int main(int argc, char* argv[])
{
int i, j, c, iter, ITER=200;
unsigned long int seed=0;
int bcycle = 0;
float bstart = 0;
int acycle = 0;
float astart = 0;
int burnin = 0;
stable_t *ST = NULL;
int useN = DIM*2;
MAXN = 1;
MAXT = MAXSTAB;
/*
* default values for args
*/
while ( (c=getopt(argc, argv,"a:b:B:C:I:hH:I:N:P:S:s:T:v"))>=0 ) {
switch ( c ) {
case 'h':
usage(burnin?burnin:ITER/2, ITER, useN);
exit(0);
case 'b':
if ( !optarg || sscanf(optarg,"%f,%f",&bpar, &bstart)<1 )
yaps_quit("Need a valid 'b' argument\n");
break;
case 'a':
if ( !optarg || sscanf(optarg,"%f,%f",&apar,&astart)<1 )
yaps_quit("Need a valid 'a' argument\n");
break;
case 'H':
if ( !optarg || sscanf(optarg,"%d",&bcycle)!=1 )
yaps_quit("Need a valid 'G' argument\n");
break;
case 'T':
if ( !optarg || sscanf(optarg,"%d",&MAXT)!=1 )
yaps_quit("Need a valid 'T' argument\n");
break;
case 'I':
if ( !optarg || sscanf(optarg,"%d",&acycle)!=1 )
yaps_quit("Need a valid 'H' argument\n");
break;
case 'N':
if ( !optarg || sscanf(optarg,"%d",&useN)!=1 )
yaps_quit("Need a valid 'N' argument\n");
break;
case 'C':
if ( !optarg || sscanf(optarg,"%d",&ITER)!=1 )
yaps_quit("Need a valid 'C' argument\n");
break;
case 'B':
if ( !optarg || sscanf(optarg,"%d",&burnin)!=1 )
yaps_quit("Need a valid 'B' argument\n");
break;
case 's':
if ( !optarg || sscanf(optarg,"%lu",&seed)!=1 )
yaps_quit("Need a valid 's' argument\n");
break;
case 'v':
verbose++;
break;
#ifdef S_USE_THREADS
case 'P':
if ( !optarg || sscanf(optarg,"%u",&threads)!=1 )
yaps_quit("Need a valid 'P' argument\n");
break;
#endif
default:
yaps_message("Bad command line argument\n\n");
usage(burnin?burnin:ITER/2, ITER, useN);
exit(0);
}
}
if ( useN>=MAXDATA )
yaps_quit("N too large\n");
if ( burnin==0 )
burnin = ITER/2;
else
if ( burnin>=ITER-1 )
yaps_quit("Burnin %d too large for cycles %d\n", burnin, ITER);
yaps_message("Configuration details\n");
yaps_message("=====================\n");
/*
* set random number generator
*/
if ( seed ) {
rng_seed(rng,seed);
} else {
rng_time(rng,&seed);
}
yaps_message("Setting seed for data = %lu\n", seed);
if ( acycle && apar==0 )
apar = 0.5;
yaps_message("Setting a=%f, b=%f, N=%d, D=%d\n",
apar, bpar, useN, NUMMN);
yaps_message(" burnin=%d,", burnin);
yaps_message(" cycles=%d\n", ITER);
/*
* fix pointers
*/
for (j=0; j<NUMMN; j++) {
n[j] = &n_data[j*DIM];
t[j] = &t_data[j*DIM];
tave[j] = &tave_data[j*DIM];
}
/*
* initialise everything
*/
for (j=0; j<NUMMN; j++) {
N[j] = useN;
T[j] = 0;
Tave[j] = 0;
for (i=0; i<DIM; i++) {
n[j][i] = 0;
t[j][i] = 0;
tave[j][i] = 0;
}
}
/*
* fix base distribution, uniform
*/
{
for (i=0; i<DIM; i++) {
H[i] = 1.0/DIM;
}
}
/*
* create data using a CRP to get initialisation for n[]
*/
