name, device=device) + elif backend == Backend.TORCHSCRIPT: + from mmdeploy.backend.torchscript import TorchscriptWrapper + return TorchscriptWrapper( + model=backend_files[0], + input_names=input_names, + output_names=output_names) else: raise NotImplementedError(f'Unknown backend type: {backend.value}') diff --git a/mmdeploy/codebase/mmcls/deploy/classification_model.py b/mmdeploy/codebase/mmcls/deploy/classification_model.py index f7f3bbfe73..263d097ecd 100644 --- a/mmdeploy/codebase/mmcls/deploy/classification_model.py +++ b/mmdeploy/codebase/mmcls/deploy/classification_model.py @@ -56,6 +56,7 @@ def _init_wrapper(self, backend: Backend, backend_files: Sequence[str], backend=backend, backend_files=backend_files, device=device, + input_names=[self.input_name], output_names=output_names, deploy_cfg=self.deploy_cfg) diff --git a/mmdeploy/codebase/mmcls/models/backbones/shufflenet_v2.py b/mmdeploy/codebase/mmcls/models/backbones/shufflenet_v2.py index 58e7030cab..2d26318af6 100644 --- a/mmdeploy/codebase/mmcls/models/backbones/shufflenet_v2.py +++ b/mmdeploy/codebase/mmcls/models/backbones/shufflenet_v2.py @@ -3,13 +3,17 @@ from mmcls.models.utils import channel_shuffle from mmdeploy.core import FUNCTION_REWRITER +from mmdeploy.utils import Backend # torch.chunk will export dynamic shape slice, which will lead integer input # on ncnn backend. So the model needs to rewrite. @FUNCTION_REWRITER.register_rewriter( func_name='mmcls.models.backbones.shufflenet_v2.InvertedResidual.forward', - backend='ncnn') + backend=Backend.NCNN.value) +@FUNCTION_REWRITER.register_rewriter( + func_name='mmcls.models.backbones.shufflenet_v2.InvertedResidual.forward', + backend=Backend.TORCHSCRIPT.value) def shufflenetv2_backbone__forward__ncnn(ctx, self, x): """Rewrite `forward` of InvertedResidual used in shufflenet_v2 for ncnn backend. diff --git a/mmdeploy/codebase/mmdet/core/bbox/delta_xywh_bbox_coder.py b/mmdeploy/codebase/mmdet/core/bbox/delta_xywh_bbox_coder.py index 1b9b7904a5..56d71499cb 100644 --- a/mmdeploy/codebase/mmdet/core/bbox/delta_xywh_bbox_coder.py +++ b/mmdeploy/codebase/mmdet/core/bbox/delta_xywh_bbox_coder.py @@ -92,48 +92,42 @@ def delta2bbox(ctx, bboxes (Tensor): Boxes with shape (N, num_classes * 4) or (N, 4), where 4 represent tl_x, tl_y, br_x, br_y. """ - means = deltas.new_tensor(means).view(1, - -1).repeat(1, - deltas.size(-1) // 4) - stds = deltas.new_tensor(stds).view(1, -1).repeat(1, deltas.size(-1) // 4) - denorm_deltas = deltas * stds + means - dx = denorm_deltas[..., 0::4] - dy = denorm_deltas[..., 1::4] - dw = denorm_deltas[..., 2::4] - dh = denorm_deltas[..., 3::4] + means = deltas.new_tensor(means).view(1, -1) + stds = deltas.new_tensor(stds).view(1, -1) + delta_shape = deltas.shape + reshaped_deltas = deltas.view(delta_shape[:-1] + (-1, 4)) + denorm_deltas = reshaped_deltas * stds + means - x1, y1 = rois[..., 0], rois[..., 1] - x2, y2 = rois[..., 2], rois[..., 3] - # Compute center of each roi - px = ((x1 + x2) * 0.5).unsqueeze(-1).expand_as(dx) - py = ((y1 + y2) * 0.5).unsqueeze(-1).expand_as(dy) - # Compute width/height of each roi - pw = (x2 - x1).unsqueeze(-1).expand_as(dw) - ph = (y2 - y1).unsqueeze(-1).expand_as(dh) + dxy = denorm_deltas[..., :2] + dwh = denorm_deltas[..., 2:] - dx_width = pw * dx - dy_height = ph * dy + xy1 = rois[..., None, :2] + xy2 = rois[..., None, 2:] + + pxy = (xy1 + xy2) * 0.5 + pwh = xy2 - xy1 + dxy_wh = pwh * dxy max_ratio = np.abs(np.log(wh_ratio_clip)) if add_ctr_clamp: - dx_width = torch.clamp(dx_width, max=ctr_clamp, min=-ctr_clamp) - dy_height = torch.clamp(dy_height, max=ctr_clamp, min=-ctr_clamp) - dw = torch.clamp(dw, max=max_ratio) - dh = torch.clamp(dh, max=max_ratio) + dxy_wh = torch.clamp(dxy_wh, max=ctr_clamp, min=-ctr_clamp) + dwh = torch.clamp(dwh, max=max_ratio) else: - dw = dw.clamp(min=-max_ratio, max=max_ratio) - dh = dh.clamp(min=-max_ratio, max=max_ratio) + dwh = dwh.clamp(min=-max_ratio, max=max_ratio) + # Use exp(network energy) to enlarge/shrink each roi - gw = pw * dw.exp() - gh = ph * dh.exp() + half_gwh = pwh * dwh.exp() * 0.5 # Use network energy to shift the center of each roi - gx = px + dx_width - gy = py + dy_height + gxy = pxy + dxy_wh + # Convert center-xy/width/height to top-left, bottom-right - x1 = gx - gw * 0.5 - y1 = gy - gh * 0.5 - x2 = gx + gw * 0.5 - y2 = gy + gh * 0.5 + xy1 = gxy - half_gwh + xy2 = gxy + half_gwh + + x1 = xy1[..., 0] + y1 = xy1[..., 1] + x2 = xy2[..., 0] + y2 = xy2[..., 1] if clip_border and max_shape is not None: from mmdeploy.codebase.mmdet.deploy import clip_bboxes @@ -190,68 +184,42 @@ def delta2bbox__ncnn(ctx, or (N, num_classes * 4) or (N, 4), where 4 represent tl_x, tl_y, br_x, br_y. """ - means = deltas.new_tensor(means).view(1, 1, - -1).repeat(1, deltas.size(-2), - deltas.size(-1) // 4).data - stds = deltas.new_tensor(stds).view(1, 1, - -1).repeat(1, deltas.size(-2), - deltas.size(-1) // 4).data - denorm_deltas = deltas * stds + means - if denorm_deltas.shape[-1] == 4: - dx = denorm_deltas[..., 0:1] - dy = denorm_deltas[..., 1:2] - dw = denorm_deltas[..., 2:3] - dh = denorm_deltas[..., 3:4] - else: - dx = denorm_deltas[..., 0::4] - dy = denorm_deltas[..., 1::4] - dw = denorm_deltas[..., 2::4] - dh = denorm_deltas[..., 3::4] + means = deltas.new_tensor(means).view(1, 1, 1, -1).data + stds = deltas.new_tensor(stds).view(1, 1, 1, -1).data + delta_shape = deltas.shape + reshaped_deltas = deltas.view(delta_shape[:-1] + (-1, 4)) + denorm_deltas = reshaped_deltas * stds + means - x1, y1 = rois[..., 0:1], rois[..., 1:2] - x2, y2 = rois[..., 2:3], rois[..., 3:4] + dxy = denorm_deltas[..., :2] + dwh = denorm_deltas[..., 2:] - # Compute center of each roi - px = (x1 + x2) * 0.5 - py = (y1 + y2) * 0.5 - # Compute width/height of each roi - pw = x2 - x1 - ph = y2 - y1 + xy1 = rois[..., None, :2] + xy2 = rois[..., None, 2:] - # do not use expand unless necessary - # since expand is a custom ops - if px.shape[-1] != 4: - px = px.expand_as(dx) - if py.shape[-1] != 4: - py = py.expand_as(dy) - if pw.shape[-1] != 4: - pw = pw.expand_as(dw) - if px.shape[-1] != 4: - ph = ph.expand_as(dh) - - dx_width = pw * dx - dy_height = ph * dy + pxy = (xy1 + xy2) * 0.5 + pwh = xy2 - xy1 + dxy_wh = pwh * dxy max_ratio = np.abs(np.log(wh_ratio_clip)) if add_ctr_clamp: - dx_width = torch.clamp(dx_width, max=ctr_clamp, min=-ctr_clamp) - dy_height = torch.clamp(dy_height, max=ctr_clamp, min=-ctr_clamp) - dw = torch.clamp(dw, max=max_ratio) - dh = torch.clamp(dh, max=max_ratio) + dxy_wh = torch.clamp(dxy_wh, max=ctr_clamp, min=-ctr_clamp) + dwh = torch.clamp(dwh, max=max_ratio) else: - dw = dw.clamp(min=-max_ratio, max=max_ratio) - dh = dh.clamp(min=-max_ratio, max=max_ratio) + dwh = dwh.clamp(min=-max_ratio, max=max_ratio) + # Use exp(network energy) to enlarge/shrink each roi - gw = pw * dw.exp() - gh = ph * dh.exp() + half_gwh = pwh * dwh.exp() * 0.5 # Use network energy to shift the center of each roi - gx = px + dx_width - gy = py + dy_height + gxy = pxy + dxy_wh + # Convert center-xy/width/height to top-left, bottom-right - x1 = gx - gw * 0.5 - y1 = gy - gh * 0.5 - x2 = gx + gw * 0.5 - y2 = gy + gh * 0.5 + xy1 = gxy - half_gwh + xy2 = gxy + half_gwh + + x1 = xy1[..., 0] + y1 = xy1[..., 1] + x2 = xy2[..., 0] + y2 = xy2[..., 1] if clip_border and max_shape is not None: from mmdeploy.codebase.mmdet.deploy import clip_bboxes diff --git a/mmdeploy/codebase/mmdet/core/ops/detection_output.py b/mmdeploy/codebase/mmdet/core/ops/detection_output.py index 48d9f84415..67809660de 100644 --- a/mmdeploy/codebase/mmdet/core/ops/detection_output.py +++ b/mmdeploy/codebase/mmdet/core/ops/detection_output.py @@ -1,3 +1,4 @@ +# Copyright (c) OpenMMLab. All rights reserved. import torch diff --git a/mmdeploy/codebase/mmdet/core/ops/prior_box.py b/mmdeploy/codebase/mmdet/core/ops/prior_box.py index 24efb02dee..28d76f95bf 100644 --- a/mmdeploy/codebase/mmdet/core/ops/prior_box.py +++ b/mmdeploy/codebase/mmdet/core/ops/prior_box.py @@ -1,3 +1,4 @@ +# Copyright (c) OpenMMLab. All rights reserved. import torch diff --git a/mmdeploy/codebase/mmdet/core/post_processing/bbox_nms.py b/mmdeploy/codebase/mmdet/core/post_processing/bbox_nms.py index 2580d37ef8..ee7a1403d7 100644 --- a/mmdeploy/codebase/mmdet/core/post_processing/bbox_nms.py +++ b/mmdeploy/codebase/mmdet/core/post_processing/bbox_nms.py @@ -5,7 +5,7 @@ import mmdeploy from mmdeploy.core import FUNCTION_REWRITER, mark from mmdeploy.mmcv.ops import ONNXNMSop, TRTBatchedNMSop -from mmdeploy.utils import is_dynamic_batch +from mmdeploy.utils import Backend, is_dynamic_batch def select_nms_index(scores: torch.Tensor, @@ -269,3 +269,60 @@ def multiclass_nms(*args, **kwargs): """Wrapper function for `_multiclass_nms`.""" return mmdeploy.codebase.mmdet.core.post_processing._multiclass_nms( *args, **kwargs) + + +@FUNCTION_REWRITER.register_rewriter( + func_name='mmdeploy.codebase.mmdet.core.post_processing._multiclass_nms', + backend=Backend.TORCHSCRIPT.value) +def multiclass_nms__torchscript(ctx, + boxes: Tensor, + scores: Tensor, + max_output_boxes_per_class: int = 1000, + iou_threshold: float = 0.5, + score_threshold: float = 0.05, + pre_top_k: int = -1, + keep_top_k: int = -1): + """rewrite for torchscript batched nms. + + Use batched_nms from torchvision instead of custom nms. + """ + # TODO: simplify inference for non-batch model + from torchvision.ops import batched_nms + batch_size = scores.shape[0] + num_boxes = scores.shape[1] + num_classes = scores.shape[2] + box_per_cls = len(boxes.shape) == 4 + scores = torch.where(scores > score_threshold, scores, scores.new_zeros(1)) + + # pre-topk + if pre_top_k > 0: + max_scores, _ = scores.max(-1) + _, topk_inds = max_scores.topk(pre_top_k) + batch_inds = torch.arange(batch_size).view( + -1, 1).expand_as(topk_inds).long() + boxes = boxes[batch_inds, topk_inds, ...] + scores = scores[batch_inds, topk_inds, :] + num_boxes = scores.shape[1] + + idxs = torch.arange(0, batch_size, device=scores.device).unsqueeze(1) + idxs = idxs.repeat(1, num_boxes).view(-1) + + keeps = [None] * num_classes + for cls_id in range(num_classes): + box = boxes if not box_per_cls else boxes[:, :, cls_id, :] + score = scores[:, :, cls_id] + box = box.view(-1, 4) + score = score.view(-1) + box_keep = batched_nms(box, score, idxs, iou_threshold=iou_threshold) + box_keep = box_keep[:max_output_boxes_per_class * batch_size] + batch_keep = idxs[box_keep] + cls_keep = torch.ones_like(box_keep) * cls_id + box_keep = box_keep - batch_keep * num_boxes + keeps[cls_id] = torch.stack([batch_keep, cls_keep, box_keep], dim=1) + + keeps = torch.cat(keeps) + scores = scores.permute(0, 2, 1) + dets, labels = select_nms_index( + scores, boxes, keeps, batch_size, keep_top_k=keep_top_k) + + return dets, labels diff --git a/mmdeploy/codebase/mmdet/deploy/object_detection_model.py b/mmdeploy/codebase/mmdet/deploy/object_detection_model.py index 79cdae9ecc..b368d10972 100644 --- a/mmdeploy/codebase/mmdet/deploy/object_detection_model.py +++ b/mmdeploy/codebase/mmdet/deploy/object_detection_model.py @@ -78,6 +78,7 @@ def _init_wrapper(self, backend: Backend, backend_files: Sequence[str], backend=backend, backend_files=backend_files, device=device, + input_names=[self.input_name], output_names=output_names, deploy_cfg=self.deploy_cfg) @@ -424,13 +425,14 @@ def _init_wrapper(self, backend, backend_files, device): backend, backend_files[0:n], device, - partition0_output_names, + output_names=partition0_output_names, deploy_cfg=self.deploy_cfg) self.second_wrapper = BaseBackendModel._build_wrapper( backend, backend_files[n:2 * n], - device, ['cls_score', 'bbox_pred'], + device, + output_names=['cls_score', 'bbox_pred'], deploy_cfg=self.deploy_cfg) def partition0_postprocess(self, x: Sequence[torch.Tensor], diff --git a/mmdeploy/codebase/mmdet/deploy/utils.py b/mmdeploy/codebase/mmdet/deploy/utils.py index 1ecd451e2f..860cb54239 100644 --- a/mmdeploy/codebase/mmdet/deploy/utils.py +++ b/mmdeploy/codebase/mmdet/deploy/utils.py @@ -5,6 +5,8 @@ import torch from torch import Tensor +from mmdeploy.core import FUNCTION_REWRITER +from mmdeploy.core.rewriters.rewriter_utils import LibVersionChecker from mmdeploy.utils import load_config @@ -69,6 +71,33 @@ def clip_bboxes(x1: Tensor, y1: Tensor, x2: Tensor, y2: Tensor, return x1, y1, x2, y2 +@FUNCTION_REWRITER.register_rewriter( + func_name='mmdeploy.codebase.mmdet.deploy.utils.clip_bboxes', + backend='tensorrt', + extra_checkers=LibVersionChecker('tensorrt', min_version='8')) +def clip_bboxes__trt8(ctx, x1: Tensor, y1: Tensor, x2: Tensor, y2: Tensor, + max_shape: Union[Tensor, Sequence[int]]): + """Clip bboxes for onnx. From TensorRT 8 we can do the operators on the + tensors directly. + + Args: + ctx (ContextCaller): The context with additional information. + x1 (Tensor): The x1 for bounding boxes. + y1 (Tensor): The y1 for bounding boxes. + x2 (Tensor): The x2 for bounding boxes. + y2 (Tensor): The y2 for bounding boxes. + max_shape (Tensor | Sequence[int]): The (H,W) of original image. + Returns: + tuple(Tensor): The clipped x1, y1, x2, y2. + """ + assert len(max_shape) == 2, '`max_shape` should be [h, w]' + x1 = torch.clamp(x1, 0, max_shape[1]) + y1 = torch.clamp(y1, 0, max_shape[0]) + x2 = torch.clamp(x2, 0, max_shape[1]) + y2 = torch.clamp(y2, 0, max_shape[0]) + return x1, y1, x2, y2 + + def pad_with_value(x: Tensor, pad_dim: int, pad_size: int, diff --git a/mmdeploy/codebase/mmdet/models/backbones.py b/mmdeploy/codebase/mmdet/models/backbones.py index 2012df0241..6520702adb 100644 --- a/mmdeploy/codebase/mmdet/models/backbones.py +++ b/mmdeploy/codebase/mmdet/models/backbones.py @@ -1,3 +1,4 @@ +# Copyright (c) OpenMMLab. All rights reserved. import torch from mmdeploy.core import FUNCTION_REWRITER diff --git a/mmdeploy/codebase/mmdet/models/dense_heads/__init__.py b/mmdeploy/codebase/mmdet/models/dense_heads/__init__.py index 080fe6c7ba..9043d44264 100644 --- a/mmdeploy/codebase/mmdet/models/dense_heads/__init__.py +++ b/mmdeploy/codebase/mmdet/models/dense_heads/__init__.py @@ -2,6 +2,7 @@ from .base_dense_head import (base_dense_head__get_bbox, base_dense_head__get_bboxes__ncnn) from .fovea_head import fovea_head__get_bboxes +from .gfl_head import gfl_head__get_bbox from .rpn_head import rpn_head__get_bboxes, rpn_head__get_bboxes__ncnn from .ssd_head import ssd_head__get_bboxes__ncnn from .yolo_head import yolov3_head__get_bboxes, yolov3_head__get_bboxes__ncnn @@ -12,5 +13,6 @@ 'yolov3_head__get_bboxes', 'yolov3_head__get_bboxes__ncnn', 'yolox_head__get_bboxes', 'base_dense_head__get_bbox', 'fovea_head__get_bboxes', 'base_dense_head__get_bboxes__ncnn', - 'ssd_head__get_bboxes__ncnn', 'yolox_head__get_bboxes__ncnn' + 'ssd_head__get_bboxes__ncnn', 'yolox_head__get_bboxes__ncnn', + 'gfl_head__get_bbox' ] diff --git a/mmdeploy/codebase/mmdet/models/dense_heads/base_dense_head.py b/mmdeploy/codebase/mmdet/models/dense_heads/base_dense_head.py index d4413182a9..3c94c16250 100644 --- a/mmdeploy/codebase/mmdet/models/dense_heads/base_dense_head.py +++ b/mmdeploy/codebase/mmdet/models/dense_heads/base_dense_head.py @@ -1,3 +1,4 @@ +# Copyright (c) OpenMMLab. All rights reserved. import torch from mmdet.core.bbox.coder import (DeltaXYWHBBoxCoder, DistancePointBBoxCoder, TBLRBBoxCoder) diff --git a/mmdeploy/codebase/mmdet/models/dense_heads/gfl_head.py b/mmdeploy/codebase/mmdet/models/dense_heads/gfl_head.py new file mode 100644 index 0000000000..8dba8b5666 --- /dev/null +++ b/mmdeploy/codebase/mmdet/models/dense_heads/gfl_head.py @@ -0,0 +1,185 @@ +# Copyright (c) OpenMMLab. All rights reserved. +import torch +import torch.nn.functional as F + +from mmdeploy.codebase.mmdet import (get_post_processing_params, + multiclass_nms, pad_with_value) +from mmdeploy.core import FUNCTION_REWRITER +from mmdeploy.utils import Backend, get_backend, is_dynamic_shape + + +@FUNCTION_REWRITER.register_rewriter( + func_name='mmdet.models.dense_heads.gfl_head.' + 'GFLHead.get_bboxes') +def gfl_head__get_bbox(ctx, + self, + cls_scores, + bbox_preds, + score_factors=None, + img_metas=None, + cfg=None, + rescale=False, + with_nms=True, + **kwargs): + """Rewrite `get_bboxes` of `GFLHead` for default backend. + + Rewrite this function to deploy model, transform network output for a + batch into bbox predictions. + + Args: + ctx (ContextCaller): The context with additional information. + self: The instance of the original class. + cls_scores (list[Tensor]): Classification scores for all + scale levels, each is a 4D-tensor, has shape + (batch_size, num_priors * num_classes, H, W). + bbox_preds (list[Tensor]): Box energies / deltas for all + scale levels, each is a 4D-tensor, has shape + (batch_size, num_priors * 4, H, W). + score_factors (list[Tensor], Optional): Score factor for + all scale level, each is a 4D-tensor, has shape + (batch_size, num_priors * 1, H, W). Default None. + img_metas (list[dict], Optional): Image meta info. Default None. + cfg (mmcv.Config, Optional): Test / postprocessing configuration, + if None, test_cfg would be