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
|
//
// Copyright © 2021 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "Network.hpp"
#include "NetworkQuantizerUtils.hpp"
#include "StaticRangeStrategy.hpp"
#include <armnn/utility/NumericCast.hpp>
#include <armnn/utility/PolymorphicDowncast.hpp>
namespace armnn
{
class QuantizerStrategy : public IStrategy
{
public :
QuantizerStrategy(const RangeTracker& rangeTracker,
const IQuantizationScheme* quantizationScheme,
bool preserveType);
~QuantizerStrategy() = default;
void ExecuteStrategy(const armnn::IConnectableLayer* layer,
const BaseDescriptor& descriptor,
const std::vector<armnn::ConstTensor>& constants,
const char* name,
const armnn::LayerBindingId id) override;
/// Extract the quantized network
INetworkPtr RetrieveFinalNetwork() { return std::move(m_QuantizedNetwork); }
private:
/// Connects the layer to preceeding layers and sets the quantization parameters based on recorded ranges
void SetQuantizedInputConnections(const IConnectableLayer* srcLayer, IConnectableLayer* quantizedLayer);
/// Record the guids so we can easily find the layers later
void RecordLayer(const IConnectableLayer* srcLayer, IConnectableLayer* qLayer);
/// Sets the bias quantization scale based on input and weight scales
ConstTensor CreateQuantizedBias(const IConnectableLayer* srcLayer,
const ConstTensor& weights,
const Optional<ConstTensor>& biases,
std::vector<int32_t>& weightsBacking);
/// Reference to the static range visitor used to retrieve the quantization ranges
const RangeTracker& m_Ranges;
/// Quantized version of the model we are building up
INetworkPtr m_QuantizedNetwork;
/// Mapping from input network guids to quantized network guids
std::unordered_map<LayerGuid, LayerGuid> m_OriginalToQuantizedGuidMap;
/// Mapping from guid to layer in quantized network
std::unordered_map<LayerGuid, IConnectableLayer*> m_QuantizedGuidToLayerMap;
const IQuantizationScheme* m_QuantizationScheme;
const bool m_PreserveType;
};
} //namespace armnn
|
