Tensorflow 对回归问题给出随机答案

Question

我正在尝试使用 Tensorflow.js 使用 Node.js 重现 Python 练习。

目标是使用机器学习简单地将摄氏温度转换为华氏温度。

但是，我是 Tensorflow.js 的菜鸟，它总是给我随机的答案。

我尝试了多种方法，例如许多不同的形状。 我检查了 Python 和 Node.js 是否具有相同的模型。他们都有以下型号：

Layer (type)                 Output shape              Param #   
=================================================================
dense_Dense1 (Dense)         [null,1]                  2         
=================================================================
Total params: 2
Trainable params: 2
Non-trainable params: 0

const tf = require("@tensorflow/tfjs-node")

function convert(c){
    return (c*1.8)+32 // Convert celsius to fahrenheit
} 

var celsius = []
var fahrenheit = []

for (let i = 0; i < 20; i++) {
    var r = 100; // Keeping this only value to ensure that Tf knows the answer I also have tried with 20 different values but doesn't work
    celsius.push([r]) // Shape [20,1]
    fahrenheit.push([convert(r)]) // Push the answer (212) to the fahrenheit array
}

var model = tf.sequential();
model.add(tf.layers.dense({inputShape:[1], units: 1}))

async function trainModel(model, inputs, labels) {
    // Prepare the model for training.  
    model.compile({
      optimizer: tf.train.adam(),
      loss: tf.losses.meanSquaredError,
      metrics: ['accuracy'], // Accuracy = 0
    });

    model.summary(); 

    const epochs = 500;

    return await model.fit(inputs, labels, {
      epochs,
      batchSize: 20, 
      verbose: false // Nothing interesting with verbose
    });
  }

c = tf.tensor(celsius)
f = tf.tensor(fahrenheit)

var training = trainModel(model, c, f)

training.then(function(args){
    var prediction = model.predict(tf.tensor([[100]]));
    prediction.print(); // Prints a random number
    console.log("Real answer = "+convert(100)) 
})

输出的张量值每次都是随机变化的。 这是一个例子：

Tensor
     [[65.9411697],]
Real answer = 212

Answer 1

似乎主要问题是优化器。 - 如果使用 SGD 优化器进行训练。预测工作正常。

const tf = require("@tensorflow/tfjs-node")
const nr_epochs=500; 

function convert(c){
  return (c*1.8)+32 // Convert celsius to fahrenheit
} 


let celsius = []
let fahrenheit = []

for (let i = 0; i < 100; i++) {
  var r = 100; // Keeping this only value to ensure that Tf knows the answer
  celsius.push(i) // Shape [20,1]
  fahrenheit.push(convert(i)) // Push the answer (212) to the fahrenheit array
}

const train = async (xy, ys) => {
  const model = tf.sequential();

  model.add(tf.layers.dense({units: 1, inputShape: [1]}));

  model.compile({loss: 'meanSquaredError', optimizer: 'sgd'});
  await model.fit(xs,ys,{epochs: nr_epochs})
  return model;
}

const predict =  (model, n) => {
  const predicted =  model.predict(tf.tensor2d([n],[1,1])); 
  return predicted;
}

const xs = tf.tensor2d(celsius.slice (0,15), [15,1]);
const ys = tf.tensor2d(fahrenheit.slice (0,15), [15,1]);
(async () => {
  let trained = await train (xs,ys);
  for (let n of [4,6,12]) {
    let predicted = predict (trained, n).dataSync ();
    console.log (`Value: ${n} Predicted: ${predicted [0]}`)
  }
})()

日志：

Value: 4 Predicted: 38.01055908203125
Value: 6 Predicted: 42.033267974853516
Value: 12 Predicted: 54.101402282714844

Answer 2

当我向模型添加另外 3 个更密集的层时，Adam 优化器起作用。但我让它只用一层就可以在 python 上的 adam 上工作。

xs = []
ys = []

for (var i = -100; i < 100; i++) {
  xs.push(i)
  ys.push( i*1.8 + 32)
}

console.log(xs,ys)

model = tf.sequential({
  layers: [
    tf.layers.dense({
            units: 4,
            inputShape: [1]
        }),
        tf.layers.dense({
            units: 4,
            inputShape: [4]
        }),
        tf.layers.dense({
            units: 4,
            inputShape: [4]
        }),
        tf.layers.dense({
            units: 1,
            inputShape: [4]
        })
  ]
})


model.compile({
  loss: 'meanSquaredError',
  optimizer: 'adam'
})

var tfxs = tf.tensor2d(xs,[xs.length,1])
var tfys = tf.tensor2d(ys,[xs.length,1])


model.fit(tfxs, tfys,{epochs: 500}).then(function() {
  model.predict(tfxs).print()
})