-
Notifications
You must be signed in to change notification settings - Fork 2
/
sample-fixed-ncrp.h
127 lines (93 loc) · 4.88 KB
/
sample-fixed-ncrp.h
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
/*
Copyright 2010 Joseph Reisinger
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
*/
// Samples from the Nested Chinese Restaurant Process using a fixed topic
// structure. This model is more expressive in that the topic structure is not
// constrained to be a tree, but rather a digraph.
#ifndef SAMPLE_GEM_FIXED_NCRP_H_
#define SAMPLE_GEM_FIXED_NCRP_H_
#include <string>
#include <vector>
#include "ncrp-base.h"
// The GEM(m, \pi) distribution hyperparameter m, controls the "proportion of
// general words relative to specific words"
DECLARE_double(gem_m);
// The GEM(m, \pi) hyperparameter \pi: reflects how strictly we expect the
// documents to adhere to the m proportions.
DECLARE_double(gem_pi);
// The file path from which to load the topic structure. The file must be
// encoded as one connection per line, child <tab> parent.
DECLARE_string(tree_structure_file);
// Whether or not to use the GEM sampler. The Multinomial sampler currently is
// more flexible as it allows the tree structure to be a DAG; the GEM sampler
// might not work yet with DAGs.
DECLARE_bool(gem_sampler);
// If unset, then just throw away extra edges that cause nodes to have multiple
// parents. Enforcing a tree topology.
DECLARE_bool(use_dag);
// Should non-WN class nodes have words assigned to them? If not, then all
// topics will start with wn_
DECLARE_bool(fold_non_wn);
// Should we perform variable selection (i.e. attribute rejection) based on
// adding a single "REJECT" node with a uniform distribution over the
// vocabulary to each topic list?
DECLARE_bool(use_reject_option);
// Should the hyperparameters on the vocabulary Dirichlet (eta) be learned. For
// now this uses moment matchin to perform the updates.
DECLARE_bool(learn_eta);
// Should all the path combinations to the root be separated out into different
// documents? DAG only.
DECLARE_bool(separate_path_assignments);
// Should we try to learn a single best sense from a list of senses?
DECLARE_bool(sense_selection);
typedef google::dense_hash_map<unsigned, DocToTopicChain> DocSenseToTopicChain;
typedef google::dense_hash_map<unsigned, google::dense_hash_map<unsigned, DocToWordCountMap> > DocSenseWordToCount;
typedef google::dense_hash_map<unsigned, google::dense_hash_map<CRP*,double> > NodeLogFrequencyMap;
// This version differs from the normal GEM sampler in that the tree structure
// is fixed a priori. Hence there is no resampling of c, the path allocations.
class GEMNCRPFixed : public NCRPBase {
public:
GEMNCRPFixed(double m, double pi);
~GEMNCRPFixed() { /* TODO: free memory! */ }
string current_state();
void load_tree_structure(const string& filename);
void load_precomputed_tree_structure(const string& filename);
protected:
void resample_posterior();
void resample_posterior_z_for(unsigned d, bool remove) { resample_posterior_z_for(d, _c[d], _z[d]); }
void resample_posterior_z_for(unsigned d, vector<CRP*>& cd, WordToCountMap& zd);
void resample_posterior_c_for(unsigned d); // used in sense selection
void resample_posterior_eta();
double compute_log_likelihood();
void contract_tree();
void build_path_assignments(CRP* node, vector<CRP*>* c, int sense_index);
void build_separate_path_assignments(CRP* node, vector< vector<CRP*> >* paths);
// Assume that all the words for document d have been assigned using the level
// assignment zd, now remove them all.
void remove_all_words_from(unsigned d, vector<CRP*>& cd, WordToCountMap& zd);
// Assume that all the words for document d have been removed, now add them
// back using the level assignment zd, now remove them all.
void add_all_words_from(unsigned d, vector<CRP*>& cd, WordToCountMap& zd);
// Returns the (unnormalized) path probability for document d given the current
// set of _z assignments
double compute_path_probability_for(unsigned d, vector<CRP*>& cd);
protected:
double _gem_m;
double _pi;
// These hold other possible sense attachments that are not currently
// in use.
DocSenseWordToCount _z_shadow; // level assignments per document, word
DocSenseToTopicChain _c_shadow; // CRP nodes for a document m
NodeLogFrequencyMap _log_node_freq; // Gives the frequency of a sense attachment
unsigned _maxL;
};
#endif // SAMPLE_GEM_FIXED_NCRP_H_