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
|
//
// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#pragma once
#include "Types.hpp"
#include "armnn/ArmNN.hpp"
#include <armnn/Logging.hpp>
#include <armnn_delegate.hpp>
#include <DelegateOptions.hpp>
#include <DelegateUtils.hpp>
#include <Profiling.hpp>
#include <tensorflow/lite/builtin_ops.h>
#include <tensorflow/lite/c/builtin_op_data.h>
#include <tensorflow/lite/c/common.h>
#include <tensorflow/lite/optional_debug_tools.h>
#include <tensorflow/lite/kernels/builtin_op_kernels.h>
#include <tensorflow/lite/interpreter.h>
#include <tensorflow/lite/kernels/register.h>
#include <string>
#include <vector>
namespace common
{
/**
* @brief Used to load in a network through Tflite Interpreter,
* register Armnn Delegate file to it, and run inference
* on it against a given backend.
* currently it is assumed that the input data will be
* cv:MAT (Frame), the assumption is implemented in
* PrepareTensors method, it can be generalized later
*
*/
template <typename Tout>
class ArmnnNetworkExecutor
{
private:
std::unique_ptr<tflite::Interpreter> m_interpreter;
std::unique_ptr<tflite::FlatBufferModel> m_model;
Profiling m_profiling;
void PrepareTensors(const void* inputData, const size_t dataBytes);
template <typename Enumeration>
auto log_as_int(Enumeration value)
-> typename std::underlying_type<Enumeration>::type
{
return static_cast<typename std::underlying_type<Enumeration>::type>(value);
}
public:
ArmnnNetworkExecutor() = delete;
/**
* @brief Initializes the network with the given input data.
*
*
* * @param[in] modelPath - Relative path to the model file
* * @param[in] backends - The list of preferred backends to run inference on
*/
ArmnnNetworkExecutor(std::string& modelPath,
std::vector<armnn::BackendId>& backends,
bool isProfilingEnabled = false);
/**
* @brief Returns the aspect ratio of the associated model in the order of width, height.
*/
Size GetImageAspectRatio();
/**
* @brief Returns the data type of the associated model.
*/
armnn::DataType GetInputDataType() const;
float GetQuantizationScale();
int GetQuantizationOffset();
float GetOutputQuantizationScale(int tensorIndex);
int GetOutputQuantizationOffset(int tensorIndex);
/**
* @brief Runs inference on the provided input data, and stores the results
* in the provided InferenceResults object.
*
* @param[in] inputData - input frame data
* @param[in] dataBytes - input data size in bytes
* @param[out] outResults - Vector of DetectionResult objects used to store the output result.
*/
bool Run(const void *inputData, const size_t dataBytes,
InferenceResults<Tout> &outResults);
};
template <typename Tout>
ArmnnNetworkExecutor<Tout>::ArmnnNetworkExecutor(std::string& modelPath,
std::vector<armnn::BackendId>& preferredBackends,
bool isProfilingEnabled):
m_profiling(isProfilingEnabled)
{
m_profiling.ProfilingStart();
armnn::OptimizerOptions optimizerOptions;
m_model = tflite::FlatBufferModel::BuildFromFile(modelPath.c_str());
if (m_model == nullptr)
{
const std::string errorMessage{"ArmnnNetworkExecutor: Failed to build the model"};
ARMNN_LOG(error) << errorMessage;
throw armnn::Exception(errorMessage);
}
m_profiling.ProfilingStopAndPrintUs("Loading the model took");
m_profiling.ProfilingStart();
tflite::ops::builtin::BuiltinOpResolver resolver;
tflite::InterpreterBuilder(*m_model, resolver)(&m_interpreter);
if (m_interpreter->AllocateTensors() != kTfLiteOk)
{
const std::string errorMessage{"ArmnnNetworkExecutor: Failed to alloc tensors"};
ARMNN_LOG(error) << errorMessage;
throw armnn::Exception(errorMessage);
}
m_profiling.ProfilingStopAndPrintUs("Create the tflite interpreter");
/* create delegate options */
m_profiling.ProfilingStart();
/* enable fast math optimization */
