int main(int argc, char** argv) {
po::variables_map cfg;
if (!init_params(argc,argv,&cfg)) return 1;
if (cfg.count("random_seed"))
rng.reset(new MT19937(cfg["random_seed"].as<uint32_t>()));
else
rng.reset(new MT19937);
// setup decoder
Decoder* decoder = setupDecoder(cfg);
if (!decoder) {
cerr << "error while loading decoder with" << cfg["decoder_config"].as<string>() << "!\n";
return 1;
}
TrainingObserver observer;
// get reference to decoder weights
vector<weight_t>& decoder_weights = decoder->CurrentWeightVector();
// setup weights
WeightVector w, w_hope, w_fear;
// the SMT weights (to be optimized)
Weights::InitFromFile(cfg["weights"].as<string>(), &decoder_weights);
Weights::InitSparseVector(decoder_weights, &w);
loadWeights(cfg["rweights"].as<string>(), w_hope);
WeightVector w_inv = w*-1;
WeightVector w_hope_inv = w_hope*-1;
//cerr << "W " << w << endl;
//cerr << "WINV " << w_inv << endl;
//cerr << "R " << w_hope << endl;
//cerr << "RINV " << w_hope_inv << endl;
const string input = decoder->GetConf()["input"].as<string>();
//cerr << "Reading input from " << ((input == "-") ? "STDIN" : input.c_str()) << endl << endl;
ReadFile in_read(input);
istream *in = in_read.stream();
assert(*in);
string id, sentence;
std::vector<HypergraphSampler::Hypothesis> samples;
while(*in >> id) {
in->ignore(1, '\t');
getline(*in, sentence);
if (sentence.empty() || id.empty()) continue;
//decoder->SetId(id);
decoder->Decode(sentence, &observer); // decode with decoder_weights
Hypergraph hg = observer.GetCurrentForest();
// get max model score
double max_tscore = ViterbiFeatures(hg).dot(w);
// get min model score
hg.Reweight(w_inv);
double min_tscore = -ViterbiFeatures(hg).dot(w_inv);
// get max rel score
hg.Reweight(w_hope);
double max_rscore = ViterbiFeatures(hg).dot(w_hope);
// get min rel_score
hg.Reweight(w_hope_inv);
double min_rscore = -ViterbiFeatures(hg).dot(w_hope_inv);
//cerr << max_tscore << " " << min_tscore << " " << max_rscore << " " << min_rscore << endl;
if (cfg.count("sample")) {
HypergraphSampler::sample_hypotheses(hg, cfg["sample"].as<int>(), &(*rng), &samples);
for (unsigned s=0;s<samples.size();++s) {
const HypergraphSampler::Hypothesis& h = samples[s];
cout << id << "\t" << "S\t" << vscale(h.fmap.dot(w), min_tscore, max_tscore) <<
"\t" << vscale(h.fmap.dot(w_hope), min_rscore, max_rscore) <<
"\t" << TD::GetString(h.words) << endl;
}
} else if (cfg.count("kbest")) {
typedef KBest::KBestDerivations<vector<WordID>, ESentenceTraversal,KBest::FilterUnique> K;
// get kbest model score derivations
hg.Reweight(w);
K kbest2(hg,cfg["kbest"].as<int>());
for (int i = 0; i < cfg["kbest"].as<int>(); ++i) {
typename K::Derivation *d = kbest2.LazyKthBest(hg.nodes_.size() - 1, i);
if (!d) break;
cout << id << "\t" << "KBT\t" << vscale(d->feature_values.dot(w), min_tscore, max_tscore) <<
"\t" << vscale(d->feature_values.dot(w_hope), min_rscore, max_rscore) <<
"\t" << TD::GetString(d->yield) << endl;
}
// get kworst model score derivations
hg.Reweight(w_inv);
K kbest3(hg,cfg["kbest"].as<int>());
for (int i = 0; i < cfg["kbest"].as<int>(); ++i) {
typename K::Derivation *d = kbest3.LazyKthBest(hg.nodes_.size() - 1, i);
if (!d) break;
cout << id << "\t" << "KWT\t" << vscale(d->feature_values.dot(w), min_tscore, max_tscore) <<
