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
|
#ifndef NEURALNETWORK_H
#define NEURALNETWORK_H
#include <vector>
#include <boost/shared_ptr.hpp>
//#include <../3rdparty/Eigen/Dense>
#include <Eigen/Dense>
#include "util.h"
#include "model.h"
#include "propagator.h"
#include "neuralClasses.h"
namespace nplm
{
class neuralNetwork
{
protected:
boost::shared_ptr<model> m;
private:
bool normalization;
double weight;
propagator prop;
std::size_t cache_size;
Eigen::Matrix<int,Dynamic,Dynamic> cache_keys;
std::vector<double> cache_values;
int cache_lookups, cache_hits;
public:
neuralNetwork()
: m(new model()),
normalization(false),
weight(1.),
prop(*m, 1),
cache_size(0)
{
}
void set_normalization(bool value) { normalization = value; }
void set_log_base(double value) { weight = 1./std::log(value); }
// This must be called if the underlying model is resized.
void resize() {
if (cache_size)
{
cache_keys.resize(m->ngram_size, cache_size);
cache_keys.fill(-1);
}
prop.resize();
}
void set_width(int width)
{
prop.resize(width);
}
template <typename Derived>
double lookup_ngram(const Eigen::MatrixBase<Derived> &ngram)
{
assert (ngram.rows() == m->ngram_size);
assert (ngram.cols() == 1);
std::size_t hash;
if (cache_size)
{
// First look in cache
hash = Eigen::hash_value(ngram) % cache_size; // defined in util.h
cache_lookups++;
if (cache_keys.col(hash) == ngram)
{
cache_hits++;
return cache_values[hash];
}
}
// Make sure that we're single threaded. Multithreading doesn't help,
// and in some cases can hurt quite a lot
int save_threads = omp_get_max_threads();
omp_set_num_threads(1);
int save_eigen_threads = Eigen::nbThreads();
Eigen::setNbThreads(1);
#ifdef __INTEL_MKL__
int save_mkl_threads = mkl_get_max_threads();
mkl_set_num_threads(1);
#endif
prop.fProp(ngram.col(0));
int output = ngram(m->ngram_size-1, 0);
double log_prob;
start_timer(3);
if (normalization)
{
Eigen::Matrix<double,Eigen::Dynamic,1> scores(m->output_vocab_size);
if (prop.skip_hidden)
prop.output_layer_node.param->fProp(prop.first_hidden_activation_node.fProp_matrix, scores);
else
prop.output_layer_node.param->fProp(prop.second_hidden_activation_node.fProp_matrix, scores);
double logz = logsum(scores.col(0));
log_prob = weight * (scores(output, 0) - logz);
}
else
{
if (prop.skip_hidden)
log_prob = weight * prop.output_layer_node.param->fProp(prop.first_hidden_activation_node.fProp_matrix, output, 0);
else
log_prob = weight * prop.output_layer_node.param->fProp(prop.second_hidden_activation_node.fProp_matrix, output, 0);
}
stop_timer(3);
if (cache_size)
{
// Update cache
cache_keys.col(hash) = ngram;
cache_values[hash] = log_prob;
}
#ifdef __INTEL_MKL__
mkl_set_num_threads(save_mkl_threads);
#endif
Eigen::setNbThreads(save_eigen_threads);
omp_set_num_threads(save_threads);
return log_prob;
}
// Look up many n-grams in parallel.
template <typename DerivedA, typename DerivedB>
void lookup_ngram(const Eigen::MatrixBase<DerivedA> &ngram, const Eigen::MatrixBase<DerivedB> &log_probs_const)
{
UNCONST(DerivedB, log_probs_const, log_probs);
assert (ngram.rows() == m->ngram_size);
//assert (ngram.cols() <= prop.get_minibatch_size());
prop.fProp(ngram);
if (normalization)
{
Eigen::Matrix<double,Eigen::Dynamic,Eigen::Dynamic> scores(m->output_vocab_size, ngram.cols());
if (prop.skip_hidden)
prop.output_layer_node.param->fProp(prop.first_hidden_activation_node.fProp_matrix, scores);
else
prop.output_layer_node.param->fProp(prop.second_hidden_activation_node.fProp_matrix, scores);
// And softmax and loss
Matrix<double,Dynamic,Dynamic> output_probs(m->output_vocab_size, ngram.cols());
double minibatch_log_likelihood;
SoftmaxLogLoss().fProp(scores.leftCols(ngram.cols()), ngram.row(m->ngram_size-1), output_probs, minibatch_log_likelihood);
for (int j=0; j<ngram.cols(); j++)
{
int output = ngram(m->ngram_size-1, j);
log_probs(0, j) = weight * output_probs(output, j);
}
}
else
{
for (int j=0; j<ngram.cols(); j++)
{
int output = ngram(m->ngram_size-1, j);
if (prop.skip_hidden)
log_probs(0, j) = weight * prop.output_layer_node.param->fProp(prop.first_hidden_activation_node.fProp_matrix, output, j);
else
log_probs(0, j) = weight * prop.output_layer_node.param->fProp(prop.second_hidden_activation_node.fProp_matrix, output, j);
}
}
}
int get_order() const { return m->ngram_size; }
void read(const std::string &filename)
{
m->read(filename);
resize();
// this is faster but takes more memory
//m->premultiply();
}
void set_cache(std::size_t cache_size)
{
this->cache_size = cache_size;
cache_keys.resize(m->ngram_size, cache_size);
cache_keys.fill(-1); // clears cache
cache_values.resize(cache_size);
cache_lookups = cache_hits = 0;
}
double cache_hit_rate()
{
return static_cast<double>(cache_hits)/cache_lookups;
}
void premultiply()
{
if (!m->premultiplied)
{
m->premultiply();
}
}
};
} // namespace nplm
#endif
|
