Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2017 ARM Limited. |
| 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | #ifndef __ARM_COMPUTE_TEST_MODEL_OBJECTS_ALEXNET_H__ |
| 25 | #define __ARM_COMPUTE_TEST_MODEL_OBJECTS_ALEXNET_H__ |
| 26 | |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 27 | #include "arm_compute/runtime/NEON/NEScheduler.h" |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 28 | #include "arm_compute/runtime/Tensor.h" |
| 29 | |
Moritz Pflanzer | a09de0c | 2017-09-01 20:41:12 +0100 | [diff] [blame] | 30 | #include "tests/AssetsLibrary.h" |
| 31 | #include "tests/Globals.h" |
| 32 | #include "tests/Utils.h" |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 33 | |
| 34 | #include <memory> |
| 35 | |
| 36 | namespace arm_compute |
| 37 | { |
| 38 | namespace test |
| 39 | { |
| 40 | namespace networks |
| 41 | { |
| 42 | /** AlexNet model object */ |
| 43 | template <typename ITensorType, |
| 44 | typename TensorType, |
| 45 | typename SubTensorType, |
| 46 | typename Accessor, |
| 47 | typename ActivationLayerFunction, |
| 48 | typename ConvolutionLayerFunction, |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 49 | typename DirectConvolutionLayerFunction, |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 50 | typename FullyConnectedLayerFunction, |
| 51 | typename NormalizationLayerFunction, |
| 52 | typename PoolingLayerFunction, |
| 53 | typename SoftmaxLayerFunction> |
| 54 | class AlexNetNetwork |
| 55 | { |
| 56 | public: |
| 57 | void init(DataType data_type, int fixed_point_position, int batches, bool reshaped_weights = false) |
| 58 | { |
| 59 | _data_type = data_type; |
| 60 | _fixed_point_position = fixed_point_position; |
| 61 | _batches = batches; |
| 62 | _reshaped_weights = reshaped_weights; |
| 63 | |
| 64 | // Initialize weights and biases |
| 65 | if(!_reshaped_weights) |
| 66 | { |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 67 | w[0].allocator()->init(TensorInfo(TensorShape(11U, 11U, 3U, 96U), 1, _data_type, _fixed_point_position)); |
| 68 | b[0].allocator()->init(TensorInfo(TensorShape(96U), 1, _data_type, _fixed_point_position)); |
| 69 | w[1].allocator()->init(TensorInfo(TensorShape(5U, 5U, 48U, 256U), 1, _data_type, _fixed_point_position)); |
| 70 | b[1].allocator()->init(TensorInfo(TensorShape(256U), 1, _data_type, _fixed_point_position)); |
| 71 | w[2].allocator()->init(TensorInfo(TensorShape(3U, 3U, 256U, 384U), 1, _data_type, _fixed_point_position)); |
| 72 | b[2].allocator()->init(TensorInfo(TensorShape(384U), 1, _data_type, _fixed_point_position)); |
| 73 | w[3].allocator()->init(TensorInfo(TensorShape(3U, 3U, 192U, 384U), 1, _data_type, _fixed_point_position)); |
| 74 | b[3].allocator()->init(TensorInfo(TensorShape(384U), 1, _data_type, _fixed_point_position)); |
| 75 | w[4].allocator()->init(TensorInfo(TensorShape(3U, 3U, 192U, 256U), 1, _data_type, _fixed_point_position)); |
| 76 | b[4].allocator()->init(TensorInfo(TensorShape(256U), 1, _data_type, _fixed_point_position)); |
| 77 | w[5].allocator()->init(TensorInfo(TensorShape(9216U, 4096U), 1, _data_type, _fixed_point_position)); |
| 78 | b[5].allocator()->init(TensorInfo(TensorShape(4096U), 1, _data_type, _fixed_point_position)); |
| 79 | w[6].allocator()->init(TensorInfo(TensorShape(4096U, 4096U), 1, _data_type, _fixed_point_position)); |
| 80 | b[6].allocator()->init(TensorInfo(TensorShape(4096U), 1, _data_type, _fixed_point_position)); |
| 81 | w[7].allocator()->init(TensorInfo(TensorShape(4096U, 1000U), 1, _data_type, _fixed_point_position)); |
| 82 | b[7].allocator()->init(TensorInfo(TensorShape(1000U), 1, _data_type, _fixed_point_position)); |
| 83 | |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 84 | w11 = std::unique_ptr<SubTensorType>(new SubTensorType(&w[1], TensorShape(5U, 5U, 48U, 128U), Coordinates())); |
| 85 | w12 = std::unique_ptr<SubTensorType>(new SubTensorType(&w[1], TensorShape(5U, 5U, 48U, 128U), Coordinates(0, 0, 0, 128))); |
| 86 | b11 = std::unique_ptr<SubTensorType>(new SubTensorType(&b[1], TensorShape(128U), Coordinates())); |
| 87 | b12 = std::unique_ptr<SubTensorType>(new SubTensorType(&b[1], TensorShape(128U), Coordinates(128))); |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 88 | |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 89 | w31 = std::unique_ptr<SubTensorType>(new SubTensorType(&w[3], TensorShape(3U, 3U, 192U, 192U), Coordinates())); |
| 90 | w32 = std::unique_ptr<SubTensorType>(new SubTensorType(&w[3], TensorShape(3U, 3U, 192U, 192U), Coordinates(0, 0, 0, 192))); |
| 91 | b31 = std::unique_ptr<SubTensorType>(new SubTensorType(&b[3], TensorShape(192U), Coordinates())); |
