昨天有講到除了可以使用Layer API來建立Model之外
其實還有一種Core API可以建立model並進行訓練
Core API跟前幾天學的不太一樣
而且也不是很易於使用
自己要處理的事情變得很多
而且也不能儲存或載入模型，感覺好像在找自己麻煩
不過今天都花時間了解了
而且Google說這樣可以創造更有彈性和更能掌控的模型

我就以Core API改寫一下

var w1 = tf.variable(tf.randomNormal([1, 20]));
var b1 = tf.variable(tf.randomNormal([20]));
var w2 = tf.variable(tf.randomNormal([20, 50]));
var w3 = tf.variable(tf.randomNormal([50, 10]));
var w4 = tf.variable(tf.randomNormal([10, 1]));
var b4 = tf.variable(tf.randomNormal([1]));

function createCoreModel(x) {
    return function (x) {
        return x.mul(w1).add(b1).matMul(w2).sigmoid().matMul(w3).sigmoid().matMul(w4).add(b4)
    };
}

視覺化結果

let history=[];
async function visualize_loss(size,epochs){
    const data=Array(epochs).fill(0);
    history.map((element)=>{
        data[element.epoch]+=element.loss;
    })
    const avg_loss=data.map((x)=>{
        return x/size;
    })
    const series1 = avg_loss.map((y, x) => ({ x, y, }));
    const plot = { values: [series1] }
    const surface = { name: 'Zoomed Line Chart', tab: 'Visor' };
    tfvis.render.linechart(surface, plot, { zoomToFit: true });
}

再來是訓練過程

async function trainCoreModel(model, data) {
    const batchSize = 16;
    const epochs = 5;

    const inputs = tf.data.array(data.inputs.arraySync());
    const labels = tf.data.array(data.labels.arraySync());
    const size = data.inputs.arraySync().length;
    const dataset = tf.data.zip({
        inputs,
        labels
    }).shuffle(size).batch(batchSize);
    const optimizer = tf.train.adam();
    for (let epoch = 0; epoch <epochs; epoch++) {
        await dataset.forEachAsync(({
            inputs,
            labels
        }) => {
            const batchdata = [inputs.arraySync(), labels.arraySync()];
            let index = 0;
            batchdata[0].map(() => {
                optimizer.minimize(() => {
                    const predYs = model(tf.tensor(batchdata[0][index])).as1D();
                    const labels = tf.tensor(batchdata[1][index])
                    const loss = tf.losses.meanSquaredError(labels, predYs);
                    history.push({epoch:epoch,loss:loss.arraySync()});
                    return loss;
                }, [w1, b1, w2, w3, w4, b4]);
                index++;
            })
        });
    }
    visualize_loss(size,epochs);
}

再把這裡改成

var model;
async function runTensorFlow() {
    // model = createModel();
    // const data = await getData();
    // const tensorData = convertToTensor(data);
    // await trainModel(model, tensorData.inputs, tensorData.labels);
    // console.log('Done Training');
    // const predictedData = getPrediction(model, tensorData);
    // visualiztionPrediction(data, predictedData);
    model = createCoreModel();
    const data = await getData();
    const tensorData = convertToTensor(data);
    await trainCoreModel(model,tensorData );
    // await trainModel(model, tensorData.inputs, tensorData.labels);
    // console.log('Done Training');
    // const predictedData = getPrediction(model, tensorData);
    // visualiztionPrediction(data, predictedData);
}
document.addEventListener('DOMContentLoaded', runTensorFlow);

得到預測結果

function getPrediction(model, normalizationData) {
    const {
        inputMax,
        inputMin,
        labelMin,
        labelMax
    } = normalizationData;
    return tf.tidy(() => {
        //tf.linspace(start_value,end_value,number_of_value);
        const input_x_tensor = tf.linspace(0, 1, 100);
        const input_x = input_x_tensor.arraySync();
        const preds = tf.tensor(input_x.map((x) => {
            return model(tf.tensor(x)).arraySync();
        }))

        //轉回原本的數= 數字*(最大值-最小值)+最小值
        const toOrignalX = input_x_tensor
            .mul(inputMax.sub(inputMin))
            .add(inputMin);
        const toOrignalY = preds
            .mul(labelMax.sub(labelMin))
            .add(labelMin);
        //tensor.dataSync() return data from tensor to array
        return [toOrignalX.dataSync(), toOrignalY.dataSync()];
    });
}

組在一起

var model;
async function runTensorFlow() {
    model = createCoreModel();
    const data = await getData();
    const tensorData = convertToTensor(data);
    await trainCoreModel(model, tensorData);
    const predictedData = getPrediction(model, tensorData);
    visualiztionPrediction(data, predictedData);
}
document.addEventListener('DOMContentLoaded', runTensorFlow);

使用Core API真的超麻煩的
可能是我對TensorFlow.js不太熟
網路上找到的範例也不多
雖然能跑得出答案
但明顯我這種作法肯定有問題
變得很慢
而且很容易當掉

在接下來的幾天，我應該能不用就不用Core API
寫一個範例要花太多時間了