使用MindSpore Lite实现图像分类（C/C++）
场景说明
开发者可以使用MindSpore，在UI代码中直接集成MindSpore Lite能力，快速部署AI算法，进行AI模型推理，实现图像分类的应用。
图像分类可实现对图像中物体的识别，在医学影像分析、自动驾驶、电子商务、人脸识别等有广泛的应用。
基本概念
- N-API：用于构建ArkTS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ArkTS模块。
开发流程
- 选择图像分类模型。
- 在端侧使用MindSpore Lite推理模型，实现对选择的图片进行分类。
环境准备
安装DevEco Studio，要求版本 >= 4.1，并更新SDK到API 11或以上。
开发步骤
本文以对相册的一张图片进行推理为例，提供使用MindSpore Lite实现图像分类的开发指导。
选择模型
本示例程序中使用的图像分类模型文件为mobilenetv2.ms，放置在entry/src/main/resources/rawfile工程目录下。
如果开发者有其他图像分类的预训练模型，请参考MindSpore Lite 模型转换介绍，将原始模型转换成.ms格式。
编写代码
图像输入和预处理
-
此处以获取相册图片为例，调用@ohos.file.picker 实现相册图片文件的选择。
import { photoAccessHelper } from '@kit.MediaLibraryKit'; import { BusinessError } from '@kit.BasicServicesKit'; let uris: Array<string> = []; // 创建图片文件选择实例 let photoSelectOptions = new photoAccessHelper.PhotoSelectOptions(); // 设置选择媒体文件类型为IMAGE，设置选择媒体文件的最大数目 photoSelectOptions.MIMEType = photoAccessHelper.PhotoViewMIMETypes.IMAGE_TYPE; photoSelectOptions.maxSelectNumber = 1; // 创建图库选择器实例，调用select()接口拉起图库界面进行文件选择。文件选择成功后，返回photoSelectResult结果集。 let photoPicker = new photoAccessHelper.PhotoViewPicker(); photoPicker.select(photoSelectOptions, async ( err: BusinessError, photoSelectResult: photoAccessHelper.PhotoSelectResult) => { if (err) { console.error('MS_LITE_ERR: PhotoViewPicker.select failed with err: ' + JSON.stringify(err)); return; } console.info('MS_LITE_LOG: PhotoViewPicker.select successfully, ' + 'photoSelectResult uri: ' + JSON.stringify(photoSelectResult)); uris = photoSelectResult.photoUris; console.info('MS_LITE_LOG: uri: ' + uris); }) -
根据模型的输入尺寸，调用@ohos.multimedia.image （实现图片处理）、@ohos.file.fs （实现基础文件操作） API对选择图片进行裁剪、获取图片buffer数据，并进行标准化处理。
import { image } from '@kit.ImageKit'; import { fileIo } from '@kit.CoreFileKit'; let modelInputHeight: number = 224; let modelInputWidth: number = 224; // 使用fileIo.openSync接口，通过uri打开这个文件得到fd let file = fileIo.openSync(this.uris[0], fileIo.OpenMode.READ_ONLY); console.info('MS_LITE_LOG: file fd: ' + file.fd); // 通过fd使用fileIo.readSync接口读取这个文件内的数据 let inputBuffer = new ArrayBuffer(4096000); let readLen = fileIo.readSync(file.fd, inputBuffer); console.info('MS_LITE_LOG: readSync data to file succeed and inputBuffer size is:' + readLen); // 通过PixelMap预处理 let imageSource = image.createImageSource(file.fd); imageSource.createPixelMap().then((pixelMap) => { pixelMap.getImageInfo().then((info) => { console.info('MS_LITE_LOG: info.width = ' + info.size.width); console.info('MS_LITE_LOG: info.height = ' + info.size.height); // 根据模型输入的尺寸，将图片裁剪为对应的size，获取图片buffer数据readBuffer pixelMap.scale(256.0 / info.size.width, 256.0 / info.size.height).then(() => { pixelMap.crop( { x: 16, y: 16, size: { height: modelInputHeight, width: modelInputWidth } } ).then(async () => { let info = await pixelMap.getImageInfo(); console.info('MS_LITE_LOG: crop info.width = ' + info.size.width); console.info('MS_LITE_LOG: crop info.height = ' + info.size.height); // 需要创建的像素buffer大小 let readBuffer = new ArrayBuffer(modelInputHeight * modelInputWidth * 4); await pixelMap.readPixelsToBuffer(readBuffer); console.info('MS_LITE_LOG: Succeeded in reading image pixel data, buffer: ' + readBuffer.byteLength); // 处理readBuffer，转换成float32格式，并进行标准化处理 const