discriminative-supervision.cc
16.4 KB
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
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
// nnet3/discriminative-supervision.cc
// Copyright 2012-2015 Johns Hopkins University (author: Daniel Povey)
// 2014-2015 Vimal Manohar
// See ../../COPYING for clarification regarding multiple authors
//
// 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
//
// THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
// KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
// WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
// MERCHANTABLITY OR NON-INFRINGEMENT.
// See the Apache 2 License for the specific language governing permissions and
// limitations under the License.
#include "nnet3/discriminative-supervision.h"
#include "lat/lattice-functions.h"
namespace kaldi {
namespace discriminative {
DiscriminativeSupervision::DiscriminativeSupervision(
const DiscriminativeSupervision &other):
weight(other.weight), num_sequences(other.num_sequences),
frames_per_sequence(other.frames_per_sequence),
num_ali(other.num_ali), den_lat(other.den_lat) { }
void DiscriminativeSupervision::Swap(DiscriminativeSupervision *other) {
std::swap(weight, other->weight);
std::swap(num_sequences, other->num_sequences);
std::swap(frames_per_sequence, other->frames_per_sequence);
std::swap(num_ali, other->num_ali);
std::swap(den_lat, other->den_lat);
}
bool DiscriminativeSupervision::operator == (
const DiscriminativeSupervision &other) const {
return ( weight == other.weight &&
num_sequences == other.num_sequences &&
frames_per_sequence == other.frames_per_sequence &&
num_ali == other.num_ali &&
fst::Equal(den_lat, other.den_lat) );
}
void DiscriminativeSupervision::Write(std::ostream &os, bool binary) const {
WriteToken(os, binary, "<DiscriminativeSupervision>");
WriteToken(os, binary, "<Weight>");
WriteBasicType(os, binary, weight);
WriteToken(os, binary, "<NumSequences>");
WriteBasicType(os, binary, num_sequences);
WriteToken(os, binary, "<FramesPerSeq>");
WriteBasicType(os, binary, frames_per_sequence);
KALDI_ASSERT(frames_per_sequence > 0 &&
num_sequences > 0);
WriteToken(os, binary, "<NumAli>");
WriteIntegerVector(os, binary, num_ali);
WriteToken(os, binary, "<DenLat>");
if (!WriteLattice(os, binary, den_lat)) {
// We can't return error status from this function so we
// throw an exception.
KALDI_ERR << "Error writing denominator lattice to stream";
}
WriteToken(os, binary, "</DiscriminativeSupervision>");
}
void DiscriminativeSupervision::Read(std::istream &is, bool binary) {
ExpectToken(is, binary, "<DiscriminativeSupervision>");
ExpectToken(is, binary, "<Weight>");
ReadBasicType(is, binary, &weight);
ExpectToken(is, binary, "<NumSequences>");
ReadBasicType(is, binary, &num_sequences);
ExpectToken(is, binary, "<FramesPerSeq>");
ReadBasicType(is, binary, &frames_per_sequence);
KALDI_ASSERT(frames_per_sequence > 0 &&
num_sequences > 0);
ExpectToken(is, binary, "<NumAli>");
ReadIntegerVector(is, binary, &num_ali);
ExpectToken(is, binary, "<DenLat>");
{
Lattice *lat = NULL;
if (!ReadLattice(is, binary, &lat) || lat == NULL) {
// We can't return error status from this function so we
// throw an exception.