| 92 | b32 = std::unique_ptr<SubTensorType>(new SubTensorType(&b[3], TensorShape(192U), Coordinates(192))); |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 93 | |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 94 | w41 = std::unique_ptr<SubTensorType>(new SubTensorType(&w[4], TensorShape(3U, 3U, 192U, 128U), Coordinates())); |
| 95 | w42 = std::unique_ptr<SubTensorType>(new SubTensorType(&w[4], TensorShape(3U, 3U, 192U, 128U), Coordinates(0, 0, 0, 128))); |
| 96 | b41 = std::unique_ptr<SubTensorType>(new SubTensorType(&b[4], TensorShape(128U), Coordinates())); |
| 97 | b42 = std::unique_ptr<SubTensorType>(new SubTensorType(&b[4], TensorShape(128U), Coordinates(128))); |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 98 | } |
| 99 | else |
| 100 | { |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 101 | auto reshape = [&](unsigned int width, unsigned int height, bool convolution_layer) -> TensorShape |
Moritz Pflanzer | 95643d8 | 2017-08-31 17:10:18 +0100 | [diff] [blame] | 102 | { |
Moritz Pflanzer | 80373f6 | 2017-09-15 10:42:58 +0100 | [diff] [blame] | 103 | const bool is_optimised = std::is_same<ITensorType, ITensor>::value && NEScheduler::get().cpu_info().CPU >= CPUTarget::ARMV7 && data_type == DataType::F32; |
Moritz Pflanzer | 95643d8 | 2017-08-31 17:10:18 +0100 | [diff] [blame] | 104 | |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 105 | if(convolution_layer && is_optimised) |
| 106 | { |
| 107 | return TensorShape{ height, width }; |
| 108 | } |
| 109 | else |
| 110 | { |
| 111 | const int interleave_width = 16 / arm_compute::data_size_from_type(_data_type); |
| 112 | |
| 113 | return TensorShape{ width * interleave_width, static_cast<unsigned int>(std::ceil(static_cast<float>(height) / interleave_width)) }; |
| 114 | } |
Moritz Pflanzer | 95643d8 | 2017-08-31 17:10:18 +0100 | [diff] [blame] | 115 | }; |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 116 | |
| 117 | // Create tensor for the reshaped weights |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 118 | w[0].allocator()->init(TensorInfo(reshape(366U, 96U, true), 1, _data_type, _fixed_point_position)); |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 119 | |
| 120 | // Configure the direct convolution's weights. Direct convolution doesn't need reshape weights |
| 121 | if(!_is_direct_conv) |
| 122 | { |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 123 | auto w11_tensor = std::unique_ptr<TensorType>(new TensorType()); |
| 124 | auto w12_tensor = std::unique_ptr<TensorType>(new TensorType()); |
| 125 | auto w31_tensor = std::unique_ptr<TensorType>(new TensorType()); |
| 126 | auto w32_tensor = std::unique_ptr<TensorType>(new TensorType()); |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 127 | auto w41_tensor = std::unique_ptr<TensorType>(new TensorType()); |
| 128 | auto w42_tensor = std::unique_ptr<TensorType>(new TensorType()); |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 129 | w11_tensor->allocator()->init(TensorInfo(reshape(1248U, 128U, true), 1, _data_type, _fixed_point_position)); |
| 130 | w12_tensor->allocator()->init(TensorInfo(reshape(1248U, 128U, true), 1, _data_type, _fixed_point_position)); |
| 131 | w31_tensor->allocator()->init(TensorInfo(reshape(1920U, 192U, true), 1, _data_type, _fixed_point_position)); |
| 132 | w32_tensor->allocator()->init(TensorInfo(reshape(1920U, 192U, true), 1, _data_type, _fixed_point_position)); |
| 133 | w41_tensor->allocator()->init(TensorInfo(reshape(1920U, 128U, true), 1, _data_type, _fixed_point_position)); |
| 134 | w42_tensor->allocator()->init(TensorInfo(reshape(1920U, 128U, true), 1, _data_type, _fixed_point_position)); |
| 135 | w[2].allocator()->init(TensorInfo(reshape(2560U, 384U, true), 1, _data_type, _fixed_point_position)); |
| 136 | w11 = std::move(w11_tensor); |
| 137 | w12 = std::move(w12_tensor); |
| 138 | w31 = std::move(w31_tensor); |
| 139 | w32 = std::move(w32_tensor); |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 140 | w41 = std::move(w41_tensor); |
| 141 | w42 = std::move(w42_tensor); |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 142 | } |
| 143 | else |
| 144 | { |
Gian Marco Iodice | 2e44868 | 2017-08-22 10:40:47 +0100 | [diff] [blame] | 145 | w[1].allocator()->init(TensorInfo(TensorShape(5U, 5U, 48U, 256U), 1, _data_type, _fixed_point_position)); |
| 146 | b[1].allocator()->init(TensorInfo(TensorShape(256U), 1, _data_type, _fixed_point_position)); |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 147 | w[2].allocator()->init(TensorInfo(TensorShape(3U, 3U, 256U, 384U), 1, _data_type, _fixed_point_position)); |
