Ascend 310 AI处理器上使用MindIR模型进行推理

Linux Ascend 推理应用 初级 中级 高级

查看源文件

概述

Ascend 310是面向边缘场景的高能效高集成度AI处理器。Atlas 200开发者套件又称Atlas 200 Developer Kit(以下简称Atlas 200 DK),是以Atlas 200 AI加速模块为核心的开发者板形态的终端类产品,集成了海思Ascend 310 AI处理器,可以实现图像、视频等多种数据分析与推理计算,可广泛用于智能监控、机器人、无人机、视频服务器等场景。

本教程介绍如何在Atlas 200 DK上使用MindSpore基于MindIR模型文件执行推理,主要包括以下流程:

  1. 开发环境准备,包括制作Atlas 200 DK的SD卡 、配置Python环境和刷配套开发软件包。

  2. 导出MindIR模型文件,这里以ResNet-50模型为例。

  3. 编译推理代码,生成可执行main文件。

  4. 加载保存的MindIR模型,执行推理并查看结果。

开发环境准备

参考Ascend 310 AI处理器上使用AIR进行推理#开发环境准备安装设备环境,然后参考安装指导安装MindSpore。

导出MindIR模型文件

在Ascend 910的机器上训练好目标网络,并保存为CheckPoint文件,通过网络和CheckPoint文件导出对应的MindIR格式模型文件,导出流程参见导出MindIR格式文件

这里提供使用ResNet-50模型导出的示例MindIR文件resnet50_imagenet.mindir

推理目录结构介绍

创建目录放置推理代码工程,例如/home/HwHiAiUser/Ascend/ascend-toolkit/20.0.RC1/acllib_linux.arm64/sample/acl_execute_model/ascend310_resnet50_preprocess_sample,可以从官网示例下载样例代码model目录用于存放上述导出的MindIR模型文件,test_data目录用于存放待分类的图片,推理代码工程目录结构如下:

└─ascend310_resnet50_preprocess_sample
    ├── CMakeLists.txt                    // 构建脚本
    ├── README.md                         // 使用说明
    ├── main.cc                           // 主函数
    ├── model
    │   └── resnet50_imagenet.mindir      // MindIR模型文件
    └── test_data
        ├── ILSVRC2012_val_00002138.JPEG  // 输入样本图片1
        ├── ILSVRC2012_val_00003014.JPEG  // 输入样本图片2
        ├── ...                           // 输入样本图片n

推理代码介绍

使用CPU算子数据预处理

推理代码样例:https://gitee.com/mindspore/docs/blob/r1.2/tutorials/tutorial_code/ascend310_resnet50_preprocess_sample/main.cc

引用mindsporemindspore::dataset的名字空间。

namespace ms = mindspore;
namespace ds = mindspore::dataset;

环境初始化,指定硬件为Ascend 310,DeviceID为0:

auto context = std::make_shared<ms::Context>();
auto ascend310_info = std::make_shared<ms::Ascend310DeviceInfo>();
ascend310_info->SetDeviceID(0);
context->MutableDeviceInfo().push_back(ascend310_info);

加载模型文件:

// Load MindIR model
ms::Graph graph;
ms::Status ret = ms::Serialization::Load(resnet_file, ms::ModelType::kMindIR, &graph);
// Build model with graph object
ms::Model resnet50;
ret = resnet50.Build(ms::GraphCell(graph), context);

获取模型所需输入信息:

std::vector<ms::MSTensor> model_inputs = resnet50.GetInputs();

加载图片文件:

// Readfile is a function to read images
ms::MSTensor ReadFile(const std::string &file);
auto image = ReadFile(image_file);

图片预处理(使用CPU算子):

// Create the CPU operator provided by MindData to get the function object

// Decode the input to RGB format
std::shared_ptr<ds::TensorTransform> decode(new ds::vision::Decode());
// Resize the image to the given size
std::shared_ptr<ds::TensorTransform> resize(new ds::vision::Resize({256}));
// Normalize the input
std::shared_ptr<ds::TensorTransform> normalize(new ds::vision::Normalize(
    {0.485 * 255, 0.456 * 255, 0.406 * 255}, {0.229 * 255, 0.224 * 255, 0.225 * 255}));
// Crop the input image at the center
std::shared_ptr<ds::TensorTransform> center_crop(new ds::vision::CenterCrop({224, 224}));
// shape (H, W, C) to shape (C, H, W)
std::shared_ptr<ds::TensorTransform> hwc2chw(new ds::vision::HWC2CHW());

