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
|
//
// Copyright © 2020 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
%{
#include "armnn/LstmParams.hpp"
#include "armnn/QuantizedLstmParams.hpp"
%}
namespace armnn
{
%feature("docstring",
"
Long Short-Term Memory layer input parameters.
See `INetwork.AddLstmLayer()`.
Operation described by the following equations:
\[i_t=\sigma(W_{xi}x_t+W_{hi}h_{t-1}+W_{ci}C_{t-1}+b_i) \\\\
f_t=\sigma(W_{xf}x_t+W_{hf}h_{t-1}+W_{cf}C_{t-1}+b_f) \\\\
C_t=clip(f_t \odot C_{t-1} + i_t \odot g(W_{xc}x_t+W_{hc}h_{t-1}+b_c),\ t_{cell}) \\\\
o_t = \sigma(W_{xo}x_t+W_{ho}h_{t-1}+W_{co}C_t+b_o) \\\\
h_t = clip(W_{proj}(o_t \odot g(C_t))+b_{proj},\ t_{proj})\ if\ there\ is\ a\ projection; \\\\
h_t = o_t \odot g(C_t)\ otherwise. \]
Where:
\(x_t\) - input;
\(i_t\) - input gate;
\(f_t\) - forget gate;
\(C_t\) - cell state;
\(o_t\) - output;
\(h_t\) - output state;
\(\sigma\) - logistic sigmoid function;
\(g\) - cell input and cell output activation function, see `LstmDescriptor.m_ActivationFunc`;
\(t_{cell}\) - threshold for clipping the cell state, see `LstmDescriptor.m_ClippingThresCell`;
\(t_{proj}\) - threshold for clipping the projected output, see `LstmDescriptor.m_ClippingThresProj`;
Contains:
m_InputToInputWeights (ConstTensor): \(W_{xi}\), input-to-input weight matrix.
m_InputToForgetWeights (ConstTensor): \(W_{xf}\), input-to-forget weight matrix.
m_InputToCellWeights (ConstTensor): \(W_{xc}\), input-to-cell weight matrix.
m_InputToOutputWeights (ConstTensor): \(W_{xo}\), input-to-output weight matrix.
m_RecurrentToInputWeights (ConstTensor): \(W_{hi}\), recurrent-to-input weight matrix.
m_RecurrentToForgetWeights (ConstTensor): \(W_{hf}\), recurrent-to-forget weight matrix.
m_RecurrentToCellWeights (ConstTensor): \(W_{hc}\), recurrent-to-cell weight matrix.
m_RecurrentToOutputWeights (ConstTensor): \(W_{ho}\), recurrent-to-output weight matrix.
m_CellToInputWeights (ConstTensor): \(W_{ci}\), cell-to-input weight matrix. Has effect if `LstmDescriptor.m_PeepholeEnabled`.
m_CellToForgetWeights (ConstTensor): \(W_{cf}\), cell-to-forget weight matrix. Has effect if `LstmDescriptor.m_PeepholeEnabled`.
m_CellToOutputWeights (ConstTensor): \(W_{co}\), cell-to-output weight matrix. Has effect if `LstmDescriptor.m_PeepholeEnabled`.
m_InputGateBias (ConstTensor): \(b_i\), input gate bias.
m_ForgetGateBias (ConstTensor): \(b_f\), forget gate bias.
m_CellBias (ConstTensor): \(b_c\), cell bias.
m_OutputGateBias (ConstTensor): \(b_o\), output gate bias.
m_ProjectionWeights (ConstTensor): \(W_{proj}\), projection weight matrix.
Has effect if `LstmDescriptor.m_ProjectionEnabled` is set to True.
m_ProjectionBias (ConstTensor): \(b_{proj}\), projection bias.
Has effect if `LstmDescriptor.m_ProjectionEnabled` is set to True.
m_InputLayerNormWeights (ConstTensor): normalisation weights for input,
has effect if `LstmDescriptor.m_LayerNormEnabled` set to True.
m_ForgetLayerNormWeights (ConstTensor): normalisation weights for forget gate,
has effect if `LstmDescriptor.m_LayerNormEnabled` set to True.
m_CellLayerNormWeights (ConstTensor): normalisation weights for current cell,
has effect if `LstmDescriptor.m_LayerNormEnabled` set to True.
m_OutputLayerNormWeights (ConstTensor): normalisation weights for output gate,
has effect if `LstmDescriptor.m_LayerNormEnabled` set to True.