used. Default None. + rescale (bool): If True, return boxes in original image space. + Default False. + with_nms (bool): If True, do nms before return boxes. + Default True. + + Returns: + If with_nms == True: + tuple[Tensor, Tensor]: tuple[Tensor, Tensor]: (dets, labels), + `dets` of shape [N, num_det, 5] and `labels` of shape + [N, num_det]. + Else: + tuple[Tensor, Tensor, Tensor]: batch_mlvl_bboxes, + batch_mlvl_scores, batch_mlvl_centerness + """ + deploy_cfg = ctx.cfg + is_dynamic_flag = is_dynamic_shape(deploy_cfg) + backend = get_backend(deploy_cfg) + num_levels = len(cls_scores) + + featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores] + mlvl_priors = self.prior_generator.grid_priors( + featmap_sizes, dtype=bbox_preds[0].dtype, device=bbox_preds[0].device) + + mlvl_cls_scores = [cls_scores[i].detach() for i in range(num_levels)] + mlvl_bbox_preds = [bbox_preds[i].detach() for i in range(num_levels)] + if score_factors is None: + with_score_factors = False + mlvl_score_factor = [None for _ in range(num_levels)] + else: + with_score_factors = True + mlvl_score_factor = [ + score_factors[i].detach() for i in range(num_levels) + ] + mlvl_score_factors = [] + assert img_metas is not None + img_shape = img_metas[0]['img_shape'] + + assert len(cls_scores) == len(bbox_preds) == len(mlvl_priors) + batch_size = cls_scores[0].shape[0] + cfg = self.test_cfg + pre_topk = cfg.get('nms_pre', -1) + + mlvl_valid_bboxes = [] + mlvl_valid_scores = [] + mlvl_valid_priors = [] + + for cls_score, bbox_pred, score_factors, priors, stride in zip( + mlvl_cls_scores, mlvl_bbox_preds, mlvl_score_factor, mlvl_priors, + self.prior_generator.strides): + assert cls_score.size()[-2:] == bbox_pred.size()[-2:] + assert stride[0] == stride[1] + + scores = cls_score.permute(0, 2, 3, 1).reshape(batch_size, -1, + self.cls_out_channels) + if self.use_sigmoid_cls: + scores = scores.sigmoid() + nms_pre_score = scores + else: + scores = scores.softmax(-1) + nms_pre_score = scores + if with_score_factors: + score_factors = score_factors.permute(0, 2, 3, + 1).reshape(batch_size, + -1).sigmoid() + score_factors = score_factors.unsqueeze(2) + bbox_pred = batched_integral(self.integral, + bbox_pred.permute(0, 2, 3, 1)) * stride[0] + if not is_dynamic_flag: + priors = priors.data + priors = priors.expand(batch_size, -1, priors.size(-1)) + if pre_topk > 0: + if with_score_factors: + nms_pre_score = nms_pre_score * score_factors + if backend == Backend.TENSORRT: + priors = pad_with_value(priors, 1, pre_topk) + bbox_pred = pad_with_value(bbox_pred, 1, pre_topk) + scores = pad_with_value(scores, 1, pre_topk, 0.) + nms_pre_score = pad_with_value(nms_pre_score, 1, pre_topk, 0.) + if with_score_factors: + score_factors = pad_with_value(score_factors, 1, pre_topk, + 0.) + + # Get maximum scores for foreground classes. + if self.use_sigmoid_cls: + max_scores, _ = nms_pre_score.max(-1) + else: + max_scores, _ = nms_pre_score[..., :-1].max(-1) + _, topk_inds = max_scores.topk(pre_topk) + batch_inds = torch.arange( + batch_size, + device=bbox_pred.device).view(-1, 1).expand_as(topk_inds) + priors = priors[batch_inds, topk_inds, :] + bbox_pred = bbox_pred[batch_inds, topk_inds, :] + scores = scores[batch_inds, topk_inds, :] + if with_score_factors: + score_factors = score_factors[batch_inds, topk_inds, :] + + mlvl_valid_bboxes.append(bbox_pred) + mlvl_valid_scores.append(scores) + priors = self.anchor_center(priors) + mlvl_valid_priors.append(priors) + if with_score_factors: + mlvl_score_factors.append(score_factors) + + batch_mlvl_bboxes_pred = torch.cat(mlvl_valid_bboxes, dim=1) + batch_scores = torch.cat(mlvl_valid_scores, dim=1) + batch_priors = torch.cat(mlvl_valid_priors, dim=1) + batch_bboxes = self.bbox_coder.decode( + batch_priors, batch_mlvl_bboxes_pred, max_shape=img_shape) + if with_score_factors: + batch_score_factors = torch.cat(mlvl_score_factors, dim=1) + + if not self.use_sigmoid_cls: + batch_scores = batch_scores[..., :self.num_classes] + + if with_score_factors: + batch_scores = batch_scores * batch_score_factors + if not with_nms: + return batch_bboxes, batch_scores + post_params = get_post_processing_params(deploy_cfg) + max_output_boxes_per_class = post_params.max_output_boxes_per_class + iou_threshold = cfg.nms.get('iou_threshold', post_params.iou_threshold) + score_threshold = cfg.get('score_thr', post_params.score_threshold) + pre_top_k = post_params.pre_top_k + keep_top_k = cfg.get('max_per_img', post_params.keep_top_k) + return multiclass_nms( + batch_bboxes, + batch_scores, + max_output_boxes_per_class, + iou_threshold=iou_threshold, + score_threshold=score_threshold, + pre_top_k=pre_top_k, + keep_top_k=keep_top_k) + + +def batched_integral(intergral, x): + batch_size = x.size(0) + x = F.softmax(x.reshape(batch_size, -1, intergral.reg_max + 1), dim=2) + x = F.linear(x, + intergral.project.type_as(x).unsqueeze(0)).reshape( + batch_size, -1, 4) + return x diff --git a/mmdeploy/codebase/mmdet/models/necks.py b/mmdeploy/codebase/mmdet/models/necks.py index f430124381..2931de9b0a 100644 --- a/mmdeploy/codebase/mmdet/models/necks.py +++ b/mmdeploy/codebase/mmdet/models/necks.py @@ -1,3 +1,4 @@ +# Copyright (c) OpenMMLab. All rights reserved. import torch from mmdeploy.core import FUNCTION_REWRITER diff --git a/mmdeploy/codebase/mmdet/models/roi_heads/single_level_roi_extractor.py b/mmdeploy/codebase/mmdet/models/roi_heads/single_level_roi_extractor.py index c997a5ab6a..de0ea22dad 100644 --- a/mmdeploy/codebase/mmdet/models/roi_heads/single_level_roi_extractor.py +++ b/mmdeploy/codebase/mmdet/models/roi_heads/single_level_roi_extractor.py @@ -4,6 +4,8 @@ from mmdeploy.core.optimizers import mark from mmdeploy.core.rewriters import FUNCTION_REWRITER +from mmdeploy.utils import get_backend +from mmdeploy.utils.constants import Backend class MultiLevelRoiAlign(Function): @@ -98,14 +100,18 @@ def single_roi_extractor__forward(ctx, roi_scale_factor=None): """Rewrite `forward` of SingleRoIExtractor for default backend. - Rewrite this function to enable exporting to onnx even though the input + Rewrite this function to: + 1. enable exporting to IR even though the input image contains no targets. Note that, `ScatterND` of onnx may conflict with `Reshape` if a tensor have a dim size of 0. Thus, we have to cat zeros to the dim 0 of `roi_feats` and recover back after all roi align finished. - Besides, this function adds mark for roi_extractor forward and remove - unnecessary code of origin forward function. + 2. this function adds mark for roi_extractor forward and remove + unnecessary code of origin forward function when using ONNX as IR. + + 3. use the roi align in torhcvision to accelerate the inference. """ + backend = get_backend(ctx.cfg) out_size = self.roi_layers[0].output_size num_levels = len(feats) roi_feats = feats[0].new_zeros(rois.shape[0], self.out_channels, *out_size) @@ -118,29 +124,30 @@ def single_roi_extractor__forward(ctx, if roi_scale_factor is not None: rois = self.roi_rescale(rois, roi_scale_factor) - # concat len num_levels * 2 of zero tensors to dim 0 of roi_feats + # concate zeros to rois and roi_feats for empty tensor cases roi_feats = torch.cat( (roi_feats.new_zeros(num_levels * 2, *roi_feats.shape[-3:]), roi_feats)) + rois = torch.cat((rois.new_zeros(num_levels * 2, 5), rois)) + _tmp = torch.linspace( + 0, + num_levels - 1, + num_levels, + dtype=target_lvls.dtype, + device=target_lvls.device) + target_lvls = torch.cat((_tmp, _tmp, target_lvls)) for i in range(num_levels): + # use the roi align in torhcvision to accelerate the inference + # roi_align in MMCV is same as torchvision when pool mode is 'avg' + if backend == Backend.TORCHSCRIPT or self.roi_layers[ + i].pool_mode == 'avg': + self.roi_layers[i].use_torchvision = True mask = target_lvls == i inds = mask.nonzero(as_tuple=False).squeeze(1) - - # concat len 2 zero tensors to dim 0 of roi_feats - rois_i = torch.cat((rois.new_zeros(2, 5), rois[inds])) - - roi_feats_t = self.roi_layers[i](feats[i], rois_i) - - # correspondingly change the inds - inds = torch.cat([ - torch.tensor([2 * i, 2 * i + 1], - device=inds.device, - dtype=inds.dtype), inds + num_levels * 2 - ]) + roi_feats_t = self.roi_layers[i](feats[i], rois[inds]) roi_feats[inds] = roi_feats_t - - # slice and recover tensors - roi_feats = roi_feats[num_levels * (2):] + # slice to recover original size + roi_feats = roi_feats[num_levels * 2:] return roi_feats diff --git a/mmdeploy/codebase/mmedit/deploy/super_resolution_model.py b/mmdeploy/codebase/mmedit/deploy/super_resolution_model.py index 903899b6ee..ade5d0beea 100644 --- a/mmdeploy/codebase/mmedit/deploy/super_resolution_model.py +++ b/mmdeploy/codebase/mmedit/deploy/super_resolution_model.py @@ -54,6 +54,7 @@ def _init_wrapper(self, backend: Backend, backend_files: Sequence[str], backend=backend, backend_files=backend_files, device=device, + input_names=[self.input_name], output_names=output_names, deploy_cfg=self.deploy_cfg) @@ -204,11 +205,7 @@ def forward(self, output = self.wrapper.invoke([img])[0] if test_mode: output = torch.from_numpy(output) - output = torch.permute(output, ( - 2, - 0, - 1, - )) + output = output.permute(2, 0, 1) output = output / 255. results = self.test_post_process([output], lq, gt) return results diff --git a/mmdeploy/codebase/mmocr/deploy/text_detection_model.py b/mmdeploy/codebase/mmocr/deploy/text_detection_model.py index 31861b66e4..d6917161d9 100644 --- a/mmdeploy/codebase/mmocr/deploy/text_detection_model.py +++ b/mmdeploy/codebase/mmocr/deploy/text_detection_model.py @@ -67,6 +67,7 @@ def _init_wrapper(self, backend: Backend, backend_files: Sequence[str], backend=backend, backend_files=backend_files, device=device, + input_names=[self.input_name], output_names=output_names, deploy_cfg=self.deploy_cfg) diff --git a/mmdeploy/codebase/mmocr/deploy/text_recognition_model.py b/mmdeploy/codebase/mmocr/deploy/text_recognition_model.py index 7f07dbba63..de9d18154f 100644 --- a/mmdeploy/codebase/mmocr/deploy/text_recognition_model.py +++ b/mmdeploy/codebase/mmocr/deploy/text_recognition_model.py @@ -72,6 +72,7 @@ def _init_wrapper(self, backend: Backend, backend_files: Sequence[str], backend=backend, backend_files=backend_files, device=device, + input_names=[self.input_name], output_names=output_names, deploy_cfg=self.deploy_cfg) diff --git a/mmdeploy/codebase/mmseg/deploy/segmentation_model.py b/mmdeploy/codebase/mmseg/deploy/segmentation_model.py index 46e0789031..a57cb9a70b 100644 --- a/mmdeploy/codebase/mmseg/deploy/segmentation_model.py +++ b/mmdeploy/codebase/mmseg/deploy/segmentation_model.py @@ -59,6 +59,7 @@ def _init_wrapper(self, backend, backend_files, device): backend=backend, backend_files=backend_files, device=device, + input_names=[self.input_name], output_names=output_names, deploy_cfg=self.deploy_cfg) diff --git a/mmdeploy/core/optimizers/function_marker.py b/mmdeploy/core/optimizers/function_marker.py index 98a46f6e73..5ad0501593 100644 --- a/mmdeploy/core/optimizers/function_marker.py +++ b/mmdeploy/core/optimizers/function_marker.py @@ -5,7 +5,7 @@ import torch from mmdeploy.core.rewriters import FUNCTION_REWRITER -from mmdeploy.utils import cfg_apply_marks, get_partition_config +from mmdeploy.utils import IR, cfg_apply_marks, get_partition_config MARK_FUNCTION_COUNT = dict() @@ -180,6 +180,20 @@ def impl(ys, prefix, level): return impl(xs, (), level) +@FUNCTION_REWRITER.register_rewriter( + 'mmdeploy.core.optimizers.function_marker.mark_tensors', ir=IR.TORCHSCRIPT) +def remove_mark__torchscript(ctx, xs: Any, *args, **kwargs): + """Disable all marks for TorchScript backend. + + As the Node `mark` is not able to be traced, we just return original input + for the function `mark_tensors`. + + Args: + xs (Any): Input structure which contains tensor. + """ + return xs + + def mark(func_name: Optional[str] = None, inputs: Optional[Sequence[str]] = None, outputs: Optional[Sequence[str]] = None, diff --git a/mmdeploy/core/rewriters/function_rewriter.py b/mmdeploy/core/rewriters/function_rewriter.py index 674361f634..e80ed41d06 100644 --- a/mmdeploy/core/rewriters/function_rewriter.py +++ b/mmdeploy/core/rewriters/function_rewriter.py @@ -1,8 +1,9 @@ # Copyright (c) OpenMMLab. All rights reserved. -from typing import Callable, Dict +from typing import Callable, Dict, List, Optional, Union -from mmdeploy.utils import Backend, get_root_logger -from .rewriter_utils import ContextCaller, RewriterRegistry, import_function +from mmdeploy.utils import IR, Backend, get_root_logger +from .rewriter_utils import (Checker, ContextCaller, RewriterRegistry, + import_function) def _set_func(origin_func_path: str, rewrite_func: Callable): @@ -66,32 +67,33 @@ class FunctionRewriter: def __init__(self): self._registry = RewriterRegistry() - def add_backend(self, backend: str): - """Add a backend by calling the _registry.add_backend.""" - self._registry.add_backend(backend) - - def register_rewriter(self, - func_name: str, - backend: str = Backend.DEFAULT.value, - **kwargs): + def register_rewriter( + self, + func_name: str, + backend: str = Backend.DEFAULT.value, + ir: IR = IR.DEFAULT, + extra_checkers: Optional[Union[Checker, List[Checker]]] = None, + **kwargs): """The interface of function rewriter decorator. Args: func_name (str): The function name/path to rewrite. - backend (str): The inference engine name. + backend (str): The rewriter will be activated on which backend. + ir (IR): The rewriter will be activated on which IR. + extra_checkers (Checker | List[Checker] | None): Other requirements + defined by Checker. + Returns: Callable: The process of registering function. """ - return self._registry.register_object(func_name, backend, **kwargs) + return self._registry.register_object(func_name, backend, ir, + extra_checkers, **kwargs) - def enter(self, - cfg: Dict = dict(), - backend: str = Backend.DEFAULT.value, - **kwargs): + def enter(self, cfg: Dict = dict(), env: Dict = dict(), **kwargs): """The implementation of function rewrite.""" # Get current records - functions_records = self._registry.get_records(backend) + functions_records = self._registry.get_records(env) self._origin_functions = list() self._additional_functions = list() diff --git a/mmdeploy/core/rewriters/module_rewriter.py b/mmdeploy/core/rewriters/module_rewriter.py index 43720443c6..d0961809a0 100644 --- a/mmdeploy/core/rewriters/module_rewriter.py +++ b/mmdeploy/core/rewriters/module_rewriter.py @@ -1,11 +1,13 @@ # Copyright (c) OpenMMLab. All rights reserved. import inspect +from typing import Dict, List, Optional, Union import mmcv from torch import nn -from mmdeploy.utils.constants import Backend -from .rewriter_utils import RewriterRegistry, eval_with_import +from mmdeploy.utils.constants import IR, Backend +from .rewriter_utils import (Checker, RewriterRegistry, collect_env, + eval_with_import) class ModuleRewriter: @@ -26,29 +28,33 @@ class ModuleRewriter: def __init__(self): self._registry = RewriterRegistry() - def add_backend(self, backend: str): - """Add a backend by calling the _registry.add_backend.""" - self._registry.add_backend(backend) - - def register_rewrite_module(self, - module_type: str, - backend: str = Backend.DEFAULT.value, - **kwargs): + def register_rewrite_module( + self, + module_type: str, + backend: str = Backend.DEFAULT.value, + ir: IR = IR.DEFAULT, + extra_checkers: Optional[Union[Checker, List[Checker]]] = None, + **kwargs): """The interface of module rewriter decorator. Args: module_type (str): The module type name to rewrite. - backend (str): The inference engine name. + backend (str): The rewriter will be activated on which backend. + ir (IR): The rewriter will be activated on which IR. + extra_checkers (Checker | List[Checker] | None): Other requirements + defined by Checker. Returns: - nn.Module: THe rewritten model. + nn.Module: The rewritten model. """ - return self._registry.register_object(module_type, backend, **kwargs) + return self._registry.register_object(module_type, backend, ir, + extra_checkers, **kwargs) def patch_model(self, model: nn.Module, cfg: mmcv.Config, backend: str = Backend.DEFAULT.value, + ir: IR = IR.DEFAULT, recursive: bool = True, **kwargs) -> nn.Module: """Replace the models that was registered. @@ -57,6 +63,7 @@ def patch_model(self, model (torch.nn.Module): The model to patch. cfg (Dict): Config dictionary of deployment. backend (str): The inference engine name. + ir (IR): The intermeditate representation name. recursive (bool): The flag to enable recursive patching. Returns: @@ -67,7 +74,9 @@ def patch_model(self, >>> patched_model = patch_model(model, cfg=deploy_cfg, >>> backend=backend) """ - self._collect_record(backend) + # TODO: Make the type of parameter backend to Backend + env = collect_env(Backend.get(backend), ir) + self._collect_record(env) return self._replace_module(model, cfg, recursive, **kwargs) def _replace_one_module(self, module, cfg, **kwargs): @@ -103,9 +112,9 @@ def _replace_module_impl(model, cfg, **kwargs): return _replace_module_impl(model, cfg, **kwargs) - def _collect_record(self, backend: str): + def _collect_record(self, env: Dict): """Collect models in registry.""" self._records = {} - records = self._registry.get_records(backend) + records = self._registry.get_records(env) for name, kwargs in records: self._records[eval_with_import(name)] = kwargs diff --git a/mmdeploy/core/rewriters/rewriter_manager.py b/mmdeploy/core/rewriters/rewriter_manager.py index df7e82703d..de3acaffd2 100644 --- a/mmdeploy/core/rewriters/rewriter_manager.py +++ b/mmdeploy/core/rewriters/rewriter_manager.py @@ -4,9 +4,10 @@ import mmcv import torch.nn as nn -from mmdeploy.utils.constants import Backend +from mmdeploy.utils.constants import IR, Backend from .function_rewriter import FunctionRewriter from .module_rewriter import ModuleRewriter +from .rewriter_utils import collect_env from .symbolic_rewriter import SymbolicRewriter @@ -18,20 +19,8 @@ def __init__(self): self.function_rewriter = FunctionRewriter() self.symbolic_rewriter = SymbolicRewriter() - def add_backend(self, backend: str): - """Add backend to all rewriters. - - Args: - backend (str): The backend to support. - """ - self.module_rewriter.add_backend(backend) - self.function_rewriter.add_backend(backend) - self.symbolic_rewriter.add_backend(backend) - REWRITER_MANAGER = RewriterManager() -for backend in Backend: - REWRITER_MANAGER.add_backend(backend.value) MODULE_REWRITER = REWRITER_MANAGER.module_rewriter FUNCTION_REWRITER = REWRITER_MANAGER.function_rewriter @@ -41,6 +30,7 @@ def add_backend(self, backend: str): def patch_model(model: nn.Module, cfg: mmcv.Config, backend: str = Backend.DEFAULT.value, + ir: IR = IR.DEFAULT, recursive: bool = True, **kwargs) -> nn.Module: """Patch the model, replace the modules that can be rewritten. Note that @@ -50,6 +40,7 @@ def patch_model(model: nn.Module, model (torch.nn.Module): The model to patch. cfg (Dict): Config dictionary of deployment. backend (str): The inference engine name. + ir (IR): The intermeditate representation name. recursive (bool): The flag to enable recursive patching. Returns: @@ -59,7 +50,7 @@ def patch_model(model: nn.Module, >>> from mmdeploy.core import patch_model >>> patched_model = patch_model(model, cfg=deploy_cfg, backend=backend) """ - return MODULE_REWRITER.patch_model(model, cfg, backend, recursive, + return MODULE_REWRITER.patch_model(model, cfg, backend, ir, recursive, **kwargs) @@ -71,6 +62,7 @@ class RewriterContext: Args: cfg (Dict): Config dictionary of deployment. backend (str): The inference engine name. + ir (IR): The intermeditate representation name. rewrite_manager (RewriterManager): An RewriteManager that consists of several rewriters @@ -84,20 +76,19 @@ class RewriterContext: def __init__(self, cfg: Dict = dict(), backend: str = Backend.DEFAULT.value, + ir: IR = IR.DEFAULT, rewriter_manager: RewriterManager = REWRITER_MANAGER, **kwargs): self._cfg = cfg - self._backend = backend self._kwargs = kwargs self._rewriter_manager = rewriter_manager + self._env = collect_env(Backend.get(backend), ir) def enter(self): """Call the enter() of rewriters.""" - self._rewriter_manager.function_rewriter.enter(self._cfg, - self._backend, + self._rewriter_manager.function_rewriter.enter(self._cfg, self._env, **self._kwargs) - self._rewriter_manager.symbolic_rewriter.enter(self._cfg, - self._backend, + self._rewriter_manager.symbolic_rewriter.enter(self._cfg, self._env, **self._kwargs) def exit(self): diff --git a/mmdeploy/core/rewriters/rewriter_utils.py b/mmdeploy/core/rewriters/rewriter_utils.py index 701078144a..a80fd84738 100644 --- a/mmdeploy/core/rewriters/rewriter_utils.py +++ b/mmdeploy/core/rewriters/rewriter_utils.py @@ -1,8 +1,11 @@ # Copyright (c) OpenMMLab. All rights reserved. import inspect -from typing import Any, Callable, Dict, List, Optional, Tuple +import warnings +from abc import ABCMeta, abstractmethod +from typing import Any, Callable, Dict, List, Optional, Tuple, Union -from mmdeploy.utils.constants import Backend +import mmdeploy +from mmdeploy.utils.constants import IR, Backend def eval_with_import(path: str) -> Any: @@ -56,6 +59,127 @@ def import_function(path: str) -> Tuple[Callable, Optional[type]]: return obj, None +def collect_env(backend: Backend, ir: IR, **kwargs) -> Dict: + """Collect current environment information, including backend, ir, codebase + version, etc. Rewriters will be checked according to env infos. + + Args: + backend (Backend): Current backend. + ir (IR): Current IR. + + Returns: + Dict: Record the value of Backend and IR as well as the versions of + libraries. + """ + from mmdeploy.utils import get_backend_version, get_codebase_version + env = dict(backend=backend, ir=ir) + env['mmdeploy'] = mmdeploy.__version__ + env.update(get_backend_version()) + env.update(get_codebase_version()) + env.update(kwargs) + return env + + +class Checker(metaclass=ABCMeta): + """The interface for checking whether a rewriter is valid.""" + + def __init__(self): + pass + + @abstractmethod + def check(self, env: Dict) -> bool: + """Check the if the rewriter is valid according to environment. + + Args: + env (Dict): The backend, IR info and version info. + """ + pass + + +class BackendChecker(Checker): + """Checker that determines which backend the rewriter must run on. + + Args: + required_backend (Backend): The rewriter will be activated on + which backend. + """ + + def __init__(self, required_backend: Backend): + super().__init__() + self.required_backend = required_backend + + def check(self, env: Dict) -> bool: + """Check the if the rewriter is valid according to backend. + + Args: + env (Dict): The backend, IR info and version info. + """ + return env['backend'] == self.required_backend + + +class IRChecker(Checker): + """Checker that determines which IR the rewriter must run on. + + Args: + required_ir (IR): The rewriter will be activated on which IR. + """ + + def __init__(self, required_ir: IR): + super().__init__() + self.required_ir = required_ir + + def check(self, env: Dict) -> bool: + """Check the if the rewriter is valid according to IR. + + Args: + env (Dict): The backend, IR info and version info. + """ + return env['ir'] == self.required_ir + + +class LibVersionChecker(Checker): + """Checker that determines which IR the rewriter must run on. + + Args: + lib (str): The name of library. + min_version (str | None): The rewriter should no lower than which + version. Default to `None`. + max_version (str | None): The rewriter should no greater than which + version. Default to `None`. + """ + + def __init__(self, + lib: str, + min_version: Optional[str] = None, + max_version: Optional[str] = None): + super().__init__() + self.lib = lib + self.min_version = min_version + self.max_version = max_version + + def check(self, env: Dict) -> bool: + """Check the if the rewriter is valid according to library version. + + Args: + env (Dict): The backend, IR info and version info. + """ + # If the library has not been installed + if env[self.lib] is None: + return False + + from packaging import version + valid = True + # The version should no less than min version and no greater than + # max version. + if self.min_version is not None: + if version.parse(env[self.lib]) < version.parse(self.min_version): + valid = False + if self.max_version is not None: + if version.parse(env[self.lib]) > version.parse(self.max_version): + valid = False + return valid + + class RewriterRegistry: """A registry that recoreds rewrite objects. @@ -75,58 +199,128 @@ class RewriterRegistry: >>> records = FUNCTION_REGISTRY.get_record("default") """ - # TODO: replace backend string with "Backend" constant def __init__(self): self._rewrite_records = dict() - self.add_backend(Backend.DEFAULT.value) - - def _check_backend(self, backend: str): - """Check if a backend has been supported.""" - if backend not in self._rewrite_records: - raise Exception('Backend is not supported by registry.') - - def add_backend(self, backend: str): - """Add a backend dictionary.""" - if backend not in self._rewrite_records: - self._rewrite_records[backend] = dict() - - def get_records(self, backend: str) -> List: - """Get all registered records in record table.""" - self._check_backend(backend) - - if backend != Backend.DEFAULT.value: - # Update dict A with dict B. - # Then convert the result dict to a list, while keeping the order - # of A and B: the elements only belong to B should alwarys come - # after the elements only belong to A. - # The complexity is O(n + m). - dict_a = self._rewrite_records[Backend.DEFAULT.value] - dict_b = self._rewrite_records[backend] - records = [] - for k, v in dict_a.items(): - if k in dict_b: - records.append((k, dict_b[k])) + + def get_records(self, env: Dict) -> List: + """Get all registered records that are valid in the given environment + from record table. + + If the backend and IR of rewriter are set to 'default', then the + rewriter is regarded as default rewriter. The default rewriter will be + activated only when all other rewriters are not valid. If there are + multiple rewriters are valid (except default rewriter), we will + activate the first one (The order is determined by the time when + rewriters are loaded). + + Args: + env (dict): Environment dictionary that includes backend, IR, + codebase version, etc. + + Returns: + List: A list that includes valid records. + """ + default_records = list() + records = list() + + for origin_function, rewriter_records in self._rewrite_records.items(): + default_rewriter = None + final_rewriter = None + for record in rewriter_records: + # Get the checkers of current rewriter + checkers: List[Checker] = record['_checkers'] + + # Check if the rewriter is default rewriter + if len(checkers) == 0: + # Process the default rewriter exceptionally + if default_rewriter is None: + default_rewriter = record + else: + warnings.warn( + 'Detect multiple valid rewriters for' + f'{origin_function}, use the first rewriter.') else: - records.append((k, v)) - for k, v in dict_b.items(): - if k not in dict_a: - records.append((k, v)) - else: - records = list( - self._rewrite_records[Backend.DEFAULT.value].items()) - return records - - def _register(self, name: str, backend: str, **kwargs): + # Check if the checker is valid. + # The checker is valid only if all the checks are passed + valid = True + for checker in checkers: + if not checker.check(env): + valid = False + break + + if valid: + # Check if there are multiple valid rewriters + if final_rewriter is not None: + warnings.warn( + 'Detect multiple valid rewriters for' + f'{origin_function}, use the first rewriter.') + else: + final_rewriter = record + + # Append final rewriter. + # If there is no valid rewriter, try not apply default rewriter + if final_rewriter is not None: + records.append((origin_function, final_rewriter)) + elif default_rewriter is not None: + default_records.append((origin_function, default_rewriter)) + + # Make the default records como to the front of list because we may + # want the non-default records to override them. + return default_records + records + + def _register(self, name: str, backend: Backend, ir: IR, + extra_checkers: List[Checker], **kwargs): """The implementation of register.""" - self._check_backend(backend) - self._rewrite_records[backend][name] = kwargs - def register_object(self, name: str, backend: str, **kwargs) -> Callable: - """The decorator to register an object.""" - self._check_backend(backend) + # Merge checkers to kwargs + record_dict = kwargs + + # Try to create a checker according to 'backend' field + if backend != Backend.DEFAULT: + extra_checkers.append(BackendChecker(backend)) + + # Try to create a checker according to 'ir' field + if ir != IR.DEFAULT: + extra_checkers.append(IRChecker(ir)) + + record_dict['_checkers'] = extra_checkers + + # There may be multiple rewriters of a function/module. We use a list + # to store the rewriters of a function/module. + if name not in self._rewrite_records: + self._rewrite_records[name] = list() + self._rewrite_records[name].append(record_dict) + + def register_object(self, + name: str, + backend: str, + ir: IR, + extra_checkers: Optional[Union[Checker, + List[Checker]]] = None, + **kwargs) -> Callable: + """The decorator to register an object. + + Args: + name (str): The import path to access the function/module. + backend (str): The rewriter will be activated on which backend. + ir (IR): The rewriter will be activated on which ir. + extra_chekcers (None | Checker | List[Checker]): Other requirements + for the rewriters. Default to `None`. + + Returns: + Callable: The decorator. + """ + + if extra_checkers is None: + extra_checkers = [] + elif isinstance(extra_checkers, Checker): + extra_checkers = [extra_checkers] + + backend = Backend.get(backend) def decorator(object): - self._register(name, backend, _object=object, **kwargs) + self._register( + name, backend, ir, extra_checkers, _object=object, **kwargs) return object return decorator diff --git a/mmdeploy/core/rewriters/symbolic_rewriter.py b/mmdeploy/core/rewriters/symbolic_rewriter.py index c9c16d071d..dd47cd8d58 100644 --- a/mmdeploy/core/rewriters/symbolic_rewriter.py +++ b/mmdeploy/core/rewriters/symbolic_rewriter.py @@ -1,13 +1,14 @@ # Copyright (c) OpenMMLab. All rights reserved. -from typing import Callable, Dict, Optional, Sequence +from typing import Callable, Dict, List, Optional, Sequence, Union from torch.autograd import Function from torch.onnx.symbolic_helper import parse_args from torch.onnx.symbolic_registry import _registry as pytorch_registry from torch.onnx.symbolic_registry import register_op -from mmdeploy.utils import Backend, get_root_logger -from .rewriter_utils import ContextCaller, RewriterRegistry, eval_with_import +from mmdeploy.utils import IR, Backend, get_root_logger +from .rewriter_utils import (Checker, ContextCaller, RewriterRegistry, + eval_with_import) class SymbolicRewriter: @@ -35,25 +36,27 @@ class SymbolicRewriter: def __init__(self) -> None: self._registry = RewriterRegistry() - def add_backend(self, backend: str): - """Add a backend by calling the _registry.add_backend.""" - self._registry.add_backend(backend) - def register_symbolic(self, func_name: str, backend: str = Backend.DEFAULT.value, is_pytorch: bool = False, arg_descriptors: Optional[Sequence[str]] = None, + ir: IR = IR.DEFAULT, + extra_checkers: Optional[Union[ + Checker, List[Checker]]] = None, **kwargs) -> Callable: """The decorator of the custom symbolic. Args: func_name (str): The function name/path to override the symbolic. - backend (str): The inference engine name. + backend (str): The rewriter will be activated on which backend. is_pytorch (bool): Enable this flag if func_name is the name of \ a pytorch builtin function. arg_descriptors (Sequence[str]): The argument descriptors of the \ symbol. + ir (IR): The rewriter will be activated on which IR. + extra_checkers (Checker | List[Checker] | None): Other requirements + defined by Checker. Returns: Callable: The process of registered symbolic. @@ -61,18 +64,20 @@ def register_symbolic(self, return self._registry.register_object( func_name, backend, + ir, + extra_checkers, is_pytorch=is_pytorch, arg_descriptors=arg_descriptors, **kwargs) def enter(self, cfg: Dict = dict(), - backend: str = Backend.DEFAULT.value, + env: Dict = dict(), opset: int = 11, **kwargs): """The implementation of symbolic register.""" # Get current records - symbolic_records = self._registry.get_records(backend) + symbolic_records = self._registry.get_records(env) self._pytorch_symbolic = list() self._extra_symbolic = list() diff --git a/mmdeploy/pytorch/functions/interpolate.py b/mmdeploy/pytorch/functions/interpolate.py index 2d9632ea30..2b619b9535 100644 --- a/mmdeploy/pytorch/functions/interpolate.py +++ b/mmdeploy/pytorch/functions/interpolate.py @@ -90,7 +90,8 @@ def forward(g, input, scale_factor, align_corners): 'tensor. Which is not available for custom ops. Computed scale' '_factor might be the right way to get final shape.') scale_factor = [ - s_out / s_in for s_out, s_in in zip(size, input_size[2:]) + float(s_out / s_in) + for s_out, s_in in zip(size, input_size[2:]) ] return BicubicInterpolate.apply(input, scale_factor, align_corners) else: diff --git a/mmdeploy/pytorch/ops/__init__.py b/mmdeploy/pytorch/ops/__init__.py index 48f9c90137..0608aadf79 100644 --- a/mmdeploy/pytorch/ops/__init__.py +++ b/mmdeploy/pytorch/ops/__init__.py @@ -3,6 +3,7 @@ adaptive_avg_pool2d__default, adaptive_avg_pool3d__default) from .grid_sampler import grid_sampler__default +from .hardsigmoid import hardsigmoid__default from .instance_norm import instance_norm__tensorrt from .lstm import generic_rnn__ncnn from .squeeze import squeeze__default @@ -10,5 +11,6 @@ __all__ = [ 'adaptive_avg_pool1d__default', 'adaptive_avg_pool2d__default', 'adaptive_avg_pool3d__default', 'grid_sampler__default', - 'instance_norm__tensorrt', 'generic_rnn__ncnn', 'squeeze__default' + 'hardsigmoid__default', 'instance_norm__tensorrt', 