imageArr = new Uint8Array( readBuffer.slice(0, modelInputHeight * modelInputWidth * 4)); console.info('MS_LITE_LOG: imageArr length: ' + imageArr.length); let means = [0.485, 0.456, 0.406]; let stds = [0.229, 0.224, 0.225]; let float32View = new Float32Array(modelInputHeight * modelInputWidth * 3); let index = 0; for (let i = 0; i < imageArr.length; i++) { if ((i + 1) % 4 == 0) { float32View[index] = (imageArr[i - 3] / 255.0 - means[0]) / stds[0]; // B float32View[index+1] = (imageArr[i - 2] / 255.0 - means[1]) / stds[1]; // G float32View[index+2] = (imageArr[i - 1] / 255.0 - means[2]) / stds[2]; // R index += 3; } } console.info('MS_LITE_LOG: float32View length: ' + float32View.length); let printStr = 'float32View data:'; for (let i = 0; i < 20; i++) { printStr += ' ' + float32View[i]; } console.info('MS_LITE_LOG: float32View data: ' + printStr); }) }) }); });
编写推理代码
调用MindSpore实现端侧推理，推理代码流程如下。
-
引用对应的头文件
#include <iostream> #include <sstream> #include <stdlib.h> #include <hilog/log.h> #include <rawfile/raw_file_manager.h> #include <mindspore/types.h> #include <mindspore/model.h> #include <mindspore/context.h> #include <mindspore/status.h> #include <mindspore/tensor.h> #include "napi/native_api.h" -
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("MS_LITE_ERR: Open model file failed"); return nullptr; } long fileSize = OH_ResourceManager_GetRawFileSize(rawFile); void *modelBuffer = malloc(fileSize); if (modelBuffer == nullptr) { LOGE("MS_LITE_ERR: OH_ResourceManager_ReadRawFile failed"); } int ret = OH_ResourceManager_ReadRawFile(rawFile, modelBuffer, fileSize); if (ret == 0) { LOGI("MS_LITE_LOG: OH_ResourceManager_ReadRawFile failed"); OH_ResourceManager_CloseRawFile(rawFile); return nullptr; } OH_ResourceManager_CloseRawFile(rawFile); *modelSize = fileSize; return modelBuffer; } -
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_ContextHandle CreateMSLiteContext(void *modelBuffer) { // Set executing context for model. auto context = OH_AI_ContextCreate(); if (context == nullptr) { DestroyModelBuffer(&modelBuffer); LOGE("MS_LITE_ERR: Create MSLite context failed.\n"); return nullptr; } auto cpu_device_info = OH_AI_DeviceInfoCreate(OH_AI_DEVICETYPE_CPU); OH_AI_DeviceInfoSetEnableFP16(cpu_device_info, true); OH_AI_ContextAddDeviceInfo(context, cpu_device_info); LOGI("MS_LITE_LOG: Build MSLite context success.\n"); return context; } OH_AI_ModelHandle CreateMSLiteModel(void *modelBuffer, size_t modelSize, OH_AI_ContextHandle context) { // Create model auto model = OH_AI_ModelCreate(); if (model == nullptr) { DestroyModelBuffer(&modelBuffer); LOGE("MS_LITE_ERR: Allocate MSLite Model failed.\n"); return nullptr; } // Build model object 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("MS_LITE_ERR: Build MSLite model failed.\n"); return nullptr; } LOGI("MS_LITE_LOG: Build MSLite model success.\n"); return model; } -
设置模型输入数据，执行模型推理。