"\t" << vscale(d->feature_values.dot(w_hope), min_rscore, max_rscore) <<
"\t" << TD::GetString(d->yield) << endl;
}
// get kbest rel score derivations
hg.Reweight(w_hope);
K kbest4(hg,cfg["kbest"].as<int>());
for (int i = 0; i < cfg["kbest"].as<int>(); ++i) {
typename K::Derivation *d = kbest4.LazyKthBest(hg.nodes_.size() - 1, i);
if (!d) break;
cout << id << "\t" << "KBR\t" << vscale(d->feature_values.dot(w), min_tscore, max_tscore) <<
"\t" << vscale(d->feature_values.dot(w_hope), min_rscore, max_rscore) <<
"\t" << TD::GetString(d->yield) << endl;
}
// get kbest model score derivations
hg.Reweight(w_hope_inv);
K kbest(hg,cfg["kbest"].as<int>());
for (int i = 0; i < cfg["kbest"].as<int>(); ++i) {
typename K::Derivation *d = kbest.LazyKthBest(hg.nodes_.size() - 1, i);
if (!d) break;
cout << id << "\t" << "KWR\t" << vscale(d->feature_values.dot(w), min_tscore, max_tscore) <<
"\t" << vscale(d->feature_values.dot(w_hope), min_rscore, max_rscore) <<
"\t" << TD::GetString(d->yield) << endl;
}
}
}
delete decoder;
return 0;
}
int main(int argc, char** argv) {
po::variables_map cfg;
if (!init_params(argc,argv,&cfg)) return 1;
if (cfg.count("random_seed"))
rng.reset(new MT19937(cfg["random_seed"].as<uint32_t>()));
else
rng.reset(new MT19937);
// set variables
lr = cfg["learningrate"].as<double>();
hope_select = cfg["hope"].as<int>();
fear_select = cfg["fear"].as<int>();
optimizer = cfg["optimizer"].as<int>();
freeze = cfg.count("freeze");
if (freeze) {
const vector<string>& ffstrs = cfg["freeze"].as<vector<string> >();
stringstream ffss;
ffss << "frozen features: ";
for (vector<string>::const_iterator ffit=ffstrs.begin();ffit!=ffstrs.end();++ffit) {
frozen_features.push_back(FD::Convert(*ffit));
ffss << *ffit << " ";
}
cerr << ffss.str() << endl;
}
scaling = cfg["scaling"].as<int>();
scalingfactor = cfg["scalingfactor"].as<double>();
cerr << "scaling="<< scaling << " scalingfactor=" << scalingfactor << endl;
// setup decoder
Decoder* decoder = setupDecoder(cfg);
if (!decoder) {
cerr << "error while loading decoder with" << cfg["decoder_config"].as<string>() << "!\n";
return 1;
}
TrainingObserver observer;
// get reference to decoder weights
vector<weight_t>& decoder_weights = decoder->CurrentWeightVector();
// the SMT weights (to be optimized)
if (cfg.count("weights")) {
Weights::InitFromFile(cfg["weights"].as<string>(), &decoder_weights);
Weights::InitSparseVector(decoder_weights, &w);
} else {
cerr << "starting with EMPTY weights!\n";
}
// the weight vector that gives the oracle
loadRelevanceWeights(cfg["rweights"].as<string>(), relw);
negrelw -= relw;
relw_scaled = relw;
// initial scaling
if (scaling != 0) scaleRelevanceWeights(scalingfactor);
// output some vector stats
cerr << "W_REL=" << relw << endl;
cerr << "W_REL_SCALED=" << relw_scaled << endl;
cerr << "|W_REL|=" << relw_scaled.size() << endl;
cerr << "|W_SMT|=" << w.size() << endl;
cerr << "hope selection: " << hope_select << endl;
const string input = decoder->GetConf()["input"].as<string>();
cerr << "Reading input from " << ((input == "-") ? "STDIN" : input.c_str()) << endl;
ReadFile in_read(input);
istream *in = in_read.stream();
assert(*in);
string id, sentence;
int cur_sent = 0;
unsigned lc = 0; // line count
double objective=0;
double tot_loss = 0;
