-- Main.simonk - 21 Nov 2005

Using the Graphical Models Toolkit (GMTK) under Condor

Known limitations

Currently, GMTK makes system calls to the C preprocessor cpp when loading files, so it can therefore only be used under the Condor Vanilla Universe. We are hoping Jeff will fix this before too long...

Using Condor/GMTK with Python

There is a locally written version of the Perl Condor API in Python, so that Condor can be controlled entirely from within Python; also Python functions can be used as callbacks from Condor. On top of this, there are some routines that easily handle submission of single or parallel jobs to Condor, monitoring the return values. Examples of the use of these scripts are available: we have used them for performing parallel training/decoding in GMTK. More details to follow. In the meantime, contact Peter Bell (s0566164@inf) if you'd like to use them.

Implementing EM Training Schedules with DAGs

DAGs (directed acyclic graphs) can be used to schedule an EM training schedule.

There is an example of EM iteration until log-likelihood has converged in ~s0565860/projects/em_scheduling/. Submitting the DAG in submit/converge.dag will start things off - everything is still pointing to my work though so just look rather than run! I'll update when there's a publicly runnable version. There is also the issue that when another job appears in the DAG after the while loop, it doesn't run because the loop body appears to have failed.

A for loop is demonstrated in submit/for_test.dag in the same directory - it makes use of scripts/unroll.pl. The managing condor_dagman job doesn't seem to die though, so the solution is ready to use just yet.

Local GMTK tools

Checkout a copy from the CVS repository "gmtk_tools". Apply for access at https://www.inf.ed.ac.uk/systems/cvs/new/

Edit | Attach | Print version | History: r9 < r8 < r7 < r6 < r5 | Backlinks | Raw View | Raw edit | More topic actions...
Topic revision: r8 - 15 Feb 2006 - 15:41:13 - Main.s0565860
 
This site is powered by the TWiki collaboration platformCopyright © by the contributing authors. All material on this collaboration platform is the property of the contributing authors.
Ideas, requests, problems regarding TWiki? Send feedback
This Wiki uses Cookies