aboutsummaryrefslogtreecommitdiff
path: root/src/armnn/layers/LstmLayer.cpp
blob: 296c973d33810708f049c18ff65337ae3066333c (plain)
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
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "LstmLayer.hpp"

#include "LayerCloneBase.hpp"

#include <armnn/TypesUtils.hpp>
#include <backends/CpuTensorHandle.hpp>
#include <backends/WorkloadFactory.hpp>

namespace armnn
{

LstmLayer::LstmLayer(const LstmDescriptor& param, const char* name)
        : LayerWithParameters(3, 4, LayerType::Lstm, param, name)
{
}

std::unique_ptr<IWorkload> LstmLayer::CreateWorkload(const Graph& graph, const IWorkloadFactory& factory) const
{
    LstmQueueDescriptor descriptor;

    // Basic parameters
    descriptor.m_InputToForgetWeights = m_BasicParameters.m_InputToForgetWeights.get();
    descriptor.m_InputToCellWeights = m_BasicParameters.m_InputToCellWeights.get();
    descriptor.m_InputToOutputWeights = m_BasicParameters.m_InputToOutputWeights.get();
    descriptor.m_RecurrentToForgetWeights = m_BasicParameters.m_RecurrentToForgetWeights.get();
    descriptor.m_RecurrentToCellWeights = m_BasicParameters.m_RecurrentToCellWeights.get();
    descriptor.m_RecurrentToOutputWeights = m_BasicParameters.m_RecurrentToOutputWeights.get();
    descriptor.m_ForgetGateBias = m_BasicParameters.m_ForgetGateBias.get();
    descriptor.m_CellBias = m_BasicParameters.m_CellBias.get();
    descriptor.m_OutputGateBias = m_BasicParameters.m_OutputGateBias.get();

    // Cifg parameters
    if (!m_Param.m_CifgEnabled)
    {
        descriptor.m_InputToInputWeights = m_CifgParameters.m_InputToInputWeights.get();
        descriptor.m_RecurrentToInputWeights = m_CifgParameters.m_RecurrentToInputWeights.get();
        descriptor.m_CellToInputWeights = m_CifgParameters.m_CellToInputWeights.get();
        descriptor.m_InputGateBias = m_CifgParameters.m_InputGateBias.get();
    }

    // Projection parameters
    if (m_Param.m_ProjectionEnabled)
    {
        descriptor.m_ProjectionWeights = m_ProjectionParameters.m_ProjectionWeights.get();
        descriptor.m_ProjectionBias    = m_ProjectionParameters.m_ProjectionBias.get();
    }

    // Peephole parameters
    if (m_Param.m_PeepholeEnabled)
    {
        descriptor.m_CellToForgetWeights = m_PeepholeParameters.m_CellToForgetWeights.get();
        descriptor.m_CellToOutputWeights = m_PeepholeParameters.m_CellToOutputWeights.get();
    }
    return factory.CreateLstm(descriptor, PrepInfoAndDesc(descriptor, graph));
}

LstmLayer* LstmLayer::Clone(Graph& graph) const
{
    auto layer = CloneBase<LstmLayer>(graph, m_Param, GetName());

