/* UCollectData: Collect data from segStore for each stream s of each
state n and store in seqMat[n][s]*/
void UCollectData(Sequence **seqMat)
{
int i,j,s,numSegs,segLen;
long int n;
float obsPerState;
Observation obs;
Ptr p;
numSegs = NumSegs(segStore);
for (i=1;i<=numSegs;i++) {
segLen=SegLength(segStore,i);
obsPerState=((float) segLen)/((float) (nStates-2));
if (obsPerState < 1.0)
HError(2122,"UCollectData: segment too short[%d]",segLen);
for (j=1;j<=segLen;j++) {
obs = GetSegObs(segStore,i,j);
n = (long int)(((float)(j-1)/obsPerState)+2);
for (s=1; s<=nStreams; s++){
if (hset.hsKind==DISCRETEHS){
p = (Ptr)((long int)obs.vq[s]);
StoreItem(seqMat[n][s],p);
}else
StoreItem(seqMat[n][s],obs.fv[s]);
}
}
}
}
/* UCollectData: Collect data from segStore for each stream s of each
state n and store in seqMat[n][s]*/
void UCollectData(Sequence ***seqMat)
{
int i,j,k,n,s,numSegs,segLen,order;
float obsPerState;
Observation obs;
Ptr p;
numSegs = NumSegs(segStore);
for (i=1;i<=numSegs;i++) {
segLen=SegLength(segStore,i);
obsPerState=((float) segLen)/((float) (nStates-2));
if (obsPerState < 1.0)
HError(2122,"UCollectData: segment too short[%d]",segLen);
for (j=1;j<=segLen;j++) {
obs = GetSegObs(segStore,i,j);
n = (int)(((float)(j-1)/obsPerState)+2);
for (s=1; s<=nStreams; s++){
if (hset.hsKind==DISCRETEHS){
p = (Ptr)((long)obs.vq[s]);
StoreItem(seqMat[n][s][1],p);
} else if(hset.msdflag[s]){
order = SpaceOrder(obs.fv[s]);
if ((k = IncludeSpace(msdInfo[n][s],order)))
StoreItem(seqMat[n][s][k],obs.fv[s]);
else if(!ignOutVec)
HError(2122,"UCollectData: no space corresponded to order[%d]",order);
}else
StoreItem(seqMat[n][s][1],obs.fv[s]);
}
}
}
}
/* EstimateModel: top level of iterative estimation process */
void EstimateModel(void)
{
LogFloat totalP,newP,delta;
Boolean converged = FALSE;
int i,iter,numSegs,segLen;
IntVec states; /* array[1..numSegs] of State */
IntVec *mixes; /* array[1..S][1..numSegs] of MixComp */
if (trace&T_TOP) printf("Starting Estimation Process\n");
if (newModel){
UniformSegment();
}
totalP=LZERO;
for (iter=1; !converged && iter<=maxIter; iter++){
ZeroAccs(&hset, uFlags); /* Clear all accumulators */
numSegs = NumSegs(segStore);
/* Align on each training segment and accumulate stats */
for (newP=0.0,i=1;i<=numSegs;i++) {
segLen = SegLength(segStore,i);
states = CreateIntVec(&gstack,segLen);
mixes = (hset.hsKind==DISCRETEHS)?NULL:
CreateMixes(&gstack,segLen);
newP += ViterbiAlign(i,segLen,states,mixes);
if (trace&T_ALN) ShowAlignment(i,segLen,states,mixes);
UpdateCounts(i,segLen,states,mixes);
FreeIntVec(&gstack,states); /* disposes mixes too */
}
/* Update parameters or quit */
newP /= (float)numSegs;
delta = newP - totalP;
converged = (iter>1) && (fabs(delta) < epsilon);
if (!converged)
UpdateParameters();
totalP = newP;
if (trace & T_TOP){
printf("Iteration %d: Average LogP =%12.5f",iter,totalP);
if (iter > 1)
printf(" Change =%12.5f\n",delta);
else
printf("\n");
fflush(stdout);
}
}
if (trace&T_TOP) {
if (converged)
printf("Estimation converged at iteration %d\n",iter);
else
printf("Estimation aborted at iteration %d\n",iter);
fflush(stdout);
}
}
int main(int argc, char *argv[])
{
char *datafn, *s;
int nSeg;
void Initialise(void);
void LoadFile(char *fn);
void EstimateModel(void);
void SaveModel(char *outfn);
if(InitShell(argc,argv,hinit_version,hinit_vc_id)<SUCCESS)
