inline
Tp1 dot_product(const WeightVector<Tp1, Alloc1>& x, const WeightVector<Tp2, Alloc2>& y)
{
const size_t size = utils::bithack::min(x.size(), y.size());
return std::inner_product(x.begin(), x.begin() + size, y.begin(), Tp1());
}
int main(int argc, char** argv) {
// handle parameters
po::variables_map cfg;
if (!init_params(argc,argv, &cfg)) exit(1); // something is wrong
// init weights
DenseWeightVector dense_weights;
WeightVector weights;
if (cfg.count("weights")) Weights::InitFromFile(cfg["weights"].as<string>(), &dense_weights);
Weights::InitSparseVector(dense_weights, &weights);
cerr << "Current Weight Vector:\n";
for (WeightVector::iterator i=weights.begin(); i!=weights.end(); ++i)
cerr << i->first << " " << FD::Convert(i->first) << "=" << i->second << endl;
/*cerr << "\nDense Weights:\n";
for (int i = 0 ; i < dense_weights.size(); ++i)
cerr << i << " " << dense_weights[i] << endl;*/
cerr << "# of features: " << FD::NumFeats() << " (-1 dummy feature @ idx 0)\n\n";
// load instances
vector<TrainingInstance> instances;
loadInstances(cfg["input"].as<string>(), instances);
// setup output directory
//MkDirP(cfg["output"].as<string>());
//stringstream outss;
//outss << cfg["output"].as<string>() << "/";
//const string out_path = outss.str();
// setup loss function
ListwiseLossFunction* lossfunc = set_loss(&cfg);
cerr << "listwise loss function: " << cfg["loss"].as<string>() << "\n";
// run AdaRank optimizer
AdaRank adarank(
instances,
instances.size(),
cfg["iterations"].as<int>(),
cfg["epsilon"].as<double>(),
dense_weights,
lossfunc,
cfg.count("verbose")
);
adarank.run();
// write output weight vector
DenseWeightVector new_dense_weights = adarank.GetWeightVector();
WeightVector new_weights;
Weights::InitSparseVector(new_dense_weights, &new_weights);
cerr << "Final Weight Vector:\n";
for (WeightVector::iterator i=new_weights.begin(); i!=new_weights.end(); ++i)
cerr << i->first << " " << FD::Convert(i->first) << "=" << i->second << endl;
Weights::WriteToFile(cfg["output"].as<string>(), new_dense_weights, true, NULL);
}
void loadRelevanceWeights(const string& fname, WeightVector& rw) {
rw.clear();
ReadFile in_file(fname);
istream& in = *in_file.stream();
assert(in);
string f;
double v;
while(in>>f) {
in>>v;
if (f.empty()) continue;
rw.set_value(FD::Convert(f), v);
}
}
int main(int argc, char** argv) {
if (argc < 2 || argc > 3) {
cerr << "USAGE: view-instances <binary instance file> [<weights>]\n";
exit(1);
}
bool has_w = (argc==3);
// load instances
vector<Instance> instances;
loadInstances(string(argv[1]), instances);
// weights
WeightVector weights;
if (has_w) {
DenseWeightVector dense_weights;
Weights::InitFromFile(string(argv[2]), &dense_weights);
Weights::InitSparseVector(dense_weights, &weights);
cerr << "Current Weight Vector:\n";
for (WeightVector::iterator i=weights.begin(); i!=weights.end(); ++i)
cerr << i->first << " " << FD::Convert(i->first) << "=" << i->second << endl;
}
double likelihood = 0.0;
double likelihood_i=0.0;
for (int i=0;i<instances.size();++i) {
if (instances[i].ir_sorted) {
if (has_w)
likelihood_i = PlackettLuce::pl_likelihood(instances[i], weights);
else
likelihood_i = PlackettLuce::pl_likelihood(instances[i]);
cout << "P(y|x;";
if (has_w) cout << "w)=";
else cout << "D)=";
cout << likelihood_i << "\n";
}
if (has_w)
cout << instances[i].AsString(weights) << endl;
else
cout << instances[i].AsString() << endl;
likelihood +=likelihood_i;
}
cerr << "Likelihood=" << likelihood << "\n";
}
boost::shared_ptr<HypothesisInfo> MakeHypothesisInfo(Hypergraph& hg) {
/*
* create an HypothesisInfo with feature vector, translation and its relevance
* relevance feature values are removed (and optionally any frozen features)
*/
boost::shared_ptr<HypothesisInfo> h(new HypothesisInfo);
h->features = ViterbiFeatures(hg);
h->rel = h->features.dot(relw);
// clean relevance weights from feature vector
for (WeightVector::iterator it=relw.begin(); it!=relw.end(); ++it) { h->features.set_value(it->first, .0); }
ViterbiESentence(hg, &(h->hyp));
if (freeze) { for (unsigned x=0;x<frozen_features.size();++x) { h->features.set_value(frozen_features[x], .0); } }
// for rel scaling:
if (h->rel > MAX_REL) MAX_REL = h->rel;
if (h->rel < MIN_REL) MIN_REL = h->rel;
return h;
}
inline
Tp1 dot_product(const WeightVector<Tp1, Alloc1>& x, const WeightVector<Tp2, Alloc2>& y, BinaryOp op)
{
typedef WeightVector<Tp1, Alloc1> weight_vector1_type;
typedef WeightVector<Tp2, Alloc2> weight_vector2_type;
const size_t size = utils::bithack::min(x.size(), y.size());
Tp1 __dot = std::inner_product(x.begin(), x.begin() + size, y.begin(), Tp1(), std::plus<Tp1>(), op);
if (x.size() > y.size()) {
typename weight_vector1_type::const_iterator iter_end = x.end();
for (typename weight_vector1_type::const_iterator iter = x.begin() + size; iter != iter_end; ++ iter)
__dot += op(*iter, Tp2());
} else {
typename weight_vector2_type::const_iterator iter_end = y.end();
for (typename weight_vector2_type::const_iterator iter = y.begin() + size; iter != iter_end; ++ iter)
__dot += op(Tp1(), *iter);
}
return __dot;
}
/// Function that runs one it of the ParEGO algorithm.