    layer->m_BasicParameters.m_InputToForgetWeights = m_BasicParameters.m_InputToForgetWeights ?
            std::make_unique<ScopedCpuTensorHandle>(*m_BasicParameters.m_InputToForgetWeights)
                : nullptr;
    layer->m_BasicParameters.m_InputToCellWeights = m_BasicParameters.m_InputToCellWeights ?
            std::make_unique<ScopedCpuTensorHandle>(*m_BasicParameters.m_InputToCellWeights) : nullptr;
    layer->m_BasicParameters.m_InputToOutputWeights = m_BasicParameters.m_InputToOutputWeights ?
            std::make_unique<ScopedCpuTensorHandle>(*m_BasicParameters.m_InputToOutputWeights) : nullptr;
    layer->m_BasicParameters.m_RecurrentToForgetWeights = m_BasicParameters.m_RecurrentToForgetWeights ?
            std::make_unique<ScopedCpuTensorHandle>(*m_BasicParameters.m_RecurrentToForgetWeights) : nullptr;
    layer->m_BasicParameters.m_RecurrentToCellWeights = m_BasicParameters.m_RecurrentToCellWeights ?
            std::make_unique<ScopedCpuTensorHandle>(*m_BasicParameters.m_RecurrentToCellWeights) : nullptr;
    layer->m_BasicParameters.m_RecurrentToOutputWeights = m_BasicParameters.m_RecurrentToOutputWeights ?
            std::make_unique<ScopedCpuTensorHandle>(*m_BasicParameters.m_RecurrentToOutputWeights) : nullptr;
    layer->m_BasicParameters.m_ForgetGateBias = m_BasicParameters.m_ForgetGateBias ?
            std::make_unique<ScopedCpuTensorHandle>(*m_BasicParameters.m_ForgetGateBias) : nullptr;
    layer->m_BasicParameters.m_CellBias = m_BasicParameters.m_CellBias ?
            std::make_unique<ScopedCpuTensorHandle>(*m_BasicParameters.m_CellBias) : nullptr;
    layer->m_BasicParameters.m_OutputGateBias = m_BasicParameters.m_OutputGateBias ?
            std::make_unique<ScopedCpuTensorHandle>(*m_BasicParameters.m_OutputGateBias) : nullptr;

    if (!m_Param.m_CifgEnabled)
    {
        layer->m_CifgParameters.m_InputToInputWeights = m_CifgParameters.m_InputToInputWeights ?
                std::make_unique<ScopedCpuTensorHandle>(*m_CifgParameters.m_InputToInputWeights) : nullptr;
        layer->m_CifgParameters.m_RecurrentToInputWeights = m_CifgParameters.m_RecurrentToInputWeights ?
                std::make_unique<ScopedCpuTensorHandle>(*m_CifgParameters.m_RecurrentToInputWeights) : nullptr;
        layer->m_CifgParameters.m_CellToInputWeights = m_CifgParameters.m_CellToInputWeights ?
                std::make_unique<ScopedCpuTensorHandle>(*m_CifgParameters.m_CellToInputWeights) : nullptr;
        layer->m_CifgParameters.m_InputGateBias = m_CifgParameters.m_InputGateBias ?
                std::make_unique<ScopedCpuTensorHandle>(*m_CifgParameters.m_InputGateBias) : nullptr;
    }

    if (m_Param.m_ProjectionEnabled)
    {
        layer->m_ProjectionParameters.m_ProjectionWeights = m_ProjectionParameters.m_ProjectionWeights ?
               std::make_unique<ScopedCpuTensorHandle>(*m_ProjectionParameters.m_ProjectionWeights) : nullptr;
        layer->m_ProjectionParameters.m_ProjectionBias = m_ProjectionParameters.m_ProjectionBias ?
               std::make_unique<ScopedCpuTensorHandle>(*m_ProjectionParameters.m_ProjectionBias) : nullptr;
    }

    if (m_Param.m_PeepholeEnabled)
    {
        layer->m_PeepholeParameters.m_CellToForgetWeights = m_PeepholeParameters.m_CellToForgetWeights ?
               std::make_unique<ScopedCpuTensorHandle>(*m_PeepholeParameters.m_CellToForgetWeights) : nullptr;
        layer->m_PeepholeParameters.m_CellToOutputWeights = m_PeepholeParameters.m_CellToOutputWeights ?
               std::make_unique<ScopedCpuTensorHandle>(*m_PeepholeParameters.m_CellToOutputWeights) : nullptr;
    }

    return std::move(layer);
}

std::vector<TensorShape> LstmLayer::InferOutputShapes(const std::vector<TensorShape>& inputShapes) const
{
    BOOST_ASSERT(inputShapes.size() == 3);