c = 0;
for (j=0; j<NUMMN; j++) {
int cc;
i = sampleH();
data[c++] = i;
// first entry always adds a table
n[j][i]++;
t[j][i]++;
T[j]++;
for (cc=1; cc<N[j]; cc++) {
float val = (cc+bpar)*rng_unit(rng);
val -= T[j]*apar+bpar;
if ( val<=0 ) {
// new table
i = sampleH();
t[j][i]++;
T[j]++;
} else {
for (i=0; i<DIM; i++) {
val -= n[j][i] - t[j][i]*apar;
if ( val<0 )
break;
}
}
assert(i<DIM);
n[j][i]++;
data[c++] = i;
}
}
binit = bpar;
/*
* record maximum entries in data
* do this where possible so that one can get the table
* sizes right
*
*/
MAXN = n[0][0]+1;
MAXT = 1;
for (j=0; j<NUMMN; j++) {
for (i=0; i<DIM; i++) {
if ( MAXN<=n[j][i] ) MAXN = n[j][i]+1;
if ( MAXT<t[j][i] ) MAXT = t[j][i]*1.1+1;
}
}
if ( MAXT>MAXN )
MAXT = MAXN;
yaps_message("Making S for N=%d M=%d a=%lf\n", MAXN,MAXT,apar);
ST = S_make(MAXN, MAXT, MAXN, MAXTAB,
apar, S_STABLE | S_UVTABLE);
if ( ST==NULL )
yaps_quit("Making S failed!\n");
S_report(ST,stdout);
/*
* the seed only sets the data/sample,
* the seed for the simulation/Gibbs is always random
*/
rng_free(rng);
rng_time(rng,&seed);
//yaps_message("Resetting seed = %lu\n", seed);
/*
* report on initial data statistics
*/
yaps_message("\nData sampled\n");
yaps_message("============\n");
for (j=0; j<NUMMN; j++) {
yaps_message("n[%d] =", j);
for (i=0; i<DIM; i++)
yaps_message(" %d", n[j][i]);
yaps_message(" = %d\n", N[j]);
yaps_message("t[%d] =",j);
for (i=0; i<DIM; i++)
yaps_message(" %d", t[j][i]);
yaps_message(" = %d\n", T[j]);
}
/*
* set the hyperparameters used in Gibbs,
* can be different to data
*/
if ( bstart==0 )
bstart = bpar;
if ( astart==0 )
astart = apar;
// initialise latent stats and reporting info
for (j=0; j<NUMMN; j++) {
T[j] = 0;
Tave[j] = 0;
}
tcnt = 0;
bave = 0;
bcnt = 0;
aave = 0;
acnt = 0;
bpar = bstart;
if ( verbose && bcycle!=0 )
yaps_message("Starting with initial b=%f\n", bpar);
apar = astart;
if ( verbose && acycle!=0 )
yaps_message("Starting with initial a=%f\n", apar);
for (j=0; j<NUMMN; j++) {
for (i=0; i<DIM; i++) {
tave[j][i] = 0;
t[j][i] = 0;
if ( n[j][i]>0 ) {
/*
* initialise to a single table
*/
t[j][i] = 1;
T[j]++;
}
}
}
for ( iter=0; iter<ITER; iter++) {
/*
* sampling with table indicators
*/
c = 0;
for (j=0; j<NUMMN; j++) {
int cc;
for (cc=0; cc<N[j]; cc++) {
float one;
i = data[c++];
assert(n[j][i]);
if ( n[j][i]==1 )
// this indicator must always be 1, no sampling
continue;
// sample whether it contributes to a table
if ( t[j][i]>1 &&
(n[j][i]-1)*rng_unit(rng)<(t[j][i]-1) ) {
t[j][i]--;
T[j]--;
}
assert(t[j][i]<n[j][i]);
// sample new table indicator
one = H[i] * (bpar + T[j]*apar) * (t[j][i]) / (n[j][i]-t[j][i]+1)
* S_V(ST, n[j][i],t[j][i]+1);
if ( rng_unit(rng) < one/(one+1.0) ) {
t[j][i]++;
T[j]++;
