From 2b9cfc501f3d753c267a8273038cc36f40b488aa Mon Sep 17 00:00:00 2001 From: Chris Dyer Date: Mon, 21 Mar 2011 22:10:02 -0400 Subject: add support for normalized 'summary features'- seemingly sound way of dealing with normalization problems in embedded crf translation models --- decoder/decoder.cc | 22 +++++++++++++++++++++- 1 file changed, 21 insertions(+), 1 deletion(-) (limited to 'decoder/decoder.cc') diff --git a/decoder/decoder.cc b/decoder/decoder.cc index 95ff6270..8a03c5c9 100644 --- a/decoder/decoder.cc +++ b/decoder/decoder.cc @@ -141,12 +141,13 @@ inline shared_ptr make_fsa_ff(string const& ffp,bool verbose // and then prune the resulting (rescored) hypergraph. All feature values from previous // passes are carried over into subsequent passes (where they may have different weights). struct RescoringPass { - RescoringPass() : density_prune(), beam_prune() {} + RescoringPass() : fid_summary(), density_prune(), beam_prune() {} shared_ptr models; shared_ptr inter_conf; vector ffs; shared_ptr w; // null == use previous weights vector weight_vector; + int fid_summary; // 0 == no summary feature double density_prune; // 0 == don't density prune double beam_prune; // 0 == don't beam prune }; @@ -155,6 +156,7 @@ ostream& operator<<(ostream& os, const RescoringPass& rp) { os << "[num_fn=" << rp.ffs.size(); if (rp.inter_conf) { os << " int_alg=" << *rp.inter_conf; } if (rp.w) os << " new_weights"; + if (rp.fid_summary) os << " summary_feature=" << FD::Convert(rp.fid_summary); if (rp.density_prune) os << " density_prune=" << rp.density_prune; if (rp.beam_prune) os << " beam_prune=" << rp.beam_prune; os << ']'; @@ -361,18 +363,21 @@ DecoderImpl::DecoderImpl(po::variables_map& conf, int argc, char** argv, istream ("weights,w",po::value(),"Feature weights file (initial forest / pass 1)") ("feature_function,F",po::value >()->composing(), "Pass 1 additional feature function(s) (-L for list)") ("intersection_strategy,I",po::value()->default_value("cube_pruning"), "Pass 1 intersection strategy for incorporating finite-state features; values include Cube_pruning, Full") + ("summary_feature", po::value(), "Compute a 'summary feature' at the end of the pass (before any pruning) with name=arg and value=inside-outside/Z") ("density_prune", po::value(), "Pass 1 pruning: keep no more than this many times the number of edges used in the best derivation tree (>=1.0)") ("beam_prune", po::value(), "Pass 1 pruning: Prune paths from scored forest, keep paths within exp(alpha>=0)") ("weights2",po::value(),"Optional pass 2") ("feature_function2",po::value >()->composing(), "Optional pass 2") ("intersection_strategy2",po::value()->default_value("cube_pruning"), "Optional pass 2") + ("summary_feature2", po::value(), "Optional pass 2") ("density_prune2", po::value(), "Optional pass 2") ("beam_prune2", po::value(), "Optional pass 2") ("weights3",po::value(),"Optional pass 3") ("feature_function3",po::value >()->composing(), "Optional pass 3") ("intersection_strategy3",po::value()->default_value("cube_pruning"), "Optional pass 3") + ("summary_feature3", po::value(), "Optional pass 3") ("density_prune3", po::value(), "Optional pass 3") ("beam_prune3", po::value(), "Optional pass 3") @@ -559,6 +564,7 @@ DecoderImpl::DecoderImpl(po::variables_map& conf, int argc, char** argv, istream for (int pass = 0; pass < MAX_PASSES; ++pass) { string ws = "weights" + StringSuffixForRescoringPass(pass); string ff = "feature_function" + StringSuffixForRescoringPass(pass); + string sf = "summary_feature" + StringSuffixForRescoringPass(pass); string bp = "beam_prune" + StringSuffixForRescoringPass(pass); string dp = "density_prune" + StringSuffixForRescoringPass(pass); bool first_pass_condition = ((pass == 0) && (conf.count(ff) || conf.count(bp) || conf.count(dp))); @@ -583,6 +589,11 @@ DecoderImpl::DecoderImpl(po::variables_map& conf, int argc, char** argv, istream if (p->IsStateful()) { has_stateful = true; } } } + if (conf.count(sf)) { + rp.fid_summary = FD::Convert(conf[sf].as()); + assert(rp.fid_summary > 0); + // TODO assert that weights for this pass have coef(fid_summary) == 0.0? + } if (conf.count(bp)) { rp.beam_prune = conf[bp].as(); } if (conf.count(dp)) { rp.density_prune = conf[dp].as(); } int palg = (has_stateful ? 