WeightVector avg_w = w;
//SparseVector<double> tot;
//SparseVector<double> oldw = w;
//tot.clear();
//tot += w;
while(*in >> id) {
in->ignore(1, '\t');
getline(*in, sentence);
if (sentence.empty() || id.empty()) continue;
cerr << "\nID="<<id << endl;
decoder->SetId(cur_sent);
decoder->Decode(sentence, &observer); // decode with decoder_weights
cur_sent = observer.GetCurrentSent();
Hypergraph hg = observer.GetCurrentForest();
vector<boost::shared_ptr<HypothesisInfo> > S;
MAX_REL = std::numeric_limits<double>::lowest();
MIN_REL = std::numeric_limits<double>::max();
// get viterbi
boost::shared_ptr<HypothesisInfo> viterbi = MakeHypothesisInfo(hg);
// get the true oracle (sets max_rel)
hg.Reweight(relw);
boost::shared_ptr<HypothesisInfo> oracle = MakeHypothesisInfo(hg);
oracle->oracle = oracle;
oracle->computeCost();
// get the worst derivation (to get min_rel)
hg.Reweight(negrelw);
boost::shared_ptr<HypothesisInfo> worst = MakeHypothesisInfo(hg);
worst->oracle = oracle;
worst->computeCost();
if (hope_select == 1) { // hope
hg.Reweight(w + relw_scaled);
S.push_back(MakeHypothesisInfo(hg));
S[0]->oracle = oracle;
S[0]->computeCost();
} else { // true oracle
S.push_back(oracle);
}
// S contains now ONE (hope/oracle) hypothesis
S[0]->computeLoss();
boost::shared_ptr<HypothesisInfo> good = S[0];
viterbi->oracle = oracle;
viterbi->computeCost();
viterbi->computeLoss();
cerr << "min_rel=" << MIN_REL << " max_rel=" << MAX_REL << endl;
cerr << "S[0]=" << S[0] << endl;
boost::shared_ptr<HypothesisInfo> fear;
if (optimizer == 4) { // PA update (single dual coordinate step)
cerr << "PA MIRA (single dual coordinate step)\n";
hg.Reweight(w - relw_scaled);
fear = MakeHypothesisInfo(hg);
fear->oracle = oracle;
fear->computeCost();
fear->computeLoss();
cerr << "LOSS: " << fear->loss;
if (fear->loss > 0.0) {
double diffsqnorm = (good->features - fear->features).l2norm_sq();
double delta;
if (diffsqnorm > 0) {
delta = fear->loss / (diffsqnorm);
if (delta > lr) delta = lr;
w += good->features * delta;
w -= fear->features * delta;
}
}
} else if (optimizer == 1) {// sgd - nonadapted step size
cerr << "SGD\n";
if (fear_select == 1) {
hg.Reweight(w - relw_scaled);
fear = MakeHypothesisInfo(hg);
} else if (fear_select == 2) {
fear = worst;
} else if (fear_select == 3) {
fear = viterbi;
}
w += good->features * lr;
w -= fear->features * lr;
} else if (optimizer == 2) { // PA MIRA with selection from cutting plane
cerr << "PA MIRA with Selection from Cutting Plane\n";
hg.Reweight(w - relw_scaled);
fear = MakeHypothesisInfo(hg);
fear->oracle = oracle;
fear->computeCost();
fear->computeLoss();
if (fear->loss < 0) {
cerr << "FEAR LOSS < 0! THIS SHOULD NOT HAPPEN!\n";
abort();
}
if (fear->loss > good->loss + SMO_EPS) {
S.push_back(fear);
OptimizeSet(S, 1); // only one iteration with a set of two constraints
} else { cerr << "constraint not violated. fear loss:" << fear->loss << "\n"; }
} else if (optimizer == 3) { // Cutting Plane MIRA
cerr << "Cutting Plane MIRA\n";
unsigned cp_iter=0; // Cutting Plane Iteration
bool again = true;
while (again && cp_iter<CP_ITER) {
again = false;
cerr << "CuttingPlane: " << cp_iter << endl;
// find a fear derivation
hg.Reweight(w - relw_scaled);
fear = MakeHypothesisInfo(hg);
fear->oracle = oracle;