}
}
}
/*
* one major cycle of Gibbs sampler finished
*/
if ( verbose>1 ) {
for (j=0; j<NUMMN; j++) {
for (i=0; i<DIM; i++)
printf(" %d", t[j][i]);
printf(" = %d\n", T[j]);
}
}
/*
* sample & record b
*/
if ( bcycle!=0 && iter%bcycle==0 ) {
// Gibbs on bpar (concentration par) too
if ( bcycle<0 ) {
int bc = -bcycle;
for (bc-- ; bc>0; bc--)
bpar = sampleb(bpar, 1, PB_shape, PB_scale, N, T,
apar, rng, 1, 1);
}
bpar = sampleb(bpar, 1, PB_shape, PB_scale, N, T,
apar, rng, 1, 1);
if ( iter>=burnin ) {
bave += bpar;
bcnt ++;
}
}
/*
* sample & record a
*/
if ( acycle!=0 && iter%acycle==0 ) {
int dimI[NUMMN];
double dimb[NUMMN];
for (j=0; j<NUMMN; j++) {
dimI[j] = DIM;
dimb[j] = bpar;
}
// Gibbs on apar (discount par) too
if ( acycle<0 ) {
int bc = -acycle;
for (bc-- ; bc>0; bc--)
apar = samplea(apar, NUMMN, dimI, T, n, t, NULL, dimb, rng, 1, 1);
}
apar = samplea(apar, NUMMN, dimI, T, n, t, NULL, dimb, rng, 1, 1);
if ( iter>=burnin ) {
aave += apar;
acnt ++;
}
if ( verbose>1 )
yaps_message("Extending S for a=%lf\n", apar);
if ( S_remake(ST,apar) )
yaps_message("Extending S failed\n");
}
/*
* full statistics collection
*/
if ( iter>=burnin ) {
for (j=0; j<NUMMN; j++) {
for (i=0; i<DIM; i++) {
tave[j][i] += t[j][i];
}
Tave[j] += T[j];
}
tcnt ++;
}
}
/*
* report for this experiment
*/
yaps_message("\nEstimates\n");
yaps_message("=========\n");
for (j=0; j<NUMMN; j++) {
yaps_message("t[%d] = ", j);
for (i=0; i<DIM; i++)
yaps_message(" %.2f", tave[j][i]/tcnt);
yaps_message("\nT[%d]=%.2f\n", j, Tave[j]/tcnt);
}
if ( bcycle!=0 && bcnt>0 )
yaps_message("\nb=%.2f", bave/bcnt);
if ( acycle!=0 && acnt>0 )
yaps_message("\na=%.3f", aave/acnt);
yaps_message("\n");
S_free(ST);
rng_free(rng);
return 0;
}
/*==========================================
* main
*========================================== */
int main(int argc, char* argv[])
{
int c, iter, ITER=0, seed=0;
enum dataType data = LdaC;
enum dataType testdata = LdaC;
int dots = 0;
enum GibbsType fix_hold = GibbsNone;
char *stem;
char *resstem;
int topwords = 20;
int noerrorlog = 0;
int displayed = 0;
int load_vocab = 0;
int checkpoint = 0;
int restart = 0;
int dopmi = 0;
int restart_hca = 0;
int load_phi = 0;
int load_mu = 0;
int procs = 1;
int maxW = 0;
enum ScoreType score=ST_idf;
double BM0val=0, BM1val =0, BP0val=0, BP1val=0;
clock_t t1=0, t2=0, t3=0;
double tot_time = 0;
double psample_time = 0;
enum ParType par;
/*
* default values
*/
ddN.T = 10;
ITER = 100;
ddN.TEST = 0;
pctl_init();
while ( (c=getopt(argc, argv,"b:c:C:d:ef:F:g:G:h:K:l:L:N:o:pq:vr:s:S:t:T:vVW:"))>=0 ) {
switch ( c ) {
case 'b':
if ( !optarg || sscanf(optarg,"%d",&ddP.back)!=1 )
yap_quit("Need a valid 'b' argument\n");
break;
case 'c':
