parazit-ir / deepspark_java

Java based Convolutional Neural Network package running on Apache Spark framework

DeepSpark_java

Java Convolutional Neural Network (CNN) package combined with Apache Spark framework

DeepSpark_java is an early version of ongoing DeepSpark project (https://github.com/deepspark/deepspark) implemented in pure java and jBlas. It provides GPU Acceleration using jCublas. (gpuAccel option)

DeepSpark_java also supports local training running on single machine and distributed (sync & async) training aided by Apache Spark (http://spark.apache.org/)

USAGE INSTRUCTIONS

Data format

Class	Description
Tensor	Base class for Tensor. Implemented using jBlas
Weight	Class for representing Network parameters
Sample	Class for representing Data container

Weight and Sample class are implemented using Tensor.

To load custom dataset, users should create own data loader to be compatible with Sample .

We provide built-in Mnist/CIFAR/ImageNet loader (See examples on src/main/java/org/acl/deepspark/utils/)

Layer format

Layer	Description
Layer	Base interface for layers
BaseLayer	Abstract class implementing Layer interface
ConvolutionLayer	Convolutional layer
PoolingLayer	Pooling (subsampling) layer
FullyConnectedLayer	Normal fully connected layer

Users should define LayerConf to specify layer details (LayerType, kernel width/height, stride, padding etc.)

To add more options, check on src/main/java/org/acl/deepspark/nn/conf/LayerConf

Training

Layer	Description
NeuralNet	Class for representing overall Network. Provides methods for initializing, training and inference
NeuralNetRunner	Runner of NeuralNet on local machine
DistNeuralNetRunner	Runner of NeuralNet in synchronous distributed setting
DistAsyncNeuralNetRunner	Runner of NeuralNet in asynchronous distributed setting

Users should define NeuralNetConf to specify training details (lr, l2_lambda, momentum, gpuAccel etc.)

To add more options, check on src/main/java/org/acl/deepspark/nn/conf/NeuralNetConf

For asynchronous update, simple ParameterServer/Client class are implemented. Check on src/main/java/org/acl/deepspark/nn/async

Examples

For actual usage code, see examples on src/test/java/org/acl/deepspark/nn/driver

Type	Path
Single Machine	MnistTest.java / CIFARTest.java
Distributed (sync)	SyncMnistTest.java
Distributed (async)	AsyncMnistTest.java

Publications

Kim, Hanjoo, Jaehong Park, Jaehee Jang, and Sungroh Yoon. "DeepSpark: Spark-Based Deep Learning Supporting Asynchronous Updates and Caffe Compatibility." arXiv preprint arXiv:1602.08191 (2016).