------------------------------------------------------------------------------------- ENTRIES ------------------------------------------------------------------------------------- 2014-05-29 12:00:00 -- Feature analysis 'set100' 2014-05-30 12:00:00 -- GMTK observation files 2014-05-31 12:00:00 -- GMTK data management files (triangulation and train) 2014-06-01 12:00:00 -- Graphical Model structure decisions 2014-06-02 12:00:00 -- GMTK model structure and parameters files 2014-06-03 12:00:00 -- GMTK pre-training commands 2014-06-04 12:00:00 -- GMTK triangulation command 2014-06-05 12:00:00 -- GMTK initial training 2014-06-09 02:12:22 -- Readings on GM-I 2014-06-12 00:17:30 -- GMTK training experiments-I 2014-06-12 00:28:00 -- GMTK (parallel) training experiments-I (2) 2014-06-12 22:33:23 -- GMTK training experiments-I (3) - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 2014-06-14 22:03:11 -- GMTK training experiment - II - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 2017-11-05 22:00:06 -- Resume w/ IA2 proj. (n.a.) ------------------------------------------------------------------------------------- CONTENTS ------------------------------------------------------------------------------------- 2014-05-29 12:00:00 -- Feature analysis 'set100' > Extract MFCCs from 'set100' dataset, using HCopy (HTK) > Generate the SCP file: bash script 'job_generate-SCP-MFCC' >>>>> DIY is hard only at the beginning, ... what a pity! 2014-05-30 12:00:00 -- GMTK observation files > see parameters (HTK format observation file: octave script 'seeParams.m' 2014-05-31 12:00:00 -- GMTK data management files (triangulation and train) > Correct sentence sent-00013.wav was "corrupted" ... ok > Get length files sent-?????.mfc : script 'mfccLenghts.m' > Generate sent-?????.lab files and the respective SCP and conversion table ('phonesXtd2index.txt') files: script 'generateLabScp.pl' (debugged) > Data management (basic) files built :) 2014-06-01 12:00:00 -- Graphical Model structure decisions > Decisions on the GM structure: version-0.0 based on the 'timit' example 2014-06-02 12:00:00 -- GMTK model structure and parameters files > Define TM structure and params (write/copy from timit ex -> mavGmtk-0.0/strPrm) 2014-06-03 12:00:00 -- GMTK pre-training commands > Learning/using gmtkParmConvert ... ok > Learning/using gmtkModelInfo ... ok > Learning/using gmtkTriangulate : results with different heuristics (S|N options+) at 'triang-heuristic.ods' 2014-06-04 12:00:00 -- GMTK triangulation command > Continue learning/using (other options) gmtkTriangulate ... ok 2014-06-05 12:00:00 -- GMTK initial training > Run initial training iterations (w/ baseline config) using gmtkEMtrain ... ok 2014-06-09 02:12:22 -- Readings on GM-I > Continue selected readings (hcopy LB - pp.27) docs: - GM struct, parm ... ok - GMTK triang, jt, ... ~ (2cont) ::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::: . «always improve your intellectual property as solid as you think it is . . appropriate, then for sure nobody can deprive you of creating and . . building, anytime and anyplace, really useful things.» . . Jos\'e dos Anjos de Pera (quoting my papa, he IS the greatest!) . ::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::: 2014-06-12 00:17:30 -- GMTK training experiments-I > exp20140611: JT (jt_info.txt) EM training (config) options > LL(it), means*(it), vars*(it) > scripts: traincommand > ( getLoglik | par2seq > plotseq.m ) 2014-06-12 00:28:00 -- GMTK (parallel) training experiments-I (2) > exp20140612: traincommand_* generate partial (data subset *) EM accumulators; run (parallel) using [multiple terminals] or using [nohup ./traincommand_* &] or ... traincommand at the end bundle accumulators and update params > scripts: traincommand_* traincommand 2014-06-12 22:33:23 -- GMTK training experiments-I (3) > exp20140613: 1) beam options (1st contact): cbeam, cpbeam, ebeam, sbeam 2) training only part of the objects > scripts: 1; or other gmtkEMtrain options tuning) traincommand (after define 'paramValues') > getTimes > plotTimes.m 2) traincommand.objsNot2train (after define 'objsNot2train') - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 2014-06-14 22:03:11 -- GMTK training experiment - II > exp20140616: Execute training schedule (suggested in the "GMTK Reference"): Step 0: Start training w/ 1 GC per GM ... ok ........................................................................... experiment w/ basic model and very little data -> adapt the following ... ........................................................................... Step 1: ?-EM iter w/o split/vanish until 2% conv Step 2: 1-EM iter w/ split all GCs Step 3: ?-EM iter w/o split/vanish until 2% conv Repeat steps 1-3 a few times (typically 5) Step 4: 1-EM iter w/ mcvr = 10, mcsr = 1 (some GCs possibly too sharp => vanishing) Step 5: ?-EM iter w/o split/vanish until 2% conv Repeat steps 4-5 a few times (when some GCs too sharp ... next step) Step 6: 1-EM iter w/ mcvr = 10, mcsr = 1 Step 7: 1-EM iter w/ mcvr = 10, mcsr = 1e10 (just to kill off some of the weak GCs) Step 8: ?-EM iter w/ split/vanish until 2% conv Repeat steps 6-8 a few times Step 9: 2-EM iter w/ mcvr = 10, mcsr = 5 Step 10: ?-EM iter w/ no split/vanish until 0.2% conv. ------------------------------------------------------------------------------------- Useful vim commands: :Cagora - print date & time :Cvai - jump to content (to use in the "ENTRIES" section) -------------------------------------------------------------------------------------