| 148 | b[2].allocator()->init(TensorInfo(TensorShape(384U), 1, _data_type, _fixed_point_position)); |
| 149 | w[3].allocator()->init(TensorInfo(TensorShape(3U, 3U, 192U, 384U), 1, _data_type, _fixed_point_position)); |
| 150 | b[3].allocator()->init(TensorInfo(TensorShape(384U), 1, _data_type, _fixed_point_position)); |
| 151 | w[4].allocator()->init(TensorInfo(TensorShape(3U, 3U, 192U, 256U), 1, _data_type, _fixed_point_position)); |
| 152 | b[4].allocator()->init(TensorInfo(TensorShape(256U), 1, _data_type, _fixed_point_position)); |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 153 | w11 = std::unique_ptr<SubTensorType>(new SubTensorType(&w[1], TensorShape(5U, 5U, 48U, 128U), Coordinates())); |
| 154 | w12 = std::unique_ptr<SubTensorType>(new SubTensorType(&w[1], TensorShape(5U, 5U, 48U, 128U), Coordinates(0, 0, 0, 128))); |
| 155 | b11 = std::unique_ptr<SubTensorType>(new SubTensorType(&b[1], TensorShape(128U), Coordinates())); |
| 156 | b12 = std::unique_ptr<SubTensorType>(new SubTensorType(&b[1], TensorShape(128U), Coordinates(128))); |
Gian Marco Iodice | 2e44868 | 2017-08-22 10:40:47 +0100 | [diff] [blame] | 157 | |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 158 | w31 = std::unique_ptr<SubTensorType>(new SubTensorType(&w[3], TensorShape(3U, 3U, 192U, 192U), Coordinates())); |
| 159 | w32 = std::unique_ptr<SubTensorType>(new SubTensorType(&w[3], TensorShape(3U, 3U, 192U, 192U), Coordinates(0, 0, 0, 192))); |
| 160 | b31 = std::unique_ptr<SubTensorType>(new SubTensorType(&b[3], TensorShape(192U), Coordinates())); |
| 161 | b32 = std::unique_ptr<SubTensorType>(new SubTensorType(&b[3], TensorShape(192U), Coordinates(192))); |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 162 | |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 163 | w41 = std::unique_ptr<SubTensorType>(new SubTensorType(&w[4], TensorShape(3U, 3U, 192U, 128U), Coordinates())); |
| 164 | w42 = std::unique_ptr<SubTensorType>(new SubTensorType(&w[4], TensorShape(3U, 3U, 192U, 128U), Coordinates(0, 0, 0, 128))); |
| 165 | b41 = std::unique_ptr<SubTensorType>(new SubTensorType(&b[4], TensorShape(128U), Coordinates())); |
| 166 | b42 = std::unique_ptr<SubTensorType>(new SubTensorType(&b[4], TensorShape(128U), Coordinates(128))); |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 167 | } |
| 168 | |
| 169 | b[5].allocator()->init(TensorInfo(TensorShape(4096U), 1, _data_type, _fixed_point_position)); |
| 170 | b[6].allocator()->init(TensorInfo(TensorShape(4096U), 1, _data_type, _fixed_point_position)); |
| 171 | b[7].allocator()->init(TensorInfo(TensorShape(1000U), 1, _data_type, _fixed_point_position)); |
| 172 | |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 173 | if(_batches > 1 && std::is_same<TensorType, Tensor>::value) |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 174 | { |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 175 | w[5].allocator()->init(TensorInfo(reshape(9216U, 4096U, false), 1, _data_type, _fixed_point_position)); |
| 176 | w[6].allocator()->init(TensorInfo(reshape(4096U, 4096U, false), 1, _data_type, _fixed_point_position)); |
| 177 | w[7].allocator()->init(TensorInfo(reshape(4096U, 1000U, false), 1, _data_type, _fixed_point_position)); |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 178 | } |
| 179 | else |
| 180 | { |
| 181 | w[5].allocator()->init(TensorInfo(TensorShape(4096U, 9216U), 1, _data_type, _fixed_point_position)); |
| 182 | w[6].allocator()->init(TensorInfo(TensorShape(4096U, 4096U), 1, _data_type, _fixed_point_position)); |
| 183 | w[7].allocator()->init(TensorInfo(TensorShape(1000U, 4096U), 1, _data_type, _fixed_point_position)); |
| 184 | } |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 185 | } |
| 186 | } |
| 187 | |
| 188 | void build() |
| 189 | { |
| 190 | input.allocator()->init(TensorInfo(TensorShape(227U, 227U, 3U, _batches), 1, _data_type, _fixed_point_position)); |
| 191 | output.allocator()->init(TensorInfo(TensorShape(1000U, _batches), 1, _data_type, _fixed_point_position)); |
| 192 | |
| 193 | // Initialize intermediate tensors |
| 194 | // Layer 1 |
| 195 | conv1_out.allocator()->init(TensorInfo(TensorShape(55U, 55U, 96U, _batches), 1, _data_type, _fixed_point_position)); |
| 196 | act1_out.allocator()->init(TensorInfo(TensorShape(55U, 55U, 96U, _batches), 1, _data_type, _fixed_point_position)); |
| 197 | norm1_out.allocator()->init(TensorInfo(TensorShape(55U, 55U, 96U, _batches), 1, _data_type, _fixed_point_position)); |
| 198 | pool1_out.allocator()->init(TensorInfo(TensorShape(27U, 27U, 96U, _batches), 1, _data_type, _fixed_point_position)); |
