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
|
//
// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "TfliteExecutor.hpp"
TfLiteExecutor::TfLiteExecutor(const ExecuteNetworkParams& params) : m_Params(params)
{
std::unique_ptr<tflite::FlatBufferModel> model =
tflite::FlatBufferModel::BuildFromFile(m_Params.m_ModelPath.c_str());
m_TfLiteInterpreter = std::make_unique<Interpreter>();
tflite::ops::builtin::BuiltinOpResolver resolver;
tflite::InterpreterBuilder builder(*model, resolver);
builder(&m_TfLiteInterpreter);
m_TfLiteInterpreter->AllocateTensors();
int status = kTfLiteError;
if (m_Params.m_TfLiteExecutor == ExecuteNetworkParams::TfLiteExecutor::ArmNNTfLiteDelegate)
{
// Create the Armnn Delegate
// Populate a DelegateOptions from the ExecuteNetworkParams.
armnnDelegate::DelegateOptions delegateOptions = m_Params.ToDelegateOptions();
delegateOptions.SetExternalProfilingParams(delegateOptions.GetExternalProfilingParams());
std::unique_ptr<TfLiteDelegate, decltype(&armnnDelegate::TfLiteArmnnDelegateDelete)>
theArmnnDelegate(armnnDelegate::TfLiteArmnnDelegateCreate(delegateOptions),
armnnDelegate::TfLiteArmnnDelegateDelete);
// Register armnn_delegate to TfLiteInterpreter
status = m_TfLiteInterpreter->ModifyGraphWithDelegate(std::move(theArmnnDelegate));
if (status == kTfLiteError)
{
LogAndThrow("Could not register ArmNN TfLite Delegate to TfLiteInterpreter");
}
}
else
{
std::cout << "Running on TfLite without ArmNN delegate\n";
}
armnn::Optional<std::string> dataFile = m_Params.m_GenerateTensorData
? armnn::EmptyOptional()
: armnn::MakeOptional<std::string>(m_Params.m_InputTensorDataFilePaths[0]);
const size_t numInputs = m_Params.m_InputNames.size();
for(unsigned int inputIndex = 0; inputIndex < numInputs; ++inputIndex)
{
int input = m_TfLiteInterpreter->inputs()[inputIndex];
TfLiteIntArray* inputDims = m_TfLiteInterpreter->tensor(input)->dims;
unsigned int inputSize = 1;
for (unsigned int dim = 0; dim < static_cast<unsigned int>(inputDims->size); ++dim)
{
inputSize *= inputDims->data[dim];
}
const auto& inputName = m_TfLiteInterpreter->input_tensor(input)->name;
const auto& dataType = m_TfLiteInterpreter->input_tensor(input)->type;
switch (dataType)
{
case kTfLiteFloat32:
{
auto inputData = m_TfLiteInterpreter->typed_tensor<float>(input);
PopulateTensorWithData(inputData, inputSize, dataFile, inputName);
break;
}
case kTfLiteInt32:
{
auto inputData = m_TfLiteInterpreter->typed_tensor<int>(input);
PopulateTensorWithData(inputData, inputSize, dataFile, inputName);
break;
}
case kTfLiteUInt8:
{
auto inputData = m_TfLiteInterpreter->typed_tensor<uint8_t>(input);
PopulateTensorWithData(inputData, inputSize, dataFile, inputName);
break;
}
case kTfLiteInt16:
{
auto inputData = m_TfLiteInterpreter->typed_tensor<int16_t>(input);
PopulateTensorWithData(inputData, inputSize, dataFile, inputName);
break;
}
case kTfLiteInt8:
{
auto inputData = m_TfLiteInterpreter->typed_tensor<int8_t>(input);
PopulateTensorWithData(inputData, inputSize, dataFile, inputName);
break;
}
default:
{
LogAndThrow("Unsupported input tensor data type");
}
}
}
}
std::vector<const void *> TfLiteExecutor::Execute()
{
int status = 0;
std::vector<const void*> results;
for (size_t x = 0; x < m_Params.m_Iterations; x++)
{
// Start timer to record inference time in milliseconds.
const auto start_time = armnn::GetTimeNow();
// Run the inference
status = m_TfLiteInterpreter->Invoke();
const auto duration = armnn::GetTimeDuration(start_time);
if (m_Params.m_DontPrintOutputs || m_Params.m_ReuseBuffers)
{
break;
}
// Print out the output
for (unsigned int outputIndex = 0; outputIndex < m_TfLiteInterpreter->outputs().size(); ++outputIndex)
{
auto tfLiteDelegateOutputId = m_TfLiteInterpreter->outputs()[outputIndex];
TfLiteIntArray* outputDims = m_TfLiteInterpreter->tensor(tfLiteDelegateOutputId)->dims;
// If we've been asked to write to a file then set a file output stream. Otherwise use stdout.
