iOS11新加入了非常有科技感的元素：AR以及機(jī)器學(xué)習(xí)。趁著有機(jī)器學(xué)習(xí)的經(jīng)驗(yàn)，先來(lái)體驗(yàn)一次。

準(zhǔn)備工作

• Xcode 9
• CoreML Model (Inceptionv3)

主要流程

1. 打開(kāi)Xcode創(chuàng)建一個(gè)空的模板工程

• 將下載好的.mlmodel文件拖入工程，并確保在Target->Build Phases->Complie Sources中看見(jiàn)它。??否則無(wú)法使用??
• 基于(懶)快速上手的原因，直接在ViewController中創(chuàng)建一個(gè)UIButton,綁定彈出相冊(cè)方法

  lazy var btn:UIButton = {[unowned self] in
        let btn = UIButton(type: .custom)
        btn.setTitle("select", for: .normal)
        btn.setTitleColor(.black, for: .normal)
        btn.frame.size = CGSize(width: 100, height: 100)
        btn.center = self.view.center
        btn.addTarget(self, action: #selector(handle), for: .touchUpInside)
        return btn
    }()

    @objc private func handle(){
        showAblum()
    }

• 遵守相關(guān)協(xié)議并實(shí)現(xiàn)可選方法獲得所選圖像
  此處自定義了一個(gè)協(xié)議 *** canShowAblum*** 給UIViewController默認(rèn)實(shí)現(xiàn)一個(gè)彈出方法。

protocol canShowAblum : class,UIImagePickerControllerDelegate,UINavigationControllerDelegate{
    
}
extension canShowAblum where Self : UIViewController{
    func showAblum(){
        let picker = UIImagePickerController()
        picker.delegate = self
        picker.sourceType = .photoLibrary
        picker.allowsEditing = true
        present(picker, animated: true, completion: nil)
    }
}

extension ViewController{
    func imagePickerController(_ picker: UIImagePickerController, didFinishPickingMediaWithInfo info: [String : Any]) {
        if let img = info["UIImagePickerControllerOriginalImage"] as? UIImage{
            _ = PictrueLearning(with: img)
        }
        dismiss(animated: true, completion: nil)
    }
}

• 傳入圖像給訓(xùn)練模型，得到結(jié)果
  訓(xùn)練模型封裝在結(jié)構(gòu)體PictureLearning中：

struct PictrueLearning {
    
    private init(){}
    
    init(with image:UIImage){
        self.init()
        train(with: image)//.scale(to: CGSize(width: 299, height: 299)))
    }
    
    fileprivate func train(with image:UIImage){
        //???確保模型初始化成功
        guard let img = image.cgImage,let model = try? VNCoreMLModel(for: Inception().model) else {
            return
        }
        //創(chuàng)建請(qǐng)求
        let request = VNCoreMLRequest(model: model) {
            guard $1 == nil,let resArr = $0.results,let res = resArr.first as? VNClassificationObservation else { return }
            print(res.identifier,res.confidence)
        }
        DispatchQueue.global().async {
            //執(zhí)行
            try? VNImageRequestHandler(cgImage: img).perform([request])
        }
    }
}
extension UIImage{
    func scale(to size:CGSize)->UIImage{
        UIGraphicsBeginImageContext(size)
        draw(in: CGRect(x: 0, y: 0, width: size.width, height: size.height))
        let scaleImage = UIGraphicsGetImageFromCurrentImageContext()
        UIGraphicsEndImageContext()
        return scaleImage ?? UIImage()
    }
}

結(jié)果分析

1.嘗試了多張圖片，發(fā)現(xiàn)綜合結(jié)果較差。 因?yàn)闊o(wú)法窺探此模型的內(nèi)含，只知道是Neural Network Classifier，影響神經(jīng)網(wǎng)絡(luò)的精度的參數(shù)有很多，準(zhǔn)備之后換自己的模型繼續(xù)嘗試。
2.此模型的輸入?yún)?shù)要求為

image.png