armnn::BackendOptions modelOptionGpu("GpuAcc", {{"FastMathEnabled", true}});
optimizerOptions.m_ModelOptions.push_back(modelOptionGpu);
armnn::BackendOptions modelOptionCpu("CpuAcc", {{"FastMathEnabled", true}});
optimizerOptions.m_ModelOptions.push_back(modelOptionCpu);
/* enable reduce float32 to float16 optimization */
optimizerOptions.m_ReduceFp32ToFp16 = true;
armnnDelegate::DelegateOptions delegateOptions(preferredBackends, optimizerOptions);
/* create delegate object */
std::unique_ptr<TfLiteDelegate, decltype(&armnnDelegate::TfLiteArmnnDelegateDelete)>
theArmnnDelegate(armnnDelegate::TfLiteArmnnDelegateCreate(delegateOptions),
armnnDelegate::TfLiteArmnnDelegateDelete);
/* Register the delegate file */
m_interpreter->ModifyGraphWithDelegate(std::move(theArmnnDelegate));
m_profiling.ProfilingStopAndPrintUs("Create and load ArmNN Delegate");
}
template<typename Tout>
void ArmnnNetworkExecutor<Tout>::PrepareTensors(const void *inputData, const size_t dataBytes)
{
size_t inputTensorSize = m_interpreter->input_tensor(0)->bytes;
auto * inputTensorPtr = m_interpreter->input_tensor(0)->data.raw;
assert(inputTensorSize >= dataBytes);
if (inputTensorPtr != nullptr)
{
memcpy(inputTensorPtr, inputData, inputTensorSize);
}
else
{
const std::string errorMessage{"ArmnnNetworkExecutor: input tensor is null"};
ARMNN_LOG(error) << errorMessage;
throw armnn::Exception(errorMessage);
}
}
template <typename Tout>
bool ArmnnNetworkExecutor<Tout>::Run(const void *inputData, const size_t dataBytes,
InferenceResults<Tout>& outResults)
{
bool ret = false;
m_profiling.ProfilingStart();
PrepareTensors(inputData, dataBytes);
if (m_interpreter->Invoke() == kTfLiteOk)
{
ret = true;
// Extract the output tensor data.
outResults.clear();
outResults.reserve(m_interpreter->outputs().size());
for (int index = 0; index < m_interpreter->outputs().size(); index++)
{
size_t size = m_interpreter->output_tensor(index)->bytes / sizeof(Tout);
const Tout *p_Output = m_interpreter->typed_output_tensor<Tout>(index);
if (p_Output != nullptr) {
InferenceResult<float> outRes(p_Output, p_Output + size);
outResults.emplace_back(outRes);
}
else
{
const std::string errorMessage{"ArmnnNetworkExecutor: p_Output tensor is null"};
ARMNN_LOG(error) << errorMessage;
ret = false;
}
}
}
else
{
const std::string errorMessage{"ArmnnNetworkExecutor: Invoke has failed"};
ARMNN_LOG(error) << errorMessage;
}
m_profiling.ProfilingStopAndPrintUs("Perform inference");
return ret;
}
template <typename Tout>
Size ArmnnNetworkExecutor<Tout>::GetImageAspectRatio()
{
assert(m_interpreter->tensor(m_interpreter->inputs()[0])->dims->size == 4);
return Size(m_interpreter->tensor(m_interpreter->inputs()[0])->dims->data[2],
m_interpreter->tensor(m_interpreter->inputs()[0])->dims->data[1]);
}
template <typename Tout>
armnn::DataType ArmnnNetworkExecutor<Tout>::GetInputDataType() const
{
return GetDataType(*(m_interpreter->tensor(m_interpreter->inputs()[0])));
}
template <typename Tout>
float ArmnnNetworkExecutor<Tout>::GetQuantizationScale()
{
return m_interpreter->tensor(m_interpreter->inputs()[0])->params.scale;
}
template <typename Tout>
int ArmnnNetworkExecutor<Tout>::GetQuantizationOffset()
{
return m_interpreter->tensor(m_interpreter->inputs()[0])->params.zero_point;
}
template <typename Tout>
float ArmnnNetworkExecutor<Tout>::GetOutputQuantizationScale(int tensorIndex)
{
assert(m_interpreter->outputs().size() > tensorIndex);
return m_interpreter->tensor(m_interpreter->outputs()[tensorIndex])->params.scale;
}
template <typename Tout>
int ArmnnNetworkExecutor<Tout>::GetOutputQuantizationOffset(int tensorIndex)
{
assert(m_interpreter->outputs().size() > tensorIndex);
return m_interpreter->tensor(m_interpreter->outputs()[tensorIndex])->params.zero_point;
}
}// namespace common
|
