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);
}
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;
}
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;
}