cuda-word2vec

cuda implementation of CBOW model

Features

• parallel speedup with Nvidia GPU
• requires constant memory for arbitrarily large training set
• stream implementation requires no training data random access
• automatic validation and display validation data negative loglikelihood during training
• support custom word binary tree construction. improve model performance with your own binary tree constructed from any other unsupervised model

Updates 0.1

• optimize GPU memory IO efficiency exploiting locality
• optimize memory performance with cuda texture
• optimize disk to memory IO efficiency and save parsing overhead with binary preprocessing of training data
• back propagation gradient-check test completed
• implement automatic cuda memory management utilities to manage GPU resource

Requirement

• cuda 5.5
• cuda Thrust
• nvcc compiler
• 1GB or more GPU memory ( 2GB+ is recommended)

Compilation

• Linux: if nvcc command is available, compile with makefile
• Windows: compile the source file manually with -arch=compute_30 and -code=sm_30
• Source Files: language_network_main.cu language_network_utility.cu language_network_kernel.cu includes/optParser/getopt_pp.cpp

Input Format

• corpus should be preprocessed into a text file with each line a document.
• words in document should be convert into ints by any bijective mapping M [1-V] <-> vocabulary (store separtely as a convertable)
• Store word binary tree in two files: tree_point and tree_code.
• line N of tree_point should be a path through internal nodes from root to leaf word-N.
• line N of tree_code should be a binary sequence cooresponding to the navigation of the above path (from root to leaf-N)
• each word id n must have a code line and a point line in the tree files
• a java implementation of huffman tree file constructor is provided, use if new tree construction is not needed

Usage

command line options:

• --train path to training data
• --code path to tree_code data
• --point path to tree_point data
• --pFeature saving path of tree point feature
• --wFeature saving path of word feature
• --batch training batch size (length of corpus segment by words, use suitable batch size that fits in your GPU memory)
• --vRatio validation ratio ( integral value, validation will be performed after every vRatio training batch)
• --iter iterations of the training pass through training data
• --learningRate constant leanrning rate
• --binaryTraining cache path for binary preprocessing of the training set( delete manually after training)
• --binaryWindow cache path for random windows data (delete manually after training)
• --vocabSize vocabulary size
• --treeSize size of inernal nodes of the binary tree

Typedefs and Numerical Issues

all types are found in src/includes/language_network_typedefs.h, modify if needed

Benchmarks and Profiling

coming soon