Introduction

MyDL (simply stands for "my deep learning") is a simple deep learning libary base on java. Its matrix operations are based on EJML(Efficient Java Matrix Library) and its part of APIs design learns from Keras.

Features

  • 100% Java. It is written in Java without any native methods and its dependency, EJML, is also fully written in Java. This provides a completely cross-platform compatibility. No matter what platform, as long as it support Java, you can use MyDL on it.
  • New implementation for tensor. We build a set of tensor classes and operations based on EJML to establish a data type friendly to deep learning on Java.
  • A lightweight framework. It is easy to deploy, all you need is a single JAR.
  • Minimalist design. Components of this framework are easy and clear. Many API designs take reference from Keras, making it easy for users to get started.

Compents implemented

  • Model: abstract model, sequential
  • Layer: abstract layer
    • Activation: abstract activation, ReLU, Sigmoid, Tanh, Softmax
    • Fully-connected: Dense(for high dimension), Linear1D
    • Reshape layer
  • Loss: abstract loss, SSE, MSE, binary cross-entropy, categorical cross-entropy
  • Optimizer: abstract optimizer, mini-batch GD/SGD
  • Tensor: abstract Tensor, Tensor1D, Tensor2D, Tensor3D
    • Tensor_size(record the shape of tensor)
  • Dataset: APIs to load MNIST dataset(included in JAR)
  • Utils: Some unclassified tools
    • Data(class for loading data into models, easy to convert)
    • IsAttribute(Determin if the object has such attribute)

Acknowlegde

  1. We wrote the initial version of this libary as a course project under the guidance of Prof. Bing Xiang in JAVA2760, School of Maths Sciences, NanKai University.
  2. Special thanks to SaucerHi who gave us inspiration on some key points in project desgin.
  3. Part of APIs refer to joelnet and Keras.

License

MyDL is is GPL-3.0 licensed.

Contribution

Due to limited time, this framework is actually far from complete, so the current release may be unstable. If you meet any problems, please feel free to raise an issue or contact us and we will fix it as soon as possible.

The initial contributors are undergraduates in Nankai Univ and honesty their skills and time is limited, so we warmly welcome anyone who wants to contribute to this project. You can:

  • Raise an issue of suggestion for improvement.
  • Create a pull request and provide description if you implement something new for this project.
  • Contact Alexhaoge and become a collaborator of this repository for long term contribution.

Video of Overview