diff options
Diffstat (limited to 'utils/maxent.cpp')
-rw-r--r-- | utils/maxent.cpp | 427 |
1 files changed, 426 insertions, 1 deletions
diff --git a/utils/maxent.cpp b/utils/maxent.cpp index 0f49ee9d..fd772e08 100644 --- a/utils/maxent.cpp +++ b/utils/maxent.cpp @@ -3,12 +3,15 @@ */ #include "maxent.h" + +#include <vector> +#include <iostream> #include <cmath> #include <cstdio> -#include "lbfgs.h" using namespace std; +namespace maxent { double ME_Model::FunctionGradient(const vector<double>& x, vector<double>& grad) { assert((int)_fb.Size() == x.size()); @@ -601,6 +604,428 @@ vector<double> ME_Model::classify(ME_Sample& mes) const { return vp; } +// template<class FuncGrad> +// std::vector<double> +// perform_LBFGS(FuncGrad func_grad, const std::vector<double> & x0); + +std::vector<double> perform_LBFGS( + double (*func_grad)(const std::vector<double> &, std::vector<double> &), + const std::vector<double> &x0); + +std::vector<double> perform_OWLQN( + double (*func_grad)(const std::vector<double> &, std::vector<double> &), + const std::vector<double> &x0, const double C); + +const int LBFGS_M = 10; + +const static int M = LBFGS_M; +const static double LINE_SEARCH_ALPHA = 0.1; +const static double LINE_SEARCH_BETA = 0.5; + +// stopping criteria +int LBFGS_MAX_ITER = 300; +const static double MIN_GRAD_NORM = 0.0001; + +// LBFGS + +double ME_Model::backtracking_line_search(const Vec& x0, const Vec& grad0, + const double f0, const Vec& dx, + Vec& x, Vec& grad1) { + double t = 1.0 / LINE_SEARCH_BETA; + + double f; + do { + t *= LINE_SEARCH_BETA; + x = x0 + t * dx; + f = FunctionGradient(x.STLVec(), grad1.STLVec()); + // cout << "*"; + } while (f > f0 + LINE_SEARCH_ALPHA * t * dot_product(dx, grad0)); + + return f; +} + +// +// Jorge Nocedal, "Updating Quasi-Newton Matrices With Limited Storage", +// Mathematics of Computation, Vol. 35, No. 151, pp. 773-782, 1980. +// +Vec approximate_Hg(const int iter, const Vec& grad, const Vec s[], + const Vec y[], const double z[]) { + int offset, bound; + if (iter <= M) { + offset = 0; + bound = iter; + } else { + offset = iter - M; + bound = M; + } + + Vec q = grad; + double alpha[M], beta[M]; + for (int i = bound - 1; i >= 0; i--) { + const int j = (i + offset) % M; + alpha[i] = z[j] * dot_product(s[j], q); + q += -alpha[i] * y[j]; + } + if (iter > 0) { + const int j = (iter - 1) % M; + const double gamma = ((1.0 / z[j]) / dot_product(y[j], y[j])); + // static double gamma; + // if (gamma == 0) gamma = ((1.0 / z[j]) / dot_product(y[j], y[j])); + q *= gamma; + } + for (int i = 0; i <= bound - 1; i++) { + const int j = (i + offset) % M; + beta[i] = z[j] * dot_product(y[j], q); + q += s[j] * (alpha[i] - beta[i]); + } + + return q; +} + +vector<double> ME_Model::perform_LBFGS(const vector<double>& x0) { + const size_t dim = x0.size(); + Vec x = x0; + + Vec grad(dim), dx(dim); + double f = FunctionGradient(x.STLVec(), grad.STLVec()); + + Vec s[M], y[M]; + double z[M]; // rho + + for (int iter = 0; iter < LBFGS_MAX_ITER; iter++) { + + fprintf(stderr, "%3d obj(err) = %f (%6.4f)", iter + 1, -f, _train_error); + if (_nheldout > 0) { + const double heldout_logl = heldout_likelihood(); + fprintf(stderr, " heldout_logl(err) = %f (%6.4f)", heldout_logl, + _heldout_error); + } + fprintf(stderr, "\n"); + + if (sqrt(dot_product(grad, grad)) < MIN_GRAD_NORM) break; + + dx = -1 * approximate_Hg(iter, grad, s, y, z); + + Vec x1(dim), grad1(dim); + f = backtracking_line_search(x, grad, f, dx, x1, grad1); + + s[iter % M] = x1 - x; + y[iter % M] = grad1 - grad; + z[iter % M] = 1.0 / dot_product(y[iter % M], s[iter % M]); + x = x1; + grad = grad1; + } + + return x.STLVec(); +} + +// OWLQN + +// stopping criteria +int OWLQN_MAX_ITER = 300; + +Vec approximate_Hg(const int iter, const Vec& grad, const Vec s[], + const Vec y[], const double z[]); + +inline