void
Feature::
createLam(int which, ECString tmp, ECString path)
{
ECString npath("/ltmp/discrim/FOLDS/");
int b,f;
int tot = Feature::total[which];
for(int i = 0 ; i < 20 ; i++)
{
ECString ftstr(npath);
ftstr += intToString(i) + "/";
ftstr += tmp;
ftstr += ".lambdas";
ifstream fts(ftstr.c_str());
assert(fts);
for(b = 1; b < 15 ; b++)
{
int bb;
if(!fts)
{
cerr << "Trouble reading lambs for " << which << " in " << ftstr
<< endl;
assert(fts);
}
fts >> bb ;
//cerr << bb << endl;
if(bb != b)
{
cerr << tmp << " " << b << " " << bb << endl;
assert(bb == b);
}
for(f = 2; f <= tot ; f++)
{
float lam;
assert(fts);
fts >> lam;
//cerr << which << " " << f << " " << b << " " << lam << endl;
if(i == 0) Feature::setLambda(which,f,b,lam);
else lamVal(which,f,b) += lam;
}
}
}
for(b = 1; b < 15 ; b++)
for(f = 2; f <= tot ; f++) lamVal(which,f,b) /= 20.0;
ECString ftstr(path);
ftstr += tmp;
ftstr += ".lambdas";
ofstream res(ftstr.c_str());
assert(res);
res.precision(3);
printLambdas(res);
}
void
Feature::
readLam(int which, ECString tmp, ECString path)
{
ECString ftstr(path);
ftstr += tmp;
ftstr += ".lambdas";
ifstream fts(ftstr.c_str());
assert(fts);
int b,f;
// wul: lambdas文件有14行，每行的格式如下(第一列为序号）：
// wul: 2 0 0.329 0.336 0.00178 0.353 0.00129 0.158 0.417
for(b = 1; b < 15 ; b++)
{
int bb;
assert(fts);
fts >> bb ;
//cerr << bb << endl;
assert(bb == b);
// wul: 读取后面的各列（不同的Which，如ru, u...)有不同的列。根据它们的Feature数目
// 而定
for(f = 2; f <= Feature::total[which] ; f++)
{
float lam;
assert(fts);
fts >> lam;
//cerr << which << " " << f << " " << b << " " << lam << endl;
// wul: Lambda Table有点数据库。在此记录下。b表示一个实例（行号）
Feature::setLambda(which,f,b,lam);
}
}
}
void
MeChart::
init(ECString path)
{
Feat::Usage = PARSE;
addEdgeSubFeatureFns();
addSubFeatureFns();
ECString tmpA[MAXNUMCALCS] = {"r","h","u","m","l","lm","ru","rm","tt",
"s","t","ww","dummy","dummy","dummy"};
for(int which = 0 ; which < Feature::numCalcs ; which++)
{
ECString tmp = tmpA[which];
Feature::init(path, tmp);
if(tmp == "s" || tmp == "t") continue;
ECString ftstr(path);
ftstr += tmp;
ftstr += ".g";
ifstream fts(ftstr.c_str());
if(!fts) cerr << "could not find " << ftstr << endl;
assert(fts);
FeatureTree* ft = new FeatureTree(fts); //puts it in root;
if(tmp == "ww") continue;
Feature::readLam(which, tmp, path);
}
int cntxSzReq = Feature::total[TCALC];
int scSz = Feature::total[SCALC];
if(scSz > cntxSzReq) cntxSzReq = scSz;
//assert(CntxArray::sz == cntxSzReq);
//assert(CntxArray::sz == (Feature::total[UCALC] -1));
}
void
Feature::
readLam(int which, ECString tmp, ECString path)
{
ECString ftstr(path);
ftstr += tmp;
ftstr += ".lambdas";
ifstream fts(ftstr.c_str());
assert(fts);
int b,f;
int tot = Feature::total[which];
/* The standard training programs never read in lambdas. Only
getProbs does, and it uses the new bucketing, which uses the
next for loop
*/
for(f = 2 ; f <= tot ; f++)
{
float logBase;
fts >> logBase;
/*JT: this used to be logFacs[f][which] = 1.0/log(logBase), but that
occurs in bucket
*/
logFacs[which][f] = logBase;
}
for(b = 1; b < 15 ; b++)
{
int bb;
if(!fts)
{
