This project is based on the NOMAD project of Hyokun Yun [1]. It is modified/rewritten for my own research purpose.

Information of the original version:

• NOMAD: Non-locking, stOchastic Multi-machine algorithm for Asynchronous and Decentralized matrix completion. NOMAD is a package for large-scale distributed matrix completion. Please refer to the paper [1] for detailed discussion on the algorithm.

• The original readme file is renamed as "README-old.md". The code is put into the branch named "version-of-Yun".

  • In order to compile the original code in modern compiler, you need to remove the line #include <tbb/compat/thread> and replace all tbb::tick_count::interval_t with std::chrono::duration<double> in the "nomad_body.hpp" file.

• NOMAD is under Apache License ver 2.0; see LICENSE file for detailed information.

Information of my version.

• The code is fully reorganized for readability.

• Fix several bugs

• Add network controlling functions

• Add fault tolerance functions

Installation

Prerequisites

• MPI library, with multi-threading support
• Intel Thread Building Block (at least 4.1)
• CMake (at least 2.6)
• Boost library (at least 1.49)
• A C++ compiler which supports C++11

Compile

To compile NOMAD, move to the root of this project and execute the make command. The executive files will be stored in the bin/ folder.

Alternatively, you can move the the code directory of nomad ./Code/nomad and use the cmake command to generate your own compiling folder.

Usage

a. Data Processing

To use NOMAD, you need to convert a text data file to a binary format NOMAD can read. The original text data for training set should be in tab-delimited form, as follows:

$ cat ./Data/tutorial/train.txt

user_A item_1 2.0 user_A item_2 3.0 user_B item_1 4.0 user_B item_3 7.0

...

Test dataset should be prepared accordingly. Then, you can execute the conversion script:

$ python ./Scripts/convert.py ./Data/tutorial/train.txt ./Data/tutorial/test.txt ./Data/tutorial/

This will generate 'train.dat' and 'test.dat' on the destination directory './Data/tutorial/'.

Execution

NOMAD has two executables, 'nomad_float' and 'nomad_double'. The former uses single-precision, while the latter uses double-precision.

You can execute nomad with --help command to see the list of options

$ ./nomad_double --help nomad options: -h [ --help ] produce help message --nthreads arg (=4) number of threads to use (0: automatic) -l [ --lrate ] arg (=0.001) learning rate -d [ --drate ] arg (=0.10000000000000001) decay rate -r [ --reg ] arg (=1) regularization parameter lambda -s [ --seed ] arg (=12345) seed value of random number generator -t [ --timeout ] arg (=10.0) timeout seconds until completion -p [ --ptoken ] arg (=1024) number of tokens in the pipeline -d [ --dim ] arg (=100) dimension of latent space --reuse arg (=1) number of column reuse --pause arg (=1) number of column reuse --r0delay arg (=1) arbitrary network delay added to communication of rank 0 machine --output arg path of the file the result will be printed into --path arg path of data

[1] NOMAD: Non-locking, stOchastic Multi-machine algorithm for asynchronous and Decentralized Matrix Completion (Hyokun Yun, Hsiang-Fu Yu, Cho-Jui Hsieh, S.V.N. Vishwanathan, Inderjit Dhillon)