JRNNLM: Recurrent neural network implementation in JAVA.

Recurrent neural network implementation in JAVA.

The code is refactored for simplicity and extensibility.

Current version:

• No class support
• No direct connection support

How to test JRNNLM

Small size language model

java -cp "target/*:lib/*" jrnnlm.test.training.SmallFileTest

Normal size language model

java -cp "target/*:lib/*" jrnnlm.test.training.NormalFileTest

Parallel training (Still unstable)

java -cp "target/*:lib/*" jrnnlm.test.training.SmallParallelTest

How to train a model for specific task

This code example shows how to train a simple language model. Change the parameters in RNNLMConfiguration for different tasks.

RNNLMConfiguration conf = new RNNLMConfiguration();
conf.trainFile = new File("data/your_training_file");
conf.validFile = new File("data/your_validation_file");
conf.hiddenSize = 50; // Size of hidden layer
conf.maxIters = 50; // Max iterations

RNNLM lm = new RNNLM(conf);
lm.train();

Size of Hidden Layer

• < 1M words: 50 - 200
• 1m - 10m words: 200 - 300
• more words: the vocabulary size should be limited, rare words should be mapped to "<unk>"

Raphael Shu (2014) (still in process)