if ( !optarg || sscanf(optarg,"%d",&checkpoint)!=1 )
yap_quit("Need a valid 'c' argument\n");
break;
case 'C':
if ( !optarg || sscanf(optarg,"%d",&ITER)!=1 )
yap_quit("Need a valid 'C' argument\n");
break;
case 'd':
if ( !optarg || sscanf(optarg,"%d",&dots)!=1 )
yap_quit("Need a valid 'd' argument\n");
break;
case 'e':
noerrorlog++;
break;
case 'f':
if ( strcmp(optarg,"witdit")==0 )
data = WitDit;
else if ( strcmp(optarg,"docword")==0 )
data = Docword;
else if ( strcmp(optarg,"ldac")==0 )
data = LdaC;
else if ( strcmp(optarg,"bag")==0 )
data = TxtBag;
else if ( strcmp(optarg,"lst")==0 )
data = SeqTxtBag;
else
yap_quit("Illegal data type for -f\n");
break;
case 'F':
if ( strcmp(optarg,"all")==0 ) {
for (par=ParAM; par<=ParBB; par++)
ddT[par].fix = 1;
} else {
par = findpar(optarg);
if ( par==ParNone )
yap_quit("Illegal arg for -F\n");
ddT[par].fix = 1;
}
break;
case 'g':
{
char var[100];
int st=0;
if ( !optarg || sscanf(optarg,"%[^, ],%d", &var[0], &st)<1 )
yap_quit("Need a valid 'g' argument\n");
par = findpar(var);
if ( par==ParBP1 )
ddP.kbatch = st;
else
yap_quit("Illegal var for -g\n");
}
break;
case 'G':
{
char var[100];
int st=0, cy=0;
if ( !optarg || sscanf(optarg,"%[^, ],%d,%d",
&var[0], &cy, &st)<2 || st<0 || cy<0 )
yap_quit("Need a valid 'G' argument\n");
par = findpar(var);
if ( par==ParNone || par==ParB0P || par==ParB0M )
yap_quit("Illegal var for -G\n");
ddT[par].fix = 0;
ddT[par].start = st;
ddT[par].cycles = cy;
}
break;
case 'h':
{
fix_hold = GibbsHold;
if ( !optarg )
yap_quit("Need a valid 'h' argument\n");
if ( strncmp(optarg,"dict,",5)==0 ) {
if ( sscanf(&optarg[5],"%d",&ddP.hold_dict)<1 || ddP.hold_dict<2 )
yap_quit("Need a valid 'hdict' argument\n");
} else if ( strncmp(optarg,"fract,",6)==0 ) {
if ( sscanf(&optarg[6],"%lf",&ddP.hold_fraction)<1
|| ddP.hold_fraction<=0 || ddP.hold_fraction>=1 )
yap_quit("Need a valid 'hfract' argument\n");
} else if ( strncmp(optarg,"doc,",4)==0 ) {
if ( sscanf(&optarg[4],"%d",&ddP.hold_every)<1 || ddP.hold_every<2 )
yap_quit("Need a valid 'hdoc' argument\n");
} else
yap_quit("Need a valid 'h' argument\n");
}
break;
case 'K':
if ( !optarg || sscanf(optarg,"%d",&ddN.T)!=1 )
yap_quit("Need a valid 'K' argument\n");
break;
case 'l':
if ( !optarg )
yap_quit("Need a valid 'l ' argument\n");
if ( strncmp(optarg,"phi,",4)==0 ) {
if ( sscanf(&optarg[4],"%d,%d",&ddP.phiiter, &ddP.phiburn)<2 )
yap_quit("Need a valid 'l word,' argument\n");
} else if ( strncmp(optarg,"theta,",6)==0 ) {
if ( sscanf(&optarg[6],"%d,%d",&ddP.thetaiter, &ddP.thetaburn)<2 )
yap_quit("Need a valid 'l word,' argument\n");
} else if ( strncmp(optarg,"mu,",3)==0 ) {
if ( sscanf(&optarg[3],"%d,%d",&ddP.muiter, &ddP.muburn)<2 )
yap_quit("Need a valid 'l word,' argument\n");