1 : 0); // if there are no stateful featueres, default to FULL @@ -794,6 +805,15 @@ bool DecoderImpl::Decode(const string& input, DecoderObserver* o) { cerr << " " << passtr << " partition log(Z): " << log(z) << endl; } + if (rp.fid_summary) { + Hypergraph::EdgeProbs posteriors; + const prob_t z = forest.ComputeEdgePosteriors(1.0, &posteriors); + if (!SILENT) { cerr << " " << passtr << " adding summary feature " << FD::Convert(rp.fid_summary) << " log(Z)=" << log(z) << endl; } + assert(forest.edges_.size() == posteriors.size()); + for (int i = 0; i < posteriors.size(); ++i) + forest.edges_[i].feature_values_.set_value(rp.fid_summary, log(posteriors[i] / z)); + } + string fullbp = "beam_prune" + StringSuffixForRescoringPass(pass); string fulldp = "density_prune" + StringSuffixForRescoringPass(pass); maybe_prune(forest,conf,fullbp.c_str(),fulldp.c_str(),passtr,srclen); -- cgit v1.2.3 From 3e6605c395fa9798f4bd0b7ce2227cd5a5391a25 Mon Sep 17 00:00:00 2001 From: Chris Dyer Date: Tue, 22 Mar 2011 11:35:45 -0400 Subject: check for infs --- decoder/decoder.cc | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-) (limited to 'decoder/decoder.cc') diff --git a/decoder/decoder.cc b/decoder/decoder.cc index 8a03c5c9..a16a9b5a 100644 --- a/decoder/decoder.cc +++ b/decoder/decoder.cc @@ -810,8 +810,12 @@ bool DecoderImpl::Decode(const string& input, DecoderObserver* o) { const prob_t z = forest.ComputeEdgePosteriors(1.0, &posteriors); if (!SILENT) { cerr << " " << passtr << " adding summary feature " << FD::Convert(rp.fid_summary) << " log(Z)=" << log(z) << endl; } assert(forest.edges_.size() == posteriors.size()); - for (int i = 0; i < posteriors.size(); ++i) - forest.edges_[i].feature_values_.set_value(rp.fid_summary, log(posteriors[i] / z)); + if (!isfinite(log(z)) || isnan(log(z))) { + cerr << " " << passtr << " !!! Invalid partition detected, abandoning.\n"; + } else { + for (int i = 0; i < posteriors.size(); ++i) + forest.edges_[i].feature_values_.set_value(rp.fid_summary, log(posteriors[i] / z)); + } } string fullbp = "beam_prune" + StringSuffixForRescoringPass(pass); -- cgit v1.2.3 From dd83d3cff60ddcb49b184927b7ea2ba1214d50eb Mon Sep 17 00:00:00 2001 From: Chris Dyer Date: Tue, 22 Mar 2011 14:44:46 -0400 Subject: fix local normalizer code for summary features --- decoder/decoder.cc | 16 +++++++++++----- decoder/hg.cc | 5 +++-- decoder/hg.h | 2 +- 3 files changed, 15 insertions(+), 8 deletions(-) (limited to 'decoder/decoder.cc') diff --git a/decoder/decoder.cc b/decoder/decoder.cc index a16a9b5a..89425198 100644 --- a/decoder/decoder.cc +++ b/decoder/decoder.cc @@ -806,15 +806,21 @@ bool DecoderImpl::Decode(const string& input, DecoderObserver* o) { } if (rp.fid_summary) { - Hypergraph::EdgeProbs posteriors; - const prob_t z = forest.ComputeEdgePosteriors(1.0, &posteriors); + const prob_t z = forest.PushWeightsToGoal(1.0); if (!SILENT) { cerr << " " << passtr << " adding summary feature " << FD::Convert(rp.fid_summary) << " log(Z)=" << log(z) << endl; } - assert(forest.edges_.size() == posteriors.size()); if (!isfinite(log(z)) || isnan(log(z))) { cerr << " " << passtr << " !!! Invalid partition detected, abandoning.