/*
* This computes the function value at a given weight vector.
* This should be parallelized by dividing the training examples
* into subsets and doing each on a different thread.
*
* @param weights - the value of the weights
* @param the value of the function
*
*/
double WFST_Trainer_Local::value2(const column_vector &weights) {
cout << "LIKELIHOOD\n";
double likelihood = 0.0;
update_arc_weights(weights);
// ``filler'' FSTs to replace
// the results from the composition
// declared outside for loop for efficiency
VectorFst<LogArc> medial;
VectorFst<LogArc> final;
medial.SetInputSymbols(fst->InputSymbols());
medial.SetOutputSymbols(fst->OutputSymbols());
final.SetInputSymbols(fst->InputSymbols());
/**
* Create an FST based on an RNN
*/
void FlatBOFstBuilder::convertRNN(CRnnLM & rnnlm, VectorFst<LogArc> &fst) {
queue<NeuronFstHistory> q;
VectorFst<LogArc> new_fst;
NeuronFstHistory fsth(rnnlm.getHiddenLayerSize(),getNumBins());
FstIndex id = 0;
NeuronFstHistory new_fsth(rnnlm.getHiddenLayerSize(),getNumBins());
FstIndex new_id;
NeuronFstHistory min_backoff(rnnlm.getHiddenLayerSize(),getNumBins());
set<NeuronFstHistory>set_min_backoff;
NeuronFstHistory bo_fsth(rnnlm.getHiddenLayerSize(),getNumBins());
bool backoff = false;
vector<FstIndex> deleted;
real p = 0.00;
real p_joint = 0.00;
real entropy = 0.0;
real delta = 0.0;
vector<real> all_prob(rnnlm.getVocabSize());
vector<real> posterior(10);
map< FstIndex,set<FstIndex> > pred;
vector<bool> non_bo_pred(rnnlm.getVocabSize());
vector<int> to_be_added;
vector<int> to_be_removed;
for (int i = 0; i < rnnlm.getVocabSize(); i++) {
to_be_removed.push_back(i);
}
vector<real> to_be_added_prob;
FstIndex n_added = 0;
FstIndex n_processed = 0;
FstIndex next_n_added = 0;
FstIndex next_n_processed = 0;
FstIndex n_backoff = 0;
FstIndex n_only_backoff = 0;
int v = rnnlm.getVocabSize();
int w = 0;
// Initialize
rnnlm.copyHiddenLayerToInput();
// printNeurons(rnnlm.getInputLayer(),0,10);
// Initial state ( 0 | hidden layer after </s>)
printNeurons(rnnlm.getHiddenLayer(),0,10);
fsth.setFstHistory(rnnlm, *dzer);
fsth.setLastWord(0);
q.push(fsth);
addFstState(id, new NeuronFstHistory(fsth), fst);
fst.SetStart(INIT_STATE);
// Final state (don't care about the associated discrete representation)
fst.AddState();
fst.SetFinal(FINAL_STATE, LogWeight::One());
/*posterior.at(INIT_STATE) = MY_LOG_ONE;*/
min_backoff.setLastWord(-1);
computeEntropyAndConditionals(entropy, all_prob, rnnlm, min_backoff);
min_backoff = getBackoff(rnnlm, min_backoff, set_min_backoff, all_prob, to_be_removed);
cout << "MIN BACKOFF " << min_backoff.toString() << endl;
set_min_backoff.insert(min_backoff);
// addFstState(id, min_backoff, fst);
// q.push(min_backoff);
// Estimate number of backoff loop to bound the backoff path length
// float ratioa = 0.0;
// float ratiob = 0.0;
float ratio = 0.0;
// for (int i=0; i < min_backoff.getNumDims(); i++) {
// if (min_backoff.getDim(i) == 1) {
// ratioa++;
// }
// if (fsth.getDim(i) == 1) {
// ratiob++;
// }
// }
// ratioa /= min_backoff.getNumDims();
// ratiob /= min_backoff.getNumDims();
// ratio = (ratioa*(1.0-ratiob))+(ratiob*(1.0-ratioa));
ratio=1.0;
// printf("ratio=%f\t%i BO loops\n", ratio, n_bo_loops);