    // Get input values for validation
    unsigned int batchSize = inputShapes[0][0];
    unsigned int outputSize = inputShapes[1][1];
    unsigned int numUnits = inputShapes[2][1];

    std::vector<TensorShape> outShapes;
    if (!m_Param.m_CifgEnabled)
    {
        outShapes.push_back(TensorShape({batchSize, numUnits*3}));
    }
    else
    {
        outShapes.push_back(TensorShape({batchSize, numUnits*4}));
    }
    outShapes.push_back(TensorShape({batchSize, outputSize}));
    outShapes.push_back(TensorShape({batchSize, numUnits}));
    outShapes.push_back(TensorShape({batchSize, outputSize}));

    return outShapes;
}

void LstmLayer::ValidateTensorShapesFromInputs()
{
    VerifyLayerConnections(3, CHECK_LOCATION());

    auto inferredShapes = InferOutputShapes( {
        GetInputSlot(0).GetConnection()->GetTensorInfo().GetShape(),
        GetInputSlot(1).GetConnection()->GetTensorInfo().GetShape(),
        GetInputSlot(2).GetConnection()->GetTensorInfo().GetShape()}
    );

    BOOST_ASSERT(inferredShapes.size() == 4);

    // Check if the weights are nullptr
    BOOST_ASSERT_MSG(m_BasicParameters.m_InputToForgetWeights != nullptr,
                     "LstmLayer: m_BasicParameters.m_InputToForgetWeights should not be null.");
    BOOST_ASSERT_MSG(m_BasicParameters.m_InputToCellWeights != nullptr,
                     "LstmLayer: m_BasicParameters.m_InputToCellWeights should not be null.");
    BOOST_ASSERT_MSG(m_BasicParameters.m_InputToOutputWeights != nullptr,
                     "LstmLayer: m_BasicParameters.m_InputToOutputWeights should not be null.");
    BOOST_ASSERT_MSG(m_BasicParameters.m_RecurrentToForgetWeights != nullptr,
                     "LstmLayer: m_BasicParameters.m_RecurrentToForgetWeights should not be null.");
    BOOST_ASSERT_MSG(m_BasicParameters.m_RecurrentToCellWeights != nullptr,
                     "LstmLayer: m_BasicParameters.m_RecurrentToCellWeights should not be null.");
    BOOST_ASSERT_MSG(m_BasicParameters.m_RecurrentToOutputWeights != nullptr,
                     "LstmLayer: m_BasicParameters.m_RecurrentToOutputWeights should not be null.");
    BOOST_ASSERT_MSG(m_BasicParameters.m_ForgetGateBias != nullptr,
                     "LstmLayer: m_BasicParameters.m_ForgetGateBias should not be null.");
    BOOST_ASSERT_MSG(m_BasicParameters.m_CellBias != nullptr,
                     "LstmLayer: m_BasicParameters.m_CellBias should not be null.");
    BOOST_ASSERT_MSG(m_BasicParameters.m_OutputGateBias != nullptr,
                     "LstmLayer: m_BasicParameters.m_OutputGateBias should not be null.");

    if (!m_Param.m_CifgEnabled)
    {
        BOOST_ASSERT_MSG(m_CifgParameters.m_InputToInputWeights != nullptr,
                         "LstmLayer: m_CifgParameters.m_InputToInputWeights should not be null.");
        BOOST_ASSERT_MSG(m_CifgParameters.m_RecurrentToInputWeights != nullptr,
                         "LstmLayer: m_CifgParameters.m_RecurrentToInputWeights should not be null.");
        BOOST_ASSERT_MSG(m_CifgParameters.m_InputGateBias != nullptr,
                         "LstmLayer: m_CifgParameters.m_InputGateBias should not be null.");

