使用MindSpore Lite Native API开发AI应用
使用场景
开发者可使用MindSpore Lite提供的Native API来部署AI算法，并提供高层接口供UI层调用，进行AI模型推理。典型场景如：AI套件SDK开发。
基本概念
- N-API：用于构建JS本地化组件的一套接口。可利用N-API，将C/C++开发的库封装成JS模块。
环境准备
- 安装DevEco Studio，要求版本 >= 3.1.0.500，并更新SDK到API 10或以上。
开发步骤
1. 新建Native工程
打开DevEco Studio，新建工程，依次点击 File -> New -> Create Project 创建 Native C++ 模板工程。在创建出的工程 entry/src/main/ 目录下会默认包含 cpp/ 目录，可以在此目录放置C/C++代码，并提供JS API供UI调用。
2. 编写C++推理代码
假设开发者已准备好.ms格式模型。
在使用MindSpore Lite Native API进行开发前，需要先引用对应的头文件。
#include <iostream>
#include <mindspore/model.h>
#include <mindspore/context.h>
#include <mindspore/status.h>
#include <mindspore/tensor.h>
#include <rawfile/raw_file_manager.h>
(1). 读取模型文件。
#define LOGI(...) ((void)OH_LOG_Print(LOG_APP, LOG_INFO, LOG_DOMAIN, "[MSLiteNapi]", __VA_ARGS__))
#define LOGD(...) ((void)OH_LOG_Print(LOG_APP, LOG_DEBUG, LOG_DOMAIN, "[MSLiteNapi]", __VA_ARGS__))
#define LOGW(...) ((void)OH_LOG_Print(LOG_APP, LOG_WARN, LOG_DOMAIN, "[MSLiteNapi]", __VA_ARGS__))
#define LOGE(...) ((void)OH_LOG_Print(LOG_APP, LOG_ERROR, LOG_DOMAIN, "[MSLiteNapi]", __VA_ARGS__))
void *ReadModelFile(NativeResourceManager *nativeResourceManager, const std::string &modelName, size_t *modelSize) {
auto rawFile = OH_ResourceManager_OpenRawFile(nativeResourceManager, modelName.c_str());
if (rawFile == nullptr) {
LOGE("Open model file failed");
return nullptr;
}
long fileSize = OH_ResourceManager_GetRawFileSize(rawFile);
void *modelBuffer = malloc(fileSize);
if (modelBuffer == nullptr) {
LOGE("Get model file size failed");
}
int ret = OH_ResourceManager_ReadRawFile(rawFile, modelBuffer, fileSize);
if (ret == 0) {
LOGE("Read model file failed");
OH_ResourceManager_CloseRawFile(rawFile);
return nullptr;
}
OH_ResourceManager_CloseRawFile(rawFile);
*modelSize = fileSize;
return modelBuffer;
}
(2). 创建上下文，设置线程数、设备类型等参数，并加载模型。
void DestroyModelBuffer(void **buffer) {
if (buffer == nullptr) {
return;
}
free(*buffer);
*buffer = nullptr;
}
OH_AI_ModelHandle CreateMSLiteModel(void *modelBuffer, size_t modelSize) {
// 创建上下文
auto context = OH_AI_ContextCreate();
if (context == nullptr) {
DestroyModelBuffer(&modelBuffer);
LOGE("Create MSLite context failed.\n");
return nullptr;
}
auto cpu_device_info = OH_AI_DeviceInfoCreate(OH_AI_DEVICETYPE_CPU);
OH_AI_ContextAddDeviceInfo(context, cpu_device_info);
// 加载.ms模型文件
auto model = OH_AI_ModelCreate();
if (model == nullptr) {
DestroyModelBuffer(&modelBuffer);
LOGE("Allocate MSLite Model failed.\n");
return nullptr;
}
auto build_ret = OH_AI_ModelBuild(model, modelBuffer, modelSize, OH_AI_MODELTYPE_MINDIR, context);
DestroyModelBuffer(&modelBuffer);
if (build_ret != OH_AI_STATUS_SUCCESS) {
OH_AI_ModelDestroy(&model);
LOGE("Build MSLite model failed.\n");
return nullptr;
