From sheila_reynolds-0lzFnAlw030Aa7HRyrKG0wuBvc1PGuF@mailblocks.com Fri Nov  5 07:14:44 2004
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id iA5FEiQP006420
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@ssli.ee.washington.edu>; Fri, 5 Nov 2004 07:14:44 -0800
Received: from smtp1.mailblocks.com (smtp1.mailblocks.com [140.174.9.69])
	by seagull.ee.washington.edu (8.13.1/8.13.0) with SMTP id
	iA5FEg4X003121
	for <gmtk-users@ssli.ee.washington.edu>; Fri, 5 Nov 2004 07:14:42 -0800
Received: (qmail 8156 invoked from network); 5 Nov 2004 15:14:41 -0000
Received: from 10.10.0.51 (HELO mailblocks.com) (10.10.0.51)
	by 10.10.0.7 with SMTP; 5 Nov 2004 15:14:41 -0000
Date: Fri, 05 Nov 2004 07:15:35 -0800
Message-Id: <sheila_reynolds-0lzFnAls030AkRXM83Biyx8Z1cQ2Gtn@mailblocks.com>
From: <sheila@ee.washington.edu>
References: <416B2D44.9050605@ssli.ee.washington.edu>
Received: from 207.229.102.132 by app2.mailblocks.com (192.168.225.51) with
	HTTP (Mailblocks); Fri, 05 Nov 2004 07:15:35 -0800
X-MB-Message-Source: WebUI
X-MB-Message-Type: User
In-Reply-To: <416B2D44.9050605@ssli.ee.washington.edu>
Content-Type: text/plain; charset="us-ascii"; format=flowed
MIME-Version: 1.0
To: gmtk-users@ssli.ee.washington.edu
X-Priority: 3
X-Spam-Status: No, hits=-16.4 required=5.0 tests=AWL,MSGID_FROM_MTA_HEADER,
	NO_REAL_NAME,PRIORITY_NO_NAME autolearn=no version=2.64
X-Spam-Checker-Version: SpamAssassin 2.64 (2004-01-11) on 
	seagull.ee.washington.edu
Subject: [Gmtk-users] problem with new GMTK
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Fri, 05 Nov 2004 15:14:44 -0000

Hi all,

I had not updated my local version of GMTK for a while but decided that 
this morning would be a good time to do it.

I get a compile error when compiling GMTK_ObsSkmeans.cc (undefined 
reference to ObservationMatrix::openFiles) which I don't think I use so 
that's not really what I'm asking about.

When I try to re-triangulate a graph (that used to work fine) I get the 
following error :

Error: random variable 'tag' (template frame 0, line 29) specifies a 
parent 'word(0)' that has the wrong cardinality when unrolling template 
0 times (in unrolled network, variable at frame 0 asked for parent 
'word' at frame '0' which is has cardinality 0, but cardinality in 
template parent is 10827)


Any clues, anyone ?

thanks,

Sheila


From bilmes@cuba.ee.washington.edu Fri Nov  5 12:17:45 2004
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id iA5KHisa024065
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@crow.ee.washington.edu>; Fri, 5 Nov 2004 12:17:45 -0800
Received: from cuba.ee.washington.edu (cuba.ee.washington.edu [128.95.167.41])
	by seagull.ee.washington.edu (8.13.1/8.13.0) with ESMTP id
	iA5KHgGI016156 for <gmtk-users>; Fri, 5 Nov 2004 12:17:42 -0800
Message-Id: <200411052017.iA5KHgGI016156@seagull.ee.washington.edu>
To: gmtk-users@seagull.ee.washington.edu
X-Mailer: MH-E 7.82; nmh 1.0.4; GNU Emacs 21.2.1
Date: Fri, 05 Nov 2004 12:17:42 -0800
From: Jeff Bilmes <bilmes@ssli-mail.ee.washington.edu>
X-Spam-Status: No, hits=-0.8 required=5.0 tests=AWL, TO_MALFORMED autolearn=no 
	version=2.64
X-Spam-Checker-Version: SpamAssassin 2.64 (2004-01-11) on 
	seagull.ee.washington.edu
Cc: 
Subject: [Gmtk-users] stuff
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
Reply-To: "Jeff Bilmes \(206/221-5236\)" <bilmes@ee.washington.edu>
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Fri, 05 Nov 2004 20:17:45 -0000


Hi Everyone,

There hasn't been much traffic on this list for a while, but please
don't think that there aren't major gmtk improvements planned (I've just
been really busy). More significant speedups are on the way!!

I do, however, want to remind everyone that Evan has completed a fairly
stable version of gmtkViz, the gmtk visualization tool that already is
quite useful. If you check out the latest version, you can type 'make
gmtkViz' in the tksrc directory to compile it (at least on ssli/linux
for now).

Here is a quick list of non-obvious features:

  1) the program is not ment for creating graphs, rather it is for creating
     a nice picture of your graph once you have edited the .str file. It will
     create a position file for each structure file that saves the position
     of each node. 
  2) The position editing is quite good, its sort of a combination
     of some of the features in microsoft Powerpoint and Adobe Illustrator. It
     will display the three partitions (P,C,E) and any frames within those partitions.
  3) you can layout one frame/partition, and then copy the layout of one frame/partition
     to another frame/partition. This feature is really crucial, as it allows you to
     quickly layout even quite complicated graphs.
  4) the program can save to postscript, for inclusion in latex/msword documents.
  5) It has lots of options for changing edge/node color, arrow display, control-point size, etc. 
     Edges are rendered as splines with control points added to control curvature.

  A few non-obvious features (from Evan):
    a) Pushing the delete key deletes any selected control points on
       an edge. 
    b) Right-clicking a control point adds a control point to an edge.
    c) If a line has no control points (except its endpoints), clicking on the
       line adds a control point.

I've already used it a few times and found it a great tool to display a graph
you want to explain to someone. Also, its a good way to check that your graph
does what you want. 

Please try it out, and if you find bugs, please send them to gmtk-bugs
(but note that Evan is now gone, so the bugs won't be fixed until we can
find a replacement).

A couple of known problems:
  1) if you read in a graph that doesn't parse, the entire program will die. 
  2) sometimes the program crashes for no apparent reason (not as bad as it sounds).

Therefore, save your work OFTEN when using it!


			-- Jeff



+======================================================================+
|            Jeff A. Bilmes, Assistant Professor                       |
| Dept. of EE              Voice: (206) 221-5236                       |
| University of Washington FAX: (206) 543-3842                         |
| Box 352500               Email: bilmes@ee.washington.edu             |
| Seattle, WA  98195-2500  http://www.ee.washington.edu/faculty/bilmes |
+======================================================================+

From bilmes@cuba.ee.washington.edu Fri Dec 24 09:10:29 2004
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id iBOHATC1001994
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK);
	Fri, 24 Dec 2004 09:10:29 -0800
Received: from cuba.ee.washington.edu (cuba.ee.washington.edu [128.95.167.41])
	by seagull.ee.washington.edu (8.13.1/8.13.0) with ESMTP id
	iBOHARKZ022073; Fri, 24 Dec 2004 09:10:27 -0800
Message-Id: <200412241710.iBOHARKZ022073@seagull.ee.washington.edu>
To: gmtk-users@seagull.ee.washington.edu
X-Mailer: MH-E 7.82; nmh 1.0.4; GNU Emacs 21.3.1
Date: Fri, 24 Dec 2004 09:10:27 -0800
From: Jeff Bilmes <bilmes@ssli-mail.ee.washington.edu>
X-Spam-Status: No, hits=-3.0 required=5.0 tests=AWL,BAYES_00,TO_MALFORMED 
	autolearn=ham version=2.64
X-Spam-Checker-Version: SpamAssassin 2.64 (2004-01-11) on 
	seagull.ee.washington.edu
Cc: mukundn@seagull.ee.washington.edu
Subject: [Gmtk-users] GMTK dev tag updated
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
Reply-To: "Jeff Bilmes \(206/221-5236\)" <bilmes@ee.washington.edu>
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Fri, 24 Dec 2004 17:10:29 -0000


Major changes:

**  well-defined template restrictions. Now insists on:

    In the left interface case (with M=S=1), we must have:

      [P | C E] == [P | C C E] == [P C | C E]

    What this means is this: The vertical bar '|' cuts the graph (via edges)
    into a left and a right portion.  In the left interface case, the nodes
    on the right of the edge cut must be the same (relatively) with respect
    to each other (i.e., they are the same nodes, by at most a shift by
    S*numframesin(C) frames).  Note that when P is empty, we don't need to
    check the interface at all since we are guaranteed E does not reach to
    the left beyond one C, so there is only one total interface, The C-C
    interface, in C | C E.

    In the right interface case (with M=S=1), we must have:

       [P C | E] == [P C C | E] == [P C | C E]

    Here, we check that the nodes on the *left* of the cut are relatively
    the same.

    The above is now the only graphical restriction paced on the
    template (other than the obvious no directed cycles).  This is done
    so that the boundary provided by boundary algorithm can be validly
    used to compute the interface for the P-C boundary, the C-C
    boundary, and the C-E boundary.

    As an example, we can now have a variable in E ask for a parent in,
    say, P, if the above restrictions are followed (which might means
    that C has to ask for the same variable in P, and that the variable
    in P being asked for also exists in C).

    Just in case, here is a complete picture of the above, for arbitrary M and S.
    Left interface:
	    P | C(1) C(2) ... C(M) E
	    P | C(1)  C(2)   ...     C(M+S) E
	    P  C(1) ...  C(S) | C(S+1) ... C(M+S) E
    Right interface:
	    P C(1) C(2) ... C(M) | E
	    P C(1)  C(2)   ...     C(M+S) | E
	    P C(1) ...  C(S) | C(S+1) ... C(M+S) E

    Therefore, you can increase M and/or S to relax further the restriction above
    (but at the cost of more restricted length T, see below).


** Now officially supports higher-order Markov chains (i.e. variables
   with parents -2, +2, -3, +3, N-chunk-crossing, N>1, etc. as long as
   the template abides by the restrictions mentioned above). You will
   need to retriangulate to get this working (i.e., the trifile is what
   the problem was before).

   Please test this out and let me know if you find any problems!!!
  
** As a consequence, we can now use the case of unrolling by zero (i.e.,
   using the basic .str template). From the basic user template partitions of
   P, C, and E with frame lengths p, c, and e, this says that we may
   process T frames, where
            T = p + (M+kS)*c + e
   where k>=0 (before the requirement was k>=1). 

   This also means that GMTK can be used as a static network inference. I.e.,
   just make P empty, put your graph in C E, do a quick triangulation of C' (say
   completed), and do a careful triangulation of E'.

   (note: soon, we'll also be supporting disconnected networks so that you can
    put the static network in C as an option, and have neither a P nor an E, right
    now you need at least one of P and C).

   Also note: as a consequence, on positive graphs, you'll also see a space
   reduction by about one half on the C storage. On mixed stochastic/deterministic
   graphs you might see a small space saving.


Minor changes:

** More/updated comments and error messages

** Compiles again on cygwin. Should compile on both cygwin and linux
   without code change (hopefully:)

** gmtkTime supports a '-times' argument in non-multitest mode, to
   run a timing multiple times.



			-- Jeff



+======================================================================+
|            Jeff A. Bilmes, Assistant Professor                       |
| Dept. of EE              Voice: (206) 221-5236                       |
| University of Washington FAX: (206) 543-3842                         |
| Box 352500               Email: bilmes@ee.washington.edu             |
| Seattle, WA  98195-2500  http://www.ee.washington.edu/faculty/bilmes |
+======================================================================+

From bilmes@cuba.ee.washington.edu Mon Jan  3 23:41:45 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j047fjLQ025584
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@crow.ee.washington.edu>; Mon, 3 Jan 2005 23:41:45 -0800
Received: from cuba.ee.washington.edu (cuba.ee.washington.edu [128.95.167.41])
	by seagull.ee.washington.edu (8.13.1/8.13.0) with ESMTP id
	j047fixH026081 for <gmtk-users>; Mon, 3 Jan 2005 23:41:44 -0800
Message-Id: <200501040741.j047fixH026081@seagull.ee.washington.edu>
To: gmtk-users@seagull.ee.washington.edu
X-Mailer: MH-E 7.82; nmh 1.0.4; GNU Emacs 21.3.1
Date: Mon, 03 Jan 2005 23:41:44 -0800
From: Jeff Bilmes <bilmes@ssli-mail.ee.washington.edu>
X-Spam-Status: No, hits=-0.5 required=5.0 tests=AWL, BETTERMEMORY,
	TO_MALFORMED autolearn=no version=2.64
X-Spam-Checker-Version: SpamAssassin 2.64 (2004-01-11) on 
	seagull.ee.washington.edu
Cc: 
Subject: [Gmtk-users] dev tag update
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
Reply-To: "Jeff Bilmes \(206/221-5236\)" <bilmes@ee.washington.edu>
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Tue, 04 Jan 2005 07:41:45 -0000


Folks,

Two new GMTK dev-tag updates of mentionable worth are now in.

Summary:

1) Re-written Decision Trees: new DTs are now much faster and take less memory.
   Also, depending on how you write your DT, things can get faster still.

2) Virtual Evidence Separators: Can speed up exact inference when you've got
   a stochastic graph with some constraints. 

Extended:
============================================================================
============================================================================
1) Re-written Decision Trees

At long last, I've re-done the code for the dreaded decision tree
non-leaf nodes.

Interestingly, I've found that this change gives anywhere from a 3 to 30% inference
speedup on the graphs (depending on how much and how determinism is
specified).  They also use significantly less memory, and take much less
time to load in (for some DTs, reading time goes from 1/2-hour to
seconds). It should now be possible to use some of those big LVCSR DTs
that people have wanted to use.

Basically, the reason for the speedup is better memory management and
implementation of DT queries. The new version also supports a new syntax
for DTs which is much faster (especially to read).  All old DTs should
still work as is however (and with a speedup). Basically, whenever you
have a construct like:


  0 16 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 default
     -1 50
     -1 39
     etc.

This *ascending* sequential consecutive order of ints will be fast since it will use
direct indexing. Note that the order right now needs to be both
*ascending*, *consecutive* and *integer* (no individual ranges) to get
direct indexing. In fact, there is now a short cut for this that you can
use:

  0 16 0 ... 14 default
     -1 50
     -1 39
     etc.

If you've got any long consecutive strings of numbers, (and some DTs
I've seen have many 10s of thousands of such numbers), this will load in
faster and take *significantly* less memory. Also, the first position
can be anything, as long as the rest are ascending and consecutive.
I.e., you can do:

  0 16 5 ... 19 default
     -1 50
     -1 39
     etc.


which will have the same speed.

Put another way, it will be beneficial to whenever possible express your
DTs in this way. If the sets are all integers but the order is not
ascending or they ints are not sequential, you'll still get a speedup
since now a hash table is used to do the mapping (rather than
the old way of binary searches with ranges).

In other words, this:

  0 4 0 1 2 default
     -1 50
     -1 39
     etc.

will be the same speed as this:

  0 4 0 ... 2 default
     -1 50
     -1 39
     etc.

which will be faster than this (uses a hash table):

  0 4 1 0 2 default
     -1 39
     -1 50
     etc.

which will be faster than this (which uses int range objects):

  0 4 1:1 0:0 2:0 default
     -1 39
     -1 50
     etc.

I.e., the first two use direct mappings, the 2nd uses a hash table, and
the 3rd case uses the more general range objects.

Indeed, more smarts can be done to do things like detect when a list of
ranges is really a ascending consecutive list, or when a hash could be
the same. Perhaps someday this will be done, but for now at least if you
know what you are doing you can get the speedup and memory savings
(which can be substantial. On some LVCSR graphs, the old DTs alone were
about 500MB RSS, now they will probably be only about 10-20MB RSS).

============================================================================
============================================================================
2) Virtual Evidence Separators.

GMTK now supports "Virtual Evidence Separators". Basically, whenever
GMTK finds a case of either:

case 1) an child observed = 1 and deterministically related to parents
        (and if there is no switching, iterable-ness, and if the child
        is constant immediate observed, meaning the observation value
        has to come from the str file, and can't be any of the
        "variable" keywords such as "frame", etc.). I.e.,

      P1 P2 ... PN
        \ \  / /     
           C=1

case 2) a grandchild observed = 1, randomly related to a child (the
        parent of grandchild), and if the child is deterministically
        related to parents (and otherwise the same conditions as (1)
        hold). I.e.,

          P1 P2 ... PN
           \ \  / /          
               C
               |
               v
              G=1

     Note that C = f(P_{1:N}) (i.e., deterministic), and G = random(C) but
     random() can't be an iterable function (one that changes from utterance
     to utterance) nor can random() be a VECPT (which is essentially iterable
     since it changes from frame to frame).

Then it GMTK can optionally use VE separators to deal with the implicit
constraints imposed on the parents by the (case 1) child or (case 2)
grandchild.

To do this, GMTK offline iterates through all parent values and builds
a table of which parents satisfy the child (or grant child) with non-zero
probability. Then when iterating through the variables, it iterates
through all satisfying parents simultaneously (since the are dependent
conditioned on the child (or grandchild)) rather than separately
like in the old version.

This set of parent values is either computed anew once each time GMTK is
run, or is additionally optionally stored on disk so that next time GMTK
runs it will load rather than re-generate the tables.  The reason for
the disk file is that when there are many parent values that need to be
checked, computing this table can take a long time.

Note you are not guaranteed that your graphs have any VE seps in them
(it might only benefit for some graphs), and even if your graph is
"VE-seperable", speedups aren't guaranteed. In some cases I've seen a
speedup but even then it wasn't big. The speedup you get I think will be
significantly influenced by the triangulation, so soon to VEsep or not
will be integrated into the triangulation search step.

The default behavior of GMTK currently has veseps turned off.  The
relevant programs options are:

-useVESeparators
  Use Virtual Evidence (VE) Separators (if any are available) during inference.
  The option is "-useVESeparators n", where n = 0x0, 0x1, 0x2, or 0x3

  The option is a 2-bit bitmask on what kind of VE sep to use. 0x1
  corresponds to the direct child (parents and child), while 0x2
  corresponds to parents, a deterministic child, and a random
  grandchild (see above)


-veSepWhere
   Where to use VE seps. Bitwise or of 0x1 (P), 0x2 (C), 0x4 (E)
   Usage is "-veSepwhere n" where n = 0x0 ... 0x7, and the
   argument is a 3-bit bitmask saying where (if any are available) the veseps
   should be used. 0x1 is P, 0x2 is in C, and 0x4 is in E. 

   If either -useVESeparators or -veSepWhere is 0, you turn off VEseps.

-veSepFileName
   Name of VE separators file to store VE sep/read previous VE sep info.
   There is a default file name, and if you use ve seps it will right now
   always write to this file.

-veSepRecompute
   Force a re-compute of VE separator information. This is necessary when
   you change some of the options above. I.e., if you decide to use VEseps
   in E, you'll need to recompute the file (otherwise you'll get an error
   message saying that the vesep file doesn't match the current configuration)

-veSepLogProdCardLimit
   The log (base 10) upper limit on a VE sep variable cardinality product.
   As mentioned above, sometimes generating the VEsep tables can take a while.
   This option limits VE seps to be only those where the (base 10) log of the product of the 
   cardinalities of the parents is below some threshold, and the threshold
   is given by this arg.

