diff --git a/README_CN.md b/README_CN.md old mode 100644 new mode 100755 index 6c2dc5170d..8814215fee --- a/README_CN.md +++ b/README_CN.md @@ -68,8 +68,8 @@
-
- +
+ Python SDK快速开始（点开查看详情）
#### 快速安装 @@ -131,7 +131,7 @@ cv2.imwrite("vis_image.jpg", vis_im)
- + C++ SDK快速开始（点开查看详情）
diff --git a/README_EN.md b/README_EN.md old mode 100644 new mode 100755 diff --git a/cmake/opencv.cmake b/cmake/opencv.cmake index 912cf2917d..87f8c8bcd9 100755 --- a/cmake/opencv.cmake +++ b/cmake/opencv.cmake @@ -41,10 +41,12 @@ elseif(IOS) else() if(CMAKE_HOST_SYSTEM_PROCESSOR MATCHES "aarch64") set(OPENCV_FILENAME "opencv-linux-aarch64-3.4.14") - elseif(TARGET_ABI MATCHES "armhf") - set(OPENCV_FILENAME "opencv-armv7hf") else() - set(OPENCV_FILENAME "opencv-linux-x64-3.4.16") + if(ENABLE_TIMVX) + set(OPENCV_FILENAME "opencv-armv7hf") + else() + set(OPENCV_FILENAME "opencv-linux-x64-3.4.16") + endif() endif() if(ENABLE_OPENCV_CUDA) if(CMAKE_HOST_SYSTEM_PROCESSOR MATCHES "aarch64") @@ -57,7 +59,7 @@ endif() set(OPENCV_INSTALL_DIR ${THIRD_PARTY_PATH}/install/) if(ANDROID) set(OPENCV_URL_PREFIX "https://bj.bcebos.com/fastdeploy/third_libs") -elseif(TARGET_ABI MATCHES "armhf") +elseif(ENABLE_TIMVX) set(OPENCV_URL_PREFIX "https://bj.bcebos.com/fastdeploy/test") else() # TODO: use fastdeploy/third_libs instead. set(OPENCV_URL_PREFIX "https://bj.bcebos.com/paddle2onnx/libs") @@ -185,7 +187,7 @@ else() file(RENAME ${THIRD_PARTY_PATH}/install/${OPENCV_FILENAME}/ ${THIRD_PARTY_PATH}/install/opencv) set(OPENCV_FILENAME opencv) set(OpenCV_DIR ${THIRD_PARTY_PATH}/install/${OPENCV_FILENAME}) - if(TARGET_ABI MATCHES "armhf") + if(ENABLE_TIMVX) set(OpenCV_DIR ${OpenCV_DIR}/lib/cmake/opencv4) endif() if (WIN32) diff --git a/cmake/paddlelite.cmake b/cmake/paddlelite.cmake index d86ccf2d1d..12a069f6e4 100755 --- a/cmake/paddlelite.cmake +++ b/cmake/paddlelite.cmake @@ -59,10 +59,12 @@ elseif(ANDROID) else() # Linux if(CMAKE_HOST_SYSTEM_PROCESSOR MATCHES "aarch64") set(PADDLELITE_URL "${PADDLELITE_URL_PREFIX}/lite-linux-arm64-20220920.tgz") - elseif(TARGET_ABI MATCHES "armhf") - set(PADDLELITE_URL "https://bj.bcebos.com/fastdeploy/test/lite-linux_armhf_1101.tgz") else() - message(FATAL_ERROR "Only support Linux aarch64 now, x64 is not supported with backend Paddle Lite.") + if(ENABLE_TIMVX) + set(PADDLELITE_URL "https://bj.bcebos.com/fastdeploy/test/lite-linux_armhf_1130.tgz") + else() + message(FATAL_ERROR "Only support Linux aarch64 or ENABLE_TIMVX now, x64 is not supported with backend Paddle Lite.") + endif() endif() endif() diff --git a/cmake/timvx.cmake b/cmake/timvx.cmake index 153f0de190..c6a7d54d27 100755 --- a/cmake/timvx.cmake +++ b/cmake/timvx.cmake @@ -1,11 +1,10 @@ -if (NOT DEFINED TARGET_ABI) +if (NOT DEFINED CMAKE_SYSTEM_PROCESSOR) set(CMAKE_SYSTEM_NAME Linux) set(CMAKE_SYSTEM_PROCESSOR arm) set(CMAKE_C_COMPILER "arm-linux-gnueabihf-gcc") set(CMAKE_CXX_COMPILER "arm-linux-gnueabihf-g++") set(CMAKE_CXX_FLAGS "-march=armv7-a -mfloat-abi=hard -mfpu=neon-vfpv4 ${CMAKE_CXX_FLAGS}") set(CMAKE_C_FLAGS "-march=armv7-a -mfloat-abi=hard -mfpu=neon-vfpv4 ${CMAKE_C_FLAGS}" ) - set(TARGET_ABI armhf) set(CMAKE_BUILD_TYPE MinSizeRel) else() if(NOT ${ENABLE_LITE_BACKEND}) diff --git a/docs/cn/build_and_install/README.md b/docs/cn/build_and_install/README.md index a40d571f5a..8be1f745a8 100755 --- a/docs/cn/build_and_install/README.md +++ b/docs/cn/build_and_install/README.md @@ -8,10 +8,10 @@ ## 自行编译安装 - [GPU部署环境](gpu.md) - [CPU部署环境](cpu.md) -- [CPU部署环境](ipu.md) +- [IPU部署环境](ipu.md) - [Jetson部署环境](jetson.md) - [Android平台部署环境](android.md) -- [瑞芯微RK1126部署环境](rk1126.md) +- [瑞芯微RV1126部署环境](rv1126.md) ## FastDeploy编译选项说明 @@ -22,6 +22,7 @@ | ENABLE_PADDLE_BACKEND | 默认OFF，是否编译集成Paddle Inference后端(CPU/GPU上推荐打开) | | ENABLE_LITE_BACKEND | 默认OFF，是否编译集成Paddle Lite后端(编译Android库时需要设置为ON) | | ENABLE_RKNPU2_BACKEND | 默认OFF，是否编译集成RKNPU2后端(RK3588/RK3568/RK3566上推荐打开) | +| ENABLE_TIMVX | 默认OFF，需要在RV1126/RV1109上部署时，需设置为ON | | ENABLE_TRT_BACKEND | 默认OFF，是否编译集成TensorRT后端(GPU上推荐打开) | | ENABLE_OPENVINO_BACKEND | 默认OFF，是否编译集成OpenVINO后端(CPU上推荐打开) | | ENABLE_VISION | 默认OFF，是否编译集成视觉模型的部署模块 | diff --git a/docs/cn/build_and_install/rk1126.md b/docs/cn/build_and_install/rk1126.md deleted file mode 100755 index 4eda019810..0000000000 --- a/docs/cn/build_and_install/rk1126.md +++ /dev/null @@ -1,63 +0,0 @@ -# 瑞芯微 RK1126 部署环境编译安装 - -FastDeploy基于 Paddle-Lite 后端支持在瑞芯微（Rockchip）Soc 上进行部署推理。 -更多详细的信息请参考：[PaddleLite部署示例](https://paddle-lite.readthedocs.io/zh/develop/demo_guides/verisilicon_timvx.html)。 - -本文档介绍如何编译基于 PaddleLite 的 C++ FastDeploy 交叉编译库。 - -相关编译选项说明如下： -|编译选项|默认值|说明|备注| -|:---|:---|:---|:---| -|ENABLE_LITE_BACKEND|OFF|编译RK库时需要设置为ON| - | - -更多编译选项请参考[FastDeploy编译选项说明](./README.md) - -## 交叉编译环境搭建 - -### 宿主机环境需求 -- os：Ubuntu == 16.04 -- cmake： version >= 3.10.0 - -### 环境搭建 -```bash - # 1. Install basic software -apt update -apt-get install -y --no-install-recommends \ - gcc g++ git make wget python unzip - -# 2. Install arm gcc toolchains -apt-get install -y --no-install-recommends \ - g++-arm-linux-gnueabi gcc-arm-linux-gnueabi \ - g++-arm-linux-gnueabihf gcc-arm-linux-gnueabihf \ - gcc-aarch64-linux-gnu g++-aarch64-linux-gnu - -# 3. Install cmake 3.10 or above -wget -c https://mms-res.cdn.bcebos.com/cmake-3.10.3-Linux-x86_64.tar.gz && \ - tar xzf cmake-3.10.3-Linux-x86_64.tar.gz && \ - mv cmake-3.10.3-Linux-x86_64 /opt/cmake-3.10 && \ - ln -s /opt/cmake-3.10/bin/cmake /usr/bin/cmake && \ - ln -s /opt/cmake-3.10/bin/ccmake /usr/bin/ccmake -``` - -## 基于 PaddleLite 的 FastDeploy 交叉编译库编译 -搭建好交叉编译环境之后，编译命令如下： -```bash -# Download the latest source code -git clone https://github.com/PaddlePaddle/FastDeploy.git -cd FastDeploy -mkdir build && cd build - -# CMake configuration with RK toolchain -cmake -DCMAKE_TOOLCHAIN_FILE=./../cmake/timvx.cmake \ - -DENABLE_TIMVX=ON \ - -DCMAKE_INSTALL_PREFIX=fastdeploy-tmivx \ - -DENABLE_VISION=ON \ # 是否编译集成视觉模型的部署模块，可选择开启 - -Wno-dev .. - -# Build FastDeploy RK1126 C++ SDK -make -j8 -make install -``` -编译完成之后，会生成 fastdeploy-tmivx 目录，表示基于 PadddleLite TIM-VX 的 FastDeploy 库编译完成。 - -RK1126 上部署 PaddleClas 分类模型请参考：[PaddleClas RK1126开发板 C++ 部署示例](../../../examples/vision/classification/paddleclas/rk1126/README.md) diff --git a/docs/cn/build_and_install/rv1126.md b/docs/cn/build_and_install/rv1126.md new file mode 100755 index 0000000000..f3cd4ed6a7 --- /dev/null +++ b/docs/cn/build_and_install/rv1126.md @@ -0,0 +1,100 @@ +# 瑞芯微 RV1126 部署环境编译安装 + +FastDeploy基于 Paddle-Lite 后端支持在瑞芯微（Rockchip）Soc 上进行部署推理。 +更多详细的信息请参考：[PaddleLite部署示例](https://www.paddlepaddle.org.cn/lite/develop/demo_guides/verisilicon_timvx.html)。 + +本文档介绍如何编译基于 PaddleLite 的 C++ FastDeploy 交叉编译库。 + +相关编译选项说明如下： +|编译选项|默认值|说明|备注| +|:---|:---|:---|:---| +|ENABLE_LITE_BACKEND|OFF|编译RK库时需要设置为ON| - | +|ENABLE_TIMVX|OFF|编译RK库时需要设置为ON| - | + +更多编译选项请参考[FastDeploy编译选项说明](./README.md) + +## 交叉编译环境搭建 + +### 宿主机环境需求 +- os：Ubuntu == 16.04 +- cmake： version >= 3.10.0 + +### 环境搭建 +```bash + # 1. Install basic software +apt update +apt-get install -y --no-install-recommends \ + gcc g++ git make wget python unzip + +# 2. Install arm gcc toolchains +apt-get install -y --no-install-recommends \ + g++-arm-linux-gnueabi gcc-arm-linux-gnueabi \ + g++-arm-linux-gnueabihf gcc-arm-linux-gnueabihf \ + gcc-aarch64-linux-gnu g++-aarch64-linux-gnu + +# 3. Install cmake 3.10 or above +wget -c https://mms-res.cdn.bcebos.com/cmake-3.10.3-Linux-x86_64.tar.gz && \ + tar xzf cmake-3.10.3-Linux-x86_64.tar.gz && \ + mv cmake-3.10.3-Linux-x86_64 /opt/cmake-3.10 && \ + ln -s /opt/cmake-3.10/bin/cmake /usr/bin/cmake && \ + ln -s /opt/cmake-3.10/bin/ccmake /usr/bin/ccmake +``` + +## 基于 PaddleLite 的 FastDeploy 交叉编译库编译 +搭建好交叉编译环境之后，编译命令如下： +```bash +# Download the latest source code +git clone https://github.com/PaddlePaddle/FastDeploy.git +cd FastDeploy +mkdir build && cd build + +# CMake configuration with RK toolchain +cmake -DCMAKE_TOOLCHAIN_FILE=./../cmake/timvx.cmake \ + -DENABLE_TIMVX=ON \ + -DCMAKE_INSTALL_PREFIX=fastdeploy-tmivx \ + -DENABLE_VISION=ON \ # 是否编译集成视觉模型的部署模块，可选择开启 + -Wno-dev .. + +# Build FastDeploy RV1126 C++ SDK +make -j8 +make install +``` +编译完成之后，会生成 fastdeploy-tmivx 目录，表示基于 PadddleLite TIM-VX 的 FastDeploy 库编译完成。 + +## 准备设备运行环境 +部署前要保证芯原 Linux Kernel NPU 驱动 galcore.so 版本及所适用的芯片型号与依赖库保持一致，在部署前，请登录开发板，并通过命令行输入以下命令查询 NPU 驱动版本，Rockchip建议的驱动版本为： 6.4.6.5 +```bash +dmesg | grep Galcore +``` + +如果当前版本不符合上述，请用户仔细阅读以下内容，以保证底层 NPU 驱动环境正确。 + +有两种方式可以修改当前的 NPU 驱动版本： +1. 手动替换 NPU 驱动版本。（推荐） +2. 刷机，刷取 NPU 驱动版本符合要求的固件。 + +### 手动替换 NPU 驱动版本 +1. 使用如下命令下载解压 PaddleLite demo，其中提供了现成的驱动文件 +```bash +wget https://paddlelite-demo.bj.bcebos.com/devices/generic/PaddleLite-generic-demo.tar.gz +tar -xf PaddleLite-generic-demo.tar.gz +``` +2. 使用 `uname -a` 查看 `Linux Kernel` 版本，确定为 `Linux` 系统 4.19.111 版本， +3. 将 `PaddleLite-generic-demo/libs/PaddleLite/linux/armhf/lib/verisilicon_timvx/viv_sdk_6_4_6_5/lib/1126/4.19.111/` 路径下的 `galcore.ko` 上传至开发板。 + +4. 登录开发板，命令行输入 `sudo rmmod galcore` 来卸载原始驱动，输入 `sudo insmod galcore.ko` 来加载传上设备的驱动。（是否需要 sudo 根据开发板实际情况，部分 adb 链接的设备请提前 adb root）。此步骤如果操作失败，请跳转至方法 2。 +5. 在开发板中输入 `dmesg | grep Galcore` 查询 NPU 驱动版本，确定为：6.4.6.5 + +### 刷机 +根据具体的开发板型号，向开发板卖家或官网客服索要 6.4.6.5 版本 NPU 驱动对应的固件和刷机方法。 + +更多细节请参考：[PaddleLite准备设备环境](https://www.paddlepaddle.org.cn/lite/develop/demo_guides/verisilicon_timvx.html#zhunbeishebeihuanjing) + +## 基于 FastDeploy 在 RV1126 上的部署示例 +1. RV1126 上部署 PaddleClas 分类模型请参考：[PaddleClas 分类模型在 RV1126 上的 C++ 部署示例](../../../examples/vision/classification/paddleclas/rv1126/README.md) + +2. RV1126 上部署 PPYOLOE 检测模型请参考：[PPYOLOE 检测模型在 RV1126 上的 C++ 部署示例](../../../examples/vision/detection/paddledetection/rv1126/README.md) + +3. RV1126 上部署 YOLOv5 检测模型请参考：[YOLOv5 检测模型在 RV1126 上的 C++ 部署示例](../../../examples/vision/detection/yolov5/rv1126/README.md) + +4. RV1126 上部署 PP-LiteSeg 分割模型请参考：[PP-LiteSeg 分割模型在 RV1126 上的 C++ 部署示例](../../../examples/vision/segmentation/paddleseg/rv1126/README.md) diff --git a/examples/vision/classification/paddleclas/quantize/README.md b/examples/vision/classification/paddleclas/quantize/README.md old mode 100644 new mode 100755 index 0a814e0e37..e55f95e261 --- a/examples/vision/classification/paddleclas/quantize/README.md +++ b/examples/vision/classification/paddleclas/quantize/README.md @@ -4,7 +4,7 @@ FastDeploy已支持部署量化模型,并提供一键模型自动化压缩的工 ## FastDeploy一键模型自动化压缩工具 FastDeploy 提供了一键模型自动化压缩工具, 能够简单地通过输入一个配置文件, 对模型进行量化. -详细教程请见: [一键模型自动化压缩工具](../../../../../tools/auto_compression/) +详细教程请见: [一键模型自动化压缩工具](../../../../../tools/common_tools/auto_compression/) 注意: 推理量化后的分类模型仍然需要FP32模型文件夹下的inference_cls.yaml文件, 自行量化的模型文件夹内不包含此yaml文件, 用户从FP32模型文件夹下复制此yaml文件到量化后的模型文件夹内即可。 ## 下载量化完成的PaddleClas模型 diff --git a/examples/vision/classification/paddleclas/quantize/cpp/README.md b/examples/vision/classification/paddleclas/quantize/cpp/README.md old mode 100644 new mode 100755 index 3dced039ca..999cc7fde3 --- a/examples/vision/classification/paddleclas/quantize/cpp/README.md +++ b/examples/vision/classification/paddleclas/quantize/cpp/README.md @@ -8,7 +8,7 @@ ### 量化模型准备 - 1. 用户可以直接使用由FastDeploy提供的量化模型进行部署. -- 2. 用户可以使用FastDeploy提供的[一键模型自动化压缩工具](../../../../../../tools/auto_compression/),自行进行模型量化, 并使用产出的量化模型进行部署.(注意: 推理量化后的分类模型仍然需要FP32模型文件夹下的inference_cls.yaml文件, 自行量化的模型文件夹内不包含此yaml文件, 用户从FP32模型文件夹下复制此yaml文件到量化后的模型文件夹内即可.) +- 2. 用户可以使用FastDeploy提供的[一键模型自动化压缩工具](../../../../../../tools/common_tools/auto_compression/),自行进行模型量化, 并使用产出的量化模型进行部署.