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
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
|
//
// Copyright © 2017,2019-2023 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "QuantizedLstmLayer.hpp"
#include "LayerCloneBase.hpp"
#include <armnn/QuantizedLstmParams.hpp>
#include <armnn/TypesUtils.hpp>
#include <armnn/backends/TensorHandle.hpp>
#include <armnn/backends/WorkloadFactory.hpp>
namespace armnn
{
QuantizedLstmLayer::QuantizedLstmLayer(const char* name)
: Layer(3, 2, LayerType::QuantizedLstm, name)
{
}
std::unique_ptr<IWorkload> QuantizedLstmLayer::CreateWorkload(const IWorkloadFactory& factory) const
{
QuantizedLstmQueueDescriptor descriptor;
// QuantizedLstmLayer parameters - there are no optional params
descriptor.m_InputToInputWeights = m_QuantizedLstmParameters.m_InputToInputWeights.get();
descriptor.m_InputToForgetWeights = m_QuantizedLstmParameters.m_InputToForgetWeights.get();
descriptor.m_InputToCellWeights = m_QuantizedLstmParameters.m_InputToCellWeights.get();
descriptor.m_InputToOutputWeights = m_QuantizedLstmParameters.m_InputToOutputWeights.get();
descriptor.m_RecurrentToInputWeights = m_QuantizedLstmParameters.m_RecurrentToInputWeights.get();
descriptor.m_RecurrentToForgetWeights = m_QuantizedLstmParameters.m_RecurrentToForgetWeights.get();
descriptor.m_RecurrentToCellWeights = m_QuantizedLstmParameters.m_RecurrentToCellWeights.get();
descriptor.m_RecurrentToOutputWeights = m_QuantizedLstmParameters.m_RecurrentToOutputWeights.get();
descriptor.m_InputGateBias = m_QuantizedLstmParameters.m_InputGateBias.get();
descriptor.m_ForgetGateBias = m_QuantizedLstmParameters.m_ForgetGateBias.get();
descriptor.m_CellBias = m_QuantizedLstmParameters.m_CellBias.get();
descriptor.m_OutputGateBias = m_QuantizedLstmParameters.m_OutputGateBias.get();
SetAdditionalInfo(descriptor);
return factory.CreateWorkload(LayerType::QuantizedLstm, descriptor, PrepInfoAndDesc(descriptor));
}
QuantizedLstmLayer* QuantizedLstmLayer::Clone(Graph& graph) const
{
auto layer = CloneBase<QuantizedLstmLayer>(graph, GetName());
layer->m_QuantizedLstmParameters.m_InputToInputWeights = m_QuantizedLstmParameters.m_InputToInputWeights ?
m_QuantizedLstmParameters.m_InputToInputWeights : nullptr;
layer->m_QuantizedLstmParameters.m_InputToForgetWeights = m_QuantizedLstmParameters.m_InputToForgetWeights ?
m_QuantizedLstmParameters.m_InputToForgetWeights : nullptr;
layer->m_QuantizedLstmParameters.m_InputToCellWeights = m_QuantizedLstmParameters.m_InputToCellWeights ?
m_QuantizedLstmParameters.m_InputToCellWeights : nullptr;
layer->m_QuantizedLstmParameters.m_InputToOutputWeights = m_QuantizedLstmParameters.m_InputToOutputWeights ?
m_QuantizedLstmParameters.m_InputToOutputWeights : nullptr;
layer->m_QuantizedLstmParameters.m_RecurrentToInputWeights = m_QuantizedLstmParameters.m_RecurrentToInputWeights ?
m_QuantizedLstmParameters.m_RecurrentToInputWeights : nullptr;
layer->m_QuantizedLstmParameters.m_RecurrentToForgetWeights = m_QuantizedLstmParameters.m_RecurrentToForgetWeights
? m_QuantizedLstmParameters.m_RecurrentToForgetWeights : nullptr;
layer->m_QuantizedLstmParameters.m_RecurrentToCellWeights = m_QuantizedLstmParameters.m_RecurrentToCellWeights ?
m_QuantizedLstmParameters.m_RecurrentToCellWeights : nullptr;
layer->m_QuantizedLstmParameters.m_RecurrentToOutputWeights = m_QuantizedLstmParameters.m_RecurrentToOutputWeights
? m_QuantizedLstmParameters.m_RecurrentToOutputWeights : nullptr;
layer->m_QuantizedLstmParameters.m_InputGateBias = m_QuantizedLstmParameters.m_InputGateBias ?
m_QuantizedLstmParameters.m_InputGateBias : nullptr;
layer->m_QuantizedLstmParameters.m_ForgetGateBias = m_QuantizedLstmParameters.m_ForgetGateBias ?
m_QuantizedLstmParameters.m_ForgetGateBias : nullptr;
layer->m_QuantizedLstmParameters.m_CellBias = m_QuantizedLstmParameters.m_CellBias ?
m_QuantizedLstmParameters.m_CellBias : nullptr;
layer->m_QuantizedLstmParameters.m_OutputGateBias = m_QuantizedLstmParameters.m_OutputGateBias ?