| 199 | pool11_out = std::unique_ptr<SubTensorType>(new SubTensorType(&pool1_out, TensorShape(27U, 27U, 48U, _batches), Coordinates())); |
| 200 | pool12_out = std::unique_ptr<SubTensorType>(new SubTensorType(&pool1_out, TensorShape(27U, 27U, 48U, _batches), Coordinates(0, 0, 48))); |
| 201 | // Layer 2 |
| 202 | conv2_out.allocator()->init(TensorInfo(TensorShape(27U, 27U, 256U, _batches), 1, _data_type, _fixed_point_position)); |
| 203 | conv21_out = std::unique_ptr<SubTensorType>(new SubTensorType(&conv2_out, TensorShape(27U, 27U, 128U, _batches), Coordinates())); |
| 204 | conv22_out = std::unique_ptr<SubTensorType>(new SubTensorType(&conv2_out, TensorShape(27U, 27U, 128U, _batches), Coordinates(0, 0, 128))); |
| 205 | act2_out.allocator()->init(TensorInfo(TensorShape(27U, 27U, 256U, _batches), 1, _data_type, _fixed_point_position)); |
| 206 | norm2_out.allocator()->init(TensorInfo(TensorShape(27U, 27U, 256U, _batches), 1, _data_type, _fixed_point_position)); |
| 207 | pool2_out.allocator()->init(TensorInfo(TensorShape(13U, 13U, 256U, _batches), 1, _data_type, _fixed_point_position)); |
| 208 | // Layer 3 |
| 209 | conv3_out.allocator()->init(TensorInfo(TensorShape(13U, 13U, 384U, _batches), 1, _data_type, _fixed_point_position)); |
| 210 | act3_out.allocator()->init(TensorInfo(TensorShape(13U, 13U, 384U, _batches), 1, _data_type, _fixed_point_position)); |
| 211 | act31_out = std::unique_ptr<SubTensorType>(new SubTensorType(&act3_out, TensorShape(13U, 13U, 192U, _batches), Coordinates())); |
| 212 | act32_out = std::unique_ptr<SubTensorType>(new SubTensorType(&act3_out, TensorShape(13U, 13U, 192U, _batches), Coordinates(0, 0, 192))); |
| 213 | // Layer 4 |
| 214 | conv4_out.allocator()->init(TensorInfo(TensorShape(13U, 13U, 384U, _batches), 1, _data_type, _fixed_point_position)); |
| 215 | conv41_out = std::unique_ptr<SubTensorType>(new SubTensorType(&conv4_out, TensorShape(13U, 13U, 192U, _batches), Coordinates())); |
| 216 | conv42_out = std::unique_ptr<SubTensorType>(new SubTensorType(&conv4_out, TensorShape(13U, 13U, 192U, _batches), Coordinates(0, 0, 192))); |
| 217 | act4_out.allocator()->init(TensorInfo(TensorShape(13U, 13U, 384U, _batches), 1, _data_type, _fixed_point_position)); |
| 218 | act41_out = std::unique_ptr<SubTensorType>(new SubTensorType(&act4_out, TensorShape(13U, 13U, 192U, _batches), Coordinates())); |
| 219 | act42_out = std::unique_ptr<SubTensorType>(new SubTensorType(&act4_out, TensorShape(13U, 13U, 192U, _batches), Coordinates(0, 0, 192))); |
| 220 | // Layer 5 |
| 221 | conv5_out.allocator()->init(TensorInfo(TensorShape(13U, 13U, 256U, _batches), 1, _data_type, _fixed_point_position)); |
| 222 | conv51_out = std::unique_ptr<SubTensorType>(new SubTensorType(&conv5_out, TensorShape(13U, 13U, 128U, _batches), Coordinates())); |
| 223 | conv52_out = std::unique_ptr<SubTensorType>(new SubTensorType(&conv5_out, TensorShape(13U, 13U, 128U, _batches), Coordinates(0, 0, 128))); |
| 224 | act5_out.allocator()->init(TensorInfo(TensorShape(13U, 13U, 256U, _batches), 1, _data_type, _fixed_point_position)); |
| 225 | pool5_out.allocator()->init(TensorInfo(TensorShape(6U, 6U, 256U, _batches), 1, _data_type, _fixed_point_position)); |
| 226 | // Layer 6 |
| 227 | fc6_out.allocator()->init(TensorInfo(TensorShape(4096U, _batches), 1, _data_type, _fixed_point_position)); |
| 228 | act6_out.allocator()->init(TensorInfo(TensorShape(4096U, _batches), 1, _data_type, _fixed_point_position)); |
| 229 | // Layer 7 |
| 230 | fc7_out.allocator()->init(TensorInfo(TensorShape(4096U, _batches), 1, _data_type, _fixed_point_position)); |
| 231 | act7_out.allocator()->init(TensorInfo(TensorShape(4096U, _batches), 1, _data_type, _fixed_point_position)); |
| 232 | // Layer 8 |
| 233 | fc8_out.allocator()->init(TensorInfo(TensorShape(1000U, _batches), 1, _data_type, _fixed_point_position)); |
| 234 | |
| 235 | // Configure Layers |
| 236 | // Layer 1 |
| 237 | TensorType *b0 = _reshaped_weights ? nullptr : &b[0]; |
Gian Marco Iodice | 559d771 | 2017-08-08 08:38:09 +0100 | [diff] [blame] | 238 | conv1.configure(&input, &w[0], b0, &conv1_out, PadStrideInfo(4, 4, 0, 0), WeightsInfo(_reshaped_weights, 11U, 11U, 96U)); |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 239 | act1.configure(&conv1_out, &act1_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); |
| 240 | norm1.configure(&act1_out, &norm1_out, NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f)); |
| 241 | pool1.configure(&norm1_out, &pool1_out, PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0))); |
| 242 | // Layer 2 |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 243 | conv21.configure(pool11_out.get(), w11.get(), b11.get(), conv21_out.get(), PadStrideInfo(1, 1, 2, 2), WeightsInfo(_reshaped_weights, 5U, 5U, 128U)); |