(注意: 推理量化后的分类模型仍然需要FP32模型文件夹下的inference_cls.yaml文件, 自行量化的模型文件夹内不包含此yaml文件, 用户从FP32模型文件夹下复制此yaml文件到量化后的模型文件夹内即可.) ## 以量化后的ResNet50_Vd模型为例, 进行部署，支持此模型需保证FastDeploy版本0.7.0以上(x.x.x>=0.7.0) 在本目录执行如下命令即可完成编译,以及量化模型部署. diff --git a/examples/vision/classification/paddleclas/quantize/python/README.md b/examples/vision/classification/paddleclas/quantize/python/README.md old mode 100644 new mode 100755 index 6b874e77ab..a7fc4f9d3a --- a/examples/vision/classification/paddleclas/quantize/python/README.md +++ b/examples/vision/classification/paddleclas/quantize/python/README.md @@ -8,7 +8,7 @@ ### 量化模型准备 - 1. 用户可以直接使用由FastDeploy提供的量化模型进行部署. -- 2. 用户可以使用FastDeploy提供的[一键模型自动化压缩工具](../../../../../../tools/auto_compression/),自行进行模型量化, 并使用产出的量化模型进行部署.(注意: 推理量化后的分类模型仍然需要FP32模型文件夹下的inference_cls.yaml文件, 自行量化的模型文件夹内不包含此yaml文件, 用户从FP32模型文件夹下复制此yaml文件到量化后的模型文件夹内即可.) +- 2. 用户可以使用FastDeploy提供的[一键模型自动化压缩工具](../../../../../../tools/common_tools/auto_compression/),自行进行模型量化, 并使用产出的量化模型进行部署.(注意: 推理量化后的分类模型仍然需要FP32模型文件夹下的inference_cls.yaml文件, 自行量化的模型文件夹内不包含此yaml文件, 用户从FP32模型文件夹下复制此yaml文件到量化后的模型文件夹内即可.) ## 以量化后的ResNet50_Vd模型为例, 进行部署 diff --git a/examples/vision/classification/paddleclas/rk1126/README.md b/examples/vision/classification/paddleclas/rv1126/README.md similarity index 61% rename from examples/vision/classification/paddleclas/rk1126/README.md rename to examples/vision/classification/paddleclas/rv1126/README.md index bac6d7bf8d..065939a278 100755 --- a/examples/vision/classification/paddleclas/rk1126/README.md +++ b/examples/vision/classification/paddleclas/rv1126/README.md @@ -1,11 +1,11 @@ -# PaddleClas 量化模型在 RK1126 上的部署 -目前 FastDeploy 已经支持基于 PaddleLite 部署 PaddleClas 量化模型到 RK1126 上。 +# PaddleClas 量化模型在 RV1126 上的部署 +目前 FastDeploy 已经支持基于 PaddleLite 部署 PaddleClas 量化模型到 RV1126 上。 模型的量化和量化模型的下载请参考：[模型量化](../quantize/README.md) ## 详细部署文档 -在 RK1126 上只支持 C++ 的部署。 +在 RV1126 上只支持 C++ 的部署。 - [C++部署](cpp) diff --git a/examples/vision/classification/paddleclas/rk1126/cpp/CMakeLists.txt b/examples/vision/classification/paddleclas/rv1126/cpp/CMakeLists.txt similarity index 100% rename from examples/vision/classification/paddleclas/rk1126/cpp/CMakeLists.txt rename to examples/vision/classification/paddleclas/rv1126/cpp/CMakeLists.txt diff --git a/examples/vision/classification/paddleclas/rk1126/cpp/README.md b/examples/vision/classification/paddleclas/rv1126/cpp/README.md similarity index 65% rename from examples/vision/classification/paddleclas/rk1126/cpp/README.md rename to examples/vision/classification/paddleclas/rv1126/cpp/README.md index 6d8ecd1512..feaba462fe 100755 --- a/examples/vision/classification/paddleclas/rk1126/cpp/README.md +++ b/examples/vision/classification/paddleclas/rv1126/cpp/README.md @@ -1,22 +1,22 @@ -# PaddleClas RK1126开发板 C++ 部署示例 -本目录下提供的 `infer.cc`，可以帮助用户快速完成 PaddleClas 量化模型在 RK1126 上的部署推理加速。 +# PaddleClas RV1126 开发板 C++ 部署示例 +本目录下提供的 `infer.cc`，可以帮助用户快速完成 PaddleClas 量化模型在 RV1126 上的部署推理加速。 ## 部署准备 ### FastDeploy 交叉编译环境准备 -- 1. 软硬件环境满足要求，以及交叉编译环境的准备，请参考：[FastDeploy 交叉编译环境准备](../../../../../../docs/cn/build_and_install/rk1126.md#交叉编译环境搭建) +- 1. 软硬件环境满足要求，以及交叉编译环境的准备，请参考：[FastDeploy 交叉编译环境准备](../../../../../../docs/cn/build_and_install/rv1126.md#交叉编译环境搭建) ### 量化模型准备 - 1. 用户可以直接使用由 FastDeploy 提供的量化模型进行部署。 -- 2. 用户可以使用 FastDeploy 提供的[一键模型自动化压缩工具](../../../../../../tools/auto_compression/)，自行进行模型量化, 并使用产出的量化模型进行部署。(注意: 推理量化后的分类模型仍然需要FP32模型文件夹下的inference_cls.yaml文件, 自行量化的模型文件夹内不包含此 yaml 文件, 用户从 FP32 模型文件夹下复制此 yaml 文件到量化后的模型文件夹内即可.) +- 2. 用户可以使用 FastDeploy 提供的[一键模型自动化压缩工具](../../../../../../tools/common_tools/auto_compression/)，自行进行模型量化, 并使用产出的量化模型进行部署。(注意: 推理量化后的分类模型仍然需要FP32模型文件夹下的inference_cls.yaml文件, 自行量化的模型文件夹内不包含此 yaml 文件, 用户从 FP32 模型文件夹下复制此 yaml 文件到量化后的模型文件夹内即可.) - 更多量化相关相关信息可查阅[模型量化](../../quantize/README.md) -## 在 RK1126 上部署量化后的 ResNet50_Vd 分类模型 -请按照以下步骤完成在 RK1126 上部署 ResNet50_Vd 量化模型： -1. 交叉编译编译 FastDeploy 库，具体请参考：[交叉编译 FastDeploy](../../../../../../docs/cn/build_and_install/rk1126.md#基于-paddlelite-的-fastdeploy-交叉编译库编译) +## 在 RV1126 上部署量化后的 ResNet50_Vd 分类模型 +请按照以下步骤完成在 RV1126 上部署 ResNet50_Vd 量化模型： +1. 交叉编译编译 FastDeploy 库，具体请参考：[交叉编译 FastDeploy](../../../../../../docs/cn/build_and_install/rv1126.md#基于-paddlelite-的-fastdeploy-交叉编译库编译) 2. 将编译后的库拷贝到当前目录，可使用如下命令： ```bash -cp -r FastDeploy/build/fastdeploy-tmivx/ FastDeploy/examples/vision/classification/paddleclas/rk1126/cpp/ +cp -r FastDeploy/build/fastdeploy-tmivx/ FastDeploy/examples/vision/classification/paddleclas/rv1126/cpp/ ``` 3. 在当前路径下载部署所需的模型和示例图片： @@ -32,7 +32,7 @@ cp -r ILSVRC2012_val_00000010.jpeg images 4. 编译部署示例，可使入如下命令： ```bash mkdir build && cd build -cmake -DCMAKE_TOOLCHAIN_FILE=../fastdeploy-tmivx/timvx.cmake -DFASTDEPLOY_INSTALL_DIR=fastdeploy-tmivx .. +cmake -DCMAKE_TOOLCHAIN_FILE=${PWD}/../fastdeploy-tmivx/timvx.cmake -DFASTDEPLOY_INSTALL_DIR=${PWD}/../fastdeploy-tmivx .. make -j8 make install # 成功编译之后，会生成 install 文件夹，里面有一个运行 demo 和部署所需的库 @@ -41,7 +41,7 @@ make install 5. 基于 adb 工具部署 ResNet50_vd 分类模型到 Rockchip RV1126，可使用如下命令： ```bash # 进入 install 目录 -cd FastDeploy/examples/vision/classification/paddleclas/rk1126/cpp/build/install/ +cd FastDeploy/examples/vision/classification/paddleclas/rv1126/cpp/build/install/ # 如下命令表示：bash run_with_adb.sh 需要运行的demo 模型路径 图片路径 设备的DEVICE_ID bash run_with_adb.sh infer_demo ResNet50_vd_infer ILSVRC2012_val_00000010.jpeg $DEVICE_ID ``` @@ -50,4 +50,4 @@ bash run_with_adb.sh infer_demo ResNet50_vd_infer ILSVRC2012_val_00000010.jpeg $ -需要特别注意的是，在 RK1126 上部署的模型需要是量化后的模型，模型的量化请参考：[模型量化](../../../../../../docs/cn/quantize.md) +需要特别注意的是，在 RV1126 上部署的模型需要是量化后的模型，模型的量化请参考：[模型量化](../../../../../../docs/cn/quantize.md) diff --git a/examples/vision/classification/paddleclas/rk1126/cpp/infer.cc b/examples/vision/classification/paddleclas/rv1126/cpp/infer.cc similarity index 100% rename from examples/vision/classification/paddleclas/rk1126/cpp/infer.cc rename to examples/vision/classification/paddleclas/rv1126/cpp/infer.cc diff --git a/examples/vision/classification/paddleclas/rk1126/cpp/run_with_adb.sh b/examples/vision/classification/paddleclas/rv1126/cpp/run_with_adb.sh similarity index 100% rename from examples/vision/classification/paddleclas/rk1126/cpp/run_with_adb.sh rename to examples/vision/classification/paddleclas/rv1126/cpp/run_with_adb.sh diff --git a/examples/vision/detection/paddledetection/quantize/README.md b/examples/vision/detection/paddledetection/quantize/README.md old mode 100644 new mode 100755 index b041b34684..f0caf81123 --- a/examples/vision/detection/paddledetection/quantize/README.md +++ b/examples/vision/detection/paddledetection/quantize/README.md @@ -4,7 +4,7 @@ FastDeploy已支持部署量化模型,并提供一键模型自动化压缩的工 ## FastDeploy一键模型自动化压缩工具 FastDeploy 提供了一键模型自动化压缩工具, 能够简单地通过输入一个配置文件, 对模型进行量化. -详细教程请见: [一键模型自动化压缩工具](../../../../../tools/auto_compression/) +详细教程请见: [一键模型自动化压缩工具](../../../../../tools/common_tools/auto_compression/) ## 下载量化完成的PP-YOLOE-l模型 用户也可以直接下载下表中的量化模型进行部署.(点击模型名字即可下载) diff --git a/examples/vision/detection/paddledetection/quantize/cpp/README.md b/examples/vision/detection/paddledetection/quantize/cpp/README.md old mode 100644 new mode 100755 index 793fc3938e..4511eb5089 --- a/examples/vision/detection/paddledetection/quantize/cpp/README.md +++ b/examples/vision/detection/paddledetection/quantize/cpp/README.md @@ -9,7 +9,7 @@ ### 量化模型准备 - 1. 用户可以直接使用由FastDeploy提供的量化模型进行部署. -- 2. 用户可以使用FastDeploy提供的[一键模型自动化压缩工具](../../../../../../tools/auto_compression/),自行进行模型量化, 并使用产出的量化模型进行部署.(注意: 推理量化后的分类模型仍然需要FP32模型文件夹下的infer_cfg.yml文件, 自行量化的模型文件夹内不包含此yaml文件, 用户从FP32模型文件夹下复制此yaml文件到量化后的模型文件夹内即可.) +- 2. 用户可以使用FastDeploy提供的[一键模型自动化压缩工具](../../../../../../tools/common_tools/auto_compression/),自行进行模型量化, 并使用产出的量化模型进行部署.(注意: 推理量化后的检测模型仍然需要FP32模型文件夹下的infer_cfg.yml文件, 自行量化的模型文件夹内不包含此yaml文件, 用户从FP32模型文件夹下复制此yaml文件到量化后的模型文件夹内即可.) ## 以量化后的PP-YOLOE-l模型为例, 进行部署。支持此模型需保证FastDeploy版本0.7.0以上(x.x.x>=0.7.0) 在本目录执行如下命令即可完成编译,以及量化模型部署. diff --git a/examples/vision/detection/paddledetection/quantize/python/README.md b/examples/vision/detection/paddledetection/quantize/python/README.md old mode 100644 new mode 100755 index 3efcd232b9..15e0d463e0 --- a/examples/vision/detection/paddledetection/quantize/python/README.md +++ b/examples/vision/detection/paddledetection/quantize/python/README.md @@ -8,7 +8,7 @@ ### 量化模型准备 - 1. 用户可以直接使用由FastDeploy提供的量化模型进行部署. -- 2. 用户可以使用FastDeploy提供的[一键模型自动化压缩工具](../../../../../../tools/auto_compression/),自行进行模型量化, 并使用产出的量化模型进行部署.(注意: 推理量化后的分类模型仍然需要FP32模型文件夹下的infer_cfg.yml文件, 自行量化的模型文件夹内不包含此yaml文件, 用户从FP32模型文件夹下复制此yaml文件到量化后的模型文件夹内即可.) +- 2. 用户可以使用FastDeploy提供的[一键模型自动化压缩工具](../../../../../../tools/common_tools/auto_compression/),自行进行模型量化, 并使用产出的量化模型进行部署.(注意: 推理量化后的分类模型仍然需要FP32模型文件夹下的infer_cfg.yml文件, 自行量化的模型文件夹内不包含此yaml文件, 用户从FP32模型文件夹下复制此yaml文件到量化后的模型文件夹内即可.) ## 以量化后的PP-YOLOE-l模型为例, 进行部署 diff --git a/examples/vision/detection/paddledetection/rk1126/download_models_and_libs.sh b/examples/vision/detection/paddledetection/rk1126/download_models_and_libs.sh deleted file mode 100755 index e7a0fca7a2..0000000000 --- a/examples/vision/detection/paddledetection/rk1126/download_models_and_libs.sh +++ /dev/null @@ -1,25 +0,0 @@ -#!/bin/bash - -set -e - -DETECTION_MODEL_DIR="$(pwd)/picodet_detection/models" -LIBS_DIR="$(pwd)" - -DETECTION_MODEL_URL="https://paddlelite-demo.bj.bcebos.com/Paddle-Lite-Demo/models/picodetv2_relu6_coco_no_fuse.tar.gz" -LIBS_URL="https://paddlelite-demo.bj.bcebos.com/Paddle-Lite-Demo/Paddle-Lite-libs.tar.gz" - -download_and_uncompress() { - local url="$1" - local dir="$2" - - echo "Start downloading ${url}" - curl -L ${url} > ${dir}/download.tar.gz - cd ${dir} - tar -zxvf download.tar.gz - rm -f download.tar.gz -} - -download_and_uncompress "${DETECTION_MODEL_URL}" "${DETECTION_MODEL_DIR}" -download_and_uncompress "${LIBS_URL}" "${LIBS_DIR}" - -echo "Download successful!" diff --git a/examples/vision/detection/paddledetection/rk1126/picodet_detection/CMakeLists.txt b/examples/vision/detection/paddledetection/rk1126/picodet_detection/CMakeLists.txt deleted file mode 100755 index 575db78adf..0000000000 --- a/examples/vision/detection/paddledetection/rk1126/picodet_detection/CMakeLists.txt +++ /dev/null @@ -1,62 +0,0 @@ -cmake_minimum_required(VERSION 3.10) -set(CMAKE_SYSTEM_NAME Linux) -if(TARGET_ARCH_ABI STREQUAL "armv8") - set(CMAKE_SYSTEM_PROCESSOR aarch64) - set(CMAKE_C_COMPILER "aarch64-linux-gnu-gcc") - set(CMAKE_CXX_COMPILER "aarch64-linux-gnu-g++") -elseif(TARGET_ARCH_ABI STREQUAL "armv7hf") - set(CMAKE_SYSTEM_PROCESSOR arm) - set(CMAKE_C_COMPILER "arm-linux-gnueabihf-gcc") - set(CMAKE_CXX_COMPILER "arm-linux-gnueabihf-g++") -else() - message(FATAL_ERROR "Unknown arch abi ${TARGET_ARCH_ABI}, only support armv8 and armv7hf.") - return() -endif() - -project(object_detection_demo) - -message(STATUS "TARGET ARCH ABI: ${TARGET_ARCH_ABI}") -message(STATUS "PADDLE LITE DIR: ${PADDLE_LITE_DIR}") -include_directories(${PADDLE_LITE_DIR}/include) -link_directories(${PADDLE_LITE_DIR}/libs/${TARGET_ARCH_ABI}) -set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11") -if(TARGET_ARCH_ABI STREQUAL "armv8") - set(CMAKE_CXX_FLAGS "-march=armv8-a ${CMAKE_CXX_FLAGS}") - set(CMAKE_C_FLAGS "-march=armv8-a ${CMAKE_C_FLAGS}") -elseif(TARGET_ARCH_ABI STREQUAL "armv7hf") - set(CMAKE_CXX_FLAGS "-march=armv7-a -mfloat-abi=hard -mfpu=neon-vfpv4 ${CMAKE_CXX_FLAGS}") - set(CMAKE_C_FLAGS "-march=armv7-a -mfloat-abi=hard -mfpu=neon-vfpv4 ${CMAKE_C_FLAGS}" ) -endif() - -include_directories(${PADDLE_LITE_DIR}/libs/${TARGET_ARCH_ABI}/third_party/yaml-cpp/include) -link_directories(${PADDLE_LITE_DIR}/libs/${TARGET_ARCH_ABI}/third_party/yaml-cpp) - -find_package(OpenMP REQUIRED) -if(OpenMP_FOUND OR OpenMP_CXX_FOUND) - set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${OpenMP_C_FLAGS}") - set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OpenMP_CXX_FLAGS}") - message(STATUS "Found OpenMP ${OpenMP_VERSION} ${OpenMP_CXX_VERSION}") - message(STATUS "OpenMP C flags: ${OpenMP_C_FLAGS}") - message(STATUS "OpenMP CXX flags: ${OpenMP_CXX_FLAGS}") - message(STATUS "OpenMP OpenMP_CXX_LIB_NAMES: ${OpenMP_CXX_LIB_NAMES}") - message(STATUS "OpenMP OpenMP_CXX_LIBRARIES: ${OpenMP_CXX_LIBRARIES}") -else() - message(FATAL_ERROR "Could not found OpenMP!") - return() -endif() -find_package(OpenCV REQUIRED) -if(OpenCV_FOUND OR OpenCV_CXX_FOUND) - include_directories(${OpenCV_INCLUDE_DIRS}) - message(STATUS "OpenCV library status:") - message(STATUS " version: ${OpenCV_VERSION}") - message(STATUS " libraries: ${OpenCV_LIBS}") - message(STATUS " include path: ${OpenCV_INCLUDE_DIRS}") -else() - message(FATAL_ERROR "Could not found OpenCV!") - return() -endif() - - -add_executable(object_detection_demo object_detection_demo.cc) - -target_link_libraries(object_detection_demo paddle_full_api_shared dl ${OpenCV_LIBS} yaml-cpp) diff --git a/examples/vision/detection/paddledetection/rk1126/picodet_detection/README.md b/examples/vision/detection/paddledetection/rk1126/picodet_detection/README.md deleted file mode 100755 index ddebdeb7ad..0000000000 --- a/examples/vision/detection/paddledetection/rk1126/picodet_detection/README.md +++ /dev/null @@ -1,343 +0,0 @@ -# 目标检测 C++ API Demo 使用指南 - -在 ARMLinux 上实现实时的目标检测功能，此 Demo 有较好的的易用性和扩展性，如在 Demo 中跑自己训练好的模型等。 - - 如果该开发板使用搭载了芯原 NPU （瑞芯微、晶晨、JQL、恩智浦）的 Soc，将有更好的加速效果。 - -## 如何运行目标检测 Demo - -### 环境准备 - -* 准备 ARMLiunx 开发版，将系统刷为 Ubuntu，用于 Demo 编译和运行。请注意，本 Demo 是使用板上编译，而非交叉编译，因此需要图形界面的开发板操作系统。 -* 如果需要使用 芯原 NPU 的计算加速，对 NPU 驱动版本有严格要求，请务必注意事先参考 [芯原 TIM-VX 部署示例](https://paddle-lite.readthedocs.io/zh/develop/demo_guides/verisilicon_timvx.html#id6)，将 NPU 驱动改为要求的版本。 -* Paddle Lite 当前已验证的开发板为 Khadas VIM3（芯片为 Amlogic A311d）、荣品 RV1126、荣品RV1109，其它平台用户可自行尝试； - - Khadas VIM3：由于 VIM3 出厂自带 Android 系统，请先刷成 Ubuntu 系统，在此提供刷机教程：[VIM3/3L Linux 文档](https://docs.khadas.com/linux/zh-cn/vim3)，其中有详细描述刷机方法。以及系统镜像：VIM3 Linux：VIM3_Ubuntu-gnome-focal_Linux-4.9_arm64_EMMC_V1.0.7-210625：[官方链接](http://dl.khadas.com/firmware/VIM3/Ubuntu/EMMC/VIM3_Ubuntu-gnome-focal_Linux-4.9_arm64_EMMC_V1.0.7-210625.img.xz)；[百度云备用链接](https://paddlelite-demo.bj.bcebos.com/devices/verisilicon/firmware/khadas/vim3/VIM3_Ubuntu-gnome-focal_Linux-4.9_arm64_EMMC_V1.0.7-210625.img.xz) - - 荣品 RV1126、1109：由于出场自带 buildroot 系统，如果使用 GUI 界面的 demo，请先刷成 Ubuntu 系统，在此提供刷机教程：[RV1126/1109 教程](https://paddlelite-demo.bj.bcebos.com/Paddle-Lite-Demo/os_img/rockchip/RV1126-RV1109%E4%BD%BF%E7%94%A8%E6%8C%87%E5%AF%BC%E6%96%87%E6%A1%A3-V3.0.pdf)，[刷机工具](https://paddlelite-demo.bj.bcebos.com/Paddle-Lite-Demo/os_img/rockchip/RKDevTool_Release.zip)，以及镜像：[1126镜像](https://paddlelite-demo.bj.bcebos.com/Paddle-Lite-Demo/os_img/update-pro-rv1126-ubuntu20.04-5-720-1280-v2-20220505.img)，[1109镜像](https://paddlelite-demo.bj.bcebos.com/Paddle-Lite-Demo/os_img/update-pro-rv1109-ubuntu20.04-5.5-720-1280-v2-20220429.img)。完整的文档和各种镜像请参考[百度网盘链接](https://pan.baidu.com/s/1Id0LMC0oO2PwR2YcYUAaiQ#list/path=%2F&parentPath=%2Fsharelink2521613171-184070898837664)，密码：2345。 -* 准备 usb camera，注意使用 openCV capture 图像时，请注意 usb camera 的 video序列号作为入参。 -* 请注意，瑞芯微芯片不带有 HDMI 接口，图像显示是依赖 MIPI DSI，所以请准备好 MIPI 显示屏（我们提供的镜像是 720*1280 分辨率，网盘中有更多分辨率选择，注意：请选择 camera-gc2093x2 的镜像）。 -* 配置开发板的网络。如果是办公网络红区，可以将开发板和PC用以太网链接，然后PC共享网络给开发板。 -* gcc g++ opencv cmake 的安装（以下所有命令均在设备上操作） - -```bash -$ sudo apt-get update -$ sudo apt-get install gcc g++ make wget unzip libopencv-dev pkg-config -$ wget https://www.cmake.org/files/v3.10/cmake-3.10.3.tar.gz -$ tar -zxvf cmake-3.10.3.tar.gz -$ cd cmake-3.10.3 -$ ./configure -$ make -$ sudo make install -``` - -### 部署步骤 - -1. 将本 repo 上传至 VIM3 开发板，或者直接开发板上下载或者 git clone 本 repo -2. 目标检测 Demo 位于 `Paddle-Lite-Demo/object_detection/linux/picodet_detection` 目录 -3. 进入 `Paddle-Lite-Demo/object_detection/linux` 目录, 终端中执行 `download_models_and_libs.sh` 脚本自动下载模型和 Paddle Lite 预测库 - -```shell -cd Paddle-Lite-Demo/object_detection/linux # 1. 终端中进入 Paddle-Lite-Demo/object_detection/linux -sh download_models_and_libs.sh # 2. 执行脚本下载依赖项 （需要联网） -``` - -下载完成后会出现提示： `Download successful!` -4. 执行用例(保证 ARMLinux 环境准备完成) - -```shell -cd picodet_detection # 1. 终端中进入 -sh build.sh armv8 # 2. 编译 Demo 可执行程序，默认编译 armv8，如果是 32bit 环境，则改成 sh build.sh armv7hf。 -sh run.sh armv8 # 3. 执行物体检测（picodet 模型） demo，会直接开启摄像头，启动图形界面并呈现检测结果。如果是 32bit 环境，则改成 sh run.sh armv7hf -``` - -### Demo 结果如下:（注意，示例的 picodet 仅使用 coco 数据集，在实际场景中效果一般，请使用实际业务场景重新训练） - - demo_view - -## 更新预测库 - -* Paddle Lite 项目：https://github.com/PaddlePaddle/Paddle-Lite - * 参考 [芯原 TIM-VX 部署示例](https://paddle-lite.readthedocs.io/zh/develop/demo_guides/verisilicon_timvx.html#tim-vx)，编译预测库 - * 编译最终产物位于 `build.lite.xxx.xxx.xxx` 下的 `inference_lite_lib.xxx.xxx` - * 替换 c++ 库 - * 头文件 - 将生成的 `build.lite.linux.armv8.gcc/inference_lite_lib.armlinux.armv8.nnadapter/cxx/include` 文件夹替换 Demo 中的 `Paddle-Lite-Demo/object_detection/linux/Paddle-Lite/include` - * armv8 - 将生成的 `build.lite.linux.armv8.gcc/inference_lite_lib.armlinux.armv8.nnadapter/cxx/libs/libpaddle_full_api_shared.so、libnnadapter.so、libtim-vx.so、libverisilicon_timvx.so` 库替换 Demo 中的 `Paddle-Lite-Demo/object_detection/linux/Paddle-Lite/libs/armv8/` 目录下同名 so - * armv7hf - 将生成的 `build.lite.linux.armv7hf.gcc/inference_lite_lib.armlinux.armv7hf.nnadapter/cxx/libs/libpaddle_full_api_shared.so、libnnadapter.so、libtim-vx.so、libverisilicon_timvx.so` 库替换 Demo 中的 `Paddle-Lite-Demo/object_detection/linux/Paddle-Lite/libs/armv7hf/` 目录下同名 so - -## Demo 内容介绍 - -先整体介绍下目标检测 Demo 的代码结构，然后再简要地介绍 Demo 每部分功能. - -1. `object_detection_demo.cc`： C++ 预测代码 - -```shell -# 位置： -Paddle-Lite-Demo/object_detection/linux/picodet_detection/object_detection_demo.cc -``` - -2. `models` : 模型文件夹 (执行 download_models_and_libs.sh 后会下载 picodet Paddle 模型), label 使用 Paddle-Lite-Demo/object_detection/assets/labels 目录下 coco_label_list.txt - -```shell -# 位置： -Paddle-Lite-Demo/object_detection/linux/picodet_detection/models/picodetv2_relu6_coco_no_fuse -Paddle-Lite-Demo/object_detection/assets/labels/coco_label_list.txt -``` - -3. `Paddle-Lite`：内含 Paddle-Lite 头文件和 动态库，默认带有 timvx 加速库，以及第三方库 yaml-cpp 用于解析 yml 配置文件（执行 download_models_and_libs.sh 后会下载） - -```shell -# 位置 -# 如果要替换动态库 so，则将新的动态库 so 更新到此目录下 -Paddle-Lite-Demo/object_detection/linux/Paddle-Lite/libs/armv8 -Paddle-Lite-Demo/object_detection/linux/Paddle-Lite/include -``` - -4. `CMakeLists.txt` : C++ 预测代码的编译脚本，用于生成可执行文件 - -```shell -# 位置 -Paddle-Lite-Demo/object_detection/linux/picodet_detection/CMakeLists.txt -# 如果有cmake 编译选项更新，可以在 CMakeLists.txt 进行修改即可，默认编译 armv8 可执行文件； -``` - -5. `build.sh` : 编译脚本 - -```shell -# 位置 -Paddle-Lite-Demo/object_detection/linux/picodet_detection/build.sh -``` - -6. `run.sh` : 运行脚本，请注意设置 arm-aarch，armv8 或者 armv7hf。默认为armv8 - -```shell -# 位置 -Paddle-Lite-Demo/object_detection/linux/picodet_detection/run.sh -``` -- 请注意，运行需要5个元素：测试程序、模型、label 文件、异构配置、yaml 文件。 - -## 代码讲解 （使用 Paddle Lite `C++ API` 执行预测） - -ARMLinux 示例基于 C++ API 开发，调用 Paddle Lite `C++s API` 包括以下五步。更详细的 `API` 描述参考：[Paddle Lite C++ API ](https://paddle-lite.readthedocs.io/zh/latest/api_reference/cxx_api_doc.html)。 - -```c++ -#include -// 引入 C++ API -#include "include/paddle_api.h" -#include "include/paddle_use_ops.h" -#include "include/paddle_use_kernels.h" - -// 使用在线编译模型的方式（等价于使用 opt 工具） - -// 1. 设置 CxxConfig -paddle::lite_api::CxxConfig cxx_config; -std::vector valid_places; -valid_places.push_back( - paddle::lite_api::Place{TARGET(kARM), PRECISION(kInt8)}); -valid_places.push_back( - paddle::lite_api::Place{TARGET(kARM), PRECISION(kFloat)}); -// 如果只需要 cpu 计算，那到此结束即可，下面是设置 NPU 的代码段 -valid_places.push_back( - paddle::lite_api::Place{TARGET(kNNAdapter), PRECISION(kInt8)}); -cxx_config.set_valid_places(valid_places); -std::string device = "verisilicon_timvx"; -cxx_config.set_nnadapter_device_names({device}); -// 设置定制化的异构策略 （如需要） -cxx_config.set_nnadapter_subgraph_partition_config_buffer( - nnadapter_subgraph_partition_config_string); - -// 2. 生成 nb 模型 （等价于 opt 工具的产出） -std::shared_ptr predictor = nullptr; -predictor = paddle::lite_api::CreatePaddlePredictor(cxx_config); -predictor->SaveOptimizedModel( - model_path, paddle::lite_api::LiteModelType::kNaiveBuffer); - -// 3. 设置 MobileConfig -MobileConfig config; -config.set_model_from_file(modelPath); // 设置 NaiveBuffer 格式模型路径 -config.set_power_mode(LITE_POWER_NO_BIND); // 设置 CPU 运行模式 -config.set_threads(4); // 设置工作线程数 - -// 4. 创建 PaddlePredictor -predictor = CreatePaddlePredictor(config); - -// 5. 设置输入数据，注意，如果是带后处理的 picodet ，则是有两个输入 -std::unique_ptr input_tensor(std::move(predictor->GetInput(0))); -input_tensor->Resize({1, 3, 416, 416}); -auto* data = input_tensor->mutable_data(); -// scale_factor tensor -auto scale_factor_tensor = predictor->GetInput(1); -scale_factor_tensor->Resize({1, 2}); -auto scale_factor_data = scale_factor_tensor->mutable_data(); -scale_factor_data[0] = 1.0f; -scale_factor_data[1] = 1.0f; - -// 6. 执行预测 -predictor->run(); - -// 7. 获取输出数据 -std::unique_ptr output_tensor(std::move(predictor->GetOutput(0))); - -``` - -## 如何更新模型和输入/输出预处理 - -### 更新模型 -1. 请参考 PaddleDetection 中 [picodet 重训和全量化文档](https://github.com/PaddlePaddle/PaddleDetection/blob/develop/configs/picodet/FULL_QUANTIZATION.md)，基于用户自己数据集重训并且重新全量化 -2. 将模型存放到目录 `object_detection_demo/models/` 下； -3. 模型名字跟工程中模型名字一模一样，即均是使用 `model`、`params`； - -```shell -# shell 脚本 `object_detection_demo/run.sh` -TARGET_ABI=armv8 # for 64bit, such as Amlogic A311D -#TARGET_ABI=armv7hf # for 32bit, such as Rockchip 1109/1126 -if [ -n "$1" ]; then - TARGET_ABI=$1 -fi -export LD_LIBRARY_PATH=../Paddle-Lite/libs/$TARGET_ABI/ -export GLOG_v=0 # Paddle-Lite 日志等级 -export VSI_NN_LOG_LEVEL=0 # TIM-VX 日志等级 -export VIV_VX_ENABLE_GRAPH_TRANSFORM=-pcq:1 # NPU 开启 perchannel 量化模型 -export VIV_VX_SET_PER_CHANNEL_ENTROPY=100 # 同上 -build/object_detection_demo models/picodetv2_relu6_coco_no_fuse ../../assets/labels/coco_label_list.txt models/picodetv2_relu6_coco_no_fuse/subgraph.txt models/picodetv2_relu6_coco_no_fuse/picodet.yml # 执行 Demo 程序，4个 arg 分别为：模型、 label 文件、 自定义异构配置、 yaml -``` - -- 如果需要更新 `label_list` 或者 `yaml` 文件，则修改 `object_detection_demo/run.sh` 中执行命令的第二个和第四个 arg 指定为新的 label 文件和 yaml 配置文件； - -```shell -# 代码文件 `object_detection_demo/rush.sh` -export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:${PADDLE_LITE_DIR}/libs/${TARGET_ARCH_ABI} -build/object_detection_demo {模型} {label} {自定义异构配置文件} {yaml} -``` - -### 更新输入/输出预处理 - -1. 更新输入预处理 -预处理完全根据 yaml 文件来，如果完全按照 PaddleDetection 中 picodet 重训，只需要替换 yaml 文件即可 - -2. 