m_QuantizedLstmParameters.m_OutputGateBias : nullptr;
return std::move(layer);
}
std::vector<TensorShape> QuantizedLstmLayer::InferOutputShapes(const std::vector<TensorShape>& inputShapes) const
{
ARMNN_ASSERT(inputShapes.size() == 3);
// Get input values for validation
unsigned int numBatches = inputShapes[0][0];
unsigned int outputSize = inputShapes[1][1];
std::vector<TensorShape> outShapes;
outShapes.push_back(TensorShape({numBatches, outputSize})); // cellStateOut
outShapes.push_back(TensorShape({numBatches, outputSize})); // output
return outShapes;
}
void QuantizedLstmLayer::ValidateTensorShapesFromInputs()
{
VerifyLayerConnections(3, CHECK_LOCATION());
const TensorShape& outputShape = GetOutputSlot(0).GetTensorInfo().GetShape();
VerifyShapeInferenceType(outputShape, m_ShapeInferenceMethod);
auto inferredShapes = InferOutputShapes(
{
GetInputSlot(0).GetTensorInfo().GetShape(), // input
GetInputSlot(1).GetTensorInfo().GetShape(), // previousCellStateIn
GetInputSlot(2).GetTensorInfo().GetShape() // previousOutputIn
});
ARMNN_ASSERT(inferredShapes.size() == 2);
// Check weights and bias for nullptr
ARMNN_ASSERT_MSG(m_QuantizedLstmParameters.m_InputToInputWeights != nullptr,
"QuantizedLstmLayer: m_QuantizedLstmParameters.m_InputToInputWeights should not be null.");
ARMNN_ASSERT_MSG(m_QuantizedLstmParameters.m_InputToForgetWeights != nullptr,
"QuantizedLstmLayer: m_QuantizedLstmParameters.m_InputToForgetWeights should not be null.");
ARMNN_ASSERT_MSG(m_QuantizedLstmParameters.m_InputToCellWeights != nullptr,
"QuantizedLstmLayer: m_QuantizedLstmParameters.m_InputToCellWeights should not be null.");
ARMNN_ASSERT_MSG(m_QuantizedLstmParameters.m_InputToOutputWeights != nullptr,
"QuantizedLstmLayer: m_QuantizedLstmParameters.m_InputToOutputWeights should not be null.");
ARMNN_ASSERT_MSG(m_QuantizedLstmParameters.m_RecurrentToInputWeights != nullptr,
"QuantizedLstmLayer: m_QuantizedLstmParameters.m_RecurrentToInputWeights should not be null.");
ARMNN_ASSERT_MSG(m_QuantizedLstmParameters.m_RecurrentToForgetWeights != nullptr,
"QuantizedLstmLayer: m_QuantizedLstmParameters.m_RecurrentToForgetWeights should not be null.");
ARMNN_ASSERT_MSG(m_QuantizedLstmParameters.m_RecurrentToCellWeights != nullptr,
"QuantizedLstmLayer: m_QuantizedLstmParameters.m_RecurrentToCellWeights should not be null.");
ARMNN_ASSERT_MSG(m_QuantizedLstmParameters.m_RecurrentToOutputWeights != nullptr,
"QuantizedLstmLayer: m_QuantizedLstmParameters.m_RecurrentToOutputWeights should not be null.");
ARMNN_ASSERT_MSG(m_QuantizedLstmParameters.m_InputGateBias != nullptr,
"QuantizedLstmLayer: m_QuantizedLstmParameters.m_InputGateBias should not be null.");
ARMNN_ASSERT_MSG(m_QuantizedLstmParameters.m_ForgetGateBias != nullptr,
"QuantizedLstmLayer: m_QuantizedLstmParameters.m_ForgetGateBias should not be null.");
ARMNN_ASSERT_MSG(m_QuantizedLstmParameters.m_CellBias != nullptr,
"QuantizedLstmLayer: m_QuantizedLstmParameters.m_CellBias should not be null.");
ARMNN_ASSERT_MSG(m_QuantizedLstmParameters.m_OutputGateBias != nullptr,
"QuantizedLstmLayer: m_QuantizedLstmParameters.m_OutputGateBias should not be null.");
// Check output TensorShape(s) match inferred shape
ValidateAndCopyShape(outputShape, inferredShapes[0], m_ShapeInferenceMethod, "QuantizedLstmLayer");
ValidateAndCopyShape(GetOutputSlot(1).GetTensorInfo().GetShape(),
inferredShapes[1],
m_ShapeInferenceMethod,
"QuantizedLstmLayer",
1);
}
Layer::ImmutableConstantTensors QuantizedLstmLayer::GetConstantTensorsByRef() const
{
// For API stability DO NOT ALTER order and add new members to the end of vector
return
{
m_QuantizedLstmParameters.m_InputToInputWeights,
m_QuantizedLstmParameters.m_InputToForgetWeights,
m_QuantizedLstmParameters.m_InputToCellWeights,
m_QuantizedLstmParameters.m_InputToOutputWeights,
m_QuantizedLstmParameters.m_RecurrentToInputWeights,
m_QuantizedLstmParameters.m_RecurrentToForgetWeights,
m_QuantizedLstmParameters.m_RecurrentToCellWeights,