'generic_rnn__ncnn', + 'squeeze__default' ] diff --git a/mmdeploy/pytorch/ops/hardsigmoid.py b/mmdeploy/pytorch/ops/hardsigmoid.py new file mode 100644 index 0000000000..a4d14173ed --- /dev/null +++ b/mmdeploy/pytorch/ops/hardsigmoid.py @@ -0,0 +1,12 @@ +# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved +# Modified from: +# https://github.com/pytorch/pytorch/blob/9ade03959392e5a90b74261012de1d806cab2253/torch/onnx/symbolic_opset9.py +from mmdeploy.core import SYMBOLIC_REWRITER + + +@SYMBOLIC_REWRITER.register_symbolic( + 'hardsigmoid', is_pytorch=True, arg_descriptors=['v']) +def hardsigmoid__default(ctx, g, self): + """Support export hardsigmoid This rewrite enable export hardsigmoid in + torch<=1.8.2.""" + return g.op('HardSigmoid', self, alpha_f=1 / 6) diff --git a/mmdeploy/utils/__init__.py b/mmdeploy/utils/__init__.py index 03543f9d5f..4847ba7b09 100644 --- a/mmdeploy/utils/__init__.py +++ b/mmdeploy/utils/__init__.py @@ -6,9 +6,10 @@ get_model_inputs, get_onnx_config, get_partition_config, get_task_type, is_dynamic_batch, is_dynamic_shape, load_config) -from .constants import SDK_TASK_MAP, Backend, Codebase, Task +from .constants import IR, SDK_TASK_MAP, Backend, Codebase, Task from .device import parse_cuda_device_id, parse_device_id -from .utils import get_root_logger, target_wrapper +from .env import get_backend_version, get_codebase_version, get_library_version +from .utils import get_file_path, get_root_logger, target_wrapper __all__ = [ 'is_dynamic_batch', 'is_dynamic_shape', 'get_task_type', 'get_codebase', @@ -18,5 +19,6 @@ 'get_model_inputs', 'cfg_apply_marks', 'get_input_shape', 'parse_device_id', 'parse_cuda_device_id', 'get_codebase_config', 'get_backend_config', 'get_root_logger', 'get_dynamic_axes', - 'target_wrapper', 'SDK_TASK_MAP' + 'target_wrapper', 'SDK_TASK_MAP', 'get_library_version', + 'get_codebase_version', 'get_backend_version', 'IR', 'get_file_path' ] diff --git a/mmdeploy/utils/config_utils.py b/mmdeploy/utils/config_utils.py index f1842a6672..3aab4e29e4 100644 --- a/mmdeploy/utils/config_utils.py +++ b/mmdeploy/utils/config_utils.py @@ -4,6 +4,7 @@ import mmcv from .constants import Backend, Codebase, Task +from .utils import deprecate def load_config(*args) -> List[mmcv.Config]: @@ -126,6 +127,7 @@ def get_ir_config(deploy_cfg: Union[str, mmcv.Config]) -> Dict: return ir_config +@deprecate(dst_obj=get_ir_config) def get_onnx_config(deploy_cfg: Union[str, mmcv.Config]) -> Dict: """Get the onnx parameters in export() from config. @@ -135,7 +137,6 @@ def get_onnx_config(deploy_cfg: Union[str, mmcv.Config]) -> Dict: Returns: Dict: The config dictionary of onnx parameters """ - onnx_config = get_ir_config(deploy_cfg=deploy_cfg) ir_type = onnx_config.get('type', None) assert ir_type is None or ir_type == 'onnx', 'Expect IR type is ONNX,'\ @@ -193,6 +194,10 @@ def is_dynamic_shape(deploy_cfg: Union[str, mmcv.Config], bool: Is config set dynamic shape (axis 2 and 3). """ + # Always dynamic for exporting torchscript + if get_backend(deploy_cfg) == Backend.TORCHSCRIPT: + return True + deploy_cfg = load_config(deploy_cfg)[0] ir_config = get_ir_config(deploy_cfg) @@ -353,18 +358,21 @@ def get_dynamic_axes( Dictionary with dynamic axes. """ deploy_cfg = load_config(deploy_cfg)[0] + ir_config = get_ir_config(deploy_cfg) + + # TODO onnx will be deprecated in the future onnx_config = deploy_cfg.get('onnx_config', None) - if onnx_config is None: + if onnx_config is None and ir_config == {}: raise KeyError( 'Field \'onnx_config\' was not found in \'deploy_cfg\'.') - dynamic_axes = onnx_config.get('dynamic_axes', None) + dynamic_axes = ir_config.get('dynamic_axes', None) if dynamic_axes and not isinstance(dynamic_axes, Dict): if axes_names is None: axes_names = [] - input_names = onnx_config.get('input_names', None) + input_names = ir_config.get('input_names', None) if input_names: axes_names += input_names - output_names = onnx_config.get('output_names', None) + output_names = ir_config.get('output_names', None) if output_names: axes_names += output_names if not axes_names: diff --git a/mmdeploy/utils/constants.py b/mmdeploy/utils/constants.py index da07cb28e7..c36e610fea 100644 --- a/mmdeploy/utils/constants.py +++ b/mmdeploy/utils/constants.py @@ -37,6 +37,13 @@ class Codebase(AdvancedEnum): MMPOSE = 'mmpose' +class IR(AdvancedEnum): + """Define intermediate representation enumerations.""" + ONNX = 'onnx' + TORCHSCRIPT = 'torchscript' + DEFAULT = 'default' + + class Backend(AdvancedEnum): """Define backend enumerations.""" PYTORCH = 'pytorch' @@ -46,6 +53,7 @@ class Backend(AdvancedEnum): NCNN = 'ncnn' OPENVINO = 'openvino' SDK = 'sdk' + TORCHSCRIPT = 'torchscript' DEFAULT = 'default' diff --git a/mmdeploy/utils/env.py b/mmdeploy/utils/env.py new file mode 100644 index 0000000000..8cc2cbd3d5 --- /dev/null +++ b/mmdeploy/utils/env.py @@ -0,0 +1,49 @@ +# Copyright (c) OpenMMLab. All rights reserved. +import importlib + +from mmdeploy.utils import Codebase + + +def get_library_version(lib): + """Try to get the version of a library if it has been installed. + + Args: + lib (str): The name of library. + + Returns: + None | str: If the library has been installed, return version. + """ + try: + lib = importlib.import_module(lib) + except Exception: + version = None + else: + version = lib.__version__ + + return version + + +def get_codebase_version(): + """Get the version dictionary of all supported codebases. + + Returns: + Dict: The name and the version of supported codebases. + """ + version_dict = dict() + for enum in Codebase: + codebase = enum.value + version_dict[codebase] = get_library_version(codebase) + return version_dict + + +def get_backend_version(): + """Get the version dictionary of some supported backend. + + Returns: + Dict: The name and the version of some supported backend. + """ + backend_library_list = ['tensorrt', 'onnxruntime', 'ncnn'] + version_dict = dict() + for backend in backend_library_list: + version_dict[backend] = get_library_version(backend) + return version_dict diff --git a/mmdeploy/utils/export_info.py b/mmdeploy/utils/export_info.py index 8c5466f892..2eaf38da1f 100644 --- a/mmdeploy/utils/export_info.py +++ b/mmdeploy/utils/export_info.py @@ -1,14 +1,14 @@ # Copyright (c) OpenMMLab. All rights reserved. import importlib +import re from typing import Dict, List, Tuple, Union import mmcv from mmdeploy.apis import build_task_processor from mmdeploy.utils import (Backend, Task, get_backend, get_codebase, - get_common_config, get_onnx_config, - get_root_logger, get_task_type, is_dynamic_batch, - load_config) + get_common_config, get_ir_config, get_root_logger, + get_task_type, is_dynamic_batch, load_config) from mmdeploy.utils.constants import SDK_TASK_MAP as task_map @@ -89,12 +89,25 @@ def get_models(deploy_cfg: Union[str, mmcv.Config], """ name, _ = get_model_name_customs(deploy_cfg, model_cfg, work_dir) precision = 'FP32' - onnx_name = get_onnx_config(deploy_cfg)['save_file'] - net = onnx_name + ir_name = get_ir_config(deploy_cfg)['save_file'] + net = ir_name weights = '' backend = get_backend(deploy_cfg=deploy_cfg) + + def replace_suffix(file_name: str, dst_suffix: str) -> str: + """Replace the suffix to the destination one. + + Args: + file_name (str): The file name to be operated. + dst_suffix (str): The destination suffix. + + Return: + str: The file name of which the suffix has been replaced. + """ + return re.sub(r'\.[a-z]+', dst_suffix, file_name) + if backend == Backend.TENSORRT: - net = onnx_name.replace('.onnx', '.engine') + net = replace_suffix(ir_name, '.engine') common_cfg = get_common_config(deploy_cfg) fp16_mode = common_cfg.get('fp16_mode', False) int8_mode = common_cfg.get('int8_mode', False) @@ -104,15 +117,15 @@ def get_models(deploy_cfg: Union[str, mmcv.Config], precision = 'INT8' elif backend == Backend.PPLNN: precision = 'FP16' - weights = onnx_name.replace('.onnx', '.json') - net = onnx_name + weights = replace_suffix(ir_name, '.json') + net = ir_name elif backend == Backend.OPENVINO: - net = onnx_name.replace('.onnx', '.xml') - weights = onnx_name.replace('.onnx', '.bin') + net = replace_suffix(ir_name, '.xml') + weights = replace_suffix(ir_name, '.bin') elif backend == Backend.NCNN: - net = onnx_name.replace('.onnx', '.param') - weights = onnx_name.replace('.onnx', '.bin') - elif backend == Backend.ONNXRUNTIME: + net = replace_suffix(ir_name, '.param') + weights = replace_suffix(ir_name, '.bin') + elif backend in [Backend.ONNXRUNTIME, Backend.TORCHSCRIPT]: pass else: raise NotImplementedError(f'Not supported backend: {backend.value}.') @@ -149,8 +162,8 @@ def get_inference_info(deploy_cfg: mmcv.Config, model_cfg: mmcv.Config, module = 'Net' input = ['prep_output'] output = ['infer_output'] - onnx_config = get_onnx_config(deploy_cfg) - input_names = onnx_config.get('input_names', None) + ir_config = get_ir_config(deploy_cfg) + input_names = ir_config.get('input_names', None) input_name = input_names[0] if input_names else 'input' input_map = dict(img=input_name) return dict( @@ -320,14 +333,14 @@ def get_detail(deploy_cfg: mmcv.Config, model_cfg: mmcv.Config, codebase['pth'] = pth codebase['config'] = model_cfg.filename codebase_config = deploy_cfg.get('codebase_config', dict()) - onnx_config = get_onnx_config(deploy_cfg) + ir_config = get_ir_config(deploy_cfg) backend_config = deploy_cfg.get('backend_config', dict()) calib_config = deploy_cfg.get('calib_config', dict()) return dict( version=version, codebase=codebase, codebase_config=codebase_config, - onnx_config=onnx_config, + onnx_config=ir_config, backend_config=backend_config, calib_config=calib_config) diff --git a/mmdeploy/utils/test.py b/mmdeploy/utils/test.py index 3fd7457f0c..c912f14821 100644 --- a/mmdeploy/utils/test.py +++ b/mmdeploy/utils/test.py @@ -14,7 +14,7 @@ import mmdeploy.codebase # noqa: F401,F403 from mmdeploy.core import RewriterContext, patch_model -from mmdeploy.utils import (Backend, get_backend, get_dynamic_axes, +from mmdeploy.utils import (IR, Backend, get_backend, get_dynamic_axes, get_ir_config, get_onnx_config) @@ -93,6 +93,8 @@ def check_backend(backend: Backend, require_plugin: bool = False): from mmdeploy.apis.ncnn import is_plugin_available elif backend == Backend.OPENVINO: from mmdeploy.apis.openvino import is_available + elif backend == Backend.TORCHSCRIPT: + from mmdeploy.backend.torchscript import ops_available as is_available else: warnings.warn('The backend checker is not available') return @@ -375,14 +377,40 @@ def get_onnx_model(wrapped_model: nn.Module, return onnx_file_path -def get_backend_outputs(onnx_file_path: str, +def get_ts_model(wrapped_model: nn.Module, + model_inputs: Dict[str, Union[Tuple, List, torch.Tensor]], + deploy_cfg: mmcv.Config) -> str: + """To get path to onnx model after export. + + Args: + wrapped_model (nn.Module): The input model. + model_inputs (dict): Inputs for model. + deploy_cfg (mmcv.Config): Deployment config. + + Returns: + str: The path to the TorchScript model file. + """ + ir_file_path = tempfile.NamedTemporaryFile(suffix='.pt').name + backend = get_backend(deploy_cfg) + patched_model = patch_model( + wrapped_model, cfg=deploy_cfg, backend=backend.value) + + from mmdeploy.apis.pytorch2torchscript import torch2torchscript_impl + torch2torchscript_impl( + patched_model, [v for _, v in model_inputs.items()], + deploy_cfg=deploy_cfg, + output_file=ir_file_path) + return ir_file_path + + +def get_backend_outputs(ir_file_path: str, model_inputs: Dict[str, Union[Tuple, List, torch.Tensor]], deploy_cfg: mmcv.Config) -> Union[Any, None]: """To get backend outputs of model. Args: - onnx_file_path (str): The path to the ONNX file. + ir_file_path (str): The path to the IR file. model_inputs (dict): Inputs for model. deploy_cfg (mmcv.Config): Deployment config. @@ -408,7 +436,7 @@ def get_backend_outputs(onnx_file_path: str, trt_file_path, 0, deploy_cfg=deploy_cfg, - onnx_model=onnx_file_path) + onnx_model=ir_file_path) backend_files = [trt_file_path] for k, v in model_inputs.items(): model_inputs[k] = model_inputs[k].cuda() @@ -441,7 +469,7 @@ def get_backend_outputs(onnx_file_path: str, backend_feats[input_names[i]] = feature_list[i] else: backend_feats[str(i)] = feature_list[i] - backend_files = [onnx_file_path] + backend_files = [ir_file_path] device = 'cpu' elif backend == Backend.NCNN: import mmdeploy.apis.ncnn as ncnn_apis @@ -449,8 +477,8 @@ def get_backend_outputs(onnx_file_path: str, return None work_dir = tempfile.TemporaryDirectory().name param_path, bin_path = ncnn_apis.get_output_model_file( - onnx_file_path, work_dir) - ncnn_apis.onnx2ncnn(onnx_file_path, param_path, bin_path) + ir_file_path, work_dir) + ncnn_apis.onnx2ncnn(ir_file_path, param_path, bin_path) backend_files = [param_path, bin_path] backend_feats = flatten_model_inputs device = 'cpu' @@ -461,25 +489,34 @@ def get_backend_outputs(onnx_file_path: str, return None openvino_work_dir = tempfile.TemporaryDirectory().name openvino_file_path = openvino_apis.get_output_model_file( - onnx_file_path, openvino_work_dir) + ir_file_path, openvino_work_dir) input_info = { name: value.shape for name, value in flatten_model_inputs.items() } - openvino_apis.onnx2openvino(input_info, output_names, onnx_file_path, + openvino_apis.onnx2openvino(input_info, output_names, ir_file_path, openvino_work_dir) backend_files = [openvino_file_path] backend_feats = flatten_model_inputs device = 'cpu' + elif backend == Backend.DEFAULT: return None + elif backend == Backend.TORCHSCRIPT: + backend_files = [ir_file_path] + device = 'cpu' + backend_feats = [v for _, v in model_inputs.items()] else: raise NotImplementedError( f'Unimplemented backend type: {backend.value}') from mmdeploy.codebase.base import BaseBackendModel - backend_model = BaseBackendModel._build_wrapper(backend, backend_files, - device, output_names) + backend_model = BaseBackendModel._build_wrapper( + backend, + backend_files, + device, + input_names=input_names, + output_names=output_names) with torch.no_grad(): backend_outputs = backend_model(backend_feats) backend_outputs = backend_model.output_to_list(backend_outputs) @@ -511,11 +548,15 @@ def get_rewrite_outputs(wrapped_model: nn.Module, cfg=deploy_cfg, backend=backend.value, opset=11), torch.no_grad(): ctx_outputs = wrapped_model(**model_inputs) - onnx_file_path = get_onnx_model(wrapped_model, model_inputs, deploy_cfg) + ir_type = get_ir_config(deploy_cfg).get('type', None) + if ir_type == IR.TORCHSCRIPT.value: + ir_file_path = get_ts_model(wrapped_model, model_inputs, deploy_cfg) + else: # TODO onnx as default, make it strict when more IR types involved + ir_file_path = get_onnx_model(wrapped_model, model_inputs, deploy_cfg) backend_outputs = None if run_with_backend: - backend_outputs = get_backend_outputs(onnx_file_path, model_inputs, + backend_outputs = get_backend_outputs(ir_file_path, model_inputs, deploy_cfg) if backend_outputs is None: diff --git a/mmdeploy/utils/utils.py b/mmdeploy/utils/utils.py index 9917dd4770..10d6d02bd0 100644 --- a/mmdeploy/utils/utils.py +++ b/mmdeploy/utils/utils.py @@ -1,8 +1,10 @@ # Copyright (c) OpenMMLab. All rights reserved. +import glob import logging +import os import sys import traceback -from typing import Callable, Optional +from typing import Callable, Optional, Union import torch.multiprocessing as mp from mmcv.utils import get_logger @@ -56,3 +58,68 @@ def get_root_logger(log_file=None, log_level=logging.INFO) -> logging.Logger: name='mmdeploy', log_file=log_file, log_level=log_level) return logger + + +def deprecate(status: str = 'future', + dst_obj: Optional[Union[object, str]] = None, + msg: str = '', + *args, + **kwargs) -> None: + """Deprecate a function or a class. + + Args: + status (str, optional): The status of the function or class. + Defaults to future. + dst_obj (str, object, optional): The object that will replace + the original one. Defaults to None. + msg (str): Additional message to be printed. + + Examples: + >>> from math import ceil + >>> from mmdeploy.utils.utils import deprecate + >>> @deprecate(status='past', dst_obj=ceil, msg='') + >>> def my_ceil(num): + >>> num = num if(num==int(num)) else int(num) + 1 + >>> return num + """ + logger = get_root_logger() + + def _register(src_obj): + + def fun(*args, **kwargs): + if status == 'future': + logger.warning( + f'DeprecationWarning: {src_obj.__name__} will be ' + f'deprecated in the future. {msg}') + elif status == 'past': + assert dst_obj is not None, 'for deprecated object, there' + ' must be a destination object' + logger.warning( + f'DeprecationWarning: {src_obj.__name__} was deprecated,' + f' use {dst_obj.