场景说明
开发者可以使用@ohos.ai.mindSporeLite，在UI代码中集成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、人脸识别等领域有广泛的应用。
若需基于本Demo适配自有模型，请优先选择静态Shape模型。注意：ArkTS接口不支持NPU后端动态Shape模型推理。
接口说明
这里给出MindSpore Lite推理的通用开发流程中涉及的一些接口，具体请见下列表格。更多接口及详细内容，请见@ohos.ai.mindSporeLite (端侧AI框架)。
| 接口名 | 描述 |
|---|---|
| loadModelFromFile(model: string, context?: Context): Promise<Model> | 从路径加载模型。 |
| loadModelFromBuffer(model: ArrayBuffer, context?: Context): Promise<Model> | 从内存加载模型。 |
| getInputs(): MSTensor[] | 获取模型的输入。 |
| predict(inputs: MSTensor[]): Promise<MSTensor[]> | 推理模型。 |
| getData(): ArrayBuffer | 获取张量的数据。 |
| setData(inputArray: ArrayBuffer): void | 设置张量的数据。 |
[h2]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。
[h2]编写推理代码
-
工程默认设备定义的能力集可能不包含MindSporeLite。需在DevEco Studio工程的entry/src/main目录下，手动创建syscap.json文件。其中，general的取值需与module.json5中的deviceTypes保持一致。配置示例如下：
{ "devices": { "general": [ "default" ] }, "development": { "addedSysCaps": [ "SystemCapability.AI.MindSporeLite" ] } } -
调用@ohos.ai.mindSporeLite实现端侧推理。具体开发过程及细节如下：
- 创建上下文，设置线程数、设备类型等参数。本样例模型，不支持使用NNRt推理。
- 加载模型。本文从内存加载模型。
- 加载数据。模型执行之前需要先获取输入，再向输入的张量中填充数据。
- 执行推理。使用predict接口进行模型推理。
import { mindSporeLite } from '@kit.MindSporeLiteKit' import { hilog } from '@kit.PerformanceAnalysisKit'; export default async function modelPredict( modelBuffer: ArrayBuffer, inputsBuffer: ArrayBuffer[]): Promise<mindSporeLite.MSTensor[]> { // 1.创建上下文，设置线程数、设备类型等参数。本样例模型，不支持配置context.target = ["nnrt"]。 let context: mindSporeLite.Context = {}; context.target = ['cpu']; context.cpu = {} context.cpu.threadNum = 2; context.cpu.threadAffinityMode = 1; context.cpu.precisionMode = 'enforce_fp32'; // 2.从内存加载模型。 let msLiteModel: mindSporeLite.Model = await mindSporeLite.loadModelFromBuffer(modelBuffer, context); // 3.设置输入数据。 let modelInputs: mindSporeLite.MSTensor[] = msLiteModel.getInputs(); for (let i = 0; i < inputsBuffer.length; i++) { let inputBuffer = inputsBuffer[i]; if (inputBuffer != null) { modelInputs[i].setData(inputBuffer as ArrayBuffer); } } // 4.执行推理。 hilog.info(0xFF00, 'MindSporeLiteArkTSDemo', '%{public}s', `=========MS_LITE_LOG: MS_LITE predict start=====`); let modelOutputs: mindSporeLite.MSTensor[] = await msLiteModel.predict(modelInputs); return modelOutputs; }
[h2]实现图像输入和预处理，并执行推理
-
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。
-
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。
-
加载模型文件，调用推理函数，对相册选择的图片进行推理，并对推理结果进行处理。
import modelPredict from './model'; import { photoAccessHelper } from '@kit.MediaLibraryKit'; import { BusinessError } from '@kit.BasicServicesKit'; import { image } from '@kit.ImageKit'; import { fileIo } from '@kit.CoreFileKit'; import { abilityAccessCtrl, Permissions } from '@kit.AbilityKit'; import { hilog } from '@kit.PerformanceAnalysisKit'; const PERMISSIONS: Permissions[] = ['ohos.permission.READ_IMAGEVIDEO']; @Entry @Component struct Index { @State modelPredict: string = 'MindSporeLite ArkTS Demo'; @State modelName: string = 'mobilenetv2.ms'; @State modelInputHeight: number = 224; @State modelInputWidth: number = 224; @State uris: Array<string> = []; @State max: number = 0; @State maxIndex: number = 0; @State maxArray: Array<number> = []; @State maxIndexArray: Array<number> = []; // ... build() { Row() { Column() { Text(this.modelPredict) // ... Button() { Text('photo') .fontSize(30) .fontWeight(FontWeight.Bold) } // ... .onClick(() => { let resMgr = this.getUIContext()?.getHostContext()?.getApplicationContext().resourceManager; if (resMgr === null || resMgr === undefined){ hilog.error(0xFF00, 'MindSporeLiteArkTSDemo', '%{public}s', `MS_LITE_ERR: get resMgr failed.