From 75d4c0f6dbe2b238953cbf864c256de3c9a55f7a Mon Sep 17 00:00:00 2001 From: Jack Zhou <136876878@qq.com> Date: Thu, 7 Nov 2019 15:30:12 +0800 Subject: [PATCH] Update CMakeLists.txt to adapt to the directory structure of win v1.6 inference lib (#20) 2. change code style 3. add error messages: display detail about input errors and config file errors --- inference/CMakeLists.txt | 29 +- inference/README.md | 13 +- inference/conf/detection_rcnn.yaml | 2 +- inference/detection_demo.cpp | 10 +- inference/docs/configuration.md | 8 +- inference/docs/linux_build.md | 42 +- inference/docs/windows_vs2015_build.md | 58 ++- inference/docs/windows_vs2019_build.md | 37 +- inference/predictor/detection_predictor.cpp | 238 ++++++---- inference/predictor/detection_predictor.h | 51 ++- inference/preprocessor/preprocessor.cpp | 21 +- inference/preprocessor/preprocessor.h | 49 +- .../preprocessor/preprocessor_detection.cpp | 208 +++++---- .../preprocessor/preprocessor_detection.h | 30 +- inference/tools/detection_result_pb2.py | 28 +- inference/tools/vis.py | 8 +- inference/utils/conf_parser.h | 432 ++++++++++-------- inference/utils/utils.h | 176 +++---- 18 files changed, 830 insertions(+), 610 deletions(-) diff --git a/inference/CMakeLists.txt b/inference/CMakeLists.txt index ed610da04..d168639f6 100644 --- a/inference/CMakeLists.txt +++ b/inference/CMakeLists.txt @@ -1,7 +1,6 @@ cmake_minimum_required(VERSION 3.0) project(cpp_inference_demo CXX C) -message("cmake module path: ${CMAKE_MODULE_PATH}") -message("cmake root path: ${CMAKE_ROOT}") + option(WITH_MKL "Compile demo with MKL/OpenBlas support,defaultuseMKL." ON) option(WITH_GPU "Compile demo with GPU/CPU, default use CPU." ON) option(WITH_STATIC_LIB "Compile demo with static/shared library, default use static." ON) @@ -70,6 +69,7 @@ link_directories("${CMAKE_CURRENT_BINARY_DIR}/ext/yaml-cpp/lib") link_directories("${CMAKE_CURRENT_BINARY_DIR}") if (WIN32) include_directories("${PADDLE_DIR}/paddle/fluid/inference") + include_directories("${PADDLE_DIR}/paddle/include") link_directories("${PADDLE_DIR}/paddle/fluid/inference") include_directories("${OPENCV_DIR}/build/include") include_directories("${OPENCV_DIR}/opencv/build/include") @@ -134,6 +134,7 @@ if(WITH_MKL) else () set(MATH_LIB ${PADDLE_DIR}/third_party/install/mklml/lib/libmklml_intel${CMAKE_SHARED_LIBRARY_SUFFIX} ${PADDLE_DIR}/third_party/install/mklml/lib/libiomp5${CMAKE_SHARED_LIBRARY_SUFFIX}) + execute_process(COMMAND cp -r ${PADDLE_DIR}/third_party/install/mklml/lib/libmklml_intel${CMAKE_SHARED_LIBRARY_SUFFIX} /usr/lib) endif () set(MKLDNN_PATH "${PADDLE_DIR}/third_party/install/mkldnn") if(EXISTS ${MKLDNN_PATH}) @@ -148,22 +149,22 @@ else() set(MATH_LIB ${PADDLE_DIR}/third_party/install/openblas/lib/libopenblas${CMAKE_STATIC_LIBRARY_SUFFIX}) endif() +if(WIN32) + if(EXISTS "${PADDLE_DIR}/paddle/fluid/inference/libpaddle_fluid${CMAKE_STATIC_LIBRARY_SUFFIX}") + set(DEPS + ${PADDLE_DIR}/paddle/fluid/inference/libpaddle_fluid${CMAKE_STATIC_LIBRARY_SUFFIX}) + else() + set(DEPS + ${PADDLE_DIR}/paddle/lib/libpaddle_fluid${CMAKE_STATIC_LIBRARY_SUFFIX}) + endif() +endif() + if(WITH_STATIC_LIB) - if (WIN32) - set(DEPS - ${PADDLE_DIR}/paddle/fluid/inference/libpaddle_fluid${CMAKE_STATIC_LIBRARY_SUFFIX}) - else () set(DEPS ${PADDLE_DIR}/paddle/lib/libpaddle_fluid${CMAKE_STATIC_LIBRARY_SUFFIX}) - endif() else() - if (WIN32) - set(DEPS - ${PADDLE_DIR}/paddle/fluid/inference/libpaddle_fluid${CMAKE_STATIC_LIBRARY_SUFFIX}) - else () set(DEPS ${PADDLE_DIR}/paddle/lib/libpaddle_fluid${CMAKE_SHARED_LIBRARY_SUFFIX}) - endif() endif() if (NOT WIN32) @@ -242,8 +243,6 @@ if (NOT WIN32) set(DEPS ${DEPS} ${OPENCV_3RD_LIB_DIR}/libippicv${CMAKE_STATIC_LIBRARY_SUFFIX}) endif() endif() -# message(${CMAKE_CXX_FLAGS}) -# set(CMAKE_CXX_FLAGS "-g ${CMAKE_CXX_FLAGS}") SET(PADDLESEG_INFERENCE_SRCS preprocessor/preprocessor.cpp preprocessor/preprocessor_detection.cpp predictor/detection_predictor.cpp @@ -265,7 +264,7 @@ if (WIN32) COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mkldnn/lib/mkldnn.dll ./mkldnn.dll COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/mklml.dll ./release/mklml.dll COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/libiomp5md.dll ./release/libiomp5md.dll - COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mkldnn/lib/mkldnn.dll ./mkldnn.dll + COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mkldnn/lib/mkldnn.dll ./release/mkldnn.dll ) endif() diff --git a/inference/README.md b/inference/README.md index 54faa0f0d..71c6c5d30 100644 --- a/inference/README.md +++ b/inference/README.md @@ -65,7 +65,7 @@ deploy 完成编译后，便生成了需要的可执行文件和链接库。这里以我们基于`faster rcnn`检测模型为例，介绍部署图像检测模型的通用流程。 -### 1. 下载模型文件 +### 4.1. 下载模型文件 我们提供faster rcnn，faster rcnn+fpn模型用于预测coco17数据集，可在以下链接下载：[faster rcnn示例模型下载地址](https://paddleseg.bj.bcebos.com/inference/faster_rcnn_pp50.zip)， [faster rcnn + fpn示例模型下载地址](https://paddleseg.bj.bcebos.com/inference/faster_rcnn_pp50_fpn.zip)。 @@ -83,7 +83,7 @@ faster_rcnn_pp50/ **假设**`Linux`上对应的路径则为`/root/projects/models/faster_rcnn_pp50/`。 -### 2. 修改配置 +### 4.2. 修改配置 `inference`源代码(即本目录)的`conf`目录下提供了示例基于faster rcnn的配置文件`detection_rcnn.yaml`, 相关的字段含义和说明如下： @@ -118,7 +118,7 @@ DEPLOY: # 预测模式，支持 NATIVE 和 ANALYSIS PREDICTOR_MODE: "ANALYSIS" # 每次预测的 batch_size - BATCH_SIZE : 3 + BATCH_SIZE : 3 # 长边伸缩的最大长度，-1代表无限制。 RESIZE_MAX_SIZE: 1333 # 输入的tensor数量。 @@ -127,7 +127,9 @@ DEPLOY: ``` 修改字段`MODEL_PATH`的值为你在**上一步**下载并解压的模型文件所放置的目录即可。更多配置文件字段介绍，请参考文档[预测部署方案配置文件说明](./docs/configuration.md)。 -### 3. 执行预测 +**注意**在使用CPU版本预测库时，`USE_GPU`的值必须设为0，否则无法正常预测。 + +### 4.3. 执行预测 在终端中切换到生成的可执行文件所在目录为当前目录(Windows系统为`cmd`)。 @@ -155,7 +157,7 @@ DEPLOY: 运行可视化脚本时，只需输入命令行参数图片路径、检测结果pb文件路径、目标框阈值以及类别-标签映射文件路径即可得到可视化的图片`X.png` (tools目录下提供coco17的类别标签映射文件coco17.json)。 -```bash +```bash python vis.py --img_path=../build/images/detection_rcnn/000000087038.jpg --img_result_path=../build/images/detection_rcnn/000000087038.jpg.pb --threshold=0.1 --c2l_path=coco17.json ``` @@ -168,3 +170,4 @@ python vis.py --img_path=../build/images/detection_rcnn/000000087038.jpg --img_r ```检测结果图：``` ![检测结果](./demo_images/000000087038.jpg.png) + diff --git a/inference/conf/detection_rcnn.yaml b/inference/conf/detection_rcnn.yaml index 50c23fbb3..a53698d7f 100644 --- a/inference/conf/detection_rcnn.yaml +++ b/inference/conf/detection_rcnn.yaml @@ -13,6 +13,6 @@ DEPLOY: CHANNELS : 3 PRE_PROCESSOR: "DetectionPreProcessor" PREDICTOR_MODE: "ANALYSIS" - BATCH_SIZE : 3 + BATCH_SIZE : 1 RESIZE_MAX_SIZE: 1333 FEEDS_SIZE: 3 diff --git a/inference/detection_demo.cpp b/inference/detection_demo.cpp index 7e711ed69..32921fbd1 100644 --- a/inference/detection_demo.cpp +++ b/inference/detection_demo.cpp @@ -23,18 +23,24 @@ int main(int argc, char** argv) { // 0. parse args google::ParseCommandLineFlags(&argc, &argv, true); if (FLAGS_conf.empty() || FLAGS_input_dir.empty()) { - std::cout << "Usage: ./predictor --conf=/config/path/to/your/model --input_dir=/directory/of/your/input/images"; + std::cout << "Usage: ./predictor --conf=/config/path/to/your/model " + << "--input_dir=/directory/of/your/input/images" << std::endl; return -1; } // 1. create a predictor and init it with conf PaddleSolution::DetectionPredictor predictor; if (predictor.init(FLAGS_conf) != 0) { + #ifdef _WIN32 + std::cerr << "Fail to init predictor" << std::endl; + #else LOG(FATAL) << "Fail to init predictor"; + #endif return -1; } // 2. get all the images with extension '.jpeg' at input_dir - auto imgs = PaddleSolution::utils::get_directory_images(FLAGS_input_dir, ".jpeg|.jpg|.JPEG|.JPG|.bmp|.BMP|.png|.PNG"); + auto imgs = PaddleSolution::utils::get_directory_images(FLAGS_input_dir, + ".jpeg|.jpg|.JPEG|.JPG|.bmp|.BMP|.png|.PNG"); // 3. predict predictor.predict(imgs); diff --git a/inference/docs/configuration.md b/inference/docs/configuration.md index 45c8605c9..1d5360a61 100644 --- a/inference/docs/configuration.md +++ b/inference/docs/configuration.md @@ -16,7 +16,7 @@ ```yaml # 预测部署时所有配置字段需在DEPLOY字段下 -DEPLOY: +DEPLOY: # 类型：required int # 含义：是否使用GPU预测。 0:不使用 1:使用 USE_GPU: 1 @@ -70,6 +70,6 @@ DEPLOY: # 含义: 输入张量的个数。大部分模型不需要设置。 默认值为1. FEEDS_SIZE: 2 # 类型: optional int - # 含义: 将图像的边变为该字段的值的整数倍。默认值为1。 - COARSEST_STRIDE: 32 -``` + # 含义: 将图像的边变为该字段的值的整数倍。在使用fpn模型时需要设为32。默认值为1。 + COARSEST_STRIDE: 32 +``` \ No newline at end of file diff --git a/inference/docs/linux_build.md b/inference/docs/linux_build.md index 2ad9e4638..012f3dee7 100644 --- a/inference/docs/linux_build.md +++ b/inference/docs/linux_build.md @@ -5,22 +5,30 @@ ## 前置条件 * G++ 4.8.2 ~ 4.9.4 -* CUDA 8.0/ CUDA 9.0 +* CUDA 9.0 / CUDA 10.0, cudnn 7+ （仅在使用GPU版本的预测库时需要） * CMake 3.0+ 请确保系统已经安装好上述基本软件，**下面所有示例以工作目录为 `/root/projects/`演示**。 ### Step1: 下载代码 -1. `mkdir -p /root/projects/paddle_models && cd /root/projects/paddle_models` -2. `git clone https://github.com/PaddlePaddle/models.git` +1. `git clone https://github.com/PaddlePaddle/PaddleDetection.git` -`C++`预测代码在`/root/projects/paddle_models/models/PaddleCV/PaddleDetection/inference` 目录，该目录不依赖任何`PaddleDetection`下其他目录。 +`C++`预测代码在`/root/projects/PaddleDetection/inference` 目录，该目录不依赖任何`PaddleDetection`下其他目录。 ### Step2: 下载PaddlePaddle C++ 预测库 fluid_inference -目前仅支持`CUDA 8` 和 `CUDA 9`，请点击 [PaddlePaddle预测库下载地址](https://www.paddlepaddle.org.cn/documentation/docs/zh/develop/advanced_usage/deploy/inference/build_and_install_lib_cn.html)下载对应的版本（develop版本）。 +PaddlePaddle C++ 预测库主要分为CPU版本和GPU版本。其中，针对不同的CUDA版本，GPU版本预测库又分为两个版本预测库：CUDA 9.0和CUDA 10.0版本预测库。以下为各版本C++预测库的下载链接： + +| 版本 | 链接 | +| ---- | ---- | +| CPU版本 | [fluid_inference.tgz](https://bj.bcebos.com/paddlehub/paddle_inference_lib/fluid_inference_linux_cpu_1.6.1.tgz) | +| CUDA 9.0版本 | [fluid_inference.tgz](https://bj.bcebos.com/paddlehub/paddle_inference_lib/fluid_inference_linux_cuda97_1.6.1.tgz) | +| CUDA 10.0版本 | [fluid_inference.tgz](https://bj.bcebos.com/paddlehub/paddle_inference_lib/fluid_inference_linux_cuda10_1.6.1.tgz) | + + +针对不同的CPU类型、不同的指令集，官方提供更多可用的预测库版本，目前已经推出1.6版本的预测库。其余版本具体请参考以下链接:[C++预测库下载列表](https://www.paddlepaddle.org.cn/documentation/docs/zh/develop/advanced_usage/deploy/inference/build_and_install_lib_cn.html) 下载并解压后`/root/projects/fluid_inference`目录包含内容为： @@ -53,25 +61,33 @@ make install ### Step4: 编译 -`CMake`编译时，涉及到四个编译参数用于指定核心依赖库的路径, 他们的定义如下: +`CMake`编译时，涉及到四个编译参数用于指定核心依赖库的路径, 他们的定义如下:（带*表示仅在使用**GPU版本**预测库时指定，其中CUDA库版本尽量对齐，**使用9.0、10.0版本，不使用9.2、10.1版本CUDA库**） | 参数名 | 含义 | | ---- | ---- | -| CUDA_LIB | cuda的库路径 | -| CUDNN_LIB | cuDnn的库路径| -| OPENCV_DIR | OpenCV的安装路径， | +| * CUDA_LIB | CUDA的库路径 | +| * CUDNN_LIB | cudnn的库路径| +| OPENCV_DIR | OpenCV的安装路径 | | PADDLE_DIR | Paddle预测库的路径 | -执行下列操作时，**注意**把对应的参数改为你的上述依赖库实际路径： +在使用**GPU版本**预测库进行编译时，可执行下列操作。**注意**把对应的参数改为你的上述依赖库实际路径： ```shell -cd /root/projects/paddle_models/models/PaddleCV/PaddleDetection/inference - +cd /root/projects/PaddleDetection/inference mkdir build && cd build -cmake .. -DWITH_GPU=ON -DPADDLE_DIR=/root/projects/fluid_inference -DCUDA_LIB=/usr/local/cuda/lib64/ -DOPENCV_DIR=/root/projects/opencv3/ -DCUDNN_LIB=/usr/local/cuda/lib64/ +cmake .. -DWITH_GPU=ON -DPADDLE_DIR=/root/projects/fluid_inference -DCUDA_LIB=/usr/local/cuda/lib64/ -DOPENCV_DIR=/root/projects/opencv3/ -DCUDNN_LIB=/usr/local/cuda/lib64/ -DWITH_STATIC_LIB=OFF make ``` +在使用**CPU版本**预测库进行编译时，可执行下列操作： + +```shell +cd /root/projects/PaddleDetection/inference + +mkdir build && cd build +cmake .. -DWITH_GPU=OFF -DPADDLE_DIR=/root/projects/fluid_inference -DOPENCV_DIR=/root/projects/opencv3/ -DWITH_STATIC_LIB=OFF +make +``` ### Step5: 预测及可视化 diff --git a/inference/docs/windows_vs2015_build.md b/inference/docs/windows_vs2015_build.md index b20b21995..b7b8b04e4 100644 --- a/inference/docs/windows_vs2015_build.md +++ b/inference/docs/windows_vs2015_build.md @@ -5,27 +5,28 @@ ## 前置条件 * Visual Studio 2015 -* CUDA 8.0/ CUDA 9.0 +* CUDA 9.0 / CUDA 10.0，cudnn 7+ （仅在使用GPU版本的预测库时需要） * CMake 3.0+ 请确保系统已经安装好上述基本软件，**下面所有示例以工作目录为 `D:\projects`演示**。 ### Step1: 下载代码 -1. 打开`cmd`, 执行 `cd D:\projects\paddle_models` -2. `git clone https://github.com/PaddlePaddle/models.git` +1. 打开`cmd`, 执行 `cd D:\projects` +2. `git clone https://github.com/PaddlePaddle/PaddleDetection.git` -`C++`预测库代码在`D:\projects\paddle_models\models\PaddleCV\PaddleDetection\inference` 目录，该目录不依赖任何`PaddleDetection`下其他目录。 +`C++`预测库代码在`D:\projects\PaddleDetection\inference` 目录，该目录不依赖任何`PaddleDetection`下其他目录。 ### Step2: 下载PaddlePaddle C++ 预测库 fluid_inference -根据Windows环境，下载相应版本的PaddlePaddle预测库，并解压到`D:\projects\`目录 +PaddlePaddle C++ 预测库主要分为两大版本：CPU版本和GPU版本。其中，针对不同的CUDA版本，GPU版本预测库又分为两个版本预测库：CUDA 9.0和CUDA 10.0版本预测库。根据Windows环境，下载相应版本的PaddlePaddle预测库，并解压到`D:\projects\`目录。以下为各版本C++预测库的下载链接： -| CUDA | GPU | 下载地址 | -|------|------|--------| -| 8.0 | Yes | [fluid_inference.zip](https://bj.bcebos.com/v1/paddleseg/fluid_inference_win.zip) | -| 9.0 | Yes | [fluid_inference_cuda90.zip](https://paddleseg.bj.bcebos.com/fluid_inference_cuda9_cudnn7.zip) | +| 版本 | 链接 | +| ---- | ---- | +| CPU版本 | [fluid_inference_install_dir.zip](https://bj.bcebos.com/paddlehub/paddle_inference_lib/fluid_install_dir_win_cpu_1.6.zip) | +| CUDA 9.0版本 | [fluid_inference_install_dir.zip](https://bj.bcebos.com/paddlehub/paddle_inference_lib/fluid_inference_install_dir_win_cuda9_1.6.1.zip) | +| CUDA 10.0版本 | [fluid_inference_install_dir.zip](https://bj.bcebos.com/paddlehub/paddle_inference_lib/fluid_inference_install_dir_win_cuda10_1.6.1.zip) | 解压后`D:\projects\fluid_inference`目录包含内容为： ``` @@ -42,9 +43,9 @@ fluid_inference 1. 在OpenCV官网下载适用于Windows平台的3.4.6版本， [下载地址](https://sourceforge.net/projects/opencvlibrary/files/3.4.6/opencv-3.4.6-vc14_vc15.exe/download) 2. 运行下载的可执行文件，将OpenCV解压至指定目录，如`D:\projects\opencv` 3. 配置环境变量，如下流程所示 - - 我的电脑->属性->高级系统设置->环境变量 + - 我的电脑->属性->高级系统设置->环境变量 - 在系统变量中找到Path（如没有，自行创建），并双击编辑 - - 新建，将opencv路径填入并保存，如`D:\projects\opencv\build\x64\vc14\bin` + - 新建，将opencv路径填入并保存，如`D:\projects\opencv\build\x64\vc14\bin` ### Step4: 以VS2015为例编译代码 @@ -56,35 +57,52 @@ fluid_inference ``` call "C:\Program Files (x86)\Microsoft Visual Studio 14.0\VC\vcvarsall.bat" amd64 ``` + +三个编译参数的含义说明如下（带*表示仅在使用**GPU版本**预测库时指定, 其中CUDA库版本尽量对齐，**使用9.0、10.0版本，不使用9.2、10.1等版本CUDA库**）： -* CMAKE编译工程 - * PADDLE_DIR: fluid_inference预测库路径 - * CUDA_LIB: CUDA动态库目录, 请根据实际安装情况调整 - * OPENCV_DIR: OpenCV解压目录 +| 参数名 | 含义 | +| ---- | ---- | +| *CUDA_LIB | CUDA的库路径 | +| OPENCV_DIR | OpenCV的安装路径， | +| PADDLE_DIR | Paddle预测库的路径 | +在使用**GPU版本**预测库进行编译时，可执行下列操作。**注意**把对应的参数改为你的上述依赖库实际路径： + +```bash +# 切换到预测库所在目录 +cd /d D:\projects\PaddleDetection\inference +# 创建构建目录, 重新构建只需要删除该目录即可 +mkdir build +cd build +# cmake构建VS项目 +D:\projects\PaddleDetection\inference\build> cmake .. -G "Visual Studio 14 2015 Win64" -DWITH_GPU=ON -DPADDLE_DIR=D:\projects\fluid_inference -DCUDA_LIB=D:\projects\cudalib\v9.0\lib\x64 -DOPENCV_DIR=D:\projects\opencv -T host=x64 ``` + +在使用**CPU版本**预测库进行编译时，可执行下列操作： +```bash # 切换到预测库所在目录 -cd /d D:\projects\paddle_models\models\PaddleCV\PaddleDetection\inference +cd /d D:\projects\PaddleDetection\inference # 创建构建目录, 重新构建只需要删除该目录即可 mkdir build cd build # cmake构建VS项目 -D:\projects\paddle_models\models\PaddleCV\PaddleDetection\inference\build> cmake .. -G "Visual Studio 14 2015 Win64" -DWITH_GPU=ON -DPADDLE_DIR=D:\projects\fluid_inference -DCUDA_LIB=D:\projects\cudalib\v9.0\lib\x64 -DOPENCV_DIR=D:\projects\opencv -T host=x64 +D:\projects\PaddleDetection\inference\build> cmake .. -G "Visual Studio 14 2015 Win64" -DWITH_GPU=OFF -DPADDLE_DIR=D:\projects\fluid_inference -DOPENCV_DIR=D:\projects\opencv -T host=x64 ``` + 这里的`cmake`参数`-G`, 表示生成对应的VS版本的工程，可以根据自己的`VS`版本调整，具体请参考[cmake文档](https://cmake.org/cmake/help/v3.15/manual/cmake-generators.7.html) * 生成可执行文件 ``` -D:\projects\paddle_models\models\PaddleCV\PaddleDetection\inference\build> msbuild /m /p:Configuration=Release cpp_inference_demo.sln +D:\projects\PaddleDetection\inference\build> msbuild /m /p:Configuration=Release cpp_inference_demo.sln ``` ### Step5: 预测及可视化 上述`Visual Studio 2015`编译产出的可执行文件在`build\release`目录下，切换到该目录： ``` -cd /d D:\projects\paddle_models\models\PaddleCV\PaddleDetection\inference\build\release +cd /d D:\projects\PaddleDetection\inference\build\release ``` 之后执行命令： @@ -93,4 +111,4 @@ cd /d D:\projects\paddle_models\models\PaddleCV\PaddleDetection\inference\build\ detection_demo.exe --conf=/path/to/your/conf --input_dir=/path/to/your/input/data/directory ``` -更详细说明请参考ReadMe文档： [预测和可视化部分](../README.md) +更详细说明请参考ReadMe文档： [预测和可视化部分](../README.md) \ No newline at end of file diff --git a/inference/docs/windows_vs2019_build.md b/inference/docs/windows_vs2019_build.md index 6a467af45..475a65b16 100644 --- a/inference/docs/windows_vs2019_build.md +++ b/inference/docs/windows_vs2019_build.md @@ -6,7 +6,7 @@ Windows 平台下，我们使用`Visual Studio 2015` 和 `Visual Studio 2019 Com ## 前置条件 * Visual Studio 2019 -* CUDA 8.0/ CUDA 9.0 +* CUDA 9.0 / CUDA 10.0，cudnn 7+ （仅在使用GPU版本的预测库时需要） * CMake 3.0+ 请确保系统已经安装好上述基本软件，我们使用的是`VS2019`的社区版。 @@ -15,20 +15,21 @@ Windows 平台下，我们使用`Visual Studio 2015` 和 `Visual Studio 2019 Com ### Step1: 下载代码 -1. 点击下载源代码：[下载地址](https://github.com/PaddlePaddle/models/archive/develop.zip) -2. 解压，解压后目录重命名为`paddle_models` +1. 点击下载源代码：[下载地址](https://github.com/PaddlePaddle/PaddleDetection/archive/master.zip) +2. 解压，解压后目录重命名为`PaddleDetection` -以下代码目录路径为`D:\projects\paddle_models` 为例。 +以下代码目录路径为`D:\projects\PaddleDetection` 为例。 ### Step2: 下载PaddlePaddle C++ 预测库 fluid_inference -根据Windows环境，下载相应版本的PaddlePaddle预测库，并解压到`D:\projects\`目录 +PaddlePaddle C++ 预测库主要分为两大版本：CPU版本和GPU版本。其中，针对不同的CUDA版本，GPU版本预测库又分为三个版本预测库：CUDA 9.0和CUDA 10.0版本预测库。根据Windows环境，下载相应版本的PaddlePaddle预测库，并解压到`D:\projects\`目录。以下为各版本C++预测库的下载链接： -| CUDA | GPU | 下载地址 | -|------|------|--------| -| 8.0 | Yes | [fluid_inference.zip](https://bj.bcebos.com/v1/paddleseg/fluid_inference_win.zip) | -| 9.0 | Yes | [fluid_inference_cuda90.zip](https://paddleseg.bj.bcebos.com/fluid_inference_cuda9_cudnn7.zip) | +| 版本 | 链接 | +| ---- | ---- | +| CPU版本 | [fluid_inference_install_dir.zip](https://bj.bcebos.com/paddlehub/paddle_inference_lib/fluid_install_dir_win_cpu_1.6.zip) | +| CUDA 9.0版本 | [fluid_inference_install_dir.zip](https://bj.bcebos.com/paddlehub/paddle_inference_lib/fluid_inference_install_dir_win_cuda9_1.6.1.zip) | +| CUDA 10.0版本 | [fluid_inference_install_dir.zip](https://bj.bcebos.com/paddlehub/paddle_inference_lib/fluid_inference_install_dir_win_cuda10_1.6.1.zip) | 解压后`D:\projects\fluid_inference`目录包含内容为： ``` @@ -39,16 +40,15 @@ fluid_inference | └── version.txt # 版本和编译信息 ``` -**注意：** `CUDA90`版本解压后目录名称为`fluid_inference_cuda90`。 ### Step3: 安装配置OpenCV 1. 在OpenCV官网下载适用于Windows平台的3.4.6版本， [下载地址](https://sourceforge.net/projects/opencvlibrary/files/3.4.6/opencv-3.4.6-vc14_vc15.exe/download) 2. 运行下载的可执行文件，将OpenCV解压至指定目录，如`D:\projects\opencv` 3. 配置环境变量，如下流程所示 - - 我的电脑->属性->高级系统设置->环境变量 + - 我的电脑->属性->高级系统设置->环境变量 - 在系统变量中找到Path（如没有，自行创建），并双击编辑 - - 新建，将opencv路径填入并保存，如`D:\projects\opencv\build\x64\vc14\bin` + - 新建，将opencv路径填入并保存，如`D:\projects\opencv\build\x64\vc14\bin` ### Step4: 使用Visual Studio 2019直接编译CMake @@ -67,16 +67,17 @@ fluid_inference 4. 点击`浏览`，分别设置编译选项指定`CUDA`、`OpenCV`、`Paddle预测库`的路径 -![step4](https://paddleseg.bj.bcebos.com/inference/vs2019_step5.png) - -三个编译参数的含义说明如下： +三个编译参数的含义说明如下（带*表示仅在使用**GPU版本**预测库时指定, 其中CUDA库版本尽量对齐，**使用9.0、10.0版本，不使用9.2、10.1等版本CUDA库**）： | 参数名 | 含义 | | ---- | ---- | -| CUDA_LIB | cuda的库路径 | +| *CUDA_LIB | CUDA的库路径 | | OPENCV_DIR | OpenCV的安装路径， | | PADDLE_DIR | Paddle预测库的路径 | +**注意**在使用CPU版本预测库时，需要把CUDA_LIB的勾去掉。 +![step4](https://paddleseg.bj.bcebos.com/inference/vs2019_step5.png) + **设置完成后**, 点击上图中`保存并生成CMake缓存以加载变量`。 5. 点击`生成`->`全部生成` @@ -89,7 +90,7 @@ fluid_inference 上述`Visual Studio 2019`编译产出的可执行文件在`out\build\x64-Release`目录下，打开`cmd`，并切换到该目录： ``` -cd D:\projects\paddle_models\models\PaddleCV\PaddleDetection\inference\build\x64-Release +cd D:\projects\PaddleDetection\inference\out\build\x64-Release ``` 之后执行命令： @@ -98,4 +99,4 @@ cd D:\projects\paddle_models\models\PaddleCV\PaddleDetection\inference\build\x64 detection_demo.exe --conf=/path/to/your/conf --input_dir=/path/to/your/input/data/directory ``` -更详细说明请参考ReadMe文档： [预测和可视化部分](../README.md) +更详细说明请参考ReadMe文档： [预测和可视化部分](../README.md) \ No newline at end of file diff --git a/inference/predictor/detection_predictor.cpp b/inference/predictor/detection_predictor.cpp index ba07e3b6c..e47125c5e 100644 --- a/inference/predictor/detection_predictor.cpp +++ b/inference/predictor/detection_predictor.cpp @@ -17,31 +17,40 @@ #include #include #include "utils/detection_result.pb.h" +#undef min namespace PaddleSolution { /* lod_buffer: every item in lod_buffer is an image matrix after preprocessing * input_buffer: same data with lod_buffer after flattening to 1-D vector and padding, needed to be empty before using this function */ - void padding_minibatch(const std::vector> &lod_buffer, std::vector &input_buffer, - std::vector &resize_heights, std::vector &resize_widths, int channels, int coarsest_stride = 1) { + void padding_minibatch(const std::vector> &lod_buffer, + std::vector &input_buffer, + std::vector &resize_heights, + std::vector &resize_widths, + int channels, int coarsest_stride = 1) { int batch_size = lod_buffer.size(); int max_h = -1; int max_w = -1; - for(int i = 0; i < batch_size; ++i) { + for (int i = 0; i < batch_size; ++i) { max_h = (max_h > resize_heights[i])? max_h:resize_heights[i]; max_w = (max_w > resize_widths[i])? max_w:resize_widths[i]; } - max_h = static_cast(ceil(static_cast(max_h) / static_cast(coarsest_stride)) * coarsest_stride); - max_w = static_cast(ceil(static_cast(max_w) / static_cast(coarsest_stride)) * coarsest_stride); + max_h = static_cast(ceil(static_cast(max_h) + / static_cast(coarsest_stride)) * coarsest_stride); + max_w = static_cast(ceil(static_cast(max_w) + / static_cast(coarsest_stride)) * coarsest_stride); std::cout << "max_w: " << max_w << " max_h: " << max_h << std::endl; - input_buffer.insert(input_buffer.end(), batch_size * channels * max_h * max_w, 0); + input_buffer.insert(input_buffer.end(), + batch_size * channels * max_h * max_w, 0); // flatten tensor and padding - for(int i = 0; i < lod_buffer.size(); ++i) { - float *input_buffer_ptr = input_buffer.data() + i * channels * max_h * max_w; + for (int i = 0; i < lod_buffer.size(); ++i) { + float *input_buffer_ptr = input_buffer.data() + + i * channels * max_h * max_w; const float *lod_ptr = lod_buffer[i].data(); - for(int c = 0; c < channels; ++c) { - for(int h = 0; h < resize_heights[i]; ++h) { - memcpy(input_buffer_ptr, lod_ptr, resize_widths[i] * sizeof(float)); + for (int c = 0; c < channels; ++c) { + for (int h = 0; h < resize_heights[i]; ++h) { + memcpy(input_buffer_ptr, lod_ptr, + resize_widths[i] * sizeof(float)); lod_ptr += resize_widths[i]; input_buffer_ptr += max_w; } @@ -49,44 +58,65 @@ namespace PaddleSolution { } } // change resize w, h - for(int i = 0; i < batch_size; ++i){ + for (int i = 0; i < batch_size; ++i) { resize_widths[i] = max_w; resize_heights[i] = max_h; } } - void output_detection_result(const float* out_addr, const std::vector> &lod_vector, const std::vector &imgs_batch){ - for(int i = 0; i < lod_vector[0].size() - 1; ++i) { + void output_detection_result(const float* out_addr, + const std::vector> &lod_vector, + const std::vector &imgs_batch) { + for (int i = 0; i < lod_vector[0].size() - 1; ++i) { DetectionResult detection_result; detection_result.set_filename(imgs_batch[i]); std::cout << imgs_batch[i] << ":" << std::endl; for (int j = lod_vector[0][i]; j < lod_vector[0][i+1]; ++j) { DetectionBox *box_ptr = detection_result.add_detection_boxes(); - box_ptr->set_class_(static_cast(round(out_addr[0 + j * 6]))); + box_ptr->set_class_( + static_cast(round(out_addr[0 + j * 6]))); box_ptr->set_score(out_addr[1 + j * 6]); box_ptr->set_left_top_x(out_addr[2 + j * 6]); box_ptr->set_left_top_y(out_addr[3 + j * 6]); box_ptr->set_right_bottom_x(out_addr[4 + j * 6]); box_ptr->set_right_bottom_y(out_addr[5 + j * 6]); printf("Class %d, score = %f, left top = [%f, %f], right bottom = [%f, %f]\n", - static_cast(round(out_addr[0 + j * 6])), out_addr[1 + j * 6], out_addr[2 + j * 6], - out_addr[3 + j * 6], out_addr[4 + j * 6], out_addr[5 + j * 6]); + static_cast(round(out_addr[0 + j * 6])), + out_addr[1 + j * 6], + out_addr[2 + j * 6], + out_addr[3 + j * 6], + out_addr[4 + j * 6], + out_addr[5 + j * 6]); } printf("\n"); - std::ofstream output(imgs_batch[i] + ".pb", std::ios::out | std::ios::trunc | std::ios::binary); + std::ofstream output(imgs_batch[i] + ".pb", + std::ios::out | std::ios::trunc | std::ios::binary); detection_result.SerializeToOstream(&output); output.close(); } } - + int DetectionPredictor::init(const std::string& conf) { if (!_model_config.load_config(conf)) { - LOG(FATAL) << "Fail to load config file: [" << conf << "]"; + #ifdef _WIN32 + std::cerr << "Fail to load config file: [" << conf << "], " + << "please check whether the config file path is correct" + << std::endl; + #else + LOG(FATAL) << "Fail to load config file: [" << conf << "], " + << "please check whether the config file path is correct"; + #endif return -1; } _preprocessor = PaddleSolution::create_processor(conf); if (_preprocessor == nullptr) { - LOG(FATAL) << "Failed to create_processor"; + #ifdef _WIN32 + std::cerr << "Failed to create_processor, please check whether you" + << " write a correct config file." << std::endl; + #else + LOG(FATAL) << "Failed to create_processor, please check whether" + << " you write a correct config file."; + #endif return -1; } @@ -116,20 +146,32 @@ namespace PaddleSolution { config.SetModel(prog_file, param_file); config.SwitchUseFeedFetchOps(false); config.SwitchSpecifyInputNames(true); - config.EnableMemoryOptim(); + config.EnableMemoryOptim(); + // config.SwitchIrOptim(true); + // config.EnableTensorRtEngine(1<<4, 30, 3); _main_predictor = paddle::CreatePaddlePredictor(config); } else { return -1; } return 0; - } int DetectionPredictor::predict(const std::vector& imgs) { + if (imgs.size() == 0) { + #ifdef _WIN32 + std::cerr << "No image found! Please check whether the images path" + << " is correct or the format of images is correct\n" + << "Supporting format: [.jpeg|.jpg|.JPEG|.JPG|.bmp|.BMP|.png|.PNG]" << std::endl; + #else + LOG(ERROR) << "No image found! Please check whether the images path" + << " is correct or the format of images is correct\n" + << "Supporting format: [.jpeg|.jpg|.JPEG|.JPG|.bmp|.BMP|.png|.PNG]"; + #endif + return -1; + } if (_model_config._predictor_mode == "NATIVE") { return native_predict(imgs); - } - else if (_model_config._predictor_mode == "ANALYSIS") { + } else if (_model_config._predictor_mode == "ANALYSIS") { return analysis_predict(imgs); } return -1; @@ -142,21 +184,24 @@ namespace PaddleSolution { int eval_width = _model_config._resize[0]; int eval_height = _model_config._resize[1]; std::size_t total_size = imgs.size(); - int default_batch_size = std::min(config_batch_size, (int)total_size); - int batch = total_size / default_batch_size + ((total_size % default_batch_size) != 0); - int batch_buffer_size = default_batch_size * channels * eval_width * eval_height; + int default_batch_size = std::min(config_batch_size, + static_cast(total_size)); + int batch = total_size / default_batch_size + + ((total_size % default_batch_size) != 0); + int batch_buffer_size = default_batch_size * channels + * eval_width * eval_height; auto& input_buffer = _buffer; auto& imgs_batch = _imgs_batch; float sr; - // DetectionResultsContainer result_container; for (int u = 0; u < batch; ++u) { int batch_size = default_batch_size; if (u == (batch - 1) && (total_size % default_batch_size)) { batch_size = total_size % default_batch_size; } - int real_buffer_size = batch_size * channels * eval_width * eval_height; + int real_buffer_size = batch_size * channels + * eval_width * eval_height; std::vector feeds; input_buffer.clear(); imgs_batch.clear(); @@ -175,58 +220,70 @@ namespace PaddleSolution { resize_heights.resize(batch_size); scale_ratios.resize(batch_size); std::vector> lod_buffer(batch_size); - if (!