-
Notifications
You must be signed in to change notification settings - Fork 1
/
manager.cc
318 lines (300 loc) · 10.8 KB
/
manager.cc
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
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
/* -*- C++ -*-
*
* Copyright (c) 2014
* Spoken Language Systems Group
* MIT Computer Science and Artificial Intelligence Laboratory
* Massachusetts Institute of Technology
*
* All Rights Reserved
./manager.cc
* FILE: cluster.cc *
* *
* *
* Chia-ying (Jackie) Lee <chiaying@csail.mit.edu> *
* Feb 2014 *
*********************************************************************/
#include <fstream>
#include <iostream>
#include <sstream>
#include <cstdlib>
#include <cmath>
#include <set>
#include <omp.h>
#include <ctime>
#include "cluster.h"
#include "gmm.h"
#include "manager.h"
#include "counter.h"
#include "sampler.h"
Manager::Manager(Config* config) {
_config = config;
_model = new Model(_config);
_total_frame_num = 0;
}
bool Manager::LoadSilenceModel(string& fn_sil) {
ifstream fsil(fn_sil.c_str(), ios::binary);
if (!fsil.is_open()) {
return false;
}
// load silence model
int num_state = _config -> num_sil_states();
int num_mixture = _config -> num_sil_mix();
Cluster* sil_cluster = new Cluster(_config, num_state, num_mixture, 0);
vector<vector<float> > trans_probs;
for (int i = 0; i < num_state; ++i) {
vector<float> trans_prob(num_state + 1, 0);
trans_prob[0]= log(_config -> sil_self_trans_prob());
trans_prob[1] = log(1 - _config -> sil_self_trans_prob());
trans_probs.push_back(trans_prob);
}
sil_cluster -> set_transition_probs(trans_probs);
for (int i = 0 ; i < num_state; ++i) {
vector<float> weights;
for (int m = 0; m < num_mixture; ++m) {
float weight;
fsil.read(reinterpret_cast<char*> (&weight), sizeof(float));
weights.push_back(weight);
vector<float> mean(_config -> dim(), 0);
vector<float> pre(_config -> dim(), 0);
fsil.read(reinterpret_cast<char*> (&mean[0]), sizeof(float) * _config -> dim());
fsil.read(reinterpret_cast<char*> (&pre[0]), sizeof(float) * _config -> dim());
(sil_cluster -> emission(i)).mixture(m).set_mean(mean);
(sil_cluster -> emission(i)).mixture(m).set_pre(pre);
(sil_cluster -> emission(i)).mixture(m).set_det();
}
(sil_cluster -> emission(i)).set_weight(weights);
}
fsil.close();
sil_cluster -> set_is_fixed(true);
_model -> AddSilenceCluster(sil_cluster);
return true;
}
void Manager::InitializeModel() {
_model -> Initialize();
cout << "There are " << (_model -> clusters()).size() << " clusters." << endl;
}
void Manager::InitializeModel(const string& fn_snapshot) {
_model -> LoadSnapshot(fn_snapshot);
}
void Manager::ParallelInference(int batch_size, int n_iter, const string& basedir) {
cout << "Doing parallel inference" << endl;
Counter global_counter(_config);
Sampler global_sampler(_config);
// Initialization
global_sampler.SampleModelParams(_model);
for (int i = 0; i <= n_iter; ++i) {
time_t start_time = time(NULL);
vector<Datum*> batch = i < 10 ? _datum : randomizer.GetBatch(batch_size);
LoadDataIntoMatrix(batch);
if (_config -> precompute()) {
_model -> PreCompute(&_features[0], _total_frame_num);
}
omp_set_num_threads(24);
#pragma omp parallel
{
Sampler local_sampler(_config);
Counter local_counter(_config);
#pragma omp for schedule(dynamic, 1)
for (vector<Datum*>::iterator d_iter = batch.begin(); \
d_iter < batch.end(); ++d_iter) {
if (!((*d_iter) -> is_corrupted())) {
local_sampler.SampleSegments((*d_iter), _model, &local_counter);
}
else {
cout << "!!!!!!!!!! " << (*d_iter) -> tag() << " is corrupted." << endl;
}
}
#pragma omp critical
global_counter += local_counter;
}
global_sampler.SampleModelParams(_model, &global_counter);
if (i % 100 == 0) {
stringstream n;
n << i;
string output_dir = basedir + "/" + n.str() + "/";
SaveData(output_dir);
SaveModel(output_dir);
}
cout << "Done the " << i << "th iteration." << endl;
time_t end_time = time(NULL);
cout << "It took " << end_time - start_time << " seconds to finish one iteration" << endl;
}
}
void Manager::SerielInference(int batch_size, int n_iter, const string& basedir) {
Counter global_counter(_config);
Sampler global_sampler(_config);
// Initialization
cout << "Sampling global model" << endl;
global_sampler.SampleModelParams(_model);
for (int i = 0; i <= n_iter; ++i) {
time_t start_time = time(NULL);
vector<Datum*> batch = i < 10 ? _datum : randomizer.GetBatch(batch_size);
LoadDataIntoMatrix(batch);
if (_config -> precompute()) {
cout << "precomputing" << endl;
_model -> PreCompute(&_features[0], _total_frame_num);
}