`); return } resMgr?.getRawFileContent(this.modelName).then(modelBuffer => { // 获取相册图片 // 1.创建图片文件选择实例 let photoSelectOptions = new photoAccessHelper.PhotoSelectOptions(); // 2.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; // 过滤选择媒体文件类型为IMAGE photoSelectOptions.maxSelectNumber = 1; // 选择媒体文件的最大数目 // 3.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) { hilog.error(0xFF00, 'MindSporeLiteArkTSDemo', '%{public}s', `MS_LITE_ERR: PhotoViewPicker.select failed with err: ${JSON.stringify(err)}`); return; } hilog.info(0xFF00, 'MindSporeLiteArkTSDemo', '%{public}s', `MS_LITE_LOG: PhotoViewPicker.select successfully, uri: ${JSON.stringify(photoSelectResult)}`); this.uris = photoSelectResult.photoUris; hilog.info(0xFF00, 'MindSporeLiteArkTSDemo', '%{public}s', `MS_LITE_LOG: uri: ${this.uris}`); // 预处理图片数据 try { // 1.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); hilog.info(0xFF00, 'MindSporeLiteArkTSDemo', '%{public}s', `MS_LITE_LOG: file fd: ${file.fd}`); // 2.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); hilog.info(0xFF00, 'MindSporeLiteArkTSDemo', '%{public}s', `MS_LITE_LOG: readSync data to file succeed and inputBuffer size is: ${readLen}`); // 3.通过PixelMap预处理 let imageSource = image.createImageSource(file.fd); if (imageSource === undefined) { hilog.error(0xFF00, 'MindSporeLiteArkTSDemo', '%{public}s', `MS_LITE_ERR: createImageSource failed.`); return } imageSource.createPixelMap().then((pixelMap) => { pixelMap.getImageInfo().then((info) => { hilog.info(0xFF00, 'MindSporeLiteArkTSDemo', '%{public}s', `MS_LITE_LOG: info.width = ${info.size.width}`); hilog.info(0xFF00, 'MindSporeLiteArkTSDemo', '%{public}s', `MS_LITE_LOG: info.height = ${info.size.height}`); // 4.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: this.modelInputHeight, width: this.modelInputWidth } } ).then(async () => { let info = await pixelMap.getImageInfo(); hilog.info(0xFF00, 'MindSporeLiteArkTSDemo', '%{public}s', `MS_LITE_LOG: crop info.width = ${info.size.width}`); hilog.info(0xFF00, 'MindSporeLiteArkTSDemo', '%{public}s', `MS_LITE_LOG: crop info.height = ${info.size.height}`); // 需要创建的像素buffer大小 let readBuffer = new ArrayBuffer(this.modelInputHeight * this.modelInputWidth * 4); await pixelMap.readPixelsToBuffer(readBuffer); hilog.info(0xFF00, 'MindSporeLiteArkTSDemo', '%{public}s', `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, this.modelInputHeight * this.modelInputWidth * 4)); hilog.info(0xFF00, 'MindSporeLiteArkTSDemo', '%{public}s', `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(this.modelInputHeight * this.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; } } hilog.info(0xFF00, 'MindSporeLiteArkTSDemo', '%{public}s', `MS_LITE_LOG: float32View length: ${float32View.length}`); let printStr = 'float32View data:'; for (let i = 0; i < 20; i++) { printStr += ' ' + float32View[i]; } hilog.info(0xFF00, 'MindSporeLiteArkTSDemo', '%{public}s', `MS_LITE_LOG: float32View data: ${printStr}`); let inputs: ArrayBuffer[] = [float32View.buffer]; // predict modelPredict(modelBuffer.buffer.slice(0), inputs).then(outputs => { hilog.info(0xFF00, 'MindSporeLiteArkTSDemo', '%{public}s', `=========MS_LITE_LOG: MS_LITE predict success=====`); // 结果打印 for (let i = 0; i < outputs.length; i++) { let out = new Float32Array(outputs[i].getData()); let printStr = outputs[i].name + ':'; for (let j = 0; j < out.length; j++) { printStr += out[j].toString() + ','; } hilog.info(0xFF00, 'MindSporeLiteArkTSDemo', '%{public}s', `MS_LITE_LOG: ${printStr}`); // 取分类占比的最大值top5 this.max = 0; this.maxIndex = 0; this.maxArray = []; this.maxIndexArray = []; let newArray = out.filter(value => value !