| 244 | conv22.configure(pool12_out.get(), w12.get(), b12.get(), conv22_out.get(), PadStrideInfo(1, 1, 2, 2), WeightsInfo(_reshaped_weights, 5U, 5U, 128U)); |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 245 | act2.configure(&conv2_out, &act2_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); |
| 246 | norm2.configure(&act2_out, &norm2_out, NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f)); |
| 247 | pool2.configure(&norm2_out, &pool2_out, PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0))); |
| 248 | // Layer 3 |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 249 | TensorType *b2 = (_reshaped_weights && !_is_direct_conv) ? nullptr : &b[2]; |
Gian Marco Iodice | 559d771 | 2017-08-08 08:38:09 +0100 | [diff] [blame] | 250 | conv3.configure(&pool2_out, &w[2], b2, &conv3_out, PadStrideInfo(1, 1, 1, 1), WeightsInfo(_reshaped_weights, 3U, 3U, 384U)); |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 251 | act3.configure(&conv3_out, &act3_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); |
| 252 | // Layer 4 |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 253 | conv41.configure(act31_out.get(), w31.get(), b31.get(), conv41_out.get(), PadStrideInfo(1, 1, 1, 1), WeightsInfo(_reshaped_weights, 3U, 3U, 192U)); |
| 254 | conv42.configure(act32_out.get(), w32.get(), b32.get(), conv42_out.get(), PadStrideInfo(1, 1, 1, 1), WeightsInfo(_reshaped_weights, 3U, 3U, 192U)); |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 255 | act4.configure(&conv4_out, &act4_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); |
| 256 | // Layer 5 |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 257 | conv51.configure(act41_out.get(), w41.get(), b41.get(), conv51_out.get(), PadStrideInfo(1, 1, 1, 1), WeightsInfo(_reshaped_weights, 3U, 3U, 128U)); |
| 258 | conv52.configure(act42_out.get(), w42.get(), b42.get(), conv52_out.get(), PadStrideInfo(1, 1, 1, 1), WeightsInfo(_reshaped_weights, 3U, 3U, 128U)); |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 259 | act5.configure(&conv5_out, &act5_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); |
| 260 | pool5.configure(&act5_out, &pool5_out, PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0))); |
| 261 | // Layer 6 |
| 262 | fc6.configure(&pool5_out, &w[5], &b[5], &fc6_out, true, _reshaped_weights); |
| 263 | act6.configure(&fc6_out, &act6_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); |
| 264 | // Layer 7 |
| 265 | fc7.configure(&act6_out, &w[6], &b[6], &fc7_out, true, _reshaped_weights); |
| 266 | act7.configure(&fc7_out, &act7_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); |
| 267 | // Layer 8 |
| 268 | fc8.configure(&act7_out, &w[7], &b[7], &fc8_out, true, _reshaped_weights); |
| 269 | // Softmax |
| 270 | smx.configure(&fc8_out, &output); |
| 271 | } |
| 272 | |
| 273 | void allocate() |
| 274 | { |
| 275 | input.allocator()->allocate(); |
| 276 | output.allocator()->allocate(); |
| 277 | |
| 278 | if(!_reshaped_weights) |
| 279 | { |
| 280 | for(auto &wi : w) |
| 281 | { |
| 282 | wi.allocator()->allocate(); |
| 283 | } |
| 284 | |
| 285 | for(auto &bi : b) |
| 286 | { |
| 287 | bi.allocator()->allocate(); |
| 288 | } |
| 289 | } |
| 290 | else |
| 291 | { |
| 292 | w[0].allocator()->allocate(); |
| 293 | w[2].allocator()->allocate(); |
| 294 | w[5].allocator()->allocate(); |
| 295 | w[6].allocator()->allocate(); |
| 296 | w[7].allocator()->allocate(); |
| 297 | |
| 298 | b[5].allocator()->allocate(); |
| 299 | b[6].allocator()->allocate(); |
| 300 | b[7].allocator()->allocate(); |
| 301 | |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 302 | if(!_is_direct_conv) |
| 303 | { |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 304 | dynamic_cast<TensorType *>(w11.get())->allocator()->allocate(); |
| 305 | dynamic_cast<TensorType *>(w12.get())->allocator()->allocate(); |
| 306 | dynamic_cast<TensorType *>(w31.get())->allocator()->allocate(); |
| 307 | dynamic_cast<TensorType *>(w32.get())->allocator()->allocate(); |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 308 | dynamic_cast<TensorType *>(w41.get())->allocator()->allocate(); |
| 309 | dynamic_cast<TensorType *>(w42.get())->allocator()->allocate(); |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 310 | } |
| 311 | else |
| 312 | { |
Gian Marco Iodice | 2e44868 | 2017-08-22 10:40:47 +0100 | [diff] [blame] | 313 | b[1].allocator()->allocate(); |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 314 | b[2].allocator()->allocate(); |
| 315 | b[3].allocator()->allocate(); |