__name__} instead. {msg}') + else: + raise KeyError(f'Unexpected key {status}') + result = src_obj(*args, **kwargs) + return result + + return fun + + return _register + + +def get_file_path(prefix, candidates) -> str: + """Search for file in candidates. + + Args: + prefix (str): Prefix of the paths. + cancidates (str): Candidate paths + Returns: + str: file path or '' if not found + """ + for candidate in candidates: + wildcard = os.path.abspath(os.path.join(prefix, candidate)) + paths = glob.glob(wildcard) + if paths: + lib_path = paths[0] + return lib_path + return '' diff --git a/mmdeploy/version.py b/mmdeploy/version.py index 106378f190..821f44df9c 100644 --- a/mmdeploy/version.py +++ b/mmdeploy/version.py @@ -1,7 +1,7 @@ # Copyright (c) OpenMMLab. All rights reserved. from typing import Tuple -__version__ = '0.2.0' +__version__ = '0.4.0' short_version = __version__ diff --git a/tests/test_apis/test_onnx2ncnn.py b/tests/test_apis/test_onnx2ncnn.py index 57199e37ff..8073b77548 100644 --- a/tests/test_apis/test_onnx2ncnn.py +++ b/tests/test_apis/test_onnx2ncnn.py @@ -1,3 +1,4 @@ +# Copyright (c) OpenMMLab. All rights reserved. import os.path as osp import tempfile diff --git a/tests/test_apis/test_torch2onnx.py b/tests/test_apis/test_torch2onnx.py index 16dc2b0d50..349a9c642a 100644 --- a/tests/test_apis/test_torch2onnx.py +++ b/tests/test_apis/test_torch2onnx.py @@ -53,9 +53,8 @@ def get_deploy_cfg(input_name, output_name, dynamic_axes): input_names=[input_name], output_names=[output_name], input_shape=None), - codebase_config=dict(type='mmedit', task=''), # useless - backend_config=dict(type='onnxruntime') # useless - )) + codebase_config=dict(type='mmedit', task=''), + backend_config=dict(type='onnxruntime'))) @pytest.mark.parametrize('input_name', [input_name]) diff --git a/tests/test_apis/test_torch2torchscript.py b/tests/test_apis/test_torch2torchscript.py new file mode 100644 index 0000000000..4bb1c5c998 --- /dev/null +++ b/tests/test_apis/test_torch2torchscript.py @@ -0,0 +1,87 @@ +# Copyright (c) OpenMMLab. All rights reserved. +import importlib +import os.path as osp +import tempfile + +import mmcv +import pytest + +from mmdeploy.apis import torch2torchscript +from mmdeploy.utils import IR, Backend +from mmdeploy.utils.test import get_random_name + +ts_file = tempfile.NamedTemporaryFile(suffix='.pt').name +input_name = get_random_name() +output_name = get_random_name() + + +def get_deploy_cfg(input_name, output_name): + return mmcv.Config( + dict( + ir_config=dict( + type=IR.TORCHSCRIPT.value, + input_names=[input_name], + output_names=[output_name], + input_shape=None), + codebase_config=dict(type='mmedit', task='SuperResolution'), + backend_config=dict(type=Backend.TORCHSCRIPT.value))) + + +def get_model_cfg(): + return mmcv.Config( + dict( + model=dict( + pretrained=None, + type='BasicRestorer', + generator=dict( + type='RRDBNet', + in_channels=3, + out_channels=3, + mid_channels=64, + num_blocks=23, + growth_channels=32), + pixel_loss=dict( + type='L1Loss', loss_weight=1.0, reduction='mean')), + test_cfg=dict(metrics='PSNR'), + test_pipeline=[ + dict( + type='LoadImageFromFile', + io_backend='disk', + key='lq', + flag='unchanged'), + dict( + type='LoadImageFromFile', + io_backend='disk', + key='gt', + flag='unchanged'), + dict(type='RescaleToZeroOne', keys=['lq', 'gt']), + dict( + type='Normalize', + keys=['lq', 'gt'], + mean=[0, 0, 0], + std=[1, 1, 1], + to_rgb=True), + dict( + type='Collect', + keys=['lq', 'gt'], + meta_keys=['lq_path', 'lq_path']), + dict(type='ImageToTensor', keys=['lq', 'gt']) + ])) + + +@pytest.mark.parametrize('input_name', [input_name]) +@pytest.mark.parametrize('output_name', [output_name]) +@pytest.mark.skipif( + not importlib.util.find_spec('mmedit'), reason='requires mmedit') +def test_torch2torchscript(input_name, output_name): + import numpy as np + deploy_cfg = get_deploy_cfg(input_name, output_name) + torch2torchscript( + np.random.rand(8, 8, 3), + '', + ts_file, + deploy_cfg, + model_cfg=get_model_cfg(), + device='cpu') + + assert osp.exists(ts_file) diff --git a/tests/test_backend/test_wrapper.py b/tests/test_backend/test_wrapper.py index f42d915131..b177a2ee58 100644 --- a/tests/test_backend/test_wrapper.py +++ b/tests/test_backend/test_wrapper.py @@ -10,6 +10,7 @@ from mmdeploy.utils.test import check_backend onnx_file = tempfile.NamedTemporaryFile(suffix='.onnx').name +ts_file = tempfile.NamedTemporaryFile(suffix='.pt').name test_img = torch.rand(1, 3, 8, 8) output_names = ['output'] input_names = ['input'] @@ -44,6 +45,18 @@ def generate_onnx_file(): dynamic_axes=None) +@pytest.fixture(autouse=True, scope='module') +def generate_torchscript_file(): + import mmcv + + from mmdeploy.apis import torch2torchscript_impl + deploy_cfg = mmcv.Config( + {'backend_config': dict(type=Backend.TORCHSCRIPT.value)}) + with torch.no_grad(): + torch2torchscript_impl(model, torch.rand(1, 3, 8, 8), deploy_cfg, + ts_file) + + def onnx2backend(backend, onnx_file): if backend == Backend.TENSORRT: from mmdeploy.backend.tensorrt import (create_trt_engine, @@ -107,6 +120,11 @@ def create_wrapper(backend, model_files): from mmdeploy.backend.openvino import OpenVINOWrapper openvino_model = OpenVINOWrapper(model_files, output_names) return openvino_model + elif backend == Backend.TORCHSCRIPT: + from mmdeploy.backend.torchscript import TorchscriptWrapper + torchscript_model = TorchscriptWrapper( + model_files, input_names=input_names, output_names=output_names) + return torchscript_model else: raise NotImplementedError(f'Unknown backend type: {backend.value}') @@ -134,20 +152,26 @@ def run_wrapper(backend, wrapper, input): results = wrapper({'input': input})['output'] results = results.detach().cpu() return results + elif backend == Backend.TORCHSCRIPT: + results = wrapper({'input': input})['output'] + return results else: raise NotImplementedError(f'Unknown backend type: {backend.value}') ALL_BACKEND = [ Backend.TENSORRT, Backend.ONNXRUNTIME, Backend.PPLNN, Backend.NCNN, - Backend.OPENVINO + Backend.OPENVINO, Backend.TORCHSCRIPT ] @pytest.mark.parametrize('backend', ALL_BACKEND) def test_wrapper(backend): check_backend(backend, True) - model_files = onnx2backend(backend, onnx_file) + if backend == Backend.TORCHSCRIPT: + model_files = ts_file + else: + model_files = onnx2backend(backend, onnx_file) assert model_files is not None wrapper = create_wrapper(backend, model_files) assert wrapper is not None diff --git a/tests/test_codebase/test_mmdet/test_mmdet_models.py b/tests/test_codebase/test_mmdet/test_mmdet_models.py index 9cfe6e83d9..f299e18641 100644 --- a/tests/test_codebase/test_mmdet/test_mmdet_models.py +++ b/tests/test_codebase/test_mmdet/test_mmdet_models.py @@ -11,6 +11,7 @@ from mmdeploy.codebase import import_codebase from mmdeploy.utils import Backend, Codebase +from mmdeploy.utils.config_utils import get_ir_config from mmdeploy.utils.test import (WrapModel, check_backend, get_model_outputs, get_rewrite_outputs) @@ -157,6 +158,31 @@ def get_single_roi_extractor(): return model +def get_gfl_head_model(): + test_cfg = mmcv.Config( + dict( + nms_pre=1000, + min_bbox_size=0, + score_thr=0.05, + nms=dict(type='nms', iou_threshold=0.6), + max_per_img=100)) + anchor_generator = dict( + type='AnchorGenerator', + scales_per_octave=1, + octave_base_scale=8, + ratios=[1.0], + strides=[8, 16, 32, 64, 128]) + from mmdet.models.dense_heads import GFLHead + model = GFLHead( + num_classes=3, + in_channels=256, + reg_max=3, + test_cfg=test_cfg, + anchor_generator=anchor_generator) + model.requires_grad_(False) + return model + + def test_focus_forward_ncnn(): backend_type = Backend.NCNN check_backend(backend_type) @@ -196,6 +222,7 @@ def test_l2norm_forward(backend_type): dict( backend_config=dict(type=backend_type.value), onnx_config=dict(input_shape=None))) + seed_everything(1234) feat = torch.rand(1, 16, s, s) model_outputs = [l2norm_neck.forward(feat)] wrapped_model = WrapModel(l2norm_neck, 'forward') @@ -349,6 +376,88 @@ def test_get_bboxes_of_rpn_head(backend_type: Backend): assert rewrite_outputs is not None +@pytest.mark.parametrize('backend_type', [Backend.ONNXRUNTIME]) +def test_get_bboxes_of_gfl_head(backend_type): + check_backend(backend_type) + head = get_gfl_head_model() + head.cpu().eval() + s = 4 + img_metas = [{ + 'scale_factor': np.ones(4), + 'pad_shape': (s, s, 3), + 'img_shape': (s, s, 3) + }] + output_names = ['dets'] + deploy_cfg = mmcv.Config( + dict( + backend_config=dict(type=backend_type.value), + onnx_config=dict(output_names=output_names, input_shape=None), + codebase_config=dict( + type='mmdet', + task='ObjectDetection', + model_type='ncnn_end2end', + post_processing=dict( + score_threshold=0.05, + iou_threshold=0.5, + max_output_boxes_per_class=200, + pre_top_k=5000, + keep_top_k=100, + background_label_id=-1, + )))) + + seed_everything(1234) + cls_score = [ + torch.rand(1, 3, pow(2, i), pow(2, i)) for i in range(5, 0, -1) + ] + seed_everything(5678) + bboxes = [torch.rand(1, 16, pow(2, i), pow(2, i)) for i in range(5, 0, -1)] + + # to get outputs of onnx model after rewrite + img_metas[0]['img_shape'] = torch.Tensor([s, s]) + wrapped_model = WrapModel( + head, 'get_bboxes', img_metas=img_metas, with_nms=True) + rewrite_inputs = { + 'cls_scores': cls_score, + 'bbox_preds': bboxes, + } + # do not run with ncnn backend + run_with_backend = False if backend_type in [Backend.NCNN] else True + rewrite_outputs, is_backend_output = get_rewrite_outputs( + wrapped_model=wrapped_model, + model_inputs=rewrite_inputs, + deploy_cfg=deploy_cfg, + run_with_backend=run_with_backend) + assert rewrite_outputs is not None + + +@pytest.mark.parametrize('backend_type', [Backend.ONNXRUNTIME]) +def test_forward_of_gfl_head(backend_type): + check_backend(backend_type) + head = get_gfl_head_model() + head.cpu().eval() + deploy_cfg = mmcv.Config( + dict( + backend_config=dict(type=backend_type.value), + onnx_config=dict(input_shape=None))) + feats = [torch.rand(1, 256, pow(2, i), pow(2, i)) for i in range(5, 0, -1)] + model_outputs = [head.forward(feats)] + wrapped_model = WrapModel(head, 'forward') + rewrite_inputs = { + 'feats': feats, + } + rewrite_outputs, is_backend_output = get_rewrite_outputs( + wrapped_model=wrapped_model, + model_inputs=rewrite_inputs, + deploy_cfg=deploy_cfg) + model_outputs[0] = [*model_outputs[0][0], *model_outputs[0][1]] + for model_output, rewrite_output in zip(model_outputs[0], + rewrite_outputs[0]): + model_output = model_output.squeeze().cpu().numpy() + rewrite_output = rewrite_output.squeeze() + assert np.allclose( + model_output, rewrite_output, rtol=1e-03, atol=1e-05) + + def _replace_r50_with_r18(model): """Replace ResNet50 with ResNet18 in config.""" model = copy.deepcopy(model) @@ -1119,26 +1228,14 @@ def test_get_bboxes_of_vfnet_head(backend_type: Backend): assert rewrite_outputs is not None -@pytest.mark.parametrize('backend_type', - [Backend.ONNXRUNTIME, Backend.OPENVINO]) -def test_base_dense_head_get_bboxes(backend_type: Backend): - """Test get_bboxes rewrite of base dense head.""" - check_backend(backend_type) - anchor_head = get_anchor_head_model() - anchor_head.cpu().eval() - s = 128 - img_metas = [{ - 'scale_factor': np.ones(4), - 'pad_shape': (s, s, 3), - 'img_shape': (s, s, 3) - }] - - output_names = ['dets', 'labels'] - - deploy_cfg = mmcv.Config( +def get_deploy_cfg(backend_type: Backend, ir_type: str): + return mmcv.Config( dict( backend_config=dict(type=backend_type.value), - onnx_config=dict(output_names=output_names, input_shape=None), + onnx_config=dict( + type=ir_type, + output_names=['dets', 'labels'], + input_shape=None), codebase_config=dict( type='mmdet', task='ObjectDetection', @@ -1151,6 +1248,26 @@ def test_base_dense_head_get_bboxes(backend_type: Backend): background_label_id=-1, )))) + +@pytest.mark.parametrize('backend_type, ir_type', + [(Backend.ONNXRUNTIME, 'onnx'), + (Backend.OPENVINO, 'onnx'), + (Backend.TORCHSCRIPT, 'torchscript')]) +def test_base_dense_head_get_bboxes(backend_type: Backend, ir_type: str): + """Test get_bboxes rewrite of base dense head.""" + check_backend(backend_type) + anchor_head = get_anchor_head_model() + anchor_head.cpu().eval() + s = 128 + img_metas = [{ + 'scale_factor': np.ones(4), + 'pad_shape': (s, s, 3), + 'img_shape': (s, s, 3) + }] + + deploy_cfg = get_deploy_cfg(backend_type, ir_type) + output_names = get_ir_config(deploy_cfg).get('output_names', None) + # the cls_score's size: (1, 36, 32, 32), (1, 36, 16, 16), # (1, 36, 8, 8), (1, 36, 4, 4), (1, 36, 2, 2). # the bboxes's size: (1, 36, 32, 32), (1, 36, 16, 16), diff --git a/tests/test_core/test_function_rewriter.py b/tests/test_core/test_function_rewriter.py index b9b43fb688..97a814e929 100644 --- a/tests/test_core/test_function_rewriter.py +++ b/tests/test_core/test_function_rewriter.py @@ -3,7 +3,8 @@ from mmdeploy.core import FUNCTION_REWRITER, RewriterContext from mmdeploy.core.rewriters.function_rewriter import FunctionRewriter -from mmdeploy.utils.constants import Backend +from mmdeploy.core.rewriters.rewriter_utils import collect_env +from mmdeploy.utils.constants import IR, Backend def test_function_rewriter(): @@ -97,7 +98,6 @@ def test_rewrite_homonymic_functions(self): assert package.module.func() == 1 function_rewriter = FunctionRewriter() - function_rewriter.add_backend(Backend.NCNN.value) @function_rewriter.register_rewriter(func_name=path1) def func_2(ctx): @@ -108,7 +108,7 @@ def func_2(ctx): def func_3(ctx): return 3 - function_rewriter.enter(backend=Backend.NCNN.value) + function_rewriter.enter(env=collect_env(Backend.NCNN, ir=IR.DEFAULT)) # This is a feature assert package.func() == 2 assert package.module.func() == 3 @@ -118,7 +118,6 @@ def func_3(ctx): assert package.module.func() == 1 function_rewriter2 = FunctionRewriter() - function_rewriter2.add_backend(Backend.NCNN.value) @function_rewriter2.register_rewriter( func_name=path1, backend=Backend.NCNN.value) @@ -129,7 +128,7 @@ def func_4(ctx): def func_5(ctx): return 5 - function_rewriter2.enter(backend=Backend.NCNN.value) + function_rewriter2.enter(env=collect_env(Backend.NCNN, ir=IR.DEFAULT)) # This is a feature assert package.func() == 4 assert package.module.func() == 5 @@ -146,7 +145,6 @@ def test_rewrite_homonymic_methods(self): c = package.C() function_rewriter = FunctionRewriter() - function_rewriter.add_backend(Backend.NCNN.value) assert c.method() == 1 @@ -159,14 +157,13 @@ def func_2(ctx, self): def func_3(ctx, self): return 3 - function_rewriter.enter(backend=Backend.NCNN.value) + function_rewriter.enter(env=collect_env(Backend.NCNN, ir=IR.DEFAULT)) assert c.method() == 3 function_rewriter.exit() assert c.method() == 1 function_rewriter2 = FunctionRewriter() - function_rewriter2.add_backend(Backend.NCNN.value) @function_rewriter2.register_rewriter( func_name=path1, backend=Backend.NCNN.value) @@ -177,7 +174,7 @@ def func_4(ctx, self): def func_5(ctx, self): return 5 - function_rewriter2.enter(backend=Backend.NCNN.value) + function_rewriter2.enter(env=collect_env(Backend.NCNN, ir=IR.DEFAULT)) assert c.method() == 4 function_rewriter2.exit() @@ -196,7 +193,6 @@ def test_rewrite_derived_methods(): assert derived_obj.method() == 1 function_rewriter = FunctionRewriter() - function_rewriter.add_backend(Backend.NCNN.value) @function_rewriter.register_rewriter(func_name=path1) def func_2(ctx, self): @@ -207,12 +203,12 @@ def func_2(ctx, self): def func_3(ctx, self): return 3 - function_rewriter.enter() + function_rewriter.enter(env=collect_env(Backend.DEFAULT, ir=IR.DEFAULT)) assert base_obj.method() == 2 assert derived_obj.method() == 2 function_rewriter.exit() - function_rewriter.enter(backend=Backend.NCNN.value) + function_rewriter.enter(env=collect_env(Backend.NCNN, ir=IR.DEFAULT)) assert base_obj.method() == 2 assert derived_obj.method() == 3 function_rewriter.exit() @@ -221,7 +217,7 @@ def func_3(ctx, self): assert derived_obj.method() == 1 # Check if the recovery is correct - function_rewriter.enter() + function_rewriter.enter(env=collect_env(Backend.DEFAULT, ir=IR.DEFAULT)) assert base_obj.method() == 2 assert derived_obj.method() == 2 function_rewriter.exit() diff --git a/tests/test_core/test_mark.py b/tests/test_core/test_mark.py index 5c0990f90f..fb85472ee1 100644 --- a/tests/test_core/test_mark.py +++ b/tests/test_core/test_mark.py @@ -4,8 +4,9 @@ import onnx import torch -from mmdeploy.core import mark +from mmdeploy.core import RewriterContext, mark from mmdeploy.core.optimizers import attribute_to_dict +from mmdeploy.utils.constants import IR, Backend output_file = tempfile.NamedTemporaryFile(suffix='.onnx').name @@ -68,3 +69,9 @@ def forward(self, x, y): type='output', name='c', shape=[2, 3, 4]) + + with RewriterContext( + cfg=None, backend=Backend.TORCHSCRIPT.value, + ir=IR.TORCHSCRIPT), torch.no_grad(), torch.jit.optimized_execution( + True): + torch.jit.trace(model, (x, y)) diff --git a/tests/test_core/test_rewriter_registry.py b/tests/test_core/test_rewriter_registry.py deleted file mode 100644 index b577d02623..0000000000 --- a/tests/test_core/test_rewriter_registry.py +++ /dev/null @@ -1,59 +0,0 @@ -# Copyright (c) OpenMMLab. All rights reserved. -import pytest - -from mmdeploy.core.rewriters.rewriter_utils import RewriterRegistry -from mmdeploy.utils.constants import Backend - - -def test_check_backend(): - with pytest.raises(Exception): - registry = RewriterRegistry() - registry._check_backend(Backend.ONNXRUNTIME.value) - - -def test_add_backend(): - registry = RewriterRegistry() - registry.add_backend(Backend.ONNXRUNTIME.value) - assert Backend.ONNXRUNTIME.value in registry._rewrite_records - assert Backend.DEFAULT.value in registry._rewrite_records - assert Backend.TENSORRT.value not in registry._rewrite_records - - -def test_register_object(): - registry = RewriterRegistry() - - @registry.register_object('add', backend=Backend.DEFAULT.value) - def add(a, b): - return a + b - - records = registry._rewrite_records[Backend.DEFAULT.value] - assert records is not None - assert records['add'] is not None - assert records['add']['_object'] is not None - add_func = records['add']['_object'] - assert add_func(123, 456) == 123 + 456 - - -def test_get_records(): - registry = RewriterRegistry() - registry.add_backend(Backend.TENSORRT.value) - - @registry.register_object('add', backend=Backend.DEFAULT.value) - def add(a, b): - return a + b - - @registry.register_object('minus', backend=Backend.DEFAULT.value) - def minus(a, b): - return a - b - - @registry.register_object('add', backend=Backend.TENSORRT.value) - def fake_add(a, b): - return a * b - - default_records = dict(registry.get_records(Backend.DEFAULT.value)) - assert default_records['add']['_object'](1, 1) == 2 - assert default_records['minus']['_object'](1, 1) == 0 - - tensorrt_records = dict(registry.get_records(Backend.TENSORRT.value)) - assert tensorrt_records['add']['_object'](1, 1) == 1 - assert tensorrt_records['minus']['_object'](1, 1) == 0 diff --git a/tests/test_core/test_rewriter_utils.py b/tests/test_core/test_rewriter_utils.py new file mode 100644 index 0000000000..4954a573d8 --- /dev/null +++ b/tests/test_core/test_rewriter_utils.py @@ -0,0 +1,112 @@ +# Copyright (c) OpenMMLab. All rights reserved. +import mmdeploy +import mmdeploy.core.rewriters.rewriter_utils as rewriter_utils +from mmdeploy.core.rewriters.rewriter_utils import (BackendChecker, + RewriterRegistry, + collect_env) +from mmdeploy.utils.constants import IR, Backend + + +def test_collect_env(): + env_dict = collect_env(Backend.ONNXRUNTIME, IR.ONNX, version='1.0') + assert env_dict['backend'] == Backend.ONNXRUNTIME + assert env_dict['ir'] == IR.ONNX + assert env_dict['version'] == '1.0' + assert env_dict['mmdeploy'] == mmdeploy.