}
LOGI("Build MSLite model success.\n");
return model;
}
(3). 设置模型输入数据，执行模型推理并获取输出数据。
#define GET_PARAMS(env, info, num) \
size_t argc = num; \
napi_value argv[num] = {nullptr}; \
napi_value thisVar = nullptr; \
void *data = nullptr; \
napi_get_cb_info(env, info, &argc, argv, &thisVar, &data)
constexpr int kNumPrintOfOutData = 10;
constexpr int RANDOM_RANGE = 128;
void FillTensorWithRandom(OH_AI_TensorHandle msTensor) {
auto size = OH_AI_TensorGetDataSize(msTensor);
char *data = (char *)OH_AI_TensorGetMutableData(msTensor);
for (size_t i = 0; i < size; i++) {
data[i] = (char)(rand() / RANDOM_RANGE);
}
}
// fill data to inputs tensor
int FillInputTensors(OH_AI_TensorHandleArray &inputs) {
for (size_t i = 0; i < inputs.handle_num; i++) {
FillTensorWithRandom(inputs.handle_list[i]);
}
return OH_AI_STATUS_SUCCESS;
}
void RunMSLiteModel(OH_AI_ModelHandle model) {
// 设置模型输入数据
auto inputs = OH_AI_ModelGetInputs(model);
FillInputTensors(inputs);
auto outputs = OH_AI_ModelGetOutputs(model);
// 执行推理并打印输出
auto predict_ret = OH_AI_ModelPredict(model, inputs, &outputs, nullptr, nullptr);
if (predict_ret != OH_AI_STATUS_SUCCESS) {
OH_AI_ModelDestroy(&model);
LOGE("Predict MSLite model error.\n");
return;
}
LOGI("Run MSLite model success.\n");
LOGI("Get model outputs:\n");
for (size_t i = 0; i < outputs.handle_num; i++) {
auto tensor = outputs.handle_list[i];
LOGI("- Tensor %{public}d name is: %{public}s.\n", static_cast<int>(i), OH_AI_TensorGetName(tensor));
LOGI("- Tensor %{public}d size is: %{public}d.\n", static_cast<int>(i), (int)OH_AI_TensorGetDataSize(tensor));
auto out_data = reinterpret_cast<const float *>(OH_AI_TensorGetData(tensor));
std::cout << "Output data is:";
for (int i = 0; (i < OH_AI_TensorGetElementNum(tensor)) && (i <= kNumPrintOfOutData); i++) {
std::cout << out_data[i] << " ";
}
std::cout << std::endl;
}
OH_AI_ModelDestroy(&model);
}
(4). 调用以上3个方法，实现完整的模型推理流程。
static napi_value RunDemo(napi_env env, napi_callback_info info)
{
LOGI("Enter runDemo()");
GET_PARAMS(env, info, 2);
napi_value error_ret;
napi_create_int32(env, -1, &error_ret);
const std::string modelName = "ml_headpose.ms";
size_t modelSize;
auto resourcesManager = OH_ResourceManager_InitNativeResourceManager(env, argv[1]);
auto modelBuffer = ReadModelFile(resourcesManager, modelName, &modelSize);
if (modelBuffer == nullptr) {
LOGE("Read model failed");
return error_ret;
}
LOGI("Read model file success");
auto model = CreateMSLiteModel(modelBuffer, modelSize);
if (model == nullptr) {
OH_AI_ModelDestroy(&model);
LOGE("MSLiteFwk Build model failed.\n");
return error_ret;
}
RunMSLiteModel(model);
napi_value success_ret;
napi_create_int32(env, 0, &success_ret);
LOGI("Exit runDemo()");
return success_ret;