KALDI_ERR << "Error reading Lattice from stream";
}
den_lat = *lat;
delete lat;
TopSort(&den_lat);
}
ExpectToken(is, binary, "</DiscriminativeSupervision>");
}
bool DiscriminativeSupervision::Initialize(const std::vector<int32> &num_ali,
const Lattice &den_lat,
BaseFloat weight) {
if (num_ali.size() == 0) return false;
if (den_lat.NumStates() == 0) return false;
this->weight = weight;
this->num_sequences = 1;
this->frames_per_sequence = num_ali.size();
this->num_ali = num_ali;
this->den_lat = den_lat;
KALDI_ASSERT(TopSort(&(this->den_lat)));
// Checks if num frames in alignment matches lattice
Check();
return true;
}
void DiscriminativeSupervision::Check() const {
int32 num_frames_subsampled = num_ali.size();
KALDI_ASSERT(num_frames_subsampled ==
num_sequences * frames_per_sequence);
{
std::vector<int32> state_times;
int32 max_time = LatticeStateTimes(den_lat, &state_times);
KALDI_ASSERT(max_time == num_frames_subsampled);
}
}
DiscriminativeSupervisionSplitter::DiscriminativeSupervisionSplitter(
const SplitDiscriminativeSupervisionOptions &config,
const TransitionModel &tmodel,
const DiscriminativeSupervision &supervision):
config_(config), tmodel_(tmodel), supervision_(supervision) {
if (supervision_.num_sequences != 1) {
KALDI_WARN << "Splitting already-reattached sequence (only expected in "
<< "testing code)";
}
KALDI_ASSERT(supervision_.num_sequences == 1); // For now, don't allow splitting already merged examples
den_lat_ = supervision_.den_lat;
PrepareLattice(&den_lat_, &den_lat_scores_);
int32 num_states = den_lat_.NumStates(),
num_frames = supervision_.frames_per_sequence * supervision_.num_sequences;
KALDI_ASSERT(num_states > 0);
int32 start_state = den_lat_.Start();
// Lattice should be top-sorted and connected, so start-state must be 0.
KALDI_ASSERT(start_state == 0 && "Expecting start-state to be 0");
KALDI_ASSERT(num_states == den_lat_scores_.state_times.size());
KALDI_ASSERT(den_lat_scores_.state_times[start_state] == 0);
KALDI_ASSERT(den_lat_scores_.state_times.back() == num_frames);
}
// Make sure that for any given pdf-id and any given frame, the den-lat has
// only one transition-id mapping to that pdf-id, on the same frame.
// It helps us to more completely minimize the lattice. Note: we
// can't do this if the criterion is MPFE, because in that case the
// objective function will be affected by the phone-identities being
// different even if the pdf-ids are the same.
void DiscriminativeSupervisionSplitter::CollapseTransitionIds(
const std::vector<int32> &state_times, Lattice *lat) const {
typedef Lattice::StateId StateId;
typedef Lattice::Arc Arc;
int32 num_frames = state_times.back(); // TODO: Check if this is always true
StateId num_states = lat->NumStates();
std::vector<std::map<int32, int32> > pdf_to_tid(num_frames);
for (StateId s = 0; s < num_states; s++) {
int32 t = state_times[s];
for (fst::MutableArcIterator<Lattice> aiter(lat, s);
!aiter.Done(); aiter.Next()) {
KALDI_ASSERT(t >= 0 && t < num_frames);
Arc arc = aiter.Value();
KALDI_ASSERT(arc.ilabel != 0 && arc.ilabel == arc.olabel);
int32 pdf = tmodel_.TransitionIdToPdf(arc.ilabel);
if (pdf_to_tid[t].count(pdf) != 0) {
arc.ilabel = arc.olabel = pdf_to_tid[t][pdf];
aiter.SetValue(arc);
} else {
pdf_to_tid[t][pdf] = arc.ilabel;
}
}
}
}
void DiscriminativeSupervisionSplitter::LatticeInfo::Check() const {
// Check if all the vectors are of size num_states
KALDI_ASSERT(state_times.size() == alpha.size() &&
state_times.size() == beta.size());
// Check that the states are ordered in increasing order of state_times.
// This must be true since the states are in breadth-first search order.
KALDI_ASSERT(IsSorted(state_times));
}
void DiscriminativeSupervisionSplitter::GetFrameRange(int32 begin_frame, int32 num_frames, bool normalize,
DiscriminativeSupervision *out_supervision) const {
int32 end_frame = begin_frame + num_frames;
// Note: end_frame is not included in the range of frames that the
// output supervision object covers; it's one past the end.
