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
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "Network.hpp"
#include "QuantizerVisitor.hpp"
#include "StaticRangeVisitor.hpp"
#include "armnn/TypesUtils.hpp"
#include <cmath>
#include <stdint.h>
#include <limits>
namespace armnn
{
namespace {
std::pair<int, float> ComputeQAsymmParams(int numBits, double min, double max)
{
BOOST_ASSERT_MSG(min < max, "Min >= max will result in invalid quantization.");
double highest = (1 << numBits)-1;
min = std::min(0.0, min); // min <= 0.0
max = std::max(0.0, max); // max >= 0.0
// assumes quantization range [0-highest]
double scale = (max-min) / highest;
double offset = -min / scale;
// clamp offset [0-highest]
offset = std::max(0.0, std::min(highest, offset));
return std::make_pair(static_cast<int>(std::round(offset)), static_cast<float>(scale));
}
template<typename srcType>
void Quantize(const srcType* src, uint8_t* dst, size_t numElements, float &scale, int &offset)
{
BOOST_ASSERT(src);
BOOST_ASSERT(dst);
float min = std::numeric_limits<srcType>::max();
float max = std::numeric_limits<srcType>::lowest();
for (size_t i = 0; i < numElements; ++i)
{
min = std::min(min, src[i]);
max = std::max(max, src[i]);
}
auto qParams = ComputeQAsymmParams(8, min, max);
offset = qParams.first;
scale = qParams.second;
for (size_t i = 0; i < numElements; ++i)
{
dst[i] = armnn::Quantize<uint8_t>(src[i], scale, offset);
}
}
ConstTensor CreateQuantizedConst(const ConstTensor& tensor, std::vector<uint8_t> &backing)
{
float scale = 0.0f;
int offset = 0;
// Reserve the backing memory
backing.resize(tensor.GetInfo().GetNumElements());
DataType type = tensor.GetInfo().GetDataType();
switch(type)
{
case DataType::Float32:
{
Quantize(static_cast<const float*>( tensor.GetMemoryArea()),
backing.data(),
backing.size(),
scale,
offset);
}
break;
default:
BOOST_ASSERT_MSG(false, "Can't quantize unsupported data type");
}
TensorInfo qInfo(tensor.GetInfo().GetShape(), DataType::QuantisedAsymm8, scale, offset);
return ConstTensor(qInfo, backing);
}
} // namespace
QuantizerVisitor::QuantizerVisitor(armnn::StaticRangeVisitor* ranges)
: m_Ranges(ranges)
, 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_OldToNewGuidMap.find(layerToFind.GetGuid());
if (found != m_OldToNewGuidMap.end())
{
// Connect the slots in the quantized model
IConnectableLayer* prevQuantizedLayer = m_GuidToLayerMap[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(), i);
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);
}
else
{
// error in graph traversal order
BOOST_ASSERT_MSG(false, "Error in graph traversal");
}
}
}
void QuantizerVisitor::RecordLayer(const IConnectableLayer* srcLayer, IConnectableLayer* quantizedLayer)
{
m_OldToNewGuidMap[srcLayer->GetGuid()] = quantizedLayer->GetGuid();
m_GuidToLayerMap[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::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);
}
} //namespace armnn
|