        ConditionalThrowIfNotEqual<LayerValidationException>(
                "LstmLayer: TensorShape set on OutputSlot[0] does not match the inferred shape.",
                GetOutputSlot(0).GetTensorInfo().GetShape(),
                inferredShapes[0]);
    }
    else
    {
        BOOST_ASSERT_MSG(m_CifgParameters.m_InputToInputWeights == nullptr,
            "LstmLayer: m_CifgParameters.m_InputToInputWeights should not have a value when CIFG is enabled.");
        BOOST_ASSERT_MSG(m_CifgParameters.m_RecurrentToInputWeights == nullptr,
            "LstmLayer: m_CifgParameters.m_RecurrentToInputWeights should not have a value when CIFG is enabled.");
        BOOST_ASSERT_MSG(m_CifgParameters.m_CellToInputWeights == nullptr,
             "LstmLayer: m_CifgParameters.m_CellToInputWeights should not have a value when CIFG is enabled.");
        BOOST_ASSERT_MSG(m_CifgParameters.m_InputGateBias == nullptr,
            "LstmLayer: m_CifgParameters.m_InputGateBias should not have a value when CIFG is enabled.");

        ConditionalThrowIfNotEqual<LayerValidationException>(
                "LstmLayer: TensorShape set on OutputSlot[0] does not match the inferred shape.",
                GetOutputSlot(0).GetTensorInfo().GetShape(),
                inferredShapes[0]);
    }

    if (m_Param.m_ProjectionEnabled)
    {
        BOOST_ASSERT_MSG(m_ProjectionParameters.m_ProjectionWeights != nullptr,
                         "LstmLayer: m_ProjectionParameters.m_ProjectionWeights should not be null.");
    }

    if (m_Param.m_PeepholeEnabled)
    {
        BOOST_ASSERT_MSG(m_PeepholeParameters.m_CellToForgetWeights != nullptr,
                         "LstmLayer: m_PeepholeParameters.m_CellToForgetWeights should not be null.");
        BOOST_ASSERT_MSG(m_PeepholeParameters.m_CellToOutputWeights != nullptr,
                         "LstmLayer: m_PeepholeParameters.m_CellToOutputWeights should not be null.");
    }

    ConditionalThrowIfNotEqual<LayerValidationException>(
            "LstmLayer: TensorShape set on OutputSlot[1] does not match the inferred shape.",
            GetOutputSlot(1).GetTensorInfo().GetShape(),
            inferredShapes[1]);
    ConditionalThrowIfNotEqual<LayerValidationException>(
            "LstmLayer: TensorShape set on OutputSlot[2] does not match the inferred shape.",
            GetOutputSlot(2).GetTensorInfo().GetShape(),
            inferredShapes[2]);
    ConditionalThrowIfNotEqual<LayerValidationException>(
            "LstmLayer: TensorShape set on OutputSlot[3] does not match the inferred shape.",
            GetOutputSlot(3).GetTensorInfo().GetShape(),
            inferredShapes[3]);
}

Layer::ConstantTensors LstmLayer::GetConstantTensorsByRef()
{
    return {m_BasicParameters.m_InputToForgetWeights,
            m_BasicParameters.m_InputToCellWeights,
            m_BasicParameters.m_InputToOutputWeights,
            m_BasicParameters.m_RecurrentToForgetWeights,
            m_BasicParameters.m_RecurrentToCellWeights,
            m_BasicParameters.m_RecurrentToOutputWeights,
            m_BasicParameters.m_ForgetGateBias,
            m_BasicParameters.m_CellBias,
            m_BasicParameters.m_OutputGateBias,

            // Cifg parameters
            m_CifgParameters.m_InputToInputWeights,
            m_CifgParameters.m_RecurrentToInputWeights,
            m_CifgParameters.m_CellToInputWeights,
            m_CifgParameters.m_InputGateBias,

            // Projection parameters
            m_ProjectionParameters.m_ProjectionWeights,
            m_ProjectionParameters.m_ProjectionBias,

            // Peephole parameters
            m_PeepholeParameters.m_CellToForgetWeights,
            m_PeepholeParameters.m_CellToOutputWeights};
}

} // namespace armnn