} else if ( strncmp(optarg,"prog,",5)==0 ) {
if ( sscanf(&optarg[5],"%d,%d",&ddP.progiter, &ddP.progburn)<2 )
yap_quit("Need a valid 'l prog,' argument\n");
} else
yap_quit("Need a valid DIAG code in 'l' argument\n");
break;
case 'L':
if ( !optarg )
yap_quit("Need a valid 'L ' argument\n");
if ( strncmp(optarg,"like,",5)==0 ) {
if ( sscanf(&optarg[5],"%d,%d",&ddP.mltiter, &ddP.mltburn)<1 )
yap_quit("Need a valid 'L like' argument\n");
} else
yap_quit("Need a valid DIAG code in 'L' argument\n");
break;
case 'N':
if ( !optarg || sscanf(optarg,"%d,%d", &ddP.maxN, &ddP.maxM)<1 )
yap_quit("Need a valid 'N' argument\n");
break;
case 'o':
{
char *ptr = strchr(optarg, ',');
int len = strlen(optarg);
if ( ptr )
len = ptr - optarg;
if ( strncmp(optarg,"idf",len)==0 )
score = ST_idf;
else if ( strncmp(optarg,"count",len)==0 )
score = ST_count;
else if ( strncmp(optarg,"Q",len)==0 )
score = ST_Q;
else if ( strncmp(optarg,"cost",len)==0 )
score = ST_cost;
else
yap_quit("Need a valid parameter for 'o' argument\n");
if ( ptr ) {
/* there was a second arg */
if ( sscanf(ptr+1, "%d", &topwords) != 1)
yap_quit("Need a valid second 'o' argument\n");
}
break;
}
break;
case 'p':
dopmi++;
break;
case 'q':
if(!optarg || sscanf(optarg, "%d", &procs) != 1)
yap_quit("Need a valid 'q' argument\n");
break;
case 'r':
if(!optarg )
yap_quit("Need a valid 'r' argument\n");
if ( strcmp(optarg,"tca")==0 )
restart++;
else if ( strcmp(optarg,"hca")==0 )
restart_hca++;
else if ( strcmp(optarg,"phi")==0 )
load_phi++;
else if ( strcmp(optarg,"mu")==0 )
load_mu++;
else
yap_quit("Need a valid 'r' argument\n");
break;
case 's':
if ( !optarg || sscanf(optarg,"%d",&seed)!=1 )
yap_quit("Need a valid 's' argument\n");
break;
case 'S':
{
char var[100];
double vin=0;
if ( !optarg || sscanf(optarg,"%[^=, ]=%lf",
&var[0], &vin)<2 )
yap_quit("Need a valid 'S' argument\n");
par = findpar(var);
if ( par==ParNone )
yap_quit("Illegal var for -S\n");
else if ( par==ParBM0 )
BM0val = vin;
else if ( par==ParBM1 )
BM1val = vin;
else if ( par==ParBP0 )
BP0val = vin;
else if ( par==ParBP1 )
BP1val = vin;
else
*(ddT[par].ptr) = vin;
}
break;
case 't':
if ( !optarg || sscanf(optarg,"%d",&ddP.training)!=1 )
yap_quit("Need a valid 't' argument\n");
break;
case 'T':
if ( !optarg )
yap_quit("Need a valid 'T' argument\n");
{
char *tname = data_name(optarg,data);
FILE *fp = fopen(tname,"r");
if ( fp==NULL ) {
free(tname);
tname = data_name(optarg,testdata);
fp = fopen(tname,"r");
} else {
testdata = data;
}
free(tname);
if ( fp!=NULL ) {
/* its a valid test filename */
ddP.teststem = optarg;
fclose(fp);
} else if ( sscanf(optarg,"%d",&ddN.TEST)!=1 )
yap_quit("Need a valid 'T' argument\n");
}
break;
