From 9ed514b2d10c9507a8399623306582da45b78800 Mon Sep 17 00:00:00 2001 From: Chris Dyer Date: Mon, 7 Nov 2011 18:09:47 -0500 Subject: updates to pro to support regularization to previous weight vectors, regualarization normalization, disable broken regularization tuning --- pro-train/mr_pro_reduce.cc | 82 ++++++++++++++++++++++++++-------------------- 1 file changed, 47 insertions(+), 35 deletions(-) (limited to 'pro-train/mr_pro_reduce.cc') diff --git a/pro-train/mr_pro_reduce.cc b/pro-train/mr_pro_reduce.cc index aff410a0..98cddba2 100644 --- a/pro-train/mr_pro_reduce.cc +++ b/pro-train/mr_pro_reduce.cc @@ -23,11 +23,12 @@ void InitCommandLine(int argc, char** argv, po::variables_map* conf) { po::options_description opts("Configuration options"); opts.add_options() ("weights,w", po::value(), "Weights from previous iteration (used as initialization and interpolation") - ("interpolation,p",po::value()->default_value(0.9), "Output weights are p*w + (1-p)*w_prev") + ("regularize_to_weights,y",po::value()->default_value(0.0), "Differences in learned weights to previous weights are penalized with an l2 penalty with this strength; 0.0 = no effect") + ("interpolate_with_weights,p",po::value()->default_value(1.0), "Output weights are p*w + (1-p)*w_prev; 1.0 = no effect") ("memory_buffers,m",po::value()->default_value(200), "Number of memory buffers (LBFGS)") - ("sigma_squared,s",po::value()->default_value(0.1), "Sigma squared for Gaussian prior") - ("min_reg,r",po::value()->default_value(1e-8), "When tuning (-T) regularization strength, minimum regularization strenght") - ("max_reg,R",po::value()->default_value(10.0), "When tuning (-T) regularization strength, maximum regularization strenght") + ("regularization_strength,C",po::value()->default_value(1.0), "l2 regularization strength") + ("min_reg,r",po::value()->default_value(0.01), "When tuning (-T) regularization strength, minimum regularization strenght") + ("max_reg,R",po::value()->default_value(1e6), "When tuning (-T) regularization strength, maximum regularization strenght") ("testset,t",po::value(), "Optional held-out test set") ("tune_regularizer,T", "Use the held out test set (-t) to tune the regularization strength") ("help,h", "Help"); @@ -95,6 +96,27 @@ void GradAdd(const SparseVector& v, const double scale, vector& weights, + const vector& prev_weights, + vector* g) { + assert(weights.size() == g->size()); + double reg = 0; + for (size_t i = 0; i < weights.size(); ++i) { + const double prev_w_i = (i < prev_weights.size() ? prev_weights[i] : 0.0); + const double& w_i = weights[i]; + double& g_i = (*g)[i]; + reg += C * w_i * w_i; + g_i += 2 * C * w_i; + + const double diff_i = w_i - prev_w_i; + reg += T * diff_i * diff_i; + g_i += 2 * T * diff_i; + } + return reg; +} + double TrainingInference(const vector& x, const vector > >& corpus, vector* g = NULL) { @@ -134,8 +156,10 @@ double TrainingInference(const vector& x, // return held-out log likelihood double LearnParameters(const vector > >& training, const vector > >& testing, - const double sigsq, + const double C, + const double T, const unsigned memory_buffers, + const vector& prev_x, vector* px) { vector& x = *px; vector vg(FD::NumFeats(), 0.0); @@ -157,26 +181,12 @@ double LearnParameters(const vector > >& train } // handle regularizer -#if 1 - double norm = 0; - for (int i = 1; i < x.size(); ++i) { - const double mean_i = 0.0; - const double param = (x[i] - mean_i); - norm += param * param; - vg[i] += param / sigsq; - } - const double reg = norm / (2.0 * sigsq); -#else - double reg = 0; -#endif + double reg = ApplyRegularizationTerms(C, T, x, prev_x, &vg); cll += reg; - cerr << cll << " (REG=" << reg << ")\tPPL=" << ppl << "\t TEST_PPL=" << tppl << "\t"; + cerr << cll << " (REG=" << reg << ")\tPPL=" << ppl << "\t TEST_PPL=" << tppl << "\t" << endl; try { - vector old_x = x; - do { - opt.Optimize(cll, vg, &x); - converged = opt.HasConverged(); - } while (!converged && x == old_x); + opt.Optimize(cll, vg, &x); + converged = opt.HasConverged(); } catch (...) { cerr << "Exception caught, assuming convergence is close enough...