   Note that VEtable computation can be significantly sped up by using a
   separate structure file to use just one time for table generatio. In this
   structure, some of the parents should be observed to the values
   that you know are guaranteed to be the ones that satisfy the child (i.e.,
   if all satisfying parents have say the third parent == 5, the special
   structure file should have that parent observed to be 5). 

   In still other words, Lets say you're in case 1 above. If there are
   any parents that have only one possible value that (along with other
   parent values) will explain C=1, then make those parents observed in
   this extra .str file. That way the exact same set of constraints will
   be generated in the table.

   Once the tables are generated and stored to disk, go back to the old
   str file.

Currently, a RV child that is all deterministic and uses switching will
not constitute a VEsep, but that will be added in a couple of weeks.

Best,



			-- Jeff



+======================================================================+
|            Jeff A. Bilmes, Assistant Professor                       |
| Dept. of EE              Voice: (206) 221-5236                       |
| University of Washington FAX: (206) 543-3842                         |
| Box 352500               Email: bilmes@ee.washington.edu             |
| Seattle, WA  98195-2500  http://www.ee.washington.edu/faculty/bilmes |
+======================================================================+

From bilmes@cuba.ee.washington.edu Mon Jan  3 23:56:42 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j047ugtl026456
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@crow.ee.washington.edu>; Mon, 3 Jan 2005 23:56:42 -0800
Received: from cuba.ee.washington.edu (cuba.ee.washington.edu [128.95.167.41])
	by seagull.ee.washington.edu (8.13.1/8.13.0) with ESMTP id
	j047ufXP026612 for <gmtk-users>; Mon, 3 Jan 2005 23:56:41 -0800
Message-Id: <200501040756.j047ufXP026612@seagull.ee.washington.edu>
To: gmtk-users@seagull.ee.washington.edu
X-Mailer: MH-E 7.82; nmh 1.0.4; GNU Emacs 21.3.1
Date: Mon, 03 Jan 2005 23:56:41 -0800
From: Jeff Bilmes <bilmes@ssli-mail.ee.washington.edu>
X-Spam-Status: No, hits=-0.7 required=5.0 tests=AWL, TO_MALFORMED autolearn=no 
	version=2.64
X-Spam-Checker-Version: SpamAssassin 2.64 (2004-01-11) on 
	seagull.ee.washington.edu
Cc: 
Subject: [Gmtk-users] another change dev tag update
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
Reply-To: "Jeff Bilmes \(206/221-5236\)" <bilmes@ee.washington.edu>
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Tue, 04 Jan 2005 07:56:42 -0000


Oh, I forgot to mention that the debugging output has been cleaned up
and more information is displayed in a more concise way. I.e., options
of interest are:

  -verb 50  - just shows message passings in JT.
  -verb 60  - shows above + clique insertions
  -verb 65  - show above + iteration starts (including separator, unassigned, & assigned starts).
  -verb 70  - show above + all iterations, parent values summarized 
  -verb 80  - show above but with parent values explicitly shown.


			-- Jeff



+======================================================================+
|            Jeff A. Bilmes, Assistant Professor                       |
| Dept. of EE              Voice: (206) 221-5236                       |
| University of Washington FAX: (206) 543-3842                         |
| Box 352500               Email: bilmes@ee.washington.edu             |
| Seattle, WA  98195-2500  http://www.ee.washington.edu/faculty/bilmes |
+======================================================================+

From karim@cs.washington.edu Thu Jan  6 14:46:52 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j06MkqeR030491
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@ssli.ee.washington.edu>; Thu, 6 Jan 2005 14:46:52 -0800
Received: from mx4.cs.washington.edu (mx4.cs.washington.edu [128.208.4.190])
	by seagull.ee.washington.edu (8.13.1/8.13.0) with ESMTP id
	j06MkpS7019332
	(version=TLSv1/SSLv3 cipher=EDH-RSA-DES-CBC3-SHA bits=168 verify=OK)
	for <gmtk-users@ssli.ee.washington.edu>; Thu, 6 Jan 2005 14:46:51 -0800
Received: from ernesto.cs.washington.edu (ernesto.cs.washington.edu
	[128.208.4.36])
	by mx4.cs.washington.edu (8.12.11/8.12.11/1.4) with ESMTP id
	j06Mkp9E007405
	(version=TLSv1/SSLv3 cipher=EDH-RSA-DES-CBC3-SHA bits=168 verify=NO)
	for <gmtk-users@ssli.ee.washington.edu>; Thu, 6 Jan 2005 14:46:51 -0800
	(envelope-from karim@cs.washington.edu)
Received: from localhost (karim@localhost)
	by ernesto.cs.washington.edu (8.13.1/8.13.1/Submit/1.7) with ESMTP id
	j06MkpOp011157; Thu, 6 Jan 2005 14:46:51 -0800
	(envelope-from karim@cs.washington.edu)
X-Authentication-Warning: ernesto.cs.washington.edu: karim owned process doing
	-bs
Date: Thu, 6 Jan 2005 14:46:51 -0800 (PST)
From: Karim Filali <karim@cs.washington.edu>
To: gmtk-users@ssli.ee.washington.edu
Message-ID: <Pine.LNX.4.58L.0501060846090.8571@ernesto.cs.washington.edu>
MIME-Version: 1.0
Content-Type: TEXT/PLAIN; charset=US-ASCII
X-Spam-Status: No, hits=-9.9 required=5.0 tests=AWL autolearn=no version=2.64
X-Spam-Checker-Version: SpamAssassin 2.64 (2004-01-11) on 
	seagull.ee.washington.edu
Cc: Karim Filali <karim@cs.washington.edu>
Subject: [Gmtk-users] "New" GMTK observation file options summary
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Thu, 06 Jan 2005 22:46:53 -0000

Summary of the "new" observation options in GMTK:
------------------------------------------------


I) Length adjustment:
--------------------
a) [-fdiffactX <str>] Automatically adjust segment lengths: 

   The following options are supported to adjust the length of
   segments so that they match across all supplied observation files:
    
     <str> =
      er : report an error (the default)
      rl : repeat the last frame of shorter segments to match the
      length of the longest corresponding segment across all
      observation input files
      rf : repeat the first frame
      se : expand segmentally i.e. repeat each frame of the shorter
      segments uniformely to match the length of the longest one 
      ts : truncate longer segments from the beginning so that their
      length matches the length of the shortest corresponding segment
      across all observation input files
      te : truncate from the end

b) [-sdiffactX <str>] Automatically adjust file lengths: 

   The following options are supported to adjust the length of
   input files so that they are the same:

     <str> =
      er : report an error
      te : truncate longer files from the end (the default)
      rl : repeat the last segment of shorter files
      wa : wrap around i.e. shorter files will cycle back to the
      beginning when a segment number larger than the size of the file
      is requested.


These length adjustments take effect after all per-file transformations
are applied and before any global transformation or feature
combination is applied.


II) Transformations:
------------------

The following transformations are supported.  They are listed in the
order in which they take effect:

a) [-preprX str] Per-stream ("stream" and "file" are used
interchangeably.) frame range: select a range of frames in each
segment of observation file number X (to achieve downsampling for
example or remove parts of the data)

b) [-transX str] Per-stream transformations string.  They are applied
to float features only.

   A transformation string has the following format

   TRANS1[<float>|<int>|<@filename>][_TRANS2[<float>|<int>|<@filename>]][...]

   where TRANSN can take the following values

   X :        empty transformation.  Don't do anything.
   O<float>:  add a constant offset to all features.  Example: O3.4
   M<float>:  multiply by a constant.  Example: M12.5
   N:         perform mean and variance normalization at the segment level
   E:         perform mean substraction only
   F<@filename>: Apply the filter in filename.  The filter is a vector
   of floats.  Example F@"filter_file"
   UH<int>:   upsample-hold by <int>.  Example: UH2 repeats each frame twice
   US<int>:   upsample-smooth by <int>.  Example: US1 repeats each frame once
   and performs smoothing
   R<int>:    applies an ARMA filter of order <int>.  Example: R3

c) [-postprX str] Per-stream frame range after the -transX
transformations are applied

--->  Length adjustments go here  <---

--->  Feature combinations (see below) go here <--- 


d) [-posttrans str] Final transformation string.  Same format as the
per-stream transformation except it is applied over the global
observation matrix.

e) [-gpr str]  Global final frame range


III) Feature combinations
--------------------------
TWO streams can be combined feature-wise using the option
[-comb str].  Only float features are affected.
The following operations are supported: addition (add), substraction
(sub), multiplication (mul), and division (div).

WARNING: The behavior of this feature when more than two streams are
combined is not well defined.  Same thing for division by zero.

The default is to allow the combination of two streams that have a
different number of float features.  The shorter stream is padded with
zeros.  This can be changed by defining
ALLOW_VARIABLE_DIM_COMBINED_STREAMS to be 0 in
GMTK_ObservationMatrix.h

--------------------------------------------------------------------


Examples:
---------

> obs-print -i1 file1 -ifmt1 ascii -nf1 3 -ni1 2

0 0 0.000000 1.100000 2.200000 1 0
0 1 1.000000 10.100000 23.200001 1 1
0 2 2.000000 100.099998 19.200001 6 7
1 0 33.000000 1.100000 -7.000000 10 4
1 1 -1.000000 -2.100000 23.200001 2 2
2 0 400.000000 71.099998 98.000000 10 4
2 1 -1.000000 0.100000 -56.200001 66 100
2 2 0.040000 0.031000 19.200001 0 9

> obs-print -i1 file2 -ifmt1 ascii -nf1 2 -ni1 2

0 0 555.000000 5763.986816 6 7
0 1 -1.000000 0.100000 334 87
0 2 -0.003400 17.000000 99 56
0 3 -986.500000 88.000000 23 8
1 0 6.000000 100.900002 22 8
1 1 0.900000 0.100000 8 7
1 2 10.900000 9.000000 1 8


*** Unequal segment lengths ***

> obs-print -i1 file1 -ifmt1 ascii -nf1 3 -ni1 2 -i2 file2 -ifmt2
  ascii -nf2 2 -ni2 2

ERROR: The number of frames for segment 0 is not the same for
  observation files 'file1' and 'file2' (3 vs. 4).  Use the -fdiff
  option.

> obs-print -i1 file1 -ifmt1 ascii -nf1 3 -ni1 2 -i2 file2 -ifmt2
  ascii -nf2 2 -ni2 2 -fdiff1 rl

0 0 0.000000 1.100000 2.200000 555.000000 5763.986816 1 0 6 7
0 1 1.000000 10.100000 23.200001 -1.000000 0.100000 1 1 334 87
0 2 2.000000 100.099998 19.200001 -0.003400 17.000000 6 7 99 56
0 3 2.000000 100.099998 19.200001 -986.500000 88.000000 6 7 23 8
1 0 33.000000 1.100000 -7.000000 6.000000 100.900002 10 4 22 8
1 1 -1.000000 -2.100000 23.200001 0.900000 0.100000 2 2 8 7
1 2 -1.000000 -2.100000 23.200001 10.900000 9.000000 2 2 1 8

*** Unequal file lengths  ***

By default the longer file is truncated (see above:  result has 2
segments but file1 has 3)

> obs-print -i1 file1 -ifmt1 ascii -nf1 3 -ni1 2 -i2 file2 -ifmt2
  ascii -nf2 2 -ni2 2 -fdiff1 rl -sdiff1 wa -sdiff2 wa

0 0 0.000000 1.100000 2.200000 555.000000 5763.986816 1 0 6 7
0 1 1.000000 10.100000 23.200001 -1.000000 0.100000 1 1 334 87
0 2 2.000000 100.099998 19.200001 -0.003400 17.000000 6 7 99 56
0 3 2.000000 100.099998 19.200001 -986.500000 88.000000 6 7 23 8
1 0 33.000000 1.100000 -7.000000 6.000000 100.900002 10 4 22 8
1 1 -1.000000 -2.100000 23.200001 0.900000 0.100000 2 2 8 7
1 2 -1.000000 -2.100000 23.200001 10.900000 9.000000 2 2 1 8
2 0 400.000000 71.099998 98.000000 555.000000 5763.986816 10 4 6 7
2 1 -1.000000 0.100000 -56.200001 -1.000000 0.100000 66 100 334 87
2 2 0.040000 0.031000 19.200001 -0.003400 17.000000 0 9 99 56
2 3 0.040000 0.031000 19.200001 -986.500000 88.000000 0 9 23 8


*** Transformations ***

> obs-print -i1 file1 -ifmt1 ascii -nf1 3 -ni1 2 -i2 file2 -ifmt2 ascii
-nf2 2 -ni2 2 -fdiff1 rl -sdiff1 wa -sdiff2 wa -pre1 0:1 -trans2
"N_M2.7_O5_F@filter_US1" -postpr2 0:3 -posttrans N_O10 -gpr 0:1

0 0 8.500000 8.500000 8.500000 10.603625 11.139734 1 0 6 7
0 1 10.500000 10.500000 10.500000 10.498142 10.397820 1 1 6 7
1 0 11.500000 11.500000 8.500000 10.845489 11.307186 10 4 22 8
1 1 9.500000 9.500000 10.500000 10.476808 10.257824 2 2 22 8
2 0 11.500000 11.500000 11.500000 10.603625 11.139734 10 4 6 7
2 1 9.500000 9.500000 9.500000 10.498142 10.397820 66 100 6 7

where the file file "filter" contains the line
1.2 2.4 1.2
 

*** Combinations  ***

> obs-print -i1 file1 -ifmt1 ascii -nf1 3 -ni1 2 -i2 file2 -ifmt2
  ascii -nf2 2 -ni2 2 -fdiff1 rl -sdiff1 wa -sdiff2 wa -pre1 0:1
  -trans2 "N_M2.7_O5_F@filter_US1" -postpr2 0:3 -posttrans N_O10 -gpr
  0:1 -comb add

0 0 10.494615 10.692422 8.500000 1 0 6 7
0 1 10.556338 10.830382 10.500000 1 1 6 7
1 0 11.481083 11.348219 8.500000 10 4 28 15
1 1 9.651711 10.170420 10.500000 2 2 28 15
2 0 11.499442 11.481358 11.500000 10 4 34 22
2 1 9.521022 9.694231 9.500000 66 100 34 22


From bilmes@cuba.ee.washington.edu Thu Jan  6 14:54:44 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j06MsioK031152
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@crow.ee.washington.edu>; Thu, 6 Jan 2005 14:54:44 -0800
Received: from cuba.ee.washington.edu (cuba.ee.washington.edu [128.95.167.41])
	by seagull.ee.washington.edu (8.13.1/8.13.0) with ESMTP id
	j06Msg7i019822 for <gmtk-users>; Thu, 6 Jan 2005 14:54:42 -0800
Message-Id: <200501062254.j06Msg7i019822@seagull.ee.washington.edu>
To: gmtk-users@seagull.ee.washington.edu
Subject: Karim Filali: [Gmtk-users] "New" GMTK observation file options summary
X-Mailer: MH-E 7.82; nmh 1.0.4; GNU Emacs 21.3.1
MIME-Version: 1.0
Content-Type: multipart/mixed; boundary="=-=-="
Date: Thu, 06 Jan 2005 14:54:42 -0800
From: Jeff Bilmes <bilmes@ssli-mail.ee.washington.edu>
X-Spam-Status: No, hits=-0.7 required=5.0 tests=AWL, TO_MALFORMED autolearn=no 
	version=2.64
X-Spam-Checker-Version: SpamAssassin 2.64 (2004-01-11) on 
	seagull.ee.washington.edu
Cc: 
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
Reply-To: "Jeff Bilmes \(206/221-5236\)" <bilmes@ee.washington.edu>
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Thu, 06 Jan 2005 22:54:45 -0000

--=-=-=


Just in case this isn't clear, Karim has done a bunch of work updating
the observation file handling in GMTK as well as a bunch of new programs
(the obs-* programs for handling observation files, which can be any
combination of pfiles, htk, ascii, or binary files).

GMTK now supports the combining of an arbitrary (rather than 3) number
of files online. This is the capital "X" in the argument names. Also, it
can combine observations of different lengths and duplicate observations
when segments in the files have different lengths, so that you don't
have to. Therefore, "X" is a number, so "-iX" really means
"-i1", "-i2", etc.

Lastly, when you don't need it, many options are no longer required
(e.g., the number of floats being specified in a pfile).

More examples will be forthcoming.



--=-=-=
Content-Type: message/rfc822
Content-Disposition: attachment; filename=233
Content-Description: forwarded message

Replied: Thu, 06 Jan 2005 14:51:12 -0800
Replied: "Karim Filali <karim@cs.washington.edu> "
Return-Path: gmtk-users-bounces@ssli.ee.washington.edu
Delivery-Date: Thu Jan 06 14:46:58 2005
Delivery-Date: Thu, 06 Jan 2005 14:46:56 -0800
Received: from crow.ee.washington.edu (localhost [127.0.0.1])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j06MktUI030494; 
	Thu, 6 Jan 2005 14:46:55 -0800
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j06MkqeR030491
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@ssli.ee.washington.edu>; Thu, 6 Jan 2005 14:46:52 -0800
Received: from mx4.cs.washington.edu (mx4.cs.washington.edu [128.208.4.190])
	by seagull.ee.washington.edu (8.13.1/8.13.0) with ESMTP id
	j06MkpS7019332
	(version=TLSv1/SSLv3 cipher=EDH-RSA-DES-CBC3-SHA bits=168 verify=OK)
	for <gmtk-users@ssli.ee.washington.edu>; Thu, 6 Jan 2005 14:46:51 -0800
Received: from ernesto.cs.washington.edu (ernesto.cs.washington.edu
	[128.208.4.36])
	by mx4.cs.washington.edu (8.12.11/8.12.11/1.4) with ESMTP id
	j06Mkp9E007405
	(version=TLSv1/SSLv3 cipher=EDH-RSA-DES-CBC3-SHA bits=168 verify=NO)
	for <gmtk-users@ssli.ee.washington.edu>; Thu, 6 Jan 2005 14:46:51 -0800
	(envelope-from karim@cs.washington.edu)
Received: from localhost (karim@localhost)
	by ernesto.cs.washington.edu (8.13.1/8.13.1/Submit/1.7) with ESMTP id
	j06MkpOp011157; Thu, 6 Jan 2005 14:46:51 -0800
	(envelope-from karim@cs.washington.edu)
X-Authentication-Warning: ernesto.cs.washington.edu: karim owned process doing
	-bs
Date: Thu, 6 Jan 2005 14:46:51 -0800 (PST)
From: Karim Filali <karim@cs.washington.edu>
To: gmtk-users@ssli.ee.washington.edu
Message-ID: <Pine.LNX.4.58L.0501060846090.8571@ernesto.cs.washington.edu>
MIME-Version: 1.0
Content-Type: TEXT/PLAIN; charset=US-ASCII
X-Spam-Status: No, hits=-9.9 required=5.0 tests=AWL autolearn=no version=2.64
X-Spam-Checker-Version: SpamAssassin 2.64 (2004-01-11) on 
	seagull.ee.washington.edu
Cc: Karim Filali <karim@cs.washington.edu>
Subject: [Gmtk-users] "New" GMTK observation file options summary
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
Sender: gmtk-users-bounces@ssli-mail.ee.washington.edu
Errors-To: gmtk-users-bounces@ssli-mail.ee.washington.edu
X-IMAPbase: 1105051630 1
Status: O
X-Status: 
X-Keywords: 
X-UID: 1

Summary of the "new" observation options in GMTK:
------------------------------------------------


I) Length adjustment:
--------------------
a) [-fdiffactX <str>] Automatically adjust segment lengths: 

   The following options are supported to adjust the length of
   segments so that they match across all supplied observation files:
    
     <str> =
      er : report an error (the default)
      rl : repeat the last frame of shorter segments to match the
      length of the longest corresponding segment across all
      observation input files
      rf : repeat the first frame
      se : expand segmentally i.e. repeat each frame of the shorter
      segments uniformely to match the length of the longest one 
      ts : truncate longer segments from the beginning so that their
      length matches the length of the shortest corresponding segment
      across all observation input files
      te : truncate from the end

b) [-sdiffactX <str>] Automatically adjust file lengths: 

   The following options are supported to adjust the length of
   input files so that they are the same:

     <str> =
      er : report an error
      te : truncate longer files from the end (the default)
      rl : repeat the last segment of shorter files
      wa : wrap around i.e. shorter files will cycle back to the
      beginning when a segment number larger than the size of the file
      is requested.