") LstmInputParams;
struct LstmInputParams
{
LstmInputParams();
const armnn::ConstTensor* m_InputToInputWeights;
const armnn::ConstTensor* m_InputToForgetWeights;
const armnn::ConstTensor* m_InputToCellWeights;
const armnn::ConstTensor* m_InputToOutputWeights;
const armnn::ConstTensor* m_RecurrentToInputWeights;
const armnn::ConstTensor* m_RecurrentToForgetWeights;
const armnn::ConstTensor* m_RecurrentToCellWeights;
const armnn::ConstTensor* m_RecurrentToOutputWeights;
const armnn::ConstTensor* m_CellToInputWeights;
const armnn::ConstTensor* m_CellToForgetWeights;
const armnn::ConstTensor* m_CellToOutputWeights;
const armnn::ConstTensor* m_InputGateBias;
const armnn::ConstTensor* m_ForgetGateBias;
const armnn::ConstTensor* m_CellBias;
const armnn::ConstTensor* m_OutputGateBias;
const armnn::ConstTensor* m_ProjectionWeights;
const armnn::ConstTensor* m_ProjectionBias;
const armnn::ConstTensor* m_InputLayerNormWeights;
const armnn::ConstTensor* m_ForgetLayerNormWeights;
const armnn::ConstTensor* m_CellLayerNormWeights;
const armnn::ConstTensor* m_OutputLayerNormWeights;
};
%feature("docstring",
"
Quantized Long Short-Term Memory layer input parameters.
See `INetwork.AddQuantizedLstmLayer()`.
Operation described by the following equations:
\[i_t=\sigma(W_{xi}x_t+W_{hi}h_{t-1}+W_{ci}C_{t-1}+b_i) \\\\
f_t=\sigma(W_{xf}x_t+W_{hf}h_{t-1}+W_{cf}C_{t-1}+b_f) \\\\
C_t=clip(f_t \odot C_{t-1} + i_t \odot g(W_{xc}x_t+W_{hc}h_{t-1}+b_c),\ t_{cell}) \\\\
o_t = \sigma(W_{xo}x_t+W_{ho}h_{t-1}+W_{co}C_t+b_o) \\\\
h_t = clip(W_{proj}(o_t \odot g(C_t))+b_{proj},\ t_{proj})\ if\ there\ is\ a\ projection; \\\\
h_t = o_t \odot g(C_t)\ otherwise. \]
Where:
\(x_t\) - input;
\(i_t\) - input gate;
\(f_t\) - forget gate;
\(C_t\) - cell state;
\(o_t\) - output;
\(h_t\) - output state;
\(\sigma\) - logistic sigmoid function;
\(g\) - cell input and cell output activation function, see `LstmDescriptor.m_ActivationFunc`;
\(t_{cell}\) - threshold for clipping the cell state, see `LstmDescriptor.m_ClippingThresCell`;
\(t_{proj}\) - threshold for clipping the projected output, see `LstmDescriptor.m_ClippingThresProj`;
Contains:
m_InputToInputWeights (ConstTensor): \(W_{xi}\), input-to-input weight matrix.
m_InputToForgetWeights (ConstTensor): \(W_{xf}\), input-to-forget weight matrix.
m_InputToCellWeights (ConstTensor): \(W_{xc}\), input-to-cell weight matrix.
m_InputToOutputWeights (ConstTensor): \(W_{xo}\), input-to-output weight matrix.
m_RecurrentToInputWeights (ConstTensor): \(W_{hi}\), recurrent-to-input weight matrix.
m_RecurrentToForgetWeights (ConstTensor): \(W_{hf}\), recurrent-to-forget weight matrix.
m_RecurrentToCellWeights (ConstTensor): \(W_{hc}\), recurrent-to-cell weight matrix.
m_RecurrentToOutputWeights (ConstTensor): \(W_{ho}\), recurrent-to-output weight matrix.
m_InputGateBias (ConstTensor): \(b_i\), input gate bias.
m_ForgetGateBias (ConstTensor): \(b_f\), forget gate bias.
m_CellBias (ConstTensor): \(b_c\), cell bias.
m_OutputGateBias (ConstTensor): \(b_o\), output gate bias.
") QuantizedLstmInputParams;
struct QuantizedLstmInputParams
{
QuantizedLstmInputParams();
const armnn::ConstTensor* m_InputToInputWeights;
const armnn::ConstTensor* m_InputToForgetWeights;
const armnn::ConstTensor* m_InputToCellWeights;
const armnn::ConstTensor* m_InputToOutputWeights;
const armnn::ConstTensor* m_RecurrentToInputWeights;
const armnn::ConstTensor* m_RecurrentToForgetWeights;
const armnn::ConstTensor* m_RecurrentToCellWeights;
const armnn::ConstTensor* m_RecurrentToOutputWeights;
const armnn::ConstTensor* m_InputGateBias;
const armnn::ConstTensor* m_ForgetGateBias;
const armnn::ConstTensor* m_CellBias;
const armnn::ConstTensor* m_OutputGateBias;
};
}
|
