offline-nemo-enc-dec-ctc-model.cc
4.0 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
// sherpa-onnx/csrc/offline-nemo-enc-dec-ctc-model.cc
//
// Copyright (c) 2023 Xiaomi Corporation
#include "sherpa-onnx/csrc/offline-nemo-enc-dec-ctc-model.h"
#include "sherpa-onnx/csrc/macros.h"
#include "sherpa-onnx/csrc/onnx-utils.h"
#include "sherpa-onnx/csrc/session.h"
#include "sherpa-onnx/csrc/text-utils.h"
#include "sherpa-onnx/csrc/transpose.h"
namespace sherpa_onnx {
class OfflineNemoEncDecCtcModel::Impl {
public:
explicit Impl(const OfflineModelConfig &config)
: config_(config),
env_(ORT_LOGGING_LEVEL_ERROR),
sess_opts_(GetSessionOptions(config)),
allocator_{} {
Init();
}
std::pair<Ort::Value, Ort::Value> Forward(Ort::Value features,
Ort::Value features_length) {
std::vector<int64_t> shape =
features_length.GetTensorTypeAndShapeInfo().GetShape();
Ort::Value out_features_length = Ort::Value::CreateTensor<int64_t>(
allocator_, shape.data(), shape.size());
const int64_t *src = features_length.GetTensorData<int64_t>();
int64_t *dst = out_features_length.GetTensorMutableData<int64_t>();
for (int64_t i = 0; i != shape[0]; ++i) {
dst[i] = src[i] / subsampling_factor_;
}
// (B, T, C) -> (B, C, T)
features = Transpose12(allocator_, &features);
std::array<Ort::Value, 2> inputs = {std::move(features),
std::move(features_length)};
auto out =
sess_->Run({}, input_names_ptr_.data(), inputs.data(), inputs.size(),
output_names_ptr_.data(), output_names_ptr_.size());
return {std::move(out[0]), std::move(out_features_length)};
}
int32_t VocabSize() const { return vocab_size_; }
int32_t SubsamplingFactor() const { return subsampling_factor_; }
OrtAllocator *Allocator() const { return allocator_; }
std::string FeatureNormalizationMethod() const { return normalize_type_; }
private:
void Init() {
auto buf = ReadFile(config_.nemo_ctc.model);
sess_ = std::make_unique<Ort::Session>(env_, buf.data(), buf.size(),
sess_opts_);
GetInputNames(sess_.get(), &input_names_, &input_names_ptr_);
GetOutputNames(sess_.get(), &output_names_, &output_names_ptr_);
// get meta data
Ort::ModelMetadata meta_data = sess_->GetModelMetadata();
if (config_.debug) {
std::ostringstream os;
PrintModelMetadata(os, meta_data);
SHERPA_ONNX_LOGE("%s\n", os.str().c_str());
}
Ort::AllocatorWithDefaultOptions allocator; // used in the macro below
SHERPA_ONNX_READ_META_DATA(vocab_size_, "vocab_size");
SHERPA_ONNX_READ_META_DATA(subsampling_factor_, "subsampling_factor");
SHERPA_ONNX_READ_META_DATA_STR(normalize_type_, "normalize_type");
}
private:
OfflineModelConfig config_;
Ort::Env env_;
Ort::SessionOptions sess_opts_;
Ort::AllocatorWithDefaultOptions allocator_;
std::unique_ptr<Ort::Session> sess_;
std::vector<std::string> input_names_;
std::vector<const char *> input_names_ptr_;
std::vector<std::string> output_names_;
std::vector<const char *> output_names_ptr_;
int32_t vocab_size_ = 0;
int32_t subsampling_factor_ = 0;
std::string normalize_type_;
};
OfflineNemoEncDecCtcModel::OfflineNemoEncDecCtcModel(
const OfflineModelConfig &config)
: impl_(std::make_unique<Impl>(config)) {}
OfflineNemoEncDecCtcModel::~OfflineNemoEncDecCtcModel() = default;
std::pair<Ort::Value, Ort::Value> OfflineNemoEncDecCtcModel::Forward(
Ort::Value features, Ort::Value features_length) {
return impl_->Forward(std::move(features), std::move(features_length));
}
int32_t OfflineNemoEncDecCtcModel::VocabSize() const {
return impl_->VocabSize();
}
int32_t OfflineNemoEncDecCtcModel::SubsamplingFactor() const {
return impl_->SubsamplingFactor();
}
OrtAllocator *OfflineNemoEncDecCtcModel::Allocator() const {
return impl_->Allocator();
}
std::string OfflineNemoEncDecCtcModel::FeatureNormalizationMethod() const {
return impl_->FeatureNormalizationMethod();
}
} // namespace sherpa_onnx