== this.max) for (let n = 0; n < 5; n++) { this.max = out[0]; this.maxIndex = 0; // 取最大值 for (let m = 0; m < newArray.length; m++) { if (newArray[m] > this.max) { this.max = newArray[m]; this.maxIndex = m; } } this.maxArray.push(Math.round(this.max * 10000)); this.maxIndexArray.push(this.maxIndex); // filter数组过滤函数 newArray = newArray.filter(value => value !== this.max) } hilog.info(0xFF00, 'MindSporeLiteArkTSDemo', '%{public}s', `MS_LITE_LOG: max: ${this.maxArray}`); hilog.info(0xFF00, 'MindSporeLiteArkTSDemo', '%{public}s', `MS_LITE_LOG: maxIndex: ${this.maxIndexArray}`); } hilog.info(0xFF00, 'MindSporeLiteArkTSDemo', '%{public}s', `=========MS_LITE_LOG END=========`); }) }).catch((error: BusinessError) => { hilog.error(0xFF00, 'MindSporeLiteArkTSDemo', '%{public}s', `MS_LITE_ERR: getRawFileContent promise error is: ${error}`); }) }) }) }) // 5.关闭文件 fileIo.closeSync(file); } catch (err) { hilog.error(0xFF00, 'MindSporeLiteArkTSDemo', '%{public}s', `MS_LITE_ERR: uri: open file fd failed. ${err}`); } }) }) }) } .width('100%') } .height('100%') } }
[h2]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.mindsporelitearktsdemo $ hdc shell aa force-stop com.samples.mindsporelitearktsdemo $ hdc shell mkdir data/local/tmp/xxx $ hdc file send C:\Users\xxx\MindSporeLiteArkTSDemo\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.mindsporelitearktsdemo
-
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：
08-06 03:24:33.743 22547-22547 A03d00/JSAPP com.sampl...liteark+ I MS_LITE_LOG: PhotoViewPicker.select successfully, photoSelectResult uri: {"photoUris":["file://media/Photo/13/IMG_1501955351_012/plant.jpg"]} 08-06 03:24:33.795 22547-22547 A03d00/JSAPP com.sampl...liteark+ I MS_LITE_LOG: readSync data to file succeed and inputBuffer size is:32824 08-06 03:24:34.147 22547-22547 A03d00/JSAPP com.sampl...liteark+ I MS_LITE_LOG: crop info.width = 224 08-06 03:24:34.147 22547-22547 A03d00/JSAPP com.sampl...liteark+ I MS_LITE_LOG: crop info.height = 224 08-06 03:24:34.160 22547-22547 A03d00/JSAPP com.sampl...liteark+ I MS_LITE_LOG: Succeeded in reading image pixel data, buffer: 200704 08-06 03:24:34.970 22547-22547 A03d00/JSAPP com.sampl...liteark+ I =========MS_LITE_LOG: MS_LITE predict start===== 08-06 03:24:35.432 22547-22547 A03d00/JSAPP com.sampl...liteark+ I =========MS_LITE_LOG: MS_LITE predict success===== 08-06 03:24:35.447 22547-22547 A03d00/JSAPP com.sampl...liteark+ I MS_LITE_LOG: Default/head-MobileNetV2Head/Sigmoid-op466:0.0000034338463592575863,0.000014028532859811094,9.119685273617506e-7,0.000049100715841632336,9.502661555416125e-7,3.945370394831116e-7,0.04346757382154465,0.00003971960904891603... 08-06 03:24:35.499 22547-22547 A03d00/JSAPP com.sampl...liteark+ I MS_LITE_LOG: max:9497,7756,1970,435,46 08-06 03:24:35.499 22547-22547 A03d00/JSAPP com.sampl...liteark+ I MS_LITE_LOG: maxIndex:323,46,13,6,349 08-06 03:24:35.499 22547-22547 A03d00/JSAPP com.sampl...liteark+ I =========MS_LITE_LOG END=========