constexpr int K_NUM_PRINT_OF_OUT_DATA = 20; // 设置模型输入数据 int FillInputTensor(OH_AI_TensorHandle input, std::vector<float> input_data) { if (OH_AI_TensorGetDataType(input) == OH_AI_DATATYPE_NUMBERTYPE_FLOAT32) { float *data = (float *)OH_AI_TensorGetMutableData(input); for (size_t i = 0; i < OH_AI_TensorGetElementNum(input); i++) { data[i] = input_data[i]; } return OH_AI_STATUS_SUCCESS; } else { return OH_AI_STATUS_LITE_ERROR; } } // 执行模型推理 int RunMSLiteModel(OH_AI_ModelHandle model, std::vector<float> input_data) { // Set input data for model. auto inputs = OH_AI_ModelGetInputs(model); auto ret = FillInputTensor(inputs.handle_list[0], input_data); if (ret != OH_AI_STATUS_SUCCESS) { LOGE("MS_LITE_ERR: RunMSLiteModel set input error.\n"); return OH_AI_STATUS_LITE_ERROR; } // Get model output. auto outputs = OH_AI_ModelGetOutputs(model); // Predict model. auto predict_ret = OH_AI_ModelPredict(model, inputs, &outputs, nullptr, nullptr); if (predict_ret != OH_AI_STATUS_SUCCESS) { LOGE("MS_LITE_ERR: MSLite Predict error.\n"); return OH_AI_STATUS_LITE_ERROR; } LOGI("MS_LITE_LOG: Run MSLite model Predict success.\n"); // Print output tensor data. LOGI("MS_LITE_LOG: Get model outputs:\n"); for (size_t i = 0; i < outputs.handle_num; i++) { auto tensor = outputs.handle_list[i]; LOGI("MS_LITE_LOG: - Tensor %{public}d name is: %{public}s.\n", static_cast<int>(i), OH_AI_TensorGetName(tensor)); LOGI("MS_LITE_LOG: - Tensor %{public}d size is: %{public}d.\n", static_cast<int>(i), (int)OH_AI_TensorGetDataSize(tensor)); LOGI("MS_LITE_LOG: - Tensor data is:\n"); auto out_data = reinterpret_cast<const float *>(OH_AI_TensorGetData(tensor)); std::stringstream outStr; for (int i = 0; (i < OH_AI_TensorGetElementNum(tensor)) && (i <= K_NUM_PRINT_OF_OUT_DATA); i++) { outStr << out_data[i] << " "; } LOGI("MS_LITE_LOG: %{public}s", outStr.str().c_str()); } return OH_AI_STATUS_SUCCESS; } -
调用以上方法，实现完整的模型推理流程。
static napi_value RunDemo(napi_env env, napi_callback_info info) { LOGI("MS_LITE_LOG: Enter runDemo()"); napi_value error_ret; napi_create_int32(env, -1, &error_ret); // 传入数据处理 size_t argc = 2; napi_value argv[2] = {nullptr}; napi_get_cb_info(env, info, &argc, argv, nullptr, nullptr); bool isArray = false; napi_is_array(env, argv[0], &isArray); uint32_t length = 0; // 获取数组的长度 napi_get_array_length(env, argv[0], &length); LOGI("MS_LITE_LOG: argv array length = %{public}d", length); std::vector<float> input_data; double param = 0; for (int i = 0; i < length; i++) { napi_value value; napi_get_element(env, argv[0], i, &value); napi_get_value_double(env, value, ¶m); input_data.push_back(static_cast<float>(param)); } std::stringstream outstr; for (int i = 0; i < K_NUM_PRINT_OF_OUT_DATA; i++) { outstr << input_data[i] << " "; } LOGI("MS_LITE_LOG: input_data = %{public}s", outstr.str().c_str()); // Read model file const std::string modelName = "mobilenetv2.ms"; LOGI("MS_LITE_LOG: Run model: %{public}s", modelName.c_str()); size_t modelSize; auto resourcesManager = OH_ResourceManager_InitNativeResourceManager(env, argv[1]); auto modelBuffer = ReadModelFile(resourcesManager, modelName, &modelSize); if (modelBuffer == nullptr) { LOGE("MS_LITE_ERR: Read model failed"); return error_ret; } LOGI("MS_LITE_LOG: Read model file success"); auto context = CreateMSLiteContext(modelBuffer); if (context == nullptr) { LOGE("MS_LITE_ERR: MSLiteFwk Build context failed.\n"); return error_ret; } auto model = CreateMSLiteModel(modelBuffer, modelSize, context); if (model == nullptr) { OH_AI_ContextDestroy(&context); LOGE("MS_LITE_ERR: MSLiteFwk Build model failed.