bgnkim / scalanetwork   1.3.0

GNU General Public License v2.0 only GitHub

A Neural network implementation with Scala

Scala versions: 2.11 2.10

ScalaNetwork 1.3.0

Currently, we don't maintain this repository any more. The code had not designed well, hence we re-designed the entire logic. For the latest library, check DeepSpark.

A Neural Network implementation with Scala, Breeze & Spark

Spark Network follows GPL v2 license.

Features

Network

ScalaNetwork supports following layered neural network implementation:

  • Fully-connected Neural Network : f(Wx + b)
  • Fully-connected Rank-3 Tensor Network : f(v1TQ[1:k]v2 + L[1:k]v + b)
  • Fully-connected Auto Encoder
  • Fully-connected Stacked Auto Encoder

Also you can implement following Recursive Network via training tools.

  • Traditional Recursive Auto Encoder (RAE)
  • Standard Recursive Auto Encoder (RAE)
  • Unfolding Recursive Auto Encoder (RAE) [EXPERIMENTAL]

Training Methodology

ScalaNetwork supports following training methodologies:

  • Stochastic Gradient Descent w/ L1-, L2-regularization, Momentum.
  • AdaGrad
  • AdaDelta

ScalaNetwork supports following environments:

  • Single-Threaded Training Environment.
  • Spark-based Distributed Environment, with modified version of Downpour SGD in DistBelief

Also you can add negative examples with Trainer.setNegativeSampler().

Activation Function

ScalaNetwork supports following activation functions:

  • Linear
  • Sigmoid
  • HyperbolicTangent
  • Rectifier
  • Softplus
  • HardSigmoid
  • HardTanh
  • Softmax

And also you can make new activation function using several operations.

Usage

Here is some examples for basic usage. If you want to extend this package or use it more precisely, please refer ScalaDoc

Download

Currently ScalaNetwork supports Scala version 2.10 ~ 2.11.

  • Stable Release is 1.3.0

If you are using SBT, add a dependency as described below:

libraryDependencies += "kr.ac.kaist.ir" %% "scalanetwork" % "1.3.0"

If you are using Maven, add a dependency as described below:

<dependency>
  <groupId>kr.ac.kaist.ir</groupId>
  <artifactId>scalanetwork_${your.scala.version}</artifactId>
  <version>1.3.0</version>
</dependency>

Simple Example

Network.apply(Activation, Int*) generates fully-connected network:

// Define 2 -> 4 -> 1 Layered, Fully connected network.
val net = Network(Sigmoid, 2, 4, 1)
// Define Manipulation Type. VectorType, AEType, RAEType, StandardRAEType, URAEType, and StringToVectorType.
val operation = new VectorType(
   corrupt = GaussianCorruption(variance = 0.1)
)
// Define Training Style. SingleThreadTrainStyle, MultiThreadTrainStyle, & DistBeliefTrainStyle
val style = new SingleThreadTrainStyle(
  net = net,
  algorithm = new StochasticGradientDescent(l2decay = 0.0001f),
  make = operation,
  param = SimpleTrainingCriteria(miniBatchFraction = 0.01f))
// Define Trainer
val train = new Trainer(
  style = style,
  stops = StoppingCriteria(maxIter = 100000))
// Do Train
train.train(set, valid)

Network Creation

To create network, you can choose one of the followings:

  • Most simplest : Using sugar syntax, Network.apply
// Network(Activation, SizeOfLayer1, SizeOfLayer2, SizeOfLayer3, ...)
Network(Sigmoid, 2, 4, 1)
Network(HyperbolicTangent, 4, 10, 7)
Network(Rectifier, 30, 10, 5)
Network(Softplus, 100, 50, 30, 10, 1)
  • If you want different activation functions for each layer,
val layer1 = new BasicLayer(10 -> 7, Sigmoid)
val layer2 = new SplitTensorLayer((3, 4) -> 2, Rectifier)
new BasicNetwork(Seq(layer1, layer2), 0.95)

Second argument of Basic Network indicates presence probability, i.e. 1 - (neuron drop-out probability for drop-out training). Default is 1.

