Learn about our RFC process, Open RFC meetings & more.Join in the discussion! »

autoencoder

0.0.2 • Public • Published

Autoencoder

A simple autoencoder module for experimentation and dimensionality reduction. Supports automatic scaling

Autoencoder

Install and load

Install from npm:

npm install autoencoder

Then load with require:

const Autoencoder = require('autoencoder')

Create new autoencoder

Autoencoder supports two ways of a model initialization:

  • Provide the number of layers, input size, encoder output size (number of latent variables) and the activation function name
const ae = new Autoencoder({
  '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
const ae = new Autoencoder({
  'encoder': [
    {'nOut': 10, 'activation': 'tanh'},
    {'nOut': 2, 'activation': 'tanh'}
  ],
  'decoder': [
    {'nOut': 2, 'activation': 'tanh'},
    {'nOut': 10}
  ]
})

Activation functions: relu, tanh, sigmoid

As other neural nets, autoencoder 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 takes true or false

Train autoencoder

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

Optimization methods: sgd, adagrad, adam

Encode, Decode, Predict

const Y = ae.encode(X)
const Xd = ae.decode(Y)
const Xp = ae.predict(X) // Similar to ae.decode(ae.encode(X))

Install

npm i autoencoder

DownloadsWeekly Downloads

4

Version

0.0.2

License

MIT

Unpacked Size

43.6 kB

Total Files

6

Last publish

Collaborators

  • avatar