HError(2100,"HInit: InitShell failed");
InitMem(); InitLabel();
InitMath(); InitSigP();
InitWave(); InitAudio();
InitVQ(); InitModel();
if(InitParm()<SUCCESS)
HError(2100,"HInit: InitParm failed");
InitTrain(); InitUtil();
if (!InfoPrinted() && NumArgs() == 0)
ReportUsage();
if (NumArgs() == 0) Exit(0);
SetConfParms();
CreateHMMSet(&hset,&gstack,FALSE);
while (NextArg() == SWITCHARG) {
s = GetSwtArg();
if (strlen(s)!=1)
HError(2119,"HInit: Bad switch %s; must be single letter",s);
switch(s[0]){
case 'e':
epsilon = GetChkedFlt(0.0,1.0,s); break;
case 'i':
maxIter = GetChkedInt(0,100,s); break;
case 'l':
if (NextArg() != STRINGARG)
HError(2119,"HInit: Segment label expected");
segLab = GetStrArg();
break;
case 'm':
minSeg = GetChkedInt(1,1000,s); break;
case 'n':
newModel = FALSE; break;
case 'o':
outfn = GetStrArg();
break;
case 'u':
SetuFlags(); break;
case 'v':
minVar = GetChkedFlt(0.0,10.0,s); break;
case 'w':
mixWeightFloor = MINMIX * GetChkedFlt(0.0,10000.0,s);
break;
case 'B':
saveBinary = TRUE;
break;
case 'F':
if (NextArg() != STRINGARG)
HError(2119,"HInit: Data File format expected");
if((dff = Str2Format(GetStrArg())) == ALIEN)
HError(-2189,"HInit: Warning ALIEN Data file format set");
break;
case 'G':
if (NextArg() != STRINGARG)
HError(2119,"HInit: Label File format expected");
if((lff = Str2Format(GetStrArg())) == ALIEN)
HError(-2189,"HInit: Warning ALIEN Label file format set");
break;
case 'H':
if (NextArg() != STRINGARG)
HError(2119,"HInit: HMM macro file name expected");
AddMMF(&hset,GetStrArg());
break;
case 'I':
if (NextArg() != STRINGARG)
HError(2119,"HInit: MLF file name expected");
LoadMasterFile(GetStrArg());
break;
case 'L':
if (NextArg()!=STRINGARG)
HError(2119,"HInit: Label file directory expected");
labDir = GetStrArg(); break;
case 'M':
if (NextArg()!=STRINGARG)
HError(2119,"HInit: Output macro file directory expected");
outDir = GetStrArg();
break;
case 'T':
if (NextArg() != INTARG)
HError(2119,"HInit: Trace value expected");
trace = GetChkedInt(0,01777,s);
break;
case 'X':
if (NextArg()!=STRINGARG)
HError(2119,"HInit: Label file extension expected");
labExt = GetStrArg(); break;
default:
HError(2119,"HInit: Unknown switch %s",s);
}
}
if (NextArg()!=STRINGARG)
HError(2119,"HInit: source HMM file name expected");
hmmfn = GetStrArg();
Initialise();
do {
if (NextArg()!=STRINGARG)
HError(2119,"HInit: training data file name expected");
datafn = GetStrArg();
LoadFile(datafn);
} while (NumArgs()>0);
nSeg = NumSegs(segStore);
if (nSeg < minSeg)
HError(2121,"HInit: Too Few Observation Sequences [%d]",nSeg);
if (trace&T_TOP) {
printf("%d Observation Sequences Loaded\n",nSeg);
fflush(stdout);
}
EstimateModel();
SaveModel(outfn);
if (trace&T_TOP)
printf("Output written to directory %s\n",
outDir==NULL?"current":outDir);
Exit(0);
return (0); /* never reached -- make compiler happy */
}
int mainHInit(int argc, char *argv[])
{
static int ft=1;
char *datafn, *s;
int nSeg;
zwangModify();
if(InitShell(argc,argv,hinit_version,hinit_vc_id)<SUCCESS)
HError(2100,"HInit: InitShell failed");
if(ft)
{
if(isMemInit==0)
{
InitMem();
isMemInit=1;
}
InitLabel();
InitMath(); InitSigP();
InitWave(); InitAudio();
InitVQ(); InitModel();
if(InitParm()<SUCCESS)
HError(2100,"HInit: InitParm failed");
InitTrain(); InitUtil();
ft=0;