These length adjustments take effect after all per-file transformations
are applied and before any global transformation or feature
combination is applied.


II) Transformations:
------------------

The following transformations are supported.  They are listed in the
order in which they take effect:

a) [-preprX str] Per-stream ("stream" and "file" are used
interchangeably.) frame range: select a range of frames in each
segment of observation file number X (to achieve downsampling for
example or remove parts of the data)

b) [-transX str] Per-stream transformations string.  They are applied
to float features only.

   A transformation string has the following format

   TRANS1[<float>|<int>|<@filename>][_TRANS2[<float>|<int>|<@filename>]][...]

   where TRANSN can take the following values

   X :        empty transformation.  Don't do anything.
   O<float>:  add a constant offset to all features.  Example: O3.4
   M<float>:  multiply by a constant.  Example: M12.5
   N:         perform mean and variance normalization at the segment level
   E:         perform mean substraction only
   F<@filename>: Apply the filter in filename.  The filter is a vector
   of floats.  Example F@"filter_file"
   UH<int>:   upsample-hold by <int>.  Example: UH2 repeats each frame twice
   US<int>:   upsample-smooth by <int>.  Example: US1 repeats each frame once
   and performs smoothing
   R<int>:    applies an ARMA filter of order <int>.  Example: R3

c) [-postprX str] Per-stream frame range after the -transX
transformations are applied

--->  Length adjustments go here  <---

--->  Feature combinations (see below) go here <--- 


d) [-posttrans str] Final transformation string.  Same format as the
per-stream transformation except it is applied over the global
observation matrix.

e) [-gpr str]  Global final frame range


III) Feature combinations
--------------------------
TWO streams can be combined feature-wise using the option
[-comb str].  Only float features are affected.
The following operations are supported: addition (add), substraction
(sub), multiplication (mul), and division (div).

WARNING: The behavior of this feature when more than two streams are
combined is not well defined.  Same thing for division by zero.

The default is to allow the combination of two streams that have a
different number of float features.  The shorter stream is padded with
zeros.  This can be changed by defining
ALLOW_VARIABLE_DIM_COMBINED_STREAMS to be 0 in
GMTK_ObservationMatrix.h

--------------------------------------------------------------------


Examples:
---------

> obs-print -i1 file1 -ifmt1 ascii -nf1 3 -ni1 2

0 0 0.000000 1.100000 2.200000 1 0
0 1 1.000000 10.100000 23.200001 1 1
0 2 2.000000 100.099998 19.200001 6 7
1 0 33.000000 1.100000 -7.000000 10 4
1 1 -1.000000 -2.100000 23.200001 2 2
2 0 400.000000 71.099998 98.000000 10 4
2 1 -1.000000 0.100000 -56.200001 66 100
2 2 0.040000 0.031000 19.200001 0 9

> obs-print -i1 file2 -ifmt1 ascii -nf1 2 -ni1 2

0 0 555.000000 5763.986816 6 7
0 1 -1.000000 0.100000 334 87
0 2 -0.003400 17.000000 99 56
0 3 -986.500000 88.000000 23 8
1 0 6.000000 100.900002 22 8
1 1 0.900000 0.100000 8 7
1 2 10.900000 9.000000 1 8


*** Unequal segment lengths ***

> obs-print -i1 file1 -ifmt1 ascii -nf1 3 -ni1 2 -i2 file2 -ifmt2
  ascii -nf2 2 -ni2 2

ERROR: The number of frames for segment 0 is not the same for
  observation files 'file1' and 'file2' (3 vs. 4).  Use the -fdiff
  option.

> obs-print -i1 file1 -ifmt1 ascii -nf1 3 -ni1 2 -i2 file2 -ifmt2
  ascii -nf2 2 -ni2 2 -fdiff1 rl

0 0 0.000000 1.100000 2.200000 555.000000 5763.986816 1 0 6 7
0 1 1.000000 10.100000 23.200001 -1.000000 0.100000 1 1 334 87
0 2 2.000000 100.099998 19.200001 -0.003400 17.000000 6 7 99 56
0 3 2.000000 100.099998 19.200001 -986.500000 88.000000 6 7 23 8
1 0 33.000000 1.100000 -7.000000 6.000000 100.900002 10 4 22 8
1 1 -1.000000 -2.100000 23.200001 0.900000 0.100000 2 2 8 7
1 2 -1.000000 -2.100000 23.200001 10.900000 9.000000 2 2 1 8

*** Unequal file lengths  ***

By default the longer file is truncated (see above:  result has 2
segments but file1 has 3)

> obs-print -i1 file1 -ifmt1 ascii -nf1 3 -ni1 2 -i2 file2 -ifmt2
  ascii -nf2 2 -ni2 2 -fdiff1 rl -sdiff1 wa -sdiff2 wa

0 0 0.000000 1.100000 2.200000 555.000000 5763.986816 1 0 6 7
0 1 1.000000 10.100000 23.200001 -1.000000 0.100000 1 1 334 87
0 2 2.000000 100.099998 19.200001 -0.003400 17.000000 6 7 99 56
0 3 2.000000 100.099998 19.200001 -986.500000 88.000000 6 7 23 8
1 0 33.000000 1.100000 -7.000000 6.000000 100.900002 10 4 22 8
1 1 -1.000000 -2.100000 23.200001 0.900000 0.100000 2 2 8 7
1 2 -1.000000 -2.100000 23.200001 10.900000 9.000000 2 2 1 8
2 0 400.000000 71.099998 98.000000 555.000000 5763.986816 10 4 6 7
2 1 -1.000000 0.100000 -56.200001 -1.000000 0.100000 66 100 334 87
2 2 0.040000 0.031000 19.200001 -0.003400 17.000000 0 9 99 56
2 3 0.040000 0.031000 19.200001 -986.500000 88.000000 0 9 23 8


*** Transformations ***

> obs-print -i1 file1 -ifmt1 ascii -nf1 3 -ni1 2 -i2 file2 -ifmt2 ascii
-nf2 2 -ni2 2 -fdiff1 rl -sdiff1 wa -sdiff2 wa -pre1 0:1 -trans2
"N_M2.7_O5_F@filter_US1" -postpr2 0:3 -posttrans N_O10 -gpr 0:1

0 0 8.500000 8.500000 8.500000 10.603625 11.139734 1 0 6 7
0 1 10.500000 10.500000 10.500000 10.498142 10.397820 1 1 6 7
1 0 11.500000 11.500000 8.500000 10.845489 11.307186 10 4 22 8
1 1 9.500000 9.500000 10.500000 10.476808 10.257824 2 2 22 8
2 0 11.500000 11.500000 11.500000 10.603625 11.139734 10 4 6 7
2 1 9.500000 9.500000 9.500000 10.498142 10.397820 66 100 6 7

where the file file "filter" contains the line
1.2 2.4 1.2
 

*** Combinations  ***

> obs-print -i1 file1 -ifmt1 ascii -nf1 3 -ni1 2 -i2 file2 -ifmt2
  ascii -nf2 2 -ni2 2 -fdiff1 rl -sdiff1 wa -sdiff2 wa -pre1 0:1
  -trans2 "N_M2.7_O5_F@filter_US1" -postpr2 0:3 -posttrans N_O10 -gpr
  0:1 -comb add

0 0 10.494615 10.692422 8.500000 1 0 6 7
0 1 10.556338 10.830382 10.500000 1 1 6 7
1 0 11.481083 11.348219 8.500000 10 4 28 15
1 1 9.651711 10.170420 10.500000 2 2 28 15
2 0 11.499442 11.481358 11.500000 10 4 34 22
2 1 9.521022 9.694231 9.500000 66 100 34 22

_______________________________________________
Gmtk-users mailing list
Gmtk-users@ssli.ee.washington.edu
https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users

--=-=-=--

From bilmes@cuba.ee.washington.edu Tue Jan 11 05:51:49 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j0BDpmfr006420
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@crow.ee.washington.edu>; Tue, 11 Jan 2005 05:51:49 -0800
Received: from cuba.ee.washington.edu (cuba.ee.washington.edu [128.95.167.41])
	by seagull.ee.washington.edu (8.13.1/8.13.0) with ESMTP id
	j0BDpmhC013377 for <gmtk-users>; Tue, 11 Jan 2005 05:51:48 -0800
Message-Id: <200501111351.j0BDpmhC013377@seagull.ee.washington.edu>
To: gmtk-users@seagull.ee.washington.edu
X-Mailer: MH-E 7.82; nmh 1.0.4; GNU Emacs 21.3.1
Date: Tue, 11 Jan 2005 05:51:47 -0800
From: Jeff Bilmes <bilmes@ssli-mail.ee.washington.edu>
X-Spam-Status: No, hits=-0.6 required=5.0 tests=AWL, TO_MALFORMED autolearn=no 
	version=2.64
X-Spam-Checker-Version: SpamAssassin 2.64 (2004-01-11) on 
	seagull.ee.washington.edu
Cc: 
Subject: [Gmtk-users] dev tag updated
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
Reply-To: "Jeff Bilmes \(206/221-5236\)" <bilmes@ee.washington.edu>
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Tue, 11 Jan 2005 13:51:49 -0000


Major changes:
  
- GMTK now supports arbitrary disconnected networks, where the network
  can either be disconnected within a partition, or across partitions.
  Also, we may now have empty P and/or E partitions, and disconnected
  C partitions (so if only a C is given, this can act as a static
  network that gets replicated as the observations get longer). A static
  now either be just a C (without having a P or an E), or can be
  a E'=CE, and unrolling by zero.  

- As a result of GMTK supporting disconnected networks, triangulations
   are now formed from UGMs that do not necessarily have observed
   parents connected to their children. This might significantly speedup
   some graphs where you have observed variables that are both children
   and parents.

-  Unfortunately, I changed the trifile format, since old format graph-ID had a bug,
   namely it didn't store the chunk information, nor the end of the graph (so    
   if you truncated frames in a graph that might have caused problems). We now use a special
   string at the end of the id. The tri-file information is still valid, so you can
   use the same trifiles after a quick by-hand edit. 

   To go from the old to the new format, you just need to add the chunk
   frames (two integers) and the special string
   "@@@!!!TRIFILE_END_OF_ID_STRING!!!@@@" at the end of the graph id.

   In other words, the beginning of every trifile contains a condensed version
   of the .str file that is used to ensure that the trifile belongs with the .str file.
   Here is an example of the beginning of one such file for auroraTutorial, in old format:


         ===========================================================
         % Structure File Identification Information
         0 skipSil 0 2 D 0 1 0 
         1 wordCounter 0 22 D 0 1 1 skipSil 0 
         2 word 0 13 D 0 1 1 wordCounter 0 
         3 wordPosition 0 8 D 0 1 0 
         4 wholeWordState 0 91 D 0 1 2 word 0 wordPosition 0 
         5 obs 0 91 C 0 1 1 wholeWordState 0 
         6 phoneTransition 0 2 D 0 1 1 wholeWordState 0 
         7 wordTransition 0 2 D 0 1 3 word 0 wordPosition 0 phoneTransition 0 
         8 skipSil 1 2 D 0 1 0 
         9 wordCounter 1 22 D 0 1 3 wordCounter -1 wordTransition -1 skipSil 0 
         10 word 1 13 D 0 1 1 wordCounter 0 
         11 wordPosition 1 8 D 0 1 3 wordTransition -1 phoneTransition -1 wordPosition -1 
         12 wholeWordState 1 91 D 0 1 2 word 0 wordPosition 0 
         13 obs 1 91 C 0 1 1 wholeWordState 0 
         14 phoneTransition 1 2 D 0 1 1 wholeWordState 0 
         15 wordTransition 1 2 D 0 1 3 word 0 wordPosition 0 phoneTransition 0 
         16 skipSil 2 2 D 0 1 0 
         17 wordCounter 2 22 D 0 1 3 wordCounter -1 wordTransition -1 skipSil 0 
         18 word 2 13 D 0 1 1 wordCounter 0 
         19 wordPosition 2 8 D 0 1 3 wordTransition -1 phoneTransition -1 wordPosition -1 
         20 wholeWordState 2 91 D 0 1 2 word 0 wordPosition 0 
         21 obs 2 91 C 0 1 1 wholeWordState 0 
         22 phoneTransition 2 2 D 0 1 1 wholeWordState 0 
         23 wordTransition 2 2 D 0 1 3 word 0 wordPosition 0 phoneTransition 0 
         24 endOfUtteranceObservation 2 2 D 0 1 2 wordTransition 0 wordCounter 0 

	 ...<stuff>...

         ===========================================================

   where <stuff> is the rest of the trifile.  The new file format for
   the trifile must have the chunk numbers and the end string at the end
   of the ID section. I.e., if the .str file had "chunk 1:1" at the end,
   the above id would turn into:


         ===========================================================
         % Structure File Identification Information
         0 skipSil 0 2 D 0 1 0 
         1 wordCounter 0 22 D 0 1 1 skipSil 0 
         2 word 0 13 D 0 1 1 wordCounter 0 
         3 wordPosition 0 8 D 0 1 0 
         4 wholeWordState 0 91 D 0 1 2 word 0 wordPosition 0 
         5 obs 0 91 C 0 1 1 wholeWordState 0 
         6 phoneTransition 0 2 D 0 1 1 wholeWordState 0 
         7 wordTransition 0 2 D 0 1 3 word 0 wordPosition 0 phoneTransition 0 
         8 skipSil 1 2 D 0 1 0 
         9 wordCounter 1 22 D 0 1 3 wordCounter -1 wordTransition -1 skipSil 0 
         10 word 1 13 D 0 1 1 wordCounter 0 
         11 wordPosition 1 8 D 0 1 3 wordTransition -1 phoneTransition -1 wordPosition -1 
         12 wholeWordState 1 91 D 0 1 2 word 0 wordPosition 0 
         13 obs 1 91 C 0 1 1 wholeWordState 0 
         14 phoneTransition 1 2 D 0 1 1 wholeWordState 0 
         15 wordTransition 1 2 D 0 1 3 word 0 wordPosition 0 phoneTransition 0 
         16 skipSil 2 2 D 0 1 0 
         17 wordCounter 2 22 D 0 1 3 wordCounter -1 wordTransition -1 skipSil 0 
         18 word 2 13 D 0 1 1 wordCounter 0 
         19 wordPosition 2 8 D 0 1 3 wordTransition -1 phoneTransition -1 wordPosition -1 
         20 wholeWordState 2 91 D 0 1 2 word 0 wordPosition 0 
         21 obs 2 91 C 0 1 1 wholeWordState 0 
         22 phoneTransition 2 2 D 0 1 1 wholeWordState 0 
         23 wordTransition 2 2 D 0 1 3 word 0 wordPosition 0 phoneTransition 0 
         24 endOfUtteranceObservation 2 2 D 0 1 2 wordTransition 0 wordCounter 0 
         1 1
         @@@!!!TRIFILE_END_OF_ID_STRING!!!@@@ 

	 ...<same stuff afterwards> ...

         ===========================================================


- For graph debugging: Early print warnings as soon as we start getting zeros, either for cliques,
  or for separators (in latter case if any even minor degree of sep pruning is on).
  This is enabled with '-verb 40', so you'll see exactly the frame where
  you start getting zeros.

- VECPT syntax change again (this will be the last one). This is on top
  of Karim's recent change. The VECPT now has the following format; The
  first set of options match that of any other CPT. Namely, we have
     1) a name, 
     2) num parents (which must be 1 in this case), 
     3) parent cardinality
     4) self cardinaltiy (which must be 2 in this case).
     5) file observation name for the VECPT
   
  Next, we have a set of optional arguments that the user may give
  to the observation code. If an optional argument is not given,
  then the default value may be used.  These optional arguments
  consist of lines of a "flag : value" syntax, where "flag"
  indicates the current argument, and "value" is its value.
  The VECPT and optional arguments *MUST* end with the string "END"
  
  Here is an example:

         0
         VECPT0  % name of VECPT
         1 % num par
         2 % par card
         2 % self card
         VECPT0_FILE % file to read in.
         nfs:2 % nfloats
         nis:0 % nints
         frs:all % float range
         irs:all % int range
         pr:all % must be all
         fmt:ascii
         swap:F % endian swapping condition
         preTransforms:X
         postTransforms:X
         sentRange:all
         END

  but many are optional, so this is the same as:

         0
         VECPT0  % name of VECPT
         1 % num par
         2 % par card
         2 % self card
         VECPT0_FILE % file to read in.
         nfs:2 % nfloats
         nis:0 % nints
         fmt:ascii
         swap:F % endian swapping condition
         END

  The shortest would be:

         0
         VECPT0  % name of VECPT
         1 % num par
         2 % par card
         2 % self card
         VECPT0_FILE % file to read in.
         END


Minor changes:

-  verbose messages now working for reading parameters. Try -debug 90 and you'll
   see what parameters are being read in.

-  Names P1, Co, E1 are now printed P',C',E', where for left interface
   P'=P, C'=C, E'=[CE], and for right interface, P'=[PC],C'=C,E'=E. -verb 80 output   
   further prettied.

-  Removed almost all of the old and unused code from the CVS distribution.

-  Some minor additional optimizations.

Bug fixes

  - various bug fixes, including a few regarding switching parents and EM training.



