对神经网络的可能性印象深刻,我决定在使用任何库之前,我都想了解它们的工作原理。因此,我编写了一个简单的培训应用,该应用使用了3层网络,每个网络有2个神经元。有一张400x400的画布。给定鼠标在画布<0; 399>上的x,y坐标,结果应为坐标/ 400 <0; 1>(因此对于100,300,应给出0.25,0.75)。
培训看起来很合理。
但是当我切换到预测模式时,网络每次训练批次都会一直得到相同的结果。无论输入是什么,它都会给出相同的结果。 然后,经过更多的训练,输出会发生变化,但每个输入仍然相同。
它是用TypeScript编写的。 我没有粘贴整个Web培训页面,而是制作了培训脚本,以便您可以更清楚地了解正在发生的事情。
let sigmoid: ActivationFunction = {
func: (x: number) => (1 / (1 + Math.exp(-x))),
derivative: (z: number) => {
return sigmoid.func(z) * (1 - sigmoid.func(z));
}
};
import Matrix from './matrix';
class NeutralNetwork {
layers: Array<number>;
weights: Matrix[];
biases: Matrix[];
activation_function: ActivationFunction;
learning_rate: number;
constructor(...layers: Array<number>) {
this.layers = layers;
this.activation_function = sigmoid;
//Initialize neural network with random weigths and biases [-1;1]
this.weights = [];
for(let i=0; i<this.layers.length - 1; i++){
this.weights.push(new Matrix(this.layers[i+1], this.layers[i]));
this.weights[i].randomize();
}
this.biases = [];
for(let i=1; i<this.layers.length; i++){
this.biases.push(new Matrix(this.layers[i], 1));
this.biases[i-1].randomize();
}
this.setActivationFunction();
this.setLearningRate();
}
feedForward(originalInput: Array<number>): Array<number> {
if(originalInput.length != this.layers[0]) throw new Error("corrupt input data");
let input : Matrix = Matrix.createFromArray(originalInput);
for(let i = 0; i < this.layers.length - 1; i++){
let output = Matrix.multiply(this.weights[i], input);
output.add(this.biases[i]);
output.map(this.activation_function.func);
input = output;
}
return input.toArray();
}
train(originalInput: Array<number>, originalTarget: Array<number>) {
if(originalInput.length != this.layers[0]) throw new Error("corrupt training data");
if(originalTarget.length != this.layers[this.layers.length - 1]) throw new Error("corrupt training data");
let outputs : Matrix[] = [];
let input : Matrix = Matrix.createFromArray(originalInput);
for(let i = 0; i < this.layers.length - 1; i++){
let output = Matrix.multiply(this.weights[i], input);
output.add(this.biases[i]);
output.map(this.activation_function.func);
input = output;
outputs.push(output);
}
let target = Matrix.createFromArray(originalTarget);
let errors = Matrix.subtract(target, outputs[this.layers.length - 2]);
for(let i = this.layers.length - 2; i>=0; i--){
let gradients = Matrix.map(outputs[i], this.activation_function.derivative);
gradients.multiply(errors);
gradients.multiply(this.learning_rate);
let outputsOfLayerBeforeTransposed = Matrix.transpose(i > 0 ? outputs[i-1] : Matrix.createFromArray(originalInput));
let deltas = Matrix.multiply(gradients, outputsOfLayerBeforeTransposed);
this.weights[i].add(deltas);
this.biases[i].add(gradients);
let weightsTransposed = Matrix.transpose(this.weights[i]);
errors = Matrix.multiply(weightsTransposed, errors);
}
return outputs[outputs.length - 1].toArray();
}
setActivationFunction(activationFunction = sigmoid) {
this.activation_function = activationFunction;
}
setLearningRate(learning_rate = 0.1) {
this.learning_rate = learning_rate;
}
}
interface ActivationFunction {
func(x: number): number;
derivative(x: number): number;
}
export = NeutralNetwork;
let NN = require('./index');
let n = new NN(2,2,2);
let data = generateTrainingData();
data.forEach(d => n.train(d.i, d.o));
//check how well is it trained
let index = 0
let t = setInterval(()=>{
let pred = n.feedForward(data[index].i);
console.log(`PREDICTED - ${pred} EXPECTED = ${data[index].o} COST - ${Math.pow(pred[0]-data[index].o[0],2)+Math.pow(pred[1]-data[index].o[1],2)}`)
if(index++ == 1000) clearInterval(t);
}, 500);
function generateTrainingData(){
let data = [];
for(let i=0;i<1000;i++){
let x = Math.floor(Math.random() * 400);
let y = Math.floor(Math.random() * 400);
data.push({
i : [x,y],
o : [x/400, y/400]
})
}
return data;
}
export default class Matrix {
rows;
columns;
data: Array<Array<number>>;
constructor(rows, columns) {
this.rows = rows;
this.columns = columns;
this.data = new Array(this.rows).fill().map(() => Array(this.columns).fill(0));
}
static map(matrix, f) : Matrix{
let m = new Matrix(matrix.rows, matrix.columns);
m.map((v,i,j) => f(matrix.data[i][j], i, j));
return m;
}
map(f) {
for (let i = 0; i < this.rows; i++) {
for (let j = 0; j < this.columns; j++) {
this.data[i][j] = f(this.data[i][j], i, j);
}
}
}
randomize() {
this.map(() => Math.random() * 2 - 1);
}
add(n) {
if (n instanceof Matrix) {
if (this.rows !== n.rows || this.columns !== n.columns) {
throw new Error('Size of both matrices must match!');
}
return this.map((v, i, j) => v + n.data[i][j]);
} else {
return this.map(v => v + n);
}
}
static subtract(a, b) : Matrix{
if (a.rows !== b.rows || a.columns !== b.columns) {
throw new Error('Size of both matrices must match!');
}
let m = new Matrix(a.rows, a.columns);
m.map((_, i, j) => a.data[i][j] - b.data[i][j]);
return m;
}
static multiply(a, b) {
if (a.columns !== b.rows) {
throw new Error('a.columns !== b.rows');
}
let m = new Matrix(a.rows, b.columns)
m.map((_, i, j) => {
let sum = 0;
for (let k = 0; k < a.cols; k++) {
sum += a.data[i][k] * b.data[k][j];
}
return sum;
});
return m;
}
multiply(n) {
if (n instanceof Matrix) {
if (this.rows !== n.rows || this.columns !== n.columns) {
throw new Error('Size of both matrices must match!');
}
return this.map((v, i, j) => v * n.data[i][j]);
} else {
return this.map(v => v * n);
}
}
toArray() {
let arr = [];
for (let i = 0; i < this.rows; i++) {
for (let j = 0; j < this.columns; j++) {
arr.push(this.data[i][j]);
}
}
return arr;
}
static transpose(matrix) : Matrix {
let m = new Matrix(matrix.columns, matrix.rows)
m.map((_, i, j) => matrix.data[j][i]);
return m;
}
static createFromArray(arr): Matrix {
let m = new Matrix(arr.length, 1);
m.map((v, i) => arr[i]);
return m;
}
}
我不太确定是什么原因造成的。我已经尝试调试了几天,但我认为我经验不足并不能让我在这里看到问题。非常感谢您的所有帮助。
答案 0 :(得分:1)
Matrix.multiply
类方法中有一个错误。它应该是a.columns
而不是a.cols
。因此,gradients
和deltas
无法正确更新。