\n"; converged = true; @@ -201,13 +211,14 @@ int main(int argc, char** argv) { } const double min_reg = conf["min_reg"].as(); const double max_reg = conf["max_reg"].as(); - double sigsq = conf["sigma_squared"].as(); // will be overridden if parameter is tuned - assert(sigsq > 0.0); + double C = conf["regularization_strength"].as(); // will be overridden if parameter is tuned + const double T = conf["regularize_to_weights"].as(); + assert(C > 0.0); assert(min_reg > 0.0); assert(max_reg > 0.0); assert(max_reg > min_reg); - const double psi = conf["interpolation"].as(); - if (psi < 0.0 || psi > 1.0) { cerr << "Invalid interpolation weight: " << psi << endl; } + const double psi = conf["interpolate_with_weights"].as(); + if (psi < 0.0 || psi > 1.0) { cerr << "Invalid interpolation weight: " << psi << endl; return 1; } ReadCorpus(&cin, &training); if (conf.count("testset")) { ReadFile rf(conf["testset"].as()); @@ -231,14 +242,15 @@ int main(int argc, char** argv) { vector > sp; vector smoothed; if (tune_regularizer) { - sigsq = min_reg; + C = min_reg; const double steps = 18; double sweep_factor = exp((log(max_reg) - log(min_reg)) / steps); cerr << "SWEEP FACTOR: " << sweep_factor << endl; - while(sigsq < max_reg) { - tppl = LearnParameters(training, testing, sigsq, conf["memory_buffers"].as(), &x); - sp.push_back(make_pair(sigsq, tppl)); - sigsq *= sweep_factor; + while(C < max_reg) { + cerr << "C=" << C << "\tT=" <(), prev_x, &x); + sp.push_back(make_pair(C, tppl)); + C *= sweep_factor; } smoothed.resize(sp.size(), 0); smoothed[0] = sp[0].second; @@ -257,16 +269,16 @@ int main(int argc, char** argv) { best_i = i; } } - sigsq = sp[best_i].first; + C = sp[best_i].first; } // tune regularizer - tppl = LearnParameters(training, testing, sigsq, conf["memory_buffers"].as(), &x); + tppl = LearnParameters(training, testing, C, T, conf["memory_buffers"].as(), prev_x, &x); if (conf.count("weights")) { for (int i = 1; i < x.size(); ++i) { x[i] = (x[i] * psi) + prev_x[i] * (1.0 - psi); } } cout.precision(15); - cout << "# sigma^2=" << sigsq << "\theld out perplexity="; + cout << "# C=" << C << "\theld out perplexity="; if (tppl) { cout << tppl << endl; } else { cout << "N/A\n"; } if (sp.size()) { cout << "# Parameter sweep:\n"; -- cgit v1.2.3 From 5c019bda38cb19e09cb16c5eefa6793d920015eb Mon Sep 17 00:00:00 2001 From: Chris Dyer Date: Thu, 10 Nov 2011 20:19:09 +0000 Subject: better defaults for pro --- pro-train/dist-pro.pl | 45 +++++++++++++++++++++------------------------ pro-train/mr_pro_reduce.cc | 8 ++++---- vest/parallelize.pl | 2 +- 3 files changed, 26 insertions(+), 29 deletions(-) (limited to 'pro-train/mr_pro_reduce.cc') diff --git a/pro-train/dist-pro.pl b/pro-train/dist-pro.pl index 4bc9cfe3..6332563f 100755 --- a/pro-train/dist-pro.pl +++ b/pro-train/dist-pro.pl @@ -63,8 +63,8 @@ my $cpbin=1; # regularization strength my $tune_regularizer = 0; -my $reg = 10; -my $reg_previous = 0; +my $reg = 500; +my $reg_previous = 5000; # Process command-line options Getopt::Long::Configure("no_auto_abbrev"); @@ -547,16 +547,12 @@ sub enseg { sub print_help { my $executable = check_output("basename $0"); chomp $executable; - print << "Help"; + print << "Help"; Usage: $executable [options] $executable [options] - Runs a complete MERT optimization and test set decoding, using - the decoder configuration in ini file. Note that many of the - options have default values that are inferred automatically - based on certain conventions. For details, refer to descriptions - of the options --decoder, --weights, and --workdir. + Runs a complete PRO optimization using the ini file specified. Required: @@ -576,6 +572,10 @@ General options: --local Run the decoder and optimizer locally with a single thread. + --use-make + Use make -j to run the optimizer commands (useful on large + shared-memory machines where qsub is unavailable). + --decode-nodes Number of decoder processes to run in parallel. [default=15] @@ -584,7 +584,7 @@ General options: --max-iterations Maximum number of iterations to run. If not specified, defaults - to 10. + to 30. --metric Metric to optimize. @@ -597,10 +597,6 @@ General options: --pmem Amount of physical memory requested for parallel decoding jobs. - --use-make - Use make -j to run the optimizer commands (useful on large - shared-memory machines where qsub is unavailable). - --workdir Directory for intermediate and output files. If not specified, the name is derived from the ini filename. Assuming that the ini @@ -611,21 +607,22 @@ General options: Regularization options: - --interpolate-with-weights - [deprecated] At each iteration the resulting weights are - interpolated with the weights from the previous iteration, with - this factor. - - --tune-regularizer - Hold out one third of the tuning data and used this to tune the - regularization parameter. [this doesn't work well] - --reg - l2 regularization strength + l2 regularization strength [default=500]. The greater this value, + the closer to zero the weights will be. --reg-previous l2 penalty for moving away from the weights from the previous - iteration. + iteration. [default=5000]. The greater this value, the closer + to the previous iteration's weights the next iteration's weights + will be. + +Deprecated options: + + --interpolate-with-weights + [deprecated] At each iteration the resulting weights are + interpolated with the weights from the previous iteration, with + this factor. [default=1.0, i.e., no effect] Help } diff --git a/pro-train/mr_pro_reduce.cc b/pro-train/mr_pro_reduce.cc index 98cddba2..6362ce47 100644 --- a/pro-train/mr_pro_reduce.cc +++ b/pro-train/mr_pro_reduce.cc @@ -23,14 +23,14 @@ void InitCommandLine(int argc, char** argv, po::variables_map* conf) { po::options_description opts("Configuration options"); opts.add_options() ("weights,w", po::value(), "Weights from previous iteration (used as initialization and interpolation") - ("regularize_to_weights,y",po::value()->default_value(0.0), "Differences in learned weights to previous weights are penalized with an l2 penalty with this strength; 0.0 = no effect") - ("interpolate_with_weights,p",po::value()->default_value(1.0), "Output weights are p*w + (1-p)*w_prev; 1.0 = no effect") - ("memory_buffers,m",po::value()->default_value(200), "Number of memory buffers (LBFGS)") - ("regularization_strength,C",po::value()->default_value(1.0), "l2 regularization strength") + ("regularization_strength,C",po::value()->default_value(500.0), "l2 regularization strength") + ("regularize_to_weights,y",po::value()->default_value(5000.0), "Differences in learned weights to previous weights are penalized with an l2 penalty with this strength; 0.0 = no effect") + ("memory_buffers,m",po::value()->default_value(100), "Number of memory buffers (LBFGS)") ("min_reg,r",po::value()->default_value(0.01), "When tuning (-T) regularization strength, minimum regularization strenght") ("max_reg,R",po::value()->default_value(1e6), "When tuning (-T) regularization strength, maximum regularization strenght") ("testset,t",po::value(), "Optional held-out test set") ("tune_regularizer,T", "Use the held out test set (-t) to tune the regularization strength") + ("interpolate_with_weights,p",po::value()->default_value(1.0), "[deprecated] Output weights are p*w + (1-p)*w_prev; 1.0 = no effect") ("help,h", "Help"); po::options_description dcmdline_options; dcmdline_options.add(opts); diff --git a/vest/parallelize.pl b/vest/parallelize.pl index 869f430b..7d0365cc 100755 --- a/vest/parallelize.pl +++ b/vest/parallelize.pl @@ -396,7 +396,7 @@ usage: $name [options] options: - --fork + --use-fork Instead of using qsub, use fork. -e, --error-dir -- cgit v1.2.3