📄 changelog
字号:
2005-02-24 Anna Langley <jal58@cam.ac.uk> * Files for autoconf installs added2005-02-08 S.E. Tranter <sej28@eng.cam.ac.uk> * HCluster.c: Added support for diagonal and block diagonal covariance representation with BIC stopping criterion. Block diagonal case is slightly different to previous local version BIC scoring info now requires trace level 1 to be output. 2004-08-04 Antti-Veikko Rosti <avir2@eng.cam.ac.uk> * HERest.c: ConvLogWt() and ConvExpWt() added after ConvDiagC() calls to convert the mixture weights into the log-scale prior to forward-backward recursions. * HHEd.c: Function ReOrderFeaturesCommand() added to sort model parameters according to a global feature element re-ordering with RP command.2004-08-02 Gunnar Evermann <ge204@eng.cam.ac.uk> * HVite.c (main): Check that number of tokens requested via -n switch doesn't exceed maximum.2004-07-120 Khe Chai Sim <kcs23@eng.cam.ac.uk> * HHEd.c: Support for generic tree name mapping. Current mapping makes use of the macRoot prefix specified with the TB command.2004-07-14 Gunnar Evermann <ge204@eng.cam.ac.uk> * HMMIRest.c (main): Employ LLFs for faster loading of lattices. (main): Add LATFILEMASK to be able to discard parts of the segment name when loading the lattice file.2004-07-14 Antti-Veikko Rosti <avir2@eng.cam.ac.uk> * HERest.c: New heap for accumulators called accStack created. This is used in AttachAccs() instead of gstack. * HMMIRest.c: Removed unused variables and added T_TIM trace flag. The heap globStack renamed to accStack.2004-07-08 Mark Gales <mjfg@eng.cam.ac.uk> * HHEd.c (RegClassesCommand): Correctly treat base classes stored in directories specified via -M.2004-07-02 B. Jia <bj214@eng.cam.ac.uk> * HMMIRest.c: Fix a bug of single pass retrain in discrimitive trainng2004-06-25 Gunnar Evermann <ge204@eng.cam.ac.uk> * *.c (traceHFB): Eliminate traceHFB variable and replace with function call.2004-06-06 B. Jia <bj214@eng.cam.ac.uk> * HMMIRest.c: New config variable UPDATEMODE to allow dumping summed accumulates.2004-04-26 K. Yu <ky219@eng.cam.ac.uk> * HMMIRest.c: Add support for MMI priors in MPE training (new config variables MMIPRIOR and MMITAUI).2004-04-17 Gunnar Evermann <ge204@eng.cam.ac.uk> * HERest.c (main): Fix for UPDATEMODE to make non-parallel mode training work again.2004-04-16 Gunnar Evermann <ge204@eng.cam.ac.uk> * HERest.c: New config variable UPDATEMODE to allow dumping summed accumulates (subsumes functionality of HAccSum).2004-04-13 B. Jia <bj214@eng.cam.ac.uk> * HMMIRest.c: New options '-rp' and '-qp' to support subdirectories of denominator and numerator lattices respectively, so that all side directories will not crowd under one directory.2004-04-02 Bin Jia <bj214@eng.cam.ac.uk> * HCompV: New option '-p' to support storing CMN/CVN files in subdirectories for use with CMEANPATHMASK and VARSCALEPATHMASK.2004-03-21 Mark Gales <mjfg@eng.cam.ac.uk> * *.c: Various fixed for use of new adaptation.2004-02-16 Bin Jia <bj214@eng.cam.ac.uk> * HDMan.c: MAXDICTS increased from 20 to 100 2003-12-17 Gunnar Evermann <ge204@eng.cam.ac.uk> * HVite.c (DoAlignment): Allow alignment of segments with empty transcriptions if an utterance boundary word (option -b) was specified.2003-12-10 S.E. Tranter <sej28@eng.cam.ac.uk> * HCluster.c: Added -y option to force splits to be binary This changed the functions: MakeInitialAssignment, MakeInitialModuloAssignment Added a -z option to allow a BIC-based stopping criterion. This is currently only implemented for a 2-way split and for a full covariance (or correlation) matrix representation - i.e. it does not yet include block-diagonal or diagonal covariance Two implementations of the BIC-based criterion are implemented - the choice of which is made on the sign of the -z parameter. Negative means a "local" decision - in that the penalty factor uses N_z, the number of frames in the cluster being split, whereas positive means a "global" decision - i.e the penalty factor uses the total number of frames input to the clusterer. The absolute value of the z parameter indicates the value of alpha in the penalty formula. This uses the new functions: BICScore, ConvertCorrelationToCovariance, CalcJointMean, CalcJointCov Some maths functions from ~mjfg/cuhtk/HTKLib/HMath.c have also been included. This is because the existing SVD implementation sometimes goes into an infinite loop, so the