// // Define a MindData preprocessor
ds::Execute preprocessor({decode, resize, normalize, center_crop, hwc2chw});

// Call the function object to get the processed image
ret = preprocessor(image, &image);

执行推理:

// Create outputs vector
std::vector<ms::MSTensor> outputs;
// Create inputs vector
std::vector<ms::MSTensor> inputs;
inputs.emplace_back(model_inputs[0].Name(), model_inputs[0].DataType(), model_inputs[0].Shape(),
                    image.Data().get(), image.DataSize());
// Call the Predict function of Model for inference
ret = resnet50.Predict(inputs, &outputs);

获取推理结果:

// Output the maximum probability to the screen
std::cout << "Image: " << image_file << " infer result: " << GetMax(outputs[0]) << std::endl;

使用Ascend 310算子数据预处理

Dvpp模块为Ascend 310芯片内置硬件解码器,相较于CPU拥有对图形处理更强劲的性能。支持JPEG图片的解码缩放等基础操作。

引用mindsporemindspore::dataset的名字空间。

namespace ms = mindspore;
namespace ds = mindspore::dataset;

环境初始化,指定硬件为Ascend 310,DeviceID为0:

auto context = std::make_shared<ms::Context>();
auto ascend310_info = std::make_shared<ms::Ascend310DeviceInfo>();
ascend310_info->SetDeviceID(0);
context->MutableDeviceInfo().push_back(ascend310_info);

加载图片文件:

// Readfile is a function to read images
ms::MSTensor ReadFile(const std::string &file);
auto image = ReadFile(image_file);

图片预处理(使用Ascend 310算子):

// Create the Dvpp operator provided by MindData to get the function object

// Decode the input to YUV420 format
std::shared_ptr<ds::TensorTransform> decode(new ds::vision::Decode());
// Resize the image to the given size
std::shared_ptr<ds::TensorTransform> resize(new ds::vision::Resize({256}));
// Normalize the input
std::shared_ptr<ds::TensorTransform> normalize(new ds::vision::Normalize(
    {0.485 * 255, 0.456 * 255, 0.406 * 255}, {0.229 * 255, 0.224 * 255, 0.225 * 255}));
// Crop the input image at the center
std::shared_ptr<ds::TensorTransform> center_crop(new ds::vision::CenterCrop({224, 224}));

图片预处理(使用Ascend 310算子, 性能为CPU算子的2.3倍),需显式指定计算硬件为Ascend 310。

// Define a MindData preprocessor, set deviceType = kAscend310, device id = 0
ds::Execute preprocessor({decode, resize, center_crop, normalize}, MapTargetDevice::kAscend310, 0);

// Call the function object to get the processed image
ret = preprocessor(image, &image);

加载模型文件: 若使用Ascend 310算子,则需要为模型插入Aipp算子。

// Load MindIR model
ms::Graph graph;
ms::Status ret = ms::Serialization::Load(resnet_file, ms::ModelType::kMindIR, &graph);
// Build model with graph object
ascend310_info->SetInsertOpConfigPath(preprocessor.AippCfgGenerator());
ms::Model resnet50;
ret = resnet50.Build(ms::GraphCell(graph), context);

获取模型所需输入信息:

std::vector<ms::MSTensor> model_inputs = resnet50.GetInputs();

执行推理:

// Create outputs vector
std::vector<ms::MSTensor> outputs;
// Create inputs vector
std::vector<ms::MSTensor> inputs;
inputs.emplace_back(model_inputs[0].Name(), model_inputs[0].DataType(), model_inputs[0].Shape(),
                    image.Data().get(), image.DataSize());
// Call the Predict function of Model for inference
ret = resnet50.Predict(inputs, &outputs);

获取推理结果:

// Output the maximum probability to the screen
std::cout << "Image: " << image_file << " infer result: " << GetMax(outputs[0]) << std::endl;

构建脚本介绍

构建脚本用于构建用户程序,样例来自于:https://gitee.com/mindspore/docs/blob/r1.2/tutorials/tutorial_code/ascend310_resnet50_preprocess_sample/CMakeLists.txt