m_QuantizedLstmParameters.m_RecurrentToOutputWeights,
m_QuantizedLstmParameters.m_InputGateBias,
m_QuantizedLstmParameters.m_ForgetGateBias,
m_QuantizedLstmParameters.m_CellBias,
m_QuantizedLstmParameters.m_OutputGateBias
};
}
void QuantizedLstmLayer::ExecuteStrategy(IStrategy& strategy) const
{
std::vector<ConstTensor> constTensors;
ManagedConstTensorHandle managedInputToInputWeights(m_QuantizedLstmParameters.m_InputToInputWeights);
ManagedConstTensorHandle managedInputToForgetWeights(m_QuantizedLstmParameters.m_InputToForgetWeights);
ManagedConstTensorHandle managedInputToCellWeights(m_QuantizedLstmParameters.m_InputToCellWeights);
ManagedConstTensorHandle managedInputToOutputWeights(m_QuantizedLstmParameters.m_InputToOutputWeights);
ManagedConstTensorHandle managedRecurrentToInputWeights(m_QuantizedLstmParameters.m_RecurrentToInputWeights);
ManagedConstTensorHandle managedRecurrentToForgetWeights(m_QuantizedLstmParameters.m_RecurrentToForgetWeights);
ManagedConstTensorHandle managedRecurrentToCellWeights(m_QuantizedLstmParameters.m_RecurrentToCellWeights);
ManagedConstTensorHandle managedRecurrentToOutputWeights(m_QuantizedLstmParameters.m_RecurrentToOutputWeights);
ManagedConstTensorHandle managedInputGateBias(m_QuantizedLstmParameters.m_InputGateBias);
ManagedConstTensorHandle managedForgetGateBias(m_QuantizedLstmParameters.m_ForgetGateBias);
ManagedConstTensorHandle managedCellBias(m_QuantizedLstmParameters.m_CellBias);
ManagedConstTensorHandle managedOutputGateBias(m_QuantizedLstmParameters.m_OutputGateBias);
// InputToX weight tensors
if (m_QuantizedLstmParameters.m_InputToInputWeights != nullptr)
{
constTensors.emplace_back(ConstTensor(managedInputToInputWeights.GetTensorInfo(),
managedInputToInputWeights.Map()));
}
if (m_QuantizedLstmParameters.m_InputToForgetWeights != nullptr)
{
constTensors.emplace_back(ConstTensor(managedInputToForgetWeights.GetTensorInfo(),
managedInputToForgetWeights.Map()));
}
if (m_QuantizedLstmParameters.m_InputToCellWeights != nullptr)
{
constTensors.emplace_back(ConstTensor(managedInputToCellWeights.GetTensorInfo(),
managedInputToCellWeights.Map()));
}
if (m_QuantizedLstmParameters.m_InputToOutputWeights != nullptr)
{
constTensors.emplace_back(ConstTensor(managedInputToOutputWeights.GetTensorInfo(),
managedInputToOutputWeights.Map()));
}
// RecurrentToX weight tensors
if (m_QuantizedLstmParameters.m_RecurrentToInputWeights != nullptr)
{
constTensors.emplace_back(ConstTensor(
managedRecurrentToInputWeights.GetTensorInfo(),
managedRecurrentToInputWeights.Map()));
}
if (m_QuantizedLstmParameters.m_RecurrentToForgetWeights != nullptr)
{
constTensors.emplace_back(ConstTensor(
managedRecurrentToForgetWeights.GetTensorInfo(),
managedRecurrentToForgetWeights.Map()));
}
if (m_QuantizedLstmParameters.m_RecurrentToCellWeights != nullptr)
{
constTensors.emplace_back(ConstTensor(
managedRecurrentToCellWeights.GetTensorInfo(),
managedRecurrentToCellWeights.Map()));
}
if (m_QuantizedLstmParameters.m_RecurrentToOutputWeights != nullptr)
{
constTensors.emplace_back(ConstTensor(
managedRecurrentToOutputWeights.GetTensorInfo(),
managedRecurrentToOutputWeights.Map()));
}
// Bias tensors
if (m_QuantizedLstmParameters.m_InputGateBias != nullptr)
{
constTensors.emplace_back(ConstTensor(managedInputGateBias.GetTensorInfo(),
managedInputGateBias.Map()));
}
if (m_QuantizedLstmParameters.m_ForgetGateBias != nullptr)
{
constTensors.emplace_back(ConstTensor(managedForgetGateBias.GetTensorInfo(),
managedForgetGateBias.Map()));
}
if (m_QuantizedLstmParameters.m_CellBias != nullptr)
{
constTensors.emplace_back(ConstTensor(managedCellBias.GetTensorInfo(),
managedCellBias.Map()));
}
if (m_QuantizedLstmParameters.m_OutputGateBias != nullptr)
{
constTensors.emplace_back(ConstTensor(managedOutputGateBias.GetTensorInfo(),
managedOutputGateBias.Map()));
}
strategy.ExecuteStrategy(this, BaseDescriptor(), constTensors, GetName());
}
} // namespace armnn
|