FILE* outputTensorFile = stdout;
if (!m_Params.m_OutputTensorFiles.empty())
{
outputTensorFile = fopen(m_Params.m_OutputTensorFiles[outputIndex].c_str(), "w");
if (outputTensorFile == NULL)
{
LogAndThrow("Specified output tensor file, \"" + m_Params.m_OutputTensorFiles[outputIndex] +
"\", cannot be created. Defaulting to stdout. Error was: " + std::strerror(errno));
}
else
{
ARMNN_LOG(info) << "Writing output " << outputIndex << "' of iteration: " << x+1 << " to file: '"
<< m_Params.m_OutputTensorFiles[outputIndex] << "'";
}
}
long outputSize = 1;
for (unsigned int dim = 0; dim < static_cast<unsigned int>(outputDims->size); ++dim)
{
outputSize *= outputDims->data[dim];
}
std::cout << m_TfLiteInterpreter->tensor(tfLiteDelegateOutputId)->name << ": ";
results.push_back(m_TfLiteInterpreter->tensor(tfLiteDelegateOutputId)->allocation);
switch (m_TfLiteInterpreter->tensor(tfLiteDelegateOutputId)->type)
{
case kTfLiteFloat32:
{
auto tfLiteDelageOutputData = m_TfLiteInterpreter->typed_tensor<float>(tfLiteDelegateOutputId);
for (int i = 0; i < outputSize; ++i)
{
fprintf(outputTensorFile, "%f ", tfLiteDelageOutputData[i]);
}
break;
}
case kTfLiteInt32:
{
auto tfLiteDelageOutputData = m_TfLiteInterpreter->typed_tensor<int32_t>(tfLiteDelegateOutputId);
for (int i = 0; i < outputSize; ++i)
{
fprintf(outputTensorFile, "%d ", tfLiteDelageOutputData[i]);
}
break;
}
case kTfLiteUInt8:
{
auto tfLiteDelageOutputData = m_TfLiteInterpreter->typed_tensor<uint8_t>(tfLiteDelegateOutputId);
for (int i = 0; i < outputSize; ++i)
{
fprintf(outputTensorFile, "%u ", tfLiteDelageOutputData[i]);
}
break;
}
case kTfLiteInt8:
{
auto tfLiteDelageOutputData = m_TfLiteInterpreter->typed_tensor<int8_t>(tfLiteDelegateOutputId);
for (int i = 0; i < outputSize; ++i)
{
fprintf(outputTensorFile, "%d ", tfLiteDelageOutputData[i]);
}
break;
}
default:
{
LogAndThrow("Unsupported output type");
}
}
std::cout << std::endl;
}
CheckInferenceTimeThreshold(duration, m_Params.m_ThresholdTime);
}
std::cout << status;
return results;
}
void TfLiteExecutor::CompareAndPrintResult(std::vector<const void*> otherOutput)
{
for (unsigned int outputIndex = 0; outputIndex < m_TfLiteInterpreter->outputs().size(); ++outputIndex)
{
auto tfLiteDelegateOutputId = m_TfLiteInterpreter->outputs()[outputIndex];
float result = 0;
switch (m_TfLiteInterpreter->tensor(tfLiteDelegateOutputId)->type)
{
case kTfLiteFloat32:
{
result = ComputeRMSE<float>(m_TfLiteInterpreter->tensor(tfLiteDelegateOutputId)->allocation,
otherOutput[outputIndex],
m_TfLiteInterpreter->tensor(tfLiteDelegateOutputId)->bytes);
break;
}
case kTfLiteInt32:
{
result = ComputeRMSE<int32_t>(m_TfLiteInterpreter->tensor(tfLiteDelegateOutputId)->allocation,
otherOutput[outputIndex],
m_TfLiteInterpreter->tensor(tfLiteDelegateOutputId)->bytes);
break;
}
case kTfLiteUInt8:
{
result = ComputeRMSE<uint8_t>(m_TfLiteInterpreter->tensor(tfLiteDelegateOutputId)->allocation,
otherOutput[outputIndex],
m_TfLiteInterpreter->tensor(tfLiteDelegateOutputId)->bytes);
break;
}
case kTfLiteInt8:
{
result = ComputeRMSE<int8_t>(m_TfLiteInterpreter->tensor(tfLiteDelegateOutputId)->allocation,
otherOutput[outputIndex],
m_TfLiteInterpreter->tensor(tfLiteDelegateOutputId)->bytes);
break;
}
default:
{
}
}
std::cout << "RMSE of "
<< m_TfLiteInterpreter->tensor(tfLiteDelegateOutputId)->name
<< ": " << result << std::endl;
}
};
|