\n"); return error_ret; } int ret = RunMSLiteModel(model, input_data); if (ret != OH_AI_STATUS_SUCCESS) { OH_AI_ModelDestroy(&model); OH_AI_ContextDestroy(&context); LOGE("MS_LITE_ERR: RunMSLiteModel failed.\n"); return error_ret; } napi_value out_data; napi_create_array(env, &out_data); auto outputs = OH_AI_ModelGetOutputs(model); OH_AI_TensorHandle output_0 = outputs.handle_list[0]; float *output0Data = reinterpret_cast<float *>(OH_AI_TensorGetMutableData(output_0)); for (size_t i = 0; i < OH_AI_TensorGetElementNum(output_0); i++) { napi_value element; napi_create_double(env, static_cast<double>(output0Data[i]), &element); napi_set_element(env, out_data, i, element); } OH_AI_ModelDestroy(&model); OH_AI_ContextDestroy(&context); LOGI("MS_LITE_LOG: Exit runDemo()"); return out_data; } -
编写CMake脚本，链接MindSpore Lite动态库。
# the minimum version of CMake. cmake_minimum_required(VERSION 3.4.1) project(MindSporeLiteCDemo) set(NATIVERENDER_ROOT_PATH ${CMAKE_CURRENT_SOURCE_DIR}) if(DEFINED PACKAGE_FIND_FILE) include(${PACKAGE_FIND_FILE}) endif() include_directories(${NATIVERENDER_ROOT_PATH} ${NATIVERENDER_ROOT_PATH}/include) add_library(entry SHARED mslite_napi.cpp) target_link_libraries(entry PUBLIC mindspore_lite_ndk) target_link_libraries(entry PUBLIC hilog_ndk.z) target_link_libraries(entry PUBLIC rawfile.z) target_link_libraries(entry PUBLIC ace_napi.z)
使用N-API将C++动态库封装成ArkTS模块
-
在 entry/src/main/cpp/types/libentry/Index.d.ts，定义ArkTS接口
runDemo()。内容如下：export const runDemo: (a: number[], b:Object) => Array<number>; -
在 oh-package.json5 文件，将API与so相关联，成为一个完整的ArkTS模块：
{ "name": "libentry.so", "types": "./Index.d.ts", "version": "1.0.0", "description": "MindSpore Lite inference module" }
调用封装的ArkTS模块进行推理并输出结果
在 entry/src/main/ets/pages/Index.ets 中，调用封装的ArkTS模块，最后对推理结果进行处理。
import msliteNapi from 'libentry.so'
import { resourceManager } from '@kit.LocalizationKit';
let resMgr: resourceManager.ResourceManager = getContext().getApplicationContext().resourceManager;
let max: number = 0;
let maxIndex: number = 0;
let maxArray: Array<number> = [];
let maxIndexArray: Array<number> = [];
// 调用c++的runDemo方法，完成图像输入和预处理后的buffer数据保存在float32View，具体可见上文图像输入和预处理中float32View的定义和处理。
console.info('MS_LITE_LOG: *** Start MSLite Demo ***');
let output: Array<number> = msliteNapi.runDemo(Array.from(float32View), resMgr);
// 取分类占比的最大值
this.max = 0;
this.maxIndex = 0;
this.maxArray = [];
this.maxIndexArray = [];
let newArray = output.filter(value => value !== max);
for (let n = 0; n < 5; n++) {
max = output[0];
maxIndex = 0;
for (let m = 0; m < newArray.length; m++) {
if (newArray[m] > max) {
max = newArray[m];
maxIndex = m;
}
}
maxArray.push(Math.round(this.max * 10000));
maxIndexArray.push(this.maxIndex);
// filter函数数组过滤函数
newArray = newArray.filter(value => value !== max);
}
console.info('MS_LITE_LOG: max:' + this.maxArray);
console.info('MS_LITE_LOG: maxIndex:' + this.maxIndexArray);
console.info('MS_LITE_LOG: *** Finished MSLite Demo ***');
调测验证
-
在DevEco Studio中连接设备，点击Run entry，编译Hap，有如下显示：