| 316 | b[4].allocator()->allocate(); |
Gian Marco Iodice | 2e44868 | 2017-08-22 10:40:47 +0100 | [diff] [blame] | 317 | w[1].allocator()->allocate(); |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 318 | w[3].allocator()->allocate(); |
| 319 | w[4].allocator()->allocate(); |
| 320 | } |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 321 | } |
| 322 | |
| 323 | conv1_out.allocator()->allocate(); |
| 324 | act1_out.allocator()->allocate(); |
| 325 | norm1_out.allocator()->allocate(); |
| 326 | pool1_out.allocator()->allocate(); |
| 327 | conv2_out.allocator()->allocate(); |
| 328 | act2_out.allocator()->allocate(); |
| 329 | norm2_out.allocator()->allocate(); |
| 330 | pool2_out.allocator()->allocate(); |
| 331 | conv3_out.allocator()->allocate(); |
| 332 | act3_out.allocator()->allocate(); |
| 333 | conv4_out.allocator()->allocate(); |
| 334 | act4_out.allocator()->allocate(); |
| 335 | conv5_out.allocator()->allocate(); |
| 336 | act5_out.allocator()->allocate(); |
| 337 | pool5_out.allocator()->allocate(); |
| 338 | fc6_out.allocator()->allocate(); |
| 339 | act6_out.allocator()->allocate(); |
| 340 | fc7_out.allocator()->allocate(); |
| 341 | act7_out.allocator()->allocate(); |
| 342 | fc8_out.allocator()->allocate(); |
| 343 | } |
| 344 | |
| 345 | /** Fills the trainable parameters and input with random data. */ |
| 346 | void fill_random() |
| 347 | { |
| 348 | library->fill_tensor_uniform(Accessor(input), 0); |
| 349 | |
| 350 | if(!_reshaped_weights) |
| 351 | { |
| 352 | for(unsigned int i = 0; i < w.size(); ++i) |
| 353 | { |
| 354 | library->fill_tensor_uniform(Accessor(w[i]), i + 1); |
| 355 | library->fill_tensor_uniform(Accessor(b[i]), i + 10); |
| 356 | } |
| 357 | } |
| 358 | else |
| 359 | { |
| 360 | library->fill_tensor_uniform(Accessor(w[0]), 1); |
| 361 | library->fill_tensor_uniform(Accessor(w[2]), 2); |
| 362 | |
| 363 | library->fill_tensor_uniform(Accessor(w[5]), 3); |
| 364 | library->fill_tensor_uniform(Accessor(b[5]), 4); |
| 365 | library->fill_tensor_uniform(Accessor(w[6]), 5); |
| 366 | library->fill_tensor_uniform(Accessor(b[6]), 6); |
| 367 | library->fill_tensor_uniform(Accessor(w[7]), 7); |
| 368 | library->fill_tensor_uniform(Accessor(b[7]), 8); |
| 369 | |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 370 | if(!_is_direct_conv) |
| 371 | { |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 372 | library->fill_tensor_uniform(Accessor(*dynamic_cast<TensorType *>(w11.get())), 9); |
| 373 | library->fill_tensor_uniform(Accessor(*dynamic_cast<TensorType *>(w12.get())), 10); |
| 374 | library->fill_tensor_uniform(Accessor(*dynamic_cast<TensorType *>(w31.get())), 11); |
| 375 | library->fill_tensor_uniform(Accessor(*dynamic_cast<TensorType *>(w32.get())), 12); |
| 376 | library->fill_tensor_uniform(Accessor(*dynamic_cast<TensorType *>(w41.get())), 13); |
| 377 | library->fill_tensor_uniform(Accessor(*dynamic_cast<TensorType *>(w42.get())), 14); |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 378 | } |
| 379 | else |
| 380 | { |
Gian Marco Iodice | 2e44868 | 2017-08-22 10:40:47 +0100 | [diff] [blame] | 381 | library->fill_tensor_uniform(Accessor(w[1]), 9); |
| 382 | library->fill_tensor_uniform(Accessor(b[1]), 10); |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 383 | library->fill_tensor_uniform(Accessor(w[3]), 11); |
| 384 | library->fill_tensor_uniform(Accessor(b[3]), 12); |
| 385 | library->fill_tensor_uniform(Accessor(w[4]), 13); |
| 386 | library->fill_tensor_uniform(Accessor(b[4]), 14); |
| 387 | } |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 388 | } |
| 389 | } |
| 390 | |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 391 | /** Fills the trainable parameters from binary files |
| 392 | * |
| 393 | * @param weights Files names containing the weights data |
| 394 | * @param biases Files names containing the bias data |
| 395 | */ |
| 396 | void fill(std::vector<std::string> weights, std::vector<std::string> biases) |
| 397 | { |
| 398 | ARM_COMPUTE_ERROR_ON(weights.size() != w.size()); |
| 399 | ARM_COMPUTE_ERROR_ON(biases.size() != b.size()); |
| 400 | ARM_COMPUTE_ERROR_ON(_reshaped_weights); |
| 401 | |
| 402 | for(unsigned int i = 0; i < weights.size(); ++i) |
| 403 | { |
| 404 | library->fill_layer_data(Accessor(w[i]), weights[i]); |
| 405 | library->fill_layer_data(Accessor(b[i]), biases[i]); |
| 406 | } |
| 407 | } |
| 408 | |
| 409 | /** Feed input to network from file. |
| 410 | * |
| 411 | * @param name File name of containing the input data. |
| 412 | */ |
| 413 | void feed(std::string name) |
| 414 | { |
| 415 | library->fill_layer_data(Accessor(input), name); |
