bulkem

R package to perform multicore and CUDA-accelerated fitting of finite mixture models.

Use cases

• You want to fit mixture models composed of inverse Gaussian components
• You have a lot of datasets to fit (thousands) and it's taking too long
• The datasets are difficult to fit, so you want to try a lot of initialisation parameters, and this is taking a long time
• You have a lot of difficult-to-fit datasets and it's taking days to run
• You have large datasets (over 10000 observations) to fit

Installation on Ubuntu 14.04 LTS (Trusty 64-bit) from a bare install

We need a newer version of R, so edit /etc/apt/sources.list and add the line

deb http://cran.rstudio.com/bin/linux/ubuntu trusty/

Then at the command line,

$ gpg --keyserver keyserver.ubuntu.com --recv-key E084DAB9
$ gpg -a --export E084DAB9 | sudo apt-key add -
$ sudo apt-get update
$ sudo apt-get install libssl-dev r-base r-cran-rcurl libxml2-dev libcurl4-openssl-dev nvidia-current git-core

Download the CUDA Toolkit:

wget http://developer.download.nvidia.com/compute/cuda/6_5/rel/installers/cuda_6.5.14_linux_64.run

Install the CUDA Toolkit. bulkem assumes that you use the default installation location of /usr/local/cuda-6.5. We only need the toolkit; the graphics driver has already been installed.

$ chmod +x cuda_6.5.14_linux_64.run
$ sudo ./cuda_6.5.14_linux_64.run -toolkit -silent

Start R:

$ R

Install the devtools package:

> install.packages('devtools')

Finally, we can install bulkem. This is manual at the moment:

$ git clone https://github.com/ihowson/bulkem
$ cd bulkem/src/cuda
$ make
$ cd ../..
$ LD_LIBRARY_PATH=/usr/local/cuda/lib64 R
> devtools::build()
> devtools::install()

demo.R provides a simple demonstration:

> source('misc/demo.R')

On my computer, this produces:

> source('misc/demo.R')
[1] "--- Using GPU: 1. 100 datasets with 2000 observations in each, 10 attempts per dataset."
Processed 100 chunks
cuda parallel: 0.423166 seconds elapsed
[1] "The last fit looks like mu=[2.783511 1.015576], lambda=[0.210135 28.845700], alpha=[0.722182 0.277818]"
[1] "    Completed in 0.511000 seconds"
[1] "--- Using GPU: 0. 100 datasets with 2000 observations in each, 10 attempts per dataset."
[1] "The last fit looks like mu=[1.015576 2.783510], lambda=[28.845905 0.210135], alpha=[0.277818 0.722182]"
[1] "    Completed in 10.214000 seconds"