online-transducer-greedy-search-decoder.cc
5.7 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
// sherpa-onnx/csrc/online-transducer-greedy-search-decoder.cc
//
// Copyright (c) 2023 Xiaomi Corporation
#include "sherpa-onnx/csrc/online-transducer-greedy-search-decoder.h"
#include <algorithm>
#include <utility>
#include <vector>
#include "sherpa-onnx/csrc/macros.h"
#include "sherpa-onnx/csrc/onnx-utils.h"
namespace sherpa_onnx {
static void UseCachedDecoderOut(
const std::vector<OnlineTransducerDecoderResult> &results,
Ort::Value *decoder_out) {
std::vector<int64_t> shape =
decoder_out->GetTensorTypeAndShapeInfo().GetShape();
float *dst = decoder_out->GetTensorMutableData<float>();
for (const auto &r : results) {
if (r.decoder_out) {
const float *src = r.decoder_out.GetTensorData<float>();
std::copy(src, src + shape[1], dst);
}
dst += shape[1];
}
}
static void UpdateCachedDecoderOut(
OrtAllocator *allocator, const Ort::Value *decoder_out,
std::vector<OnlineTransducerDecoderResult> *results) {
std::vector<int64_t> shape =
decoder_out->GetTensorTypeAndShapeInfo().GetShape();
auto memory_info =
Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
std::array<int64_t, 2> v_shape{1, shape[1]};
const float *src = decoder_out->GetTensorData<float>();
for (auto &r : *results) {
if (!r.decoder_out) {
r.decoder_out = Ort::Value::CreateTensor<float>(allocator, v_shape.data(),
v_shape.size());
}
float *dst = r.decoder_out.GetTensorMutableData<float>();
std::copy(src, src + shape[1], dst);
src += shape[1];
}
}
OnlineTransducerDecoderResult
OnlineTransducerGreedySearchDecoder::GetEmptyResult() const {
int32_t context_size = model_->ContextSize();
int32_t blank_id = 0; // always 0
OnlineTransducerDecoderResult r;
r.tokens.resize(context_size, -1);
r.tokens.back() = blank_id;
return r;
}
void OnlineTransducerGreedySearchDecoder::StripLeadingBlanks(
OnlineTransducerDecoderResult *r) const {
int32_t context_size = model_->ContextSize();
auto start = r->tokens.begin() + context_size;
auto end = r->tokens.end();
r->tokens = std::vector<int64_t>(start, end);
}
void OnlineTransducerGreedySearchDecoder::Decode(
Ort::Value encoder_out,
std::vector<OnlineTransducerDecoderResult> *result) {
std::vector<int64_t> encoder_out_shape =
encoder_out.GetTensorTypeAndShapeInfo().GetShape();
if (encoder_out_shape[0] != result->size()) {
SHERPA_ONNX_LOGE(
"Size mismatch! encoder_out.size(0) %d, result.size(0): %d",
static_cast<int32_t>(encoder_out_shape[0]),
static_cast<int32_t>(result->size()));
exit(-1);
}
int32_t batch_size = static_cast<int32_t>(encoder_out_shape[0]);
int32_t num_frames = static_cast<int32_t>(encoder_out_shape[1]);
int32_t vocab_size = model_->VocabSize();
Ort::Value decoder_out{nullptr};
bool is_batch_decoder_out_cached = true;
for (const auto &r : *result) {
if (!r.decoder_out) {
is_batch_decoder_out_cached = false;
break;
}
}
if (is_batch_decoder_out_cached) {
auto &r = result->front();
std::vector<int64_t> decoder_out_shape =
r.decoder_out.GetTensorTypeAndShapeInfo().GetShape();
decoder_out_shape[0] = batch_size;
decoder_out = Ort::Value::CreateTensor<float>(model_->Allocator(),
decoder_out_shape.data(),
decoder_out_shape.size());
UseCachedDecoderOut(*result, &decoder_out);
} else {
Ort::Value decoder_input = model_->BuildDecoderInput(*result);
decoder_out = model_->RunDecoder(std::move(decoder_input));
}
for (int32_t t = 0; t != num_frames; ++t) {
Ort::Value cur_encoder_out =
GetEncoderOutFrame(model_->Allocator(), &encoder_out, t);
Ort::Value logit =
model_->RunJoiner(std::move(cur_encoder_out), View(&decoder_out));
float *p_logit = logit.GetTensorMutableData<float>();
bool emitted = false;
for (int32_t i = 0; i < batch_size; ++i, p_logit += vocab_size) {
auto &r = (*result)[i];
if (blank_penalty_ > 0.0) {
p_logit[0] -= blank_penalty_; // assuming blank id is 0
}
auto y = static_cast<int32_t>(std::distance(
static_cast<const float *>(p_logit),
std::max_element(static_cast<const float *>(p_logit),
static_cast<const float *>(p_logit) + vocab_size)));
// blank id is hardcoded to 0
// also, it treats unk as blank
if (y != 0 && y != unk_id_) {
emitted = true;
r.tokens.push_back(y);
r.timestamps.push_back(t + r.frame_offset);
r.num_trailing_blanks = 0;
} else {
++r.num_trailing_blanks;
}
// export the per-token log scores
if (y != 0 && y != unk_id_) {
// apply temperature-scaling
for (int32_t n = 0; n < vocab_size; ++n) {
p_logit[n] /= temperature_scale_;
}
LogSoftmax(p_logit, vocab_size); // renormalize probabilities,
// save time by doing it only for
// emitted symbols
const float *p_logprob = p_logit; // rename p_logit as p_logprob,
// now it contains normalized
// probability
r.ys_probs.push_back(p_logprob[y]);
}
}
if (emitted) {
Ort::Value decoder_input = model_->BuildDecoderInput(*result);
decoder_out = model_->RunDecoder(std::move(decoder_input));
}
}
UpdateCachedDecoderOut(model_->Allocator(), &decoder_out, result);
// Update frame_offset
for (auto &r : *result) {
r.frame_offset += num_frames;
}
}
} // namespace sherpa_onnx