case 'v':
verbose++;
break;
case 'V':
load_vocab = 1;
break;
case 'W':
if ( !optarg || sscanf(optarg,"%d",&maxW)<1 )
yap_quit("Need a valid 'W' argument\n");
break;
default:
yap_quit("Unknown option '%c'\n", c);
}
}
if (argc-optind != 2) {
usage();
exit(-1);
}
if ( optind>=argc ) {
yap_quit("No arguments given\n");
}
stem = strdup(argv[optind++]);
resstem = strdup(argv[optind++]);
if ( dopmi )
load_vocab = 1;
if ( dopmi && verbose !=2 ) {
/*
* due to the use of the ".top" file
* its really multi-purpose
*/
yap_quit("When computing PMI verbose must be exactly 2\n");
}
if ( noerrorlog==0 ) {
char *wname = yap_makename(resstem, ".log");
yap_file(wname);
free(wname);
}
yap_commandline(argc, argv);
#ifdef H_THREADS
yap_message(" Threads,");
#endif
if ( restart || restart_hca ) {
char *fname = yap_makename(resstem,".par");
FILE *fp = fopen(fname,"r");
char *buf;
if ( !fp )
yap_quit("Parameter file '%s' doesn't exist\n", fname);
fclose(fp);
free(fname);
buf = readpar(resstem,"T",50);
if ( !buf )
yap_quit("Parameter file '%s' has no T\n", fname);
ddN.T = atoi(buf);
free(buf);
if ( restart ) {
buf = readpar(resstem,"E",50);
if ( !buf )
yap_quit("Parameter file '%s' has no E\n", fname);
ddN.E = atoi(buf);
free(buf);
pctl_read(resstem);
}
if ( maxW==0 ) {
buf = readpar(resstem,"W",50);
if ( buf ) {
maxW = atoi(buf);
free(buf);
}
}
if ( ddP.training==0 ) {
buf = readpar(resstem,"TRAIN",50);
if ( buf ) {
ddP.training = atoi(buf);
free(buf);
}
}
if ( ddN.TEST==0 ) {
buf = readpar(resstem,"TEST",50);
if ( buf ) {
ddN.TEST = atoi(buf);
free(buf);
}
}
}
assert(ddN.T>0);
assert(ddN.TEST>=0);
assert(restart || restart_hca || ITER>0);
if ( load_phi && ddP.phiiter>0 )
yap_quit("Options '-l phi,...' and '-r phi' incompatible\n");
if ( load_mu && ddP.muiter>0 )
yap_quit("Options '-l mu,...' and '-r mu' incompatible\n");
/*
* set random number generator
*/
if ( seed ) {
rng_seed(rngp,seed);
} else {
rng_time(rngp,&seed);
}
yap_message("Setting seed = %lu\n", seed);
/*
* read data and get dimensions
*/
{
D_bag_t *dbp = data_read(stem, data);
int training = pctl_training(dbp->D);
if ( ddP.teststem ) {
D_bag_t *dbpt = data_read(ddP.teststem, testdata);
/* need to load a separate test set, strip to bare training */
data_shrink(dbp, training);
ddN.TEST = dbpt->D;
data_append(dbp, dbpt);
free(dbpt->w); free(dbpt->d); free(dbpt);
}
if ( maxW>0 ) {
if ( dbp->W <= maxW )
dbp->W = maxW;
if ( dbp->W > maxW )
data_vocabshrink(dbp, maxW);
}
/*
* transfer into system
*/
ddN.D = dbp->D;
ddN.W = dbp->W;
ddN.N = dbp->N;
ddN.NT = dbp->N;
ddN.DT = training;
ddD.w = dbp->w;
ddD.d = dbp->d;
free(dbp);
if ( ddN.DT<ddN.D ) {
/* recompute NT */
int i;
for (i=0; i<ddN.N; i++)
if ( ddD.d[i]>=ddN.DT )
break;
