infers

Machine learning and Matrix operation library by TypeScript.

• XOR EXAMPLE
• API DOC

Installed

Make sure NPM is installed, Switch to the project directory then execute the following command.

$ npm install infers@latest

Reference in project:

import { Matrix, BPNet } from 'infers'

Examples

Matrix transpose:

let m = new Matrix([
  [1, 5, 0],
  [2, 4 , -1],
  [0, -2, 0]
])
m.T.print()
// Matrix 3x3 [
//  1, 2, 0, 
//  5, 4, -2, 
//  0, -1, 0, 
// ]

BP neural network example of XOR, three-layer network:

let xs = new Matrix([[1, 0], [0, 1], [0, 0], [1, 1]])
let ys = new Matrix([[1], [1], [0], [0]])
let model = new BPNet([2, [6, 'Tanh'], [1, 'Sigmoid']], { rate: 0.1 })
model.fit(xs, ys, {
  epochs: 5000, onEpoch: (epoch, loss) => {
    if (epoch % 100 === 0) console.log('epoch:' + epoch, 'loss:', loss)
  }
})
model.predict(xs).print()
// Matrix 4x1 [
//  0.9862025352830867, 
//  0.986128496195502, 
//  0.01443800549676924, 
//  0.014425871504885788, 
// ]

BP neural network example of addition, four-layer network:

let xs = new Matrix([[1, 4], [3, 2], [6, 5], [4, 7]])
let ys = new Matrix([[5], [5], [11], [11]])
let model = new BPNet([2, 6, 6, 1], { mode: 'bgd', rate: 0.01 })
model.fit(xs, ys, {
  epochs: 500, onEpoch: (epoch, loss) => {
    console.log('epoch:' + epoch, 'loss:', loss)
  }
})
let xs2 = new Matrix([[5, 8], [22, 6], [-5, 9], [-5, -4]])
model.predict(xs2).print()
// Matrix 2x1 [
//  12.994745740521667, 
//  27.99134620596921, 
//  3.9987224114576856, 
//  -9.000000644547901,
// ]

RNN: Recurrent neural network example:

let trainData = ['hello rnn', 'good morning', 'I love 🍎!', 'I eat 🍊!']
let net = new RNN({ trainData })
net.fit({
  epochs: 1500, onEpochs: (epoch, loss) => {
    if (epoch % 10 === 0) console.log('epoch: ', epoch, 'loss: ', loss)
  }
})
console.log(net.predict('I love'))
console.log(net.predict('I eat'))
console.log(net.predict('hel'))
console.log(net.predict('good'))
//  🍊!/n
//  🍎!/n
// lo rnn/n
//  morning/n

API:

• NetShape: [number, (number | [number, ActivationFunction]), ...(number | [number, ActivationFunction])[]]
  The hierarchical structure of the network model, It includes the number of neurons in each layer, the type of activation function and the total number of layers.
• rate: number
  The learning rate is the update step of every gradient descent, generally between 0 and 1.
• epochs: number
  All the data of the whole training set are iterated once.
• ActivationFunction: 'Sigmoid' | 'Relu' | 'Tanh' | 'Softmax'
• Mode: 'sgd' | 'bgd' | 'mbgd'

Different learning rates, iterations and network shapes are needed to deal with different problems, which need to be adjusted according to the cost function. Parameter optimization is also the process of model optimization.

Export

• class Matrix
  • Mathematical operation of matrix
  • addition, multiply, transpose, determinant, inverse
• class BPNet
  • Fully connected neural network
  • Multi-layer network model
• class RNN
  • Recurrent neural network
  • Used natural language processing