更新输出预处理 -此处需要更新 `object_detection_demo/object_detection_demo.cc` 中的 `postprocess` 方法 - -```c++ -std::vector postprocess(const float *output_data, int64_t output_size, - const std::vector &word_labels, - const float score_threshold, - cv::Mat &output_image, double time) { - std::vector results; - std::vector colors = { - cv::Scalar(237, 189, 101), cv::Scalar(0, 0, 255), - cv::Scalar(102, 153, 153), cv::Scalar(255, 0, 0), - cv::Scalar(9, 255, 0), cv::Scalar(0, 0, 0), - cv::Scalar(51, 153, 51)}; - for (int64_t i = 0; i < output_size; i += 6) { - if (output_data[i + 1] < score_threshold) { - continue; - } - int class_id = static_cast(output_data[i]); - float score = output_data[i + 1]; - RESULT result; - std::string class_name = "Unknown"; - if (word_labels.size() > 0 && class_id >= 0 && - class_id < word_labels.size()) { - class_name = word_labels[class_id]; - } - result.class_name = class_name; - result.score = score; - result.left = output_data[i + 2] / 416; // 此处416根据输入的 HW 得来 - result.top = output_data[i + 3] / 416; - result.right = output_data[i + 4] / 416; - result.bottom = output_data[i + 5] / 416; - int lx = static_cast(result.left * output_image.cols); - int ly = static_cast(result.top * output_image.rows); - int w = static_cast(result.right * output_image.cols) - lx; - int h = static_cast(result.bottom * output_image.rows) - ly; - cv::Rect bounding_box = - cv::Rect(lx, ly, w, h) & - cv::Rect(0, 0, output_image.cols, output_image.rows); - if (w > 0 && h > 0 && score <= 1) { - cv::Scalar color = colors[results.size() % colors.size()]; - cv::rectangle(output_image, bounding_box, color); - cv::rectangle(output_image, cv::Point2d(lx, ly), - cv::Point2d(lx + w, ly - 10), color, -1); - cv::putText(output_image, std::to_string(results.size()) + "." + - class_name + ":" + std::to_string(score), - cv::Point2d(lx, ly), cv::FONT_HERSHEY_PLAIN, 1, - cv::Scalar(255, 255, 255)); - results.push_back(result); - } - } - return results; -} -``` - -## 更新模型后，自定义 NPU-CPU 异构配置（如需使用 NPU 加速） -由于使用芯原 NPU 在 8bit 量化的情况下有最优的性能，因此部署时，我们往往会考虑量化 - - 由于量化可能会引入一定程度的精度问题，所以我们可以通过自定义的异构定制，来将部分有精度问题的 layer 异构至cpu，从而达到最优的精度 - -### 第一步，确定模型量化后在 arm cpu 上的精度 -如果在 arm cpu 上，精度都无法满足，那量化本身就是失败的，此时可以考虑修改训练集或者预处理。 - - 修改 Demo 程序，仅用 arm cpu 计算 -```c++ -paddle::lite_api::CxxConfig cxx_config; -std::vector valid_places; -valid_places.push_back( - paddle::lite_api::Place{TARGET(kARM), PRECISION(kInt8)}); -valid_places.push_back( - paddle::lite_api::Place{TARGET(kARM), PRECISION(kFloat)}); -// 仅用 arm cpu 计算， 注释如下代码即可 -/* -valid_places.push_back( - paddle::lite_api::Place{TARGET(kNNAdapter), PRECISION(kInt8)}); -valid_places.push_back( - paddle::lite_api::Place{TARGET(kNNAdapter), PRECISION(kFloat)}); -*/ -``` -如果 arm cpu 计算结果精度达标，则继续 - -### 第二步，获取整网拓扑信息 - - 回退第一步的修改，不再注释，使用 NPU 加速 - - 运行 Demo，如果此时精度良好，则无需参考后面步骤，模型部署和替换非常顺利，enjoy it。 - - 如果精度不行，请参考后续步骤。 - -### 第三步，获取整网拓扑信息 - - 回退第一步的修改，使用 - - 修改 run.sh ，将其中 export GLOG_v=0 改为 export GLOG_v=5 - - 运行 Demo，等摄像头启动，即可 ctrl+c 关闭 Demo - - 收集日志，搜索关键字 "subgraph operators" 随后那一段，便是整个模型的拓扑信息，其格式如下： - - 每行记录由『算子类型:输入张量名列表:输出张量名列表』组成（即以分号分隔算子类型、输入和输出张量名列表），以逗号分隔输入、输出张量名列表中的每个张量名； - - 示例说明： - ``` - op_type0:var_name0,var_name1:var_name2 表示将算子类型为 op_type0、输入张量为var_name0 和 var_name1、输出张量为 var_name2 的节点强制运行在 ARM CPU 上 - ``` - -### 第四步，修改异构配置文件 - - 首先看到示例 Demo 中 Paddle-Lite-Demo/object_detection/linux/picodet_detection/models/picodetv2_relu6_coco_no_fuse 目录下的 subgraph.txt 文件。(feed 和 fetch 分别代表整个模型的输入和输入) - ``` - feed:feed:scale_factor - feed:feed:image - - sqrt:tmp_3:sqrt_0.tmp_0 - reshape2:sqrt_0.tmp_0:reshape2_0.tmp_0,reshape2_0.tmp_1 - - matmul_v2:softmax_0.tmp_0,auto_113_:linear_0.tmp_0 - reshape2:linear_0.tmp_0:reshape2_2.tmp_0,reshape2_2.tmp_1 - - sqrt:tmp_6:sqrt_1.tmp_0 - reshape2:sqrt_1.tmp_0:reshape2_3.tmp_0,reshape2_3.tmp_1 - - matmul_v2:softmax_1.tmp_0,auto_113_:linear_1.tmp_0 - reshape2:linear_1.tmp_0:reshape2_5.tmp_0,reshape2_5.tmp_1 - - sqrt:tmp_9:sqrt_2.tmp_0 - reshape2:sqrt_2.tmp_0:reshape2_6.tmp_0,reshape2_6.tmp_1 - - matmul_v2:softmax_2.tmp_0,auto_113_:linear_2.tmp_0 - ... - ``` - - 在 txt 中的都是需要异构至 cpu 计算的 layer，在示例 Demo 中，我们把 picodet 后处理的部分异构至 arm cpu 做计算，不必担心，Paddle-Lite 的 arm kernel 性能也是非常卓越。 - - 如果新训练的模型没有额外修改 layer，则直接复制使用示例 Demo 中的 subgraph.txt 即可 - - 此时 ./run.sh 看看精度是否符合预期，如果精度符合预期，恭喜，可以跳过本章节，enjoy it。 - - 如果精度不符合预期，则将上文『第二步，获取整网拓扑信息』中获取的拓扑信息，从 "feed" 之后第一行，直到 "sqrt" 之前，都复制进 sugraph.txt。这一步代表了将大量的 backbone 部分算子放到 arm cpu 计算。 - - 此时 ./run.sh 看看精度是否符合预期，如果精度达标，那说明在 backbone 中确实存在引入 NPU 精度异常的层（再次重申，在 subgraph.txt 的代表强制在 arm cpu 计算）。 - - 逐行删除、成片删除、二分法，发挥开发人员的耐心，找到引入 NPU 精度异常的 layer，将其留在 subgraph.txt 中，按照经验，如果有 NPU 精度问题，可能会有 1~5 层conv layer 需要异构。 - - 剩余没有精度问题的 layer 在 subgraph.txt 中删除即可 diff --git a/examples/vision/detection/paddledetection/rk1126/picodet_detection/build.sh b/examples/vision/detection/paddledetection/rk1126/picodet_detection/build.sh deleted file mode 100755 index 1219c481b9..0000000000 --- a/examples/vision/detection/paddledetection/rk1126/picodet_detection/build.sh +++ /dev/null @@ -1,17 +0,0 @@ -#!/bin/bash -USE_FULL_API=TRUE -# configure -TARGET_ARCH_ABI=armv8 # for RK3399, set to default arch abi -#TARGET_ARCH_ABI=armv7hf # for Raspberry Pi 3B -PADDLE_LITE_DIR=../Paddle-Lite -THIRD_PARTY_DIR=./third_party -if [ "x$1" != "x" ]; then - TARGET_ARCH_ABI=$1 -fi - -# build -rm -rf build -mkdir build -cd build -cmake -DPADDLE_LITE_DIR=${PADDLE_LITE_DIR} -DTARGET_ARCH_ABI=${TARGET_ARCH_ABI} -DTHIRD_PARTY_DIR=${THIRD_PARTY_DIR} .. -make diff --git a/examples/vision/detection/paddledetection/rk1126/picodet_detection/object_detection_demo.cc b/examples/vision/detection/paddledetection/rk1126/picodet_detection/object_detection_demo.cc deleted file mode 100644 index 8c132dea27..0000000000 --- a/examples/vision/detection/paddledetection/rk1126/picodet_detection/object_detection_demo.cc +++ /dev/null @@ -1,411 +0,0 @@ -// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// 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 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -#include "paddle_api.h" -#include "yaml-cpp/yaml.h" -#include -#include -#include -#include -#include -#include -#include -#include -#include -#include - -int WARMUP_COUNT = 0; -int REPEAT_COUNT = 1; -const int CPU_THREAD_NUM = 2; -const paddle::lite_api::PowerMode CPU_POWER_MODE = - paddle::lite_api::PowerMode::LITE_POWER_HIGH; -const std::vector INPUT_SHAPE = {1, 3, 416, 416}; -std::vector INPUT_MEAN = {0.f, 0.f, 0.f}; -std::vector INPUT_STD = {1.f, 1.f, 1.f}; -float INPUT_SCALE = 1 / 255.f; -const float SCORE_THRESHOLD = 0.35f; - -struct RESULT { - std::string class_name; - float score; - float left; - float top; - float right; - float bottom; -}; - -inline int64_t get_current_us() { - struct timeval time; - gettimeofday(&time, NULL); - return 1000000LL * (int64_t)time.tv_sec + (int64_t)time.tv_usec; -} - -bool read_file(const std::string &filename, std::vector *contents, - bool binary = true) { - FILE *fp = fopen(filename.c_str(), binary ? "rb" : "r"); - if (!fp) - return false; - fseek(fp, 0, SEEK_END); - size_t size = ftell(fp); - fseek(fp, 0, SEEK_SET); - contents->clear(); - contents->resize(size); - size_t offset = 0; - char *ptr = reinterpret_cast(&(contents->at(0))); - while (offset < size) { - size_t already_read = fread(ptr, 1, size - offset, fp); - offset += already_read; - ptr += already_read; - } - fclose(fp); - return true; -} - -std::vector load_labels(const std::string &path) { - std::ifstream file; - std::vector labels; - file.open(path); - while (file) { - std::string line; - std::getline(file, line); - labels.push_back(line); - } - file.clear(); - file.close(); - return labels; -} - -bool load_yaml_config(std::string yaml_path) { - YAML::Node cfg; - try { - std::cout << "before loadFile" << std::endl; - cfg = YAML::LoadFile(yaml_path); - } catch (YAML::BadFile &e) { - std::cout << "Failed to load yaml file " << yaml_path - << ", maybe you should check this file." << std::endl; - return false; - } - auto preprocess_cfg = cfg["TestReader"]["sample_transforms"]; - for (const auto &op : preprocess_cfg) { - if (!op.IsMap()) { - std::cout << "Require the transform information in yaml be Map type." - << std::endl; - std::abort(); - } - auto op_name = op.begin()->first.as(); - if (op_name == "NormalizeImage") { - INPUT_MEAN = op.begin()->second["mean"].as>(); - INPUT_STD = op.begin()->second["std"].as>(); - INPUT_SCALE = op.begin()->second["scale"].as(); - } - } - return true; -} - -void preprocess(cv::Mat &input_image, std::vector &input_mean, - std::vector &input_std, float input_scale, - int input_width, int input_height, float *input_data) { - cv::Mat resize_image; - cv::resize(input_image, resize_image, cv::Size(input_width, input_height), 0, - 0); - if (resize_image.channels() == 4) { - cv::cvtColor(resize_image, resize_image, cv::COLOR_BGRA2RGB); - } - cv::Mat norm_image; - resize_image.convertTo(norm_image, CV_32FC3, input_scale); - // NHWC->NCHW - int image_size = input_height * input_width; - const float *image_data = reinterpret_cast(norm_image.data); - float32x4_t vmean0 = vdupq_n_f32(input_mean[0]); - float32x4_t vmean1 = vdupq_n_f32(input_mean[1]); - float32x4_t vmean2 = vdupq_n_f32(input_mean[2]); - float32x4_t vscale0 = vdupq_n_f32(1.0f / input_std[0]); - float32x4_t vscale1 = vdupq_n_f32(1.0f / input_std[1]); - float32x4_t vscale2 = vdupq_n_f32(1.0f / input_std[2]); - float *input_data_c0 = input_data; - float *input_data_c1 = input_data + image_size; - float *input_data_c2 = input_data + image_size * 2; - int i = 0; - for (; i < image_size - 3; i += 4) { - float32x4x3_t vin3 = vld3q_f32(image_data); - float32x4_t vsub0 = vsubq_f32(vin3.val[0], vmean0); - float32x4_t vsub1 = vsubq_f32(vin3.val[1], vmean1); - float32x4_t vsub2 = vsubq_f32(vin3.val[2], vmean2); - float32x4_t vs0 = vmulq_f32(vsub0, vscale0); - float32x4_t vs1 = vmulq_f32(vsub1, vscale1); - float32x4_t vs2 = vmulq_f32(vsub2, vscale2); - vst1q_f32(input_data_c0, vs0); - vst1q_f32(input_data_c1, vs1); - vst1q_f32(input_data_c2, vs2); - image_data += 12; - input_data_c0 += 4; - input_data_c1 += 4; - input_data_c2 += 4; - } - for (; i < image_size; i++) { - *(input_data_c0++) = (*(image_data++) - input_mean[0]) / input_std[0]; - *(input_data_c1++) = (*(image_data++) - input_mean[1]) / input_std[1]; - *(input_data_c2++) = (*(image_data++) - input_mean[2]) / input_std[2]; - } -} - -std::vector postprocess(const float *output_data, int64_t output_size, - const std::vector &word_labels, - const float score_threshold, - cv::Mat &output_image, double time) { - std::vector results; - std::vector colors = { - cv::Scalar(237, 189, 101), cv::Scalar(0, 0, 255), - cv::Scalar(102, 153, 153), cv::Scalar(255, 0, 0), - cv::Scalar(9, 255, 0), cv::Scalar(0, 0, 0), - cv::Scalar(51, 153, 51)}; - for (int64_t i = 0; i < output_size; i += 6) { - if (output_data[i + 1] < score_threshold) { - continue; - } - int class_id = static_cast(output_data[i]); - float score = output_data[i + 1]; - RESULT result; - std::string class_name = "Unknown"; - if (word_labels.size() > 0 && class_id >= 0 && - class_id < word_labels.size()) { - class_name = word_labels[class_id]; - } - result.class_name = class_name; - result.score = score; - result.left = output_data[i + 2] / 416; - result.top = output_data[i + 3] / 416; - result.right = output_data[i + 4] / 416; - result.bottom = output_data[i + 5] / 416; - int lx = static_cast(result.left * output_image.cols); - int ly = static_cast(result.top * output_image.rows); - int w = static_cast(result.right * output_image.cols) - lx; - int h = static_cast(result.bottom * output_image.rows) - ly; - cv::Rect bounding_box = - cv::Rect(lx, ly, w, h) & - cv::Rect(0, 0, output_image.cols, output_image.rows); - if (w > 0 && h > 0 && score <= 1) { - cv::Scalar color = colors[results.size() % colors.size()]; - cv::rectangle(output_image, bounding_box, color); - cv::rectangle(output_image, cv::Point2d(lx, ly), - cv::Point2d(lx + w, ly - 10), color, -1); - cv::putText(output_image, std::to_string(results.size()) + "." + - class_name + ":" + std::to_string(score), - cv::Point2d(lx, ly), cv::FONT_HERSHEY_PLAIN, 1, - cv::Scalar(255, 255, 255)); - results.push_back(result); - } - } - return results; -} - -cv::Mat process(cv::Mat &input_image, std::vector &word_labels, - std::shared_ptr &predictor) { - // Preprocess