This repository contains source code and sample data relevant to the paper "Unsupervised Learning of Cone Spectral Classes from Natural Images" by Noah C. Benson, Jeremy R. Manning, and David H. Brainard, currently submitted for publication to PLoS Computational Biology. The respository is broken into several parts, all detailed under the Organization header. In brief, the source code necessary to produce all simulations, analyses, and figures reported in the paper are included here, and are broken into three parts:
- A simulation library written in Clojure, an open-source Java Virtual Machine-based functional language
- A few Matlab functions that can be used for embedding the correlation matrices in the outputs of the simulations (big Endian binary sim[*].bin files)
- A Mathematica Notebook containing analysis and figure-generation code.
Primary Author: Noah C. Benson
Principle Investigator: David H. Brainard
This library is organized into a few directories, detailed here:
- analysis - The analysis directory contains all the code needed to post-process the simulations and plot the figures from the paper. This directory inludes both Matlab code (for embedding the correlation matrices) and Mathematica code (for flattening and classifying retinas and for generating figures).
- doc - The doc directory contains html documentation for the clojure library. This html may also be access directly here. The Matlab and Mathematica code in the analysis directory is documented internally.
- samples - The samples directory contains examples of natural image patches that were drawn from our hyperspectral database for use in simulation of a 20 x 20 retina.
- src - The src (source) directory contains clojure source code for the receptor learning retina simulation library. It is documented in the doc directory.
- resources - The resources directory contains additional resources that are not required for use in the library, but that may be useful.
This library is designed for use in simulating the responses of retinas to natural images. The expectation is that it will usually be used in conjunction with Leiningen, which provides a single script, lein, that can be downloaded and run in order to resolve and install all dependencies in this local directory. To do this, simply follow these instructions:
- place the lein script (at the time of the writing of this README, the script can be found here for Mac or Unix and here for Windows), in your path,
- clone this git repository,
- in a terminal, switch to the project's root directory (the directory containing the project.clj file),
- enter 'lein repl'.
This should result in a number of dependencies being resolved followed
by a prompt. This prompt is a clojure repl; to import the library,
(use '(brainardlab.nben.retina core constants hyperspectral
You may then access functions such as simulate-some-retinas directly.
Using the main function
Alternately, you may instead enter 'lein run' in place of 'lein repl'; this is the interface for calling the main function in the library. The main function expects 4 arguments: the name of the plan to simulate (these are listed in util.clj or brainardlab.nben.retina.util/simulation-plans, which is well-documented; note that these are keywords in the code, but this argument expects a string without the preceding :, as in "tritanopes", not ":tritanopes"), the hyperspectral cache file, the output directory (into which simulation files will be written), the number of workers processing the plan, and the id of this worker. If you are running the simulations on a single computer, then you would be using 1 worker and that worker's id would be 0. For distributed calculation of a plan (such as the rather large :standard plan), if the number of workers is n and each worker is given an id 0 ... n-1, then the plan will be divided up and executed as evenly as possible.
The simulations expect to load their images from a hyperspectral cache. This is effectively a pre-cached set of natural images, all of which have had a fair amount of pre-processing done so as to optimize the speed of simulations. Hyperspectral cache files can be built using the functions in the hyperspectral namespace (brainardlab.nben.retina.hyperspectral). The relevant functions are build-hyperspectral-cache (which requires that images be added sequentially to the cache using the push-to-cache function) and autobuild-hyperspectral-cache. The latter is recommended, as it requires only the paths of the databases used in Benson et al. (2014), and will automatically load the images and build the cache from them. For example:
(autobuild-hyperspectral-cache "data/my-hyperspectral-cache.bin" :Chakrabarti2011 "hsdbs/Chakrabarti2011/CZ_hsdb" :Foster2004 "hsdbs/Foster2004" :Parragas1998 "hsdbs/Parragas1998/brelstaff")
Note that this entire library was written in clojure partially because clojure's paradigm is optimized for multi-threaded programming. This library automatically threads across as many processors as possible when performing simulations, and it is much much more efficient to run simulations in bulk. That said, there are some perils. Primarily, if you simulate a very small retina with a very large retina, the small retina will finish scanning the entire hyperspectral cache before the large retina has advanced very far. This can result in memory problems, as the library holds in memory all hyperspectral images in the hyperspectral cache that have (1) been seen by at least one retina and (2) have not been seen by all retinas in the simulation. Because the retinas scan the cache in order, one ideally wants them to be approximately in sync with each other.
Documentation is available in the doc directory of the repository. This documentation was generated with Codox, and provides descriptions of all the public functions in the package. This documentation can also be found online here.
Copyright © 2013-2014 Noah C. Benson
Distributed under the Eclipse Public License, the same as Clojure.