KALDI_ASSERT(num_frames > 0 && begin_frame >= 0 &&
begin_frame + num_frames <=
supervision_.num_sequences * supervision_.frames_per_sequence);
CreateRangeLattice(den_lat_,
den_lat_scores_,
begin_frame, end_frame, normalize,
&(out_supervision->den_lat));
out_supervision->num_ali.clear();
std::copy(supervision_.num_ali.begin() + begin_frame,
supervision_.num_ali.begin() + end_frame,
std::back_inserter(out_supervision->num_ali));
out_supervision->num_sequences = 1;
out_supervision->weight = supervision_.weight;
out_supervision->frames_per_sequence = num_frames;
out_supervision->Check();
}
void DiscriminativeSupervisionSplitter::CreateRangeLattice(
const Lattice &in_lat, const LatticeInfo &scores,
int32 begin_frame, int32 end_frame, bool normalize,
Lattice *out_lat) const {
typedef Lattice::StateId StateId;
const std::vector<int32> &state_times = scores.state_times;
// Some checks to ensure the lattice and scores are prepared properly
KALDI_ASSERT(state_times.size() == in_lat.NumStates());
if (!in_lat.Properties(fst::kTopSorted, true))
KALDI_ERR << "Input lattice must be topologically sorted.";
std::vector<int32>::const_iterator begin_iter =
std::lower_bound(state_times.begin(), state_times.end(), begin_frame),
end_iter = std::lower_bound(begin_iter,
state_times.end(), end_frame);
KALDI_ASSERT(*begin_iter == begin_frame &&
(begin_iter == state_times.begin() ||
begin_iter[-1] < begin_frame));
// even if end_frame == supervision_.num_frames, there should be a state with
// that frame index.
KALDI_ASSERT(end_iter[-1] < end_frame &&
(end_iter < state_times.end() || *end_iter == end_frame));
StateId begin_state = begin_iter - state_times.begin(),
end_state = end_iter - state_times.begin();
KALDI_ASSERT(end_state > begin_state);
out_lat->DeleteStates();
out_lat->ReserveStates(end_state - begin_state + 2);
// Add special start state
StateId start_state = out_lat->AddState();
out_lat->SetStart(start_state);
for (StateId i = begin_state; i < end_state; i++)
out_lat->AddState();
// Add the special final-state.
StateId final_state = out_lat->AddState();
out_lat->SetFinal(final_state, LatticeWeight::One());
for (StateId state = begin_state; state < end_state; state++) {
StateId output_state = state - begin_state + 1;
if (state_times[state] == begin_frame) {
// we'd like to make this an initial state, but OpenFst doesn't allow
// multiple initial states. Instead we add an epsilon transition to it
// from our actual initial state. The weight on this
// transition is the forward probability of the said 'initial state'
LatticeWeight weight = LatticeWeight::One();
weight.SetValue1((normalize ? scores.beta[0] : 0.0) - scores.alpha[state]);
// Add negative of the forward log-probability to the graph cost score,
// since the acoustic scores would be changed later.
// Assuming that the lattice is scaled with appropriate acoustic
// scale.
// We additionally normalize using the total lattice score. Since the
// same score is added as normalizer to all the paths in the lattice,
// the relative probabilities of the paths in the lattice is not affected.
// Note: Doing a forward-backward on this split must result in a total
// score of 0 because of the normalization.
out_lat->AddArc(start_state,
LatticeArc(0, 0, weight, output_state));
} else {
KALDI_ASSERT(scores.state_times[state] < end_frame);
}
for (fst::ArcIterator<Lattice> aiter(in_lat, state);
!aiter.Done(); aiter.Next()) {
const LatticeArc &arc = aiter.Value();
StateId nextstate = arc.nextstate;
if (nextstate >= end_state) {
// A transition to any state outside the range becomes a transition to
// our special final-state.
// The weight is just the negative of the backward log-probability +
// the arc cost. We again normalize with the total lattice score.
LatticeWeight weight;
//KALDI_ASSERT(scores.beta[state] < 0);
weight.SetValue1(arc.weight.Value1() - scores.beta[nextstate]);
weight.SetValue2(arc.weight.Value2());
// Add negative of the backward log-probability to the LM score, since
// the acoustic scores would be changed later.