  • If you want single-layer AutoEncoder,
val layer = new ReconBasicLayer(10 -> 7, Sigmoid)
new AutoEncoder(layer, 0.95)

AutoEncoder only accepts Reconstructable type. Currently, ReconBasicLayer is only supported one. (Tensor layer version is planned)

  • If you want to stack autoencoders,
val net1 = new AutoEncoder(...)
val net2 = new AutoEncoder(...)
new StackedAutoEncoder(Seq(net1, net2))

Note that StackedAutoEncoder does not get any presence probability.

Training

Algorithm & Training Criteria

Before choose Training Style, you must specify algorithm and training criteria.

/* Algorithms */
new StochasticGradientDescent(rate=0.8, l1decay=0.0, l2decay=0.0001, momentum=0.0001)
new AdaGrad(rate=0.6, l1decay=0.0, l2decay=0.0001)
new AdaDelta(l1decay=0.0, l2decay=0.0001, decay=0.95, epsilon=1e-6)
/* Training Criteria */
import scala.concurrent.duration._
SimpleTrainingCriteria(miniBatchFraction=0.01f, validationSize=20)
DistBeliefCriteria(miniBatchFraction=0.01f, validationSize=20, submitInterval=1.seconds,
  updateStep=2, fetchStep=10, numCores=1, repartitionOnStart = true, storageLevel = StorageLevel.MEMORY_ONLY)

Validation size sets the number of elements used for validation phrase.

Input Options

Also you can specify input operations or options.

/* Corruptions */
NoCorruption
DroppingCorruption(presence=0.95)
GaussianCorruption(mean=0, variance=0.1)
/* Objective Functions */
SquaredErr
CrossEntropyErr // Which is Logistic Err
/* Manipulation Type : Vector input, Vector output */
// General Neural Network type
new VectorType(corrupt, objective)
// General AutoEncoder type
new AEType(corrupt, objective)

/* Manipulation Type : Tree input, Null output (AutoEncoder) */
// Train network as RAE style. 
// Every internal node regarded as reconstruction its direct children (not all leaves).
new RAEType(corrupt, objective)
new StandardRAEType(corrupt, objective)
// Experimental: Train network as URAE style. 
// With same structure, network should reconstruct all leaves from root.
new URAEType(corrupt, objective)

/* Manipulation Type : String input, Vector output */
new StringToVectorType(model, objective)

Training Style

You can choose the training style of the network.

/* Styles */
new SingleThreadTrainStyle(net, algorithm, make:ManipulationType, param)
new MultiThreadTrainStyle(net, sparkContext, algorithm, make:ManipulationType, param:DistBeliefCriteria)
new DistBeliefTrainStyle(net, sparkContext, algorithm, make:ManipulationType, param:DistBeliefCriteria)

Training

Training is done by Trainer class.

/* Stopping Criteria */
StoppingCriteria(maxIter = 100000, waitAfterUpdate=2,
  improveThreshold=0.95, lossThreshold=1e-4, validationFreq=1.0f)

/* Trainer */
new Trainer(style = style, stops = StoppingCriteria(), name = "Trainer")
  • waitAfterUpdate indicates wating time from the improvement. If network output improved on 100-th iteration, the trainer waits until Max(validationEpoch, 100 * patienceStep).
  • Improve Threshold indicates bottom line for improvement. To be regarded as improved, loss should be less than (best loss) * improveThreshold
  • Loss threshold indicates maximum loss can be accepted.
  • Validation Frequency sets the number of iterations between validations. (1 iteration does train all training examples)

Training is done by train method.

// If training and validation set are the same
trainer.train(Seq[(IN, OUT)])
trainer.train(Int => Seq[(IN, OUT)]) // With generator.

// If they are different
trainer.train(Seq[(IN, OUT)], Seq[(IN, OUT)])
trainer.train(Int => Seq[(IN, OUT)], Int => Seq[(IN, OUT)])

// If you are using RDD
trainer.train(RDD[(IN, OUT)])
trainer.train(RDD[(IN, OUT)], RDD[(IN, OUT)])

If you are using RDD, ScalaNetwork automatically caches your input sequence.

Also you can add negative examples, using trainer.setNegativeTrainingReference()