| 416 | } |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 417 | |
| 418 | /** Get the classification results. |
| 419 | * |
| 420 | * @return Vector containing the classified labels |
| 421 | */ |
| 422 | std::vector<unsigned int> get_classifications() |
| 423 | { |
| 424 | std::vector<unsigned int> classified_labels; |
| 425 | Accessor output_accessor(output); |
| 426 | |
| 427 | Window window; |
| 428 | window.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 429 | for(unsigned int d = 1; d < output_accessor.shape().num_dimensions(); ++d) |
| 430 | { |
| 431 | window.set(d, Window::Dimension(0, output_accessor.shape()[d], 1)); |
| 432 | } |
| 433 | |
| 434 | execute_window_loop(window, [&](const Coordinates & id) |
| 435 | { |
| 436 | int max_idx = 0; |
| 437 | float val = 0; |
| 438 | const void *const out_ptr = output_accessor(id); |
| 439 | for(unsigned int l = 0; l < output_accessor.shape().x(); ++l) |
| 440 | { |
| 441 | float curr_val = reinterpret_cast<const float *>(out_ptr)[l]; |
| 442 | if(curr_val > val) |
| 443 | { |
| 444 | max_idx = l; |
| 445 | val = curr_val; |
| 446 | } |
| 447 | } |
| 448 | classified_labels.push_back(max_idx); |
| 449 | }); |
| 450 | return classified_labels; |
| 451 | } |
| 452 | |
| 453 | /** Clear all allocated memory from the tensor objects */ |
| 454 | void clear() |
| 455 | { |
| 456 | // Free allocations |
| 457 | input.allocator()->free(); |
| 458 | output.allocator()->free(); |
| 459 | |
| 460 | if(!_reshaped_weights) |
| 461 | { |
| 462 | for(auto &wi : w) |
| 463 | { |
| 464 | wi.allocator()->free(); |
| 465 | } |
| 466 | |
| 467 | for(auto &bi : b) |
| 468 | { |
| 469 | bi.allocator()->free(); |
| 470 | } |
| 471 | } |
| 472 | else |
| 473 | { |
| 474 | w[0].allocator()->free(); |
| 475 | w[2].allocator()->free(); |
| 476 | w[5].allocator()->free(); |
| 477 | w[6].allocator()->free(); |
| 478 | w[7].allocator()->free(); |
| 479 | |
| 480 | b[5].allocator()->free(); |
| 481 | b[6].allocator()->free(); |
| 482 | b[7].allocator()->free(); |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 483 | |
| 484 | if(_is_direct_conv) |
| 485 | { |
| 486 | w[3].allocator()->free(); |
| 487 | w[4].allocator()->free(); |
| 488 | b[2].allocator()->free(); |
| 489 | b[3].allocator()->free(); |
| 490 | b[4].allocator()->free(); |
| 491 | } |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 492 | } |
| 493 | |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 494 | w11.reset(); |
| 495 | w12.reset(); |
| 496 | b11.reset(); |
| 497 | b11.reset(); |
| 498 | w31.reset(); |
| 499 | w32.reset(); |
| 500 | b31.reset(); |
| 501 | b32.reset(); |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 502 | w41.reset(); |
| 503 | w42.reset(); |
| 504 | b41.reset(); |
| 505 | b42.reset(); |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 506 | |
| 507 | conv1_out.allocator()->free(); |
| 508 | act1_out.allocator()->free(); |
| 509 | norm1_out.allocator()->free(); |
| 510 | pool1_out.allocator()->free(); |
| 511 | conv2_out.allocator()->free(); |
| 512 | act2_out.allocator()->free(); |
| 513 | norm2_out.allocator()->free(); |
| 514 | pool2_out.allocator()->free(); |
| 515 | conv3_out.allocator()->free(); |
| 516 | act3_out.allocator()->free(); |
| 517 | conv4_out.allocator()->free(); |
| 518 | act4_out.allocator()->free(); |
| 519 | conv5_out.allocator()->free(); |
| 520 | act5_out.allocator()->free(); |
| 521 | pool5_out.allocator()->free(); |
| 522 | fc6_out.allocator()->free(); |
| 523 | act6_out.allocator()->free(); |
| 524 | fc7_out.allocator()->free(); |
| 525 | act7_out.allocator()->free(); |
| 526 | fc8_out.allocator()->free(); |
| 527 | } |
| 528 | |
| 529 | /** Runs the model */ |
| 530 | void run() |
| 531 | { |
| 532 | // Layer 1 |
| 533 | conv1.run(); |
| 534 | act1.run(); |
| 535 | norm1.run(); |
| 536 | pool1.run(); |
| 537 | // Layer 2 |
| 538 | conv21.run(); |
| 539 | conv22.run(); |
| 540 | act2.run(); |
| 541 | norm2.run(); |
| 542 | pool2.run(); |
| 543 | // Layer 3 |
| 544 | conv3.run(); |
| 545 | act3.run(); |
| 546 | // Layer 4 |
| 547 | conv41.run(); |
| 548 | conv42.run(); |
| 549 | act4.run(); |
| 550 | // Layer 5 |
| 551 | conv51.run(); |
| 552 | conv52.run(); |
| 553 | act5.run(); |
| 554 | pool5.run(); |
| 555 | // Layer 6 |
| 556 | fc6.run(); |
| 557 | act6.run(); |
| 558 | // Layer 7 |
| 559 | fc7.run(); |
| 560 | act7.run(); |
| 561 | // Layer 8 |
| 562 | fc8.run(); |
| 563 | // Softmax |
| 564 | smx.run(); |
| 565 | } |
| 566 | |
Joel Liang | 1c5ffd6 | 2017-12-28 10:09:51 +0800 | [diff] [blame] | 567 | /** Sync the results */ |
| 568 | void sync() |
| 569 | { |