fear->computeCost();
fear->computeLoss();
if (fear->loss < 0) {
cerr << "FEAR LOSS < 0! THIS SHOULD NOT HAPPEN!\n";
//abort();
}
// find max loss hypothesis
double max_loss_in_set = (*std::max_element(S.begin(), S.end(), lossComp))->loss;
if (fear->loss > max_loss_in_set + SMO_EPS) {
cerr << "Adding new fear " << fear << " to S\n";
S.push_back(fear);
OptimizeSet(S);
again = true;
} else { cerr << "constraint not violated. fear loss:" << fear->loss << "\n"; }
cp_iter++;
// update losses
//for(unsigned i=0;i<S.size();i++) S[i]->computeLoss();
}
}
cerr << "|W|=" << w.size() << endl;
tot_loss += relscale(viterbi->rel);
//print objective after this sentence
//double w_change = (w - oldw).l2norm_sq();
//double temp_objective = 0.5 * w_change;// + max_step_size * max_fear;
for(int u=0;u!=S.size();u++) {
cerr << "alpha=" << S[u]->alpha << " loss=" << S[u]->loss << endl;
//temp_objective += S[u]->alpha * S[u]->loss;
}
//objective += temp_objective;
//cerr << "SENT OBJ: " << temp_objective << " NEW OBJ: " << objective << endl;
//tot += w;
++lc;
avg_w *= lc;
avg_w = (w + avg_w) / (lc+1);
// set decoder weights for next sentence
decoder_weights.clear();
w.init_vector(&decoder_weights);
// rescale relevance weights to balance with new model after the update
if (scaling == 2) {
scaleRelevanceWeights(scalingfactor);
cerr << "W_REL_SCALED=" << relw_scaled << endl;
}
// viterbi 2 for debugging
//hg.Reweight(w);
//boost::shared_ptr<HypothesisInfo> viterbi2 = MakeHypothesisInfo(hg);
//viterbi2->oracle = oracle;
//viterbi2->computeCost();
//viterbi2->computeLoss();
//fear->computeLoss();
//viterbi->computeLoss();
//good->computeLoss();
cerr << "FEAR : " << fear << " \n" << TD::GetString(fear->hyp) << endl;
cerr << "BEST : " << viterbi << " \n" << TD::GetString(viterbi->hyp) << endl;
//cerr << "BEST2: " << viterbi2 << " \n" << TD::GetString(viterbi2->hyp) << endl;
cerr << "HOPE : " << good << " \n" << TD::GetString(good->hyp) << endl;
cout << id << " ||| " << TD::GetString(fear->hyp) << " ||| " << TD::GetString(viterbi->hyp) << " ||| " << TD::GetString(good->hyp) << endl;
S.clear();
fear.reset();
viterbi.reset();
//viterbi2.reset();
good.reset();
worst.reset();
oracle.reset();
}
//cerr << "FINAL OBJECTIVE: "<< objective << endl;
cerr << "Translated " << lc << " sentences\n";
cerr << " [AVG METRIC LAST PASS="******"]\n";
//tot_loss = 0;
decoder_weights.clear();
w.init_vector(&decoder_weights);
//Weights::ShowLargestFeatures(decoder_weights);
// write weights
int node_id = rng->next() * 100000;
cerr << " Writing model to " << node_id << endl;
ostringstream os;
os << cfg["weights_output"].as<string>() << "/last." << node_id;
string msg = "HGMIRA tuned weights ||| " + boost::lexical_cast<std::string>(node_id) + " ||| " + boost::lexical_cast<std::string>(lc);
Weights::WriteToFile(os.str(), decoder_weights, true, &msg);
//SparseVector<double> x = tot;
//x /= lc+1;
ostringstream sa;
string msga = "HGMIRA tuned weights AVERAGED ||| " + boost::lexical_cast<std::string>(node_id) + " ||| " + boost::lexical_cast<std::string>(lc);
sa << cfg["weights_output"].as<string>() << "/avg." << node_id;
avg_w.init_vector(&decoder_weights);
Weights::WriteToFile(sa.str(), decoder_weights, true, &msga);
delete decoder;
cerr << "\ndone.\n";
return 0;
}