__version__ + + +class TestChecker: + env = collect_env(Backend.ONNXRUNTIME, IR.ONNX) + + def test_backend_checker(self): + true_checker = rewriter_utils.BackendChecker(Backend.ONNXRUNTIME) + assert true_checker.check(self.env) is True + + false_checker = rewriter_utils.BackendChecker(Backend.TENSORRT) + assert false_checker.check(self.env) is False + + def test_ir_checker(self): + true_checker = rewriter_utils.IRChecker(IR.ONNX) + assert true_checker.check(self.env) is True + + false_checker = rewriter_utils.IRChecker(IR.TORCHSCRIPT) + assert false_checker.check(self.env) is False + + def test_lib_version_checker(self): + true_checker = rewriter_utils.LibVersionChecker( + 'mmdeploy', mmdeploy.__version__, mmdeploy.__version__) + assert true_checker.check(self.env) is True + + false_checker = rewriter_utils.LibVersionChecker( + 'mmdeploy', max_version='0.0.0') + assert false_checker.check(self.env) is False + + +def test_register_object(): + registry = RewriterRegistry() + checker = rewriter_utils.BackendChecker(Backend.ONNXRUNTIME) + + @registry.register_object( + 'add', + backend=Backend.DEFAULT.value, + ir=IR.DEFAULT, + extra_checkers=checker) + def add(a, b): + return a + b + + records = registry._rewrite_records + assert records is not None + assert records['add'] is not None + assert isinstance(records['add'][0]['_checkers'], list) + assert isinstance(records['add'][0]['_checkers'][0], BackendChecker) + assert records['add'][0]['_object'] is not None + add_func = records['add'][0]['_object'] + assert add_func(123, 456) == 123 + 456 + + +def test_get_records(): + registry = RewriterRegistry() + + @registry.register_object( + 'get_num', backend=Backend.ONNXRUNTIME.value, ir=IR.ONNX) + def get_num_1(): + return 1 + + @registry.register_object( + 'get_num', backend=Backend.ONNXRUNTIME.value, ir=IR.TORCHSCRIPT) + def get_num_2(): + return 2 + + @registry.register_object( + 'get_num', backend=Backend.TENSORRT.value, ir=IR.ONNX) + def get_num_3(): + return 3 + + @registry.register_object( + 'get_num', backend=Backend.TENSORRT.value, ir=IR.TORCHSCRIPT) + def get_num_4(): + return 4 + + @registry.register_object( + 'get_num', backend=Backend.DEFAULT.value, ir=IR.DEFAULT) + def get_num_5(): + return 5 + + records = dict( + registry.get_records(collect_env(Backend.ONNXRUNTIME, IR.ONNX))) + assert records['get_num']['_object']() == 1 + + records = dict( + registry.get_records(collect_env(Backend.ONNXRUNTIME, IR.TORCHSCRIPT))) + assert records['get_num']['_object']() == 2 + + records = dict( + registry.get_records(collect_env(Backend.TENSORRT, IR.ONNX))) + assert records['get_num']['_object']() == 3 + + records = dict( + registry.get_records(collect_env(Backend.TENSORRT, IR.TORCHSCRIPT))) + assert records['get_num']['_object']() == 4 + + records = dict(registry.get_records(collect_env(Backend.NCNN, IR.ONNX))) + assert records['get_num']['_object']() == 5 diff --git a/tests/test_csrc/CMakeLists.txt b/tests/test_csrc/CMakeLists.txt index 3ff7e2d155..34cc0349dd 100644 --- a/tests/test_csrc/CMakeLists.txt +++ b/tests/test_csrc/CMakeLists.txt @@ -14,60 +14,63 @@ aux_source_directory(${CMAKE_CURRENT_SOURCE_DIR}/net NET_TC) aux_source_directory(${CMAKE_CURRENT_SOURCE_DIR}/model MODEL_TC) set(DEVICE_TC) -foreach(DEVICE IN LISTS MMDEPLOY_TARGET_DEVICES) - list(APPEND DEVICE_TC - ${CMAKE_CURRENT_SOURCE_DIR}/device/test_${DEVICE}_device.cpp) -endforeach() +foreach (DEVICE IN LISTS MMDEPLOY_TARGET_DEVICES) + list(APPEND DEVICE_TC + ${CMAKE_CURRENT_SOURCE_DIR}/device/test_${DEVICE}_device.cpp) +endforeach () set(CAPI_TC) -if("all" IN_LIST MMDEPLOY_CODEBASES) - set(TASK_LIST - "classifier;detector;segmentor;text_detector;text_recognizer;restorer;model" - ) - set(CODEBASES "mmcls;mmdet;mmseg;mmedit;mmocr") -else() - set(TASK_LIST "model") - set(CODEBASES "${MMDEPLOY_CODEBASES}") - if("mmcls" IN_LIST MMDEPLOY_CODEBASES) - list(APPEND TASK_LIST "classifier") - endif() - if("mmdet" IN_LIST MMDEPLOY_CODEBASES) - list(APPEND TASK_LIST "detector") - endif() - if("mmseg" IN_LIST MMDEPLOY_CODEBASES) - list(APPEND TASK_LIST "segmentor") - endif() - if("mmedit" IN_LIST MMDEPLOY_CODEBASES) - list(APPEND TASK_LIST "restorer") - endif() - if("mmocr" IN_LIST MMDEPLOY_CODEBASES) - list(APPEND TASK_LIST "text_detector") - list(APPEND TASK_LIST "text_recognizer") - endif() -endif() -foreach(TASK ${TASK_LIST}) - list(APPEND CAPI_TC ${CMAKE_CURRENT_SOURCE_DIR}/capi/test_${TASK}.cpp) -endforeach() +if ("all" IN_LIST MMDEPLOY_CODEBASES) + set(TASK_LIST + "classifier;detector;segmentor;text_detector;text_recognizer;restorer;model" + ) + set(CODEBASES "mmcls;mmdet;mmseg;mmedit;mmocr") +else () + set(TASK_LIST "model") + set(CODEBASES "${MMDEPLOY_CODEBASES}") + if ("mmcls" IN_LIST MMDEPLOY_CODEBASES) + list(APPEND TASK_LIST "classifier") + endif () + if ("mmdet" IN_LIST MMDEPLOY_CODEBASES) + list(APPEND TASK_LIST "detector") + endif () + if ("mmseg" IN_LIST MMDEPLOY_CODEBASES) + list(APPEND TASK_LIST "segmentor") + endif () + if ("mmedit" IN_LIST MMDEPLOY_CODEBASES) + list(APPEND TASK_LIST "restorer") + endif () + if ("mmocr" IN_LIST MMDEPLOY_CODEBASES) + list(APPEND TASK_LIST "text_detector") + list(APPEND TASK_LIST "text_recognizer") + endif () +endif () +foreach (TASK ${TASK_LIST}) + list(APPEND CAPI_TC ${CMAKE_CURRENT_SOURCE_DIR}/capi/test_${TASK}.cpp) +endforeach () # generate the header file configure_file(config/test_define.h.in - ${CMAKE_CURRENT_SOURCE_DIR}/test_define.h) + ${CMAKE_CURRENT_SOURCE_DIR}/test_define.h) set(TC_SRCS - ${TC_SRCS} - ${ARCHIVE_TC} - ${CORE_TC} - ${TRANSFORM_TC} - ${MODEL_TC} - ${NET_TC} - ${DEVICE_TC} - ${CAPI_TC}) + ${TC_SRCS} + ${ARCHIVE_TC} + ${CORE_TC} + ${TRANSFORM_TC} + ${MODEL_TC} + ${NET_TC} + ${DEVICE_TC} + ${CAPI_TC}) add_executable(mmdeploy_tests ${TC_SRCS}) target_include_directories(mmdeploy_tests - PRIVATE ${CMAKE_SOURCE_DIR}/third_party/catch2) + PRIVATE ${CMAKE_SOURCE_DIR}/third_party/catch2) target_include_directories(mmdeploy_tests PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}) -target_link_libraries( - mmdeploy_tests PRIVATE ${MMDEPLOY_LIBS} ${OpenCV_LIBS} - -Wl,--no-as-needed ${MMDEPLOY_DYNAMIC_MODULES} -Wl,--as-need - -Wl,--whole-archive ${MMDEPLOY_STATIC_MODULES} -Wl,--no-whole-archive) + +mmdeploy_load_static(mmdeploy_tests MMDeployStaticModules) +mmdeploy_load_dynamic(mmdeploy_tests MMDeployDynamicModules) +target_link_libraries(mmdeploy_tests PRIVATE + MMDeployLibs + mmdeploy_transform + mmdeploy_opencv_utils) diff --git a/tests/test_csrc/archive/test_value_archive.cpp b/tests/test_csrc/archive/test_value_archive.cpp index 9e53f12269..f46316e355 100644 --- a/tests/test_csrc/archive/test_value_archive.cpp +++ b/tests/test_csrc/archive/test_value_archive.cpp @@ -1,5 +1,12 @@ // Copyright (c) OpenMMLab. All rights reserved. +// clang-format off + +#include "catch.hpp" + +// clang-format on + +#include #include #include #include @@ -8,10 +15,9 @@ #include #include #include -#include "core/utils/formatter.h" #include "archive/value_archive.h" -#include "catch.hpp" +#include "core/utils/formatter.h" // clang-format off @@ -41,8 +47,8 @@ TEMPLATE_LIST_TEST_CASE("test array-like for value", "[value]", ArrayLikeTypes) } TEST_CASE("test native array for value archive", "[value1]") { - const int a[10] = {0,1,2,3,4,5,6,7,8,9}; - int b[10] = {0,0,0,0,0,0,0,0,0,0}; + const int a[10] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9}; + int b[10] = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0}; mmdeploy::Value value; mmdeploy::ValueOutputArchive oa(value); oa(a); diff --git a/tests/test_csrc/capi/test_classifier.cpp b/tests/test_csrc/capi/test_classifier.cpp index 5e0b0f4136..1b4ff1bbf0 100644 --- a/tests/test_csrc/capi/test_classifier.cpp +++ b/tests/test_csrc/capi/test_classifier.cpp @@ -33,11 +33,11 @@ TEST_CASE("test classifier's c api", "[classifier]") { ret = mmdeploy_classifier_apply(handle, mats.data(), (int)mats.size(), &results, &result_count); REQUIRE(ret == MM_SUCCESS); auto result_ptr = results; - INFO("model_path: {}", model_path); + MMDEPLOY_INFO("model_path: {}", model_path); for (auto i = 0; i < (int)mats.size(); ++i) { - INFO("the {}-th classification result: ", i); + MMDEPLOY_INFO("the {}-th classification result: ", i); for (int j = 0; j < *result_count; ++j, ++result_ptr) { - INFO("\t label: {}, score: {}", result_ptr->label_id, result_ptr->score); + MMDEPLOY_INFO("\t label: {}, score: {}", result_ptr->label_id, result_ptr->score); } } @@ -46,12 +46,12 @@ TEST_CASE("test classifier's c api", "[classifier]") { }; auto gResources = MMDeployTestResources::Get(); - auto img_lists = gResources.LocateImageResources("mmcls/images"); + auto img_lists = gResources.LocateImageResources(fs::path{"mmcls"} / "images"); REQUIRE(!img_lists.empty()); for (auto& backend : gResources.backends()) { DYNAMIC_SECTION("loop backend: " << backend) { - auto model_list = gResources.LocateModelResources("mmcls/" + backend); + auto model_list = gResources.LocateModelResources(fs::path{"mmcls/"} / backend); REQUIRE(!model_list.empty()); for (auto& model_path : model_list) { for (auto& device_name : gResources.device_names(backend)) { diff --git a/tests/test_csrc/capi/test_detector.cpp b/tests/test_csrc/capi/test_detector.cpp index f7a72e5410..a801d352a1 100644 --- a/tests/test_csrc/capi/test_detector.cpp +++ b/tests/test_csrc/capi/test_detector.cpp @@ -6,12 +6,13 @@ #include "apis/c/detector.h" #include "core/logger.h" +#include "core/utils/formatter.h" #include "opencv2/opencv.hpp" #include "test_resource.h" - using namespace std; TEST_CASE("test detector's c api", "[detector]") { + MMDEPLOY_INFO("test detector"); auto test = [](const string &device, const string &model_path, const vector &img_list) { mm_handle_t handle{nullptr}; auto ret = mmdeploy_detector_create_by_path(model_path.c_str(), device.c_str(), 0, &handle); @@ -32,26 +33,30 @@ TEST_CASE("test detector's c api", "[detector]") { REQUIRE(ret == MM_SUCCESS); auto result_ptr = results; for (auto i = 0; i < mats.size(); ++i) { - INFO("the '{}-th' image has '{}' objects", i, result_count[i]); + MMDEPLOY_INFO("the '{}-th' image has '{}' objects", i, result_count[i]); for (auto j = 0; j < result_count[i]; ++j, ++result_ptr) { auto &bbox = result_ptr->bbox; - INFO(" >> bbox[{}, {}, {}, {}], label_id {}, score {}", bbox.left, bbox.top, bbox.right, - bbox.bottom, result_ptr->label_id, result_ptr->score); + MMDEPLOY_INFO(" >> bbox[{}, {}, {}, {}], label_id {}, score {}", bbox.left, bbox.top, + bbox.right, bbox.bottom, result_ptr->label_id, result_ptr->score); } } mmdeploy_detector_release_result(results, result_count, (int)mats.size()); mmdeploy_detector_destroy(handle); }; - - auto gResources = MMDeployTestResources::Get(); - auto img_lists = gResources.LocateImageResources("mmdet/images"); + MMDEPLOY_INFO("get test resources"); + auto &gResources = MMDeployTestResources::Get(); + MMDEPLOY_INFO("locate image resources"); + auto img_lists = gResources.LocateImageResources(fs::path{"mmdet"} / "images"); + MMDEPLOY_INFO("{}", img_lists.size()); REQUIRE(!img_lists.empty()); for (auto &backend : gResources.backends()) { + MMDEPLOY_INFO("backend: {}", backend); DYNAMIC_SECTION("loop backend: " << backend) { - auto model_list = gResources.LocateModelResources("mmdet/" + backend); + auto model_list = gResources.LocateModelResources(fs::path{"mmdet"} / backend); REQUIRE(!model_list.empty()); for (auto &model_path : model_list) { + MMDEPLOY_INFO("model: {}", model_path); for (auto &device_name : gResources.device_names(backend)) { test(device_name, model_path, img_lists); } diff --git a/tests/test_csrc/capi/test_model.cpp b/tests/test_csrc/capi/test_model.cpp index af0a983628..d9bab881f3 100644 --- a/tests/test_csrc/capi/test_model.cpp +++ b/tests/test_csrc/capi/test_model.cpp @@ -12,7 +12,7 @@ TEST_CASE("test model c capi", "[model]") { std::string model_path; for (auto const &codebase : gResource.codebases()) { for (auto const &backend : gResource.backends()) { - if (auto _model_list = gResource.LocateModelResources(codebase + "/" + backend); + if (auto _model_list = gResource.LocateModelResources(fs::path{codebase} / backend); !