atanasster/hyperparameters

ES6 hyperparameters search for tfjs

ES6 hyperparameters optimization

⚠️ Early version subject to changes.

Features

• written in javascript - Use with tensorflow.js as a replacement to your python hyperparameters library
• use from cdn or npm - Link hpjs in your html file from a cdn, or install in your project with npm
• versatile - Utilize multiple parameters and multiple search algorithms (grid search, random, bayesian)

Installation

$ npm install hyperparameters

Parameter Expressions

import * as hpjs from 'hyperparameters';

hpjs.choice(options)

• Randomly returns one of the options

hpjs.randint(upper)

• Return a random integer in the range [0, upper)

hpjs.uniform(low, high)

• Returns a single value uniformly between low and high i.e. any value between low and high has an equal probability of being selected

hpjs.quniform(low, high, q)

• returns a quantized value of hp.uniform calculated as round(uniform(low, high) / q) * q

hpjs.loguniform(low, high)

• Returns a value exp(uniform(low, high)) so the logarithm of the return value is uniformly distributed.

hpjs.qloguniform(low, high, q)

• Returns a value round(exp(uniform(low, high)) / q) * q

hpjs.normal(mu, sigma)

• Returns a real number that's normally-distributed with mean mu and standard deviation sigma

hpjs.qnormal(mu, sigma, q)

• Returns a value round(normal(mu, sigma) / q) * q

hpjs.lognormal(mu, sigma)

• Returns a value exp(normal(mu, sigma))

hpjs.qlognormal(mu, sigma, q)

• Returns a value round(exp(normal(mu, sigma)) / q) * q

Random numbers generator

import { RandomState } from 'hyperparameters';

example:

const rng = new RandomState(12345);
console.log(rng.randrange(0, 5, 0.5));

Spaces

import { sample } from 'hyperparameters';

example:

import * as hpjs from 'hyperparameters';

const space = {
  x: hpjs.normal(0, 2),
  y: hpjs.uniform(0, 1),
  choice: hpjs.choice([
    undefined, hp.uniform('float', 0, 1),
  ]),
  array: [
    hpjs.normal(0, 2), hpjs.uniform(0, 3), hpjs.choice([false, true]),
  ],
  obj: {
    u: hpjs.uniform(0, 3),
    v: hpjs.uniform(0, 3),
    w: hpjs.uniform(-3, 0)
  }
};

console.log(hpjs.sample.randomSample(space));

fmin - find best value of a function over the arguments

import * as hpjs from 'hyperparameters';
const trials = hpjs.fmin(optimizationFunction, space, estimator, max_estimates, options); 

example:

import * as hpjs from 'hyperparameters';

const fn = x => ((x ** 2) - (x + 1));
const space = hpjs.uniform(-5, 5);
fmin(fn, space, hpjs.search.randomSearch, 1000, { rng: new hpjs.RandomState(123456) })
  .then(trials => console.log(result.argmin));

Getting started with tensorflow.js

1. include javascript file

• include (latest) version from cdn

<script src="https://cdn.jsdelivr.net/npm/hyperparameters@latest/dist/hyperparameters.min.js" />

• create search space

  const space = {
    optimizer: hpjs.choice(['sgd', 'adam', 'adagrad', 'rmsprop']),
    epochs: hpjs.quniform(50, 250, 50),
  };

• create tensorflow.js train function. Parameters are optimizer and epochs. input and output data passed as second argument

const trainModel = async ({ optimizer, epochs }, { xs, ys }) => {
  // Create a simple model.
  const model = tf.sequential();
  model.add(tf.layers.dense({ units: 1, inputShape: [1] }));
  // Prepare the model for training: Specify the loss and the optimizer.
  model.compile({
    loss: 'meanSquaredError',
    optimizer
  });
  // Train the model using the data.
  const h = await model.fit(xs, ys, { epochs });
  return { model, loss: h.history.loss[h.history.loss.length - 1] };
};

• create optimization function

const modelOpt = async ({ optimizer, epochs }, { xs, ys }) => {
  const { loss } = await trainModel({ optimizer, epochs }, { xs, ys });
  return { loss, status: hpjs.STATUS_OK };
};

• find optimal hyperparameters

const trials = await hpjs.fmin(
    modelOpt, space, hpjs.search.randomSearch, 10,
    { rng: new hpjs.RandomState(654321), xs, ys }
  );
  const opt = trials.argmin;
  console.log('best optimizer',opt.optimizer);
  console.log('best no of epochs', opt.epochs);

2. install with npm

• install hyperparameters in your package.json

$ npm install hyperparameters 

• import hyperparameters

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

• create search space

  const space = {
    optimizer: hpjs.choice(['sgd', 'adam', 'adagrad', 'rmsprop']),
    epochs: hpjs.quniform(50, 250, 50),
  };

• create tensorflow.js train function. Parameters are optimizer and epochs. input and output data passed as second argument

const trainModel = async ({ optimizer, epochs }, { xs, ys }) => {
  // Create a simple model.
  const model = tf.sequential();
  model.add(tf.layers.dense({ units: 1, inputShape: [1] }));
  // Prepare the model for training: Specify the loss and the optimizer.
  model.compile({
    loss: 'meanSquaredError',
    optimizer
  });
  // Train the model using the data.
  const h = await model.fit(xs, ys, { epochs });
  return { model, loss: h.history.loss[h.history.loss.length - 1] };
};

• create optimization function

const modelOpt = async ({ optimizer, epochs }, { xs, ys }) => {
  const { loss } = await trainModel({ optimizer, epochs }, { xs, ys });
  return { loss, status: hpjs.STATUS_OK };
};

• find optimal hyperparameters

const trials = await hpjs.fmin(
  modelOpt, space, hpjs.search.randomSearch, 10,
  { rng: new hpjs.RandomState(654321), xs, ys }
);
const opt = trials.argmin;
console.log('best optimizer',opt.optimizer);
console.log('best no of epochs', opt.epochs);

License

MIT © Atanas Stoyanov & Martin Stoyanov