VNCoreMLFeatureValueObservation 是否输出 softmax 概率？如果是这样，如何提取最高值？

Question

我正在尝试让我的第一个图像分类模型正常工作，但是， VNClassificationObservation 不工作，而 VNCoreMLFeatureValueObservation 是。

以下是关于我的模型的一些信息：

MLModelDescription: MLModelDescription inputDescriptionsByName: {

"input_1__0" = "input_1__0 : Image (Color, 299 x 299)";

} outputDescriptionsByName: {

    "output_node0__0" = "output_node0__0 : MultiArray (MLMultiArrayDataTypeDouble, 43)";

} predictedFeatureName: (null) 

根据文档：

VNClassificationObservation

This type of observation results from performing a VNCoreMLRequest image
analysis with a Core ML model whose role is classification (rather than 
prediction or image-to-image processing).
Vision infers that an MLModel object is a classifier model if that model 
predicts a single feature. 
That is, the model's modelDescription object has a non-nil value for its 
predictedFeatureName property.

起初我假设当文档说“预测”时，他们指的是具有价值预测的回归类型模型。但现在我认为他们指的是 softmax 预测概率？因此 VNClassificationObservation 不输出 softmax 预测概率。

现在，

VNCoreMLFeatureValueObservation:

Overview
This type of observation results from performing a VNCoreMLRequest image analysis with a Core ML model whose role is prediction rather than classification or image-to-image processing.

Vision infers that an MLModel object is a predictor model if that model predicts multiple features. You can tell that a model predicts multiple features when its modelDescription object has a nil value for its predictedFeatureName property, or when it inserts its output in an outputDescriptionsByName dictionary.

我对措辞感到困惑。这是否意味着多输入、多输出模型？ 不是分类，而是预测，也有点令人困惑，但我假设 由于我得到的输出，softmax probs。

当我运行下面的代码时，我得到：

let request = VNCoreMLRequest(model: model) { [weak self] request, error in
            guard let results = request.results as? [VNCoreMLFeatureValueObservation],
                let topResult = results.first else {
                    fatalError("unexpected result type from VNCoreMLRequest")
DispatchQueue.main.async { [weak self] in

                print("topResult!", topResult)

                //print(model.debugDescription.outputDescriptionsByName)
            }
        }
    let handler = VNImageRequestHandler(ciImage: image)

    DispatchQueue.global(qos: .userInteractive).async {

        do {

            try handler.perform([request])

        } catch {print(error)}

我得到一堆值：

topResult! Optional(<VNCoreMLFeatureValueObservation: 
0x1c003f0c0> C99BC0A0-7722-4DDC-8FB8-C0FEB1CEEFA5 1 "MultiArray : Double 43 
vector

[ 0.02323521859943867,0.03784361109137535,0.0327669121325016,0.02373981475830078,0.01920632272958755,0.01511944644153118,0.0268220379948616,0.00990589614957571,0.006585873663425446,0.02727104164659977,0.02337176166474819,0.0177282840013504,0.01582957617938519,0.01962342299520969,0.0335112139582634,0.01197215262800455,0.04638960584998131,0.0546870082616806,0.008390620350837708,0.02519697323441505,0.01038128975778818,0.02463733218610287,0.05725555866956711,0.02852404117584229,0.01987413503229618,0.02478211745619774,0.01224409975111485,0.03397252038121223,0.02300941571593285,0.02020683139562607,0.03740271925926208,0.01999092660844326,0.03210178017616272,0.02830206602811813,0.01122485008090734,0.01071082800626755,0.02285266295075417,0.01730070635676384,0.009790488518774509,0.01149104069918394,0.03331543132662773,0.01211327593773603,0.0193191897124052]" (1.000000))

如果这些确实是 softmax 概率，我将如何获取最大值的索引？我似乎无法使用.count 或类似的数组方法。

我尝试将其转换为数组，但这两个都不起作用 l

let values  = topResult.featureValue as Array! (Can't convert...coercion)
let values = topResult as Array!

如果这些不是 softmax 值/概率，那么我会去获取 概率。价值观。我正在尝试获取前 3 个 softmax 概率的索引。

谢谢。

!!!更新!!!!!!!!!:

在函数中尝试这样做： var localPrediction：字符串？ 让 topResult = results.first?.featureValue.multiArrayValue

 DispatchQueue.main.async { () in
            var max_value : Float32 = 0
            for i in 0..<topResult!.count{
                if max_value < topResult![i].floatValue{
                    max_value = topResult![i].floatValue
                    localPrediction = String(i)}

                                        }

Answer 1

当您的模型是分类器时，即 mlmodel 文件中的 NeuralNetworkClassifier，则输出为 VNClassificationObservation 对象。

如果您的模型不是分类器，即 NeuralNetwork 或 NeuralNetworkRegressor，则输出是一个或多个 VNCoreMLFeatureValueObservation 对象，其中包含来自最后一层的输出。

因此，如果您希望在 VNCoreMLFeatureValueObservation 中输出 softmax，那么您需要确保您的模型将 softmax 作为其最后一层。

要获取最大元素的索引和值，使用：

func argmax(_ array: UnsafePointer<Double>, count: Int) -> (Int, Double) {
  var maxValue: Double = 0
  var maxIndex: vDSP_Length = 0
  vDSP_maxviD(array, 1, &maxValue, &maxIndex, vDSP_Length(count))
  return (Int(maxIndex), maxValue)
}

要使用它，首先将 MLMultiArray 的 dataPointer 转换为 UnsafePointer<Double>，然后调用 argmax() 函数：

let featurePointer = UnsafePointer<Double>(OpaquePointer(features.dataPointer))
let (maxIndex, maxValue) = argmax(featurePointer, 43)