_model_list.empty()) { model_path = _model_list.front(); break; diff --git a/tests/test_csrc/capi/test_restorer.cpp b/tests/test_csrc/capi/test_restorer.cpp index 502d377021..4e56537174 100644 --- a/tests/test_csrc/capi/test_restorer.cpp +++ b/tests/test_csrc/capi/test_restorer.cpp @@ -40,12 +40,12 @@ TEST_CASE("test restorer's c api", "[restorer]") { }; auto gResources = MMDeployTestResources::Get(); - auto img_lists = gResources.LocateImageResources("mmedit/images"); + auto img_lists = gResources.LocateImageResources(fs::path{"mmedit"} / "images"); REQUIRE(!img_lists.empty()); for (auto &backend : gResources.backends()) { DYNAMIC_SECTION("loop backend: " << backend) { - auto model_list = gResources.LocateModelResources("mmedit/" + backend); + auto model_list = gResources.LocateModelResources(fs::path{"mmedit"} / backend); REQUIRE(!model_list.empty()); for (auto &model_path : model_list) { for (auto &device_name : gResources.device_names(backend)) { diff --git a/tests/test_csrc/capi/test_segmentor.cpp b/tests/test_csrc/capi/test_segmentor.cpp index b042d793c5..6de6150bf1 100644 --- a/tests/test_csrc/capi/test_segmentor.cpp +++ b/tests/test_csrc/capi/test_segmentor.cpp @@ -43,12 +43,12 @@ TEST_CASE("test segmentor's c api", "[segmentor]") { }; auto gResources = MMDeployTestResources::Get(); - auto img_lists = gResources.LocateImageResources("mmseg/images"); + auto img_lists = gResources.LocateImageResources(fs::path{"mmseg"} / "images"); REQUIRE(!img_lists.empty()); for (auto &backend : gResources.backends()) { DYNAMIC_SECTION("loop backend: " << backend) { - auto model_list = gResources.LocateModelResources("mmseg/" + backend); + auto model_list = gResources.LocateModelResources(fs::path{"mmseg"} / backend); REQUIRE(!model_list.empty()); for (auto &model_path : model_list) { for (auto &device_name : gResources.device_names(backend)) { diff --git a/tests/test_csrc/capi/test_text_detector.cpp b/tests/test_csrc/capi/test_text_detector.cpp index a2bdd84493..af12d14e09 100644 --- a/tests/test_csrc/capi/test_text_detector.cpp +++ b/tests/test_csrc/capi/test_text_detector.cpp @@ -34,12 +34,12 @@ TEST_CASE("test text detector's c api", "[text-detector]") { auto result_ptr = results; for (auto i = 0; i < mats.size(); ++i) { - INFO("the {}-th image has '{}' objects", i, result_count[i]); + MMDEPLOY_INFO("the {}-th image has '{}' objects", i, result_count[i]); for (auto j = 0; j < result_count[i]; ++j, ++result_ptr) { auto& bbox = result_ptr->bbox; - INFO(">> bbox[{}].score: {}, coordinate: ", i, result_ptr->score); + MMDEPLOY_INFO(">> bbox[{}].score: {}, coordinate: ", i, result_ptr->score); for (auto& _bbox : result_ptr->bbox) { - INFO(">> >> ({}, {})", _bbox.x, _bbox.y); + MMDEPLOY_INFO(">> >> ({}, {})", _bbox.x, _bbox.y); } } } @@ -49,12 +49,12 @@ TEST_CASE("test text detector's c api", "[text-detector]") { }; auto& gResources = MMDeployTestResources::Get(); - auto img_list = gResources.LocateImageResources("mmocr/images"); + auto img_list = gResources.LocateImageResources(fs::path{"mmocr"} / "images"); REQUIRE(!img_list.empty()); for (auto& backend : gResources.backends()) { DYNAMIC_SECTION("loop backend: " << backend) { - auto model_list = gResources.LocateModelResources("mmocr/textdet/" + backend); + auto model_list = gResources.LocateModelResources(fs::path{"mmocr"} / "textdet" / "backend"); REQUIRE(!model_list.empty()); for (auto& model_path : model_list) { for (auto& device_name : gResources.device_names(backend)) { diff --git a/tests/test_csrc/capi/test_text_recognizer.cpp b/tests/test_csrc/capi/test_text_recognizer.cpp index 94f01063dc..3265c4b0ec 100644 --- a/tests/test_csrc/capi/test_text_recognizer.cpp +++ b/tests/test_csrc/capi/test_text_recognizer.cpp @@ -35,7 +35,7 @@ TEST_CASE("test text recognizer's c api", "[text-recognizer]") { for (auto i = 0; i < mats.size(); ++i) { std::vector score(results[i].score, results[i].score + results[i].length); - INFO("image {}, text = {}, score = {}", i, results[i].text, score); + MMDEPLOY_INFO("image {}, text = {}, score = {}", i, results[i].text, score); } mmdeploy_text_recognizer_release_result(results, (int)mats.size()); @@ -43,12 +43,12 @@ TEST_CASE("test text recognizer's c api", "[text-recognizer]") { }; auto& gResources = MMDeployTestResources::Get(); - auto img_list = gResources.LocateImageResources("mmocr/images"); + auto img_list = gResources.LocateImageResources(fs::path{"mmocr"} / "images"); REQUIRE(!img_list.empty()); for (auto& backend : gResources.backends()) { DYNAMIC_SECTION("loop backend: " << backend) { - auto model_list = gResources.LocateModelResources("mmocr/textreg/" + backend); + auto model_list = gResources.LocateModelResources(fs::path{"mmocr"} / "textreg" / "backend"); REQUIRE(!model_list.empty()); for (auto& model_path : model_list) { for (auto& device_name : gResources.device_names(backend)) { @@ -93,7 +93,7 @@ TEST_CASE("test text detector-recognizer combo", "[text-detector-recognizer]") { for (int j = 0; j < bbox_count[i]; ++j) { auto& text = texts[offset + j]; std::vector score(text.score, text.score + text.length); - INFO("image {}, text = {}, score = {}", i, text.text, score); + MMDEPLOY_INFO("image {}, text = {}, score = {}", i, text.text, score); } offset += bbox_count[i]; } @@ -106,13 +106,15 @@ TEST_CASE("test text detector-recognizer combo", "[text-detector-recognizer]") { }; auto& gResources = MMDeployTestResources::Get(); - auto img_list = gResources.LocateImageResources("mmocr/images"); + auto img_list = gResources.LocateImageResources(fs::path{"mmocr"} / "images"); REQUIRE(!img_list.empty()); for (auto& backend : gResources.backends()) { DYNAMIC_SECTION("loop backend: " << backend) { - auto det_model_list = gResources.LocateModelResources("/mmocr/textdet/" + backend); - auto reg_model_list = gResources.LocateModelResources("/mmocr/textreg/" + backend); + auto det_model_list = + gResources.LocateModelResources(fs::path{"mmocr"} / "textdet" / backend); + auto reg_model_list = + gResources.LocateModelResources(fs::path{"mmocr"} / "textreg" / backend); REQUIRE(!det_model_list.empty()); REQUIRE(!reg_model_list.empty()); auto det_model_path = det_model_list.front(); diff --git a/tests/test_csrc/core/test_mat.cpp b/tests/test_csrc/core/test_mat.cpp index b1ae27cb35..bb3e1a8842 100644 --- a/tests/test_csrc/core/test_mat.cpp +++ b/tests/test_csrc/core/test_mat.cpp @@ -1,6 +1,8 @@ // Copyright (c) OpenMMLab. All rights reserved. +#include #include +#include #include "catch.hpp" #include "core/logger.h" diff --git a/tests/test_csrc/core/test_status_code.cpp b/tests/test_csrc/core/test_status_code.cpp index 33059862d4..1316a07952 100644 --- a/tests/test_csrc/core/test_status_code.cpp +++ b/tests/test_csrc/core/test_status_code.cpp @@ -26,13 +26,13 @@ TEST_CASE("test status_code", "[status_code]") { sqrt_of_negative().value(); } catch (const Exception& e) { REQUIRE(e.code() == eInvalidArgument); - INFO("{}", e.what()); + MMDEPLOY_INFO("{}", e.what()); } auto r = sqrt_of_negative(); REQUIRE(!r); REQUIRE(r.error() == eInvalidArgument); - INFO("{}", r.error().message().c_str()); + MMDEPLOY_INFO("{}", r.error().message().c_str()); } } // namespace mmdeploy diff --git a/tests/test_csrc/core/test_token.cpp b/tests/test_csrc/core/test_token.cpp deleted file mode 100644 index de2e0f0cb2..0000000000 --- a/tests/test_csrc/core/test_token.cpp +++ /dev/null @@ -1,36 +0,0 @@ -// Copyright (c) OpenMMLab. All rights reserved. - -#include - -#include "catch.hpp" -#include "experimental/collection.h" - -namespace token { - -using namespace mmdeploy::token; - -using batch_size = mmdeploy::Token; -using type = mmdeploy::Token; -using name = mmdeploy::Token; - -} // namespace token - -TEST_CASE("test token", "[token]") { - using namespace mmdeploy::token; - using mmdeploy::Collection; - - auto produce = [] { - Collection c; - c << token::batch_size{64} << token::type{"Resize"} << token::name("resize1"); - return c; - }; - - auto c = produce(); - - auto consume = [](token::batch_size b, token::type t) { - std::cout << b.key() << ": " << *b << "\n" << t.key() << ": " << *t << "\n"; - return std::string{"success"}; - }; - - (void)Apply(consume, c); -} diff --git a/tests/test_csrc/core/test_value.cpp b/tests/test_csrc/core/test_value.cpp index 07bfe6d7ff..0ecc1c629b 100644 --- a/tests/test_csrc/core/test_value.cpp +++ b/tests/test_csrc/core/test_value.cpp @@ -219,7 +219,7 @@ TEST_CASE("test pointer of Value", "[value]") { REQUIRE(p["object"].is_object()); REQUIRE(p["array"].is_array()); REQUIRE(p["array"].is_array()); - INFO("{}", p); + MMDEPLOY_INFO("{}", p); } TEST_CASE("test null Value", "[value]") { @@ -332,7 +332,7 @@ TEST_CASE("test speed of value", "[value]") { } auto t1 = std::chrono::high_resolution_clock::now(); auto dt = std::chrono::duration(t1 - t0).count(); - INFO("time = {}ms", (float)dt); + MMDEPLOY_INFO("time = {}ms", (float)dt); } TEST_CASE("test ctor of value", "[value]") { diff --git a/tests/test_csrc/device/test_cpu_device.cpp b/tests/test_csrc/device/test_cpu_device.cpp index 88164014cf..3109f6cd4b 100644 --- a/tests/test_csrc/device/test_cpu_device.cpp +++ b/tests/test_csrc/device/test_cpu_device.cpp @@ -10,93 +10,6 @@ using namespace mmdeploy; using namespace std::string_literals; -namespace mmdeploy { -Kernel CreateCpuKernel(std::function task); -} - -TEST_CASE("basic device", "[device]") { - Platform platform("cpu"); - REQUIRE(platform.GetPlatformName() == "cpu"s); - REQUIRE(platform.GetPlatformId() == 0); - - const Device host("cpu"); - Stream stream(host); - // REQUIRE(platform.CreateStream("cpu", &stream) == 0); - REQUIRE(stream); - - SECTION("basic stream") { - bool set_me{}; - auto kernel = CreateCpuKernel([&] { set_me = true; }); - REQUIRE(kernel); - REQUIRE(stream.Submit(kernel)); - REQUIRE(stream.Wait()); - REQUIRE(set_me); - } - - SECTION("recursive task") { - auto outer_loop = CreateCpuKernel([&] { - for (int i = 0; i < 10; ++i) { - auto inner_loop = CreateCpuKernel([&, i] { - for (int j = 0; j < 10; ++j) { - std::cerr << "(" << i << ", " << j << ") "; - } - std::cerr << "\n"; - }); - REQUIRE(stream.Submit(inner_loop)); - } - }); - REQUIRE(stream.Submit(outer_loop)); - REQUIRE(stream.Wait()); - } - - SECTION("basic event") { - Event event(host); - // REQUIRE(platform.CreateEvent("cpu", &event) == 0); - REQUIRE(event); - auto sleeping = CreateCpuKernel([&] { - std::cerr << "start sleeping\n"; - for (int i = 0; i < 5; ++i) { - std::this_thread::sleep_for(std::chrono::milliseconds(100)); - std::cerr << "0.1 second passed.\n"; - } - std::cerr << "time's up, waking up.\n"; - }); - for (int i = 0; i < 2; ++i) { - REQUIRE(stream.Submit(sleeping)); - REQUIRE(event.Record(stream)); - REQUIRE(event.Wait()); - std::cerr << "waked up.\n"; - } - } - - SECTION("event on stream") { - const int N = 10; - std::vector streams; - streams.reserve(N); - for (int i = 0; i < N; ++i) { - streams.emplace_back(host); - } - std::vector events; - events.reserve(N); - for (int i = 0; i < N; ++i) { - events.emplace_back(host); - } - for (int i = 0; i < N; ++i) { - auto kernel = CreateCpuKernel([&, i] { - std::cerr << "greatings from stream " << i << ".\n"; - std::this_thread::sleep_for(std::chrono::milliseconds(100)); - std::cerr << "0.1 second passed, goodbye.\n"; - }); - if (i) { - REQUIRE(streams[i].DependsOn(events[i - 1])); - } - REQUIRE(streams[i].Submit(kernel)); - REQUIRE(events[i].Record(streams[i])); - } - REQUIRE(events.back().Wait()); - } -} - TEST_CASE("test buffer", "[buffer]") { using namespace mmdeploy; Device device{"cpu"}; diff --git a/tests/test_csrc/model/test_directory_model.cpp b/tests/test_csrc/model/test_directory_model.cpp index 6ea1bacc99..50091383ae 100644 --- a/tests/test_csrc/model/test_directory_model.cpp +++ b/tests/test_csrc/model/test_directory_model.cpp @@ -24,14 +24,14 @@ TEST_CASE("test directory model", "[model]") { REQUIRE(!directory_model_list.empty()); auto model_dir = "sdk_models/good_model"; REQUIRE(gResource.IsDir(model_dir)); - auto model_path = gResource.resource_root_path() + "/" + model_dir; - REQUIRE(!model_impl->Init(model_path).has_error()); + auto model_path = gResource.resource_root_path() / model_dir; + REQUIRE(!model_impl->Init(model_path.string()).has_error()); REQUIRE(!model_impl->ReadFile("deploy.json").has_error()); REQUIRE(model_impl->ReadFile("not-existing-file").has_error()); model_dir = "sdk_models/bad_model"; REQUIRE(gResource.IsDir(model_dir)); - model_path = gResource.resource_root_path() + "/" + model_dir; - REQUIRE(!model_impl->Init(model_path).has_error()); + model_path = gResource.resource_root_path() / model_dir; + REQUIRE(!model_impl->Init(model_path.string()).has_error()); REQUIRE(model_impl->ReadMeta().has_error()); } diff --git a/tests/test_csrc/model/test_model.cpp b/tests/test_csrc/model/test_model.cpp index b00f8c2b5f..34bd4c9841 100644 --- a/tests/test_csrc/model/test_model.cpp +++ b/tests/test_csrc/model/test_model.cpp @@ -24,7 +24,8 @@ TEST_CASE("model constructor", "[model]") { TEST_CASE("model init", "[model]") { auto& gResource = MMDeployTestResources::Get(); for (auto& codebase : gResource.codebases()) { - if (auto img_list = gResource.LocateImageResources(codebase + "/images"); !img_list.empty()) { + if (auto img_list = gResource.LocateImageResources(fs::path{codebase} / "images"); + !img_list.empty()) { Model model; REQUIRE(model.Init(img_list.front()).has_error()); break; @@ -32,7 +33,7 @@ TEST_CASE("model init", "[model]") { } for (auto& codebase : gResource.codebases()) { for (auto& backend : gResource.backends()) { - if (auto model_list = gResource.LocateModelResources(codebase + "/" + backend); + if (auto model_list = gResource.LocateModelResources(fs::path{codebase} / backend); !model_list.empty()) { Model model; REQUIRE(!model.Init(model_list.front()).has_error()); diff --git a/tests/test_csrc/model/test_zip_model.cpp b/tests/test_csrc/model/test_zip_model.cpp index 48f787bdea..8d5cb9ca76 100644 --- a/tests/test_csrc/model/test_zip_model.cpp +++ b/tests/test_csrc/model/test_zip_model.cpp @@ -25,10 +25,10 @@ TEST_CASE("test zip model", "[zip_model]") { auto& gResource = MMDeployTestResources::Get(); SECTION("bad sdk model") { - auto zip_model_path = "sdk_models/not_zip_file"; + auto zip_model_path = fs::path{"sdk_models"} / "not_zip_file"; REQUIRE(gResource.IsFile(zip_model_path)); - auto model_path = gResource.resource_root_path() + "/" + zip_model_path; - REQUIRE(model_impl->Init(model_path).has_error()); + auto model_path = gResource.resource_root_path() / zip_model_path; + REQUIRE(model_impl->Init(model_path.string()).has_error()); } SECTION("bad zip buffer") { std::vector buffer(100); @@ -36,10 +36,10 @@ TEST_CASE("test zip model", "[zip_model]") { } SECTION("good sdk model") { - auto zip_model_path = "sdk_models/good_model.zip"; + auto zip_model_path = fs::path{"sdk_models"} / "good_model.zip"; REQUIRE(gResource.IsFile(zip_model_path)); - auto model_path = gResource.resource_root_path() + "/" + zip_model_path; - REQUIRE(!model_impl->Init(model_path).has_error()); + auto model_path = gResource.resource_root_path() / zip_model_path; + REQUIRE(!model_impl->Init(model_path.string()).has_error()); REQUIRE(!model_impl->ReadFile("deploy.json").has_error()); REQUIRE(model_impl->ReadFile("not-exist-file").has_error()); REQUIRE(!model_impl->ReadMeta().has_error()); diff --git a/tests/test_csrc/net/test_ncnn_net.cpp b/tests/test_csrc/net/test_ncnn_net.cpp index 98b348c19b..b55280c041 100644 --- a/tests/test_csrc/net/test_ncnn_net.cpp +++ b/tests/test_csrc/net/test_ncnn_net.cpp @@ -12,7 +12,7 @@ using namespace mmdeploy; TEST_CASE("test ncnn net", "[ncnn_net]") { auto& gResource = MMDeployTestResources::Get(); - auto model_list = gResource.LocateModelResources("mmcls/ncnn"); + auto model_list = gResource.LocateModelResources(fs::path{"mmcls"} / "ncnn"); REQUIRE(!model_list.empty()); Model model(model_list.front()); diff --git a/tests/test_csrc/net/test_openvino_net.cpp b/tests/test_csrc/net/test_openvino_net.cpp index f4a2f683f3..c3f82eb61e 100644 --- a/tests/test_csrc/net/test_openvino_net.cpp +++ b/tests/test_csrc/net/test_openvino_net.cpp @@ -12,7 +12,7 @@ using namespace mmdeploy; TEST_CASE("test openvino net", "[openvino_net]") { auto& gResource = MMDeployTestResources::Get(); - auto model_list = gResource.LocateModelResources("mmcls/openvino"); + auto model_list = gResource.LocateModelResources(fs::path{"mmcls"} / "openvino"); REQUIRE(!model_list.empty()); Model model(model_list.front()); diff --git a/tests/test_csrc/net/test_ort_net.cpp b/tests/test_csrc/net/test_ort_net.cpp index 506fbaf199..1162210009 100644 --- a/tests/test_csrc/net/test_ort_net.cpp +++ b/tests/test_csrc/net/test_ort_net.cpp @@ -12,7 +12,7 @@ using namespace mmdeploy; TEST_CASE("test ort net", "[ort_net]") { auto& gResource = MMDeployTestResources::Get(); - auto model_list = gResource.LocateModelResources("mmcls/ort"); + auto model_list = gResource.LocateModelResources(fs::path{"mmcls"} / "ort"); REQUIRE(!model_list.empty()); Model model(model_list.front()); diff --git a/tests/test_csrc/net/test_ppl_net.cpp b/tests/test_csrc/net/test_ppl_net.cpp index 64a6a478a1..b5d34a8ab5 100644 --- a/tests/test_csrc/net/test_ppl_net.cpp +++ b/tests/test_csrc/net/test_ppl_net.cpp @@ -12,7 +12,7 @@ using namespace mmdeploy; TEST_CASE("test pplnn net", "[ppl_net]") { auto& gResource = MMDeployTestResources::Get(); - auto model_list = gResource.LocateModelResources("mmcls/pplnn"); + auto model_list = gResource.LocateModelResources(fs::path{"mmcls"} / "pplnn"); REQUIRE(!model_list.empty()); Model model(model_list.front()); diff --git a/tests/test_csrc/net/test_trt_net.cpp b/tests/test_csrc/net/test_trt_net.cpp index 2b2841d772..c1c579b2c4 100644 --- a/tests/test_csrc/net/test_trt_net.cpp +++ b/tests/test_csrc/net/test_trt_net.cpp @@ -12,7 +12,7 @@ using namespace mmdeploy; TEST_CASE("test trt net", "[trt_net]") { auto& gResource = MMDeployTestResources::Get(); - auto model_list = gResource.LocateModelResources("mmcls/trt"); + auto model_list = gResource.LocateModelResources(fs::path{"mmcls"} / "trt"); REQUIRE(!model_list.empty()); Model model(model_list.front()); diff --git a/tests/test_csrc/preprocess/test_compose.cpp b/tests/test_csrc/preprocess/test_compose.cpp index 9b7cd4d8d1..97e8ea452d 100644 --- a/tests/test_csrc/preprocess/test_compose.cpp +++ b/tests/test_csrc/preprocess/test_compose.cpp @@ -11,7 +11,7 @@ #include "core/registry.h" #include "core/utils/formatter.h" #include "json.hpp" -#include "preprocess/cpu/opencv_utils.h" +#include "opencv_utils.h" #include "test_resource.h" #include "test_utils.h" diff --git a/tests/test_csrc/preprocess/test_crop.cpp b/tests/test_csrc/preprocess/test_crop.cpp index 836e527260..b5958b4218 100644 --- a/tests/test_csrc/preprocess/test_crop.cpp +++ b/tests/test_csrc/preprocess/test_crop.cpp @@ -4,7 +4,7 @@ #include "catch.hpp" #include "core/mat.h" #include "core/utils/device_utils.h" -#include "preprocess/cpu/opencv_utils.h" +#include "opencv_utils.h" #include "preprocess/transform/transform.h" #include "test_resource.h" #include "test_utils.h" diff --git a/tests/test_csrc/preprocess/test_image2tensor.cpp b/tests/test_csrc/preprocess/test_image2tensor.cpp index 13de3e3414..16939a09b2 100644 --- a/tests/test_csrc/preprocess/test_image2tensor.cpp +++ b/tests/test_csrc/preprocess/test_image2tensor.cpp @@ -2,7 +2,7 @@ #include "catch.hpp" #include "core/tensor.h" #include "core/utils/device_utils.h" -#include "preprocess/cpu/opencv_utils.h" +#include "opencv_utils.h" #include "preprocess/transform/transform.h" #include "test_resource.h" #include "test_utils.h" diff --git a/tests/test_csrc/preprocess/test_load.cpp b/tests/test_csrc/preprocess/test_load.cpp index fa7ef5867b..47abe91b3b 100644 --- a/tests/test_csrc/preprocess/test_load.cpp +++ b/tests/test_csrc/preprocess/test_load.cpp @@ -4,7 +4,7 @@ #include "core/mat.h" #include "core/tensor.h" #include "core/utils/device_utils.h" -#include "preprocess/cpu/opencv_utils.h" +#include "opencv_utils.h" #include "preprocess/transform/transform.h" #include "test_resource.h" #include "test_utils.h" diff --git a/tests/test_csrc/preprocess/test_normalize.cpp b/tests/test_csrc/preprocess/test_normalize.cpp index 647203c02c..bf96f55b23 100644 --- a/tests/test_csrc/preprocess/test_normalize.cpp +++ b/tests/test_csrc/preprocess/test_normalize.cpp @@ -3,7 +3,7 @@ #include "catch.hpp" #include "core/mat.h" #include "core/utils/device_utils.h" -#include "preprocess/cpu/opencv_utils.h" +#include "opencv_utils.h" #include "preprocess/transform/transform.h" #include "test_resource.h" #include "test_utils.h" diff --git a/tests/test_csrc/preprocess/test_pad.cpp b/tests/test_csrc/preprocess/test_pad.cpp index 3f1608b3b2..338be4bbaf 100644 --- a/tests/test_csrc/preprocess/test_pad.cpp +++ b/tests/test_csrc/preprocess/test_pad.cpp @@ -3,7 +3,7 @@ #include "catch.hpp" #include "core/mat.h" #include "core/utils/device_utils.h" -#include "preprocess/cpu/opencv_utils.h" +#include "opencv_utils.h" #include "preprocess/transform/transform.h" #include "test_resource.h" #include "test_utils.h" diff --git a/tests/test_csrc/preprocess/test_resize.cpp b/tests/test_csrc/preprocess/test_resize.cpp index 8c63d5a19f..e5143f3091 100644 --- a/tests/test_csrc/preprocess/test_resize.cpp +++ b/tests/test_csrc/preprocess/test_resize.cpp @@ -3,7 +3,7 @@ #include "catch.hpp" #include "core/mat.h" #include "core/utils/device_utils.h" -#include "preprocess/cpu/opencv_utils.h" +#include "opencv_utils.h" #include "preprocess/transform/transform.h" #include "test_resource.h" #include "test_utils.h" diff --git a/tests/test_csrc/test_resource.h b/tests/test_csrc/test_resource.h index 11fbd034e2..f59c79bf52 100644 --- a/tests/test_csrc/test_resource.h +++ b/tests/test_csrc/test_resource.h @@ -5,21 +5,13 @@ #include #include #include +#include #include #include +#include "core/utils/filesystem.h" #include "test_define.h" -#if __GNUC__ >= 8 -#include -namespace fs = std::filesystem; -#else - -#include - -namespace fs = std::experimental::filesystem; -#endif - using namespace std; class MMDeployTestResources { @@ -35,51 +27,51 @@ class MMDeployTestResources { } const std::vector &backends() const { return backends_; } const std::vector &codebases() const { return codebases_; } - const std::string &resource_root_path() const { return resource_root_path_; } + const fs::path &resource_root_path() const { return resource_root_path_; } bool HasDevice(const std::string &name) const { return std::any_of(devices_.begin(), devices_.end(), [&](const std::string &device_name) { return device_name == name; }); } - bool IsDir(const std::string &dir_name) const { - fs::path path{resource_root_path_ + "/" + dir_name}; + bool IsDir(const fs::path &dir_name) const { + auto path = resource_root_path_ / dir_name; return fs::is_directory(path); } - bool IsFile(const std::string &file_name) const { - fs::path path{resource_root_path_ + "/" + file_name}; + bool IsFile(const fs::path &file_name) const { + auto path = resource_root_path_ / file_name; return fs::is_regular_file(path); } public: - std::vector LocateModelResources(const std::string &sdk_model_zoo_dir) { + std::vector LocateModelResources(const fs::path &sdk_model_zoo_dir) { std::vector sdk_model_list; if (resource_root_path_.empty()) { return sdk_model_list; } - fs::path path{resource_root_path_ + "/" + sdk_model_zoo_dir}; + auto path = resource_root_path_ / sdk_model_zoo_dir; if (!fs::is_directory(path)) { return sdk_model_list; } for (auto const &dir_entry : fs::directory_iterator{path}) { fs::directory_entry entry{dir_entry.path()}; if (auto const &_path = dir_entry.path(); fs::is_directory(_path)) { - sdk_model_list.push_back(dir_entry.path()); + sdk_model_list.push_back(dir_entry.path().string()); } } return sdk_model_list; } - std::vector LocateImageResources(const std::string &img_dir) { + std::vector LocateImageResources(const fs::path &img_dir) { std::vector img_list; if (resource_root_path_.empty()) { return img_list; } - fs::path path{resource_root_path_ + "/" + img_dir}; + auto path = resource_root_path_ / img_dir; if (!fs::is_directory(path)) { return img_list; } @@ -122,15 +114,17 @@ class MMDeployTestResources { return result; } - std::string LocateResourceRootPath(const fs::path &cur_path, int max_depth) { + fs::path LocateResourceRootPath(const fs::path &cur_path, int max_depth) { if (max_depth < 0) { return ""; } for (auto const &dir_entry : fs::directory_iterator{cur_path}) { fs::directory_entry entry{dir_entry.path()}; auto const &_path = dir_entry.path(); - if (fs::is_directory(_path) && _path.filename() == "mmdeploy_test_resources") { - return _path.string(); + // filename must be checked before fs::is_directory, the latter will throw + // when _path points to a system file on Windows + if (_path.filename() == "mmdeploy_test_resources" && fs::is_directory(_path)) { + return _path; } } // Didn't find 'mmdeploy_test_resources' in current directory. @@ -143,7 +137,8 @@ class MMDeployTestResources { std::vector backends_; std::vector codebases_; std::map> backend_devices_; - std::string resource_root_path_; + fs::path resource_root_path_; + // std::string resource_root_path_; }; #endif // MMDEPLOY_TEST_RESOURCE_H diff --git a/tests/test_ops/test_ops.py b/tests/test_ops/test_ops.py index eeb86425fd..60143d878f 100644 --- a/tests/test_ops/test_ops.py +++ b/tests/test_ops/test_ops.py @@ -105,12 +105,14 @@ def wrapped_function(inputs, grid): @pytest.mark.parametrize('dynamic_export', [True, False]) @pytest.mark.parametrize('mode', ['bicubic', 'nearest']) @pytest.mark.parametrize('align_corners', [True, False]) -@pytest.mark.parametrize('scale_factor', [2, 4]) +@pytest.mark.parametrize('output_size', [[10, 20], None]) +@pytest.mark.parametrize('scale_factor', [2]) @pytest.mark.parametrize('n, c, h, w', [(2, 3, 5, 10)]) def test_bicubic_interpolate(backend, dynamic_export, mode, align_corners, + output_size, scale_factor, n, c, @@ -140,8 +142,12 @@ def test_bicubic_interpolate(backend, if mode == 'nearest': align_corners = None - resize = nn.Upsample( - scale_factor=scale_factor, mode=mode, align_corners=align_corners) + if output_size is None: + resize = nn.Upsample( + scale_factor=scale_factor, mode=mode, align_corners=align_corners) + else: + resize = nn.Upsample( + size=output_size, mode=mode, align_corners=align_corners) expected_result = resize(input).cuda() wrapped_model = WrapFunction(resize).eval() diff --git a/tests/test_pytorch/test_pytorch_ops.py b/tests/test_pytorch/test_pytorch_ops.py index 69c9e12ed7..9a03148817 100644 --- a/tests/test_pytorch/test_pytorch_ops.py +++ b/tests/test_pytorch/test_pytorch_ops.py @@ -116,3 +116,10 @@ def test_squeeze(self): nodes = get_model_onnx_nodes(model, x) assert nodes[0].attribute[0].ints == [0] assert nodes[0].op_type == 'Squeeze' + + +def test_hardsigmoid(): + x = torch.rand(1, 2, 3, 4) + model = torch.nn.Hardsigmoid().eval() + nodes = get_model_onnx_nodes(model, x) + assert nodes[0].op_type == 'HardSigmoid' diff --git a/tests/test_utils/test_util.py b/tests/test_utils/test_util.py index e9f5ad33c2..d4e6764eec 100644 --- a/tests/test_utils/test_util.py +++ b/tests/test_utils/test_util.py @@ -1,4 +1,5 @@ # Copyright (c) OpenMMLab. All rights reserved. +import importlib import logging import os import tempfile @@ -144,7 +145,8 @@ def test_get_onnx_config(self): class TestIsDynamic: - config_with_onnx_config = mmcv.Config(dict(onnx_config=dict())) + config_with_onnx_config = mmcv.Config( + dict(onnx_config=dict(), backend_config=dict(type='default'))) config_with_dynamic_axes = mmcv.Config( dict( @@ -154,7 +156,8 @@ class TestIsDynamic: 0: 'batch', 2: 'height', 3: 'width' - }}))) + }}), + backend_config=dict(type='default'))) config_with_dynamic_axes_and_input_names = mmcv.Config( dict( @@ -165,12 +168,14 @@ class TestIsDynamic: 0: 'batch', 2: 'height', 3: 'width' - }}))) + }}), + backend_config=dict(type='default'))) config_with_dynamic_axes_list = mmcv.Config( dict( onnx_config=dict( - type='onnx', input_names=['image'], dynamic_axes=[[0, 2, 3]]))) + type='onnx', input_names=['image'], dynamic_axes=[[0, 2, 3]]), + backend_config=dict(type='default'))) def test_is_dynamic_batch_none(self): assert util.is_dynamic_batch( @@ -440,3 +445,25 @@ def test_get_root_logger(): from mmdeploy.utils import get_root_logger logger = get_root_logger() logger.info('This is a test message') + + +def test_get_library_version(): + assert util.get_library_version('abcdefg') is None + try: + lib = importlib.import_module('setuptools') + except ImportError: + pass + else: + assert util.get_library_version('setuptools') == lib.__version__ + + +def test_get_codebase_version(): + versions = util.get_codebase_version() + for k, v in versions.items(): + assert v == util.get_library_version(k) + + +def test_get_backend_version(): + versions = util.get_backend_version() + for k, v in versions.items(): + assert v == util.get_library_version(k) diff --git a/tools/check_env.py b/tools/check_env.py index 68aa2799e7..3718db1bd5 100644 --- a/tools/check_env.py +++ b/tools/check_env.py @@ -4,49 +4,36 @@ from mmcv.utils import get_git_hash import mmdeploy -from mmdeploy.utils import get_root_logger +from mmdeploy.utils import (get_backend_version, get_codebase_version, + get_root_logger) def collect_env(): """Collect the information of the running environments.""" env_info = collect_base_env() - env_info['MMDeployment'] = f'{mmdeploy.__version__}+{get_git_hash()[:7]}' + env_info['MMDeploy'] = f'{mmdeploy.__version__}+{get_git_hash()[:7]}' return env_info def check_backend(): - try: - import onnxruntime as ort - except ImportError: - ort_version = None - else: - ort_version = ort.__version__ + backend_versions = get_backend_version() + ort_version = backend_versions['onnxruntime'] + trt_version = backend_versions['tensorrt'] + ncnn_version = backend_versions['ncnn'] + import mmdeploy.apis.onnxruntime as ort_apis logger = get_root_logger() - logger.info(f'onnxruntime: {ort_version} ops_is_avaliable : ' + logger.info(f'onnxruntime: {ort_version}\tops_is_avaliable : ' f'{ort_apis.is_available()}') - try: - import tensorrt as trt - except ImportError: - trt_version = None - else: - trt_version = trt.__version__ import mmdeploy.apis.tensorrt as trt_apis - logger.info( - f'tensorrt: {trt_version} ops_is_avaliable : {trt_apis.is_available()}' - ) - - try: - import ncnn - except ImportError: - ncnn_version = None - else: - ncnn_version = ncnn.__version__ + logger.info(f'tensorrt: {trt_version}\tops_is_avaliable : ' + f'{trt_apis.is_available()}') + import mmdeploy.apis.ncnn as ncnn_apis logger.info( - f'ncnn: {ncnn_version} ops_is_avaliable : {ncnn_apis.is_available()}') + f'ncnn: {ncnn_version}\tops_is_avaliable : {ncnn_apis.is_available()}') import mmdeploy.apis.pplnn as pplnn_apis logger.info(f'pplnn_is_avaliable: {pplnn_apis.is_available()}') @@ -56,45 +43,9 @@ def check_backend(): def check_codebase(): - try: - import mmcls - except ImportError: - mmcls_version = None - else: - mmcls_version = mmcls.__version__ - logger.info(f'mmcls: {mmcls_version}') - - try: - import mmdet - except ImportError: - mmdet_version = None - else: - mmdet_version = mmdet.__version__ - logger.info(f'mmdet: {mmdet_version}') - - try: - import mmedit - except ImportError: - mmedit_version = None - else: - mmedit_version = mmedit.__version__ - logger.info(f'mmedit: {mmedit_version}') - - try: - import mmocr - except ImportError: - mmocr_version = None - else: - mmocr_version = mmocr.__version__ - logger.info(f'mmocr: {mmocr_version}') - - try: - import mmseg - except ImportError: - mmseg_version = None - else: - mmseg_version = mmseg.__version__ - logger.info(f'mmseg: {mmseg_version}') + codebase_versions = get_codebase_version() + for k, v in codebase_versions.items(): + logger.info(f'{k}:\t{v}') if __name__ == '__main__': diff --git a/tools/deploy.py b/tools/deploy.py index ef854764df..ee0dd7fb43 100644 --- a/tools/deploy.py +++ b/tools/deploy.py @@ -10,9 +10,9 @@ from mmdeploy.apis import (create_calib_table, extract_model, get_predefined_partition_cfg, torch2onnx, - visualize_model) -from mmdeploy.utils import (Backend, get_backend, get_calib_filename, - get_ir_config, get_model_inputs, get_onnx_config, + torch2torchscript, visualize_model) +from mmdeploy.utils import (IR, Backend, get_backend, get_calib_filename, + get_ir_config, get_model_inputs, get_partition_config, get_root_logger, load_config, target_wrapper) from mmdeploy.utils.export_info import dump_info @@ -66,6 +66,21 @@ def create_process(name, target, args, kwargs, ret_value=None): logger.info(f'{name} success.') +def torch2ir(ir_type: IR): + """Return the conversion function from torch to the intermediate + representation. + + Args: + ir_type (IR): The type of the intermediate representation. + """ + if ir_type == IR.ONNX: + return torch2onnx + elif ir_type == IR.TORCHSCRIPT: + return torch2torchscript + else: + raise KeyError(f'Unexpected IR type {ir_type}') + + def main(): args = parse_args() set_start_method('spawn') @@ -87,18 +102,20 @@ def main(): ret_value = mp.Value('d', 0, lock=False) - # convert onnx - onnx_save_file = get_onnx_config(deploy_cfg)['save_file'] + # convert to IR + ir_config = get_ir_config(deploy_cfg) + ir_save_file = ir_config['save_file'] + ir_type = IR.get(ir_config['type']) create_process( - 'torch2onnx', - target=torch2onnx, - args=(args.img, args.work_dir, onnx_save_file, deploy_cfg_path, + f'torch2{ir_type.value}', + target=torch2ir(ir_type), + args=(args.img, args.work_dir, ir_save_file, deploy_cfg_path, model_cfg_path, checkpoint_path), kwargs=dict(device=args.device), ret_value=ret_value) # convert backend - onnx_files = [osp.join(args.work_dir, onnx_save_file)] + ir_files = [osp.join(args.work_dir, ir_save_file)] # partition model partition_cfgs = get_partition_config(deploy_cfg) @@ -112,8 +129,8 @@ def main(): partition_cfgs = get_predefined_partition_cfg( deploy_cfg, partition_cfgs['type']) - origin_onnx_file = onnx_files[0] - onnx_files = [] + origin_ir_file = ir_files[0] + ir_files = [] for partition_cfg in partition_cfgs: save_file = partition_cfg['save_file'] save_path = osp.join(args.work_dir, save_file) @@ -124,11 +141,11 @@ def main(): create_process( f'partition model {save_file} with start: {start}, end: {end}', extract_model, - args=(origin_onnx_file, start, end), + args=(origin_ir_file, start, end), kwargs=dict(dynamic_axes=dynamic_axes, save_file=save_path), ret_value=ret_value) - onnx_files.append(save_path) + ir_files.append(save_path) # calib data calib_filename = get_calib_filename(deploy_cfg) @@ -146,12 +163,12 @@ def main(): device=args.device), ret_value=ret_value) - backend_files = onnx_files + backend_files = ir_files # convert backend backend = get_backend(deploy_cfg) if backend == Backend.TENSORRT: model_params = get_model_inputs(deploy_cfg) - assert len(model_params) == len(onnx_files) + assert len(model_params) == len(ir_files) from mmdeploy.apis.tensorrt import is_available as trt_is_available from mmdeploy.apis.tensorrt import onnx2tensorrt @@ -160,7 +177,7 @@ def main(): + ' please install TensorRT and build TensorRT custom ops first.' backend_files = [] for model_id, model_param, onnx_path in zip( - range(len(onnx_files)), model_params, onnx_files): + range(len(ir_files)), model_params, ir_files): onnx_name = osp.splitext(osp.split(onnx_path)[1])[0] save_file = model_param.get('save_file', onnx_name + '.engine') @@ -186,7 +203,7 @@ def main(): from mmdeploy.apis.ncnn import get_output_model_file, onnx2ncnn backend_files = [] - for onnx_path in onnx_files: + for onnx_path in ir_files: model_param_path, model_bin_path = get_output_model_file( onnx_path, args.work_dir) create_process( @@ -207,7 +224,7 @@ def main(): get_output_model_file, onnx2openvino) openvino_files = [] - for onnx_path in onnx_files: + for onnx_path in ir_files: model_xml_path = get_output_model_file(onnx_path, args.work_dir) input_info = get_input_info_from_cfg(deploy_cfg) output_names = get_ir_config(deploy_cfg).output_names @@ -227,7 +244,7 @@ def main(): from mmdeploy.apis.pplnn import onnx2pplnn pplnn_files = [] - for onnx_path in onnx_files: + for onnx_path in ir_files: algo_file = onnx_path.replace('.onnx', '.json') model_inputs = get_model_inputs(deploy_cfg) assert 'opt_shape' in model_inputs, 'Expect opt_shape ' \ diff --git a/tools/onnx2ncnn.py b/tools/onnx2ncnn.py index 6f2bb6190b..0bddd6e036 100644 --- a/tools/onnx2ncnn.py +++ b/tools/onnx2ncnn.py @@ -1,3 +1,4 @@ +# Copyright (c) OpenMMLab. All rights reserved. import argparse import logging