}
(5). 编写CMake脚本，链接MindSpore Lite动态库libmindspore_lite_ndk.so。
cmake_minimum_required(VERSION 3.4.1)
project(OHOSMSLiteNapi)
set(NATIVERENDER_ROOT_PATH ${CMAKE_CURRENT_SOURCE_DIR})
include_directories(${NATIVERENDER_ROOT_PATH}
${NATIVERENDER_ROOT_PATH}/include)
add_library(mslite_napi SHARED mslite_napi.cpp)
target_link_libraries(mslite_napi PUBLIC mindspore_lite_ndk) # 链接MindSpore Lite动态库。
target_link_libraries(mslite_napi PUBLIC hilog_ndk.z)
target_link_libraries(mslite_napi PUBLIC rawfile.z)
target_link_libraries(mslite_napi PUBLIC ace_napi.z)
3. 使用N-API将C++动态库封装成JS模块
在 entry/src/main/cpp/types/ 新建 libmslite_api/ 子目录，并在子目录中创建 index.d.ts，内容如下：
export const runDemo: (a:String, b:Object) => number;
以上代码用于定义JS接口runDemo() 。
另外，新增 oh-package.json5 文件，将API与so相关联，成为一个完整的JS模块：
{
"name": "libmslite_napi.so",
"types": "./index.d.ts"
}
4. 在UI代码中调用封装的MindSpore模块
在 entry/src/ets/MainAbility/pages/index.ets 中，定义onClick()事件，并在事件回调中调用封装的runDemo()接口。
import hilog from '@ohos.hilog'
import msliteNapi from 'libmslite_napi.so' // 导入msliteNapi模块。
import resManager from '@ohos.resourceManager'
const TAG = 'MSLiteNativeDemo'
@Entry
@Component
struct Index {
@State message: string = 'MindSpore Lite Demo'
build() {
Row() {
Column() {
Text(this.message)
.fontSize(30)
.fontWeight(FontWeight.Bold)
.onClick(async () => {
hilog.info(0x0000, TAG, '*** Start MSLite Demo ***');
let ret: number = 0;
ret = msliteNapi.runDemo("", getContext(this).resourceManager); // 调用runDemo()，执行AI模型推理。
if (ret == -1) {
hilog.error(0x0000, TAG, 'Error when running MSLite Demo!');
}
hilog.info(0x0000, TAG, '*** Finished MSLite Demo ***');
})
}
.width('100%')
}
.height('100%')
}
}
调测验证
在DevEco Studio 中连接设备，点击Run entry运行，应用进程有如下日志：
08-08 16:55:33.766 1513-1529/com.mslite.native_demo I A00000/MSLiteNativeDemo: *** Start MSLite Demo ***
08-08 16:55:33.766 1513-1529/com.mslite.native_demo I A00000/[MSLiteNapi]: Enter runDemo()
08-08 16:55:33.772 1513-1529/com.mslite.native_demo I A00000/[MSLiteNapi]: Read model file success
08-08 16:55:33.799 1513-1529/com.mslite.native_demo I A00000/[MSLiteNapi]: Build MSLite model success.
08-08 16:55:33.818 1513-1529/com.mslite.native_demo I A00000/[MSLiteNapi]: Run MSLite model success.
08-08 16:55:33.818 1513-1529/com.mslite.native_demo I A00000/[MSLiteNapi]: Get model outputs:
08-08 16:55:33.818 1513-1529/com.mslite.native_demo I A00000/[MSLiteNapi]: - Tensor 0 name is: output_node_0.
08-08 16:55:33.818 1513-1529/com.mslite.native_demo I A00000/[MSLiteNapi]: - Tensor 0 size is: 12.
08-08 16:55:33.826 1513-1529/com.mslite.native_demo I A00000/[MSLiteNapi]: Exit runDemo()
08-08 16:55:33.827 1513-1529/com.mslite.native_demo I A00000/MSLiteNativeDemo: *** Finished MSLite Demo ***