image and fill the data of input tensor - std::unique_ptr input_tensor( - std::move(predictor->GetInput(0))); - input_tensor->Resize(INPUT_SHAPE); - int input_width = INPUT_SHAPE[3]; - int input_height = INPUT_SHAPE[2]; - auto *input_data = input_tensor->mutable_data(); -#if 1 - // scale_factor tensor - auto scale_factor_tensor = predictor->GetInput(1); - scale_factor_tensor->Resize({1, 2}); - auto scale_factor_data = scale_factor_tensor->mutable_data(); - scale_factor_data[0] = 1.0f; - scale_factor_data[1] = 1.0f; -#endif - - double preprocess_start_time = get_current_us(); - preprocess(input_image, INPUT_MEAN, INPUT_STD, INPUT_SCALE, input_width, - input_height, input_data); - double preprocess_end_time = get_current_us(); - double preprocess_time = - (preprocess_end_time - preprocess_start_time) / 1000.0f; - - double prediction_time; - // Run predictor - // warm up to skip the first inference and get more stable time, remove it in - // actual products - for (int i = 0; i < WARMUP_COUNT; i++) { - predictor->Run(); - } - // repeat to obtain the average time, set REPEAT_COUNT=1 in actual products - double max_time_cost = 0.0f; - double min_time_cost = std::numeric_limits::max(); - double total_time_cost = 0.0f; - for (int i = 0; i < REPEAT_COUNT; i++) { - auto start = get_current_us(); - predictor->Run(); - auto end = get_current_us(); - double cur_time_cost = (end - start) / 1000.0f; - if (cur_time_cost > max_time_cost) { - max_time_cost = cur_time_cost; - } - if (cur_time_cost < min_time_cost) { - min_time_cost = cur_time_cost; - } - total_time_cost += cur_time_cost; - prediction_time = total_time_cost / REPEAT_COUNT; - printf("iter %d cost: %f ms\n", i, cur_time_cost); - } - printf("warmup: %d repeat: %d, average: %f ms, max: %f ms, min: %f ms\n", - WARMUP_COUNT, REPEAT_COUNT, prediction_time, max_time_cost, - min_time_cost); - - // Get the data of output tensor and postprocess to output detected objects - std::unique_ptr output_tensor( - std::move(predictor->GetOutput(0))); - const float *output_data = output_tensor->mutable_data(); - int64_t output_size = 1; - for (auto dim : output_tensor->shape()) { - output_size *= dim; - } - cv::Mat output_image = input_image.clone(); - double postprocess_start_time = get_current_us(); - std::vector results = - postprocess(output_data, output_size, word_labels, SCORE_THRESHOLD, - output_image, prediction_time); - double postprocess_end_time = get_current_us(); - double postprocess_time = - (postprocess_end_time - postprocess_start_time) / 1000.0f; - - printf("results: %d\n", results.size()); - for (int i = 0; i < results.size(); i++) { - printf("[%d] %s - %f %f,%f,%f,%f\n", i, results[i].class_name.c_str(), - results[i].score, results[i].left, results[i].top, results[i].right, - results[i].bottom); - } - printf("Preprocess time: %f ms\n", preprocess_time); - printf("Prediction time: %f ms\n", prediction_time); - printf("Postprocess time: %f ms\n\n", postprocess_time); - - return output_image; -} - -int main(int argc, char **argv) { - if (argc < 5 || argc == 6) { - printf("Usage: \n" - "./object_detection_demo model_dir label_path [input_image_path] " - "[output_image_path]" - "use images from camera if input_image_path and input_image_path " - "isn't provided."); - return -1; - } - - std::string model_path = argv[1]; - std::string label_path = argv[2]; - std::vector word_labels = load_labels(label_path); - std::string nnadapter_subgraph_partition_config_path = argv[3]; - - std::string yaml_path = argv[4]; - if (yaml_path != "null") { - load_yaml_config(yaml_path); - } - - // Run inference by using full api with CxxConfig - paddle::lite_api::CxxConfig cxx_config; - if (1) { // combined model - cxx_config.set_model_file(model_path + "/model"); - cxx_config.set_param_file(model_path + "/params"); - } else { - cxx_config.set_model_dir(model_path); - } - cxx_config.set_threads(CPU_THREAD_NUM); - cxx_config.set_power_mode(CPU_POWER_MODE); - - std::shared_ptr predictor = nullptr; - std::vector valid_places; - valid_places.push_back( - paddle::lite_api::Place{TARGET(kARM), PRECISION(kInt8)}); - valid_places.push_back( - paddle::lite_api::Place{TARGET(kARM), PRECISION(kFloat)}); - valid_places.push_back( - paddle::lite_api::Place{TARGET(kNNAdapter), PRECISION(kInt8)}); - valid_places.push_back( - paddle::lite_api::Place{TARGET(kNNAdapter), PRECISION(kFloat)}); - cxx_config.set_valid_places(valid_places); - std::string device = "verisilicon_timvx"; - cxx_config.set_nnadapter_device_names({device}); - // cxx_config.set_nnadapter_context_properties(nnadapter_context_properties); - - // cxx_config.set_nnadapter_model_cache_dir(nnadapter_model_cache_dir); - // Set the subgraph custom partition configuration file - - if (!nnadapter_subgraph_partition_config_path.empty()) { - std::vector nnadapter_subgraph_partition_config_buffer; - if (read_file(nnadapter_subgraph_partition_config_path, - &nnadapter_subgraph_partition_config_buffer, false)) { - if (!nnadapter_subgraph_partition_config_buffer.empty()) { - std::string nnadapter_subgraph_partition_config_string( - nnadapter_subgraph_partition_config_buffer.data(), - nnadapter_subgraph_partition_config_buffer.size()); - cxx_config.set_nnadapter_subgraph_partition_config_buffer( - nnadapter_subgraph_partition_config_string); - } - } else { - printf("Failed to load the subgraph custom partition configuration file " - "%s\n", - nnadapter_subgraph_partition_config_path.c_str()); - } - } - - try { - predictor = paddle::lite_api::CreatePaddlePredictor(cxx_config); - predictor->SaveOptimizedModel( - model_path, paddle::lite_api::LiteModelType::kNaiveBuffer); - } catch (std::exception e) { - printf("An internal error occurred in PaddleLite(cxx config).\n"); - } - - paddle::lite_api::MobileConfig config; - config.set_model_from_file(model_path + ".nb"); - config.set_threads(CPU_THREAD_NUM); - config.set_power_mode(CPU_POWER_MODE); - config.set_nnadapter_device_names({device}); - predictor = - paddle::lite_api::CreatePaddlePredictor( - config); - if (argc > 5) { - WARMUP_COUNT = 1; - REPEAT_COUNT = 5; - std::string input_image_path = argv[5]; - std::string output_image_path = argv[6]; - cv::Mat input_image = cv::imread(input_image_path); - cv::Mat output_image = process(input_image, word_labels, predictor); - cv::imwrite(output_image_path, output_image); - cv::imshow("Object Detection Demo", output_image); - cv::waitKey(0); - } else { - cv::VideoCapture cap(1); - cap.set(cv::CAP_PROP_FRAME_WIDTH, 640); - cap.set(cv::CAP_PROP_FRAME_HEIGHT, 480); - if (!cap.isOpened()) { - return -1; - } - while (1) { - cv::Mat input_image; - cap >> input_image; - cv::Mat output_image = process(input_image, word_labels, predictor); - cv::imshow("Object Detection Demo", output_image); - if (cv::waitKey(1) == char('q')) { - break; - } - } - cap.release(); - cv::destroyAllWindows(); - } - return 0; -} diff --git a/examples/vision/detection/paddledetection/rk1126/picodet_detection/run.sh b/examples/vision/detection/paddledetection/rk1126/picodet_detection/run.sh deleted file mode 100755 index a6025104a1..0000000000 --- a/examples/vision/detection/paddledetection/rk1126/picodet_detection/run.sh +++ /dev/null @@ -1,15 +0,0 @@ -#!/bin/bash -#run - -TARGET_ABI=armv8 # for 64bit -#TARGET_ABI=armv7hf # for 32bit -if [ -n "$1" ]; then - TARGET_ABI=$1 -fi -export LD_LIBRARY_PATH=../Paddle-Lite/libs/$TARGET_ABI/ -export GLOG_v=0 -export VSI_NN_LOG_LEVEL=0 -export VIV_VX_ENABLE_GRAPH_TRANSFORM=-pcq:1 -export VIV_VX_SET_PER_CHANNEL_ENTROPY=100 -export TIMVX_BATCHNORM_FUSION_MAX_ALLOWED_QUANT_SCALE_DEVIATION=30000 -build/object_detection_demo models/picodetv2_relu6_coco_no_fuse ../../assets/labels/coco_label_list.txt models/picodetv2_relu6_coco_no_fuse/subgraph.txt models/picodetv2_relu6_coco_no_fuse/picodet.yml diff --git a/examples/vision/detection/paddledetection/rv1126/README.md b/examples/vision/detection/paddledetection/rv1126/README.md new file mode 100755 index 0000000000..ff1d58bab6 --- /dev/null +++ b/examples/vision/detection/paddledetection/rv1126/README.md @@ -0,0 +1,11 @@ +# PP-YOLOE 量化模型在 RV1126 上的部署 +目前 FastDeploy 已经支持基于 PaddleLite 部署 PP-YOLOE 量化模型到 RV1126 上。 + +模型的量化和量化模型的下载请参考：[模型量化](../quantize/README.md) + + +## 详细部署文档 + +在 RV1126 上只支持 C++ 的部署。 + +- [C++部署](cpp) diff --git a/examples/vision/detection/paddledetection/rv1126/cpp/CMakeLists.txt b/examples/vision/detection/paddledetection/rv1126/cpp/CMakeLists.txt new file mode 100755 index 0000000000..7a145177ee --- /dev/null +++ b/examples/vision/detection/paddledetection/rv1126/cpp/CMakeLists.txt @@ -0,0 +1,38 @@ +PROJECT(infer_demo C CXX) +CMAKE_MINIMUM_REQUIRED (VERSION 3.10) + +# 指定下载解压后的fastdeploy库路径 +option(FASTDEPLOY_INSTALL_DIR "Path of downloaded fastdeploy sdk.") + +include(${FASTDEPLOY_INSTALL_DIR}/FastDeploy.cmake) + +# 添加FastDeploy依赖头文件 +include_directories(${FASTDEPLOY_INCS}) +include_directories(${FastDeploy_INCLUDE_DIRS}) + +add_executable(infer_demo ${PROJECT_SOURCE_DIR}/infer_ppyoloe.cc) +# 添加FastDeploy库依赖 +target_link_libraries(infer_demo ${FASTDEPLOY_LIBS}) + +set(CMAKE_INSTALL_PREFIX ${CMAKE_SOURCE_DIR}/build/install) + +install(TARGETS infer_demo DESTINATION ./) + +install(DIRECTORY models DESTINATION ./) +install(DIRECTORY images DESTINATION ./) +# install(DIRECTORY run_with_adb.sh DESTINATION ./) + +file(GLOB FASTDEPLOY_LIBS ${FASTDEPLOY_INSTALL_DIR}/lib/*) +install(PROGRAMS ${FASTDEPLOY_LIBS} DESTINATION lib) + +file(GLOB OPENCV_LIBS ${FASTDEPLOY_INSTALL_DIR}/third_libs/install/opencv/lib/lib*) +install(PROGRAMS ${OPENCV_LIBS} DESTINATION lib) + +file(GLOB PADDLELITE_LIBS ${FASTDEPLOY_INSTALL_DIR}/third_libs/install/paddlelite/lib/lib*) +install(PROGRAMS ${PADDLELITE_LIBS} DESTINATION lib) + +file(GLOB TIMVX_LIBS ${FASTDEPLOY_INSTALL_DIR}/third_libs/install/paddlelite/lib/verisilicon_timvx/*) +install(PROGRAMS ${TIMVX_LIBS} DESTINATION lib) + +file(GLOB ADB_TOOLS run_with_adb.sh) +install(PROGRAMS ${ADB_TOOLS} DESTINATION ./) diff --git a/examples/vision/detection/paddledetection/rv1126/cpp/README.md b/examples/vision/detection/paddledetection/rv1126/cpp/README.md new file mode 100755 index 0000000000..5b366bd83c --- /dev/null +++ b/examples/vision/detection/paddledetection/rv1126/cpp/README.md @@ -0,0 +1,55 @@ +# PP-YOLOE 量化模型 C++ 部署示例 + +本目录下提供的 `infer.cc`，可以帮助用户快速完成 PP-YOLOE 量化模型在 RV1126 上的部署推理加速。 + +## 部署准备 +### FastDeploy 交叉编译环境准备 +- 1. 软硬件环境满足要求，以及交叉编译环境的准备，请参考：[FastDeploy 交叉编译环境准备](../../../../../../docs/cn/build_and_install/rv1126.md#交叉编译环境搭建) + +### 模型准备 +- 1. 用户可以直接使用由 FastDeploy 提供的量化模型进行部署。 +- 2. 用户可以先使用 PaddleDetection 自行导出 Float32 模型，注意导出模型模型时设置参数：use_shared_conv=False，更多细节请参考：[PP-YOLOE](https://github.com/PaddlePaddle/PaddleDetection/tree/release/2.4/configs/ppyoloe) +- 3. 用户可以使用 FastDeploy 提供的[一键模型自动化压缩工具](../../../../../../tools/common_tools/auto_compression/)，自行进行模型量化, 并使用产出的量化模型进行部署。（注意: 推理量化后的检测模型仍然需要FP32模型文件夹下的 infer_cfg.yml 文件，自行量化的模型文件夹内不包含此 yaml 文件，用户从 FP32 模型文件夹下复制此yaml文件到量化后的模型文件夹内即可。） +- 更多量化相关相关信息可查阅[模型量化](../../quantize/README.md) + +## 在 RV1126 上部署量化后的 PP-YOLOE 检测模型 +请按照以下步骤完成在 RV1126 上部署 PP-YOLOE 量化模型： +1. 交叉编译编译 FastDeploy 库，具体请参考：[交叉编译 FastDeploy](../../../../../../docs/cn/build_and_install/rv1126.md#基于-paddlelite-的-fastdeploy-交叉编译库编译) + +2. 将编译后的库拷贝到当前目录，可使用如下命令： +```bash +cp -r FastDeploy/build/fastdeploy-tmivx/ FastDeploy/examples/vision/detection/yolov5/rv1126/cpp +``` + +3. 