Index of /~rtreiman/InflCContxtOnPronV/src

      Name                    Last modified       Size  Description


[DIR] Parent Directory        28-Jun-2005 10:57      -  
[TXT] Makefile                28-Jun-2005 10:51     2k  
[   ] Mikenet.tgz             28-Jun-2005 10:51   3.7M  
[TXT] crit_eval.c             28-Jun-2005 10:51     9k  
[TXT] evaluator.c             28-Jun-2005 10:51    11k  
[TXT] make_op_pat.c           28-Jun-2005 10:51     3k  
[TXT] model.c                 28-Jun-2005 10:51     3k  
[TXT] model.h                 28-Jun-2005 10:51     1k  
[TXT] simconfig.h             28-Jun-2005 10:51     1k  
[TXT] train.c                 28-Jun-2005 10:51     3k  
[TXT] which_ortho_features.c  28-Jun-2005 10:51     1k  


This directory contains source for all the programs used to run and
analyze the model.

train - is used to train the model, it gets invoked as follows:

Arguments are learning rate (epsilon) training pattern file, and a
vestigial "info" option that used to do something but now just has to
be there or the program will refuse to run. 

> train -epsilon 0.005 -pat train.pat 

evaluator - is used to analyze the model's performance on words
(train.pat)

crit_eval - is used to analyze the model's pronunciations of
nonwords. It scores them as "crit" "typical" or "other" based on the vowel.

Each of these gets invoked by specifying a weights file and
dictionary/pat for comparison:

> evaluator -load s1_final_weights.gz -pat train.pat -key train.dict 
> crit_eval -load s1_final_weights.gz -pat test.pat -key test.dict 

You will also need model.c which contains specifications about the
size and connectivity of the model, and simconfig.h which contains an
even higher-level description of same.

which_ortho_features and make_op_pat are included for completeness,
although they aren't necessary unless you want to change the stimulus
set. The first makes the ortho.features file, which lets you prune
unnecessary letters from the input, and the second makes a pattern
file out of a .dict file (translating human-readable specifications of
orthography/phonology into vectors of input/output patterns).

Of course you'll need to futz with the Makefile so that it points to
your Mikenet directory to get the right libraries.