int sign(double x) { + if (x > 0) return 1; + if (x < 0) return -1; + return 0; +}; + +static Vec pseudo_gradient(const Vec& x, const Vec& grad0, const double C) { + Vec grad = grad0; + for (size_t i = 0; i < x.Size(); i++) { + if (x[i] != 0) { + grad[i] += C * sign(x[i]); + continue; + } + const double gm = grad0[i] - C; + if (gm > 0) { + grad[i] = gm; + continue; + } + const double gp = grad0[i] + C; + if (gp < 0) { + grad[i] = gp; + continue; + } + grad[i] = 0; + } + + return grad; +} + +double ME_Model::regularized_func_grad(const double C, const Vec& x, + Vec& grad) { + double f = FunctionGradient(x.STLVec(), grad.STLVec()); + for (size_t i = 0; i < x.Size(); i++) { + f += C * fabs(x[i]); + } + + return f; +} + +double ME_Model::constrained_line_search(double C, const Vec& x0, + const Vec& grad0, const double f0, + const Vec& dx, Vec& x, Vec& grad1) { + // compute the orthant to explore + Vec orthant = x0; + for (size_t i = 0; i < orthant.Size(); i++) { + if (orthant[i] == 0) orthant[i] = -grad0[i]; + } + + double t = 1.0 / LINE_SEARCH_BETA; + + double f; + do { + t *= LINE_SEARCH_BETA; + x = x0 + t * dx; + x.Project(orthant); + // for (size_t i = 0; i < x.Size(); i++) { + // if (x0[i] != 0 && sign(x[i]) != sign(x0[i])) x[i] = 0; + // } + + f = regularized_func_grad(C, x, grad1); + // cout << "*"; + } while (f > f0 + LINE_SEARCH_ALPHA * dot_product(x - x0, grad0)); + + return f; +} + +vector<double> ME_Model::perform_OWLQN(const vector<double>& x0, + const double C) { + const size_t dim = x0.size(); + Vec x = x0; + + Vec grad(dim), dx(dim); + double f = regularized_func_grad(C, x, grad); + + Vec s[M], y[M]; + double z[M]; // rho + + for (int iter = 0; iter < OWLQN_MAX_ITER; iter++) { + Vec pg = pseudo_gradient(x, grad, C); + + fprintf(stderr, "%3d obj(err) = %f (%6.4f)", iter + 1, -f, _train_error); + if (_nheldout > 0) { + const double heldout_logl = heldout_likelihood(); + fprintf(stderr, " heldout_logl(err) = %f (%6.4f)", heldout_logl, + _heldout_error); + } + fprintf(stderr, "\n"); + + if (sqrt(dot_product(pg, pg)) < MIN_GRAD_NORM) break; + + dx = -1 * approximate_Hg(iter, pg, s, y, z); + if (dot_product(dx, pg) >= 0) dx.Project(-1 * pg); + + Vec x1(dim), grad1(dim); + f = constrained_line_search(C, x, pg, f, dx, x1, grad1); + + s[iter % M] = x1 - x; + y[iter % M] = grad1 - grad; + z[iter % M] = 1.0 / dot_product(y[iter % M], s[iter % M]); + + x = x1; + grad = grad1; + } + + return x.STLVec(); +} + +// SGD + +// const double SGD_ETA0 = 1; +// const double SGD_ITER = 30; +// const double SGD_ALPHA = 0.85; + +//#define FOLOS_NAIVE +//#define FOLOS_LAZY +#define SGD_CP + +inline void apply_l1_penalty(const int i, const double u, vector<double>& _vl, + vector<double>& q) { + double& w = _vl[i]; + const double z = w; + double& qi = q[i]; + if (w > 0) { + w = max(0.0, w - (u + qi)); + } else if (w < 0) { + w = min(0.0, w + (u - qi)); + } + qi += w - z; +} + +static double l1norm(const vector<double>& v) { + double sum = 0; + for (size_t i = 0; i < v.size(); i++) sum += abs(v[i]); + return sum; +} + +inline void update_folos_lazy(const int iter_sample, const int k, + vector<double>& _vl, + const vector<double>& sum_eta, + vector<int>& last_updated) { + const double penalty = sum_eta[iter_sample] - sum_eta[last_updated[k]]; + double& x = _vl[k]; + if (x > 0) + x = max(0.0, x - penalty); + else + x = min(0.0, x + penalty); + last_updated[k] = iter_sample; +} + +int ME_Model::perform_SGD() { + if (_l2reg > 0) { + cerr << "error: L2 regularization is currently not supported in SGD mode." + << endl; + exit(1); + } + + cerr << "performing SGD" << endl; + + const double l1param = _l1reg; + + const int d = _fb.Size(); + + vector<int> ri(_vs.size()); + for (size_t i = 0; i < ri.size(); i++) ri[i] = i; + + vector<double> grad(d); + int iter_sample = 0; + const