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
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "Network.hpp"
#include "QuantizerVisitor.hpp"
#include "StaticRangeVisitor.hpp"
#include "NetworkQuantizerUtils.hpp"
namespace armnn
{
QuantizerVisitor::QuantizerVisitor(const RangeTracker& rangeTracker)
: m_Ranges(rangeTracker)
, m_QuantizedNetwork(INetwork::Create())
{
}
void QuantizerVisitor::SetQuantizedInputConnections(const IConnectableLayer* srcLayer,
IConnectableLayer* quantizedLayer)
{
for (unsigned int i = 0; i < srcLayer->GetNumInputSlots(); i++)
{
const IInputSlot& srcInputSlot = srcLayer->GetInputSlot(i);
const InputSlot* inputSlot = boost::polymorphic_downcast<const InputSlot*>(&srcInputSlot);
const OutputSlot* outputSlot = inputSlot->GetConnectedOutputSlot();
unsigned int slotIdx = outputSlot->CalculateIndexOnOwner();
Layer& layerToFind = outputSlot->GetOwningLayer();
auto found = m_OriginalToQuantizedGuidMap.find(layerToFind.GetGuid());
if (found == m_OriginalToQuantizedGuidMap.end())
{
// Error in graph traversal order
BOOST_ASSERT_MSG(false, "Error in graph traversal");
return;
}
// Connect the slots in the quantized model
IConnectableLayer* prevQuantizedLayer = m_QuantizedGuidToLayerMap[found->second];
IInputSlot& newInputSlot = quantizedLayer->GetInputSlot(i);
IOutputSlot& newOutputSlot = prevQuantizedLayer->GetOutputSlot(slotIdx);
newOutputSlot.Connect(newInputSlot);
// Fetch the min/max ranges that were computed earlier
auto range = m_Ranges.GetRange(layerToFind.GetGuid(), slotIdx);
auto qParams = ComputeQAsymmParams(8, range.first, range.second);
// Set the quantization params
TensorInfo info(newOutputSlot.GetTensorInfo());
info.SetDataType(DataType::QuantisedAsymm8);
info.SetQuantizationOffset(qParams.first);
info.SetQuantizationScale(qParams.second);
newOutputSlot.SetTensorInfo(info);
}
}
void QuantizerVisitor::RecordLayer(const IConnectableLayer* srcLayer, IConnectableLayer* quantizedLayer)
{
m_OriginalToQuantizedGuidMap[srcLayer->GetGuid()] = quantizedLayer->GetGuid();
m_QuantizedGuidToLayerMap[quantizedLayer->GetGuid()] = quantizedLayer;
}
void QuantizerVisitor::VisitAdditionLayer(const IConnectableLayer* layer, const char* name)
{
IConnectableLayer* newLayer = m_QuantizedNetwork->AddAdditionLayer(name);
RecordLayer(layer, newLayer);
SetQuantizedInputConnections(layer, newLayer);
}
void QuantizerVisitor::VisitActivationLayer(const IConnectableLayer* layer,
const ActivationDescriptor& activationDescriptor,
const char* name)
{
IConnectableLayer* newLayer = m_QuantizedNetwork->AddActivationLayer(activationDescriptor, name);
RecordLayer(layer, newLayer);
SetQuantizedInputConnections(layer, newLayer);
}
void QuantizerVisitor::VisitFullyConnectedLayer(const IConnectableLayer *layer,
const FullyConnectedDescriptor& desc,
const ConstTensor& weights,
const Optional<ConstTensor>& biases,
const char *name)
{
std::vector<uint8_t> weightsBacking;
ConstTensor qWeights = CreateQuantizedConst(weights, weightsBacking);
IConnectableLayer* newLayer;
if (biases.has_value())
{
std::vector<uint8_t> biasBacking;
ConstTensor qBias = CreateQuantizedConst(biases.value(), biasBacking);
newLayer = m_QuantizedNetwork->AddFullyConnectedLayer(desc, qWeights, qBias, name);
}
else
{
