Learning rate scaler callback

Learning rate scaler callback for @tensorflow/tfjs

Installation

Package available via:

• NPM npm install learning-rate-scaler

Usage

Module exports LearningRateCallback and example function sineWaveRateScaler = (min, max, frequency, maxLearningRateDecay, earningRateDecayPerEpoch) to use with it. (also a default LearningRateCallback scaler function).

• example 1

LearningRateCallback with defaults

import {
  LearningRateCallback,
  sineWaveRateScaler,
} from "learning-rate-scaler";

const learningRateCallback = new LearningRateCallback()

const inputs = input({
  shape: [1],
});
const y1 = dense({ units: 1, name: "y1" }).apply(inputs);
let layersModel = model({
  inputs: inputs,
  outputs: [y1],
});
layersModel.compile({
    optimizer: train.sgd(0.001),
    loss: {
      y1: "meanAbsoluteError",
    },
});
const history = await layersModel.fitDataset(trainingDataset, {
  epochs: 1,
  callbacks: [learningRateCallback],
});

This would produce a sine-wave-looking learning-rate curve for each sample in the batch (for batch size of 2048 items with default settings).

• example 2

const frequency = 1 / 25; // a fraction - fraction lim -> 0  produces wider periods.
const maxLearningRateDecay = 1.0001 // a number slightly higher than 1
const sineWaveRateScaler = sineWaveRateScaler(0.0001, 1, frequency, )
const learningRateCallback = new LearningRateCallback({scalerFunction: sineWaveRateScaler})
[...]
const history = await layersModel.fitDataset(trainingDataset, {
  epochs: 1,
  callbacks: [learningRateCallback],
});

This would produce a sine-wave-looking learning-rate curve with decaying of the initial max learning rate for each sample in the batch (for batch size of 2048 items with default settings).

Also it's possible to decay epoch-wise also by providing

const sineWaveRateScaler = sineWaveRateScaler(0.0001, 1, frequency, 1.0001, 1.0001)

Credits

Prepared by Webplan.pro