[Subject Prev][Subject Next][Thread Prev][Thread Next][Date Index][Thread Index]

[hts-users:04500] Re: informativeness of questions

Thanks very much, Heiga - I will give that a try!


On Thu, Mar 9, 2017 at 2:59 AM, Heiga ZEN (Byung Ha CHUN) <heigazen@xxxxxxxxxx> wrote:
Hi Erica,

HHEd has "TR" command to control log levels.  Here are log levels related to clustering.

#define T_CLUSTERS 0x0100   /* Show contents of clusters */
#define T_QST 0x0200        /* Show items specified by questions */
#define T_TREE_ANS 0x0400   /* Trace best unseen triphone tree filtering */
#define T_TREE_BESTQ 0x0800 /* Trace best question for splitting each node */
#define T_TREE_BESTM 0x1000 /* Trace best merge of terminal nodes  */
#define T_TREE_OKQ 0x2000   /* Trace all questions exceeding threshold */
#define T_TREE_ALLQ 0x4000  /* Trace all questions */
#define T_TREE_ALLM 0x8000  /* Trace all possible merges */

So if you run

TR 2049

(2049 = 0x0801)  just before TB commands, (hopefully) HTS will output both normal log plus best questions at each node.  Likewise, if you turn on T_TREE_ALL_Q, it will generate log for all questions.  You can try other values or just comment-out related lines in the HHEd.c:ValidProbNode() to generate this information.



On Wed, Mar 8, 2017 at 8:48 PM Erica Cooper <ecooper@xxxxxxxxxxxxxxx> wrote:
Hi all,

Is there a way to find out how informative a given question was in the decision tree construction, e.g. by finding out what was the likelihood gain at a given node split?  I haven't found anything like this in the training log output or in the tree files themselves, but maybe this information is available elsewhere in the training output? or would I have to modify the code to log this and re-train?

Heiga ZEN (in Japanese)
Byung Ha CHUN (in Korean)

[hts-users:04498] informativeness of questions, Erica Cooper
[hts-users:04499] Re: informativeness of questions, Heiga ZEN (Byung Ha CHUN)