Maximally Informative Dimensions wrapper in MATLAB
Contained in this README is information on how to install and use the MATLAB Maximally Informative Dimensions (MID) code wrapper. This is called a "wrapper" because the MID code itself is written in C with parameter values defined in XML. This wrapper allows users more familiar with MATLAB to use the MID code without having to touch the raw MID code. MID is a method for characterizing receptive fields of neurons responding to stimuli including naturalistic stimuli. To find out more, see https://www.princeton.edu/~wbialek/rome/refs/sharpee+al_04.pdf.
Unarchive the file "matlab_mid_wrapper.tar" in your preferred directory. This will produce a new directory called "matlab_mid_wrapper". Make sure all the files listed in the below manifest are present in the correct folders.
MANIFEST: ./matlab_mid_wrapper/README.txt ./matlab_mid_wrapper/EASY_DOWNLOAD.m ./matlab_mid_wrapper/EASY_INSTALL.m ./matlab_mid_wrapper/organize.m ./matlab_mid_wrapper/submit_jobs.py ./matlab_mid_wrapper/sample_interactive_script.m ./matlab_mid_wrapper/sample_cluster_script.m ./matlab_mid_wrapper/MATLAB_functions/generate_xml_mid.m ./matlab_mid_wrapper/MATLAB_functions/prepare_mid_for_cluster.m ./matlab_mid_wrapper/MATLAB_functions/run_mid_interactive.m
Note: The "matlab_mid_wrapper" directory does not contain all the necessary files yet to run MID. They must be downloaded from github. The following is instructions on how to acquire the remaining files. Follow the directions carefully as the directory structure is important to properly use this wrapper.
- If you do not have git already, download it from http://git-scm.com/download
From here there are two methods for downloading and installing the files. It is recommended you use method A.
A3) Open MATLAB and change your working directory to "matlab_mid_wrapper". You must change to this directory, you cannot just add a path to it.
A4) Run the script "EASY_DOWNLOAD.m". You will see download progress output in the MATLAB command window.
A5) After all the files are downloaded, make sure there is a folder called "mid" in the "matlab_mid_wrapper" directory. Then run the script "EASY_INSTALL.m". This will show the compilation progress of the raw MID code.
A6) When you stop seeing output from the compilation process in the MATLAB command window, check to make sure the word "error" hasn't appeared in the last few lines of output. If an error has occurred, the code has not been compiled and MID cannot be used until the error is fixed. Review the above steps.
B3) Open Unix/Linux terminal. Go to your "matlab_mid_wrapper" directory.
B4) Get the code: git clone git://github.com/sharpee/mid.git mid
B5) Get the demo (if you like): git clone git://github.com/sharpee/mid-demo.git mid-demo
B6) After all the files are downloaded, make sure there is a folder called "mid" in the "matlab_mid_wrapper" directory.
B7) If you do not have the mxml package on your system enter these commands:
cd mid git submodule init git submodule update cd mid/mxml-2.6 ./configure --prefix="$PWD" make make install cd ../..
B8) Compile the MID code by entering the following commands:
cd mid make make clean
That's it, the MID code is installed.
HOW TO USE MID
Refer to the two included sample scripts.
The script "sample_interactive_script.m" runs interactively; i.e. MID is run on the computer that is invoking the script and output is shown in the MATLAB terminal. Running interactively is useful for testing and analyzing a small number of data sets or in the absence of a cluster.
The other sample script, "sample_cluster_script.m" shows how to prepare shell scripts such that MID can be run simultaneously for multiple data sets on a cluster. Each data set will be submitted to a computer on the cluster. Running on a cluster is useful if you have many data sets to analyze. The code submitted to the cluster runs entirely in C thus no MATLAB licenses are used.
To get started quickly, simply copy a sample script and reassign the variables to the appropriate values for the data set(s) you would like to analyze. Heed the warnings in the comments, particularly about changing optimization parameters in the struct "fit_options". Then run the script.
If you are running an interactive script, this is all that needs to be done.
If you are running a cluster script, you must make a simple edit to the python script "submit_jobs.py". Make a list of the prefixes you used when you prepared the data for submission; e.g. prefix = ['aud1', 'aud2'] for two data sets assigned the prefix 'aud1' and 'aud2'. Run this script by entering "python submit_jobs.py" in the command terminal. This script will submit the shell scripts to the cluster. Note that you can write a shell script or a MATLAB script with the same functionality, but cluster submission servers may not have MATLAB installed.
The output of the MID code are binary files with data type 'double' ending with something similar to "prefix-ND-n2-v1-p1.dat" containing the receptive field features. They will appear in the main directory "matlab_mid_wrapper". Since this can become rather messy, run the script "organize.m" to move these files to the "Results" subdirectory. "v1" means it is the first vector fit (up to the assigned variable "nd") while "p1" is the training set used (usually up to 4 unless reassigned) and "n2" indicates that a two dimensional feature space was found (the second vector being in "prefix-ND-n2-v2-p1.dat").