为编译器添加头文件搜索路径:

option(MINDSPORE_PATH "mindspore install path" "")
include_directories(${MINDSPORE_PATH})
include_directories(${MINDSPORE_PATH}/include)

在MindSpore中查找所需动态库:

find_library(MS_LIB libmindspore.so ${MINDSPORE_PATH}/lib)
file(GLOB_RECURSE MD_LIB ${MINDSPORE_PATH}/_c_dataengine*)

使用指定的源文件生成目标可执行文件,并为目标文件链接MindSpore库:

add_executable(resnet50_sample main.cc)
target_link_libraries(resnet50_sample ${MS_LIB} ${MD_LIB})

编译推理代码

进入工程目录ascend310_resnet50_preprocess_sample,设置如下环境变量:

# control log level. 0-DEBUG, 1-INFO, 2-WARNING, 3-ERROR, default level is WARNING.
export GLOG_v=2

# Conda environmental options
LOCAL_ASCEND=/usr/local/Ascend # the root directory of run package

# lib libraries that the run package depends on
export LD_LIBRARY_PATH=${LOCAL_ASCEND}/ascend-toolkit/latest/acllib/lib64:${LOCAL_ASCEND}/ascend-toolkit/latest/atc/lib64:${LOCAL_ASCEND}/driver/lib64:${LOCAL_ASCEND}/ascend-toolkit/latest/opp/op_impl/built-in/ai_core/tbe/op_tiling:${LD_LIBRARY_PATH}

# lib libraries that the mindspore depends on, modify "pip3" according to the actual situation
export LD_LIBRARY_PATH=`pip3 show mindspore-ascend | grep Location | awk '{print $2"/mindspore/lib"}' | xargs realpath`:${LD_LIBRARY_PATH}

# Environment variables that must be configured
export TBE_IMPL_PATH=${LOCAL_ASCEND}/ascend-toolkit/latest/opp/op_impl/built-in/ai_core/tbe            # TBE operator implementation tool path
export ASCEND_OPP_PATH=${LOCAL_ASCEND}/ascend-toolkit/latest/opp                                       # OPP path
export PATH=${LOCAL_ASCEND}/ascend-toolkit/latest/atc/ccec_compiler/bin/:${PATH}                       # TBE operator compilation tool path
export PYTHONPATH=${TBE_IMPL_PATH}:${PYTHONPATH}                                                       # Python library that TBE implementation depends on

执行cmake命令,其中pip3需要按照实际情况修改:

cmake . -DMINDSPORE_PATH=`pip3 show mindspore-ascend | grep Location | awk '{print $2"/mindspore"}' | xargs realpath`

再执行make命令编译即可。

make

编译完成后,在ascend310_resnet50_preprocess_sample下会生成可执行main文件。

执行推理并查看结果

登录Atlas 200 DK开发者板环境,创建model目录放置MindIR文件resnet50_imagenet.mindir,例如/home/HwHiAiUser/Ascend/ascend-toolkit/20.0.RC1/acllib_linux.arm64/sample/acl_execute_model/ascend310_resnet50_preprocess_sample/model。 创建test_data目录放置图片,例如/home/HwHiAiUser/Ascend/ascend-toolkit/20.0.RC1/acllib_linux.arm64/sample/acl_execute_model/ascend310_resnet50_preprocess_sample/test_data。 就可以开始执行推理了:

./resnet50_sample

执行后,会对test_data目录下放置的所有图片进行推理,比如放置了9张ImageNet2012验证集中label为0的图片,可以看到推理结果如下。

Image: ./test_data/ILSVRC2012_val_00002138.JPEG infer result: 0
Image: ./test_data/ILSVRC2012_val_00003014.JPEG infer result: 0
Image: ./test_data/ILSVRC2012_val_00006697.JPEG infer result: 0
Image: ./test_data/ILSVRC2012_val_00007197.JPEG infer result: 0
Image: ./test_data/ILSVRC2012_val_00009111.JPEG infer result: 0
Image: ./test_data/ILSVRC2012_val_00009191.JPEG infer result: 0
Image: ./test_data/ILSVRC2012_val_00009346.JPEG infer result: 0
Image: ./test_data/ILSVRC2012_val_00009379.JPEG infer result: 0
Image: ./test_data/ILSVRC2012_val_00009396.JPEG infer result: 0