Gleethos / neureka

A platform independent tensor library with autograd for the JVM
https://gleethos.github.io/neureka/index.html
MIT License
69 stars 1 forks source link
deep-neural-networks java library

NEUREKA

A lightweight
platform independent
tensor library for the JVM

OpenCL accelerated nd-arrays / tensors for Java, Kotlin, Groovy, Scala, Jython, JRuby...

Current Build Code Coverage Version Code Quality Licence Size
Build Status Codacy Badge GitHub version Java Version Codacy Badge License: MIT Size

:hammer_and_wrench: Features

- [tensors & nd-arrays](docs/markdown/tensors_or_nd-arrays.md) - convolution - broadcasting - dynamic computation graph - autograd - flexible indexing and slicing - seeding - labeling - jpg, png, idx support - [highly extensible backend](docs/markdown/extending_neureka.md)

Take a quick look :eyes:

Impress me! Show me more Documentation
- [ANN in 11 Lines of Code!](docs/markdown/impressive.md) - [Neureka with Java](docs/markdown/java_example.md) :coffee: - [Neureka with Groovy](docs/markdown/groovy_example.md) :star: - [living
documentation](https://gleethos.github.io/neureka/showcase.html) - [Javadocs](https://gleethos.github.io/neureka/jdocs/index.html) :book:

:robot: Tech

Dynamic Autograd : Recording the Computation-Graph

Neureka trains your neural network using a computation graph recorder.

This is contrary to the approaches found in other frameworks such as TensorFlow, Theano, Caffe, and CNTK which require the definition of a computation graph ahead of time. This means a developer has to build a neural network structure which cannot change during runtime.

Neureka, uses the recorded computation graph in order to apply a technique called reverse-mode auto-differentiation, which allows your network structure to change during runtime arbitrarily with zero lag or overhead.
This powerful feature has been inspired by PyTorch:

Main-Package Overview

Package Description
neureka the root package containing the tensor class and the following sub-packages
neureka.devices a sub-package which enables cross platform acceleration (OpenCLDevice) and tensor persistence (FileDevice)
neureka.math a sub-package containing collections of functions and the ability to create custom ones
neureka.optimization a sub-package for weight-gradient optimization
neureka.autograd the guts of Neurekas autograd system
neureka.backend the backend containing both a consistent API and a standard implementation

Getting started with Apache Maven

<dependency>
  <groupId>com.github.gleethos</groupId>
  <artifactId>neureka</artifactId>
  <version>1.0.0</version>
</dependency>

Getting started with Gradle

Groovy DSL:

implementation 'com.github.gleethos:neureka:1.0.0'

Kotlin DSL:

implementation("com.github.gleethos:neureka:1.0.0")

Getting started with

1. Add the JitPack url in your root build.gradle at the end of repositories

allprojects {
    repositories {
        //...
        maven { url 'https://jitpack.io' }
    }
}

2. Add Neureka as dependency

...either by specifiying the version tag:

dependencies {
    implementation 'com.github.Gleethos:neureka:v1.0.0'
}

...or by using a custom commit hash instead:

dependencies {
    implementation 'com.github.Gleethos:neureka:8485bca'//Any commit hash...
}

Getting started with Groovy Grape

@GrabResolver(name = 'jitpack.io', root = 'https://jitpack.io')
@Grab('com.github.Gleethos:neureka:v1.0.0')

import neureka.*

:rocket: Building from source

Execute the following:

 gradlew build

Tests:

 gradlew check

Jar file:

 gradlew jar

Min-jar file:

 gradlew proguard

:mount_fuji: Dependencies


:book: Documentation


:microscope: Tests & Specs :scroll:


:seedling: Development [Commit activity 1 year]() Maintenance GitHub commits

Want to contribute? Don't worry:

There is plenty of developer friendly highly readable living documentation to go through which explains the inner and outer workings of this project very well!

If you want to dive right into it, start off by extending the backend for additional types of operations or data type support.

Any feedback or contribution, even as simple as a typo fix, is always greatly appreciated!


:memo: Todos - Issues


:balance_scale: License

Open Source Love


Tweet