for (vector<Datum*>::iterator d_iter = batch.begin(); \
d_iter != batch.end(); ++d_iter) {
if (!((*d_iter) -> is_corrupted())) {
if (i == 0) {
cout << "Inferencing " << (*d_iter) -> tag() << endl;
}
global_sampler.SampleSegments((*d_iter), _model, &global_counter);
}
else {
cout << "!!!!!!!!!! " << (*d_iter) -> tag() << " is corrupted." << endl;
}
}
cout << "Sampling model" << endl;
global_sampler.SampleModelParams(_model, &global_counter);
if (i % 1000 == 0) {
stringstream n;
n << i;
string output_dir = basedir + "/" + n.str() + "/";
SaveData(output_dir);
SaveModel(output_dir);
}
cout << "Done the " << i << "th iteration." << endl;
time_t end_time = time(NULL);
cout << "It took " << end_time - start_time << " seconds to finish one iteration" << endl;
}
}
void Manager::SaveModel(const string& output_dir) {
string path = output_dir + "snapshot";
_model -> Save(path);
}
void Manager::SaveData(const string& output_dir) {
#pragma omp parallel
{
#pragma omp for schedule (dynamic, 1)
for (int d = 0; d < (int) _datum.size(); ++d) {
_datum[d] -> Save(output_dir);
}
}
}
void Manager::Inference(int batch_size, int n_iter, const string& basedir) {
if (_config -> parallel()) {
ParallelInference(batch_size, n_iter, basedir);
}
else {
SerielInference(batch_size, n_iter, basedir);
}
}
string Manager::GetTag(const string& s) {
size_t found_last_slash, found_last_period;
found_last_slash = s.find_last_of("/");
found_last_period = s.find_last_of(".");
return s.substr(found_last_slash + 1, \
found_last_period - 1 - found_last_slash);
}
bool Manager::LoadData(string& fn_data_list) {
ifstream flist(fn_data_list.c_str());
while (flist.good()) {
string fn_index;
string fn_data;
flist >> fn_index;
flist >> fn_data;
if (fn_index != "" && fn_data != "") {
string tag = GetTag(fn_data);
ifstream findex(fn_index.c_str());
ifstream fdata(fn_data.c_str(), ios::binary);
if (!findex.is_open()) {
cout << "Cannot open " << fn_index << endl;
return false;
}
if (!fdata.is_open()) {
cout << "Cannot open " << fn_data << endl;
return false;
}
cout << "loading " << fn_data << endl;
Datum* datum = new Datum(_config);
datum -> set_tag(tag);
LoadBounds(datum, findex, fdata);
findex.close();
fdata.close();
_datum.push_back(datum);
}
}
flist.close();
randomizer.LoadData(_datum);
return true;
}
void Manager::LoadBounds(Datum* datum, ifstream& findex, ifstream& fdata) {
vector<Bound*> bounds;
int total_frame;
findex >> total_frame;
int start_frame = 0;
int end_frame = 0;
int label = -2; // >= 0: is labeled, is boundary. == -1: is boundary. == -2 : normal
while (end_frame != total_frame - 1) {
findex >> start_frame;
findex >> end_frame;
findex >> label;
if (start_frame > end_frame) {
datum -> set_corrupted(true);
break;
}
int frame_num = end_frame - start_frame + 1;
Bound* bound = new Bound(_config);
// float** data is deleted in bound.cc
float** data = new float* [frame_num];
for (int i = 0; i < frame_num; ++i) {
data[i] = new float [_config -> dim()];
fdata.read(reinterpret_cast<char*> (data[i]), \
sizeof(float) * _config -> dim());
}
bound -> set_data(data, frame_num);
if (label >= 0) {
if (bounds.size() > 0) {
bounds[bounds.size() - 1] -> set_is_boundary(true);
}
bound -> set_is_labeled(true);
bound -> set_is_boundary(true);
bound -> set_label(label);
}
if (label == -1) {
bound -> set_is_boundary(true);
}
bound -> set_start_frame(start_frame + _total_frame_num);
bound -> set_end_frame(end_frame + _total_frame_num);
bounds.push_back(bound);
}
if (datum -> is_corrupted()) {
vector<Bound*>::iterator b_iter = bounds.begin();
for (; b_iter != bounds.end(); ++b_iter) {
delete *b_iter;
}
}
else {
_total_frame_num += total_frame;
bounds[bounds.size() - 1] -> set_is_boundary(true);
datum -> set_bounds(bounds);
}
}
void Manager::LoadDataIntoMatrix(vector<Datum*>& batch) {
_features.clear();
_total_frame_num = 0;
vector<Datum*>::iterator d_iter = batch.begin();
for (; d_iter != batch.end(); ++d_iter) {
vector<Bound*> bounds = (*d_iter) -> bounds();
vector<Bound*>::iterator b_iter = bounds.begin();
for (; b_iter != bounds.end(); ++b_iter) {
vector<float*> f = (*b_iter) -> data();
for (unsigned int i = 0; i < f.size(); ++i) {
_features.push_back(f[i]);
}
(*b_iter) -> set_start_frame(_total_frame_num);
(*b_iter) -> set_end_frame(_total_frame_num + (*b_iter) -> frame_num() - 1);
_total_frame_num += (*b_iter) -> frame_num();
}
}
}
Manager::~Manager() {
vector<Datum*>::iterator d_iter = _datum.begin();
for (; d_iter != _datum.end(); ++d_iter) {
delete *d_iter;
}
_datum.clear();
delete _model;
}