1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
|
//TODO: non-sparse vector for all feature functions? modelset applymodels keeps track of who has what features? it's nice having FF that could generate a handful out of 10000 possible feats, though.
//TODO: actually score rule_feature()==true features once only, hash keyed on rule or modify TRule directly? need to keep clear in forest which features come from models vs. rules; then rescoring could drop all the old models features at once
#include "fast_lexical_cast.hpp"
#include <stdexcept>
#include "ff.h"
#include "tdict.h"
#include "hg.h"
using namespace std;
FeatureFunction::~FeatureFunction() {}
void FeatureFunction::FinalTraversalFeatures(const void* /* ant_state */,
SparseVector<double>* /* features */) const {
}
string FeatureFunction::usage_helper(std::string const& name,std::string const& params,std::string const& details,bool sp,bool sd) {
string r=name;
if (sp) {
r+=": ";
r+=params;
}
if (sd) {
r+="\n";
r+=details;
}
return r;
}
Features FeatureFunction::single_feature(WordID feat) {
return Features(1,feat);
}
Features ModelSet::all_features(std::ostream *warn,bool warn0) {
return ::all_features(models_,weights_,warn,warn0);
}
void show_features(Features const& ffs,DenseWeightVector const& weights_,std::ostream &out,std::ostream &warn,bool warn_zero_wt) {
out << "Weight Feature\n";
for (unsigned i=0;i<ffs.size();++i) {
WordID fid=ffs[i];
string const& fname=FD::Convert(fid);
double wt=weights_[fid];
if (warn_zero_wt && wt==0)
warn<<"WARNING: "<<fname<<" has 0 weight."<<endl;
out << wt << " " << fname<<endl;
}
}
void ModelSet::show_features(std::ostream &out,std::ostream &warn,bool warn_zero_wt)
{
// ::show_features(all_features(),weights_,out,warn,warn_zero_wt);
show_all_features(models_,weights_,out,warn,warn_zero_wt,warn_zero_wt);
}
// Hiero and Joshua use log_10(e) as the value, so I do to
WordPenalty::WordPenalty(const string& param) :
fid_(FD::Convert("WordPenalty")),
value_(-1.0 / log(10)) {
if (!param.empty()) {
cerr << "Warning WordPenalty ignoring parameter: " << param << endl;
}
}
void FeatureFunction::TraversalFeaturesImpl(const SentenceMetadata& smeta,
const Hypergraph::Edge& edge,
const std::vector<const void*>& ant_states,
SparseVector<double>* features,
SparseVector<double>* estimated_features,
void* state) const {
throw std::runtime_error("TraversalFeaturesImpl not implemented - override it or TraversalFeaturesLog.\n");
abort();
}
void WordPenalty::TraversalFeaturesImpl(const SentenceMetadata& smeta,
const Hypergraph::Edge& edge,
const std::vector<const void*>& ant_states,
SparseVector<double>* features,
SparseVector<double>* estimated_features,
void* state) const {
(void) smeta;
(void) ant_states;
(void) state;
(void) estimated_features;
features->set_value(fid_, edge.rule_->EWords() * value_);
}
SourceWordPenalty::SourceWordPenalty(const string& param) :
fid_(FD::Convert("SourceWordPenalty")),
value_(-1.0 / log(10)) {
if (!param.empty()) {
cerr << "Warning SourceWordPenalty ignoring parameter: " << param << endl;
}
}
Features SourceWordPenalty::features() const {
return single_feature(fid_);
}
Features WordPenalty::features() const {
return single_feature(fid_);
}
void SourceWordPenalty::TraversalFeaturesImpl(const SentenceMetadata& smeta,
const Hypergraph::Edge& edge,
const std::vector<const void*>& ant_states,
SparseVector<double>* features,
SparseVector<double>* estimated_features,
void* state) const {
(void) smeta;
(void) ant_states;
(void) state;
(void) estimated_features;
features->set_value(fid_, edge.rule_->FWords() * value_);
}
ArityPenalty::ArityPenalty(const std::string& param) :
value_(-1.0 / log(10)) {
string fname = "Arity_";
unsigned MAX=DEFAULT_MAX_ARITY;
using namespace boost;
if (!param.empty())
MAX=lexical_cast<unsigned>(param);
for (unsigned i = 0; i <= MAX; ++i) {
WordID fid=FD::Convert(fname+lexical_cast<string>(i));
fids_.push_back(fid);
}
while (!fids_.empty() && fids_.back()==0) fids_.pop_back(); // pretty up features vector in case FD was frozen. doesn't change anything
}
Features ArityPenalty::features() const {
return Features(fids_.begin(),fids_.end());
}
void ArityPenalty::TraversalFeaturesImpl(const SentenceMetadata& smeta,
const Hypergraph::Edge& edge,
const std::vector<const void*>& ant_states,
SparseVector<double>* features,
SparseVector<double>* estimated_features,
void* state) const {
(void) smeta;
(void) ant_states;
(void) state;
(void) estimated_features;
unsigned a=edge.Arity();
features->set_value(a<fids_.size()?fids_[a]:0, value_);
}
ModelSet::ModelSet(const vector<double>& w, const vector<const FeatureFunction*>& models) :
models_(models),
weights_(w),
state_size_(0),
model_state_pos_(models.size()) {
for (int i = 0; i < models_.size(); ++i) {
model_state_pos_[i] = state_size_;
state_size_ += models_[i]->NumBytesContext();
}
}
void ModelSet::AddFeaturesToEdge(const SentenceMetadata& smeta,
const Hypergraph& /* hg */,
const FFStates& node_states,
Hypergraph::Edge* edge,
FFState* context,
prob_t* combination_cost_estimate) const {
edge->reset_info();
context->resize(state_size_);
if (state_size_ > 0) {
memset(&(*context)[0], 0, state_size_);
}
SparseVector<double> est_vals; // only computed if combination_cost_estimate is non-NULL
if (combination_cost_estimate) *combination_cost_estimate = prob_t::One();
for (int i = 0; i < models_.size(); ++i) {
const FeatureFunction& ff = *models_[i];
void* cur_ff_context = NULL;
vector<const void*> ants(edge->tail_nodes_.size());
bool has_context = ff.NumBytesContext() > 0;
if (has_context) {
int spos = model_state_pos_[i];
cur_ff_context = &(*context)[spos];
for (int i = 0; i < ants.size(); ++i) {
ants[i] = &node_states[edge->tail_nodes_[i]][spos];
}
}
ff.TraversalFeatures(smeta, *edge, ants, &edge->feature_values_, &est_vals, cur_ff_context);
}
if (combination_cost_estimate)
combination_cost_estimate->logeq(est_vals.dot(weights_));
edge->edge_prob_.logeq(edge->feature_values_.dot(weights_));
}
void ModelSet::AddFinalFeatures(const FFState& state, Hypergraph::Edge* edge,SentenceMetadata const& smeta) const {
assert(1 == edge->rule_->Arity());
edge->reset_info();
for (int i = 0; i < models_.size(); ++i) {
const FeatureFunction& ff = *models_[i];
const void* ant_state = NULL;
bool has_context = ff.NumBytesContext() > 0;
if (has_context) {
int spos = model_state_pos_[i];
ant_state = &state[spos];
}
ff.FinalTraversalFeatures(smeta, *edge, ant_state, &edge->feature_values_);
}
edge->edge_prob_.logeq(edge->feature_values_.dot(weights_));
}
|
