NOTE: the instructions contained on this page were written for use specifically in the CLSP WS02 Lab. 
They may refer to applications, files or other materials that are not accessible from other locations.


Graphical Models ToolKit Lab
JHU Summer School on Human Language Technology 
July 9, 2002

Gang Ji (gji@crow.ee.washington.edu) and Sourin Das(sourin@clsp.jhu.edu)


In today's lab you will learn how to :

Run a Viterbi decoder using some already-existing model parameters.
Train a set of model parameters.
Modify the structure of decoder model so that it is guaranteed to decode at least one non-silence word, and no more than 7.

NOTE: The text in italics denotes commands that needs to be executed.

PART 1:
1.Set the following paths:

setenv PATH /export/apps/bin:$PATH
setenv PATH /export/ws02asr/tools/gmtk/solaris/bin:$PATH 
rehash

2.Copy the tutorial files into your home directory:
   
cp /export/ws02asr/tools/gmtk/auroraTutorial.tar.gz ~

3.Untar and unzip it:
   
tar -xvzf auroraTutorial.tar.gz

4.Change to the tutorial directory and do the following:
chmod 700 vitcommand traincommand     # Just to give write permission.

IMP NOTE: Go through the "README " file. 

5.Edit the "vitcommand" file. Change the following line
GMTKVITERBI=../../tksrc/gmtkViterbi
to
GMTKVITERBI= /export/ws02asr/tools/gmtk/solaris/bin/gmtkViterbi
and save your file.

6.Run the decoder:
 vitcommand

It will start generating the decoded output.

PART 2:

1.Edit the "traincommand" file. Change the following line 
GMTKEMTRAIN=../../tksrc/gmtkEMtrain
to
GMTKEMTRAIN=/export/ws02asr/tools/gmtk/solaris/bin/gmtkEMtrain
and save your files.
2.Run the train command to train a new set of parameters.
traincommand
NOTE: It may take few minutes to train 

3.Decode the data using your newly train parameters. Edit the "vitcommand" file. Comment out  the line ( by putting a '#' in front of it)  starting with '$GMTKVITERBI -of1'  in the first block. Then uncomment (remove the leading '#')
line starting with '$GMTKVITERBI -of1' in the second block. Save your file.

4.Run the decoder again:
vitcommand


PART 3:

In this part you will modify the structure so that network so that no more than 7 words (counting silence) will be decoded. 

1.Create a new directory:
mkdir MY_PARAMS
2.Copy the following parameter file to your new directory:
cp  PARAMS/aurora_decode.str MY_PARAMS/
cp  PARAMS/nonTrainable.master MY_PARAMS/
cp  PARAMS/commonParams  MY_PARAMS/
cp  PARAMS/learned_parms.gmp MY_PARAMS/

3.Modify the structure of your model so that the above mentioned constraint is followed. You need to modify the "aurora_decode.str" file for this.

4.Run the decoder to check your model.