_preprocessor->batch_process(imgs_batch, lod_buffer, ori_widths.data(), ori_heights.data(), - resize_widths.data(), resize_heights.data(), scale_ratios.data())) { + if (!_preprocessor->batch_process(imgs_batch, lod_buffer, + ori_widths.data(), + ori_heights.data(), + resize_widths.data(), + resize_heights.data(), + scale_ratios.data())) { return -1; } - // flatten and padding - padding_minibatch(lod_buffer, input_buffer, resize_heights, resize_widths, channels, _model_config._coarsest_stride); + // flatten and padding + padding_minibatch(lod_buffer, input_buffer, resize_heights, + resize_widths, channels, + _model_config._coarsest_stride); paddle::PaddleTensor im_tensor, im_size_tensor, im_info_tensor; im_tensor.name = "image"; - im_tensor.shape = std::vector({ batch_size, channels, resize_heights[0], resize_widths[0] }); - im_tensor.data.Reset(input_buffer.data(), input_buffer.size() * sizeof(float)); + im_tensor.shape = std::vector({ batch_size, + channels, + resize_heights[0], + resize_widths[0] }); + im_tensor.data.Reset(input_buffer.data(), + input_buffer.size() * sizeof(float)); im_tensor.dtype = paddle::PaddleDType::FLOAT32; - + std::vector image_infos; - for(int i = 0; i < batch_size; ++i) { + for (int i = 0; i < batch_size; ++i) { image_infos.push_back(resize_heights[i]); image_infos.push_back(resize_widths[i]); image_infos.push_back(scale_ratios[i]); } im_info_tensor.name = "info"; im_info_tensor.shape = std::vector({batch_size, 3}); - im_info_tensor.data.Reset(image_infos.data(), batch_size * 3 * sizeof(float)); + im_info_tensor.data.Reset(image_infos.data(), + batch_size * 3 * sizeof(float)); im_info_tensor.dtype = paddle::PaddleDType::FLOAT32; - + std::vector image_size; - for(int i = 0; i < batch_size; ++i) { + for (int i = 0; i < batch_size; ++i) { image_size.push_back(ori_heights[i]); image_size.push_back(ori_widths[i]); } std::vector image_size_f; - for(int i = 0; i < batch_size; ++i) { + for (int i = 0; i < batch_size; ++i) { image_size_f.push_back(ori_heights[i]); image_size_f.push_back(ori_widths[i]); image_size_f.push_back(1.0); } - + int feeds_size = _model_config._feeds_size; im_size_tensor.name = "im_size"; - if(feeds_size == 2) { + if (feeds_size == 2) { im_size_tensor.shape = std::vector({ batch_size, 2}); - im_size_tensor.data.Reset(image_size.data(), batch_size * 2 * sizeof(int)); + im_size_tensor.data.Reset(image_size.data(), + batch_size * 2 * sizeof(int)); im_size_tensor.dtype = paddle::PaddleDType::INT32; - } - else if(feeds_size == 3) { + } else if (feeds_size == 3) { im_size_tensor.shape = std::vector({ batch_size, 3}); - im_size_tensor.data.Reset(image_size_f.data(), batch_size * 3 * sizeof(float)); + im_size_tensor.data.Reset(image_size_f.data(), + batch_size * 3 * sizeof(float)); im_size_tensor.dtype = paddle::PaddleDType::FLOAT32; } std::cout << "Feed size = " << feeds_size << std::endl; feeds.push_back(im_tensor); - if(_model_config._feeds_size > 2) { + if (_model_config._feeds_size > 2) { feeds.push_back(im_info_tensor); } feeds.push_back(im_size_tensor); @@ -234,7 +291,11 @@ namespace PaddleSolution { auto t1 = std::chrono::high_resolution_clock::now(); if (!_main_predictor->Run(feeds, &_outputs, batch_size)) { + #ifdef _WIN32 + std::cerr << "Failed: NativePredictor->Run() return false at batch: " << u; + #else LOG(ERROR) << "Failed: NativePredictor->Run() return false at batch: " << u; + #endif continue; } auto t2 = std::chrono::high_resolution_clock::now(); @@ -249,48 +310,51 @@ namespace PaddleSolution { std::cout << _outputs[0].shape[j] << ","; } std::cout << ")" << std::endl; - + // const size_t nums = _outputs.front().data.length() / sizeof(float); // if (out_num % batch_size != 0 || out_num != nums) { // LOG(ERROR) << "outputs data size mismatch with shape size."; // return -1; // } - float* out_addr = (float *)(_outputs[0].data.data()); + float* out_addr = reinterpret_cast(_outputs[0].data.data()); output_detection_result(out_addr, _outputs[0].lod, imgs_batch); } return 0; } - int DetectionPredictor::analysis_predict(const std::vector& imgs) { - + int DetectionPredictor::analysis_predict( + const std::vector& imgs) { int config_batch_size = _model_config._batch_size; int channels = _model_config._channels; int eval_width = _model_config._resize[0]; int eval_height = _model_config._resize[1]; auto total_size = imgs.size(); - int default_batch_size = std::min(config_batch_size, (int)total_size); - int batch = total_size / default_batch_size + ((total_size % default_batch_size) != 0); - int batch_buffer_size = default_batch_size * channels * eval_width * eval_height; + int default_batch_size = std::min(config_batch_size, + static_cast(total_size)); + int batch = total_size / default_batch_size + + ((total_size % default_batch_size) != 0); + int batch_buffer_size = default_batch_size * channels + * eval_width * eval_height; auto& input_buffer = _buffer; auto& imgs_batch = _imgs_batch; - //DetectionResultsContainer result_container; for (int u = 0; u < batch; ++u) { int batch_size = default_batch_size; if (u == (batch - 1) && (total_size % default_batch_size)) { batch_size = total_size % default_batch_size; } - int real_buffer_size = batch_size * channels * eval_width * eval_height; + int real_buffer_size = batch_size * channels * + eval_width * eval_height; std::vector feeds; - //input_buffer.resize(real_buffer_size); + // input_buffer.resize(real_buffer_size); input_buffer.clear(); imgs_batch.clear(); for (int i = 0; i < batch_size; ++i) { int idx = u * default_batch_size + i; imgs_batch.push_back(imgs[idx]); } - + std::vector ori_widths; std::vector ori_heights; std::vector resize_widths; @@ -301,56 +365,64 @@ namespace PaddleSolution { resize_widths.resize(batch_size); resize_heights.resize(batch_size); scale_ratios.resize(batch_size); - + std::vector> lod_buffer(batch_size); - if (!_preprocessor->batch_process(imgs_batch, lod_buffer, ori_widths.data(), ori_heights.data(), - resize_widths.data(), resize_heights.data(), scale_ratios.data())){ + if (!_preprocessor->batch_process(imgs_batch, lod_buffer, + ori_widths.data(), + ori_heights.data(), + resize_widths.data(), + resize_heights.data(), + scale_ratios.data())) { std::cout << "Failed to preprocess!" << std::endl; return -1; } - //flatten tensor - padding_minibatch(lod_buffer, input_buffer, resize_heights, resize_widths, channels, _model_config._coarsest_stride); + // flatten tensor + padding_minibatch(lod_buffer, input_buffer, resize_heights, + resize_widths, channels, + _model_config._coarsest_stride); std::vector input_names = _main_predictor->GetInputNames(); - auto im_tensor = _main_predictor->GetInputTensor(input_names.front()); - im_tensor->Reshape({ batch_size, channels, resize_heights[0], resize_widths[0] }); + auto im_tensor = _main_predictor->GetInputTensor( + input_names.front()); + im_tensor->Reshape({ batch_size, channels, + resize_heights[0], resize_widths[0] }); im_tensor->copy_from_cpu(input_buffer.data()); - - if(input_names.size() > 2){ + + if (input_names.size() > 2) { std::vector image_infos; - for(int i = 0; i < batch_size; ++i) { + for (int i = 0; i < batch_size; ++i) { image_infos.push_back(resize_heights[i]); image_infos.push_back(resize_widths[i]); image_infos.push_back(scale_ratios[i]); - } - auto im_info_tensor = _main_predictor->GetInputTensor(input_names[1]); + } + auto im_info_tensor = _main_predictor->GetInputTensor( + input_names[1]); im_info_tensor->Reshape({batch_size, 3}); im_info_tensor->copy_from_cpu(image_infos.data()); } std::vector image_size; - for(int i = 0; i < batch_size; ++i) { + for (int i = 0; i < batch_size; ++i) { image_size.push_back(ori_heights[i]); image_size.push_back(ori_widths[i]); } std::vector image_size_f; - for(int i = 0; i < batch_size; ++i) { + for (int i = 0; i < batch_size; ++i) { image_size_f.push_back(static_cast(ori_heights[i])); image_size_f.push_back(static_cast(ori_widths[i])); image_size_f.push_back(1.0); } - - auto im_size_tensor = _main_predictor->GetInputTensor(input_names.back()); - if(input_names.size() > 2) { + + auto im_size_tensor = _main_predictor->GetInputTensor( + input_names.back()); + if (input_names.size() > 2) { im_size_tensor->Reshape({batch_size, 3}); im_size_tensor->copy_from_cpu(image_size_f.data()); - } - else{ + } else { im_size_tensor->Reshape({batch_size, 2}); im_size_tensor->copy_from_cpu(image_size.data()); } - auto t1 = std::chrono::high_resolution_clock::now(); _main_predictor->ZeroCopyRun(); @@ -374,10 +446,10 @@ namespace PaddleSolution { out_data.resize(out_num); output_t->copy_to_cpu(out_data.data()); - float* out_addr = (float *)(out_data.data()); + float* out_addr = reinterpret_cast(out_data.data()); auto lod_vector = output_t->lod(); - output_detection_result(out_addr, lod_vector, imgs_batch); + output_detection_result(out_addr, lod_vector, imgs_batch); } return 0; } -} +} // namespace PaddleSolution diff --git a/inference/predictor/detection_predictor.h b/inference/predictor/detection_predictor.h index 3bc4cfdd7..2ce6f4520 100644 --- a/inference/predictor/detection_predictor.h +++ b/inference/predictor/detection_predictor.h @@ -4,7 +4,7 @@ // you may not use this file except in compliance with the License. // You may obtain a copy of the License at // -// http://www.apache.org/licenses/LICENSE-2.0 +// http://www.apache.org/licenses/LICENSE-2.0 // // Unless required by applicable law or agreed to in writing, software // distributed under the License is distributed on an "AS IS" BASIS, @@ -14,39 +14,40 @@ #pragma once +#include +#include +#include + #include #include #include #include #include #include -#include -#include #include -#include -#include -#include -#include +#include "utils/conf_parser.h" +#include "utils/utils.h" +#include "preprocessor/preprocessor.h" namespace PaddleSolution { - class DetectionPredictor { - public: - // init a predictor with a yaml config file - int init(const std::string& conf); - // predict api - int predict(const std::vector& imgs); +class DetectionPredictor { + public: + // init a predictor with a yaml config file + int init(const std::string& conf); + // predict api + int predict(const std::vector& imgs); - private: - int native_predict(const std::vector& imgs); - int analysis_predict(const std::vector& imgs); - private: - std::vector _buffer; - std::vector _imgs_batch; - std::vector _outputs; + private: + int native_predict(const std::vector& imgs); + int analysis_predict(const std::vector& imgs); + private: + std::vector _buffer; + std::vector _imgs_batch; + std::vector _outputs; - PaddleSolution::PaddleModelConfigPaser _model_config; - std::shared_ptr _preprocessor; - std::unique_ptr _main_predictor; - }; -} + PaddleSolution::PaddleModelConfigPaser _model_config; + std::shared_ptr _preprocessor; + std::unique_ptr _main_predictor; +}; +} // namespace PaddleSolution diff --git a/inference/preprocessor/preprocessor.cpp b/inference/preprocessor/preprocessor.cpp index dbe7bcf62..941416a04 100644 --- a/inference/preprocessor/preprocessor.cpp +++ b/inference/preprocessor/preprocessor.cpp @@ -16,14 +16,18 @@ #include "preprocessor.h" #include "preprocessor_detection.h" +#include namespace PaddleSolution { std::shared_ptr create_processor(const std::string& conf_file) { - auto config = std::make_shared(); if (!config->load_config(conf_file)) { - LOG(FATAL) << "fail to laod conf file [" << conf_file << "]"; + #ifdef _WIN32 + std::cerr << "fail to load conf file [" << conf_file << "]" << std::endl; + #else + LOG(FATAL) << "fail to load conf file [" << conf_file << "]"; + #endif return nullptr; } @@ -34,10 +38,13 @@ namespace PaddleSolution { } return p; } - - - LOG(FATAL) << "unknown processor_name [" << config->_pre_processor << "]"; - + #ifdef _WIN32 + std::cerr << "unknown processor_name [" << config->_pre_processor << "]," + << "please check whether PRE_PROCESSOR is set correctly" << std::endl; + #else + LOG(FATAL) << "unknown processor_name [" << config->_pre_processor << "]," + << "please check whether PRE_PROCESSOR is set correctly"; + #endif return nullptr; } -} +} // namespace PaddleSolution diff --git a/inference/preprocessor/preprocessor.h b/inference/preprocessor/preprocessor.h index a3fb2e029..b5cfa467f 100644 --- a/inference/preprocessor/preprocessor.h +++ b/inference/preprocessor/preprocessor.h @@ -26,39 +26,52 @@ namespace PaddleSolution { class ImagePreProcessor { -protected: - ImagePreProcessor() {}; - -public: + protected: + ImagePreProcessor() {} + + public: virtual ~ImagePreProcessor() {} - virtual bool single_process(const std::string& fname, float* data, int* ori_w, int* ori_h) { + virtual bool single_process(const std::string& fname, + float* data, + int* ori_w, + int* ori_h) { return true; } - virtual bool batch_process(const std::vector& imgs, float* data, int* ori_w, int* ori_h) { + virtual bool batch_process(const std::vector& imgs, + float* data, + int* ori_w, + int* ori_h) { return true; } virtual bool single_process(const std::string& fname, float* data) { return true; } - - virtual bool batch_process(const std::vector& imgs, float* data) { + + virtual bool batch_process(const std::vector& imgs, + float* data) { return true; } - - virtual bool single_process(const std::string& fname, std::vector &data, int* ori_w, int* ori_h, int* resize_w, int* resize_h, float* scale_ratio) { - return true; - } - virtual bool batch_process(const std::vector& imgs, std::vector> &data, int* ori_w, int* ori_h, int* resize_w, int* resize_h, float* scale_ratio) { - return true; + virtual bool single_process(const std::string& fname, + std::vector &data, + int* ori_w, int* ori_h, + int* resize_w, int* resize_h, + float* scale_ratio) { + return true; } -}; // end of class ImagePreProcessor - -std::shared_ptr create_processor(const std::string &config_file); + virtual bool batch_process(const std::vector& imgs, + std::vector> &data, + int* ori_w, int* ori_h, int* resize_w, + int* resize_h, float* scale_ratio) { + return true; + } +}; // end of class ImagePreProcessor -} // end of namespace paddle_solution +std::shared_ptr + create_processor(const std::string &config_file); +} // namespace PaddleSolution diff --git a/inference/preprocessor/preprocessor_detection.cpp b/inference/preprocessor/preprocessor_detection.cpp index ba8fd0e32..15d2dc3c4 100644 --- a/inference/preprocessor/preprocessor_detection.cpp +++ b/inference/preprocessor/preprocessor_detection.cpp @@ -4,7 +4,7 @@ // you may not use this file except in compliance with the License. // You may obtain a copy of the License at // -// http://www.apache.org/licenses/LICENSE-2.0 +// http://www.apache.org/licenses/LICENSE-2.0 // // Unless required by applicable law or agreed to in writing, software // distributed under the License is distributed on an "AS IS" BASIS, @@ -12,119 +12,133 @@ // See the License for the specific language governing permissions and // limitations under the License. +#include #include #include -#include - #include "preprocessor_detection.h" #include "utils/utils.h" namespace PaddleSolution { - bool DetectionPreProcessor::single_process(const std::string& fname, std::vector &vec_data, int* ori_w, int* ori_h, int* resize_w, int* resize_h, float* scale_ratio) { - cv::Mat im1 = cv::imread(fname, -1); - cv::Mat im; - if(_config->_feeds_size == 3) { // faster rcnn - im1.convertTo(im, CV_32FC3, 1/255.0); - } - else if(_config->_feeds_size == 2){ //yolo v3 - im = im1; - } - if (im.data == nullptr || im.empty()) { - LOG(ERROR) << "Failed to open image: " << fname; - return false; - } - - int channels = im.channels(); - if (channels == 1) { - cv::cvtColor(im, im, cv::COLOR_GRAY2BGR); - } - channels = im.channels(); - if (channels != 3 && channels != 4) { - LOG(ERROR) << "Only support rgb(gray) and rgba image."; - return false; - } - *ori_w = im.cols; - *ori_h = im.rows; - cv::cvtColor(im, im, cv::COLOR_BGR2RGB); - //channels = im.channels(); +bool DetectionPreProcessor::single_process(const std::string& fname, + std::vector &vec_data, + int* ori_w, int* ori_h, + int* resize_w, int* resize_h, + float* scale_ratio) { + cv::Mat im1 = cv::imread(fname, -1); + cv::Mat im; + if (_config->_feeds_size == 3) { // faster rcnn + im1.convertTo(im, CV_32FC3, 1/255.0); + } else if (_config->_feeds_size == 2) { // yolo v3 + im = im1; + } + if (im.data == nullptr || im.empty()) { + #ifdef _WIN32 + std::cerr << "Failed to open image: " << fname << std::endl; + #else + LOG(ERROR) << "Failed to open image: " << fname; + #endif + return false; + } + int channels = im.channels(); + if (channels == 1) { + cv::cvtColor(im, im, cv::COLOR_GRAY2BGR); + } + channels = im.channels(); + if (channels != 3 && channels != 4) { + #ifdef _WIN32 + std::cerr << "Only support rgb(gray) and rgba image." << std::endl; + #else + LOG(ERROR) << "Only support rgb(gray) and rgba image."; + #endif + return false; + } + *ori_w = im.cols; + *ori_h = im.rows; + cv::cvtColor(im, im, cv::COLOR_BGR2RGB); + // channels = im.channels(); - //resize - int rw = im.cols; - int rh = im.rows; - float im_scale_ratio; - utils::scaling(_config->_resize_type, rw, rh, _config->_resize[0], _config->_resize[1], _config->_target_short_size, _config->_resize_max_size, im_scale_ratio); - cv::Size resize_size(rw, rh); - *resize_w = rw; - *resize_h = rh; - *scale_ratio = im_scale_ratio; - if (*ori_h != rh || *ori_w != rw) { - cv::Mat im_temp; - if(_config->_resize_type == utils::SCALE_TYPE::UNPADDING) { - cv::resize(im, im_temp, resize_size, 0, 