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
|
//
// Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#pragma once
#include "Types.hpp"
#include "armnn/ArmNN.hpp"
#include "armnnTfLiteParser/ITfLiteParser.hpp"
#include "armnnUtils/DataLayoutIndexed.hpp"
#include <armnn/Logging.hpp>
#include <string>
#include <vector>
namespace common
{
/**
* @brief Used to load in a network through ArmNN and run inference on it against a given backend.
*
*/
template <class Tout>
class ArmnnNetworkExecutor
{
private:
armnn::IRuntimePtr m_Runtime;
armnn::NetworkId m_NetId{};
mutable InferenceResults<Tout> m_OutputBuffer;
armnn::InputTensors m_InputTensors;
armnn::OutputTensors m_OutputTensors;
std::vector<armnnTfLiteParser::BindingPointInfo> m_outputBindingInfo;
std::vector<std::string> m_outputLayerNamesList;
armnnTfLiteParser::BindingPointInfo m_inputBindingInfo;
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. Parsed through TfLiteParser and optimized for a
* given backend.
*
* Note that the output layers names order in m_outputLayerNamesList affects the order of the feature vectors
* in output of the Run method.
*
* * @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);
/**
* @brief Returns the aspect ratio of the associated model in the order of width, height.
*/
Size GetImageAspectRatio();
armnn::DataType GetInputDataType() const;
float GetQuantizationScale();
int GetQuantizationOffset();
/**
* @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] results - Vector of DetectionResult objects used to store the output result.
*/
bool Run(const void* inputData, const size_t dataBytes, common::InferenceResults<Tout>& outResults);
};
template <class Tout>
ArmnnNetworkExecutor<Tout>::ArmnnNetworkExecutor(std::string& modelPath,
std::vector<armnn::BackendId>& preferredBackends)
: m_Runtime(armnn::IRuntime::Create(armnn::IRuntime::CreationOptions()))
{
// Import the TensorFlow lite model.
armnnTfLiteParser::ITfLiteParserPtr parser = armnnTfLiteParser::ITfLiteParser::Create();
armnn::INetworkPtr network = parser->CreateNetworkFromBinaryFile(modelPath.c_str());
std::vector<std::string> inputNames = parser->GetSubgraphInputTensorNames(0);
m_inputBindingInfo = parser->GetNetworkInputBindingInfo(0, inputNames[0]);
m_outputLayerNamesList = parser->GetSubgraphOutputTensorNames(0);
std::vector<armnn::BindingPointInfo> outputBindings;
for(const std::string& name : m_outputLayerNamesList)
{
m_outputBindingInfo.push_back(std::move(parser->GetNetworkOutputBindingInfo(0, name)));
}
std::vector<std::string> errorMessages;
// optimize the network.
armnn::IOptimizedNetworkPtr optNet = Optimize(*network,
preferredBackends,
m_Runtime->GetDeviceSpec(),
armnn::OptimizerOptions(),
armnn::Optional<std::vector<std::string>&>(errorMessages));
if (!optNet)
{
const std::string errorMessage{"ArmnnNetworkExecutor: Failed to optimize network"};
ARMNN_LOG(error) << errorMessage;
throw armnn::Exception(errorMessage);
}
// Load the optimized network onto the m_Runtime device
std::string errorMessage;
if (armnn::Status::Success != m_Runtime->LoadNetwork(m_NetId, std::move(optNet), errorMessage))
{
ARMNN_LOG(error) << errorMessage;
throw armnn::Exception(errorMessage);
}
//pre-allocate memory for output (the size of it never changes)
for (int it = 0; it < m_outputLayerNamesList.size(); ++it)
{
const armnn::DataType dataType = m_outputBindingInfo[it].second.GetDataType();
const armnn::TensorShape& tensorShape = m_outputBindingInfo[it].second.GetShape();
std::vector<Tout> oneLayerOutResult;
oneLayerOutResult.resize(tensorShape.GetNumElements(), 0);
m_OutputBuffer.emplace_back(oneLayerOutResult);
// Make ArmNN output tensors
m_OutputTensors.reserve(m_OutputBuffer.size());
for (size_t it = 0; it < m_OutputBuffer.size(); ++it)
{
m_OutputTensors.emplace_back(std::make_pair(
m_outputBindingInfo[it].first,
armnn::Tensor(m_outputBindingInfo[it].second,
m_OutputBuffer.at(it).data())
));
}
}
}
template <class Tout>
armnn::DataType ArmnnNetworkExecutor<Tout>::GetInputDataType() const
{
return m_inputBindingInfo.second.GetDataType();
}
template <class Tout>
void ArmnnNetworkExecutor<Tout>::PrepareTensors(const void* inputData, const size_t dataBytes)
{
assert(m_inputBindingInfo.second.GetNumBytes() >= dataBytes);
m_InputTensors.clear();
m_InputTensors = {{ m_inputBindingInfo.first, armnn::ConstTensor(m_inputBindingInfo.second, inputData)}};
}
template <class Tout>
bool ArmnnNetworkExecutor<Tout>::Run(const void* inputData, const size_t dataBytes, InferenceResults<Tout>& outResults)
{
/* Prepare tensors if they are not ready */
ARMNN_LOG(debug) << "Preparing tensors...";
this->PrepareTensors(inputData, dataBytes);
ARMNN_LOG(trace) << "Running inference...";
armnn::Status ret = m_Runtime->EnqueueWorkload(m_NetId, m_InputTensors, m_OutputTensors);
std::stringstream inferenceFinished;
inferenceFinished << "Inference finished with code {" << log_as_int(ret) << "}\n";
ARMNN_LOG(trace) << inferenceFinished.str();
if (ret == armnn::Status::Failure)
{
ARMNN_LOG(error) << "Failed to perform inference.";
}
outResults.reserve(m_outputLayerNamesList.size());
outResults = m_OutputBuffer;
return (armnn::Status::Success == ret);
}
template <class Tout>
float ArmnnNetworkExecutor<Tout>::GetQuantizationScale()
{
return this->m_inputBindingInfo.second.GetQuantizationScale();
}
template <class Tout>
int ArmnnNetworkExecutor<Tout>::GetQuantizationOffset()
{
return this->m_inputBindingInfo.second.GetQuantizationOffset();
}
template <class Tout>
Size ArmnnNetworkExecutor<Tout>::GetImageAspectRatio()
{
const auto shape = m_inputBindingInfo.second.GetShape();
assert(shape.GetNumDimensions() == 4);
armnnUtils::DataLayoutIndexed nhwc(armnn::DataLayout::NHWC);
return Size(shape[nhwc.GetWidthIndex()],
shape[nhwc.GetHeightIndex()]);
}
}// namespace common
|