ddN.NT = i;
}
}
data_read_epoch(stem);
/*
* at this point, dimensions are fixed, so load phi and mu if needed
*/
if ( load_phi )
pctl_loadphi(resstem);
if ( load_mu )
pctl_loadmu(resstem);
/*
* correct parameters after command line
*/
pctl_fix(ITER);
if ( BM0val>0 ) {
ddP.b_mu[0] = BM0val;
}
if ( BM1val>0 ) {
int i;
for (i=1; i<ddN.E; i++)
ddP.b_mu[i] = BM1val;
}
if ( BP0val>0 ) {
int i;
for (i=0; i<ddN.T; i++)
ddP.b_phi[0][i] = BP0val;
}
if ( BP1val>0 ) {
int i;
if ( ddN.E==1 )
yap_quit("b_phi[1] invalid when epochs==1\n");
for (i=0; i<ddN.T; i++)
ddP.b_phi[1][i] = BP1val;
}
pctl_samplereport();
/*
* all data structures
*/
data_alloc();
if ( ddP.phiiter>0 )
phi_init(resstem);
else
ddS.phi = NULL;
if ( ddP.muiter>0 )
mu_init(resstem);
else
ddS.mu = NULL;
if ( ddP.thetaiter>0 )
theta_init(resstem);
else
ddS.theta = NULL;
tca_alloc();
if ( PCTL_BURSTY() )
dmi_init(&ddM, ddS.z, ddD.w, ddD.N_dTcum,
ddN.T, ddN.N, ddN.W, ddN.D, ddN.DT,
(fix_hold==GibbsHold)?pctl_hold:NULL);
if ( load_vocab ) {
data_vocab(stem);
}
cache_init();
/*
* yap some details
*/
data_report(ITER, seed);
pctl_report();
/*
* load/init topic assignments and prepare statistics
*/
if ( restart || restart_hca) {
tca_read_z(resstem, 0, ddN.DT);
tca_rand_z(ddN.T, ddN.DT, ddN.D);
} else {
tca_rand_z(ddN.T, 0, ddN.D);
}
tca_reset_stats(resstem, restart, 0);
if ( (restart || restart_hca ) && ITER )
yap_message("Initial log_2(perp)=%lf\n", -M_LOG2E * likelihood()/ddN.NT);
if ( ITER )
yap_report("cycles: ");
for (iter=0; iter<ITER; iter++) {
int pro;
double thislp = 0;
int thisNd = 0;
int doc;
#ifdef H_THREADS
pthread_t thread[procs];
#endif
D_pargs_p parg[procs];
#ifdef MU_CACHE
mu_side_fact_reinit();
#endif
#ifdef PHI_CACHE
phi_cache_reinit();
#endif
t1 = clock();
/*
* sampling
*/
#ifdef IND_STATS
ddP.doc_ind_stats = u32tri(ddN.T,ddN.E,ddN.E);
ddP.word_ind_stats = u32tri(ddN.T,ddN.E,ddN.E);
#endif
/* a bit complex if no threads! */
doc = 0;
for (pro = 0 ; pro < procs ; pro++){
parg[pro].dots=dots;
parg[pro].procs=procs;
parg[pro].doc = &doc;
#ifndef H_THREADS
sampling_p(&parg[pro]);
#else
if ( procs==1 )
sampling_p(&parg[pro]);
else if( pthread_create(&thread[pro],NULL,sampling_p,(void*) &parg[pro]) != 0){
yap_message("thread failed %d\n",pro+1 );
}
#endif
}
#ifdef H_THREADS
if ( procs>1 ) {
//waiting for threads to finish
for (pro = 0; pro < procs; pro++){
pthread_join(thread[pro], NULL);
}
}
#endif
// getting lp, Nd and clock
for(pro = 0; pro < procs; pro++){
thislp += parg[pro].thislp;
thisNd += parg[pro].thisNd;
tot_time += parg[pro].tot_time;
}
#ifdef H_THREADS
if ( procs>1 )
tca_reset_stats(NULL,1,1);
#endif
/*
* full check
*/
#ifndef NDEBUG
{
int e, d;
check_cp_et();
for (e=0; e<ddN.E; e++)
