offline-fire-red-asr-greedy-search-decoder.cc
3.2 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
// sherpa-onnx/csrc/offline-fire-red-asr-greedy-search-decoder.cc
//
// Copyright (c) 2025 Xiaomi Corporation
#include "sherpa-onnx/csrc/offline-fire-red-asr-greedy-search-decoder.h"
#include <algorithm>
#include <tuple>
#include <utility>
#include "sherpa-onnx/csrc/macros.h"
#include "sherpa-onnx/csrc/onnx-utils.h"
namespace sherpa_onnx {
// Note: this functions works only for batch size == 1 at present
std::vector<OfflineFireRedAsrDecoderResult>
OfflineFireRedAsrGreedySearchDecoder::Decode(Ort::Value cross_k,
Ort::Value cross_v,
int32_t num_feature_frames) {
const auto &meta_data = model_->GetModelMetadata();
auto memory_info =
Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
// For multilingual models, initial_tokens contains [sot, language, task]
// - language is English by default
// - task is transcribe by default
//
// For non-multilingual models, initial_tokens contains [sot]
std::array<int64_t, 2> token_shape = {1, 1};
int64_t token = meta_data.sos_id;
int32_t batch_size = 1;
Ort::Value tokens = Ort::Value::CreateTensor(
memory_info, &token, 1, token_shape.data(), token_shape.size());
std::array<int64_t, 1> offset_shape{1};
Ort::Value offset = Ort::Value::CreateTensor<int64_t>(
model_->Allocator(), offset_shape.data(), offset_shape.size());
*(offset.GetTensorMutableData<int64_t>()) = 0;
std::vector<OfflineFireRedAsrDecoderResult> ans(1);
auto self_kv_cache = model_->GetInitialSelfKVCache();
std::tuple<Ort::Value, Ort::Value, Ort::Value, Ort::Value, Ort::Value,
Ort::Value>
decoder_out = {Ort::Value{nullptr},
std::move(self_kv_cache.first),
std::move(self_kv_cache.second),
std::move(cross_k),
std::move(cross_v),
std::move(offset)};
// assume at most 6 tokens per second
int32_t num_possible_tokens = num_feature_frames / 100.0 * 6;
num_possible_tokens =
std::min<int32_t>(num_possible_tokens, meta_data.max_len / 2);
for (int32_t i = 0; i < num_possible_tokens; ++i) {
decoder_out = model_->ForwardDecoder(View(&tokens),
std::move(std::get<1>(decoder_out)),
std::move(std::get<2>(decoder_out)),
std::move(std::get<3>(decoder_out)),
std::move(std::get<4>(decoder_out)),
std::move(std::get<5>(decoder_out)));
const auto &logits = std::get<0>(decoder_out);
const float *p_logits = logits.GetTensorData<float>();
auto logits_shape = logits.GetTensorTypeAndShapeInfo().GetShape();
int32_t vocab_size = logits_shape[2];
int32_t max_token_id = static_cast<int32_t>(std::distance(
p_logits, std::max_element(p_logits, p_logits + vocab_size)));
if (max_token_id == meta_data.eos_id) {
break;
}
ans[0].tokens.push_back(max_token_id);
token = max_token_id;
// increment offset
*(std::get<5>(decoder_out).GetTensorMutableData<int64_t>()) += 1;
}
return ans;
}
} // namespace sherpa_onnx