			-- Jeff



+======================================================================+
|            Jeff A. Bilmes, Assistant Professor                       |
| Dept. of EE              Voice: (206) 221-5236                       |
| University of Washington FAX: (206) 543-3842                         |
| Box 352500               Email: bilmes@ee.washington.edu             |
| Seattle, WA  98195-2500  http://www.ee.washington.edu/faculty/bilmes |
+======================================================================+

From gang@ssli.ee.washington.edu Tue Jan 18 16:55:18 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j0J0tIJV028417
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@ssli.ee.washington.edu>; Tue, 18 Jan 2005 16:55:18 -0800
Received: from [128.95.167.96] (anatra.ee.washington.edu [128.95.167.96])
	by seagull.ee.washington.edu (8.13.1/8.13.0) with ESMTP id
	j0J0tH3w004613
	for <gmtk-users@ssli.ee.washington.edu>; Tue, 18 Jan 2005 16:55:17 -0800
Message-ID: <41EDAFF5.3030207@ssli.ee.washington.edu>
Date: Tue, 18 Jan 2005 16:55:17 -0800
From: Gang Ji <gang@ssli-mail.ee.washington.edu>
User-Agent: Mozilla Thunderbird 1.0 (X11/20041206)
X-Accept-Language: en-us, en
MIME-Version: 1.0
To: gmtk-users@ssli.ee.washington.edu
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit
X-Spam-Status: No, hits=-13.7 required=5.0 tests=AWL autolearn=no version=2.64
X-Spam-Checker-Version: SpamAssassin 2.64 (2004-01-11) on 
	seagull.ee.washington.edu
Subject: [Gmtk-users] dev tag updated
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Wed, 19 Jan 2005 00:55:18 -0000

Hi GMTKer's,

Bug fixes: there was a bug when two or more iterators for one 
NGramCPT/FNGramCPT exist.  This bug was fixed in this version.

*** how to create vocab object ***

% in master file
VOCAB_IN_FILE inline
1 % number of vocabs
0 % index
wordVocab % name
27682 % cardinality
w.dct % file for word list


*** how to use ngram cpt ***

% in master file
NGRAM_CPT_IN_FILE inline
1 % number of ngrams
0 % index
ngram % name
2 % number of parents
27682 27682 27682 % cardinalities
bigtrigram.lm wordVocab % ARPA lm file name followed by vocab object name

% in structure file
conditionalparents: word(-2), word(-1) using NGramCPT("ngram");
conditionalparents: word(-1) using NGramCPT("ngram"); % also supports 
less parents

*** how to use fngram cpt ***

% in master file
FNGRAM_CPT_IN_FILE inline
1 % number of fngrams
0 % index
fngram % name
3 % number of parents
27680 27680 25 27680 % cardinalities
meeting.flm W-wordVocab:P-posVocab % FLM filename followed by vocab mapping

% in structure file
conditionalparents: word(-2), word(-1), pos(0) using FNGramCPT("fngram");
conditionalparents: word(-2), word(-1) using FNGramCPT("fngram", 0, 1); 
% this means uing 0th and 1st parent in FLM specification

Best,
Gang

-- 
Office:   (206) 221-5216
Fax:      (206) 543-3842
email:    gang@ee.washington.edu
homepage: welcome.to/rainier


From bilmes@cuba.ee.washington.edu Tue Mar  1 02:25:04 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j21AP4Kb015467
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@crow.ee.washington.edu>; Tue, 1 Mar 2005 02:25:04 -0800
Received: from cuba.ee.washington.edu (cuba.ee.washington.edu [128.95.167.41])
	by seagull.ee.washington.edu (8.13.1/8.13.0) with ESMTP id
	j21AP47D027207 for <gmtk-users>; Tue, 1 Mar 2005 02:25:04 -0800
Message-Id: <200503011025.j21AP47D027207@seagull.ee.washington.edu>
To: gmtk-users@seagull.ee.washington.edu
X-Mailer: MH-E 7.82; nmh 1.0.4; GNU Emacs 21.3.1
Date: Tue, 01 Mar 2005 02:25:04 -0800
From: Jeff Bilmes <bilmes@ssli-mail.ee.washington.edu>
X-Spam-Status: No, hits=-0.6 required=5.0 tests=AWL, TO_MALFORMED autolearn=no 
	version=2.64
X-Spam-Checker-Version: SpamAssassin 2.64 (2004-01-11) on 
	seagull.ee.washington.edu
Cc: 
Subject: [Gmtk-users] dev tag updated
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
Reply-To: "Jeff Bilmes \(206/221-5236\)" <bilmes@ee.washington.edu>
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Tue, 01 Mar 2005 10:25:05 -0000


Folks,

I've updated the gmtk dev tag.

Changes:
 - new option: -cpbeam, which does partial clique beam pruning. This can
   be used in combination with the other beam options (namely, -cbeam,
   -crbeam, -ckbeam, and -sbeam).  It in particular can be used to prune
   states away when Gaussians are ordered using -vcap so that they are
   early. In other words, -cpbeam will prune away a clique as it is
   being created based on an estimate of what in the future will be the
   maximum clique value. -cpbeam can also reduce memory. The argument to
   -cpbeam is of the same form as -cbeam. Lastly, -cpbeam only applies
   to the C' partition (not P' or E').
 - various minor speedups regarding clique packing, 
 - various bug fixes.



			-- Jeff



+======================================================================+
|            Jeff A. Bilmes, Assistant Professor                       |
| Dept. of EE              Voice: (206) 221-5236                       |
| University of Washington FAX: (206) 543-3842                         |
| Box 352500               Email: bilmes@ee.washington.edu             |
| Seattle, WA  98195-2500  http://www.ee.washington.edu/faculty/bilmes |
+======================================================================+

From bilmes@cuba.ee.washington.edu Thu Mar  3 00:08:39 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j2388dqg014245
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@crow.ee.washington.edu>; Thu, 3 Mar 2005 00:08:39 -0800
Received: from cuba.ee.washington.edu (cuba.ee.washington.edu [128.95.167.41])
	by seagull.ee.washington.edu (8.13.1/8.13.0) with ESMTP id
	j2388cTd001134 for <gmtk-users>; Thu, 3 Mar 2005 00:08:39 -0800
Message-Id: <200503030808.j2388cTd001134@seagull.ee.washington.edu>
To: gmtk-users@seagull.ee.washington.edu
X-Mailer: MH-E 7.82; nmh 1.0.4; GNU Emacs 21.3.1
Date: Thu, 03 Mar 2005 00:08:38 -0800
From: Jeff Bilmes <bilmes@ssli-mail.ee.washington.edu>
X-Spam-Status: No, hits=-0.6 required=5.0 tests=AWL, TO_MALFORMED autolearn=no 
	version=2.64
X-Spam-Checker-Version: SpamAssassin 2.64 (2004-01-11) on 
	seagull.ee.washington.edu
Cc: 
Subject: [Gmtk-users] on most recent dev tag
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
Reply-To: "Jeff Bilmes \(206/221-5236\)" <bilmes@ee.washington.edu>
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Thu, 03 Mar 2005 08:08:40 -0000


Folks,

A few more details on yesterdays (short) dev tag update message:

- The new "-cpbeam" option is really different than all previous beam
  options. You might ask, why is this beam pruning option different from
  all other beam pruning options?  Well, basically, all previous beam
  options construct the entire clique and only then will they start
  pruning it down, so it means you need to have enough memory to hold at
  least one clique. While the clique might be "pruned" via previous
  pruning (via the separators that have been pruned with either -sbeam
  or indirectly via previous clique -cbeams), there were still cases
  where the clique construction itself gets large.

  -cpbeam fixes that. Basically, while constructing a clique, if partial entries
  in the clique fall below the currently estimated max clique value, then the
  rest of the clique entry is not computed. Since cliques are essentially formed
  like a depth-first tree traversal, early pruning with -cpbeam can have a large
  effect in trimming off large portions of the tree (so they are never constructed).
  This is why this option can help to decrease memory requirements as well.

  There are two potential caveats, however, that I might as well mention
  now. First, since we're beam pruning a clique before we know the
  clique's max value, we have to estimate the clique's max value. This
  is done by the two preceding instances of the clique to the left. This
  can be done for C' but since there is only ever one P' or E' per
  utterance, P' and E' don't currently benefit from -cpbeam (although it
  would be possible to estimate from utterance to utterance which I'll
  give a try at some point). So far, the simple estimation algorithm,
  however, seem to be amazingly good (better than I expected) so this
  probably won't be a problem.

  Secondly, -cpbeam works by assuming that later entries in the clique
  will always *reduce* the current probability.  For example, suppose
  the clique entries consists of multiplying 5 probabilities:

       p1*p2*p3*p4*p5

  If -cpbeam finds that p1*p2*p3 falls below a threshold, it need not bother with p4 or p5,
  but this really is only true if p4 <=1 and p5 <= 1. While this is true for probabilities,
  if p4 or p5 are Gaussian scores, we might have p4 or p5 > 1. This can happen if you have
  Gaussians with very small variances but you have observations that are close to the mean.
  Therefore it is important to ensure that you're data is scaled correctly. There are several ways
  to do this:

     1) always globally variance normalize your acoustic feature vectors (unity variance should
        be sufficient, but to be safer, you might even want to normalize to say variance 10 or 100,
        it really depends on the data).
     2) using a higher variance floor parameter
     3) scaling the data on training/testing.

  Note that option 1 and 3 is easy now that we have Karim's front end in place, which can
  do all kinds of feature transformations online (see message from a few weeks ago on command
  line syntax).

  So to summarize, I do hope that -cpbeam makes everyones life much easier :-)
   
Other changes in yesterdays dev tag:

- trace messages for reading in files now turned on with -verb 59 rather than verb 90
  (-verb 60 is the level at which you start getting a ton of messages)

- -vcap now supports 'S' (which sorts if a variable is switching or not)


Optimization options (skip if you don't want details):

   - some triangulations involved variables with the "disposition 4" (meaning a
     clique continuation). It turns out that this variable disposition was
     unnecessary, and by removing them have achieved a speedup on those graphs.
     I've seen speedups of 10-15% but it could be much more, depending on the
     triangulation. In general, the triangulations with more disp4's will be speedup more.

   - I secretly updated the dev tag earlier in Feb so that clique
     packing is much faster (it turns out that you can organize the
     packing to almost optimally not span word boundaries, which is what
     is done now). In the best of cases, this can give another 20-40%
     speedup. If you haven't updated lately, you'll probably notice this
     change (except when cliques are small).


more speedups on the way ...



			-- Jeff



+======================================================================+
|            Jeff A. Bilmes, Assistant Professor                       |
| Dept. of EE              Voice: (206) 221-5236                       |
| University of Washington FAX: (206) 543-3842                         |
| Box 352500               Email: bilmes@ee.washington.edu             |
| Seattle, WA  98195-2500  http://www.ee.washington.edu/faculty/bilmes |
+======================================================================+

From bilmes@cuba.ee.washington.edu Sat Mar  5 19:21:00 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j263KxNw001258
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@crow.ee.washington.edu>; Sat, 5 Mar 2005 19:21:00 -0800
Received: from cuba.ee.washington.edu (cuba.ee.washington.edu [128.95.167.41])
	by seagull.ee.washington.edu (8.13.1/8.13.0) with ESMTP id
	j263KxEV019083 for <gmtk-users>; Sat, 5 Mar 2005 19:20:59 -0800
Message-Id: <200503060320.j263KxEV019083@seagull.ee.washington.edu>
To: gmtk-users@seagull.ee.washington.edu
X-Mailer: MH-E 7.82; nmh 1.0.4; GNU Emacs 21.3.1
Date: Sat, 05 Mar 2005 19:20:59 -0800
From: Jeff Bilmes <bilmes@ssli-mail.ee.washington.edu>
X-Spam-Status: No, hits=-0.6 required=5.0 tests=AWL, TO_MALFORMED autolearn=no 
	version=2.64
X-Spam-Checker-Version: SpamAssassin 2.64 (2004-01-11) on 
	seagull.ee.washington.edu
Cc: 
Subject: [Gmtk-users] dev tag updated (please get)
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
Reply-To: "Jeff Bilmes \(206/221-5236\)" <bilmes@ee.washington.edu>
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Sun, 06 Mar 2005 03:21:00 -0000


Yet another dev tag update (sorry).

Bug fixex:
  - the previous dev tag updated had a bug which might cause some graphs run slower.
    Update this one to fix that bug and also get ...

Enhancements/Speedups:
  - new clique sorting options ('-vcap' M, T, '+', and '.'). The new default -vcap seems
    to speed up inference by 10-20% on most graphs. 
    The other -vcap options '+' and '.' are by arbitrary and file position order (useful for debugging your models).



			-- Jeff



+======================================================================+
|            Jeff A. Bilmes, Assistant Professor                       |
| Dept. of EE              Voice: (206) 221-5236                       |
| University of Washington FAX: (206) 543-3842                         |
| Box 352500               Email: bilmes@ee.washington.edu             |
| Seattle, WA  98195-2500  http://www.ee.washington.edu/faculty/bilmes |
+======================================================================+

From bilmes@cuba.ee.washington.edu Sat Mar 12 19:30:17 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j2D3UHI7028192
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@crow.ee.washington.edu>; Sat, 12 Mar 2005 19:30:17 -0800
Received: from cuba.ee.washington.edu (cuba.ee.washington.edu [128.95.167.41])
	by seagull.ee.washington.edu (8.13.1/8.13.0) with ESMTP id
	j2D3UGJo020953 for <gmtk-users>; Sat, 12 Mar 2005 19:30:16 -0800
Message-Id: <200503130330.j2D3UGJo020953@seagull.ee.washington.edu>
To: gmtk-users@seagull.ee.washington.edu
X-Mailer: MH-E 7.82; nmh 1.0.4; GNU Emacs 21.3.1
Date: Sat, 12 Mar 2005 19:30:16 -0800
From: Jeff Bilmes <bilmes@ssli-mail.ee.washington.edu>
X-Spam-Status: No, hits=-0.6 required=5.0 tests=AWL, TO_MALFORMED autolearn=no 
	version=2.64
X-Spam-Checker-Version: SpamAssassin 2.64 (2004-01-11) on 
	seagull.ee.washington.edu
Cc: 
Subject: [Gmtk-users] comparison of the old and new GMTK (funny)
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
Reply-To: "Jeff Bilmes \(206/221-5236\)" <bilmes@ee.washington.edu>
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Sun, 13 Mar 2005 03:30:17 -0000


http://w4.pica.army.mil/ARDEC-RI/dale/gmtk.html


			-- Jeff



+======================================================================+
|            Jeff A. Bilmes, Assistant Professor                       |
| Dept. of EE              Voice: (206) 221-5236                       |
| University of Washington FAX: (206) 543-3842                         |
| Box 352500               Email: bilmes@ee.washington.edu             |
| Seattle, WA  98195-2500  http://www.ee.washington.edu/faculty/bilmes |
+======================================================================+

From sheila@ssli-mail.ee.washington.edu Sun Mar 13 13:12:30 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j2DLCUqD009330
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@crow.ee.washington.edu>; Sun, 13 Mar 2005 13:12:30 -0800
Received: from patka.ee.washington.edu (patka.ee.washington.edu
	[128.95.167.81])
	by seagull.ee.washington.edu (8.13.1/8.13.0) with ESMTP id
	j2DLCTe2020823
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=NO);
	Sun, 13 Mar 2005 13:12:29 -0800
Received: from patka.ee.washington.edu (localhost [127.0.0.1])
	by patka.ee.washington.edu (8.12.11/8.12.10) with ESMTP id
	j2DLCT9j032409; Sun, 13 Mar 2005 13:12:29 -0800
Received: from localhost (sheila@localhost)
	by patka.ee.washington.edu (8.12.11/8.12.11/Submit) with ESMTP id
	j2DLCTvb032405; Sun, 13 Mar 2005 13:12:29 -0800
Date: Sun, 13 Mar 2005 13:12:28 -0800 (PST)
From: Sheila M Reynolds <sheila@ssli-mail.ee.washington.edu>
To: "Jeff Bilmes (206/221-5236)" <bilmes@ee.washington.edu>
Subject: Re: [Gmtk-users] comparison of the old and new GMTK (funny)
In-Reply-To: <200503130330.j2D3UGJo020953@seagull.ee.washington.edu>
Message-ID: <Pine.LNX.4.58.0503131311510.32374@patka.ee.washington.edu>
References: <200503130330.j2D3UGJo020953@seagull.ee.washington.edu>
MIME-Version: 1.0
Content-Type: TEXT/PLAIN; charset=US-ASCII
X-Spam-Status: No, hits=-13.1 required=5.0 tests=AWL,BAYES_00 autolearn=ham 
	version=2.64
X-Spam-Checker-Version: SpamAssassin 2.64 (2004-01-11) on 
	seagull.ee.washington.edu
Cc: gmtk-users@seagull.ee.washington.edu
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Sun, 13 Mar 2005 21:12:30 -0000


Wow, a 2 foot drop onto concrete at 20 degrees below, very impressive ;-)

			-- Sheila



On Sat, 12 Mar 2005, Jeff Bilmes wrote:

>
> http://w4.pica.army.mil/ARDEC-RI/dale/gmtk.html
>
>
> 			-- Jeff
>
>
>
> +======================================================================+
> |            Jeff A. Bilmes, Assistant Professor                       |
> | Dept. of EE              Voice: (206) 221-5236                       |
> | University of Washington FAX: (206) 543-3842                         |
> | Box 352500               Email: bilmes@ee.washington.edu             |
> | Seattle, WA  98195-2500  http://www.ee.washington.edu/faculty/bilmes |
> +======================================================================+
> _______________________________________________
> Gmtk-users mailing list
> Gmtk-users@ssli.ee.washington.edu
> https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users
>

From sheila_reynolds-08cwiA/55yCcickalbiTd2mWyMi8Hed@mailblocks.com Sun Mar 27 14:32:23 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j2RMWN7r006820
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@ssli.ee.washington.edu>; Sun, 27 Mar 2005 14:32:23 -0800
Received: from smtp3.mailblocks.com (smtp3.mailblocks.com [140.174.9.93])
	by seagull.ee.washington.edu (8.13.1/8.13.0) with SMTP id
	j2RMWIwF000949
	for <gmtk-users@ssli.ee.washington.edu>; Sun, 27 Mar 2005 14:32:18 -0800
Received: (qmail 3157 invoked from network); 27 Mar 2005 22:32:17 -0000
Received: from 10.10.0.51 (HELO mailblocks.com) (10.10.0.51)
	by 10.10.0.93 with SMTP; 27 Mar 2005 22:32:17 -0000
Date: Sun, 27 Mar 2005 14:32:17 -0800
Message-Id: <sheila_reynolds-08cwiA/15yCcsCpy8MfLGyAPyT3/b3v@mailblocks.com>
From: <sheila@ee.washington.edu>
References: <Pine.LNX.4.58L.0501060846090.8571@ernesto.cs.washington.edu>
Received: from 207.229.102.130 by app2.mailblocks.com (192.168.225.51) with
	HTTP (Mailblocks); Sun, 27 Mar 2005 14:32:17 -0800
X-MB-Message-Source: WebUI
X-MB-Message-Type: User
In-Reply-To: <Pine.LNX.4.58L.0501060846090.8571@ernesto.cs.washington.edu>
Content-Type: text/plain; charset="us-ascii"; format=flowed
MIME-Version: 1.0
To: gmtk-users@ssli.ee.washington.edu
X-Priority: 3
X-Spam-Status: No, hits=-8.7 required=5.0 tests=AWL,MSGID_FROM_MTA_HEADER,
	NO_DNS_FOR_FROM,NO_REAL_NAME,PRIORITY_NO_NAME autolearn=no 
	version=2.64
X-Spam-Checker-Version: SpamAssassin 2.64 (2004-01-11) on 
	seagull.ee.washington.edu
Subject: [Gmtk-users] need to do a simple 2-state HMM w/ continuous
	observations
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Sun, 27 Mar 2005 22:32:25 -0000

Hey folks,

I've never worked with continuous data before and I've got some 
sequences that I want to try to segment into two different states, with 
each state modeled (at least initially) by a single Gaussian.