| 570 | sync_if_necessary<TensorType>(); |
| 571 | sync_tensor_if_necessary<TensorType>(output); |
| 572 | } |
| 573 | |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 574 | private: |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 575 | struct DirectConv |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 576 | { |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 577 | template <typename ConvolutionLayerFunction1 = ConvolutionLayerFunction, typename DirectConvolutionLayerFunction1 = DirectConvolutionLayerFunction> |
| 578 | typename std::enable_if < !std::is_same<ConvolutionLayerFunction1, DirectConvolutionLayerFunction1>::value, void >::type |
| 579 | configure(ITensorType *input, const ITensorType *weights, const ITensorType *biases, ITensorType *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info = WeightsInfo()) |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 580 | { |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 581 | _func.configure(input, weights, biases, output, conv_info); |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 582 | } |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 583 | |
| 584 | template <typename ConvolutionLayerFunction1 = ConvolutionLayerFunction, typename DirectConvolutionLayerFunction1 = DirectConvolutionLayerFunction> |
| 585 | typename std::enable_if<std::is_same<ConvolutionLayerFunction1, DirectConvolutionLayerFunction1>::value, void>::type |
| 586 | configure(ITensorType *input, const ITensorType *weights, const ITensorType *biases, ITensorType *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info = WeightsInfo()) |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 587 | { |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 588 | _func.configure(input, weights, biases, output, conv_info, weights_info); |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 589 | } |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 590 | |
| 591 | void run() |
| 592 | { |
| 593 | _func.run(); |
| 594 | } |
| 595 | |
| 596 | DirectConvolutionLayerFunction _func{}; |
| 597 | }; |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 598 | |
| 599 | DataType _data_type{ DataType::UNKNOWN }; |
| 600 | int _fixed_point_position{ 0 }; |
| 601 | unsigned int _batches{ 0 }; |
| 602 | bool _reshaped_weights{ false }; |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame] | 603 | bool _is_direct_conv{ !std::is_same<ConvolutionLayerFunction, DirectConvolutionLayerFunction>::value }; |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 604 | |
| 605 | ActivationLayerFunction act1{}, act2{}, act3{}, act4{}, act5{}, act6{}, act7{}; |
Gian Marco Iodice | 2e44868 | 2017-08-22 10:40:47 +0100 | [diff] [blame] | 606 | ConvolutionLayerFunction conv1{}; |
| 607 | DirectConv conv21{}, conv22{}, conv3{}, conv41{}, conv42{}, conv51{}, conv52{}; |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 608 | FullyConnectedLayerFunction fc6{}, fc7{}, fc8{}; |
| 609 | NormalizationLayerFunction norm1{}, norm2{}; |
| 610 | PoolingLayerFunction pool1{}, pool2{}, pool5{}; |
| 611 | SoftmaxLayerFunction smx{}; |
| 612 | |
| 613 | TensorType input{}, output{}; |
| 614 | std::array<TensorType, 8> w{ {} }, b{ {} }; |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 615 | std::unique_ptr<ITensorType> w11{ nullptr }, w12{ nullptr }, b11{ nullptr }, b12{ nullptr }; |
| 616 | std::unique_ptr<ITensorType> w31{ nullptr }, w32{ nullptr }, b31{ nullptr }, b32{ nullptr }; |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 617 | std::unique_ptr<ITensorType> w41{ nullptr }, w42{ nullptr }, b41{ nullptr }, b42{ nullptr }; |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 618 | |
| 619 | TensorType conv1_out{}, act1_out{}, norm1_out{}, pool1_out{}; |
| 620 | TensorType conv2_out{}, act2_out{}, pool2_out{}, norm2_out{}; |
| 621 | TensorType conv3_out{}, act3_out{}; |
| 622 | TensorType conv4_out{}, act4_out{}; |
| 623 | TensorType conv5_out{}, act5_out{}, pool5_out{}; |
| 624 | TensorType fc6_out{}, act6_out{}; |
| 625 | TensorType fc7_out{}, act7_out{}; |
| 626 | TensorType fc8_out{}; |
| 627 | |
| 628 | std::unique_ptr<SubTensorType> pool11_out{}, pool12_out{}; |
| 629 | std::unique_ptr<SubTensorType> conv21_out{}, conv22_out{}; |
| 630 | std::unique_ptr<SubTensorType> act31_out{}, act32_out{}; |
| 631 | std::unique_ptr<SubTensorType> conv41_out{}, conv42_out{}, act41_out{}, act42_out{}; |
| 632 | std::unique_ptr<SubTensorType> conv51_out{}, conv52_out{}; |
| 633 | }; |
| 634 | } // namespace networks |
| 635 | } // namespace test |
| 636 | } // namespace arm_compute |
| 637 | #endif //__ARM_COMPUTE_TEST_MODEL_OBJECTS_ALEXNET_H__ |