在当前路径下载部署所需的模型和示例图片： +```bash +mkdir models && mkdir images +wget https://bj.bcebos.com/fastdeploy/models/ppyoloe_noshare_qat.tar.gz +tar -xvf ppyoloe_noshare_qat.tar.gz +cp -r ppyoloe_noshare_qat models +wget https://gitee.com/paddlepaddle/PaddleDetection/raw/release/2.4/demo/000000014439.jpg +cp -r 000000014439.jpg images +``` + +4. 编译部署示例，可使入如下命令： +```bash +mkdir build && cd build +cmake -DCMAKE_TOOLCHAIN_FILE=${PWD}/../fastdeploy-tmivx/timvx.cmake -DFASTDEPLOY_INSTALL_DIR=${PWD}/../fastdeploy-tmivx .. +make -j8 +make install +# 成功编译之后，会生成 install 文件夹，里面有一个运行 demo 和部署所需的库 +``` + +5. 基于 adb 工具部署 PP-YOLOE 检测模型到 Rockchip RV1126，可使用如下命令： +```bash +# 进入 install 目录 +cd FastDeploy/examples/vision/detection/paddledetection/rv1126/cpp/build/install/ +# 如下命令表示：bash run_with_adb.sh 需要运行的demo 模型路径 图片路径 设备的DEVICE_ID +bash run_with_adb.sh infer_demo ppyoloe_noshare_qat 000000014439.jpg $DEVICE_ID +``` + +部署成功后运行结果如下： + + + +需要特别注意的是，在 RV1126 上部署的模型需要是量化后的模型，模型的量化请参考：[模型量化](../../../../../../docs/cn/quantize.md) diff --git a/examples/vision/detection/paddledetection/rv1126/cpp/infer_ppyoloe.cc b/examples/vision/detection/paddledetection/rv1126/cpp/infer_ppyoloe.cc new file mode 100755 index 0000000000..77368584f3 --- /dev/null +++ b/examples/vision/detection/paddledetection/rv1126/cpp/infer_ppyoloe.cc @@ -0,0 +1,66 @@ +// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// 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 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "fastdeploy/vision.h" +#ifdef WIN32 +const char sep = '\\'; +#else +const char sep = '/'; +#endif + +void InitAndInfer(const std::string& model_dir, const std::string& image_file) { + auto model_file = model_dir + sep + "model.pdmodel"; + auto params_file = model_dir + sep + "model.pdiparams"; + auto config_file = model_dir + sep + "infer_cfg.yml"; + auto subgraph_file = model_dir + sep + "subgraph.txt"; + + fastdeploy::RuntimeOption option; + option.UseTimVX(); + option.SetLiteSubgraphPartitionPath(subgraph_file); + + auto model = fastdeploy::vision::detection::PPYOLOE(model_file, params_file, + config_file, option); + assert(model.Initialized()); + + auto im = cv::imread(image_file); + + fastdeploy::vision::DetectionResult res; + if (!model.Predict(im, &res)) { + std::cerr << "Failed to predict." << std::endl; + return; + } + + std::cout << res.Str() << std::endl; + + auto vis_im = fastdeploy::vision::VisDetection(im, res, 0.5); + cv::imwrite("vis_result.jpg", vis_im); + std::cout << "Visualized result saved in ./vis_result.jpg" << std::endl; + +} + +int main(int argc, char* argv[]) { + if (argc < 3) { + std::cout << "Usage: infer_demo path/to/quant_model " + "path/to/image " + "run_option, " + "e.g ./infer_demo ./PPYOLOE_L_quant ./test.jpeg" + << std::endl; + return -1; + } + + std::string model_dir = argv[1]; + std::string test_image = argv[2]; + InitAndInfer(model_dir, test_image); + return 0; +} diff --git a/examples/vision/detection/paddledetection/rv1126/cpp/run_with_adb.sh b/examples/vision/detection/paddledetection/rv1126/cpp/run_with_adb.sh new file mode 100755 index 0000000000..dd7d7b47d2 --- /dev/null +++ b/examples/vision/detection/paddledetection/rv1126/cpp/run_with_adb.sh @@ -0,0 +1,47 @@ +#!/bin/bash +HOST_SPACE=${PWD} +echo ${HOST_SPACE} +WORK_SPACE=/data/local/tmp/test + +# The first parameter represents the demo name +DEMO_NAME=image_classification_demo +if [ -n "$1" ]; then + DEMO_NAME=$1 +fi + +# The second parameter represents the model name +MODEL_NAME=mobilenet_v1_fp32_224 +if [ -n "$2" ]; then + MODEL_NAME=$2 +fi + +# The third parameter indicates the name of the image to be tested +IMAGE_NAME=0001.jpg +if [ -n "$3" ]; then + IMAGE_NAME=$3 +fi + +# The fourth parameter represents the ID of the device +ADB_DEVICE_NAME= +if [ -n "$4" ]; then + ADB_DEVICE_NAME="-s $4" +fi + +# Set the environment variables required during the running process +EXPORT_ENVIRONMENT_VARIABLES="export GLOG_v=5; export SUBGRAPH_ONLINE_MODE=true; export RKNPU_LOGLEVEL=5; export RKNN_LOG_LEVEL=5; ulimit -c unlimited; export VIV_VX_ENABLE_GRAPH_TRANSFORM=-pcq:1; export VIV_VX_SET_PER_CHANNEL_ENTROPY=100; export TIMVX_BATCHNORM_FUSION_MAX_ALLOWED_QUANT_SCALE_DEVIATION=300000; export VSI_NN_LOG_LEVEL=5;" + +EXPORT_ENVIRONMENT_VARIABLES="${EXPORT_ENVIRONMENT_VARIABLES}export LD_LIBRARY_PATH=${WORK_SPACE}/lib:\$LD_LIBRARY_PATH;" + +# Please install adb, and DON'T run this in the docker. +set -e +adb $ADB_DEVICE_NAME shell "rm -rf $WORK_SPACE" +adb $ADB_DEVICE_NAME shell "mkdir -p $WORK_SPACE" + +# Upload the demo, librarys, model and test images to the device +adb $ADB_DEVICE_NAME push ${HOST_SPACE}/lib $WORK_SPACE +adb $ADB_DEVICE_NAME push ${HOST_SPACE}/${DEMO_NAME} $WORK_SPACE +adb $ADB_DEVICE_NAME push models $WORK_SPACE +adb $ADB_DEVICE_NAME push images $WORK_SPACE + +# Execute the deployment demo +adb $ADB_DEVICE_NAME shell "cd $WORK_SPACE; ${EXPORT_ENVIRONMENT_VARIABLES} chmod +x ./${DEMO_NAME}; ./${DEMO_NAME} ./models/${MODEL_NAME} ./images/$IMAGE_NAME" diff --git a/examples/vision/detection/yolov5/quantize/README.md b/examples/vision/detection/yolov5/quantize/README.md old mode 100644 new mode 100755 index 20b628d9b9..aa0a8d5431 --- a/examples/vision/detection/yolov5/quantize/README.md +++ b/examples/vision/detection/yolov5/quantize/README.md @@ -4,7 +4,7 @@ FastDeploy已支持部署量化模型,并提供一键模型自动化压缩的工 ## FastDeploy一键模型自动化压缩工具 FastDeploy 提供了一键模型自动化压缩工具, 能够简单地通过输入一个配置文件, 对模型进行量化. -详细教程请见: [一键模型自动化压缩工具](../../../../../tools/auto_compression/) +详细教程请见: [一键模型自动化压缩工具](../../../../../tools/common_tools/auto_compression/) ## 下载量化完成的YOLOv5s模型 用户也可以直接下载下表中的量化模型进行部署.(点击模型名字即可下载) diff --git a/examples/vision/detection/yolov5/quantize/README_EN.md b/examples/vision/detection/yolov5/quantize/README_EN.md old mode 100644 new mode 100755 index c704ccce2e..d897710f98 --- a/examples/vision/detection/yolov5/quantize/README_EN.md +++ b/examples/vision/detection/yolov5/quantize/README_EN.md @@ -6,7 +6,7 @@ Users can use the one-click model quantization tool to quantize and deploy the m ## FastDeploy One-Click Model Quantization Tool FastDeploy provides a one-click quantization tool that allows users to quantize a model simply with a configuration file. -For a detailed tutorial, please refer to: [One-Click Model Quantization Tool](../../../../../tools/auto_compression/) +For a detailed tutorial, please refer to: [One-Click Model Quantization Tool](../../../../../tools/common_tools/auto_compression/) ## Download Quantized YOLOv5s Model diff --git a/examples/vision/detection/yolov5/quantize/cpp/README.md b/examples/vision/detection/yolov5/quantize/cpp/README.md old mode 100644 new mode 100755 index 90334c7baf..baee6d351d --- a/examples/vision/detection/yolov5/quantize/cpp/README.md +++ b/examples/vision/detection/yolov5/quantize/cpp/README.md @@ -9,7 +9,7 @@ ### 量化模型准备 - 1. 用户可以直接使用由FastDeploy提供的量化模型进行部署. -- 2. 用户可以使用FastDeploy提供的[一键模型自动化压缩工具](../../../../../../tools/auto_compression/),自行进行模型量化, 并使用产出的量化模型进行部署. +- 2. 用户可以使用FastDeploy提供的[一键模型自动化压缩工具](../../../../../../tools/common_tools/auto_compression/),自行进行模型量化, 并使用产出的量化模型进行部署. ## 以量化后的YOLOv5s模型为例, 进行部署 在本目录执行如下命令即可完成编译,以及量化模型部署.支持此模型需保证FastDeploy版本0.7.0以上(x.x.x>=0.7.0) diff --git a/examples/vision/detection/yolov5/quantize/python/README.md b/examples/vision/detection/yolov5/quantize/python/README.md old mode 100644 new mode 100755 index c63ef712c5..9108e256e6 --- a/examples/vision/detection/yolov5/quantize/python/README.md +++ b/examples/vision/detection/yolov5/quantize/python/README.md @@ -8,7 +8,7 @@ ### 量化模型准备 - 1. 用户可以直接使用由FastDeploy提供的量化模型进行部署. -- 2. 用户可以使用FastDeploy提供的[一键模型自动化压缩工具](../../../../../../tools/auto_compression/),自行进行模型量化, 并使用产出的量化模型进行部署. +- 2. 用户可以使用FastDeploy提供的[一键模型自动化压缩工具](../../../../../../tools/common_tools/auto_compression/),自行进行模型量化, 并使用产出的量化模型进行部署. ## 以量化后的YOLOv5s模型为例, 进行部署 diff --git a/examples/vision/detection/yolov5/rv1126/README.md b/examples/vision/detection/yolov5/rv1126/README.md new file mode 100755 index 0000000000..968cc4306a --- /dev/null +++ b/examples/vision/detection/yolov5/rv1126/README.md @@ -0,0 +1,11 @@ +# YOLOv5 量化模型在 RV1126 上的部署 +目前 FastDeploy 已经支持基于 PaddleLite 部署 YOLOv5 量化模型到 RV1126 上。 + +模型的量化和量化模型的下载请参考：[模型量化](../quantize/README.md) + + +## 详细部署文档 + +在 RV1126 上只支持 C++ 的部署。 + +- [C++部署](cpp) diff --git a/examples/vision/detection/yolov5/rv1126/cpp/CMakeLists.txt b/examples/vision/detection/yolov5/rv1126/cpp/CMakeLists.txt new file mode 100755 index 0000000000..3c9eee38a2 --- /dev/null +++ b/examples/vision/detection/yolov5/rv1126/cpp/CMakeLists.txt @@ -0,0 +1,37 @@ +PROJECT(infer_demo C CXX) +CMAKE_MINIMUM_REQUIRED (VERSION 3.10) + +# 指定下载解压后的fastdeploy库路径 +option(FASTDEPLOY_INSTALL_DIR "Path of downloaded fastdeploy sdk.") + +include(${FASTDEPLOY_INSTALL_DIR}/FastDeploy.cmake) + +# 添加FastDeploy依赖头文件 +include_directories(${FASTDEPLOY_INCS}) +include_directories(${FastDeploy_INCLUDE_DIRS}) + +add_executable(infer_demo ${PROJECT_SOURCE_DIR}/infer.cc) +# 添加FastDeploy库依赖 +target_link_libraries(infer_demo ${FASTDEPLOY_LIBS}) + +set(CMAKE_INSTALL_PREFIX ${CMAKE_SOURCE_DIR}/build/install) + +install(TARGETS infer_demo DESTINATION ./) + +install(DIRECTORY models DESTINATION ./) +install(DIRECTORY images DESTINATION ./) + +file(GLOB FASTDEPLOY_LIBS ${FASTDEPLOY_INSTALL_DIR}/lib/*) +install(PROGRAMS ${FASTDEPLOY_LIBS} DESTINATION lib) + +file(GLOB OPENCV_LIBS ${FASTDEPLOY_INSTALL_DIR}/third_libs/install/opencv/lib/lib*) +install(PROGRAMS ${OPENCV_LIBS} DESTINATION lib) + +file(GLOB PADDLELITE_LIBS ${FASTDEPLOY_INSTALL_DIR}/third_libs/install/paddlelite/lib/lib*) +install(PROGRAMS ${PADDLELITE_LIBS} DESTINATION lib) + +file(GLOB TIMVX_LIBS ${FASTDEPLOY_INSTALL_DIR}/third_libs/install/paddlelite/lib/verisilicon_timvx/*) +install(PROGRAMS ${TIMVX_LIBS} DESTINATION lib) + +file(GLOB ADB_TOOLS run_with_adb.sh) +install(PROGRAMS ${ADB_TOOLS} DESTINATION ./) diff --git a/examples/vision/detection/yolov5/rv1126/cpp/README.md b/examples/vision/detection/yolov5/rv1126/cpp/README.md new file mode 100755 index 0000000000..9711577f20 --- /dev/null +++ b/examples/vision/detection/yolov5/rv1126/cpp/README.md @@ -0,0 +1,54 @@ +# YOLOv5 量化模型 C++ 部署示例 + +本目录下提供的 `infer.cc`，可以帮助用户快速完成 YOLOv5 量化模型在 RV1126 上的部署推理加速。 + +## 部署准备 +### FastDeploy 交叉编译环境准备 +- 1. 软硬件环境满足要求，以及交叉编译环境的准备，请参考：[FastDeploy 交叉编译环境准备](../../../../../../docs/cn/build_and_install/rv1126.md#交叉编译环境搭建) + +### 量化模型准备 +- 1. 用户可以直接使用由 FastDeploy 提供的量化模型进行部署。 +- 2. 用户可以使用 FastDeploy 提供的[一键模型自动化压缩工具](../../../../../../tools/common_tools/auto_compression/),自行进行模型量化, 并使用产出的量化模型进行部署。 +- 更多量化相关相关信息可查阅[模型量化](../../quantize/README.md) + +## 在 RV1126 上部署量化后的 YOLOv5 检测模型 +请按照以下步骤完成在 RV1126 上部署 YOLOv5 量化模型： +1. 交叉编译编译 FastDeploy 库，具体请参考：[交叉编译 FastDeploy](../../../../../../docs/cn/build_and_install/rv1126.md#基于-paddlelite-的-fastdeploy-交叉编译库编译) + +2. 将编译后的库拷贝到当前目录，可使用如下命令： +```bash +cp -r FastDeploy/build/fastdeploy-tmivx/ FastDeploy/examples/vision/detection/yolov5/rv1126/cpp +``` + +3. 在当前路径下载部署所需的模型和示例图片： +```bash +mkdir models && mkdir images +wget https://bj.bcebos.com/fastdeploy/models/yolov5s_ptq_model.tar.gz +tar -xvf yolov5s_ptq_model.tar.gz +cp -r yolov5s_ptq_model models +wget https://gitee.com/paddlepaddle/PaddleDetection/raw/release/2.4/demo/000000014439.jpg +cp -r 000000014439.jpg images +``` + +4. 编译部署示例，可使入如下命令： +```bash +mkdir build && cd build +cmake -DCMAKE_TOOLCHAIN_FILE=${PWD}/../fastdeploy-tmivx/timvx.cmake -DFASTDEPLOY_INSTALL_DIR=${PWD}/../fastdeploy-tmivx .. +make -j8 +make install +# 成功编译之后，会生成 install 文件夹，里面有一个运行 demo 和部署所需的库 +``` + +5. 