0, cv::INTER_LINEAR); - } - else if(_config->_resize_type == utils::SCALE_TYPE::RANGE_SCALING) { - cv::resize(im, im_temp, cv::Size(), im_scale_ratio, im_scale_ratio, cv::INTER_LINEAR); - } - im = im_temp; + // resize + int rw = im.cols; + int rh = im.rows; + float im_scale_ratio; + utils::scaling(_config->_resize_type, rw, rh, _config->_resize[0], + _config->_resize[1], _config->_target_short_size, + _config->_resize_max_size, im_scale_ratio); + cv::Size resize_size(rw, rh); + *resize_w = rw; + *resize_h = rh; + *scale_ratio = im_scale_ratio; + if (*ori_h != rh || *ori_w != rw) { + cv::Mat im_temp; + if (_config->_resize_type == utils::SCALE_TYPE::UNPADDING) { + cv::resize(im, im_temp, resize_size, 0, 0, cv::INTER_LINEAR); + } else if (_config->_resize_type == utils::SCALE_TYPE::RANGE_SCALING) { + cv::resize(im, im_temp, cv::Size(), im_scale_ratio, + im_scale_ratio, cv::INTER_LINEAR); } + im = im_temp; + } - vec_data.resize(channels * rw * rh); - float *data = vec_data.data(); + vec_data.resize(channels * rw * rh); + float *data = vec_data.data(); - float* pmean = _config->_mean.data(); - float* pscale = _config->_std.data(); - for (int h = 0; h < rh; ++h) { - const uchar* uptr = im.ptr(h); - const float* fptr = im.ptr(h); - int im_index = 0; - for (int w = 0; w < rw; ++w) { - for (int c = 0; c < channels; ++c) { - int top_index = (c * rh + h) * rw + w; - float pixel;// = static_cast(fptr[im_index]);// / 255.0; - if(_config->_feeds_size == 2){ //yolo v3 - pixel = static_cast(uptr[im_index++]) / 255.0; - } - else if(_config->_feeds_size == 3){ - pixel = fptr[im_index++]; - } - pixel = (pixel - pmean[c]) / pscale[c]; - data[top_index] = pixel; + float* pmean = _config->_mean.data(); + float* pscale = _config->_std.data(); + for (int h = 0; h < rh; ++h) { + const uchar* uptr = im.ptr(h); + const float* fptr = im.ptr(h); + int im_index = 0; + for (int w = 0; w < rw; ++w) { + for (int c = 0; c < channels; ++c) { + int top_index = (c * rh + h) * rw + w; + float pixel; + if (_config->_feeds_size == 2) { // yolo v3 + pixel = static_cast(uptr[im_index++]) / 255.0; + } else if (_config->_feeds_size == 3) { + pixel = fptr[im_index++]; } + pixel = (pixel - pmean[c]) / pscale[c]; + data[top_index] = pixel; } } - return true; } + return true; +} - bool DetectionPreProcessor::batch_process(const std::vector& imgs, std::vector> &data, int* ori_w, int* ori_h, int* resize_w, int* resize_h, float* scale_ratio) { - auto ic = _config->_channels; - auto iw = _config->_resize[0]; - auto ih = _config->_resize[1]; - std::vector threads; - for (int i = 0; i < imgs.size(); ++i) { - std::string path = imgs[i]; - int* width = &ori_w[i]; - int* height = &ori_h[i]; - int* resize_width = &resize_w[i]; - int* resize_height = &resize_h[i]; - float* sr = &scale_ratio[i]; - threads.emplace_back([this, &data, i, path, width, height, resize_width, resize_height, sr] { - std::vector buffer; - single_process(path, buffer, width, height, resize_width, resize_height, sr); - data[i] = buffer; - }); - } - for (auto& t : threads) { - if (t.joinable()) { - t.join(); - } - } - return true; +bool DetectionPreProcessor::batch_process(const std::vector& imgs, + std::vector> &data, + int* ori_w, int* ori_h, int* resize_w, + int* resize_h, float* scale_ratio) { + auto ic = _config->_channels; + auto iw = _config->_resize[0]; + auto ih = _config->_resize[1]; + std::vector threads; + for (int i = 0; i < imgs.size(); ++i) { + std::string path = imgs[i]; + int* width = &ori_w[i]; + int* height = &ori_h[i]; + int* resize_width = &resize_w[i]; + int* resize_height = &resize_h[i]; + float* sr = &scale_ratio[i]; + threads.emplace_back([this, &data, i, path, width, height, + resize_width, resize_height, sr] { + std::vector buffer; + single_process(path, buffer, width, height, resize_width, + resize_height, sr); + data[i] = buffer; + }); } - - bool DetectionPreProcessor::init(std::shared_ptr config) { - _config = config; - return true; + for (auto& t : threads) { + if (t.joinable()) { + t.join(); + } } + return true; +} +bool DetectionPreProcessor::init(std::shared_ptr config) { + _config = config; + return true; } +} // namespace PaddleSolution diff --git a/inference/preprocessor/preprocessor_detection.h b/inference/preprocessor/preprocessor_detection.h index 731329040..7ee1678e3 100644 --- a/inference/preprocessor/preprocessor_detection.h +++ b/inference/preprocessor/preprocessor_detection.h @@ -4,7 +4,7 @@ // you may not use this file except in compliance with the License. // You may obtain a copy of the License at // -// http://www.apache.org/licenses/LICENSE-2.0 +// http://www.apache.org/licenses/LICENSE-2.0 // // Unless required by applicable law or agreed to in writing, software // distributed under the License is distributed on an "AS IS" BASIS, @@ -18,19 +18,23 @@ namespace PaddleSolution { - class DetectionPreProcessor : public ImagePreProcessor { +class DetectionPreProcessor : public ImagePreProcessor { + public: + DetectionPreProcessor() : _config(nullptr) { + } - public: - DetectionPreProcessor() : _config(nullptr) { - }; + bool init(std::shared_ptr config); - bool init(std::shared_ptr config); - - bool single_process(const std::string& fname, std::vector &data, int* ori_w, int* ori_h, int* resize_w, int* resize_h, float* scale_ratio); + bool single_process(const std::string& fname, std::vector &data, + int* ori_w, int* ori_h, int* resize_w, + int* resize_h, float* scale_ratio); - bool batch_process(const std::vector& imgs, std::vector> &data, int* ori_w, int* ori_h, int* resize_w, int* resize_h, float* scale_ratio); - private: - std::shared_ptr _config; - }; + bool batch_process(const std::vector& imgs, + std::vector> &data, + int* ori_w, int* ori_h, int* resize_w, + int* resize_h, float* scale_ratio); + private: + std::shared_ptr _config; +}; -} +} // namespace PaddleSolution diff --git a/inference/tools/detection_result_pb2.py b/inference/tools/detection_result_pb2.py index 80a6a82a3..153a41df4 100644 --- a/inference/tools/detection_result_pb2.py +++ b/inference/tools/detection_result_pb2.py @@ -134,7 +134,8 @@ _DETECTIONBOX = _descriptor.Descriptor( extension_ranges=[], oneofs=[], serialized_start=43, - serialized_end=175, ) + serialized_end=175, +) _DETECTIONRESULT = _descriptor.Descriptor( name='DetectionResult', @@ -185,7 +186,8 @@ _DETECTIONRESULT = _descriptor.Descriptor( extension_ranges=[], oneofs=[], serialized_start=177, - serialized_end=267, ) + serialized_end=267, +) _DETECTIONRESULT.fields_by_name['detection_boxes'].message_type = _DETECTIONBOX DESCRIPTOR.message_types_by_name['DetectionBox'] = _DETECTIONBOX @@ -193,22 +195,20 @@ DESCRIPTOR.message_types_by_name['DetectionResult'] = _DETECTIONRESULT DetectionBox = _reflection.GeneratedProtocolMessageType( 'DetectionBox', - (_message.Message, ), - dict( - DESCRIPTOR=_DETECTIONBOX, - __module__='detection_result_pb2' - # @@protoc_insertion_point(class_scope:PaddleSolution.DetectionBox) - )) + (_message.Message,), + dict(DESCRIPTOR=_DETECTIONBOX, + __module__='detection_result_pb2' + # @@protoc_insertion_point(class_scope:PaddleSolution.DetectionBox) + )) _sym_db.RegisterMessage(DetectionBox) DetectionResult = _reflection.GeneratedProtocolMessageType( 'DetectionResult', - (_message.Message, ), - dict( - DESCRIPTOR=_DETECTIONRESULT, - __module__='detection_result_pb2' - # @@protoc_insertion_point(class_scope:PaddleSolution.DetectionResult) - )) + (_message.Message,), + dict(DESCRIPTOR=_DETECTIONRESULT, + __module__='detection_result_pb2' + # @@protoc_insertion_point(class_scope:PaddleSolution.DetectionResult) + )) _sym_db.RegisterMessage(DetectionResult) # @@protoc_insertion_point(module_scope) diff --git a/inference/tools/vis.py b/inference/tools/vis.py index d0f9fc841..f27ae1f83 100644 --- a/inference/tools/vis.py +++ b/inference/tools/vis.py @@ -85,8 +85,8 @@ if __name__ == "__main__": for box in detection_result.detection_boxes: if box.score >= Flags.threshold: box_class = getattr(box, 'class') - text_class_score_str = "%s %.2f" % ( - class2LabelMap.get(str(box_class)), box.score) + text_class_score_str = "%s %.2f" % (class2LabelMap.get( + str(box_class)), box.score) text_point = (int(box.left_top_x), int(box.left_top_y)) ptLeftTop = (int(box.left_top_x), int(box.left_top_y)) @@ -106,8 +106,8 @@ if __name__ == "__main__": text_box_left_top = (text_point[0], text_point[1] - text_size[0][1]) - text_box_right_bottom = ( - text_point[0] + text_size[0][0], text_point[1]) + text_box_right_bottom = (text_point[0] + + text_size[0][0], text_point[1]) cv2.rectangle(img, text_box_left_top, text_box_right_bottom, color, -1, 8) diff --git a/inference/utils/conf_parser.h b/inference/utils/conf_parser.h index 21944d032..97461a6e9 100644 --- a/inference/utils/conf_parser.h +++ b/inference/utils/conf_parser.h @@ -4,7 +4,7 @@ // you may not use this file except in compliance with the License. // You may obtain a copy of the License at // -// http://www.apache.org/licenses/LICENSE-2.0 +// http://www.apache.org/licenses/LICENSE-2.0 // // Unless required by applicable law or agreed to in writing, software // distributed under the License is distributed on an "AS IS" BASIS, @@ -13,225 +13,285 @@ // limitations under the License. #pragma once +#include #include #include #include #include -#include namespace PaddleSolution { - class PaddleModelConfigPaser { - std::map _scaling_map; - public: - PaddleModelConfigPaser() - :_class_num(0), - _channels(0), - _use_gpu(0), - _batch_size(1), - _target_short_size(0), - _model_file_name("__model__"), - _param_file_name("__params__"), - _scaling_map{{"UNPADDING", 0}, - {"RANGE_SCALING",1}}, - _feeds_size(1), - _coarsest_stride(1) - { - } - ~PaddleModelConfigPaser() { - } +class PaddleModelConfigPaser { + std::map _scaling_map; - void reset() { - _crop_size.clear(); - _resize.clear(); - _mean.clear(); - _std.clear(); - _img_type.clear(); - _class_num = 0; - _channels = 0; - _use_gpu = 0; - _target_short_size = 0; - _batch_size = 1; - _model_file_name = "__model__"; - _model_path = "./"; - _param_file_name="__params__"; - _resize_type = 0; - _resize_max_size = 0; - _feeds_size = 1; - _coarsest_stride = 1; - } + public: + PaddleModelConfigPaser() + :_class_num(0), + _channels(0), + _use_gpu(0), + _batch_size(1), + _target_short_size(0), + _model_file_name("__model__"), + _param_file_name("__params__"), + _scaling_map{{"UNPADDING", 0}, + {"RANGE_SCALING", 1}}, + _feeds_size(1), + _coarsest_stride(1) {} - std::string process_parenthesis(const std::string& str) { - if (str.size() < 2) { - return str; - } - std::string nstr(str); - if (str[0] == '(' && str.back() == ')') { - nstr[0] = '['; - nstr[str.size() - 1] = ']'; - } - return nstr; - } + ~PaddleModelConfigPaser() {} - template - std::vector parse_str_to_vec(const std::string& str) { - std::vector data; - auto node = YAML::Load(str); - for (const auto& item : node) { - data.push_back(item.as()); - } - return data; + void reset() { + _crop_size.clear(); + _resize.clear(); + _mean.clear(); + _std.clear(); + _img_type.clear(); + _class_num = 0; + _channels = 0; + _use_gpu = 0; + _target_short_size = 0; + _batch_size = 1; + _model_file_name = "__model__"; + _model_path = "./"; + _param_file_name = "__params__"; + _resize_type = 0; + _resize_max_size = 0; + _feeds_size = 1; + _coarsest_stride = 1; + } + + std::string process_parenthesis(const std::string& str) { + if (str.size() < 2) { + return str; + } + std::string nstr(str); + if (str[0] == '(' && str.back() == ')') { + nstr[0] = '['; + nstr[str.size() - 1] = ']'; } + return nstr; + } - bool load_config(const std::string& conf_file) { - - reset(); + template + std::vector parse_str_to_vec(const std::string& str) { + std::vector data; + auto node = YAML::Load(str); + for (const auto& item : node) { + data.push_back(item.as()); + } + return data; + } - YAML::Node config = YAML::LoadFile(conf_file); - // 1. get resize + bool load_config(const std::string& conf_file) { + reset(); + YAML::Node config; + try { + config = YAML::LoadFile(conf_file); + } catch(...) { + return false; + } + // 1. get resize + if (config["DEPLOY"]["EVAL_CROP_SIZE"].IsDefined()) { auto str = config["DEPLOY"]["EVAL_CROP_SIZE"].as(); _resize = parse_str_to_vec(process_parenthesis(str)); + } else { + std::cerr << "Please set EVAL_CROP_SIZE: (xx, xx)" << std::endl; + return false; + } + // 0. get crop_size + if (config["DEPLOY"]["CROP_SIZE"].IsDefined()) { + auto crop_str = config["DEPLOY"]["CROP_SIZE"].as(); + _crop_size = parse_str_to_vec(process_parenthesis(crop_str)); + } else { + _crop_size = _resize; + } - // 0. get crop_size - if(config["DEPLOY"]["CROP_SIZE"].IsDefined()) { - auto crop_str = config["DEPLOY"]["CROP_SIZE"].as(); - _crop_size = parse_str_to_vec(process_parenthesis(crop_str)); - } - else { - _crop_size = _resize; - } - - // 2. get mean + // 2. get mean + if (config["DEPLOY"]["MEAN"].IsDefined()) { for (const auto& item : config["DEPLOY"]["MEAN"]) { _mean.push_back(item.as()); } - - // 3. get std + } else { + std::cerr << "Please set MEAN: [xx, xx, xx]" << std::endl; + return false; + } + // 3. get std + if(config["DEPLOY"]["STD"].IsDefined()) { for (const auto& item : config["DEPLOY"]["STD"]) { _std.push_back(item.as()); } - - // 4. get image type + } else { + std::cerr << "Please set STD: [xx, xx, xx]" << std::endl; + return false; + } + // 4. get image type + if (config["DEPLOY"]["IMAGE_TYPE"].IsDefined()) { _img_type = config["DEPLOY"]["IMAGE_TYPE"].as(); - // 5. get class number + } else { + std::cerr << "Please set IMAGE_TYPE: \"rgb\" or \"rgba\"" << std::endl; + return false; + } + // 5. get class number + if (config["DEPLOY"]["NUM_CLASSES"].IsDefined()) { _class_num = config["DEPLOY"]["NUM_CLASSES"].as(); - // 7. set model path + } else { + std::cerr << "Please set NUM_CLASSES: x" << std::endl; + return false; + } + // 7. set model path + if (config["DEPLOY"]["MODEL_PATH"].IsDefined()) { _model_path = config["DEPLOY"]["MODEL_PATH"].as(); - // 8. get model file_name + } else { + std::cerr << "Please set MODEL_PATH: \"/path/to/model_dir\"" << std::endl; + return false; + } + // 8. get model file_name + if (config["DEPLOY"]["MODEL_FILENAME"].IsDefined()) { _model_file_name = config["DEPLOY"]["MODEL_FILENAME"].as(); - // 9. get model param file name - _param_file_name = config["DEPLOY"]["PARAMS_FILENAME"].as(); - // 10. get pre_processor + } else { + _model_file_name = "__model__"; + } + // 9. get model param file name + if (config["DEPLOY"]["PARAMS_FILENAME"].IsDefined()) { + _param_file_name + = config["DEPLOY"]["PARAMS_FILENAME"].as(); + } else { + _param_file_name = "__params__"; + } + // 10. get pre_processor + if (config["DEPLOY"]["PRE_PROCESSOR"].IsDefined()) { _pre_processor = config["DEPLOY"]["PRE_PROCESSOR"].as(); - // 11. use_gpu + } else { + std::cerr << "Please set PRE_PROCESSOR: \"DetectionPreProcessor\"" << std::endl; + return false; + } + // 11. use_gpu + if (config["DEPLOY"]["USE_GPU"].IsDefined()) { _use_gpu = config["DEPLOY"]["USE_GPU"].as(); - // 12. predictor_mode + } else { + _use_gpu = 0; + } + // 12. predictor_mode + if (config["DEPLOY"]["PREDICTOR_MODE"].IsDefined()) { _predictor_mode = config["DEPLOY"]["PREDICTOR_MODE"].as(); - // 13. batch_size + } else { + std::cerr << "Please set PREDICTOR_MODE: \"NATIVE\" or \"ANALYSIS\"" << std::endl; + return false; + } + // 13. batch_size + if (config["DEPLOY"]["BATCH_SIZE"].IsDefined()) { _batch_size = config["DEPLOY"]["BATCH_SIZE"].as(); - // 14. channels + } else { + _batch_size = 1; + } + // 14. channels + if (config["DEPLOY"]["CHANNELS"].IsDefined()) { _channels = config["DEPLOY"]["CHANNELS"].as(); - // 15. target_short_size - if(config["DEPLOY"]["TARGET_SHORT_SIZE"].IsDefined()) { - _target_short_size = config["DEPLOY"]["TARGET_SHORT_SIZE"].as(); - } - // 16.resize_type - if(config["DEPLOY"]["RESIZE_TYPE"].IsDefined() && - _scaling_map.find(config["DEPLOY"]["RESIZE_TYPE"].as()) != _scaling_map.end()) { - _resize_type = _scaling_map[config["DEPLOY"]["RESIZE_TYPE"].as()]; - } - else{ - _resize_type = 0; - } - // 17.resize_max_size - if(config["DEPLOY"]["RESIZE_MAX_SIZE"].IsDefined()) { - _resize_max_size = config["DEPLOY"]["RESIZE_MAX_SIZE"].as(); - } - // 18.feeds_size - if(config["DEPLOY"]["FEEDS_SIZE"].IsDefined()){ - _feeds_size = config["DEPLOY"]["FEEDS_SIZE"].as(); - } - // 19. coarsest_stride - if(config["DEPLOY"]["COARSEST_STRIDE"].IsDefined()) { - _coarsest_stride = config["DEPLOY"]["COARSEST_STRIDE"].as(); - } - return true; + } else { + std::cerr << "Please set CHANNELS: x" << std::endl; + return false; + } + // 15. target_short_size + if (config["DEPLOY"]["TARGET_SHORT_SIZE"].IsDefined()) { + _target_short_size = config["DEPLOY"]["TARGET_SHORT_SIZE"].as(); + } + // 16.resize_type + if (config["DEPLOY"]["RESIZE_TYPE"].IsDefined() && + _scaling_map.find(config["DEPLOY"]["RESIZE_TYPE"].as()) != _scaling_map.end()) { + _resize_type = _scaling_map[config["DEPLOY"]["RESIZE_TYPE"].as()]; + } else { + _resize_type = 0; + } + // 17.resize_max_size + if (config["DEPLOY"]["RESIZE_MAX_SIZE"].IsDefined()) { + _resize_max_size = config["DEPLOY"]["RESIZE_MAX_SIZE"].as(); } + // 18.feeds_size + if (config["DEPLOY"]["FEEDS_SIZE"].IsDefined()) { + _feeds_size = config["DEPLOY"]["FEEDS_SIZE"].as(); + } + // 19. coarsest_stride + if (config["DEPLOY"]["COARSEST_STRIDE"].IsDefined()) { + _coarsest_stride = config["DEPLOY"]["COARSEST_STRIDE"].as(); + } + return true; + } - void debug() const { - - std::cout << "SCALE_RESIZE: (" << _resize[0] << ", " << _resize[1] << ")" << std::endl; + void debug() const { + std::cout << "SCALE_RESIZE: (" << _resize[0] << ", " + << _resize[1] << ")" << std::endl; - std::cout << "MEAN: ["; - for (int i = 0; i < _mean.size(); ++i) { - if (i != _mean.size() - 1) { - std::cout << _mean[i] << ", "; - } else { - std::cout << _mean[i]; - } + std::cout << "MEAN: ["; + for (int i = 0; i < _mean.size(); ++i) { + if (i != _mean.size() - 1) { + std::cout << _mean[i] << ", "; + } else { + std::cout << _mean[i]; } - std::cout << "]" << std::endl; + } + std::cout << "]" << std::endl; - std::cout << "STD: ["; - for (int i = 0; i < _std.size(); ++i) { - if (i != _std.size() - 