/* Stacks for global structures requiring memory allocation */
CreateHeap(&segmentStack,"SegStore", MSTAK, 1, 0.0, 100000, LONG_MAX);
CreateHeap(&sequenceStack,"SeqStore", MSTAK, 1, 0.0, 1000, 1000);
CreateHeap(&clustSetStack,"ClustSetStore", MSTAK, 1, 0.0, 1000, 1000);
CreateHeap(&transStack,"TransStore", MSTAK, 1, 0.0, 1000, 1000);
CreateHeap(&traceBackStack,"TraceBackStore", MSTAK, 1, 0.0, 1000, 1000);
CreateHeap(&bufferStack,"BufferStore", MSTAK, 1, 0.0, 1000, 1000);
}
zwangHMemGetConf();
zwangHWaveGetConf();
zwangHLabelGetConf();
zwangHMathGetConf();
zwangHSigPGetConf();
zwangHAudioGetConf();
zwangHVQGetConf();
zwangHModelGetConf();
zwangHParmGetConf();
zwangHTrainGetConf();
zwangHUtilGetConf();
SetConfParms();
CreateHMMSet(&hset,&gstack,FALSE);
while (NextArg() == SWITCHARG) {
s = GetSwtArg();
if (strlen(s)!=1)
HError(2119,"HInit: Bad switch %s; must be single letter",s);
switch(s[0]){
case 'e':
epsilon = GetChkedFlt(0.0,1.0,s); break;
case 'i':
maxIter = GetChkedInt(0,100,s); break;
case 'l':
if (NextArg() != STRINGARG)
HError(2119,"HInit: Segment label expected");
segLab = GetStrArg();
break;
case 'm':
minSeg = GetChkedInt(1,1000,s); break;
case 'n':
newModel = FALSE; break;
case 'o':
outfn = GetStrArg();
break;
case 'u':
SetuFlags(); break;
case 'v':
minVar = GetChkedFlt(0.0,10.0,s); break;
case 'w':
mixWeightFloor = MINMIX * GetChkedFlt(0.0,10000.0,s);
break;
case 'B':
saveBinary = TRUE;
break;
case 'F':
if (NextArg() != STRINGARG)
HError(2119,"HInit: Data File format expected");
if((dff = Str2Format(GetStrArg())) == ALIEN)
HError(-2189,"HInit: Warning ALIEN Data file format set");
break;
case 'G':
if (NextArg() != STRINGARG)
HError(2119,"HInit: Label File format expected");
if((lff = Str2Format(GetStrArg())) == ALIEN)
HError(-2189,"HInit: Warning ALIEN Label file format set");
break;
case 'H':
if (NextArg() != STRINGARG)
HError(2119,"HInit: HMM macro file name expected");
AddMMF(&hset,GetStrArg());
break;
case 'I':
if (NextArg() != STRINGARG)
HError(2119,"HInit: MLF file name expected");
LoadMasterFile(GetStrArg());
break;
case 'L':
if (NextArg()!=STRINGARG)
HError(2119,"HInit: Label file directory expected");
labDir = GetStrArg(); break;
case 'M':
if (NextArg()!=STRINGARG)
HError(2119,"HInit: Output macro file directory expected");
outDir = GetStrArg();
break;
case 'T':
if (NextArg() != INTARG)
HError(2119,"HInit: Trace value expected");
trace = GetChkedInt(0,01777,s);
break;
case 'X':
if (NextArg()!=STRINGARG)
HError(2119,"HInit: Label file extension expected");
labExt = GetStrArg(); break;
default:
HError(2119,"HInit: Unknown switch %s",s);
}
}
if (NextArg()!=STRINGARG)
HError(2119,"HInit: source HMM file name expected");
hmmfn = GetStrArg();
Initialise();
do {
if (NextArg()!=STRINGARG)
HError(2119,"HInit: training data file name expected");
datafn = GetStrArg();
LoadFile(datafn);
} while (NumArgs()>0);
nSeg = NumSegs(segStore);
if (nSeg < minSeg)
HError(2121,"HInit: Too Few Observation Sequences [%d]",nSeg);
EstimateModel();
SaveModel(outfn);
Dispose(hset.hmem,hset.mtab);
ResetHeap(&gstack);
ResetHeap(&segmentStack);
ResetHeap(&sequenceStack);
ResetHeap(&clustSetStack);
ResetHeap(&transStack);
ResetHeap(&traceBackStack);
ResetHeap(&bufferStack);
zwangInitParmClear();
zwangInitShellClear();
return (0); /* never reached -- make compiler happy */
}