Big Topic Model is a fast engine for running large-scale Topic Models. It uses a hybrid data and model parallel mechanism to accelerate and performs 3~5 times faster than the state-of-the-art ones on some general dataset. It supports a set of tpoic models including LDA, DTM, MedLDA and RTM.
Big Topic Model depends on several third-party libraries including google glog, gflags, dSFMT and eigen. Currently our tests compile and run mainly based on Intel libraries(Intel® Parallel Studio XE Professional Edition for C++ Linux* 2016), but other libraries like openmpi are practicable.
Go to http://software.intel.com to download the Intel software Toolkit. And then set the compilevars:
source /opt/intel/2016/compilers_and_libraries/linux/bin/compilervars.sh intel64
Make sure you can access github.com. To get third-party dependencies, run:
$ ./set_third_party.sh
(You do not need to do this if you already have all the third-party dependencies.)
First, run build.sh under the root directory to generate release and debug folder.
$ ./build.sh
Second, using make to compile.
$ cd release/
$ make
Set the data under root direcotry, run:
$ ln -sf where/you/store/the/data/ data
To run on a cluster, data partition is necessary. Unlike other current cluster computing systems, Big Topic Model not only segments data by documents, but also by words, that is a two-dimensional partition. So the data we need is different from the traditional svm data, and we need to preprocess the data by distributing words randomly for each group before running on cluster.
Use split-input-data.sh in src folder to divide the raw data into groups of equal size for the cluster to format.
Now, we get a package of data for each machine. Use format.py in src folder to do the data partition according to the parallelism.
Modify the run.sh under release folder, to run your model.
$ cd release/
$ ./run.sh