newLayer = m_QuantizedNetwork->AddFullyConnectedLayer(desc, qWeights, name);
}
RecordLayer(layer, newLayer);
SetQuantizedInputConnections(layer, newLayer);
}
void QuantizerVisitor::VisitInputLayer(const IConnectableLayer *layer, LayerBindingId id, const char *name)
{
IConnectableLayer* newLayer = m_QuantizedNetwork->AddInputLayer(id, name);
RecordLayer(layer, newLayer);
}
void QuantizerVisitor::VisitOutputLayer(const IConnectableLayer* layer, LayerBindingId id, const char* name)
{
IConnectableLayer* newLayer = m_QuantizedNetwork->AddOutputLayer(id, name);
RecordLayer(layer, newLayer);
SetQuantizedInputConnections(layer, newLayer);
}
void QuantizerVisitor::VisitBatchNormalizationLayer(const IConnectableLayer* layer,
const BatchNormalizationDescriptor& desc,
const ConstTensor& mean,
const ConstTensor& variance,
const ConstTensor& beta,
const ConstTensor& gamma,
const char* name)
{
std::vector<uint8_t> meanBacking;
ConstTensor qMean = CreateQuantizedConst(mean, meanBacking);
std::vector<uint8_t> varianceBacking;
ConstTensor qVariance = CreateQuantizedConst(variance, varianceBacking);
std::vector<uint8_t> betaBacking;
ConstTensor qBeta = CreateQuantizedConst(beta, betaBacking);
std::vector<uint8_t> gammaBacking;
ConstTensor qGamma = CreateQuantizedConst(variance, gammaBacking);
IConnectableLayer* newLayer = m_QuantizedNetwork->AddBatchNormalizationLayer(desc,
qMean,
qVariance,
qBeta,
qGamma,
name);
RecordLayer(layer, newLayer);
SetQuantizedInputConnections(layer, newLayer);
}
void QuantizerVisitor::VisitConvolution2dLayer(const IConnectableLayer* layer,
const Convolution2dDescriptor& convolution2dDescriptor,
const ConstTensor& weights,
const Optional<ConstTensor>& biases,
const char* name)
{
std::vector<uint8_t> weightsBacking;
ConstTensor qWeights = CreateQuantizedConst(weights, weightsBacking);
IConnectableLayer* newLayer;
if (biases.has_value())
{
std::vector<uint8_t> biasesBacking;
ConstTensor qBiases = CreateQuantizedConst(biases.value(), biasesBacking);
newLayer = m_QuantizedNetwork->AddConvolution2dLayer(convolution2dDescriptor,
qWeights,
qBiases,
name);
}
else
{
newLayer = m_QuantizedNetwork->AddConvolution2dLayer(convolution2dDescriptor, qWeights, name);
}
RecordLayer(layer, newLayer);
SetQuantizedInputConnections(layer, newLayer);
}
void QuantizerVisitor::VisitDepthwiseConvolution2dLayer(const IConnectableLayer* layer,
const DepthwiseConvolution2dDescriptor& desc,
const ConstTensor& weights,
const Optional<ConstTensor>& biases,
const char* name)
{
std::vector<uint8_t> weightsBacking;
ConstTensor qWeights = CreateQuantizedConst(weights, weightsBacking);
IConnectableLayer* newLayer;
if (biases.has_value())
{
std::vector<uint8_t> biasesBacking;
ConstTensor qBiases = CreateQuantizedConst(biases.value(), biasesBacking);
newLayer = m_QuantizedNetwork->AddDepthwiseConvolution2dLayer(desc,
qWeights,
qBiases,
name);
}
else
{
newLayer = m_QuantizedNetwork->AddDepthwiseConvolution2dLayer(desc, qWeights, name);
}
RecordLayer(layer, newLayer);
SetQuantizedInputConnections(layer, newLayer);
}
void QuantizerVisitor::VisitPermuteLayer(const IConnectableLayer* layer,
const PermuteDescriptor& permuteDescriptor,
const char* name)
{
IConnectableLayer* newLayer = m_QuantizedNetwork->AddPermuteLayer(permuteDescriptor, name);
RecordLayer(layer, newLayer);
