auto-encoder.ts

auto-encoder.ts

A simple Auto Encoder typescript library for experimentation and dimensionality reduction. Supports automatic scaling.

This is a Typescript wrapper on top of autoencoder.

With additional helper functions: exportAutoEncoder(autoEncoder) and restoreAutoEncoder(json).

Features

• Build embedding model in self-supervised manner (without manually labelling data)
• Reduce data dimension based on data distribution
• Support export/restore with JSON
• Lightweight (without node-gpy, cmake, python, cuda)
• Automatic scaling/normalizing (can be turned off)
• Static Type Checking and Completion with Typescript
• Isomorphic package: works in Node.js and browsers
• Works with plain Javascript, Typescript is not mandatory

Installation

npm install auto-encoder.ts

You can also install auto-encoder.ts with pnpm, yarn, or slnpm

Usage Example

Create new Auto Encoder

There are two ways to initialize a model:

• Provide the number of layers, input size, encoder output size (number of latent variables) and the activation function name

import { createAutoEncoder } from 'auto-encoder.ts'

const model = createAutoEncoder({
  nInputs: 10,
  nHidden: 2,
  nLayers: 2, // (default 2) - number of layers in each encoder/decoder
  activation: 'relu', // (default 'relu') - applied to all, but the last layer
})

• Define each layer separately for both encoder and decoder

import { createAutoEncoder } from 'auto-encoder.ts'

const model = createAutoEncoder({
  encoder: [
    { nOut: 10, activation: 'tanh' },
    { nOut: 2, activation: 'tanh' },
  ],
  decoder: [{ nOut: 2, activation: 'tanh' }, { nOut: 10 }],
  scale: false, // (default true)
})

Activation functions: relu, tanh, sigmoid

Auto Scaling

Similar to other neural networks, auto-encoder is very sensitive to input scaling.

To make it easier the scaling is enabled by default.

you can control it with an extra parameter scale that can be true or false.

Train the Auto Encoder model

model.fit(X, {
  batchSize: 100,
  iterations: 5000,
  method: 'adagrad', // (default 'adagrad')
  stepSize: 0.01, // (default 0.05)
})

Optimization methods: sgd, adagrad, adam

Encode, Decode, Predict

const Y = model.encode(X)
const Xd = model.decode(Y)

// Similar to model.decode(model.encode(X))
const Xp = model.predict(X)

Web demo (dimensionality reduction)

Try the package in the browser on StatSim Vis. Choose a CSV file, change the Projection method to Autoencoder, then click Run.

Typescript Signature

Below are the exported function and types:

import { ActivationFunctionName, OptimizationMethodName } from 'adnn.ts'

function createAutoEncoder(options: AutoEncoderOptions): AutoEncoder

function exportAutoEncoder(autoEncoder: AutoEncoder): AutoEncoderJSON

function restoreAutoEncoder(json: AutoEncoderJSON): AutoEncoder

interface AutoEncoder {
  fit(X: BatchValues, options?: FitOptions): void
  encode(X: BatchValues): BatchValues
  decode(X: BatchValues): BatchValues
  /** @description Similar to this.decode(this.encode(X)) */
  predict(X: BatchValues): BatchValues
}

type AutoEncoderJSON = {
  scale: boolean | undefined
  max: number[]
  min: number[]
  nInputs: number
  nHidden: number
  encoder: unknown
  decoder: unknown
}

type FitOptions = {
  /** @default round(totalSize/50) */
  batchSize?: number
  /** @default 100 */
  iterations?: number
  /** @default 'adagrad' */
  method?: OptimizationMethodName
  /** @default 0.05 */
  stepSize?: number
}

type BatchValues = Values[]

type Values = number[]

type AutoEncoderOptions =
  | {
      /** @default true */
      scale?: boolean
      /** @description number of input features */
      nInputs: number
      /** @description number of embedding features */
      nHidden: number
      /**
       * @description number of layers in each encoder/decoder
       * @default 2
       */
      nLayers?: number
      /**
       * @description applied to all, but the last layer
       * @default 'relu'
       */
      activation?: ActivationFunctionName
    }
  | {
      /** @default true */
      scale?: boolean
      encoder: LayerOptions[]
      decoder: LayerOptions[]
    }

type LayerOptions = {
  nOut: number
  /** @description no activation function in the last layer of decoder gives better result */
  activation?: ActivationFunctionName
}

License

This project is licensed with BSD-2-Clause

This is free, libre, and open-source software. It comes down to four essential freedoms [ref]:

• The freedom to run the program as you wish, for any purpose
• The freedom to study how the program works, and change it so it does your computing as you wish
• The freedom to redistribute copies so you can help others
• The freedom to distribute copies of your modified versions to others