Does anyone have an example of something like this somewhere that I 
could take a look at ?  All I've done with GMTK so far has involved 
discrete (integer) data, so I need some help with this -- file formats, 
what new things I need to specify, etc.  Once I've got things defined, 
I'll use EM training and then Viterbi to find the best path.

thanks,

Sheila


From klivescu@csail.mit.edu Sun Mar 27 20:15:40 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j2S4FdTM029890
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@ssli.ee.washington.edu>; Sun, 27 Mar 2005 20:15:40 -0800
Received: from rozz.csail.mit.edu (rozz.csail.mit.edu [128.30.2.16])
	by seagull.ee.washington.edu (8.13.1/8.13.0) with ESMTP id
	j2S4Fcfm016774
	(version=TLSv1/SSLv3 cipher=EDH-RSA-DES-CBC3-SHA bits=168 verify=FAIL)
	for <gmtk-users@ssli.ee.washington.edu>; Sun, 27 Mar 2005 20:15:39 -0800
Received: from voltron.csail.mit.edu ([128.30.2.19] helo=localhost)
	by rozz.csail.mit.edu with esmtp (Exim 4.50)
	id 1DFlf7-0007X7-QA; Sun, 27 Mar 2005 23:15:37 -0500
Received: from c-65-96-167-95.hsd1.ma.comcast.net
	(c-65-96-167-95.hsd1.ma.comcast.net [65.96.167.95]) 
	by imap.csail.mit.edu (IMP) with HTTP 
	for <klivescu@imap.csail.mit.edu>; Sun, 27 Mar 2005 23:15:37 -0500
Message-ID: <1111983337.424784e9baaac@imap.csail.mit.edu>
Date: Sun, 27 Mar 2005 23:15:37 -0500
From: klivescu@csail.mit.edu
To: sheila@ee.washington.edu
Subject: Re: [Gmtk-users] need to do a simple 2-state HMM w/ continuous
	observations
References: <Pine.LNX.4.58L.0501060846090.8571@ernesto.cs.washington.edu>
	<sheila_reynolds-08cwiA/15yCcsCpy8MfLGyAPyT3/b3v@mailblocks.com>
In-Reply-To: <sheila_reynolds-08cwiA/15yCcsCpy8MfLGyAPyT3/b3v@mailblocks.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=ISO-8859-1
Content-Transfer-Encoding: 8bit
User-Agent: Internet Messaging Program (IMP) 3.2.2
X-Spam-Status: No, hits=-2.2 required=5.0 tests=AWL, NO_REAL_NAME autolearn=no 
	version=2.64
X-Spam-Checker-Version: SpamAssassin 2.64 (2004-01-11) on 
	seagull.ee.washington.edu
Cc: gmtk-users@ssli.ee.washington.edu
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Mon, 28 Mar 2005 04:15:41 -0000

Hi Sheila,

Are you still looking for this info?  If so, have you looked at Jeff's Aurora
tutorial?  It's not a simple HMM, but it's HMM-based and it's a good place to
start since it uses continuous observations and Gaussian mixtures.

Karen

Quoting sheila@ee.washington.edu:

> Hey folks,
> 
> I've never worked with continuous data before and I've got some 
> sequences that I want to try to segment into two different states, with 
> each state modeled (at least initially) by a single Gaussian.
> 
> Does anyone have an example of something like this somewhere that I 
> could take a look at ?  All I've done with GMTK so far has involved 
> discrete (integer) data, so I need some help with this -- file formats, 
> what new things I need to specify, etc.  Once I've got things defined, 
> I'll use EM training and then Viterbi to find the best path.
> 
> thanks,
> 
> Sheila
> 
> _______________________________________________
> Gmtk-users mailing list
> Gmtk-users@ssli.ee.washington.edu
> https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users
> 




From bilmes@cuba.ee.washington.edu Sun Mar 27 20:18:32 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j2S4IWeW030006
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@ssli.ee.washington.edu>; Sun, 27 Mar 2005 20:18:32 -0800
Received: from cuba.ee.washington.edu (cuba.ee.washington.edu [128.95.167.41])
	by seagull.ee.washington.edu (8.13.1/8.13.0) with ESMTP id
	j2S4IQbJ016840; Sun, 27 Mar 2005 20:18:26 -0800
Message-Id: <200503280418.j2S4IQbJ016840@seagull.ee.washington.edu>
To: klivescu@csail.mit.edu
Subject: Re: [Gmtk-users] need to do a simple 2-state HMM w/ continuous
	observations 
In-Reply-To: Message from klivescu@csail.mit.edu of "Sun,
	27 Mar 2005 23:15:37 EST."
	<1111983337.424784e9baaac@imap.csail.mit.edu> 
X-Mailer: MH-E 7.82; nmh 1.0.4; GNU Emacs 21.3.1
Date: Sun, 27 Mar 2005 20:18:26 -0800
From: Jeff Bilmes <bilmes@ssli-mail.ee.washington.edu>
X-Spam-Status: No, hits=-2.3 required=5.0 tests=AWL, BAYES_00, NO_DNS_FOR_FROM 
	autolearn=ham version=2.64
X-Spam-Checker-Version: SpamAssassin 2.64 (2004-01-11) on 
	seagull.ee.washington.edu
Cc: gmtk-users@ssli.ee.washington.edu, sheila@ee.washington.edu
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
Reply-To: "Jeff Bilmes \(206/221-5236\)" <bilmes@ee.washington.edu>
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Mon, 28 Mar 2005 04:18:32 -0000


Yes, you can get the latest one at ~/gmtk/gmtk_dev/tests_and_development/auroraTutorial04.tar.gz

			-- Jeff

In the message dated Sun, 27 Mar 2005 23:15:37 EST
  klivescu@csail.mit.edu writes:
>Hi Sheila,
>
>Are you still looking for this info?  If so, have you looked at Jeff's Aurora
>tutorial?  It's not a simple HMM, but it's HMM-based and it's a good place to
>start since it uses continuous observations and Gaussian mixtures.
>
>Karen
>
>Quoting sheila@ee.washington.edu:
>
>> Hey folks,
>> 
>> I've never worked with continuous data before and I've got some 
>> sequences that I want to try to segment into two different states, with 
>> each state modeled (at least initially) by a single Gaussian.
>> 
>> Does anyone have an example of something like this somewhere that I 
>> could take a look at ?  All I've done with GMTK so far has involved 
>> discrete (integer) data, so I need some help with this -- file formats, 
>> what new things I need to specify, etc.  Once I've got things defined, 
>> I'll use EM training and then Viterbi to find the best path.
>> 
>> thanks,
>> 
>> Sheila
>> 
>> _______________________________________________
>> Gmtk-users mailing list
>> Gmtk-users@ssli.ee.washington.edu
>> https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users
>> 
>
>
>
>_______________________________________________
>Gmtk-users mailing list
>Gmtk-users@ssli.ee.washington.edu
>https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users


+======================================================================+
|            Jeff A. Bilmes, Assistant Professor                       |
| Dept. of EE              Voice: (206) 221-5236                       |
| University of Washington FAX: (206) 543-3842                         |
| Box 352500               Email: bilmes@ee.washington.edu             |
| Seattle, WA  98195-2500  http://www.ee.washington.edu/faculty/bilmes |
+======================================================================+

From bilmes@cuba.ee.washington.edu Tue Mar 29 15:25:41 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j2TNPfH2004986
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@crow.ee.washington.edu>; Tue, 29 Mar 2005 15:25:41 -0800
Received: from cuba.ee.washington.edu (cuba.ee.washington.edu [128.95.167.41])
	by seagull.ee.washington.edu (8.13.1/8.13.0) with ESMTP id
	j2TNPZv5022594 for <gmtk-users>; Tue, 29 Mar 2005 15:25:35 -0800
Message-Id: <200503292325.j2TNPZv5022594@seagull.ee.washington.edu>
To: gmtk-users@seagull.ee.washington.edu
X-Mailer: MH-E 7.82; nmh 1.0.4; GNU Emacs 21.3.1
Date: Tue, 29 Mar 2005 15:25:35 -0800
From: Jeff Bilmes <bilmes@ssli-mail.ee.washington.edu>
X-Spam-Status: No, hits=0.0 required=5.0 tests=AWL,NO_DNS_FOR_FROM,
	TO_MALFORMED autolearn=no version=2.64
X-Spam-Checker-Version: SpamAssassin 2.64 (2004-01-11) on 
	seagull.ee.washington.edu
Cc: 
Subject: [Gmtk-users] dev tag (bug fix)
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
Reply-To: "Jeff Bilmes \(206/221-5236\)" <bilmes@ee.washington.edu>
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Tue, 29 Mar 2005 23:25:41 -0000


Folks,

In case anyone hasn't updated recently, please do. There was a bug fix
early this week that for some graphs and/or triangulations might give
you slightly different results (or even cause a crash).


			-- Jeff



+======================================================================+
|            Jeff A. Bilmes, Assistant Professor                       |
| Dept. of EE              Voice: (206) 221-5236                       |
| University of Washington FAX: (206) 543-3842                         |
| Box 352500               Email: bilmes@ee.washington.edu             |
| Seattle, WA  98195-2500  http://www.ee.washington.edu/faculty/bilmes |
+======================================================================+

From bilmes@cuba.ee.washington.edu Wed Apr  6 13:56:40 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j36KueMF002412
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@crow.ee.washington.edu>; Wed, 6 Apr 2005 13:56:40 -0700
Received: from cuba.ee.washington.edu (cuba.ee.washington.edu [128.95.167.41])
	by seagull.ee.washington.edu (8.13.1/8.13.0) with ESMTP id
	j36KuZuA001995 for <gmtk-users>; Wed, 6 Apr 2005 13:56:35 -0700
Message-Id: <200504062056.j36KuZuA001995@seagull.ee.washington.edu>
To: gmtk-users@seagull.ee.washington.edu
X-Mailer: MH-E 7.82; nmh 1.0.4; GNU Emacs 21.3.1
Date: Wed, 06 Apr 2005 13:56:35 -0700
From: Jeff Bilmes <bilmes@ssli-mail.ee.washington.edu>
X-Spam-Status: No, hits=0.0 required=5.0 tests=AWL,NO_DNS_FOR_FROM,
	TO_MALFORMED autolearn=no version=2.64
X-Spam-Checker-Version: SpamAssassin 2.64 (2004-01-11) on 
	seagull.ee.washington.edu
Cc: 
Subject: [Gmtk-users] dev tag updated
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
Reply-To: "Jeff Bilmes \(206/221-5236\)" <bilmes@ee.washington.edu>
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Wed, 06 Apr 2005 20:56:41 -0000


I've updated the dev tag on GMTK so that it now compiles both with gcc
3.4.3 and gcc 3.3 (let me know if you see any problems).



			-- Jeff



+======================================================================+
|            Jeff A. Bilmes, Assistant Professor                       |
| Dept. of EE              Voice: (206) 221-5236                       |
| University of Washington FAX: (206) 543-3842                         |
| Box 352500               Email: bilmes@ee.washington.edu             |
| Seattle, WA  98195-2500  http://www.ee.washington.edu/faculty/bilmes |
+======================================================================+

From bilmes@cuba.ee.washington.edu Fri May 27 02:33:00 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j4R9X0Ym003893
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@crow.ee.washington.edu>; Fri, 27 May 2005 02:33:00 -0700
Received: from cuba.ee.washington.edu (cuba.ee.washington.edu [128.95.167.41])
	by seagull.ee.washington.edu (8.13.1/8.13.3) with ESMTP id
	j4R9Wx5R017747 for <gmtk-users>; Fri, 27 May 2005 02:32:59 -0700
Message-Id: <200505270932.j4R9Wx5R017747@seagull.ee.washington.edu>
To: gmtk-users@seagull.ee.washington.edu
X-Mailer: MH-E 7.82; nmh 1.0.4; GNU Emacs 21.3.1
Date: Fri, 27 May 2005 02:32:59 -0700
From: Jeff Bilmes <bilmes@ssli-mail.ee.washington.edu>
X-Spam-Status: No, score=-2.0 required=5.0 tests=ALL_TRUSTED, AWL,
	TO_MALFORMED autolearn=failed version=3.0.3
X-Spam-Checker-Version: SpamAssassin 3.0.3 (2005-04-27) on 
	seagull.ee.washington.edu
Cc: 
Subject: [Gmtk-users] gmtk dev tag updated
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
Reply-To: "Jeff Bilmes \(206/221-5236\)" <bilmes@ee.washington.edu>
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Fri, 27 May 2005 09:33:00 -0000


Hi Everyone,

In case you haven't updated in a while, a number of dev tag updates have
occurred recently. Most of them have been bug fixes, but a few new features
have been added that might as well be mentioned at this point:

1) GMTK now supports HTK lattices (thanks to Gang)
2) A new all_unequal() decision tree function is available
     (i.e., all_unequal(p1,p2,p3,p4) returns one only when none of the parents are equal to each other).


In case anyone is wondering, the next set of features to be added will
be fast N-best lists, sampling, and then a re-done inference inner loop
that should give us some additional significant speedups.

Best,


			-- Jeff



+======================================================================+
|            Jeff A. Bilmes, Assistant Professor                       |
| Dept. of EE              Voice: (206) 221-5236                       |
| University of Washington FAX: (206) 543-3842                         |
| Box 352500               Email: bilmes@ee.washington.edu             |
| Seattle, WA  98195-2500  http://www.ee.washington.edu/faculty/bilmes |
+======================================================================+

From bilmes@cuba.ee.washington.edu Fri May 27 02:33:50 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j4R9Xoj0003941
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@crow.ee.washington.edu>; Fri, 27 May 2005 02:33:50 -0700
Received: from cuba.ee.washington.edu (cuba.ee.washington.edu [128.95.167.41])
	by seagull.ee.washington.edu (8.13.1/8.13.3) with ESMTP id
	j4R9XoBh017767 for <gmtk-users>; Fri, 27 May 2005 02:33:50 -0700
Message-Id: <200505270933.j4R9XoBh017767@seagull.ee.washington.edu>
To: gmtk-users@seagull.ee.washington.edu
X-Mailer: MH-E 7.82; nmh 1.0.4; GNU Emacs 21.3.1
Date: Fri, 27 May 2005 02:33:50 -0700
From: Jeff Bilmes <bilmes@ssli-mail.ee.washington.edu>
X-Spam-Status: No, score=-2.0 required=5.0 tests=ALL_TRUSTED, AWL,
	TO_MALFORMED autolearn=failed version=3.0.3
X-Spam-Checker-Version: SpamAssassin 3.0.3 (2005-04-27) on 
	seagull.ee.washington.edu
Cc: 
Subject: [Gmtk-users] Sorry, I forgot...
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
Reply-To: "Jeff Bilmes \(206/221-5236\)" <bilmes@ee.washington.edu>
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Fri, 27 May 2005 09:33:51 -0000


forgot to mention that the all_unequal() function is thanks to Chris!

:-)



			-- Jeff



+======================================================================+
|            Jeff A. Bilmes, Assistant Professor                       |
| Dept. of EE              Voice: (206) 221-5236                       |
| University of Washington FAX: (206) 543-3842                         |
| Box 352500               Email: bilmes@ee.washington.edu             |
| Seattle, WA  98195-2500  http://www.ee.washington.edu/faculty/bilmes |
+======================================================================+

From bilmes@cuba.ee.washington.edu Fri May 27 19:10:57 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j4S2AuFl020382
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@crow.ee.washington.edu>; Fri, 27 May 2005 19:10:57 -0700
Received: from cuba.ee.washington.edu (cuba.ee.washington.edu [128.95.167.41])
	by seagull.ee.washington.edu (8.13.1/8.13.3) with ESMTP id
	j4S2AtbM027656 for <gmtk-users>; Fri, 27 May 2005 19:10:55 -0700
Message-Id: <200505280210.j4S2AtbM027656@seagull.ee.washington.edu>
To: gmtk-users@seagull.ee.washington.edu
X-Mailer: MH-E 7.82; nmh 1.0.4; GNU Emacs 21.3.1
Date: Fri, 27 May 2005 19:10:55 -0700
From: Jeff Bilmes <bilmes@ssli-mail.ee.washington.edu>
X-Spam-Status: No, score=-2.0 required=5.0 tests=ALL_TRUSTED, AWL,
	TO_MALFORMED autolearn=failed version=3.0.3
X-Spam-Checker-Version: SpamAssassin 3.0.3 (2005-04-27) on 
	seagull.ee.washington.edu
Cc: 
Subject: [Gmtk-users] whoops, minor bug fix
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
Reply-To: "Jeff Bilmes \(206/221-5236\)" <bilmes@ee.washington.edu>
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Sat, 28 May 2005 02:10:57 -0000


If you checked out the latest dev tag based on this morning's email,
please do so again as in that version I had forgotten to remove a line
that caused a silly bug. Its fixed now...


			-- Jeff



+======================================================================+
|            Jeff A. Bilmes, Assistant Professor                       |
| Dept. of EE              Voice: (206) 221-5236                       |
| University of Washington FAX: (206) 543-3842                         |
| Box 352500               Email: bilmes@ee.washington.edu             |
| Seattle, WA  98195-2500  http://www.ee.washington.edu/faculty/bilmes |
+======================================================================+

From bilmes@cuba.ee.washington.edu Wed Jun  1 01:31:03 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j518V292024807
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@crow.ee.washington.edu>; Wed, 1 Jun 2005 01:31:03 -0700
Received: from cuba.ee.washington.edu (cuba.ee.washington.edu [128.95.167.41])
	by seagull.ee.washington.edu (8.13.1/8.13.3) with ESMTP id
	j518V1Z8016942 for <gmtk-users>; Wed, 1 Jun 2005 01:31:01 -0700
Message-Id: <200506010831.j518V1Z8016942@seagull.ee.washington.edu>
To: gmtk-users@seagull.ee.washington.edu
X-Mailer: MH-E 7.82; nmh 1.0.4; GNU Emacs 21.3.1
Date: Wed, 01 Jun 2005 01:31:01 -0700
From: Jeff Bilmes <bilmes@ssli-mail.ee.washington.edu>
X-Spam-Status: No, score=-2.0 required=5.0 tests=ALL_TRUSTED, AWL,
	TO_MALFORMED autolearn=failed version=3.0.3
X-Spam-Checker-Version: SpamAssassin 3.0.3 (2005-04-27) on 
	seagull.ee.washington.edu
Cc: 
Subject: [Gmtk-users] dev tag update 
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
Reply-To: "Jeff Bilmes \(206/221-5236\)" <bilmes@ee.washington.edu>
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Wed, 01 Jun 2005 08:31:03 -0000


Folks,

Another update has occurred. In this latest one, the new hash tables are
in (at last). You'll see a small but consistent speedup over the old
hash tables.

