code for fun

a mix of Matlab, Java, Perl code.
that's all i write these days.

[Graphs] [Clique tree inference] [What do you say?] [Building a lexicon from emails] [Calling Matlab from Java] [Useful utilities]

if you use my code or find problems with it,
i'd appreciate you sending me an email about it.

Graphs

Standard graph operations implemented in Matlab.

moralize.m
one of the inputs is a file with variable names. if you don't want to use this for manual verification, just comment the code that requires "flag" and "gnames" or "num2cname".
triangulate.m
a greedy implementation of triagulation. also requires the following files: choose_node.m, new_edges.m, induced_cluster.m, add_edges.m, remap.m, rmnode.m calc_weight.m get_realpos.m

Clique tree inference

Based on one of the clearest papers I've ever read! If you need to implement any kind of inference, this is the paper to read up on:
Huang and Darwiche (1994). Inference in Belief Networks: A Procedural Guide.
The following code implements the steps identified in their paper to build the secondary structure.

step 1. see the graph section.
step 2. see the graph section.
step 3. idcliques.m
step 4. buildjtree.m

Some more code to make things more operational.

validate_localconsistency.m
checks equation (2) is true in the built structure, or after every update.
validate_joinproperty.m
checks equation (3) is true in the built structure.
canonize.m
here's how to convert a dbn's dag into a canonical dbn's dag, while retaining only variables that are canonical by definition. be aware that the dag has "shifted" so the dag indices for the variables in the original dbn no longer corresponding to those in the original dag.
enforce_clique.m
give it a vector of the indices (in the new dag) and make sure those variables form a clique.
mk_converjtree.m
if you built a join tree after making it canonical, you'll need to convert the variables back to the original indices.
globalpropagation.m
uses the following:
- unmark_cliques.m
- collect_evidence.m
- distribute_evidence.m
- some helper procedures: find_unbs.m, onemsgpass.m get_sset.m
  msgpass.m please note: msgpass.m simply sends one message between X and Y through R, using the following data structure specific files (not included here): mp_addout.m (marginalizes out a set of variables), mk_init_mp0.m, mp_div.m, mp_mult.m.

What do you say?

This one is kind of cute. I wanted to write a program that generates sentences based on words that I use often. So this is kind of what it does, although you've got to tell it specific the syntax of the sentence you'd want to generate. Take a look at the readme file for more details.

genbo.pl
uses trained trigrams to generate sentences.
ranbo.pl
random, no stats considered.
cleanGrams.pl
utility for cleaning up and segmenting an arpa file.

Code needed from the section building a lexicon from emails.

Building a lexicon from emails

Some code for taking a text inbox and formatting it nicely. After that, split the text into separate files according their lexical categories. (Built-in lexicon is limited. Should expand this for real!) Take a look at the readme file for more details.

If you ever need to call Matlab from Java, here's the trick. Use the Java class developed by Kamin at MatlabControl.html. If you follow all his instructions, you should be able to get the two to work together. I had more problems because Matlab doesn't understand focus and I needed to be able to get my Java stand-alone working the same way it would with Matlab hooked up to it. So I used Kamin's class to make a MatlabJava server, and made my Java stand-alone a client. That way, I could bypass the whole focus problem because my stand-alone program doesn't need to communicate with Matlab at all. Here is the code.

the server code
start this in matlab
make sure it is listening with the following:
>> ms = MatServer;
>> ms.body

whatever Matlab scripts you need to run, put them inside processInput(). the two scripts in the code are testEE.m and testDM.m with their expected syntax. don't worry about what these two scripts do, just get them to do some disp() statements at first to get the overall infrastructure going.
having said that, you'll also find that an additional file cp.txt is required when testDM is called. this is the file that stores the output of the work from testDM inside Matlab, so that Java can parse it and make use of the output. this means that inside testDM.m, in additional to the computation, you will need to make sure that it does something like: save 'cp.txt' -ascii RESULT.
the client code
no need to "start" this anywhere, just use the test files below
no changes necessary to get it up and running. but all the matlab scripts and processing the results from the server is updated here.
the test code
start this in the command prompt (or however you run java)
this is how my stand-alone would interact with the client.
another test code
same as testClientBasic.java
more complicated cases.

Useful utilities

Some useful files...

getpapers.pl
text mode for grabbing files to your directory. i found this useful whenever the gui programs didn't work or took too many human steps. if you have the links, just stick them into a file and let this script run overnight.
chPStoTXT.pl
changing the extensions of multiple files at once. this one does .ps to .txt. you can modify it to suit your needs.
reminder.pl
takes a list of appointments and beeps when the time is up.
vi.m
simple (not optimized!) value iteration code for MDPs.

Free counter