基于 adb 工具部署 YOLOv5 检测模型到 Rockchip RV1126，可使用如下命令： +```bash +# 进入 install 目录 +cd FastDeploy/examples/vision/detection/yolov5/rv1126/cpp/build/install/ +# 如下命令表示：bash run_with_adb.sh 需要运行的demo 模型路径 图片路径 设备的DEVICE_ID +bash run_with_adb.sh infer_demo yolov5s_ptq_model 000000014439.jpg $DEVICE_ID +``` + +部署成功后，vis_result.jpg 保存的结果如下： + + + +需要特别注意的是，在 RV1126 上部署的模型需要是量化后的模型，模型的量化请参考：[模型量化](../../../../../../docs/cn/quantize.md) diff --git a/examples/vision/detection/yolov5/rv1126/cpp/infer.cc b/examples/vision/detection/yolov5/rv1126/cpp/infer.cc new file mode 100755 index 0000000000..f1cf9e8dc3 --- /dev/null +++ b/examples/vision/detection/yolov5/rv1126/cpp/infer.cc @@ -0,0 +1,64 @@ +// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// 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 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "fastdeploy/vision.h" +#ifdef WIN32 +const char sep = '\\'; +#else +const char sep = '/'; +#endif + +void InitAndInfer(const std::string& model_dir, const std::string& image_file) { + auto model_file = model_dir + sep + "model.pdmodel"; + auto params_file = model_dir + sep + "model.pdiparams"; + auto subgraph_file = model_dir + sep + "subgraph.txt"; + + fastdeploy::RuntimeOption option; + option.UseTimVX(); + option.SetLiteSubgraphPartitionPath(subgraph_file); + + auto model = fastdeploy::vision::detection::YOLOv5( + model_file, params_file, option, fastdeploy::ModelFormat::PADDLE); + assert(model.Initialized()); + + auto im = cv::imread(image_file); + + fastdeploy::vision::DetectionResult res; + if (!model.Predict(im, &res)) { + std::cerr << "Failed to predict." << std::endl; + return; + } + + std::cout << res.Str() << std::endl; + + auto vis_im = fastdeploy::vision::VisDetection(im, res); + cv::imwrite("vis_result.jpg", vis_im); + std::cout << "Visualized result saved in ./vis_result.jpg" << std::endl; +} + +int main(int argc, char* argv[]) { + if (argc < 3) { + std::cout << "Usage: infer_demo path/to/quant_model " + "path/to/image " + "run_option, " + "e.g ./infer_demo ./yolov5s_quant ./000000014439.jpg" + << std::endl; + return -1; + } + + std::string model_dir = argv[1]; + std::string test_image = argv[2]; + InitAndInfer(model_dir, test_image); + return 0; +} diff --git a/examples/vision/detection/yolov5/rv1126/cpp/run_with_adb.sh b/examples/vision/detection/yolov5/rv1126/cpp/run_with_adb.sh new file mode 100755 index 0000000000..aacaed4c51 --- /dev/null +++ b/examples/vision/detection/yolov5/rv1126/cpp/run_with_adb.sh @@ -0,0 +1,47 @@ +#!/bin/bash +HOST_SPACE=${PWD} +echo ${HOST_SPACE} +WORK_SPACE=/data/local/tmp/test + +# The first parameter represents the demo name +DEMO_NAME=image_classification_demo +if [ -n "$1" ]; then + DEMO_NAME=$1 +fi + +# The second parameter represents the model name +MODEL_NAME=mobilenet_v1_fp32_224 +if [ -n "$2" ]; then + MODEL_NAME=$2 +fi + +# The third parameter indicates the name of the image to be tested +IMAGE_NAME=0001.jpg +if [ -n "$3" ]; then + IMAGE_NAME=$3 +fi + +# The fourth parameter represents the ID of the device +ADB_DEVICE_NAME= +if [ -n "$4" ]; then + ADB_DEVICE_NAME="-s $4" +fi + +# Set the environment variables required during the running process +EXPORT_ENVIRONMENT_VARIABLES="export GLOG_v=5; export VIV_VX_ENABLE_GRAPH_TRANSFORM=-pcq:1; export VIV_VX_SET_PER_CHANNEL_ENTROPY=100; export TIMVX_BATCHNORM_FUSION_MAX_ALLOWED_QUANT_SCALE_DEVIATION=300000; export VSI_NN_LOG_LEVEL=5;" + +EXPORT_ENVIRONMENT_VARIABLES="${EXPORT_ENVIRONMENT_VARIABLES}export LD_LIBRARY_PATH=${WORK_SPACE}/lib:\$LD_LIBRARY_PATH;" + +# Please install adb, and DON'T run this in the docker. +set -e +adb $ADB_DEVICE_NAME shell "rm -rf $WORK_SPACE" +adb $ADB_DEVICE_NAME shell "mkdir -p $WORK_SPACE" + +# Upload the demo, librarys, model and test images to the device +adb $ADB_DEVICE_NAME push ${HOST_SPACE}/lib $WORK_SPACE +adb $ADB_DEVICE_NAME push ${HOST_SPACE}/${DEMO_NAME} $WORK_SPACE +adb $ADB_DEVICE_NAME push models $WORK_SPACE +adb $ADB_DEVICE_NAME push images $WORK_SPACE + +# Execute the deployment demo +adb $ADB_DEVICE_NAME shell "cd $WORK_SPACE; ${EXPORT_ENVIRONMENT_VARIABLES} chmod +x ./${DEMO_NAME}; ./${DEMO_NAME} ./models/${MODEL_NAME} ./images/$IMAGE_NAME" diff --git a/examples/vision/detection/yolov6/quantize/README.md b/examples/vision/detection/yolov6/quantize/README.md old mode 100644 new mode 100755 index bceb33afb8..57c71cedbf --- a/examples/vision/detection/yolov6/quantize/README.md +++ b/examples/vision/detection/yolov6/quantize/README.md @@ -4,7 +4,7 @@ FastDeploy已支持部署量化模型,并提供一键模型自动化压缩的工 ## FastDeploy一键模型自动化压缩工具 FastDeploy 提供了一键模型自动化压缩工具, 能够简单地通过输入一个配置文件, 对模型进行量化. -详细教程请见: [一键模型自动化压缩工具](../../../../../tools/auto_compression/) +详细教程请见: [一键模型自动化压缩工具](../../../../../tools/common_tools/auto_compression/) ## 下载量化完成的YOLOv6s模型 用户也可以直接下载下表中的量化模型进行部署.(点击模型名字即可下载) diff --git a/examples/vision/detection/yolov6/quantize/README_EN.md b/examples/vision/detection/yolov6/quantize/README_EN.md old mode 100644 new mode 100755 index b7ae61d115..7a13ef7d0e --- a/examples/vision/detection/yolov6/quantize/README_EN.md +++ b/examples/vision/detection/yolov6/quantize/README_EN.md @@ -6,7 +6,7 @@ Users can use the one-click model quantization tool to quantize and deploy the m ## FastDeploy One-Click Model Quantization Tool FastDeploy provides a one-click quantization tool that allows users to quantize a model simply with a configuration file. -For detailed tutorial, please refer to : [One-Click Model Quantization Tool](../../../../../tools/auto_compression/) +For detailed tutorial, please refer to : [One-Click Model Quantization Tool](../../../../../tools/common_tools/auto_compression/) ## Download Quantized YOLOv6s Model diff --git a/examples/vision/detection/yolov6/quantize/cpp/README.md b/examples/vision/detection/yolov6/quantize/cpp/README.md old mode 100644 new mode 100755 index ee11f315de..7ad762100b --- a/examples/vision/detection/yolov6/quantize/cpp/README.md +++ b/examples/vision/detection/yolov6/quantize/cpp/README.md @@ -9,7 +9,7 @@ ### 量化模型准备 - 1. 用户可以直接使用由FastDeploy提供的量化模型进行部署. -- 2. 用户可以使用FastDeploy提供的[一键模型自动化压缩工具](../../../../../../tools/auto_compression/),自行进行模型量化, 并使用产出的量化模型进行部署. +- 2. 用户可以使用FastDeploy提供的[一键模型自动化压缩工具](../../../../../../tools/common_tools/auto_compression/),自行进行模型量化, 并使用产出的量化模型进行部署. ## 以量化后的YOLOv6s模型为例, 进行部署 在本目录执行如下命令即可完成编译,以及量化模型部署.支持此模型需保证FastDeploy版本0.7.0以上(x.x.x>=0.7.0) diff --git a/examples/vision/detection/yolov6/quantize/python/README.md b/examples/vision/detection/yolov6/quantize/python/README.md old mode 100644 new mode 100755 index 169a0a9468..057e13f9a3 --- a/examples/vision/detection/yolov6/quantize/python/README.md +++ b/examples/vision/detection/yolov6/quantize/python/README.md @@ -8,7 +8,7 @@ ### 量化模型准备 - 1. 用户可以直接使用由FastDeploy提供的量化模型进行部署. -- 2. 用户可以使用FastDeploy提供的[一键模型自动化压缩工具](../../../../../../tools/auto_compression/),自行进行模型量化, 并使用产出的量化模型进行部署. +- 2. 用户可以使用FastDeploy提供的[一键模型自动化压缩工具](../../../../../../tools/common_tools/auto_compression/),自行进行模型量化, 并使用产出的量化模型进行部署. ## 以量化后的YOLOv6s模型为例, 进行部署 ```bash diff --git a/examples/vision/detection/yolov7/quantize/README.md b/examples/vision/detection/yolov7/quantize/README.md old mode 100644 new mode 100755 index 9c5261c342..e1bbc2f699 --- a/examples/vision/detection/yolov7/quantize/README.md +++ b/examples/vision/detection/yolov7/quantize/README.md @@ -4,7 +4,7 @@ FastDeploy已支持部署量化模型,并提供一键模型自动化压缩的工 ## FastDeploy一键模型自动化压缩工具 FastDeploy 提供了一键模型自动化压缩工具, 能够简单地通过输入一个配置文件, 对模型进行量化. -详细教程请见: [一键模型自动化压缩工具](../../../../../tools/auto_compression/) +详细教程请见: [一键模型自动化压缩工具](../../../../../tools/common_tools/auto_compression/) ## 下载量化完成的YOLOv7模型 用户也可以直接下载下表中的量化模型进行部署.(点击模型名字即可下载) diff --git a/examples/vision/detection/yolov7/quantize/README_EN.md b/examples/vision/detection/yolov7/quantize/README_EN.md old mode 100644 new mode 100755 index 4e6b2e3533..5a1add2e6c --- a/examples/vision/detection/yolov7/quantize/README_EN.md +++ b/examples/vision/detection/yolov7/quantize/README_EN.md @@ -6,7 +6,7 @@ Users can use the one-click model quantization tool to quantize and deploy the m ## FastDeploy One-Click Model Quantization Tool FastDeploy provides a one-click quantization tool that allows users to quantize a model simply with a configuration file. -For detailed tutorial, please refer to : [One-Click Model Quantization Tool](../../../../../tools/auto_compression/) +For detailed tutorial, please refer to : [One-Click Model Quantization Tool](../../../../../tools/common_tools/auto_compression/) ## Download Quantized YOLOv7 Model diff --git a/examples/vision/detection/yolov7/quantize/cpp/README.md b/examples/vision/detection/yolov7/quantize/cpp/README.md old mode 100644 new mode 100755 index 17eb9eb0fe..d59964bcc9 --- a/examples/vision/detection/yolov7/quantize/cpp/README.md +++ b/examples/vision/detection/yolov7/quantize/cpp/README.md @@ -9,7 +9,7 @@ ### 量化模型准备 - 1. 用户可以直接使用由FastDeploy提供的量化模型进行部署. -- 2. 用户可以使用FastDeploy提供的[一键模型自动化压缩工具](../../../../../../tools/auto_compression/),自行进行模型量化, 并使用产出的量化模型进行部署. +- 2. 用户可以使用FastDeploy提供的[一键模型自动化压缩工具](../../../../../../tools/common_tools/auto_compression/),自行进行模型量化, 并使用产出的量化模型进行部署. ## 以量化后的YOLOv7模型为例, 进行部署 在本目录执行如下命令即可完成编译,以及量化模型部署.支持此模型需保证FastDeploy版本0.7.0以上(x.x.x>=0.7.0) diff --git a/examples/vision/detection/yolov7/quantize/python/README.md b/examples/vision/detection/yolov7/quantize/python/README.md old mode 100644 new mode 100755 index 4389d44c18..08ee726e22 --- a/examples/vision/detection/yolov7/quantize/python/README.md +++ b/examples/vision/detection/yolov7/quantize/python/README.md @@ -8,7 +8,7 @@ ### 量化模型准备 - 1. 用户可以直接使用由FastDeploy提供的量化模型进行部署. -- 2. 用户可以使用FastDeploy提供的[一键模型自动化压缩工具](../../../../../../tools/auto_compression/),自行进行模型量化, 并使用产出的量化模型进行部署. +- 2. 用户可以使用FastDeploy提供的[一键模型自动化压缩工具](../../../../../../tools/common_tools/auto_compression/),自行进行模型量化, 并使用产出的量化模型进行部署. ## 以量化后的YOLOv7模型为例, 进行部署 ```bash diff --git a/examples/vision/segmentation/paddleseg/quantize/README.md b/examples/vision/segmentation/paddleseg/quantize/README.md old mode 100644 new mode 100755 index add706d22f..83a76e3846 --- a/examples/vision/segmentation/paddleseg/quantize/README.md +++ b/examples/vision/segmentation/paddleseg/quantize/README.md @@ -4,7 +4,7 @@ FastDeploy已支持部署量化模型,并提供一键模型自动化压缩的工 ## FastDeploy一键模型自动化压缩工具 FastDeploy 提供了一键模型自动化压缩工具, 能够简单地通过输入一个配置文件, 对模型进行量化. -详细教程请见: [一键模型自动化压缩工具](../../../../../tools/auto_compression/) +详细教程请见: [一键模型自动化压缩工具](../../../../../tools/common_tools/auto_compression/) 注意: 推理量化后的分类模型仍然需要FP32模型文件夹下的deploy.yaml文件, 自行量化的模型文件夹内不包含此yaml文件, 用户从FP32模型文件夹下复制此yaml文件到量化后的模型文件夹内即可。 ## 下载量化完成的PaddleSeg模型 diff --git a/examples/vision/segmentation/paddleseg/quantize/cpp/README.md b/examples/vision/segmentation/paddleseg/quantize/cpp/README.md old mode 100644 new mode 100755 index 0fbaac173d..bd17ec634e --- a/examples/vision/segmentation/paddleseg/quantize/cpp/README.md +++ b/examples/vision/segmentation/paddleseg/quantize/cpp/README.md @@ -8,7 +8,7 @@ ### 量化模型准备 - 1. 用户可以直接使用由FastDeploy提供的量化模型进行部署. -- 2. 用户可以使用FastDeploy提供的[一键模型自动化压缩工具](../../../../../../tools/auto_compression/),自行进行模型量化, 并使用产出的量化模型进行部署.(注意: 推理量化后的分类模型仍然需要FP32模型文件夹下的deploy.yaml文件, 自行量化的模型文件夹内不包含此yaml文件, 用户从FP32模型文件夹下复制此yaml文件到量化后的模型文件夹内即可.) +- 2. 用户可以使用FastDeploy提供的[一键模型自动化压缩工具](../../../../../../tools/common_tools/auto_compression/),自行进行模型量化, 并使用产出的量化模型进行部署.(注意: 推理量化后的分类模型仍然需要FP32模型文件夹下的deploy.yaml文件, 自行量化的模型文件夹内不包含此yaml文件, 用户从FP32模型文件夹下复制此yaml文件到量化后的模型文件夹内即可.) ## 以量化后的PP_LiteSeg_T_STDC1_cityscapes模型为例, 进行部署 在本目录执行如下命令即可完成编译,以及量化模型部署.