Moreover, you now have control over the hash table load factor so you
can tradeoff memory for speed. Its accessed using the "-hashLoadFactor
lf" argument, where 0.1 <= lf <= 0.99 is the range of reasonable
values. Increasing lf (e.g., lf=0.90) means that the hash table will run
slower but use up less memory, and decreasing lf means things will run
faster but will use up more memory. Note that if lf gets too small,
things will start running *much* slower since you'll be thrashing.

Also beware that hings can run bit unexpectedly. If you are working with
inherently small state spaces, and you increase lf, things might even
run faster since everything might start fitting in cache.

The speed doesn't appear to get that much slower when increase lf, but
in some cases the memory can be reduced quite a bit (for big state space
problems). Therefore, adding to your bag of tricks to reduce memory
usage when memory is a problem, you might try something like
"-hashLoadFactor 0.98" which at best might give you a factor 2 reduction
in memory without too much of a speed hit.


			-- Jeff



+======================================================================+
|            Jeff A. Bilmes, Assistant Professor                       |
| Dept. of EE              Voice: (206) 221-5236                       |
| University of Washington FAX: (206) 543-3842                         |
| Box 352500               Email: bilmes@ee.washington.edu             |
| Seattle, WA  98195-2500  http://www.ee.washington.edu/faculty/bilmes |
+======================================================================+

From sheila@ssli-mail.ee.washington.edu Wed Jun 15 21:21:53 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j5G4Lr79013898
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@ssli.ee.washington.edu>; Wed, 15 Jun 2005 21:21:53 -0700
Received: from patka.ee.washington.edu (patka.ee.washington.edu
	[128.95.167.81])
	by seagull.ee.washington.edu (8.13.1/8.13.3) with ESMTP id
	j5G4LqIp015593
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=NO)
	for <gmtk-users@ssli.ee.washington.edu>; Wed, 15 Jun 2005 21:21:52 -0700
Received: from patka.ee.washington.edu (localhost [127.0.0.1])
	by patka.ee.washington.edu (8.12.11/8.12.10) with ESMTP id
	j5G4Lqq8029658
	for <gmtk-users@ssli.ee.washington.edu>; Wed, 15 Jun 2005 21:21:52 -0700
Received: from localhost (sheila@localhost)
	by patka.ee.washington.edu (8.12.11/8.12.11/Submit) with ESMTP id
	j5G4Lqes029654
	for <gmtk-users@ssli.ee.washington.edu>; Wed, 15 Jun 2005 21:21:52 -0700
Date: Wed, 15 Jun 2005 21:21:52 -0700 (PDT)
From: Sheila M Reynolds <sheila@ssli-mail.ee.washington.edu>
To: gmtk-users@ssli.ee.washington.edu
Message-ID: <Pine.LNX.4.58.0506152120490.29642@patka.ee.washington.edu>
MIME-Version: 1.0
Content-Type: TEXT/PLAIN; charset=US-ASCII
X-Spam-Status: No, score=-0.8 required=5.0 tests=AWL autolearn=failed 
	version=3.0.3
X-Spam-Checker-Version: SpamAssassin 3.0.3 (2005-04-27) on 
	seagull.ee.washington.edu
Subject: [Gmtk-users] question about Gaussians
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Thu, 16 Jun 2005 04:21:53 -0000


is it possible to 'tie' the variances of different Gaussians ?  in other
words, have the variance be 'learned' via EM but force the two (or more)
Gaussians to have the same variance (different means) ?

thanks,

Sheila


From bilmes@cuba.ee.washington.edu Wed Jun 15 21:28:48 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j5G4Sme8014501
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK);
	Wed, 15 Jun 2005 21:28:48 -0700
Received: from cuba.ee.washington.edu (cuba.ee.washington.edu [128.95.167.41])
	by seagull.ee.washington.edu (8.13.1/8.13.3) with ESMTP id
	j5G4SkX1016057; Wed, 15 Jun 2005 21:28:46 -0700
Message-Id: <200506160428.j5G4SkX1016057@seagull.ee.washington.edu>
To: Sheila M Reynolds <sheila@ssli-mail.ee.washington.edu>
Subject: Re: [Gmtk-users] question about Gaussians 
In-Reply-To: Message from Sheila M Reynolds
	<sheila@ssli-mail.ee.washington.edu> of "Wed,
	15 Jun 2005 21:21:52 PDT."
	<Pine.LNX.4.58.0506152120490.29642@patka.ee.washington.edu> 
X-Mailer: MH-E 7.82; nmh 1.0.4; GNU Emacs 21.3.1
Date: Wed, 15 Jun 2005 21:28:46 -0700
From: Jeff Bilmes <bilmes@ssli-mail.ee.washington.edu>
X-Spam-Status: No, score=-4.7 required=5.0 tests=ALL_TRUSTED,AWL,BAYES_00 
	autolearn=ham version=3.0.3
X-Spam-Checker-Version: SpamAssassin 3.0.3 (2005-04-27) on 
	seagull.ee.washington.edu
Cc: gmtk-users@ssli.ee.washington.edu
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
Reply-To: "Jeff Bilmes \(206/221-5236\)" <bilmes@ee.Washington.EDU>
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Thu, 16 Jun 2005 04:28:48 -0000


Yes, this is quite possible. Just have the two Gausisans name the same
variance object. Also, when splitting components, the split "children"
components can share means, covariances, or dlinks using the
"cloneShareMeans", "cloneShareCovars", and/or "cloneShareDlinks" options
of gmtkEMTrainNew (the stuff in the old documentation about this is
actually still up to date regarding these features).


			-- Jeff

In the message dated Wed, 15 Jun 2005 21:21:52 PDT
  Sheila M Reynolds <sheila@ssli-mail.ee.washington.edu> writes:
>
>is it possible to 'tie' the variances of different Gaussians ?  in other
>words, have the variance be 'learned' via EM but force the two (or more)
>Gaussians to have the same variance (different means) ?
>
>thanks,
>
>Sheila
>
>_______________________________________________
>Gmtk-users mailing list
>Gmtk-users@ssli.ee.washington.edu
>https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users


+======================================================================+
|            Jeff A. Bilmes, Assistant Professor                       |
| Dept. of EE              Voice: (206) 221-5236                       |
| University of Washington FAX: (206) 543-3842                         |
| Box 352500               Email: bilmes@ee.washington.edu             |
| Seattle, WA  98195-2500  http://www.ee.washington.edu/faculty/bilmes |
+======================================================================+

From gang@ssli.ee.washington.edu Thu Jun 16 08:58:36 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j5GFwaqv014484
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@ssli.ee.washington.edu>; Thu, 16 Jun 2005 08:58:36 -0700
Received: from [128.95.167.96] (anatra.ee.washington.edu [128.95.167.96])
	(authenticated bits=0)
	by seagull.ee.washington.edu (8.13.1/8.13.3) with ESMTP id
	j5GFwZuW016471
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=NO)
	for <gmtk-users@ssli.ee.washington.edu>; Thu, 16 Jun 2005 08:58:35 -0700
Message-ID: <42B1A1AB.5070700@ssli.ee.washington.edu>
Date: Thu, 16 Jun 2005 08:58:35 -0700
From: Gang Ji <gang@ssli-mail.ee.washington.edu>
User-Agent: Mozilla Thunderbird 1.0.2 (X11/20050317)
X-Accept-Language: en-us, en
MIME-Version: 1.0
To: gmtk-users@ssli.ee.washington.edu
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit
X-Spam-Status: No, score=-3.5 required=5.0 tests=ALL_TRUSTED,AWL 
	autolearn=failed version=3.0.3
X-Spam-Checker-Version: SpamAssassin 3.0.3 (2005-04-27) on 
	seagull.ee.washington.edu
Subject: [Gmtk-users] printing two variables
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Thu, 16 Jun 2005 15:58:36 -0000

Hi,

Is it possible to print two random variables in gmtkViterbiNew?  This is 
*not* for debugging reasons.  I need to know two random variable 
sequences in decoding.  Rather than decode it twice, I hope this can be 
done in one pass.

Best,
Gang

-- 
Office:   (206) 221-5216
Fax:      (206) 543-3842
email:    gang@ee.washington.edu
homepage: welcome.to/rainier


From klivescu@csail.mit.edu Thu Jun 16 10:20:25 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j5GHKPOJ019729
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK);
	Thu, 16 Jun 2005 10:20:25 -0700
Received: from rozz.csail.mit.edu (rozz.csail.mit.edu [128.30.2.16])
	by seagull.ee.washington.edu (8.13.1/8.13.3) with ESMTP id
	j5GHKMQr020357
	(version=TLSv1/SSLv3 cipher=EDH-RSA-DES-CBC3-SHA bits=168 verify=FAIL); 
	Thu, 16 Jun 2005 10:20:22 -0700
Received: from voltron.csail.mit.edu ([128.30.2.19] helo=localhost)
	by rozz.csail.mit.edu with esmtp (Exim 4.50)
	id 1Diy2P-00064a-Pr; Thu, 16 Jun 2005 13:20:21 -0400
Received: from sls-370-29.csail.mit.edu (sls-370-29.csail.mit.edu
	[128.30.34.63]) by imap.csail.mit.edu (IMP) with HTTP 
	for <klivescu@imap.csail.mit.edu>; Thu, 16 Jun 2005 13:20:21 -0400
Message-ID: <1118942421.42b1b4d5b8068@imap.csail.mit.edu>
Date: Thu, 16 Jun 2005 13:20:21 -0400
From: klivescu@csail.mit.edu
To: Gang Ji <gang@ssli-mail.ee.washington.edu>
Subject: Re: [Gmtk-users] printing two variables
References: <42B1A1AB.5070700@ssli.ee.washington.edu>
In-Reply-To: <42B1A1AB.5070700@ssli.ee.washington.edu>
MIME-Version: 1.0
Content-Type: text/plain; charset=ISO-8859-1
Content-Transfer-Encoding: 8bit
User-Agent: Internet Messaging Program (IMP) 3.2.2
X-Spam-Status: No, score=-2.6 required=5.0 tests=BAYES_00,NO_REAL_NAME 
	autolearn=ham version=3.0.3
X-Spam-Checker-Version: SpamAssassin 3.0.3 (2005-04-27) on 
	seagull.ee.washington.edu
Cc: gmtk-users@ssli.ee.washington.edu
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Thu, 16 Jun 2005 17:20:25 -0000

Hey Gang,

Have you used "-showVitVals <filename>" or "-dumpNames <varlist> -ofilelist
<filelist>"?  When I need multiple variables printed, I usually do the former
and grep the output for the relevant variables, or the latter if I want to do a
whole batch of segments.

(Let me know if you find that there's actually some other way)

Karen

Quoting Gang Ji <gang@ssli-mail.ee.washington.edu>:

> Hi,
> 
> Is it possible to print two random variables in gmtkViterbiNew?  This is 
> *not* for debugging reasons.  I need to know two random variable 
> sequences in decoding.  Rather than decode it twice, I hope this can be 
> done in one pass.
> 
> Best,
> Gang
> 
> -- 
> Office:   (206) 221-5216
> Fax:      (206) 543-3842
> email:    gang@ee.washington.edu
> homepage: welcome.to/rainier
> 
> _______________________________________________
> Gmtk-users mailing list
> Gmtk-users@ssli.ee.washington.edu
> https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users
> 




From gang@ssli.ee.washington.edu Fri Jun 17 08:57:02 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j5HFv1es030045
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@ssli.ee.washington.edu>; Fri, 17 Jun 2005 08:57:02 -0700
Received: from [128.95.167.96] (anatra.ee.washington.edu [128.95.167.96])
	(authenticated bits=0)
	by seagull.ee.washington.edu (8.13.1/8.13.3) with ESMTP id
	j5HFux9v015940
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=NO);
	Fri, 17 Jun 2005 08:57:00 -0700
Message-ID: <42B2F2CB.6060102@ssli.ee.washington.edu>
Date: Fri, 17 Jun 2005 08:56:59 -0700
From: Gang Ji <gang@ssli-mail.ee.washington.edu>
User-Agent: Mozilla Thunderbird 1.0.2 (X11/20050317)
X-Accept-Language: en-us, en
MIME-Version: 1.0
To: klivescu@csail.mit.edu
Subject: Re: [Gmtk-users] printing two variables
References: <42B1A1AB.5070700@ssli.ee.washington.edu>
	<1118942421.42b1b4d5b8068@imap.csail.mit.edu>
In-Reply-To: <1118942421.42b1b4d5b8068@imap.csail.mit.edu>
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit
X-Spam-Status: No, score=-5.1 required=5.0 tests=ALL_TRUSTED,AWL,BAYES_00 
	autolearn=ham version=3.0.3
X-Spam-Checker-Version: SpamAssassin 3.0.3 (2005-04-27) on 
	seagull.ee.washington.edu
Cc: gmtk-users@ssli.ee.washington.edu
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Fri, 17 Jun 2005 15:57:02 -0000

Hi Karen,

Thanks a lot for the suggestion.  What I need is the following: when 
there is a word transition, I want to know the value of random variable 
A and random variable B.  This is a little different from what the 
"-showVitVals" does because the latter will give values at each time 
frame.  What I can do is using that output and doing some 
post-processing to get what I need.  I hope there is a direct way to do 
that.

Best,
Gang

klivescu@csail.mit.edu wrote:

>Hey Gang,
>
>Have you used "-showVitVals <filename>" or "-dumpNames <varlist> -ofilelist
><filelist>"?  When I need multiple variables printed, I usually do the former
>and grep the output for the relevant variables, or the latter if I want to do a
>whole batch of segments.
>
>(Let me know if you find that there's actually some other way)
>
>Karen
>
>Quoting Gang Ji <gang@ssli-mail.ee.washington.edu>:
>
>  
>
>>Hi,
>>
>>Is it possible to print two random variables in gmtkViterbiNew?  This is 
>>*not* for debugging reasons.  I need to know two random variable 
>>sequences in decoding.  Rather than decode it twice, I hope this can be 
>>done in one pass.
>>
>>Best,
>>Gang
>>
>>-- 
>>Office:   (206) 221-5216
>>Fax:      (206) 543-3842
>>email:    gang@ee.washington.edu
>>homepage: welcome.to/rainier
>>
>>_______________________________________________
>>Gmtk-users mailing list
>>Gmtk-users@ssli.ee.washington.edu
>>https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users
>>
>>    
>>
>
>
>
>_______________________________________________
>Gmtk-users mailing list
>Gmtk-users@ssli.ee.washington.edu
>https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users
>  
>


-- 
Office:   (206) 221-5216
Fax:      (206) 543-3842
email:    gang@ee.washington.edu
homepage: welcome.to/rainier


From karim@cs.washington.edu Fri Jun 17 10:23:47 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j5HHNloj003347
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK);
	Fri, 17 Jun 2005 10:23:47 -0700
Received: from mx1.cs.washington.edu (mx1.cs.washington.edu [128.208.5.52])
	by seagull.ee.washington.edu (8.13.1/8.13.3) with ESMTP id
	j5HHNjY3019918
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK);
	Fri, 17 Jun 2005 10:23:45 -0700
Received: from ernesto.cs.washington.edu (ernesto.cs.washington.edu
	[128.208.4.36])
	by mx1.cs.washington.edu (8.13.3/8.13.3/1.5) with ESMTP id
	j5HHNjg2031395
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=NO);
	Fri, 17 Jun 2005 10:23:45 -0700
	(envelope-from karim@cs.washington.edu)
Received: from localhost (karim@localhost)
	by ernesto.cs.washington.edu (8.13.1/8.13.1/Submit/1.7) with ESMTP id
	j5HHNji9027690; Fri, 17 Jun 2005 10:23:45 -0700
	(envelope-from karim@cs.washington.edu)
X-Authentication-Warning: ernesto.cs.washington.edu: karim owned process doing
	-bs
Date: Fri, 17 Jun 2005 10:23:45 -0700 (PDT)
From: Karim Filali <karim@cs.washington.edu>
To: Gang Ji <gang@ssli-mail.ee.washington.edu>
Subject: Re: [Gmtk-users] printing two variables
In-Reply-To: <42B2F2CB.6060102@ssli.ee.washington.edu>
Message-ID: <Pine.LNX.4.58L.0506170931530.22928@ernesto.cs.washington.edu>
References: <42B1A1AB.5070700@ssli.ee.washington.edu>
	<1118942421.42b1b4d5b8068@imap.csail.mit.edu>
	<42B2F2CB.6060102@ssli.ee.washington.edu>
MIME-Version: 1.0
Content-Type: TEXT/PLAIN; charset=US-ASCII
X-Spam-Status: No, score=-5.6 required=5.0 tests=ALL_TRUSTED,AWL,BAYES_00 
	autolearn=ham version=3.0.3
X-Spam-Checker-Version: SpamAssassin 3.0.3 (2005-04-27) on 
	seagull.ee.washington.edu
Cc: gmtk-users@ssli.ee.washington.edu
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Fri, 17 Jun 2005 17:23:47 -0000

If you don't mind changing your structure/master files and if A and B are
already in the same clique anyway, you can add a new variable C to your
graph, where C=A*(card(B))+B.  Then map output values of C to strings A_B
or something like that using the -varMap option.

