* First source release!
* Added DeepVECPT, DeepCPT, and gmtkDMLPtrain to support
hybrid dynamic graphical models / deep neural network
* Added gmtkOnline to support filtering & smoothing
* Added option to print Viterbi results according to original
sections
* Viterbi results can now be saved in any of the observation file
formats, as well a binary format for output with the new gmtkPrint
command
* Added separate Viterbi regular expression filters for the prolog,
chunk, and epilog. They can be applied to the original or modified
sections
* Added Viterbi output filters that use the decision tree leaf
node expression syntax and Viterbi output run-length compression
* Added option to save clique posteriors to any of the observation
file formats
* New constant space observation code supports island algorithm
to do true log space inference
* The island algorithm supports arbitrary roots as the logarithm base
* Added support for HDF5 observation files
* Added option to left-, right- or center-justify the observation
frames used when the number of usable frames is less than the number
of frames in a segment
* Added an option to skip zero probability segments rather than
terminating the GMTK command
* Added support for external compiled decision trees
* FNGram and NGram CPTs can now be read from compressed files
* You can now specify different verbosity levels for each component
(inference, I/O, training, etc.)
* Many bug fixes
2010-08-08
* speedups: Various speedups and bug fixes.
* island: new implementation of island algorithm and data
structures, it is much cleaner now.
* pruning: New pruning options: diversity/cluster pruning, and
predictive pruning (the predictie pruning stuff is published in
interspeech 2010). Diversity pruning has not yet been published or
described.
* significant speedups to Gaussian evaluation.
* random variables can now have symbol tables.
* gmtkViterbiNew is deprecated, gmtkViterbi is the new gmtkViterbi
program. As a result, the output file format is significantly
changed.
* added and changed quite a bit of code since the previous
version. Essentially re-wrote the core data structures and
unrolling strategy so that now, there is no long delay waiting for
STL to unroll a large graph - now there is no unrolling at all,
and the only thing that happens is data structures are allocated
as we go forward in the collect evidence phase.
2008-11-21
* added: lots has happened since 05, but this file unfortunately
has not been kept up to date. I hope to keep it up to date now.
2005-08-28
* added -cusample option, uniformaly at random sample from pruned
clique entries during CE.
2005-08-27
* new -deterministiChildrenStore option to inference programs, to
optionally not store determinisitc variables in cliques. Saves memory
but 10-20% slower.
2005-06-20
* added new beam options, -cmbeam, -cmmin
* new arguments ability (makes it easy to change arguments in one go)
2005-05-31
* added HTK lattice support (Gang Ji)
* added faster hash tables.
* E_root_clique now is maxweight clique.
2005-04-21
* fixed bug with static (1 frame) graphs on E1 partition.
2005-03-26
* fixed bug with dispositions not being re-sorted with variables in
clique.
* removed extraneous var in clique packer structure.
2005-03-09
* gmtkTime supports file line version of vcap,vpap,jcap,icap, and
trifile
2005-03-08
* speedup: faster Gaussian evaluation
* cpbeam applies to separators as well as clique
* started ChangeLog