支持此模型需保证FastDeploy版本0.7.0以上(x.x.x>=0.7.0) diff --git a/examples/vision/segmentation/paddleseg/quantize/python/README.md b/examples/vision/segmentation/paddleseg/quantize/python/README.md old mode 100644 new mode 100755 index 2e06ae145c..b9cc1c6c4a --- a/examples/vision/segmentation/paddleseg/quantize/python/README.md +++ b/examples/vision/segmentation/paddleseg/quantize/python/README.md @@ -8,7 +8,7 @@ ### 量化模型准备 - 1. 用户可以直接使用由FastDeploy提供的量化模型进行部署. -- 2. 用户可以使用FastDeploy提供的[一键模型自动化压缩工具](../../../../../../tools/auto_compression/),自行进行模型量化, 并使用产出的量化模型进行部署.(注意: 推理量化后的分类模型仍然需要FP32模型文件夹下的deploy.yaml文件, 自行量化的模型文件夹内不包含此yaml文件, 用户从FP32模型文件夹下复制此yaml文件到量化后的模型文件夹内即可.) +- 2. 用户可以使用FastDeploy提供的[一键模型自动化压缩工具](../../../../../../tools/common_tools/auto_compression/),自行进行模型量化, 并使用产出的量化模型进行部署.(注意: 推理量化后的分类模型仍然需要FP32模型文件夹下的deploy.yaml文件, 自行量化的模型文件夹内不包含此yaml文件, 用户从FP32模型文件夹下复制此yaml文件到量化后的模型文件夹内即可.) ## 以量化后的PP_LiteSeg_T_STDC1_cityscapes模型为例, 进行部署 diff --git a/examples/vision/segmentation/paddleseg/rv1126/README.md b/examples/vision/segmentation/paddleseg/rv1126/README.md new file mode 100755 index 0000000000..ed5d150807 --- /dev/null +++ b/examples/vision/segmentation/paddleseg/rv1126/README.md @@ -0,0 +1,11 @@ +# PP-LiteSeg 量化模型在 RV1126 上的部署 +目前 FastDeploy 已经支持基于 PaddleLite 部署 PP-LiteSeg 量化模型到 RV1126 上。 + +模型的量化和量化模型的下载请参考：[模型量化](../quantize/README.md) + + +## 详细部署文档 + +在 RV1126 上只支持 C++ 的部署。 + +- [C++部署](cpp) diff --git a/examples/vision/segmentation/paddleseg/rv1126/cpp/CMakeLists.txt b/examples/vision/segmentation/paddleseg/rv1126/cpp/CMakeLists.txt new file mode 100755 index 0000000000..baaf8331f1 --- /dev/null +++ b/examples/vision/segmentation/paddleseg/rv1126/cpp/CMakeLists.txt @@ -0,0 +1,38 @@ +PROJECT(infer_demo C CXX) +CMAKE_MINIMUM_REQUIRED (VERSION 3.10) + +# 指定下载解压后的fastdeploy库路径 +option(FASTDEPLOY_INSTALL_DIR "Path of downloaded fastdeploy sdk.") + +include(${FASTDEPLOY_INSTALL_DIR}/FastDeploy.cmake) + +# 添加FastDeploy依赖头文件 +include_directories(${FASTDEPLOY_INCS}) +include_directories(${FastDeploy_INCLUDE_DIRS}) + +add_executable(infer_demo ${PROJECT_SOURCE_DIR}/infer.cc) +# 添加FastDeploy库依赖 +target_link_libraries(infer_demo ${FASTDEPLOY_LIBS}) + +set(CMAKE_INSTALL_PREFIX ${CMAKE_SOURCE_DIR}/build/install) + +install(TARGETS infer_demo DESTINATION ./) + +install(DIRECTORY models DESTINATION ./) +install(DIRECTORY images DESTINATION ./) +# install(DIRECTORY run_with_adb.sh DESTINATION ./) + +file(GLOB FASTDEPLOY_LIBS ${FASTDEPLOY_INSTALL_DIR}/lib/*) +install(PROGRAMS ${FASTDEPLOY_LIBS} DESTINATION lib) + +file(GLOB OPENCV_LIBS ${FASTDEPLOY_INSTALL_DIR}/third_libs/install/opencv/lib/lib*) +install(PROGRAMS ${OPENCV_LIBS} DESTINATION lib) + +file(GLOB PADDLELITE_LIBS ${FASTDEPLOY_INSTALL_DIR}/third_libs/install/paddlelite/lib/lib*) +install(PROGRAMS ${PADDLELITE_LIBS} DESTINATION lib) + +file(GLOB TIMVX_LIBS ${FASTDEPLOY_INSTALL_DIR}/third_libs/install/paddlelite/lib/verisilicon_timvx/*) +install(PROGRAMS ${TIMVX_LIBS} DESTINATION lib) + +file(GLOB ADB_TOOLS run_with_adb.sh) +install(PROGRAMS ${ADB_TOOLS} DESTINATION ./) diff --git a/examples/vision/segmentation/paddleseg/rv1126/cpp/README.md b/examples/vision/segmentation/paddleseg/rv1126/cpp/README.md new file mode 100755 index 0000000000..6295fa728e --- /dev/null +++ b/examples/vision/segmentation/paddleseg/rv1126/cpp/README.md @@ -0,0 +1,54 @@ +# PP-LiteSeg 量化模型 C++ 部署示例 + +本目录下提供的 `infer.cc`，可以帮助用户快速完成 PP-LiteSeg 量化模型在 RV1126 上的部署推理加速。 + +## 部署准备 +### FastDeploy 交叉编译环境准备 +- 1. 软硬件环境满足要求，以及交叉编译环境的准备，请参考：[FastDeploy 交叉编译环境准备](../../../../../../docs/cn/build_and_install/rv1126.md#交叉编译环境搭建) + +### 模型准备 +- 1. 用户可以直接使用由 FastDeploy 提供的量化模型进行部署。 +- 2. 用户可以使用 FastDeploy 提供的一键模型自动化压缩工具,自行进行模型量化, 并使用产出的量化模型进行部署.(注意: 推理量化后的分类模型仍然需要FP32模型文件夹下的 deploy.yaml 文件, 自行量化的模型文件夹内不包含此 yaml 文件, 用户从FP32模型文件夹下复制此yaml文件到量化后的模型文件夹内即可.) +- 更多量化相关相关信息可查阅[模型量化](../../quantize/README.md) + +## 在 RV1126 上部署量化后的 PP-LiteSeg 分割模型 +请按照以下步骤完成在 RV1126 上部署 PP-LiteSeg 量化模型： +1. 交叉编译编译 FastDeploy 库，具体请参考：[交叉编译 FastDeploy](../../../../../../docs/cn/build_and_install/rv1126.md#基于-paddlelite-的-fastdeploy-交叉编译库编译) + +2. 将编译后的库拷贝到当前目录，可使用如下命令： +```bash +cp -r FastDeploy/build/fastdeploy-tmivx/ FastDeploy/examples/vision/segmentation/paddleseg/rv1126/cpp +``` + +3. 在当前路径下载部署所需的模型和示例图片： +```bash +mkdir models && mkdir images +wget https://bj.bcebos.com/fastdeploy/models/rk1/ppliteseg.tar.gz +tar -xvf ppliteseg.tar.gz +cp -r ppliteseg models +wget https://paddleseg.bj.bcebos.com/dygraph/demo/cityscapes_demo.png +cp -r cityscapes_demo.png images +``` + +4. 编译部署示例，可使入如下命令： +```bash +mkdir build && cd build +cmake -DCMAKE_TOOLCHAIN_FILE=${PWD}/../fastdeploy-tmivx/timvx.cmake -DFASTDEPLOY_INSTALL_DIR=${PWD}/../fastdeploy-tmivx .. +make -j8 +make install +# 成功编译之后，会生成 install 文件夹，里面有一个运行 demo 和部署所需的库 +``` + +5. 基于 adb 工具部署 PP-LiteSeg 分割模型到 Rockchip RV1126，可使用如下命令： +```bash +# 进入 install 目录 +cd FastDeploy/examples/vision/segmentation/paddleseg/rv1126/cpp/build/install/ +# 如下命令表示：bash run_with_adb.sh 需要运行的demo 模型路径 图片路径 设备的DEVICE_ID +bash run_with_adb.sh infer_demo ppliteseg cityscapes_demo.png $DEVICE_ID +``` + +部署成功后运行结果如下： + + + +需要特别注意的是，在 RV1126 上部署的模型需要是量化后的模型，模型的量化请参考：[模型量化](../../../../../../docs/cn/quantize.md) diff --git a/examples/vision/segmentation/paddleseg/rv1126/cpp/infer.cc b/examples/vision/segmentation/paddleseg/rv1126/cpp/infer.cc new file mode 100755 index 0000000000..8c9c7456ce --- /dev/null +++ b/examples/vision/segmentation/paddleseg/rv1126/cpp/infer.cc @@ -0,0 +1,66 @@ +// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// 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 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "fastdeploy/vision.h" +#ifdef WIN32 +const char sep = '\\'; +#else +const char sep = '/'; +#endif + +void InitAndInfer(const std::string& model_dir, const std::string& image_file) { + auto model_file = model_dir + sep + "model.pdmodel"; + auto params_file = model_dir + sep + "model.pdiparams"; + auto config_file = model_dir + sep + "deploy.yaml"; + auto subgraph_file = model_dir + sep + "subgraph.txt"; + + fastdeploy::RuntimeOption option; + option.UseTimVX(); + option.SetLiteSubgraphPartitionPath(subgraph_file); + + auto model = fastdeploy::vision::segmentation::PaddleSegModel( + model_file, params_file, config_file,option); + + assert(model.Initialized()); + + auto im = cv::imread(image_file); + + fastdeploy::vision::SegmentationResult res; + if (!model.Predict(im, &res)) { + std::cerr << "Failed to predict." << std::endl; + return; + } + + std::cout << res.Str() << std::endl; + + auto vis_im = fastdeploy::vision::VisSegmentation(im, res, 0.5); + cv::imwrite("vis_result.jpg", vis_im); + std::cout << "Visualized result saved in ./vis_result.jpg" << std::endl; +} + +int main(int argc, char* argv[]) { + if (argc < 3) { + std::cout << "Usage: infer_demo path/to/quant_model " + "path/to/image " + "run_option, " + "e.g ./infer_demo ./ResNet50_vd_quant ./test.jpeg" + << std::endl; + return -1; + } + + std::string model_dir = argv[1]; + std::string test_image = argv[2]; + InitAndInfer(model_dir, test_image); + return 0; +} \ No newline at end of file diff --git a/examples/vision/segmentation/paddleseg/rv1126/cpp/run_with_adb.sh b/examples/vision/segmentation/paddleseg/rv1126/cpp/run_with_adb.sh new file mode 100755 index 0000000000..aacaed4c51 --- /dev/null +++ b/examples/vision/segmentation/paddleseg/rv1126/cpp/run_with_adb.sh @@ -0,0 +1,47 @@ +#!/bin/bash +HOST_SPACE=${PWD} +echo ${HOST_SPACE} +WORK_SPACE=/data/local/tmp/test + +# The first parameter represents the demo name +DEMO_NAME=image_classification_demo +if [ -n "$1" ]; then + DEMO_NAME=$1 +fi + +# The second parameter represents the model name +MODEL_NAME=mobilenet_v1_fp32_224 +if [ -n "$2" ]; then + MODEL_NAME=$2 +fi + +# The third parameter indicates the name of the image to be tested +IMAGE_NAME=0001.jpg +if [ -n "$3" ]; then + IMAGE_NAME=$3 +fi + +# The fourth parameter represents the ID of the device +ADB_DEVICE_NAME= +if [ -n "$4" ]; then + ADB_DEVICE_NAME="-s $4" +fi + +# Set the environment variables required during the running process +EXPORT_ENVIRONMENT_VARIABLES="export GLOG_v=5; export VIV_VX_ENABLE_GRAPH_TRANSFORM=-pcq:1; export VIV_VX_SET_PER_CHANNEL_ENTROPY=100; export TIMVX_BATCHNORM_FUSION_MAX_ALLOWED_QUANT_SCALE_DEVIATION=300000; export VSI_NN_LOG_LEVEL=5;" + +EXPORT_ENVIRONMENT_VARIABLES="${EXPORT_ENVIRONMENT_VARIABLES}export LD_LIBRARY_PATH=${WORK_SPACE}/lib:\$LD_LIBRARY_PATH;" + +# Please install adb, and DON'T run this in the docker. +set -e +adb $ADB_DEVICE_NAME shell "rm -rf $WORK_SPACE" +adb $ADB_DEVICE_NAME shell "mkdir -p $WORK_SPACE" + +# Upload the demo, librarys, model and test images to the device +adb $ADB_DEVICE_NAME push ${HOST_SPACE}/lib $WORK_SPACE +adb $ADB_DEVICE_NAME push ${HOST_SPACE}/${DEMO_NAME} $WORK_SPACE +adb $ADB_DEVICE_NAME push models $WORK_SPACE +adb $ADB_DEVICE_NAME push images $WORK_SPACE + +# Execute the deployment demo +adb $ADB_DEVICE_NAME shell "cd $WORK_SPACE; ${EXPORT_ENVIRONMENT_VARIABLES} chmod +x ./${DEMO_NAME}; ./${DEMO_NAME} ./models/${MODEL_NAME} ./images/$IMAGE_NAME" diff --git a/fastdeploy/backends/lite/lite_backend.cc b/fastdeploy/backends/lite/lite_backend.cc index 42797411d8..53d8fd5426 100755 --- a/fastdeploy/backends/lite/lite_backend.cc +++ b/fastdeploy/backends/lite/lite_backend.cc @@ -72,7 +72,7 @@ void LiteBackend::BuildOption(const LiteBackendOption& option) { } } } - if(option_.enable_timvx){ + if(option_.enable_timvx) { config_.set_nnadapter_device_names({"verisilicon_timvx"}); valid_places.push_back( paddle::lite_api::Place{TARGET(kNNAdapter), PRECISION(kInt8)}); diff --git a/fastdeploy/vision/detection/contrib/yolov5/yolov5.cc b/fastdeploy/vision/detection/contrib/yolov5/yolov5.cc index 422cf501c4..92d53dd107 100755 --- a/fastdeploy/vision/detection/contrib/yolov5/yolov5.cc +++ b/fastdeploy/vision/detection/contrib/yolov5/yolov5.cc @@ -27,6 +27,7 @@ YOLOv5::YOLOv5(const std::string& model_file, const std::string& params_file, } else { valid_cpu_backends = {Backend::PDINFER, Backend::ORT, Backend::LITE}; valid_gpu_backends = {Backend::PDINFER, Backend::ORT, Backend::TRT}; + valid_timvx_backends = {Backend::LITE}; } runtime_option = custom_option; runtime_option.model_format = model_format; diff --git a/fastdeploy/vision/detection/ppdet/model.h b/fastdeploy/vision/detection/ppdet/model.h old mode 100644 new mode 100755 index a686198ebe..44f574435d --- a/fastdeploy/vision/detection/ppdet/model.h +++ b/fastdeploy/vision/detection/ppdet/model.h @@ -64,6 +64,7 @@ class FASTDEPLOY_DECL PPYOLOE : public PPDetBase { valid_cpu_backends = {Backend::OPENVINO, Backend::ORT, Backend::PDINFER, Backend::LITE}; valid_gpu_backends = {Backend::ORT, Backend::PDINFER, Backend::TRT}; + valid_timvx_backends = {Backend::LITE}; initialized = Initialize(); } diff --git a/fastdeploy/vision/segmentation/ppseg/model.cc b/fastdeploy/vision/segmentation/ppseg/model.cc old mode 100644 new mode 100755 index 1bf83131cf..ad2a7773a9 --- a/fastdeploy/vision/segmentation/ppseg/model.cc +++ b/fastdeploy/vision/segmentation/ppseg/model.cc @@ -27,6 +27,7 @@ PaddleSegModel::PaddleSegModel(const std::string& model_file, valid_cpu_backends = {Backend::OPENVINO, Backend::PDINFER, Backend::ORT, Backend::LITE}; valid_gpu_backends = {Backend::PDINFER, Backend::ORT, Backend::TRT}; valid_rknpu_backends = {Backend::RKNPU2}; + valid_timvx_backends = {Backend::LITE}; runtime_option = custom_option; runtime_option.model_format = model_format; runtime_option.model_file = model_file;