Read Me

####################################################
# Created by Jordan Barnes, Simon Fraser University#
# Dec 11, 2008. For questions email jordanb@sfu.ca #
####################################################

#Requirements: Python, Matlab


DataPrep Folder

This folder contains all the tools needed to extract letter data. 

Selector.py - allows you to select the abstract letter parts from the gridfonts.data.txt corpus. Output data is saved to 'roles.txt' in triplets of:

	figure <81 units long, encoded as a Matlab vector>
	whole <81 units long, encoded as a Matlab vector>
	ground <81 units long, encoded as a Matlab vector>

Selecting the figures from 6 different letters for example, will produce a roles.txt that is 18 lines long. 

shuffler.py - enriches the data-set by randomly increasing the size of the input file. Any sized input vectors will yield a corresponding sized output vector.

expander72.py - randomly recoded the input data to be shifted in to one of 9 possible vertical locations. Data must be 81 units long. Output will be 153 units long.

3rd party module - a.py is a freely available python module for hex and binary conversions. Needed for gridfont hex conversion.

Abstract Folder

This folder contains the network built to completely emulate the Analogator network made by Douglas Blank. This network needs tweaking! Currently it is not converging on abstract representations. The algorithm looks right but the parameters need tweaking.

wholeNet.m - Accepts two files, that use 152 bit dynamic location letter representations. wholespace.txt contains a large corpus of step one training data. wholetestset.txt, contains a smaller number of letters, for the representations in step two to be continually compared to. wholespacevalids.txt contains a test set of letters in order to determine the networks performance.

Feature Folder

This folder contains the network built to emulate the kinds of feature errors that Blank got when letters appeared in static positions. This network semi converges but not as well as Blank's did. Again, parameter tweaking is needed.

feature.m - Accepts two files, that use 81 bit static location letter representations. aroles.txt contains a large corpus of step one training data. testset.txt, contains a smaller number of letters, for the representations in step two to be continually compared to. avalids.txt contains a test set of letters in order to determine the networks performance.

Single Folder

Just a one step role dissociating network. Mainly meant to test that the Matlab NNT functions are all working as they should.

CompSingle.m - Accepts one file of 81 bit representations, aroles.txt. avalids.txt contains a test set of letters in order to determine the networks performance.

Commons

glimpse.py - allows you to view network output and visually verify any representations used. Inputs are given 81 input units per line.

graphics.py - is a widely used python module for object oriented graphics