//foreach state in the queue
while (!q.empty()) {
fsth = q.front();
q.pop();
id = h2state[&fsth];
state2h.push_back(new NeuronFstHistory(fsth));
if (id == FINAL_STATE) { continue; }
dprintf(1,"-- STUDY STATE %li = %s\n", id, fsth.toString().c_str());
/* try { posterior.at(id) = MY_LOG_ONE; }
catch (exception e) {
posterior.resize((int) (posterior.size()*1.5)+1);
posterior.at(id) = MY_LOG_ONE;
}*/
computeEntropyAndConditionals(entropy, all_prob, rnnlm, fsth);
//compute BO in advance and check if it is a min BO node
bo_fsth = getBackoff(rnnlm, fsth, set_min_backoff, all_prob, to_be_removed);
if (bo_fsth == fsth) { bo_fsth = min_backoff; }
//foreach w (ie, foreach word of each class c)
//test if the edge has to kept or removed
backoff = false; //no backoff yet since no edge has been removed
for (w=0; w < rnnlm.getVocabSize(); w++) {
p = all_prob[w];
/*p_joint = exp(-posterior[id]-p);*/
p_joint = exp(-p);
delta = -1.0*p_joint*log2(p_joint);
//accept edge if this leads to a minimum
//relative gain of the entropy
dprintf(2,"P = %e \tP_joint = %e \tH = %e \tDelta =%e \tDelta H = %.6f %%\n",exp(-p), p_joint, entropy, delta, 100.0*delta/entropy);
if (set_min_backoff.find(fsth) != set_min_backoff.end() || (delta > pruning_threshold*entropy)) {
// if ((fsth == min_backoff) || (delta > pruning_threshold*entropy)) {
next_n_added++;
to_be_added.push_back(w);
to_be_added_prob.push_back(p);
dprintf(2,"\tACCEPT [%li] -- %i (%s) / %f --> ...\t(%e > %e)\n", id, w, rnnlm.getWordString(w), p, delta, pruning_threshold*entropy);
// to_be_removed.push_back(w);
}
//backoff
else {
// to_be_removed.push_back(w);
backoff = true;
dprintf(2,"\tPRUNE [%li] -- %i / %f --> ...\n", id, w, p);
}
//print
if (next_n_processed % 100000 == 0) {
fprintf(stderr, "\rH=%.5f / N proc'd=%li / N added=%li (%.5f %%) / N bo=%li (%.5f %%) / %li/%li Nodes (%2.1f %%) / N min BO=%i", entropy, n_processed, n_added, ((float) n_added/ (float)n_processed)*100.0, n_backoff, ((float) n_backoff/ (float)n_added)*100.0, id, id+q.size(), 100.0 - (float) (100.0*id/(id+q.size())), (int) set_min_backoff.size());
}
next_n_processed++;
// }
}
//Set a part of the new FST history
new_fsth.setFstHistory(rnnlm, *dzer);
//if at least one word is backing off
if (backoff) {
n_backoff++;
if (to_be_added.size() == 0) {
n_only_backoff++;
}
if (addFstState(new_id, new NeuronFstHistory(bo_fsth), fst)) {
q.push(bo_fsth);
try { non_bo_pred.at(new_id) = false; }
catch (exception e) {
non_bo_pred.resize(new_id+(int) (non_bo_pred.size()*0.5)+1);
non_bo_pred.at(new_id) = false;
}
}
dprintf(1,"BACKOFF\t[%li]\t(%s)\n-------\t[%li]\t(%s)\n", id, fsth.toString().c_str(), new_id, bo_fsth.toString().c_str());
fst.AddArc(id, LogArc(EPSILON, EPSILON, LogWeight::Zero(), new_id));
addPred(pred, new_id, id);
}
vector<real>::iterator it_p = to_be_added_prob.begin();
for (vector<int>::iterator it = to_be_added.begin(); it != to_be_added.end(); ++it) {
w = *it;
p = *it_p;
if (w == 0) {
fst.AddArc(id, LogArc(FstWord(w),FstWord(w),p,FINAL_STATE));
dprintf(1,"EDGE [%li] (%s)\n---- %i (%s) / %f -->\n---- [%li] FINAL STATE)\n\n", id, fsth.toString().c_str(), FstWord(w), rnnlm.getWordString(w), p, FINAL_STATE);
}
//accept edge
else {
new_fsth.setLastWord(w);
//if sw not in the memory