double eta0 = SGD_ETA0; + + // cerr << "l1param = " << l1param << endl; + cerr << "eta0 = " << eta0 << " alpha = " << SGD_ALPHA << endl; + + double u = 0; + vector<double> q(d, 0); + vector<int> last_updated(d, 0); + vector<double> sum_eta; + sum_eta.push_back(0); + + for (int iter = 0; iter < SGD_ITER; iter++) { + + random_shuffle(ri.begin(), ri.end()); + + double logl = 0; + int ncorrect = 0, ntotal = 0; + for (size_t i = 0; i < _vs.size(); i++, ntotal++, iter_sample++) { + const Sample& s = _vs[ri[i]]; + +#ifdef FOLOS_LAZY + for (vector<int>::const_iterator j = s.positive_features.begin(); + j != s.positive_features.end(); j++) { + for (vector<int>::const_iterator k = _feature2mef[*j].begin(); + k != _feature2mef[*j].end(); k++) { + update_folos_lazy(iter_sample, *k, _vl, sum_eta, last_updated); + } + } +#endif + + vector<double> membp(_num_classes); + const int max_label = conditional_probability(s, membp); + + const double eta = + eta0 * pow(SGD_ALPHA, + (double)iter_sample / _vs.size()); // exponential decay + // const double eta = eta0 / (1.0 + (double)iter_sample / + // _vs.size()); + + // if (iter_sample % _vs.size() == 0) cerr << "eta = " << eta << + // endl; + u += eta * l1param; + + sum_eta.push_back(sum_eta.back() + eta * l1param); + + logl += log(membp[s.label]); + if (max_label == s.label) ncorrect++; + + // binary features + for (vector<int>::const_iterator j = s.positive_features.begin(); + j != s.positive_features.end(); j++) { + for (vector<int>::const_iterator k = _feature2mef[*j].begin(); + k != _feature2mef[*j].end(); k++) { + const double me = membp[_fb.Feature(*k).label()]; + const double ee = (_fb.Feature(*k).label() == s.label ? 1.0 : 0); + const double grad = (me - ee); + _vl[*k] -= eta * grad; +#ifdef SGD_CP + apply_l1_penalty(*k, u, _vl, q); +#endif + } + } + // real-valued features + for (vector<pair<int, double> >::const_iterator j = s.rvfeatures.begin(); + j != s.rvfeatures.end(); j++) { + for (vector<int>::const_iterator k = _feature2mef[j->first].begin(); + k != _feature2mef[j->first].end(); k++) { + const double me = membp[_fb.Feature(*k).label()]; + const double ee = (_fb.Feature(*k).label() == s.label ? 1.0 : 0); + const double grad = (me - ee) * j->second; + _vl[*k] -= eta * grad; +#ifdef SGD_CP + apply_l1_penalty(*k, u, _vl, q); +#endif + } + } + +#ifdef FOLOS_NAIVE + for (size_t j = 0; j < d; j++) { + double& x = _vl[j]; + if (x > 0) + x = max(0.0, x - eta * l1param); + else + x = min(0.0, x + eta * l1param); + } +#endif + } + logl /= _vs.size(); +// fprintf(stderr, "%4d logl = %8.3f acc = %6.4f ", iter, logl, +// (double)ncorrect / ntotal); + +#ifdef FOLOS_LAZY + if (l1param > 0) { + for (size_t j = 0; j < d; j++) + update_folos_lazy(iter_sample, j, _vl, sum_eta, last_updated); + } +#endif + + double f = logl; + if (l1param > 0) { + const double l1 = + l1norm(_vl); // this is not accurate when lazy update is used + // cerr << "f0 = " << update_model_expectation() - l1param * l1 << " + // "; + f -= l1param * l1; + int nonzero = 0; + for (int j = 0; j < d; j++) + if (_vl[j] != 0) nonzero++; + // cerr << " f = " << f << " l1 = " << l1 << " nonzero_features = " + // << nonzero << endl; + } + // fprintf(stderr, "%4d obj = %7.3f acc = %6.4f", iter+1, f, + // (double)ncorrect/ntotal); + // fprintf(stderr, "%4d obj = %f", iter+1, f); + fprintf(stderr, "%3d obj(err) = %f (%6.4f)", iter + 1, f, + 1 - (double)ncorrect / ntotal); + + if (_nheldout > 0) { + double heldout_logl = heldout_likelihood(); + // fprintf(stderr, " heldout_logl = %f acc = %6.4f\n", + // heldout_logl, 1 - _heldout_error); + fprintf(stderr, " heldout_logl(err) = %f (%6.4f)", heldout_logl, + _heldout_error); + } + fprintf(stderr, "\n"); + } + + return 0; +} + +} // namespace maxent + /* * $Log: maxent.cpp,v $ * Revision 1.1.1.1 2007/05/15 08:30:35 kyoshida |