To run the simulations simply run the script code.R. The plot output corresponding to each simulation will appear in the directory plots.
The simulations depend on the following R packages: sgmcmc (the package itself), rstan, MASS, ggplot2 (all available on CRAN). These dependencies should be automatically installed when code.R is run. The script will also check TensorFlow for R (also available on CRAN) has been installed properly, which is a dependency for sgmcmc. If it is not, it will attempt to install this itself. The installation of TensorFlow for R requires python-pip and python-virtualenv to be installed so if these are not available the script will stop and prompt you to install these.
Some of the simulations load large datasets into memory, so it’s recommended you have at least 8GB of RAM.
Simulations took approximately 3 hours on a quad core i5 laptop with 8GB RAM.
Error: Python module tensorflow was not found. – Try rerunning code.R or restarting your R session. This Error can sometimes occur when TensorFlow is installed during the code.R script and then is immediately called. Restarting the R session and rerunning the script should solve the error.
Original results were run on a laptop with Ubuntu 16.04 LTS; R version 3.2.3; Python version 2.7.12; TensorFlow version 1.7.0; TensorFlow for R version 1.5; sgmcmc version 0.2.2; rstan version 2.16.2.
While we can guarantee reproducibility on a single platform, and have ensured reproducibility across two platforms (see Original Setup Section). The TensorFlow seed setting appears to be very dependent on version and platform, this can make reproducibility across different platforms difficult. We have done our best to make everything as reproducible as possible and have given as much detail as possible of our set up in the Original Setup Section to make it easier.