如何知道 Tensorflow.js 中的模型预测是什么？

Question

我正在使用通过 MobileNetV2 的迁移学习构建的自定义模型来检测人类意识。

我训练了模型，它在测试数据上达到了 99% 的分类准确率。

有3个类即0、5、10代表认知水平

我的模型returns在 React Native 中对来自 TensorCamera 的输入图像进行了以下预测，但我对预测在这个 Tensor 对象中的实际位置感到困惑

const prediction = model.predict(tensor.reshape([1,224,224,3]));

预测输出：

Tensor {
  "dataId": Object {},
  "dtype": "float32",
  "id": 160213,
  "isDisposedInternal": false,
  "kept": false,
  "rankType": "2",
  "scopeId": 365032,
  "shape": Array [
    1,
    3,
  ],
  "size": 3,
  "strides": Array [
    3,
  ],
}

处理相机输入并每 3 帧进行预测的函数

  const handleCameraStream = imageAsTensors => {
    const verbose = true;
    console.log("Tensor input 1");
    try {
      tf.print(imageAsTensors, verbose);
    } catch (e) {
      console.log("Tensor 1 not found!");
    }
    const loop = async () => {
      if (loadedModel !== null) {
        if (frameCount % makePredictionsEveryNFrames === 0) {
          const imageTensor = imageAsTensors.next().value;
          console.log("Tensor input 2");
          tf.print(imageTensor, verbose);
          await getPrediction(imageTensor).catch(e => console.log(e));
        }
      }

      frameCount += 1;
      frameCount = frameCount % makePredictionsEveryNFrames;
      requestAnimationFrameId = requestAnimationFrame(loop);
    };
    //loop infinitely to constantly make predictions
    loop();
  };

获取预测的函数

  const getPrediction = async tensor => {
    if (!tensor) {
      console.log("Tensor not found!");
      return;
    }
    const model = await loadedModel;
    const prediction = model.predict(tensor.reshape([1, 224, 224, 3]));
    if (!prediction || prediction.length === 0) {
      console.log("No prediction available");
      return;
    }
    console.log(prediction);
    //console.log(`Predictions: ${JSON.stringify(prediction)}`);

    // Only take the predictions with a probability of 30% and greater
    //Stop looping
    cancelAnimationFrame(requestAnimationFrameId);
    //setPredictionFound(true);
    //setModelPrediction(prediction[0].className);
    tensor.dispose();
  };

Answer 1

我需要使用 dataSync() 来获取预测结果

    const preds = prediction.dataSync();
    preds.forEach((pred, i) => {
      //console.log(`x: ${i}, pred: ${pred}`);
      if (pred > 0.8) {
        console.log(`x: ${i}, pred: ${pred}`);
        setModelPrediction(prediction: pred, class: i)
      }

x: 0, pred: 0.9000627994537354
x: 1, pred: 0.023466499522328377
x: 2, pred: 0.0764707624912262

如何知道 Tensorflow.js 中的模型预测是什么？

How to know what are the model predictions in Tensorflow.js?

tensorflow

tensorflow.js

tensorflow2.0