   Karim

On Fri, 17 Jun 2005, Gang Ji wrote:

> Hi Karen,
> 
> Thanks a lot for the suggestion.  What I need is the following: when 
> there is a word transition, I want to know the value of random variable 
> A and random variable B.  This is a little different from what the 
> "-showVitVals" does because the latter will give values at each time 
> frame.  What I can do is using that output and doing some 
> post-processing to get what I need.  I hope there is a direct way to do 
> that.
> 
> Best,
> Gang
> 
> klivescu@csail.mit.edu wrote:
> 
> >Hey Gang,
> >
> >Have you used "-showVitVals <filename>" or "-dumpNames <varlist>
> >-ofilelist <filelist>"?  When I need multiple variables printed, I
> >usually do the former and grep the output for the relevant variables,
> >or the latter if I want to do a whole batch of segments.
> >
> >(Let me know if you find that there's actually some other way)
> >
> >Karen
> >
> >Quoting Gang Ji <gang@ssli-mail.ee.washington.edu>:
> >
> >  
> >
> >>Hi,
> >>
> >>Is it possible to print two random variables in gmtkViterbiNew?  This is 
> >>*not* for debugging reasons.  I need to know two random variable 
> >>sequences in decoding.  Rather than decode it twice, I hope this can be 
> >>done in one pass.
> >>
> >>Best,
> >>Gang
> >>
> >>-- 
> >>Office:   (206) 221-5216
> >>Fax:      (206) 543-3842
> >>email:    gang@ee.washington.edu
> >>homepage: welcome.to/rainier
> >>
> >>_______________________________________________
> >>Gmtk-users mailing list
> >>Gmtk-users@ssli.ee.washington.edu
> >>https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users
> >>
> >>    
> >>
> >
> >
> >
> >_______________________________________________
> >Gmtk-users mailing list
> >Gmtk-users@ssli.ee.washington.edu
> >https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users
> >  
> >
> 
> 
> -- 
> Office:   (206) 221-5216
> Fax:      (206) 543-3842
> email:    gang@ee.washington.edu
> homepage: welcome.to/rainier
> 
> _______________________________________________
> Gmtk-users mailing list
> Gmtk-users@ssli.ee.washington.edu
> https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users
> 

From gang@ssli-mail.ee.washington.edu Fri Jun 17 10:31:08 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j5HHV8CG003617
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@ssli.ee.washington.edu>; Fri, 17 Jun 2005 10:31:08 -0700
Received: from anatra.ee.washington.edu (anatra.ee.washington.edu
	[128.95.167.96])
	by seagull.ee.washington.edu (8.13.1/8.13.3) with ESMTP id
	j5HHV69m020022
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=NO)
	for <gmtk-users@ssli.ee.washington.edu>; Fri, 17 Jun 2005 10:31:06 -0700
Received: from anatra.ee.washington.edu (localhost [127.0.0.1])
	by anatra.ee.washington.edu (8.12.11/8.12.10) with ESMTP id
	j5HHV6ur010295
	for <gmtk-users@ssli.ee.washington.edu>; Fri, 17 Jun 2005 10:31:06 -0700
Received: from localhost (gang@localhost)
	by anatra.ee.washington.edu (8.12.11/8.12.11/Submit) with ESMTP id
	j5HHV6pD010291
	for <gmtk-users@ssli.ee.washington.edu>; Fri, 17 Jun 2005 10:31:06 -0700
Date: Fri, 17 Jun 2005 10:31:06 -0700 (PDT)
From: Gang Ji <gang@ssli-mail.ee.washington.edu>
To: gmtk-users@ssli.ee.washington.edu
Subject: Re: [Gmtk-users] printing two variables
In-Reply-To: <Pine.LNX.4.58L.0506170931530.22928@ernesto.cs.washington.edu>
Message-ID: <Pine.LNX.4.58.0506171029470.7566@anatra.ee.washington.edu>
References: <42B1A1AB.5070700@ssli.ee.washington.edu>
	<1118942421.42b1b4d5b8068@imap.csail.mit.edu>
	<42B2F2CB.6060102@ssli.ee.washington.edu>
	<Pine.LNX.4.58L.0506170931530.22928@ernesto.cs.washington.edu>
MIME-Version: 1.0
Content-Type: TEXT/PLAIN; charset=US-ASCII
X-Spam-Status: No, score=-0.8 required=5.0 tests=AWL autolearn=failed 
	version=3.0.3
X-Spam-Checker-Version: SpamAssassin 3.0.3 (2005-04-27) on 
	seagull.ee.washington.edu
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Fri, 17 Jun 2005 17:31:08 -0000


Well.  That will potentially create a random variable with BIG
cardinality, isn't it?

Best,
Gang

Office:   (206) 221-5216
Fax:      (206) 543-3842
email:    gang@ee.washington.edu
homepage: welcome.to/rainier

On Fri, 17 Jun 2005, Karim Filali wrote:

> If you don't mind changing your structure/master files and if A and B are
> already in the same clique anyway, you can add a new variable C to your
> graph, where C=A*(card(B))+B.  Then map output values of C to strings A_B
> or something like that using the -varMap option.
>
>    Karim
>
> On Fri, 17 Jun 2005, Gang Ji wrote:
>
> > Hi Karen,
> >
> > Thanks a lot for the suggestion.  What I need is the following: when
> > there is a word transition, I want to know the value of random variable
> > A and random variable B.  This is a little different from what the
> > "-showVitVals" does because the latter will give values at each time
> > frame.  What I can do is using that output and doing some
> > post-processing to get what I need.  I hope there is a direct way to do
> > that.
> >
> > Best,
> > Gang
> >
> > klivescu@csail.mit.edu wrote:
> >
> > >Hey Gang,
> > >
> > >Have you used "-showVitVals <filename>" or "-dumpNames <varlist>
> > >-ofilelist <filelist>"?  When I need multiple variables printed, I
> > >usually do the former and grep the output for the relevant variables,
> > >or the latter if I want to do a whole batch of segments.
> > >
> > >(Let me know if you find that there's actually some other way)
> > >
> > >Karen
> > >
> > >Quoting Gang Ji <gang@ssli-mail.ee.washington.edu>:
> > >
> > >
> > >
> > >>Hi,
> > >>
> > >>Is it possible to print two random variables in gmtkViterbiNew?  This is
> > >>*not* for debugging reasons.  I need to know two random variable
> > >>sequences in decoding.  Rather than decode it twice, I hope this can be
> > >>done in one pass.
> > >>
> > >>Best,
> > >>Gang
> > >>
> > >>--
> > >>Office:   (206) 221-5216
> > >>Fax:      (206) 543-3842
> > >>email:    gang@ee.washington.edu
> > >>homepage: welcome.to/rainier
> > >>
> > >>_______________________________________________
> > >>Gmtk-users mailing list
> > >>Gmtk-users@ssli.ee.washington.edu
> > >>https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users
> > >>
> > >>
> > >>
> > >
> > >
> > >
> > >_______________________________________________
> > >Gmtk-users mailing list
> > >Gmtk-users@ssli.ee.washington.edu
> > >https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users
> > >
> > >
> >
> >
> > --
> > Office:   (206) 221-5216
> > Fax:      (206) 543-3842
> > email:    gang@ee.washington.edu
> > homepage: welcome.to/rainier
> >
> > _______________________________________________
> > Gmtk-users mailing list
> > Gmtk-users@ssli.ee.washington.edu
> > https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users
> >
>

From bilmes@cuba.ee.washington.edu Fri Jun 17 10:49:32 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j5HHnWjr004615
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK);
	Fri, 17 Jun 2005 10:49:32 -0700
Received: from cuba.ee.washington.edu (cuba.ee.washington.edu [128.95.167.41])
	by seagull.ee.washington.edu (8.13.1/8.13.3) with ESMTP id
	j5HHnTfT021007; Fri, 17 Jun 2005 10:49:29 -0700
Message-Id: <200506171749.j5HHnTfT021007@seagull.ee.washington.edu>
To: Gang Ji <gang@ssli-mail.ee.washington.edu>
Subject: Re: [Gmtk-users] printing two variables 
In-Reply-To: Message from Gang Ji <gang@ssli-mail.ee.washington.edu> of "Fri,
	17 Jun 2005 10:31:06 PDT."
	<Pine.LNX.4.58.0506171029470.7566@anatra.ee.washington.edu> 
X-Mailer: MH-E 7.82; nmh 1.0.4; GNU Emacs 21.3.1
Date: Fri, 17 Jun 2005 10:49:29 -0700
From: Jeff Bilmes <bilmes@ssli-mail.ee.washington.edu>
X-Spam-Status: No, score=-4.7 required=5.0 tests=ALL_TRUSTED,AWL,BAYES_00 
	autolearn=ham version=3.0.3
X-Spam-Checker-Version: SpamAssassin 3.0.3 (2005-04-27) on 
	seagull.ee.washington.edu
Cc: gmtk-users@ssli.ee.washington.edu
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
Reply-To: "Jeff Bilmes \(206/221-5236\)" <bilmes@ee.Washington.EDU>
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Fri, 17 Jun 2005 17:49:33 -0000


Actually, this is really an issue with the gmtkViterbiNew front end
(which I'm planing to redo anyway soon). But for now, a quick
hack to gmtkViterbiNew.cc would do the trick (basically, all the
code is currently doing is printing out some random variables
when the value of other RVs is a particular value, the code
in gmtkViterbiNew.cc is *really* simple.


			-- Jeff

In the message dated Fri, 17 Jun 2005 10:31:06 PDT
  Gang Ji <gang@ssli-mail.ee.washington.edu> writes:
>
>Well.  That will potentially create a random variable with BIG
>cardinality, isn't it?
>
>Best,
>Gang
>
>Office:   (206) 221-5216
>Fax:      (206) 543-3842
>email:    gang@ee.washington.edu
>homepage: welcome.to/rainier
>
>On Fri, 17 Jun 2005, Karim Filali wrote:
>
>> If you don't mind changing your structure/master files and if A and B are
>> already in the same clique anyway, you can add a new variable C to your
>> graph, where C=A*(card(B))+B.  Then map output values of C to strings A_B
>> or something like that using the -varMap option.
>>
>>    Karim
>>
>> On Fri, 17 Jun 2005, Gang Ji wrote:
>>
>> > Hi Karen,
>> >
>> > Thanks a lot for the suggestion.  What I need is the following: when
>> > there is a word transition, I want to know the value of random variable
>> > A and random variable B.  This is a little different from what the
>> > "-showVitVals" does because the latter will give values at each time
>> > frame.  What I can do is using that output and doing some
>> > post-processing to get what I need.  I hope there is a direct way to do
>> > that.
>> >
>> > Best,
>> > Gang
>> >
>> > klivescu@csail.mit.edu wrote:
>> >
>> > >Hey Gang,
>> > >
>> > >Have you used "-showVitVals <filename>" or "-dumpNames <varlist>
>> > >-ofilelist <filelist>"?  When I need multiple variables printed, I
>> > >usually do the former and grep the output for the relevant variables,
>> > >or the latter if I want to do a whole batch of segments.
>> > >
>> > >(Let me know if you find that there's actually some other way)
>> > >
>> > >Karen
>> > >
>> > >Quoting Gang Ji <gang@ssli-mail.ee.washington.edu>:
>> > >
>> > >
>> > >
>> > >>Hi,
>> > >>
>> > >>Is it possible to print two random variables in gmtkViterbiNew?  This is
>> > >>*not* for debugging reasons.  I need to know two random variable
>> > >>sequences in decoding.  Rather than decode it twice, I hope this can be
>> > >>done in one pass.
>> > >>
>> > >>Best,
>> > >>Gang
>> > >>
>> > >>--
>> > >>Office:   (206) 221-5216
>> > >>Fax:      (206) 543-3842
>> > >>email:    gang@ee.washington.edu
>> > >>homepage: welcome.to/rainier
>> > >>
>> > >>_______________________________________________
>> > >>Gmtk-users mailing list
>> > >>Gmtk-users@ssli.ee.washington.edu
>> > >>https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users
>> > >>
>> > >>
>> > >>
>> > >
>> > >
>> > >
>> > >_______________________________________________
>> > >Gmtk-users mailing list
>> > >Gmtk-users@ssli.ee.washington.edu
>> > >https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users
>> > >
>> > >
>> >
>> >
>> > --
>> > Office:   (206) 221-5216
>> > Fax:      (206) 543-3842
>> > email:    gang@ee.washington.edu
>> > homepage: welcome.to/rainier
>> >
>> > _______________________________________________
>> > Gmtk-users mailing list
>> > Gmtk-users@ssli.ee.washington.edu
>> > https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users
>> >
>>
>_______________________________________________
>Gmtk-users mailing list
>Gmtk-users@ssli.ee.washington.edu
>https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users


+======================================================================+
|            Jeff A. Bilmes, Assistant Professor                       |
| Dept. of EE              Voice: (206) 221-5236                       |
| University of Washington FAX: (206) 543-3842                         |
| Box 352500               Email: bilmes@ee.washington.edu             |
| Seattle, WA  98195-2500  http://www.ee.washington.edu/faculty/bilmes |
+======================================================================+

From bilmes@cuba.ee.washington.edu Mon Jun 20 13:31:57 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j5KKVvDf008591
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK);
	Mon, 20 Jun 2005 13:31:57 -0700
Received: from cuba.ee.washington.edu (cuba.ee.washington.edu [128.95.167.41])
	by seagull.ee.washington.edu (8.13.1/8.13.3) with ESMTP id
	j5KKVtgA000708; Mon, 20 Jun 2005 13:31:55 -0700
Message-Id: <200506202031.j5KKVtgA000708@seagull.ee.washington.edu>
To: Gang Ji <gang@ssli-mail.ee.washington.edu>,
	gmtk-users@ssli.ee.washington.edu
Subject: Re: [Gmtk-users] printing two variables 
In-Reply-To: Message from Jeff Bilmes <bilmes@ssli-mail.ee.washington.edu> 
	of "Fri, 17 Jun 2005 10:49:29 PDT."
	<200506171749.j5HHnTfT021007@seagull.ee.washington.edu> 
X-Mailer: MH-E 7.82; nmh 1.0.4; GNU Emacs 21.3.1
Date: Mon, 20 Jun 2005 13:31:55 -0700
From: Jeff Bilmes <bilmes@ssli-mail.ee.washington.edu>
X-Spam-Status: No, score=-4.7 required=5.0 tests=ALL_TRUSTED,AWL,BAYES_00 
	autolearn=ham version=3.0.3
X-Spam-Checker-Version: SpamAssassin 3.0.3 (2005-04-27) on 
	seagull.ee.washington.edu
Cc: 
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
Reply-To: "Jeff Bilmes \(206/221-5236\)" <bilmes@ee.Washington.EDU>
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Mon, 20 Jun 2005 20:31:58 -0000


Actually, Gang, I think Karen is right (now that I've more carefully
read your question). -dumpNames will dump *all* values of all variables
when they are assigned to their viterbi values. Once done, you can write
a script to do whatever you want with those values (such as extracting
values of two RVs based on some other RV at a particular set of time frames (which
could be based on any RV value)).


			-- Jeff

In the message dated Fri, 17 Jun 2005 10:49:29 PDT
  Jeff Bilmes <bilmes@ssli-mail.ee.washington.edu> writes:
>
>Actually, this is really an issue with the gmtkViterbiNew front end
>(which I'm planing to redo anyway soon). But for now, a quick
>hack to gmtkViterbiNew.cc would do the trick (basically, all the
>code is currently doing is printing out some random variables
>when the value of other RVs is a particular value, the code
>in gmtkViterbiNew.cc is *really* simple.
>
>
>			-- Jeff
>
>In the message dated Fri, 17 Jun 2005 10:31:06 PDT
>  Gang Ji <gang@ssli-mail.ee.washington.edu> writes:
>>
>>Well.  That will potentially create a random variable with BIG
>>cardinality, isn't it?
>>
>>Best,
>>Gang
>>
>>Office:   (206) 221-5216
>>Fax:      (206) 543-3842
>>email:    gang@ee.washington.edu
>>homepage: welcome.to/rainier
>>
>>On Fri, 17 Jun 2005, Karim Filali wrote:
>>
>>> If you don't mind changing your structure/master files and if A and B are
>>> already in the same clique anyway, you can add a new variable C to your
>>> graph, where C=A*(card(B))+B.  Then map output values of C to strings A_B
>>> or something like that using the -varMap option.
>>>
>>>    Karim
>>>
>>> On Fri, 17 Jun 2005, Gang Ji wrote:
>>>
>>> > Hi Karen,
>>> >
>>> > Thanks a lot for the suggestion.  What I need is the following: when
>>> > there is a word transition, I want to know the value of random variable
>>> > A and random variable B.  This is a little different from what the
>>> > "-showVitVals" does because the latter will give values at each time
>>> > frame.  What I can do is using that output and doing some
>>> > post-processing to get what I need.  I hope there is a direct way to do
>>> > that.
>>> >
>>> > Best,
>>> > Gang
>>> >
>>> > klivescu@csail.mit.edu wrote:
>>> >
>>> > >Hey Gang,
>>> > >
>>> > >Have you used "-showVitVals <filename>" or "-dumpNames <varlist>
>>> > >-ofilelist <filelist>"?  When I need multiple variables printed, I
>>> > >usually do the former and grep the output for the relevant variables,
>>> > >or the latter if I want to do a whole batch of segments.
>>> > >
>>> > >(Let me know if you find that there's actually some other way)
>>> > >
>>> > >Karen
>>> > >
>>> > >Quoting Gang Ji <gang@ssli-mail.ee.washington.edu>:
>>> > >
>>> > >
>>> > >
>>> > >>Hi,
>>> > >>
>>> > >>Is it possible to print two random variables in gmtkViterbiNew?  This is
>>> > >>*not* for debugging reasons.  I need to know two random variable
>>> > >>sequences in decoding.  Rather than decode it twice, I hope this can be
>>> > >>done in one pass.
>>> > >>
>>> > >>Best,
>>> > >>Gang
>>> > >>
>>> > >>--
>>> > >>Office:   (206) 221-5216
>>> > >>Fax:      (206) 543-3842
>>> > >>email:    gang@ee.washington.edu
>>> > >>homepage: welcome.to/rainier
>>> > >>
>>> > >>_______________________________________________
>>> > >>Gmtk-users mailing list
>>> > >>Gmtk-users@ssli.ee.washington.edu
>>> > >>https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users
>>> > >>
>>> > >>
>>> > >>
>>> > >
>>> > >
>>> > >
>>> > >_______________________________________________
>>> > >Gmtk-users mailing list
>>> > >Gmtk-users@ssli.ee.washington.edu
>>> > >https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users
>>> > >
>>> > >
>>> >
>>> >
>>> > --
>>> > Office:   (206) 221-5216
>>> > Fax:      (206) 543-3842
>>> > email:    gang@ee.washington.edu
>>> > homepage: welcome.to/rainier
>>> >
>>> > _______________________________________________
>>> > Gmtk-users mailing list
>>> > Gmtk-users@ssli.ee.washington.edu
>>> > https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users
>>> >
>>>
>>_______________________________________________
>>Gmtk-users mailing list
>>Gmtk-users@ssli.ee.washington.edu
>>https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users
>
>
>+======================================================================+
>|            Jeff A. Bilmes, Assistant Professor                       |
>| Dept. of EE              Voice: (206) 221-5236                       |
>| University of Washington FAX: (206) 543-3842                         |
>| Box 352500               Email: bilmes@ee.washington.edu             |
>| Seattle, WA  98195-2500  http://www.ee.washington.edu/faculty/bilmes |
>+======================================================================+
>_______________________________________________
>Gmtk-users mailing list
>Gmtk-users@ssli.ee.washington.edu
>https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users


+======================================================================+
|            Jeff A. Bilmes, Assistant Professor                       |
| Dept. of EE              Voice: (206) 221-5236                       |
| University of Washington FAX: (206) 543-3842                         |
| Box 352500               Email: bilmes@ee.washington.edu             |
| Seattle, WA  98195-2500  http://www.ee.washington.edu/faculty/bilmes |
+======================================================================+

From gang@ssli.ee.washington.edu Wed Jul 13 15:16:34 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j6DMGYXb026487
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@ssli.ee.washington.edu>; Wed, 13 Jul 2005 15:16:34 -0700
Received: from [128.95.167.96] (anatra.ee.washington.edu [128.95.167.96])
	(authenticated bits=0)
	by seagull.ee.washington.edu (8.13.1/8.13.3) with ESMTP id
	j6DMGWfx032742
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=NO)
	for <gmtk-users@ssli.ee.washington.edu>; Wed, 13 Jul 2005 15:16:33 -0700
Message-ID: <42D592C0.9060308@ssli.ee.washington.edu>
Date: Wed, 13 Jul 2005 15:16:32 -0700
From: Gang Ji <gang@ssli-mail.ee.washington.edu>
User-Agent: Mozilla Thunderbird 1.0.2 (X11/20050317)
X-Accept-Language: en-us, en
MIME-Version: 1.0
To: gmtk-users@ssli.ee.washington.edu
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit
X-Spam-Status: No, score=-3.8 required=5.0 tests=ALL_TRUSTED,AWL 
	autolearn=failed version=3.0.3
X-Spam-Checker-Version: SpamAssassin 3.0.3 (2005-04-27) on 
	seagull.ee.washington.edu
Subject: [Gmtk-users] htk lattice support in gmtk
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Wed, 13 Jul 2005 22:16:34 -0000

Hi all,

HTK lattice in GMTK is now up and running.  Please refer to wiki page

  https://ssli.ee.washington.edu/ssliwiki/GMTK_2fLatticeCPT

for details and documentations.