SetQuantizedInputConnections(layer, newLayer);
}
void QuantizerVisitor::VisitSpaceToBatchNdLayer(const IConnectableLayer* layer,
const SpaceToBatchNdDescriptor& spaceToBatchNdDescriptor,
const char* name)
{
IConnectableLayer* newLayer = m_QuantizedNetwork->AddSpaceToBatchNdLayer(spaceToBatchNdDescriptor, name);
RecordLayer(layer, newLayer);
SetQuantizedInputConnections(layer, newLayer);
}
void QuantizerVisitor::VisitPooling2dLayer(const IConnectableLayer* layer,
const Pooling2dDescriptor& pooling2dDescriptor,
const char* name)
{
IConnectableLayer* newLayer = m_QuantizedNetwork->AddPooling2dLayer(pooling2dDescriptor, name);
RecordLayer(layer, newLayer);
SetQuantizedInputConnections(layer, newLayer);
}
void QuantizerVisitor::VisitMergerLayer(const IConnectableLayer* layer,
const OriginsDescriptor& mergerDescriptor,
const char* name)
{
IConnectableLayer* newLayer = m_QuantizedNetwork->AddMergerLayer(mergerDescriptor, name);
RecordLayer(layer, newLayer);
SetQuantizedInputConnections(layer, newLayer);
}
void QuantizerVisitor::VisitSoftmaxLayer(const IConnectableLayer* layer,
const SoftmaxDescriptor& softmaxDescriptor,
const char* name)
{
IConnectableLayer* newLayer = m_QuantizedNetwork->AddSoftmaxLayer(softmaxDescriptor, name);
RecordLayer(layer, newLayer);
SetQuantizedInputConnections(layer, newLayer);
}
void QuantizerVisitor::VisitConstantLayer(const IConnectableLayer* layer,
const ConstTensor& input,
const char* name)
{
std::vector<uint8_t> inputBacking;
ConstTensor qInput = CreateQuantizedConst(input, inputBacking);
IConnectableLayer* newLayer = m_QuantizedNetwork->AddConstantLayer(qInput, name);
RecordLayer(layer, newLayer);
}
void QuantizerVisitor::VisitSplitterLayer(const IConnectableLayer* layer,
const SplitterDescriptor& splitterDescriptor,
const char* name)
{
IConnectableLayer* newLayer = m_QuantizedNetwork->AddSplitterLayer(splitterDescriptor, name);
RecordLayer(layer, newLayer);
SetQuantizedInputConnections(layer, newLayer);
}
void QuantizerVisitor::VisitReshapeLayer(const IConnectableLayer* layer,
const ReshapeDescriptor& reshapeDescriptor,
const char* name)
{
IConnectableLayer* newLayer = m_QuantizedNetwork->AddReshapeLayer(reshapeDescriptor, name);
RecordLayer(layer, newLayer);
SetQuantizedInputConnections(layer, newLayer);
}
void QuantizerVisitor::VisitResizeBilinearLayer(const IConnectableLayer* layer,
const ResizeBilinearDescriptor& resizeDesc,
const char* name)
{
IConnectableLayer* newLayer = m_QuantizedNetwork->AddResizeBilinearLayer(resizeDesc, name);
RecordLayer(layer, newLayer);
SetQuantizedInputConnections(layer, newLayer);
}
void QuantizerVisitor::VisitStridedSliceLayer(const IConnectableLayer* layer,
const StridedSliceDescriptor& stridedSliceDescriptor,
const char* name)
{
IConnectableLayer* newLayer = m_QuantizedNetwork->AddStridedSliceLayer(stridedSliceDescriptor, name);
RecordLayer(layer, newLayer);
SetQuantizedInputConnections(layer, newLayer);
}
void QuantizerVisitor::VisitBatchToSpaceNdLayer(const IConnectableLayer* layer,
const BatchToSpaceNdDescriptor& batchToSpaceNdDescriptor,
const char* name)
{
IConnectableLayer* newLayer = m_QuantizedNetwork->AddBatchToSpaceNdLayer(batchToSpaceNdDescriptor, name);
RecordLayer(layer, newLayer);
SetQuantizedInputConnections(layer, newLayer);
}
} //namespace armnn
|
