/* 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);
}
}
/* LoadFile: load whole file or segments into segStore */
void LoadFile(char *fn)
{
BufferInfo info;
char labfn[80];
Transcription *trans;
long segStIdx,segEnIdx;
static int segIdx=1; /* Between call handle on latest seg in segStore */
static int prevSegIdx=1;
HTime tStart, tEnd;
int i,k,s,ncas,nObs=0,segLen;
LLink p;
Observation obs;
if((pbuf=OpenBuffer(&bufferStack, fn, 10, dff, FALSE_dup, FALSE_dup))==NULL)
HError(2150,"LoadFile: Config parameters invalid");
GetBufferInfo(pbuf,&info);
CheckData(fn,info);
if (firstTime) InitSegStore(&info);
if (segId == NULL) { /* load whole parameter file */
nObs = ObsInBuffer(pbuf);
tStart = 0.0;
tEnd = (info.tgtSampRate * nObs);
LoadSegment(segStore, tStart, tEnd, pbuf);
segIdx++;
}
else { /* load segment of parameter file */
MakeFN(fn,labDir,labExt,labfn);
trans = LOpen(&transStack,labfn,lff);
ncas = NumCases(trans->head,segId);
if ( ncas > 0) {
for (i=1,nObs=0; i<=ncas; i++) {
p = GetCase(trans->head,segId,i);
segStIdx = (long)(p->start/info.tgtSampRate);
segEnIdx = (long)(p->end/info.tgtSampRate);
if (segEnIdx >= ObsInBuffer(pbuf))
segEnIdx = ObsInBuffer(pbuf)-1;
if (segEnIdx - segStIdx + 1 >= nStates-2) {
LoadSegment(segStore, p->start, p->end, pbuf);
if (trace&T_LD1)
printf(" loading seg %s %f[%ld]->%f[%ld]\n",segId->name,
p->start,segStIdx,p->end,segEnIdx);
nObs += SegLength(segStore, segIdx);
segIdx++;
}else if (trace&T_LD1)
printf(" seg %s %f->%f ignored\n",segId->name,
p->start,p->end);
}
}
}
if (hset.hsKind == DISCRETEHS){
for (k=prevSegIdx; k<segIdx; k++){
segLen = SegLength(segStore, k);
for (i=1; i<=segLen; i++){
obs = GetSegObs(segStore, k, i);
for (s=1; s<=nStreams; s++){
if( (obs.vq[s] < 1) || (obs.vq[s] > maxMixInS[s]))
HError(2150,"LoadFile: Discrete data value [ %d ] out of range in stream [ %d ] in file %s",obs.vq[s],s,fn);
}
}
}
prevSegIdx=segIdx;
}
if (trace&T_LD0)
printf(" %d observations loaded from %s\n",nObs,fn);
CloseBuffer(pbuf);
ResetHeap(&transStack);
}