void universe::iterate_ParEGO()
{
int prior_it=0;
int stopcounter=0;
weights->changeWeights(iter, space->fWeightVectors);
//fprintf(stdout, "%.2lf %.2lf weightvectors \n", space->fWeightVectors[0], space->fWeightVectors[1]);
//fprintf(stdout, "fMEasureFIt\n");
for(int i=1;i<=iter;i++)
{
space->fMeasuredFit[i] = space->Tcheby(&space->fCostVectors[i][1]);
//fprintf(stdout,"%lg ", space->fMeasuredFit[i]);
if(space->fMeasuredFit[i]<model->ymin)
{
model->ymin=space->fMeasuredFit[i];
best_ever=i;
}
}
//fprintf(stdout,"\n ymin: %lg\n", model->ymin);
//fprintf(stdout,"best_ever: %d\n", best_ever);
//chose the solutions to use to update the DACE model
if(iter>11*space->fSearchSpaceDim+24)
{
model->fCorrelationSize = 11*space->fSearchSpaceDim+24;
space->chooseUpdateSolutions(iter, model->fCorrelationSize);
model->pax=&space->fSelectedXVectors;
model->pay=&space->fSelectedMeasuredFit;
}
else
{
model->fCorrelationSize=iter;
model->pax=&space->fXVectors;
model->pay=&space->fMeasuredFit;
}
model->buildDACE(weights->change, iter);
start = clock();
// BEGIN GA code
double best_imp=INFTY;
double* best_x;
best_x = (double*)calloc(space->fSearchSpaceDim+1, sizeof(double));
//could change the GA not to be an object. have to think about adv and disadv
// pop size inti 20
GeneticAlgorithm ga(20, space->fSearchSpaceDim);
ga.run(space, model, iter, best_x, &best_imp);
// END GA code
end = clock();
cpu_time_used = ((double) (end - start)) / CLOCKS_PER_SEC;
// timestr << iter << " " << cpu_time_used << endl;
//fprintf(stdout, "ax\n");
for(int d=1;d<=space->fSearchSpaceDim;d++)
{
space->fXVectors[iter+1][d]=best_x[d];
//fprintf(stdout, "%lg ", space->fXVectors[iter+1][d]);
}
//fprintf(stdout, "\n");
space->fMeasuredFit[iter+1]=space->myfit(iter+1, iter+1);
fprintf(stdout, "%d ", iter+1+prior_it);
for(int d=1; d <=space->fSearchSpaceDim; d++)
fprintf(stdout, "%lg ", space->fXVectors[iter+1][d]);
fprintf(stdout, "decision\n");
fprintf(stdout,"%d ", iter+1);
for(int i=1;i<=space->fNoObjectives;i++)
{
fprintf(stdout, "%lg ", space->fCostVectors[iter+1][i]);
fprintf(plotfile, "%lg ", space->fCostVectors[iter+1][i]);
//fprintf(plotfile, "%.5lf ", ff[iter+1][i]);
}
fprintf(plotfile, "\n");
fprintf(stdout, "objective\n");
//cout<<"ymin"<<model->ymin<<"\n";
improvements[iter+1]=improvements[iter];
if (space->fMeasuredFit[iter+1]>=model->ymin)
{
// fprintf(stdout,"No actual improver found\n");
stopcounter++;
}
else
{
improvements[iter+1]++;
model->ymin = space->fMeasuredFit[iter+1];
stopcounter=0;
best_ever=iter+1;
}
iter++;
}
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;
}
//updates relw_scaled by scaling it to the current norm of w, multiplied by the scalingfactor
void scaleRelevanceWeights(const double scalingfactor) {
relw_scaled /= relw_scaled.l2norm(); // to length 1: relw / ||relw||
const double cur_w_norm = w.l2norm(); // ||w||
if (cur_w_norm == 0) return;
relw_scaled *= cur_w_norm * scalingfactor; // scale by ||w||*scalingfactor
}