cerr << "Trouble reading lambs for " << which << " in " << ftstr
<< endl;
assert(fts);
}
fts >> bb ;
//cerr << bb << endl;
if(bb != b)
{
cerr << tmp << " " << b << " " << bb << endl;
assert(bb == b);
}
for(f = 2; f <= tot ; f++)
{
float lam;
assert(fts);
fts >> lam;
//cerr << which << " " << f << " " << b << " " << lam << endl;
Feature::setLambda(which,f,b,lam);
}
}
}
void
Feature::
readLam(int which, ECString tmp, ECString path)
{
ECString ftstr(path);
ftstr += tmp;
ftstr += ".lambdas";
ifstream fts(ftstr.c_str());
assert(fts);
int b,f;
int tot = Feature::total[which];
if(Feature::isLM or Feature::useExtraConditioning)
{
/* This for loop is removed for old bucketing; */
for(f = 2 ; f <= tot ; f++)
{
float logBase;
fts >> logBase;
logFacs[which][f] = 1.0/log(logBase);
}
}
for(b = 1; b < 15 ; b++)
{
int bb;
if(!fts)
{
cerr << "Trouble reading lambs for " << which << " in " << ftstr
<< endl;
assert(fts);
}
fts >> bb ;
//cerr << bb << endl;
if(bb != b)
{
cerr << tmp << " " << b << " " << bb << endl;
assert(bb == b);
}
for(f = 2; f <= tot ; f++)
{
float lam;
assert(fts);
fts >> lam;
//cerr << which << " " << f << " " << b << " " << lam << endl;
Feature::setLambda(which,f,b,lam);
}
}
}
int
main(int argc, char *argv[])
{
struct rlimit core_limits;
core_limits.rlim_cur = 0;
core_limits.rlim_max = 0;
setrlimit( RLIMIT_CORE, &core_limits );
ECArgs args( argc, argv );
assert(args.nargs() == 2);
conditionedType = args.arg(0);
cerr << "start trainRs: " << conditionedType << endl;
ECString path( args.arg( 1 ) );
if(args.isset('L')) Feature::setLM();
Term::init(path);
readHeadInfo(path);
Pst pst(path);
if(Feature::isLM) ClassRule::readCRules(path);
addSubFeatureFns();
Feature::init(path, conditionedType);
whichInt = Feature::whichInt;
int ceFunInt = Feature::conditionedFeatureInt[Feature::whichInt];
Feature::conditionedEvent
= SubFeature::Funs[ceFunInt];
Feat::Usage = PARSE;
ECString ftstr(path);
ftstr += conditionedType;
ftstr += ".g";
ifstream fts(ftstr.c_str());
if(!fts)
{
cerr << "Could not find " << ftstr << endl;
assert(fts);
}
tRoot = new FeatureTree(fts); //puts it in root;
cout.precision(3);
cerr.precision(3);
lamInit();
InputTree* trainingData[1001];
int usedCount = 0;
sentenceCount = 0;
for( ; ; sentenceCount++)
{
if(sentenceCount%10000 == 1)
{
// cerr << conditionedType << ".tr "
//<< sentenceCount << endl;
}
if(usedCount >= 1000) break;
InputTree* correct = new InputTree;
cin >> (*correct);
if(correct->length() == 0) break;
if(!cin) break;
EcSPairs wtList;
correct->make(wtList);
InputTree* par;
par = correct;
trainingData[usedCount++] = par;
}
if(Feature::isLM) pickLogBases(trainingData,sentenceCount);
procGSwitch = true;
for(pass = 0 ; pass < 10 ; pass++)
{
if(pass%2 == 1) cout << "Pass " << pass << endl;
goThroughSents(trainingData, sentenceCount);
updateLambdas();
//printLambdas(cout);
zeroData();
}
ECString resS(path);
resS += conditionedType;
resS += ".lambdas";
ofstream res(resS.c_str());
res.precision(3);
printLambdas(res);
printLambdas(cout);
cout << "Total params = " << FeatureTree::totParams << endl;
cout << "Done: " << (int)sbrk(0) << endl;
}