check_m_vte(e);
for (d=0; d<ddN.DT; d++)
check_n_dt(d);
}
#endif
#ifdef IND_STATS
{
char *fname = yap_makename(resstem,".istats");
FILE *ifp = fopen(fname,"a");
int e1, e2, kk;
fprintf(ifp,"Iteration %d\n", iter);
for (kk=0; kk<ddN.T; kk++) {
fprintf(ifp," Topic %d\n", kk);
for (e1=0; e1<ddN.E; e1++) {
fprintf(ifp," Epoch %d\n ", e1);
for (e2=0; e2<ddN.E; e2++)
fprintf(ifp," %u", (unsigned)ddP.doc_ind_stats[kk][e1][e2]);
fprintf(ifp,"\n ");
for (e2=0; e2<ddN.E; e2++)
fprintf(ifp," %u", (unsigned)ddP.word_ind_stats[kk][e1][e2]);
fprintf(ifp,"\n");
}
}
fclose(ifp);
free(ddP.doc_ind_stats[0][0]); free(ddP.doc_ind_stats[0]);
free(ddP.doc_ind_stats);
free(ddP.word_ind_stats[0][0]); free(ddP.word_ind_stats[0]);
free(ddP.word_ind_stats);
free(fname);
}
#endif
/*
* sample hyperparameters
*/
t3 = clock();
pctl_sample(iter, procs);
/*
* do time calcs here to remove diagnostics+reporting
*/
t2 = clock();
tot_time += (double)(t2 - t1) / CLOCKS_PER_SEC;
psample_time += (double)(t2 - t3) / CLOCKS_PER_SEC;
/*
* progress reports
*/
if ( ( iter>ddP.progburn && (iter%ddP.progiter)==0 ) || iter+1>=ITER ) {
yap_message(" %d\nlog_2(perp)=%lf,%lf", iter,
-M_LOG2E * likelihood()/ddN.NT, -M_LOG2E * thislp/thisNd);
pctl_update(iter);
if ( verbose && iter%10==0 )
yap_probs();
if ( iter>0 && verbose>1 ) {
if ( ddN.tokens ) {
tca_displaytopics(resstem,topwords,score);
displayed++;
}
}
if ( iter+1<ITER ) {
// yap_message("\n");
yap_report("cycles: ");
}
} else {
yap_message(" %d", iter);
if ( verbose>1) yap_message("\n");
}
if ( checkpoint>0 && iter>0 && iter%checkpoint==0 ) {
data_checkpoint(resstem, stem, iter+1);
yap_message(" checkpointed\n");
tca_report(resstem, stem, ITER, procs, fix_hold,
(dopmi&&displayed>0)?1:0);
}
if ( ddP.phiiter>0 && iter>ddP.phiburn && (iter%ddP.phiiter)==0 )
phi_update();
if ( ddP.thetaiter>0 && iter>ddP.thetaburn && (iter%ddP.thetaiter)==0 )
theta_update();
if ( ddP.muiter>0 && iter>ddP.muburn && (iter%ddP.muiter)==0 )
mu_update();
} // over iter
if ( ITER )
yap_report("Finished after %d cycles on average of %lf+%lf(s) per cycle\n",
iter, (tot_time-psample_time)/iter, psample_time/iter);
if ( ( verbose==1 || ((iter+1)%5!=0 && verbose>1) ) ) {
if ( ddN.tokens ) {
tca_displaytopics(resstem,topwords,score);
displayed++;
}
}
yap_probs();
if ( ITER>0 )
data_checkpoint(resstem, stem, ITER);
tca_report(resstem, stem, ITER, procs, fix_hold, (dopmi&&displayed>0)?1:0);
if ( ddP.phiiter>0 )
phi_save(resstem);
if ( ddP.thetaiter>0 )
theta_save(resstem);
if ( ddP.muiter>0 )
mu_save(resstem);
/*
* free
*/
phi_free();
theta_free();
mu_free();
cache_free();
pctl_free();
data_free();
dmi_free(&ddM);
tca_free();
free(stem);
free(resstem);
rng_free(rngp);
return 0;
}