1) { - std::cout << _std[i] << ", "; - } - else { - std::cout << _std[i]; - } + std::cout << "STD: ["; + for (int i = 0; i < _std.size(); ++i) { + if (i != _std.size() - 1) { + std::cout << _std[i] << ", "; + } else { + std::cout << _std[i]; } - std::cout << "]" << std::endl; - std::cout << "DEPLOY.TARGET_SHORT_SIZE: " << _target_short_size << std::endl; - std::cout << "DEPLOY.IMAGE_TYPE: " << _img_type << std::endl; - std::cout << "DEPLOY.NUM_CLASSES: " << _class_num << std::endl; - std::cout << "DEPLOY.CHANNELS: " << _channels << std::endl; - std::cout << "DEPLOY.MODEL_PATH: " << _model_path << std::endl; - std::cout << "DEPLOY.MODEL_FILENAME: " << _model_file_name << std::endl; - std::cout << "DEPLOY.PARAMS_FILENAME: " << _param_file_name << std::endl; - std::cout << "DEPLOY.PRE_PROCESSOR: " << _pre_processor << std::endl; - std::cout << "DEPLOY.USE_GPU: " << _use_gpu << std::endl; - std::cout << "DEPLOY.PREDICTOR_MODE: " << _predictor_mode << std::endl; - std::cout << "DEPLOY.BATCH_SIZE: " << _batch_size << std::endl; - } - //DEPLOY.COARSEST_STRIDE - int _coarsest_stride; - // DEPLOY.FEEDS_SIZE - int _feeds_size; - // DEPLOY.RESIZE_TYPE 0:unpadding 1:rangescaling Default:0 - int _resize_type; - // DEPLOY.RESIZE_MAX_SIZE - int _resize_max_size; - // DEPLOY.CROP_SIZE - std::vector _crop_size; - // DEPLOY.SCALE_RESIZE - std::vector _resize; - // DEPLOY.MEAN - std::vector _mean; - // DEPLOY.STD - std::vector _std; - // DEPLOY.IMAGE_TYPE - std::string _img_type; - // DEPLOY.TARGET_SHORT_SIZE - int _target_short_size; - // DEPLOY.NUM_CLASSES - int _class_num; - // DEPLOY.CHANNELS - int _channels; - // DEPLOY.MODEL_PATH - std::string _model_path; - // DEPLOY.MODEL_FILENAME - std::string _model_file_name; - // DEPLOY.PARAMS_FILENAME - std::string _param_file_name; - // DEPLOY.PRE_PROCESSOR - std::string _pre_processor; - // DEPLOY.USE_GPU - int _use_gpu; - // DEPLOY.PREDICTOR_MODE - std::string _predictor_mode; - // DEPLOY.BATCH_SIZE - int _batch_size; - }; - -} + } + std::cout << "]" << std::endl; + std::cout << "DEPLOY.TARGET_SHORT_SIZE: " << _target_short_size + << std::endl; + std::cout << "DEPLOY.IMAGE_TYPE: " << _img_type << std::endl; + std::cout << "DEPLOY.NUM_CLASSES: " << _class_num << std::endl; + std::cout << "DEPLOY.CHANNELS: " << _channels << std::endl; + std::cout << "DEPLOY.MODEL_PATH: " << _model_path << std::endl; + std::cout << "DEPLOY.MODEL_FILENAME: " << _model_file_name + << std::endl; + std::cout << "DEPLOY.PARAMS_FILENAME: " << _param_file_name + << std::endl; + std::cout << "DEPLOY.PRE_PROCESSOR: " << _pre_processor << std::endl; + std::cout << "DEPLOY.USE_GPU: " << _use_gpu << std::endl; + std::cout << "DEPLOY.PREDICTOR_MODE: " << _predictor_mode << std::endl; + std::cout << "DEPLOY.BATCH_SIZE: " << _batch_size << std::endl; + } + // DEPLOY.COARSEST_STRIDE + int _coarsest_stride; + // DEPLOY.FEEDS_SIZE + int _feeds_size; + // DEPLOY.RESIZE_TYPE 0:unpadding 1:rangescaling Default:0 + int _resize_type; + // DEPLOY.RESIZE_MAX_SIZE + int _resize_max_size; + // DEPLOY.CROP_SIZE + std::vector _crop_size; + // DEPLOY.SCALE_RESIZE + std::vector _resize; + // DEPLOY.MEAN + std::vector _mean; + // DEPLOY.STD + std::vector _std; + // DEPLOY.IMAGE_TYPE + std::string _img_type; + // DEPLOY.TARGET_SHORT_SIZE + int _target_short_size; + // DEPLOY.NUM_CLASSES + int _class_num; + // DEPLOY.CHANNELS + int _channels; + // DEPLOY.MODEL_PATH + std::string _model_path; + // DEPLOY.MODEL_FILENAME + std::string _model_file_name; + // DEPLOY.PARAMS_FILENAME + std::string _param_file_name; + // DEPLOY.PRE_PROCESSOR + std::string _pre_processor; + // DEPLOY.USE_GPU + int _use_gpu; + // DEPLOY.PREDICTOR_MODE + std::string _predictor_mode; + // DEPLOY.BATCH_SIZE + int _batch_size; +}; +} // namespace PaddleSolution diff --git a/inference/utils/utils.h b/inference/utils/utils.h index 63245219e..11301cedc 100644 --- a/inference/utils/utils.h +++ b/inference/utils/utils.h @@ -28,97 +28,103 @@ #endif namespace PaddleSolution { - namespace utils { - enum SCALE_TYPE{ - UNPADDING, - RANGE_SCALING - }; - inline std::string path_join(const std::string& dir, const std::string& path) { - std::string seperator = "/"; - #ifdef _WIN32 - seperator = "\\"; - #endif - return dir + seperator + path; +namespace utils { + +enum SCALE_TYPE{ + UNPADDING, + RANGE_SCALING +}; + +inline std::string path_join(const std::string& dir, const std::string& path) { + std::string seperator = "/"; + #ifdef _WIN32 + seperator = "\\"; + #endif + return dir + seperator + path; +} +#ifndef _WIN32 +// scan a directory and get all files with input extensions +inline std::vector get_directory_images(const std::string& path, + const std::string& exts) { + std::vector imgs; + struct dirent *entry; + DIR *dir = opendir(path.c_str()); + if (dir == NULL) { + closedir(dir); + return imgs; + } + while ((entry = readdir(dir)) != NULL) { + std::string item = entry->d_name; + auto ext = strrchr(entry->d_name, '.'); + if (!ext || std::string(ext) == "." || std::string(ext) == "..") { + continue; } - #ifndef _WIN32 - // scan a directory and get all files with input extensions - inline std::vector get_directory_images(const std::string& path, const std::string& exts) - { - std::vector imgs; - struct dirent *entry; - DIR *dir = opendir(path.c_str()); - if (dir == NULL) { - closedir(dir); - return imgs; - } + if (exts.find(ext) != std::string::npos) { + imgs.push_back(path_join(path, entry->d_name)); + } + } + sort(imgs.begin(), imgs.end()); + return imgs; +} +#else +// scan a directory and get all files with input extensions +inline std::vector get_directory_images(const std::string& path, + const std::string& exts) { + std::vector imgs; + for (const auto& item : + std::experimental::filesystem::directory_iterator(path)) { + auto suffix = item.path().extension().string(); + if (exts.find(suffix) != std::string::npos && suffix.size() > 0) { + auto fullname = path_join(path, item.path().filename().string()); + imgs.push_back(item.path().string()); + } + } + sort(imgs.begin(), imgs.end()); + return imgs; +} +#endif - while ((entry = readdir(dir)) != NULL) { - std::string item = entry->d_name; - auto ext = strrchr(entry->d_name, '.'); - if (!ext || std::string(ext) == "." || std::string(ext) == "..") { - continue; - } - if (exts.find(ext) != std::string::npos) { - imgs.push_back(path_join(path, entry->d_name)); - } - } - sort(imgs.begin(), imgs.end()); - return imgs; +inline int scaling(int resize_type, int &w, int &h, int new_w, int new_h, + int target_size, int max_size, float &im_scale_ratio) { + if (w <= 0 || h <= 0 || new_w <= 0 || new_h <= 0) { + return -1; + } + switch (resize_type) { + case SCALE_TYPE::UNPADDING: + { + w = new_w; + h = new_h; + im_scale_ratio = 0; } - #else - // scan a directory and get all files with input extensions - inline std::vector get_directory_images(const std::string& path, const std::string& exts) + break; + case SCALE_TYPE::RANGE_SCALING: { - std::vector imgs; - for (const auto& item : std::experimental::filesystem::directory_iterator(path)) { - auto suffix = item.path().extension().string(); - if (exts.find(suffix) != std::string::npos && suffix.size() > 0) { - auto fullname = path_join(path, item.path().filename().string()); - imgs.push_back(item.path().string()); + int im_max_size = std::max(w, h); + int im_min_size = std::min(w, h); + float scale_ratio = static_cast(target_size) + / static_cast(im_min_size); + if (max_size > 0) { + if (round(scale_ratio * im_max_size) > max_size) { + scale_ratio = static_cast(max_size) + / static_cast(im_max_size); } } - sort(imgs.begin(), imgs.end()); - return imgs; + w = round(scale_ratio * static_cast(w)); + h = round(scale_ratio * static_cast(h)); + im_scale_ratio = scale_ratio; } - #endif - - inline int scaling(int resize_type, int &w, int &h, int new_w, int new_h, int target_size, int max_size, float &im_scale_ratio) - { - if(w <= 0 || h <= 0 || new_w <= 0 || new_h <= 0){ - return -1; - } - switch(resize_type) { - case SCALE_TYPE::UNPADDING: - { - w = new_w; - h = new_h; - im_scale_ratio=0; - } - break; - case SCALE_TYPE::RANGE_SCALING: - { - int im_max_size = std::max(w, h); - int im_min_size = std::min(w, h); - float scale_ratio= static_cast(target_size) / static_cast(im_min_size); - if(max_size > 0) { - if(round(scale_ratio * im_max_size) > max_size) { - scale_ratio = static_cast(max_size) / static_cast(im_max_size); - } - } - w = round(scale_ratio * static_cast(w)); - h = round(scale_ratio * static_cast(h)); - im_scale_ratio = scale_ratio; - } - break; - default : - { - std::cout << "Can't support this type of scaling strategy." << std::endl; - std::cout << "Throw exception at file " << __FILE__ << " on line " << __LINE__ << std::endl; - throw 0; - } - break; - } - return 0; - } + break; + default : + { + std::cout << "Can't support this type of scaling strategy." + << std::endl; + std::cout << "Throw exception at file " << __FILE__ + << " on line " << __LINE__ << std::endl; + throw 0; + } + break; } + return 0; } +} // namespace utils +} // namespace PaddleSolution -- GitLab