// Note: We don't normalize here because that is already done with the
// initial cost.
out_lat->AddArc(output_state,
LatticeArc(arc.ilabel, arc.olabel, weight, final_state));
} else {
StateId output_nextstate = nextstate - begin_state + 1;
out_lat->AddArc(output_state,
LatticeArc(arc.ilabel, arc.olabel, arc.weight, output_nextstate));
}
}
}
// Get rid of the word labels and put the
// transition-ids on both sides.
fst::Project(out_lat, fst::PROJECT_INPUT);
fst::RmEpsilon(out_lat);
if (config_.collapse_transition_ids)
CollapseTransitionIds(state_times, out_lat);
if (config_.determinize) {
if (!config_.minimize) {
Lattice tmp_lat;
fst::Determinize(*out_lat, &tmp_lat);
std::swap(*out_lat, tmp_lat);
} else {
Lattice tmp_lat;
fst::Reverse(*out_lat, &tmp_lat);
fst::Determinize(tmp_lat, out_lat);
fst::Reverse(*out_lat, &tmp_lat);
fst::Determinize(tmp_lat, out_lat);
fst::RmEpsilon(out_lat);
}
}
fst::TopSort(out_lat);
std::vector<int32> state_times_tmp;
KALDI_ASSERT(LatticeStateTimes(*out_lat, &state_times_tmp) ==
end_frame - begin_frame);
// Remove the acoustic scale that was previously added
if (config_.acoustic_scale != 1.0) {
fst::ScaleLattice(fst::AcousticLatticeScale(
1 / config_.acoustic_scale), out_lat);
}
}
void DiscriminativeSupervisionSplitter::PrepareLattice(
Lattice *lat, LatticeInfo *scores) const {
// Scale the lattice to appropriate acoustic scale. It is important to
// ensure this is equal to the acoustic scale used while training. This is
// because, on splitting lattices, the initial and final costs are added
// into the graph cost.
KALDI_ASSERT(config_.acoustic_scale != 0.0);
if (config_.acoustic_scale != 1.0)
fst::ScaleLattice(fst::AcousticLatticeScale(
config_.acoustic_scale), lat);
LatticeStateTimes(*lat, &(scores->state_times));
int32 num_states = lat->NumStates();
std::vector<std::pair<int32,int32> > state_time_indexes(num_states);
for (int32 s = 0; s < num_states; s++) {
state_time_indexes[s] = std::make_pair(scores->state_times[s], s);
}
// Order the states based on the state times. This is stronger than just
// topological sort. This is required by the lattice splitting code.
std::sort(state_time_indexes.begin(), state_time_indexes.end());
std::vector<int32> state_order(num_states);
for (int32 s = 0; s < num_states; s++) {
state_order[state_time_indexes[s].second] = s;
}
fst::StateSort(lat, state_order);
ComputeLatticeScores(*lat, scores);
}
void DiscriminativeSupervisionSplitter::ComputeLatticeScores(const Lattice &lat,
LatticeInfo *scores) const {
LatticeStateTimes(lat, &(scores->state_times));
ComputeLatticeAlphasAndBetas(lat, false,
&(scores->alpha), &(scores->beta));
scores->Check();
// This check will fail if the lattice is not breadth-first search sorted
}
void MergeSupervision(const std::vector<const DiscriminativeSupervision*> &input,
DiscriminativeSupervision *output_supervision) {
KALDI_ASSERT(!input.empty());
int32 num_inputs = input.size();
if (num_inputs == 1) {
*output_supervision = *(input[0]);
return;
}
*output_supervision = *(input[num_inputs-1]);
for (int32 i = num_inputs - 2; i >= 0; i--) {
const DiscriminativeSupervision &src = *(input[i]);
KALDI_ASSERT(src.num_sequences == 1);
if (output_supervision->weight == src.weight &&
output_supervision->frames_per_sequence ==
src.frames_per_sequence) {
// Combine with current output
// append src.den_lat to output_supervision->den_lat.
fst::Concat(src.den_lat, &output_supervision->den_lat);
output_supervision->num_ali.insert(
output_supervision->num_ali.begin(),
src.num_ali.begin(), src.num_ali.end());
output_supervision->num_sequences++;
} else {
KALDI_ERR << "Mismatch weight or frames_per_sequence between inputs";
}
}
DiscriminativeSupervision &out_sup = *output_supervision;
fst::TopSort(&(out_sup.den_lat));
out_sup.Check();
}
} // namespace discriminative
} // namespace kaldi