//then add a new state for sw in the FST and push sw in the queue
if (addFstState(new_id, new NeuronFstHistory(new_fsth), fst)) {
q.push(new_fsth);
try { non_bo_pred.at(new_id) = true; }
catch (exception e) {
non_bo_pred.resize(new_id+(int) (non_bo_pred.size()*0.5)+1);
non_bo_pred.at(new_id) = true;
}
}
else { /* already exists */ }
//add the edge in the FST
non_bo_pred.at(new_id) = true;
fst.AddArc(id, LogArc(FstWord(w),FstWord(w),p,new_id));
dprintf(1,"EDGE [%li] (%s)\n---- %i (%s) / %f -->\n---- [%li] (%s)\n\n", id, fsth.toString().c_str(), FstWord(w), rnnlm.getWordString(w), p, new_id, new_fsth.toString().c_str());
// posterior.at(new_id) += posterior[id]*p;
}
/*if (posterior[id]+p < LogWeight::Zero().Value()) {
p_joint = exp(-posterior[id]-p);
entropy -= p_joint*log2(p_joint);
}*/
++it_p;
}
n_added = next_n_added;
n_processed = next_n_processed;
//reset queues
to_be_added.clear();
to_be_added_prob.clear();
// to_be_removed.clear();
}
cout << endl;
//compute backoff weights
deleted = compactBackoffNodes(fst, pred, non_bo_pred);
computeAllBackoff(fst, pred);
//remove useless nodes
removeStates(fst, new_fst, deleted);
fst.DeleteStates();
fst = new_fst;
//Fill the table of symbols
SymbolTable dic("dictionnary");
dic.AddSymbol("*", 0);
for (int i=0; i<rnnlm.getVocabSize(); i++) {
dic.AddSymbol(string(rnnlm.getWordString(i)), i+1);
}
fst.SetInputSymbols(&dic);
fst.SetOutputSymbols(&dic);
//printf("H=%.5f / N proc'd=%li / N added=%li (%.5f %%) %li/%li Nodes (%2.1f %%)\n", entropy, n_processed, n_added, ((float) n_added/ (float)n_processed)*100.0, id, id+q.size(), 100.0 - (float) (100.0*id/(id+q.size())));
cout << "END" << endl;
}
void
train_model(string eps, string s1s2_sep, string skip, int order,
string smooth, string prefix, string seq_sep, string prune,
double theta, string count_pattern)
{
namespace s = fst::script;
using fst::script::FstClass;
using fst::script::MutableFstClass;
using fst::script::VectorFstClass;
using fst::script::WeightClass;
// create symbols file
cout << "Generating symbols..." << endl;
NGramInput *ingram =
new NGramInput(prefix + ".corpus.aligned", prefix + ".corpus.syms",
"", eps, unknown_symbol, "", "");
ingram->ReadInput(0, 1);
// compile strings into a far archive
cout << "Compiling symbols into FAR archive..." << endl;
fst::FarEntryType fet = fst::StringToFarEntryType(entry_type);
fst::FarTokenType ftt = fst::StringToFarTokenType(token_type);
fst::FarType fartype = fst::FarTypeFromString(far_type);
delete ingram;
vector<string> in_fname;
in_fname.push_back(prefix + ".corpus.aligned");
fst::script::FarCompileStrings(in_fname, prefix + ".corpus.far",
arc_type, fst_type, fartype,
generate_keys, fet, ftt,
prefix + ".corpus.syms", unknown_symbol,
keep_symbols, initial_symbols,
allow_negative_labels, file_list_input,
key_prefix, key_suffix);
//count n-grams
cout << "Counting n-grams..." << endl;
NGramCounter<Log64Weight> ngram_counter(order, epsilon_as_backoff);
FstReadOptions opts;
FarReader<StdArc> *far_reader;
far_reader = FarReader<StdArc>::Open(prefix + ".corpus.far");
int fstnumber = 1;
const Fst<StdArc> *ifst = 0, *lfst = 0;
while (!far_reader->Done()) {
if (ifst)
delete ifst;
ifst = far_reader->GetFst().Copy();
if (!ifst) {
