Neoclassic Placebo Maker
Join us to discuss the challenges, solutions and best practices for in-house JavaScript code sharing. Tuesday, 12/17 at 10am PT/1pm ET.Sign up here »

@tartarus/deep

0.1.0 • Public • Published

Tartarus Deep Learning Framework

Deep learning framework for TypeScript. Run it on a browser, on AWS Lambda, or on anything that runs Node.js!

Travis CI Coverage Status Codacy Maintainability David David

Features

From-the-ground-up implementation for:

  • Math: Vector and Matrix operations, seeded randomization
  • Graph: Acyclic networks, automated data routing, support for multiple input and output layers
  • Machine Learning: Forward and back propagation, logistic regression, gradient descent, loss (cost) functions, activation functions, optimizers, metrics, dense layers, concat layers

Goals

  • From-the-ground-up implementation for all standard deep learning operations
  • Compatibility with Node.js, modern browsers, and AWS Lambda

Example

import { Model, Dense, MemoryInputFeed } from '@tartarus/deep';
 
/*
 * 1. Define a model 
 *    - 4 input nodes
 *    - hidden layer with 5 nodes and sigmoid activation
 *    - output layer with 3 nodes and softmax activation 
 */
const model = new Model({ optimizer: 'stochastic', loss: 'mean-squared-error' });
 
model
  .input(4)
  .push(new Dense({ units: 5, activation: 'sigmoid' }))
  .push(new Dense({ units: 3, activation: 'softmax' }));
 
model.compile()
  .then(
    async () => {
      /* 2. Prepare three samples of training data */
      const feed = new MemoryInputFeed();
      
      feed
        .add([1, 2, 3, 4], [1, 0, 0]) // .add(input, expected output)
        .add([4, 3, 2, 1], [0, 1, 0])
        .add([5, 6, 7, 8], [0, 0, 1]);
      
      
      /* 3. Train model */
      await model.fit(feed, { batchSize: 1, epochs: 100 });
      
      
      /* 4. Predict */
      const result = await model.predict([8, 9, 10, 11]);
      
      console.log(`Prediction: ${result.getDefaultValue().toJSON()}`);
    }
  );

Model

Layers

Dense Layer

Concat Layer

Supplementary

Activation Functions

Loss Functions

Initializers

Metrics

Input

Install

npm i @tartarus/deep

DownloadsWeekly Downloads

1

Version

0.1.0

License

MIT

Unpacked Size

615 kB

Total Files

435

Last publish

Collaborators

  • avatar