Launching com.samples.mindsporelitecdemo $ hdc shell aa force-stop com.samples.mindsporelitecdemo $ hdc shell mkdir data/local/tmp/xxx $ hdc file send C:\Users\xxx\MindSporeLiteCDemo\entry\build\default\outputs\default\entry-default-signed.hap "data/local/tmp/xxx" $ hdc shell bm install -p data/local/tmp/xxx $ hdc shell rm -rf data/local/tmp/xxx $ hdc shell aa start -a EntryAbility -b com.samples.mindsporelitecdemo -
在设备屏幕点击photo按钮，选择图片，点击确定。设备屏幕显示所选图片的分类结果，在日志打印结果中，过滤关键字”MS_LITE“，可得到如下结果：
08-05 17:15:52.001 4684-4684 A03d00/JSAPP pid-4684 I MS_LITE_LOG: PhotoViewPicker.select successfully, photoSelectResult uri: {"photoUris":["file://media/Photo/13/IMG_1501955351_012/plant.jpg"]} ... 08-05 17:15:52.627 4684-4684 A03d00/JSAPP pid-4684 I MS_LITE_LOG: crop info.width = 224 08-05 17:15:52.627 4684-4684 A03d00/JSAPP pid-4684 I MS_LITE_LOG: crop info.height = 224 08-05 17:15:52.628 4684-4684 A03d00/JSAPP pid-4684 I MS_LITE_LOG: Succeeded in reading image pixel data, buffer: 200704 08-05 17:15:52.971 4684-4684 A03d00/JSAPP pid-4684 I MS_LITE_LOG: float32View data: float32View data: 1.2385478019714355 1.308123230934143 1.4722440242767334 1.2385478019714355 1.308123230934143 1.4722440242767334 1.2385478019714355 1.308123230934143 1.4722440242767334 1.2385478019714355 1.308123230934143 1.4722440242767334 1.2385478019714355 1.308123230934143 1.4722440242767334 1.2385478019714355 1.308123230934143 1.4722440242767334 1.2385478019714355 1.308123230934143 08-05 17:15:52.971 4684-4684 A03d00/JSAPP pid-4684 I MS_LITE_LOG: *** Start MSLite Demo *** 08-05 17:15:53.454 4684-4684 A00000/[MSLiteNapi] pid-4684 I MS_LITE_LOG: Build MSLite model success. 08-05 17:15:53.753 4684-4684 A00000/[MSLiteNapi] pid-4684 I MS_LITE_LOG: Run MSLite model Predict success. 08-05 17:15:53.753 4684-4684 A00000/[MSLiteNapi] pid-4684 I MS_LITE_LOG: Get model outputs: 08-05 17:15:53.753 4684-4684 A00000/[MSLiteNapi] pid-4684 I MS_LITE_LOG: - Tensor 0 name is: Default/head-MobileNetV2Head/Sigmoid-op466. 08-05 17:15:53.753 4684-4684 A00000/[MSLiteNapi] pid-4684 I MS_LITE_LOG: - Tensor data is: 08-05 17:15:53.753 4684-4684 A00000/[MSLiteNapi] pid-4684 I MS_LITE_LOG: 3.43385e-06 1.40285e-05 9.11969e-07 4.91007e-05 9.50266e-07 3.94537e-07 0.0434676 3.97196e-05 0.00054832 0.000246202 1.576e-05 3.6494e-06 1.23553e-05 0.196977 5.3028e-05 3.29346e-05 4.90475e-07 1.66109e-06 7.03273e-06 8.83677e-07 3.1365e-06 08-05 17:15:53.781 4684-4684 A03d00/JSAPP pid-4684 W MS_LITE_WARN: output length = 500 ;value = 0.0000034338463592575863,0.000014028532859811094,9.119685273617506e-7,0.000049100715841632336,9.502661555416125e-7,3.945370394831116e-7,0.04346757382154465,0.00003971960904891603,0.0005483203567564487,0.00024620210751891136,0.000015759984307806008,0.0000036493988773145247,0.00001235533181898063,0.1969769448041916,0.000053027983085485175,0.000032934600312728435,4.904751449430478e-7,0.0000016610861166554969,0.000007032729172351537,8.836767619868624e-7 08-05 17:15:53.831 4684-4684 A03d00/JSAPP pid-4684 I MS_LITE_LOG: max:9497,7756,1970,435,46 08-05 17:15:53.831 4684-4684 A03d00/JSAPP pid-4684 I MS_LITE_LOG: maxIndex:323,46,13,6,349 08-05 17:15:53.831 4684-4684 A03d00/JSAPP pid-4684 I MS_LITE_LOG: *** Finished MSLite Demo ***
效果示意
在设备上，点击photo按钮，选择相册中的一张图片，点击确定。在图片下方显示此图片占比前4的分类信息。
相关实例
针对使用MindSpore Lite进行图像分类应用的开发，有以下相关实例可供参考：