To gmtk developers:

Please write more documents in

    https://ssli.ee.washington.edu/ssliwiki/GMTK_2fDocumentation

Let's keep it updated so other user can easily find information.

Best,
Gang

-- 
Office:   (206) 221-5216
Fax:      (206) 543-3842
email:    gang@ee.washington.edu
homepage: welcome.to/rainier


From asubram@ee.washington.edu Mon Jul 18 15:34:01 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j6IMY17j014645
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@ssli.ee.washington.edu>; Mon, 18 Jul 2005 15:34:01 -0700
Received: from maxwell-a.ee.washington.edu (maxwell-a.ee.washington.edu
	[128.95.42.91])
	by seagull.ee.washington.edu (8.13.1/8.13.3) with ESMTP id
	j6IMY0Js029737
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=NO)
	for <gmtk-users@ssli.ee.washington.edu>; Mon, 18 Jul 2005 15:34:00 -0700
Received: from localhost (edgerton2.ee.washington.edu [128.95.42.88])
	by maxwell-a.ee.washington.edu (8.12.11/8.12.11) with ESMTP id
	j6IMY0LC024581
	for <gmtk-users@ssli.ee.washington.edu>; Mon, 18 Jul 2005 15:34:00 -0700
Received: from maxwell-a.ee.washington.edu ([128.95.42.7])
	by localhost (edgerton2.ee.washington.edu [128.95.42.88]) (amavisd-new,
	port 25)
	with ESMTP id 25403-10 for <gmtk-users@ssli.ee.washington.edu>;
	Mon, 18 Jul 2005 15:33:59 -0700 (PDT)
Received: from einstein.ee.washington.edu (einstein.ee.washington.edu
	[128.95.42.38])
	by maxwell-a.ee.washington.edu (8.12.11/8.12.11) with ESMTP id
	j6IMXxfi024576
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=NOT)
	for <gmtk-users@ssli.ee.washington.edu>; Mon, 18 Jul 2005 15:33:59 -0700
Date: Mon, 18 Jul 2005 15:33:59 -0700 (PDT)
From: Amar Subramanya <asubram@ee.washington.edu>
To: gmtk-users@ssli.ee.washington.edu
Message-ID: <Pine.LNX.4.58.0507181529580.2509@einstein.ee.washington.edu>
MIME-Version: 1.0
Content-Type: TEXT/PLAIN; charset=US-ASCII
X-Edgerton-Scanned: by amavisd-new at ee.washington.edu
X-Spam-Status: No, score=-4.1 required=5.0 tests=ALL_TRUSTED,AWL 
	autolearn=failed version=3.0.3
X-Spam-Level: 
X-Spam-Checker-Version: SpamAssassin 3.0.3 (2005-04-27) on 
	seagull.ee.washington.edu
Subject: [Gmtk-users] using gmtkNGramIndex
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Mon, 18 Jul 2005 22:34:01 -0000

I am trying to convert a bigram LM file from ascii to binary in order to 
speed up loading time. If I execute

gmtkNGramIndex -lmFile newbigram.lm -vocab ../DATA/voc20k94.lst -outBin true

in the path

/n/vogel/s1/asubram/wsj/bigram

the program exits with no errors but I can't seem to find the *.idx file 
that the program is supposed to produce. Am I doing something wrong here?

Amar

From gang@ssli.ee.washington.edu Mon Jul 18 16:48:00 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j6INm00H019317
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@ssli.ee.washington.edu>; Mon, 18 Jul 2005 16:48:00 -0700
Received: from [128.95.167.96] (anatra.ee.washington.edu [128.95.167.96])
	(authenticated bits=0)
	by seagull.ee.washington.edu (8.13.1/8.13.3) with ESMTP id
	j6INlwIc032613
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=NO);
	Mon, 18 Jul 2005 16:47:58 -0700
Message-ID: <42DC3FAE.6060009@ssli.ee.washington.edu>
Date: Mon, 18 Jul 2005 16:47:58 -0700
From: Gang Ji <gang@ssli-mail.ee.washington.edu>
User-Agent: Mozilla Thunderbird 1.0.2 (X11/20050317)
X-Accept-Language: en-us, en
MIME-Version: 1.0
To: Amar Subramanya <asubram@ee.washington.edu>
Subject: Re: [Gmtk-users] using gmtkNGramIndex
References: <Pine.LNX.4.58.0507181529580.2509@einstein.ee.washington.edu>
In-Reply-To: <Pine.LNX.4.58.0507181529580.2509@einstein.ee.washington.edu>
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit
X-Spam-Status: No, score=-5.4 required=5.0 tests=ALL_TRUSTED,AWL,BAYES_00 
	autolearn=ham version=3.0.3
X-Spam-Checker-Version: SpamAssassin 3.0.3 (2005-04-27) on 
	seagull.ee.washington.edu
Cc: gmtk-users@ssli.ee.washington.edu
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Mon, 18 Jul 2005 23:48:00 -0000

Some one commented the code so that it actually does nothing because of 
"if 0" in the macro.  :)

I have a quick fix in /u/gang/research/programs/gmtk/tksrc/

Best,
Gang

Amar Subramanya wrote:

>I am trying to convert a bigram LM file from ascii to binary in order to 
>speed up loading time. If I execute
>
>gmtkNGramIndex -lmFile newbigram.lm -vocab ../DATA/voc20k94.lst -outBin true
>
>in the path
>
>/n/vogel/s1/asubram/wsj/bigram
>
>the program exits with no errors but I can't seem to find the *.idx file 
>that the program is supposed to produce. Am I doing something wrong here?
>
>Amar
>_______________________________________________
>Gmtk-users mailing list
>Gmtk-users@ssli.ee.washington.edu
>https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users
>  
>


-- 
Office:   (206) 221-5216
Fax:      (206) 543-3842
email:    gang@ee.washington.edu
homepage: welcome.to/rainier


From asubram@ee.washington.edu Mon Jul 18 17:05:22 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j6J05MVX020267
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK);
	Mon, 18 Jul 2005 17:05:22 -0700
Received: from maxwell-a.ee.washington.edu (maxwell-a.ee.washington.edu
	[128.95.42.91])
	by seagull.ee.washington.edu (8.13.1/8.13.3) with ESMTP id
	j6J05LJp001056
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=NO);
	Mon, 18 Jul 2005 17:05:21 -0700
Received: from localhost (edgerton2.ee.washington.edu [128.95.42.88])
	by maxwell-a.ee.washington.edu (8.12.11/8.12.11) with ESMTP id
	j6J05Lrl030231; Mon, 18 Jul 2005 17:05:21 -0700
Received: from maxwell-a.ee.washington.edu ([128.95.42.7])
	by localhost (edgerton2.ee.washington.edu [128.95.42.88]) (amavisd-new,
	port 25)
	with ESMTP id 06109-05; Mon, 18 Jul 2005 17:05:19 -0700 (PDT)
Received: from einstein.ee.washington.edu (einstein.ee.washington.edu
	[128.95.42.38])
	by maxwell-a.ee.washington.edu (8.12.11/8.12.11) with ESMTP id
	j6J05Jer030179
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=NOT);
	Mon, 18 Jul 2005 17:05:19 -0700
Date: Mon, 18 Jul 2005 17:05:19 -0700 (PDT)
From: Amar Subramanya <asubram@ee.washington.edu>
To: Gang Ji <gang@ssli-mail.ee.washington.edu>
Subject: Re: [Gmtk-users] using gmtkNGramIndex
In-Reply-To: <42DC3FAE.6060009@ssli.ee.washington.edu>
Message-ID: <Pine.LNX.4.58.0507181705070.11841@einstein.ee.washington.edu>
References: <Pine.LNX.4.58.0507181529580.2509@einstein.ee.washington.edu>
	<42DC3FAE.6060009@ssli.ee.washington.edu>
MIME-Version: 1.0
Content-Type: TEXT/PLAIN; charset=US-ASCII
X-Edgerton-Scanned: by amavisd-new at ee.washington.edu
X-Spam-Status: No, score=-5.5 required=5.0 tests=ALL_TRUSTED,AWL,BAYES_00 
	autolearn=ham version=3.0.3
X-Spam-Level: 
X-Spam-Checker-Version: SpamAssassin 3.0.3 (2005-04-27) on 
	seagull.ee.washington.edu
Cc: gmtk-users@ssli.ee.washington.edu
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Tue, 19 Jul 2005 00:05:23 -0000

Thanks, that works.

Amar

On Mon, 18 Jul 2005, Gang Ji wrote:

> Some one commented the code so that it actually does nothing because of 
> "if 0" in the macro.  :)
> 
> I have a quick fix in /u/gang/research/programs/gmtk/tksrc/
> 
> Best,
> Gang
> 
> Amar Subramanya wrote:
> 
> >I am trying to convert a bigram LM file from ascii to binary in order to 
> >speed up loading time. If I execute
> >
> >gmtkNGramIndex -lmFile newbigram.lm -vocab ../DATA/voc20k94.lst -outBin true
> >
> >in the path
> >
> >/n/vogel/s1/asubram/wsj/bigram
> >
> >the program exits with no errors but I can't seem to find the *.idx file 
> >that the program is supposed to produce. Am I doing something wrong here?
> >
> >Amar
> >_______________________________________________
> >Gmtk-users mailing list
> >Gmtk-users@ssli.ee.washington.edu
> >https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users
> >  
> >
> 
> 
> 

From bilmes@cuba.ee.washington.edu Sat Jul 23 00:19:12 2005
Received: from seagull.ee.washington.edu (ssli-mail.ee.washington.edu
	[128.95.167.7])
	by crow.ee.washington.edu (8.13.1/8.13.1) with ESMTP id j6N7JC2f028818
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <gmtk-users@crow.ee.washington.edu>; Sat, 23 Jul 2005 00:19:12 -0700
Received: from cuba.ee.washington.edu (cuba.ee.washington.edu [128.95.167.41])
	by seagull.ee.washington.edu (8.13.1/8.13.3) with ESMTP id
	j6N7J9If016219; Sat, 23 Jul 2005 00:19:09 -0700
Message-Id: <200507230719.j6N7J9If016219@seagull.ee.washington.edu>
To: gmtk-users@seagull.ee.washington.edu
X-Mailer: MH-E 7.82; nmh 1.0.4; GNU Emacs 21.3.1
Date: Sat, 23 Jul 2005 00:19:09 -0700
From: Jeff Bilmes <bilmes@ssli-mail.ee.washington.edu>
X-Spam-Status: No, score=-3.7 required=5.0 tests=ALL_TRUSTED,AWL,BAYES_00,
	TO_MALFORMED autolearn=ham version=3.0.3
X-Spam-Checker-Version: SpamAssassin 3.0.3 (2005-04-27) on 
	seagull.ee.washington.edu
Cc: alex@neisis.net
Subject: [Gmtk-users] GMTK dev tag update
X-BeenThere: gmtk-users@ssli.ee.washington.edu
X-Mailman-Version: 2.1.2
Precedence: list
Reply-To: "Jeff Bilmes \(206/221-5236\)" <bilmes@ee.washington.edu>
List-Id: Users of GMTK. Discussions, questions, etc.
	<gmtk-users.ssli.ee.washington.edu>
List-Unsubscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=unsubscribe>
List-Archive: <https://ssli.ee.washington.edu/mailman/private/gmtk-users>
List-Post: <mailto:gmtk-users@ssli.ee.washington.edu>
List-Help: <mailto:gmtk-users-request@ssli.ee.washington.edu?subject=help>
List-Subscribe: <https://ssli.ee.washington.edu/mailman/listinfo/gmtk-users>, 
	<mailto:gmtk-users-request@ssli.ee.washington.edu?subject=subscribe>
X-List-Received-Date: Sat, 23 Jul 2005 07:19:12 -0000


Two things of note in the GMTK world, namely Bayesian Dirichlet priors
and graph visualization.

=================================================================-

First, for those wishing just to get the latest static binaries, see:

   http://ssli.ee.washington.edu/~bilmes/gmtk/linux/FriJul22_2005/


=================================================================-

GMTK now supports Bayesian Dirichlet priors inside of EM training for
multinomial distributions.  Specifically, Dirichlet priors can be used
for DenseCPTs, DPMFs, SparseCPTs (indirectly via DPMFs), and mixture
responsibilities (again via DPMFs).

Dense CPTs:

There are two ways to use Dirichlet priors. 

Case A: One way is to have a *constant* count value added to all CPT
accumulators. This means that during learning, the expected counts for
the CPT entries will have an added constant value of 'alpha':

 P(C=j|parents) = (1/Z) ( E[ counts of j | parents ] + alpha )

where alpha is the constant Dirichlet hyperparameter (count), and where
Z is a normalization constant. Without Dirichlet smoothing, it is always
the case that alpha == 0.

Case B: Another way is to have a table of count values for the CPT, as
in:

 P(C=j|parents) = (1/Z) ( E[ counts of j | parents ] + alpha(j,parents) )

so the Dirichlet counts consist of a multi-dimensional table that is the
same size and dimensionality as the CPT. The change is entirely
backwardly compatible with existing DenseCPTs and DPMFs, so if you don't
want to use them, no changes are needed.

Here is the syntax for the above two cases:

Case A:

Here is a dense CPT with a constant Dirichlet count of 5:

---------------------------------------------
1 % cpt number 1
state1_with_state_pars % name
1 % num parents
3 3 % cards
DirichletConst 5
0.3333 0.3333 0.3333
0.3333 0.3333 0.3333
0.3333 0.3333 0.3333
---------------------------------------------

In other words, you can optionally include a 'DirichletConst v' before
the Dense CPT probabilities, where v is the shared Dirichlet
hyperparameter for all values of the child RV and for all parent
values. The value 'v' is a floating point value (so fractional counts
allowed), must be >= 0, and is a real count value (i.e., it is not in
log form).


Case B:

Here, you associate a Dirichlet Table with a CPT, as follows:

-----------------------
1 % cpt number 1
state1_with_state_pars % name
1 % num parents
3 3 % cards
DirichletTable 3x3dirichlet_tab
0.3333 0.3333 0.3333
0.3333 0.3333 0.3333
0.3333 0.3333 0.3333
-----------------------

where '3x3dirichlet_tab' is the name of a new GMTK object, of type
Dirichlet Table (these will be described below).

---------------------------
DPMFs/SparseCPTs:

Recall, that DPMFs are used for both (Gaussian) mixture
responsibilities, and also for the non-zero values of the rows of Sparse
CPTs. For using Dirichlet Priors with mixture responsibilities or
SparseCPTs, the Dirichlet prior are associated with the underlying DPMF.

Also recall that when DPMFs are used for responsibilities, they can
change length (i.e., when a mixture split or vanish occurs). Therefore,
if a split/vanish occurs, and it is currently using a DirichletTable, if
the length no longer matches, then the DirichletTable is no longer used
(i.e., GMTK will turn off smoothing). With high enough verbosity
(currently 50), you'll get a warning message when this occurs.

The syntax for DPMFs is very similar to DenseCPTs, specifically, we can
either have:

-----------------------
0 % pmf number 0
gmMixWeight0 2 % name, cardinality
DirichletConst 10
0.5 0.5
-----------------------

or

-----------------------
0 % pmf number 0
gmMixWeight0 2 % name, cardinality
DirichletTable table-name
0.5 0.5
-----------------------


---------------------------------------------

DirichletTables are a new GMTK object that store the Dirichlet prior
hyperparameters for each child RV value for all parent values. The
format is basically a N-Dimensional table. Here are two examples:


---------------------------------------------------------------------------------
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Dirichlet Tables 
DIRICHLET_TAB_IN_FILE inline
2

0 % dirichlet number 0
3dirichlet_tab % name
1 % table dimensionality
3 % dimensionalities
% meta counts
5 4 3

1 % dirichlet number 1
3x3dirichlet_tab % name
2 % table dimensionality
3 3 % dimensionalities
% meta counts
1e2 0   0
0   1e2 0
0   0  1e2
---------------------------------------------------------------------------------

The first table (named '3dirichlet_tab' defines a 1-dimensional length 3 table
with prior counts '5 4 3'. Such a table can be either used with a 3-element
DPMF, or a DenseCPT with no parents and child cardinality of 3.

The second table '3x3dirichlet_tab' defines a 2-dimensional 3x3
table. If this was a 3x3 Markov transition matrix, such priors would
encourage the states to do a self-transition (since the table is
diagonal).

When a Dirichlet Table is associate with a DenseCPT or DPMF, the
dimensionalities must match *exactly*! Also, the data order for each
dimension is exactly the same as for a DenseCPT. This means that if a
DenseCPT has d parents, the DirichletTable must have dimensionality
(d+1). All but the last dimension length of the DirichletTable must
match the cardinality of the corresponding parent of the DenseCPT, and
the last dimension length of the DirichletTable must match the
cardinality of the child used in the DenseCPT.

Note that, like 'DirichletConst', the table values are actual
real-valued counts, and that they must be >= 0.


Training: lastly, in order to get all this to work, you must use the
gmtkEMTrainNew program, and you must give the command line option
'-dirichletPriors T'. The default value for the command line parameter
is false. 

The reason for this command line option is that if you are using GMTK in
parallel training mode with Dirichlet priors, you only want to turn on priors for
one process rather than all of them so that the prior counts don't get
counted as many times as you have divided the training into parallel
chunks. Thus, your parallel script would call something like:

  gmtkEMTrainnew -dirichletPriors T  {parallel chunk 1 parameters}
  gmtkEMTrainnew -dirichletPriors F  {parallel chunk 2 parameters}
  gmtkEMTrainnew -dirichletPriors F  {parallel chunk 3 parameters}
     ...
  gmtkEMTrainnew -dirichletPriors F  {parallel chunk N parameters}


It of course doesn't matter which chunk gets the '-dirichletPriors T' as
long as only one of them does. If you are not using Dirichlet priors in
any of your CPTs for DPFMs, then this command line option has no effect.


=================================================================-

There has been substantial progress on gmtkViz, the GMTK visualization
tool.  Alex Norman <alex@neisis.net> has been working on extending and
fixing bugs.

gmtkViz is not made by default since it depends on having the library
wxWidgets available. When you check out the latest version, you can try
a 'make gmtkViz' to make the gmtkViz program.

As I mentioned before, gmtkViz does not do graph layout for you, but
makes it really easy to do so, and you can export postscript to include
into your latex/msword files. 

The latest changes have fixed many bugs (including printing) and added a
ton of new features. There is also an online quick-reference-quide when
you run the program. Please send any questions/comments to Alex Norman
<alex@neisis.net>, but only for the next few weeks since he'll be
leaving on Aug 19th. Please do try it out if you can though.



			-- Jeff



+======================================================================+
|            Jeff A. Bilmes, Associate Professor                       |
| Dept. of EE              Voice: (206) 221-5236                       |
| University of Washington FAX: (206) 543-3842                         |
| Box 352500               Email: bilmes@ee.washington.edu             |
| Seattle, WA  98195-2500  http://www.ee.washington.edu/faculty/bilmes |
+======================================================================+


