Skip to content

Latest commit

 

History

History
122 lines (82 loc) · 5.48 KB

README.md

File metadata and controls

122 lines (82 loc) · 5.48 KB
Raw

StepRampMCMC

Matlab+CUDA code to run the stepping/ramping model comparison described in Latimer et al, Science 2015.

This code takes in spike data observed over many trials, fits "stepping" and "ramping" models to the data, and compares the model fits with the Deviance Information Criterion (DIC).

Requirements:

  • MATLAB version R2013a (or above) with statistics toolbox
  • Nvidia CUDA toolkit (version 7.0 recommended)
  • Nvidia graphics card (compute version 1.3 or above. See the list provided by Nvidia at https://developer.nvidia.com/cuda-gpus At least 1GB of graphics memory is recommended. Apologies to AMD users.) The code uses double precision floating point arithmetic. Although the code will run on Nvidia cars with the proper compute version, it may be extremely slow in some lines of gaming cards. For example, the GTX 900 series runs double precision operations at approximately 1/32 the speed of single precision operations.

Recommended:

  • Ubuntu 15.04
  • 16+GB RAM
  • NVidia Titan, Titan Z, or Tesla card

Release Notes

The tools provided run in MATLAB and use Nvidia's CUDA toolkit to accelerate compution. Running this package requires an Nvidia GPU. This code has been tested with MATLAB R2013a with CUDA version 7.0 running on Ubuntu 15.04. The graphics card used was an Nvidia Titan Z.

Previous versions of this package have been tested on Windows 7 and Mac OSX. However, this release is targeted towards Ubuntu users. Compiling the CUDA tools on OSX works similarly to the compilation on Ubuntu. If running this code on Windows is absolutely necessary, the author may provide limited support in getting the GPU code functional on windows.

Setup / Installation

Make sure you have installed the CUDA toolkit from https://developer.nvidia.com/cuda-downloads We recommend installing the CUDA samples to the 'samples' folder of your CUDA install directory.

To complete the CUDA setup, you must include the following lines to the end of your .bashrc file in your home directory:

export PATH=/usr/local/cuda-7.0/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda-7.0/lib64:$LD_LIBRARY_PATH

(make sure the directories match your CUDA install directory!)

To compile the CUDA files into MEX files that can be executed by MATLAB, cd into the directory code/CUDAlib/ and then:

  1. Edit myPath.m and execute myPaths;
    This tells the compiler where to find your MATLAB and CUDA installations. The MATLAB directory should be detected automatically. However, you may need to set the correct CUDA path. If your GPU has compute version less than 3.0, you will need to change the GPUArchitecture variable to match your GPU's compute version. If executing this function reports an error, open myPaths.m and edit the relevant path name so it points to the correct directory.

  2. Execute compileAllMexFiles;
    This should compile/link all scripts.

WARNING: Using these CUDA tools and MATLAB's native GPU functions in the same session may cause a library conflict that will crash MATLAB.

Known problems and warning messages setting up the CUDA/MEX files:

  • Compiler error: "mex.h" not found
    Solution: this probably means your MATLAB directory is not set up properly in code/CUDAlib/myPath.m

  • Linking error: undefined reference to `__cxa_atexit'
    Solution: append -lm -lstdc++ to the CXXLIBS option in your mexopts.sh file

  • MATLAB's mex linker may report a gcc version warning such as Warning: You are using gcc version "4.9.1". The version currently supported with MEX is "4.4.x". Solution: We have tested that this warning can be ignored in MATLAB R2013a.

  • Compiler gives a warning about a CUDA header file such as /usr/local/cuda/samples/common/inc/exception.h Solution: This problem is associated with older versions of CUDA and can be ignored. Upgrading to version 7.0 eliminates this warning.

  • Receiving error message while running the example script Warning: Unknown problem with sampling drift variance (most likely numerical error). Solution: The CUDA files may not have compiled correctly. Check the GPUArchitecture variable in coda/CUDAlib/myPath.m to make sure it matches your GPU's compute version and recompile.

This code was made before the current MATLAB versions (R2015) with better built-in functionality between gpuArrays and mex files. Therefore, we include our own wrapper to place data on the GPU from MATLAB for mex functions along with these specific compiling functions.

Running the Model Comparison

The script exampleScript.m simulates data from one of the two models and runs the model comparison function.

This script calls code/runModelComparison.m, which takes the data in the form of a timeSeries object described in the function header. The timeSeries is a processed form of the spike data observed over all trials. Only the window of interest in each trial is included. (i.e., spikes before integration and after integration have already been pruned)

For examples of what the timeSeries structure should look like, see the functions

  • simulateRampingModel.m
  • simulateSteppingModel.m

These functions generate a simulated timeSeries object with all the necessary parts.

All settings, including how many MCMC samples to use, are given in setupMCMCParams.m