alternative code (which contains nr stuff so cant be released) has been included. However, this is not yet called anywhere, since the icc version (e.g. compiled on spot) doesn't seem to produced this problem. This involves the functions: PYTHAG, MAXX, SIGN, SVDCmp, InvSVD_cuhtk 2003-10-14 Anna Langley <jal58@eng.cam.ac.uk> * htk_htktools_nt.mkf: Use backslashes in pathnames.2003-10-14 Gunnar Evermann <ge204@eng.cam.ac.uk> * HLEd.c: Rename DELETE as Visual C++ doen't seem to like it.2003-09-10 Gunnar Evermann <ge204@eng.cam.ac.uk> * HHEd.c (ParseAlpha): Fix parsing of octal charcter codes. From Tanel Alum鋏 <tanel.alumae@aqris.com>2003-07-28 Anna Langley <jal58@eng.cam.ac.uk> * htk_htktools_nt.mkf: was not building HLRescore, applied changes supplied by Arlindo Veiga which corrects this. 2003-07-24 M. J. F. Gales <mjfg@eng.cam.ac.uk> * HERest.c: clarify the single pass retraining when the TARGETKIND does not change. 2003-07-08 Anna Langley <jal58@eng.cam.ac.uk> * HERest.c, HHEd.c, HVite.c, HMMIRest.c, HEAdaptLat.c: removed unused variables from sources. 2003-06-18 M. J. F. Gales <mjfg@eng.cam.ac.uk> * HERest.c (MLUpdateModels): maxM uninitialised. This means that the weights are not always guaranteed to be updated when there are multiple components.2003-04-05 S.E. Tranter <sej28@tigger> * HCluster.c (main): The Minimum number of frames required to attempt to do full covariance matrix inversion is set to at least the dimension of the feature vector + 1 New functions: InvertTriMatSVD to invert a TriMat covariance matrix using SVD decomposition (more stable than CovInvert) IsCovInvertableSVD : checks if SVD inversion has been relatively stable StoreCovsSVD : reordering of StoreCovs and inversion of matrix using InvertTriMatSVD, so the inversion is now done before checking if the matrix is invertable Modified functions: StoreCovsSVD now called instead of StoreCovs in main loop MakeClusterStats now calls InvertTriMatSVD not CovInvert 2003-03-30 Gunnar Evermann <ge204@eng.cam.ac.uk> * HLRescore.c (ProcessLattice): Try to make lattice node order deterministic by setting ln->score to node posterior before saving (can be turned off by setting SORTLATTICE=F).2003-03-17 Sue Tranter <sej28@ceres> * HCluster.c (SetConfParms): Bug fix to allow feature vectors of other than 13 dimension. SetConfParams now adds the "base_dimen" if it sees _D and _A, rather than adding 13. Also, it checks that the minvar is not set to a different value in the cfgfile and command line ReportUsage changes to allow "-w" option to set base_dimen on command line (_D and _A should be set in cfg file) Main : moved SetConfParms and setting dimension to after dealing with input args. (to allow -w) -w option added to set base dimension on commandline ShowParameters: ordering changed to be alphabetical and -v and -w options added 2003-03-20 Dan Povey <dp10006@eng.cam.ac.uk * HMMIRest.c: Bug fix for transition update.2003-03-13 Gunnar Evermann <ge204@eng.cam.ac.uk> * HVite.c (ProcessFile): Output current beam if alignment failed and is restartable.2003-03-12 Gunnar Evermann <ge204@eng.cam.ac.uk> * HERest.c (main): Undo change to total logprob calculation.2003-03-12 Lan Wang <lw256@eng.cam.ac.uk> * HMMIRest.c (SetConfParms): New config variables INXFORMMASK and PAXFORMMASK. 2003-03-10 Gunnar Evermann <ge204@eng.cam.ac.uk> * HERest.c: Remove FIXUP again.2003-03-10 Dan Povey <dp10006@eng.cam.ac.uk> * HMMIRest.c: new config vars: VARFLOORPERCENTILE and VARSMOOTH * HERest.c: new config vars: VARFLOORPERCENTILE and FIXUP2003-02-27 S.E. Tranter <sej28@eng.cam.ac.uk> * Makefile: Added HCluster to makefile * HCluster.c (main): printing of numargs removed printing of dimension information needs T_TOP CloseAndExit function added to enable program to exit safely at different stages Bugfix for the case where all the segs are too small for main clustering decisions (assigns them all to 1 cluster) Code which allowed an HPARM2 parameter (_Z) to be set has been commented out, since it caused a seg fault and may have been an interaction problem with a new HParm.c, since this didn't happen previously. This option was not currently being used in the config file anyway.2003-02-25 Mark Gales <mjfg@eng.cam.ac.uk> * HHEd.c: Bug fixes in regression tree clustering. From Jos
⌨️ 快捷键说明
复制代码
Ctrl + C
搜索代码
Ctrl + F
全屏模式
F11
切换主题
Ctrl + Shift + D
显示快捷键
?
增大字号
Ctrl + =
减小字号
Ctrl + -