E_FATAL("ngramcount: unable to read fst #%d\n", fstnumber);
//exit(1);
}
bool counted = false;
if (ifst->Properties(kString | kUnweighted, true)) {
counted = ngram_counter.Count(*ifst);
}
else {
VectorFst<Log64Arc> log_ifst;
Map(*ifst, &log_ifst, ToLog64Mapper<StdArc> ());
counted = ngram_counter.Count(&log_ifst);
}
if (!counted)
cout << "ngramcount: fst #" << fstnumber << endl;
if (ifst->InputSymbols() != 0) { // retain for symbol table
if (lfst)
delete lfst; // delete previously observed symbol table
lfst = ifst;
ifst = 0;
}
far_reader->Next();
++fstnumber;
}
delete far_reader;
if (!lfst) {
E_FATAL("None of the input FSTs had a symbol table\n");
//exit(1);
}
VectorFst<StdArc> vfst;
ngram_counter.GetFst(&vfst);
ArcSort(&vfst, StdILabelCompare());
vfst.SetInputSymbols(lfst->InputSymbols());
vfst.SetOutputSymbols(lfst->InputSymbols());
vfst.Write(prefix + ".corpus.cnts");
StdMutableFst *fst =
StdMutableFst::Read(prefix + ".corpus.cnts", true);
if (smooth != "no") {
cout << "Smoothing model..." << endl;
bool prefix_norm = 0;
if (smooth == "presmoothed") { // only for use with randgen counts
prefix_norm = 1;
smooth = "unsmoothed"; // normalizes only based on prefix count
}
if (smooth == "kneser_ney") {
NGramKneserNey ngram(fst, backoff, backoff_label,
norm_eps, check_consistency,
discount_D, bins);
ngram.MakeNGramModel();
fst = ngram.GetMutableFst();
}
else if (smooth == "absolute") {
NGramAbsolute ngram(fst, backoff, backoff_label,
norm_eps, check_consistency,
discount_D, bins);
ngram.MakeNGramModel();
fst = ngram.GetMutableFst();
}
else if (smooth == "katz") {
NGramKatz ngram(fst, backoff, backoff_label,
norm_eps, check_consistency, bins);
ngram.MakeNGramModel();
fst = ngram.GetMutableFst();
}
else if (smooth == "witten_bell") {
NGramWittenBell ngram(fst, backoff, backoff_label,
norm_eps, check_consistency,
witten_bell_k);
ngram.MakeNGramModel();
fst = ngram.GetMutableFst();
}
else if (smooth == "unsmoothed") {
NGramUnsmoothed ngram(fst, 1, prefix_norm, backoff_label,
norm_eps, check_consistency);
ngram.MakeNGramModel();
fst = ngram.GetMutableFst();
}
else {
E_FATAL("Bad smoothing method: %s\n", smooth.c_str());
}
}
if (prune != "no") {
cout << "Pruning model..." << endl;
if (prune == "count_prune") {
NGramCountPrune ngramsh(fst, count_pattern,
shrink_opt, total_unigram_count,
backoff_label, norm_eps,
check_consistency);
ngramsh.ShrinkNGramModel();
}
else if (prune == "relative_entropy") {
NGramRelEntropy ngramsh(fst, theta, shrink_opt,
total_unigram_count, backoff_label,
norm_eps, check_consistency);
ngramsh.ShrinkNGramModel();
}
else if (prune == "seymore") {
NGramSeymoreShrink ngramsh(fst, theta, shrink_opt,
total_unigram_count, backoff_label,
norm_eps, check_consistency);
ngramsh.ShrinkNGramModel();
}
else {
E_FATAL("Bad shrink method: %s\n", prune.c_str());
}
}
cout << "Minimizing model..." << endl;
MutableFstClass *minimized = new s::MutableFstClass(*fst);
Minimize(minimized, 0, fst::kDelta);
fst = minimized->GetMutableFst<StdArc>();
cout << "Correcting final model..." << endl;
StdMutableFst *out = new StdVectorFst();
relabel(fst, out, prefix, eps, skip, s1s2_sep, seq_sep);
cout << "Writing binary model to disk..." << endl;
out->Write(prefix + ".fst");
}