\n"; } else { - for (int i = 0; i < posteriors.size(); ++i) - forest.edges_[i].feature_values_.set_value(rp.fid_summary, log(posteriors[i] / z)); + for (int i = 0; i < forest.edges_.size(); ++i) { + const double log_prob_transition = log(forest.edges_[i].edge_prob_); // locally normalized by the edge + // head node by forest.PushWeightsToGoal + if (!isfinite(log_prob_transition) || isnan(log_prob_transition)) { + cerr << "Edge: i=" << i << " got bad inside prob: " << *forest.edges_[i].rule_ << endl; + abort(); + } + + forest.edges_[i].feature_values_.set_value(rp.fid_summary, log_prob_transition); + } } } diff --git a/decoder/hg.cc b/decoder/hg.cc index 39ac5132..a4028b0e 100644 --- a/decoder/hg.cc +++ b/decoder/hg.cc @@ -226,9 +226,9 @@ prob_t Hypergraph::PushViterbiWeightsToGoal(int fid) { } -void Hypergraph::PushWeightsToGoal(double scale) { +prob_t Hypergraph::PushWeightsToGoal(double scale) { vector posts; - ComputeEdgePosteriors(scale, &posts); + const prob_t inside_z = ComputeEdgePosteriors(scale, &posts); for (int i = 0; i < nodes_.size(); ++i) { const Hypergraph::Node& node = nodes_[i]; prob_t z = prob_t::Zero(); @@ -238,6 +238,7 @@ void Hypergraph::PushWeightsToGoal(double scale) { edges_[node.in_edges_[j]].edge_prob_ = posts[node.in_edges_[j]] / z; } } + return inside_z; } struct EdgeExistsWeightFunction { diff --git a/decoder/hg.h b/decoder/hg.h index aa1202b1..e5ef05f8 100644 --- a/decoder/hg.h +++ b/decoder/hg.h @@ -449,7 +449,7 @@ public: void PushWeightsToSource(double scale = 1.0); // same, except weights are pushed to the goal, works for HGs, // not just lattices - void PushWeightsToGoal(double scale = 1.0); + prob_t PushWeightsToGoal(double scale = 1.0); // contrary to PushWeightsToGoal, use viterbi semiring; store log(p) to fid. note that p_viterbi becomes 1; k*p_viterbi becomes k. also modifies edge_prob_ (note that the fid stored log(p) will stick around even if you reweight) // afterwards, product of edge_prob_ for a derivation will equal 1 for the viterbi (p_v before, 1 after), and in general (k*p_v before, k after). returns inside(goal) -- cgit v1.2.3 From 3ffb9a6573fc59b25ad28885234f1fc921dae51c Mon Sep 17 00:00:00 2001 From: Chris Dyer Date: Tue, 22 Mar 2011 16:48:28 -0400 Subject: reweight after weight pushing to avoid weird output --- decoder/decoder.cc | 1 + 1 file changed, 1 insertion(+) (limited to 'decoder/decoder.cc') diff --git a/decoder/decoder.cc b/decoder/decoder.cc index 89425198..ac063659 100644 --- a/decoder/decoder.cc +++ b/decoder/decoder.cc @@ -821,6 +821,7 @@ bool DecoderImpl::Decode(const string& input, DecoderObserver* o) { forest.edges_[i].feature_values_.set_value(rp.fid_summary, log_prob_transition); } + forest.Reweight(cur_weights); // reset weights } } -- cgit v1.2.3 From 6353b6e86b90d42513de2aad97c419031d01dca3 Mon Sep 17 00:00:00 2001 From: Chris Dyer Date: Wed, 23 Mar 2011 12:15:55 -0400 Subject: yet another feature attempt --- decoder/decoder.cc | 21 +++++++++++++++++++++ 1 file changed, 21 insertions(+) (limited to 'decoder/decoder.cc') diff --git a/decoder/decoder.cc b/decoder/decoder.cc index ac063659..b7774acc 100644 --- a/decoder/decoder.cc +++ b/decoder/decoder.cc @@ -806,6 +806,7 @@ bool DecoderImpl::Decode(const string& input, DecoderObserver* o) { } if (rp.fid_summary) { +#if 0 const prob_t z = forest.PushWeightsToGoal(1.0); if (!SILENT) { cerr << " " << passtr << " adding summary feature " << FD::Convert(rp.fid_summary) << " log(Z)=" << log(z) << endl; } if (!isfinite(log(z)) || isnan(log(z))) { @@ -823,6 +824,26 @@ bool DecoderImpl::Decode(const string& input, DecoderObserver* o) { } forest.Reweight(cur_weights); // reset weights } +#endif + Hypergraph::EdgeProbs posts; + const prob_t z = forest.ComputeEdgePosteriors(1.0, &posts); + if (!isfinite(log(z)) || isnan(log(z))) { + cerr << " " << passtr << " !!! Invalid partition detected, abandoning.\n"; + } else { + for (int i = 0; i < forest.nodes_.size(); ++i) { + const Hypergraph::EdgesVector& in_edges = forest.nodes_[i].in_edges_; + prob_t node_post = prob_t(0); + for (int j = 0; j < in_edges.size(); ++j) + node_post += (posts[in_edges[j]] / z); + const double log_np = log(node_post); + if (!isfinite(log_np) || isnan(log_np)) { + cerr << "got bad posterior prob for node " << i << endl; + abort(); + } + for (int j = 0; j < in_edges.size(); ++j) + forest.edges_[in_edges[j]].feature_values_.set_value(rp.fid_summary, exp(log_np)); + } + } } string fullbp = "beam_prune" + StringSuffixForRescoringPass(pass); -- cgit v1.2.3