TensorFlow.js：第1部分：使用低级API近似线性函数

import * as tf from '@tensorflow/tfjs';

export default class AbstractRegressionModel {
    
    constructor(
        width,
        height,
        optimizerFunction = tf.train.sgd,
        maxEpochPerTrainSession = 100,
        learningRate = 0.1,
        expectedLoss = 0.001
    ) {
        this.width = width;
        this.height = height;
        this.optimizerFunction = optimizerFunction;
        this.expectedLoss = expectedLoss;
        this.learningRate = learningRate;
        this.maxEpochPerTrainSession = maxEpochPerTrainSession;

        this.initModelVariables();

        this.trainSession = 0;
        this.epochNumber = 0;
        this.history = [];
    }

    initModelVariables() {
        throw Error('Model variables should be defined')
    }

    f() {
        throw Error('Model should be defined')
    }

    xNormalization = xs => xs.map(x => x / this.width);
    yNormalization = ys => ys.map(y => y / this.height);
    yDenormalization = ys => ys.map(y => y * this.height);

    /**
     * Calculate loss function as mean-squared deviation
     *
     * @param predictedValue - tensor1d - predicted values of calculated model
     * @param realValue - tensor1d - real value of experimental points
     */
    loss = (predictedValue, realValue) => {
        const loss = predictedValue.sub(realValue).square().mean();
        this.lossVal = loss.dataSync()[0];
        return loss;
    };

    /**
     * Train model until explicitly stop process via invocation of stop method
     * or loss achieve necessary accuracy, or train achieve max epoch value
     *
     * @param x - array of x coordinates
     * @param y - array of y coordinates
     * @param callback - optional, invoked after each training step
     */
    async train(x, y, callback) {
        const currentTrainSession = ++this.trainSession;
        this.lossVal = Number.POSITIVE_INFINITY;
        this.epochNumber = 0;
        this.history = [];

        // convert data into tensors
        const input = tf.tensor1d(this.xNormalization(x));
        const output = tf.tensor1d(this.yNormalization(y));

        while (
            currentTrainSession === this.trainSession
            && this.lossVal > this.expectedLoss
            && this.epochNumber <= this.maxEpochPerTrainSession
            ) {
            const optimizer = this.optimizerFunction(this.learningRate);
            optimizer.minimize(() => this.loss(this.f(input), output));
            this.history = [...this.history, {
                epoch: this.epochNumber,
                loss: this.lossVal
            }];
            callback && callback();
            this.epochNumber++;
            await tf.nextFrame();
        }
    }

    stop() {
        this.trainSession++;
    }

    /**
     * Predict value basing on trained model
     *  @param x - array of x coordinates
     *  @return Array({x: integer, y: integer}) - predicted values associated with input
     *
     * */
    predict(x) {
        const input = tf.tensor1d(this.xNormalization(x));
        const output = this.yDenormalization(this.f(input).arraySync());
        return output.map((y, i) => ({ x: x[i], y }));
    }
}

import * as tf from '@tensorflow/tfjs';
import AbstractRegressionModel from "./AbstractRegressionModel";

export default class LinearRegressionModel extends AbstractRegressionModel {
    initModelVariables() {
        this.k = tf.scalar(Math.random()).variable();
        this.b = tf.scalar(Math.random()).variable();
    }

    f = x => x.mul(this.k).add(this.b);
}

import * as tf from '@tensorflow/tfjs';
import AbstractRegressionModel from "./AbstractRegressionModel";

export default class QuadraticRegressionModel extends AbstractRegressionModel {
    initModelVariables() {
        this.a = tf.scalar(Math.random()).variable();
        this.b = tf.scalar(Math.random()).variable();
        this.c = tf.scalar(Math.random()).variable();
    }

    f = x => this.a.mul(x.square()).add(this.b.mul(x)).add(this.c);
}

import React, { useState, useEffect } from 'react';

import Canvas from './components/Canvas';
import LossPlot from './components/LossPlot_v3';
import LinearRegressionModel from './model/LinearRegressionModel';

import './RegressionModel.scss';

const WIDTH = 400;
const HEIGHT = 400;
const LINE_POINT_STEP = 5;

const predictedInput = Array.from({ length: WIDTH / LINE_POINT_STEP + 1 })
    .map((v, i) => i * LINE_POINT_STEP);

const model = new LinearRegressionModel(WIDTH, HEIGHT);

export default () => {
    const [points, changePoints] = useState([]);
    const [curvePoints, changeCurvePoints] = useState([]);
    const [lossHistory, changeLossHistory] = useState([]);

    useEffect(() => {
        if (points.length > 0) {
            const input = points.map(({ x }) => x);
            const output = points.map(({ y }) => y);

            model.train(input, output, () => {
                changeCurvePoints(() => model.predict(predictedInput));
                changeLossHistory(() => model.history);
            });
        }
    }, [points]);

    return (
        <div className="regression-low-level">
            <div className="regression-low-level__top">
                <div className="regression-low-level__workarea">
                    <div className="regression-low-level__canvas">
                        <Canvas
                            width={WIDTH}
                            height={HEIGHT}
                            points={points}
                            curvePoints={curvePoints}
                            changePoints={changePoints}
                        />
                    </div>
                    <div className="regression-low-level__toolbar">
                        <button
                            className="btn btn-red"
                            onClick={() => model.stop()}>Stop
                        </button>
                        <button
                            className="btn btn-yellow"
                            onClick={() => {
                                model.stop();
                                changePoints(() => []);
                                changeCurvePoints(() => []);
                            }}>Clear
                        </button>
                    </div>
                </div>
                <div className="regression-low-level__loss">
                    <LossPlot
                              loss={lossHistory}/>
                </div>

            </div>
        </div>
    )
}