Bookmark this: tool.io/optimization/adagrad-step/source-code
Adagrad Optimizer Step Calculator Source Code
Browse the implementation files for Adagrad Optimizer Step Calculator.tool.ts typescript
// Tool: Adagrad Optimizer Step Calculator
const parseFiniteNumber = (input: unknown): number | null => {
if (
(typeof input !== 'string' && typeof input !== 'number') ||
(typeof input === 'string' && input.trim() === '')
) {
return null;
}
const value = Number(input);
return Number.isFinite(value) ? value : null;
};
const round = (value: number): number =>
Number(value.toPrecision(12));
export const calculateAdagradStep = (
parameterInput: unknown,
gradientInput: unknown,
accumulatedSquaredGradientInput: unknown,
learningRateInput: unknown,
epsilonInput: unknown
): string => {
const parameter = parseFiniteNumber(parameterInput);
const gradient = parseFiniteNumber(gradientInput);
const accumulatedSquaredGradient = parseFiniteNumber(
accumulatedSquaredGradientInput
);
const learningRate = parseFiniteNumber(learningRateInput);
const epsilon = parseFiniteNumber(epsilonInput);
if (
parameter === null ||
gradient === null ||
accumulatedSquaredGradient === null ||
learningRate === null ||
epsilon === null ||
accumulatedSquaredGradient < 0 ||
learningRate <= 0 ||
epsilon <= 0
) {
return '';
}
const nextAccumulatedSquaredGradient =
accumulatedSquaredGradient + gradient ** 2;
const update =
(learningRate * gradient) /
(Math.sqrt(nextAccumulatedSquaredGradient) + epsilon);
const nextParameter = parameter - update;
if (
![
nextAccumulatedSquaredGradient,
nextParameter,
update
].every(Number.isFinite)
) {
return '';
}
return JSON.stringify(
{
accumulatedSquaredGradient: round(
nextAccumulatedSquaredGradient
),
parameter: round(nextParameter),
update: round(update)
},
null,
2
);
};
export const Tool: TypeTool = {
name: 'Adagrad Optimizer Step Calculator',
slug: 'optimization/adagrad-step',
intro:
'Calculate one scalar Adagrad parameter update and its accumulated gradient state.',
tool: calculateAdagradStep,
i: [
{
example: '1',
label: 'Parameter',
placeholder: 'Current parameter value',
type: 'number',
value: 1
},
{
example: '0.1',
label: 'Gradient',
placeholder: 'Current gradient',
type: 'number',
value: 0.1
},
{
example: '0',
label: 'Accumulated Squared Gradient',
placeholder: 'Previous accumulated gradient',
type: 'number',
value: 0
},
{
example: '0.01',
label: 'Learning Rate',
placeholder: 'Positive learning rate',
type: 'number',
value: 0.01
},
{
example: '0.00000001',
label: 'Epsilon',
placeholder: 'Positive stability constant',
type: 'number',
value: 0.00000001
}
],
o: [
{
example: '{ "parameter": 0.99, "update": 0.01 }',
label: 'Adagrad Step',
placeholder: 'Updated optimizer state will appear here...',
render: 'textarea',
type: 'text'
}
],
cite: [
{
title:
'Adaptive Subgradient Methods for Online Learning and Stochastic Optimization',
url: 'https://jmlr.org/papers/v12/duchi11a.html'
}
]
};
export default Tool;
Description
Browse the implementation files for Adagrad Optimizer Step Calculator.