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 | fb5aabb | 2017-07-18 14:39:55 +0100 | [diff] [blame] | 27 | #include "AssetsLibrary.h" |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 28 | #include "Globals.h" |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 29 | #include "Utils.h" |
| 30 | |
| 31 | #include <memory> |
| 32 | |
| 33 | namespace arm_compute |
| 34 | { |
| 35 | namespace test |
| 36 | { |
| 37 | namespace networks |
| 38 | { |
| 39 | /** AlexNet model object */ |
| 40 | template <typename ITensorType, |
| 41 | typename TensorType, |
| 42 | typename SubTensorType, |
| 43 | typename Accessor, |
| 44 | typename ActivationLayerFunction, |
| 45 | typename ConvolutionLayerFunction, |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame^] | 46 | typename DirectConvolutionLayerFunction, |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 47 | typename FullyConnectedLayerFunction, |
| 48 | typename NormalizationLayerFunction, |
| 49 | typename PoolingLayerFunction, |
| 50 | typename SoftmaxLayerFunction> |
| 51 | class AlexNetNetwork |
| 52 | { |
| 53 | public: |
| 54 | void init(DataType data_type, int fixed_point_position, int batches, bool reshaped_weights = false) |
| 55 | { |
| 56 | _data_type = data_type; |
| 57 | _fixed_point_position = fixed_point_position; |
| 58 | _batches = batches; |
| 59 | _reshaped_weights = reshaped_weights; |
| 60 | |
| 61 | // Initialize weights and biases |
| 62 | if(!_reshaped_weights) |
| 63 | { |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame^] | 64 | w[0].allocator()->init(TensorInfo(TensorShape(11U, 11U, 3U, 96U), 1, _data_type, _fixed_point_position)); |
| 65 | b[0].allocator()->init(TensorInfo(TensorShape(96U), 1, _data_type, _fixed_point_position)); |
| 66 | w[1].allocator()->init(TensorInfo(TensorShape(5U, 5U, 48U, 256U), 1, _data_type, _fixed_point_position)); |
| 67 | b[1].allocator()->init(TensorInfo(TensorShape(256U), 1, _data_type, _fixed_point_position)); |
| 68 | w[2].allocator()->init(TensorInfo(TensorShape(3U, 3U, 256U, 384U), 1, _data_type, _fixed_point_position)); |
| 69 | b[2].allocator()->init(TensorInfo(TensorShape(384U), 1, _data_type, _fixed_point_position)); |
| 70 | w[3].allocator()->init(TensorInfo(TensorShape(3U, 3U, 192U, 384U), 1, _data_type, _fixed_point_position)); |
| 71 | b[3].allocator()->init(TensorInfo(TensorShape(384U), 1, _data_type, _fixed_point_position)); |
| 72 | w[4].allocator()->init(TensorInfo(TensorShape(3U, 3U, 192U, 256U), 1, _data_type, _fixed_point_position)); |
| 73 | b[4].allocator()->init(TensorInfo(TensorShape(256U), 1, _data_type, _fixed_point_position)); |
| 74 | w[5].allocator()->init(TensorInfo(TensorShape(9216U, 4096U), 1, _data_type, _fixed_point_position)); |
| 75 | b[5].allocator()->init(TensorInfo(TensorShape(4096U), 1, _data_type, _fixed_point_position)); |
| 76 | w[6].allocator()->init(TensorInfo(TensorShape(4096U, 4096U), 1, _data_type, _fixed_point_position)); |
| 77 | b[6].allocator()->init(TensorInfo(TensorShape(4096U), 1, _data_type, _fixed_point_position)); |
| 78 | w[7].allocator()->init(TensorInfo(TensorShape(4096U, 1000U), 1, _data_type, _fixed_point_position)); |
| 79 | b[7].allocator()->init(TensorInfo(TensorShape(1000U), 1, _data_type, _fixed_point_position)); |
| 80 | |
| 81 | w21 = std::unique_ptr<SubTensorType>(new SubTensorType(&w[1], TensorShape(5U, 5U, 48U, 128U), Coordinates())); |
| 82 | w22 = std::unique_ptr<SubTensorType>(new SubTensorType(&w[1], TensorShape(5U, 5U, 48U, 128U), Coordinates(0, 0, 0, 128))); |
| 83 | b21 = std::unique_ptr<SubTensorType>(new SubTensorType(&b[1], TensorShape(128U), Coordinates())); |
| 84 | b22 = std::unique_ptr<SubTensorType>(new SubTensorType(&b[1], TensorShape(128U), Coordinates(128))); |
| 85 | |
| 86 | w41 = std::unique_ptr<SubTensorType>(new SubTensorType(&w[3], TensorShape(3U, 3U, 192U, 192U), Coordinates())); |
| 87 | w42 = std::unique_ptr<SubTensorType>(new SubTensorType(&w[3], TensorShape(3U, 3U, 192U, 192U), Coordinates(0, 0, 0, 192))); |
| 88 | b41 = std::unique_ptr<SubTensorType>(new SubTensorType(&b[3], TensorShape(192U), Coordinates())); |
| 89 | b42 = std::unique_ptr<SubTensorType>(new SubTensorType(&b[3], TensorShape(192U), Coordinates(192))); |
| 90 | |
| 91 | w51 = std::unique_ptr<SubTensorType>(new SubTensorType(&w[4], TensorShape(3U, 3U, 192U, 128U), Coordinates())); |
| 92 | w52 = std::unique_ptr<SubTensorType>(new SubTensorType(&w[4], TensorShape(3U, 3U, 192U, 128U), Coordinates(0, 0, 0, 128))); |
| 93 | b51 = std::unique_ptr<SubTensorType>(new SubTensorType(&b[4], TensorShape(128U), Coordinates())); |
| 94 | b52 = std::unique_ptr<SubTensorType>(new SubTensorType(&b[4], TensorShape(128U), Coordinates(128))); |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 95 | } |
| 96 | else |
| 97 | { |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame^] | 98 | const unsigned int data_type_size = 16 / arm_compute::data_size_from_type(_data_type); |
| 99 | |
| 100 | // Create tensor for the reshaped weights |
| 101 | auto w21_tensor = std::unique_ptr<TensorType>(new TensorType()); |
| 102 | auto w22_tensor = std::unique_ptr<TensorType>(new TensorType()); |
| 103 | |
| 104 | w[0].allocator()->init(TensorInfo(TensorShape(366U * data_type_size, 96U / data_type_size), 1, _data_type, _fixed_point_position)); |
| 105 | w21_tensor->allocator()->init(TensorInfo(TensorShape(1248U * data_type_size, 128U / data_type_size), 1, _data_type, _fixed_point_position)); |
| 106 | w22_tensor->allocator()->init(TensorInfo(TensorShape(1248U * data_type_size, 128U / data_type_size), 1, _data_type, _fixed_point_position)); |
| 107 | w21 = std::move(w21_tensor); |
| 108 | w22 = std::move(w22_tensor); |
| 109 | |
| 110 | // Configure the direct convolution's weights. Direct convolution doesn't need reshape weights |
| 111 | if(!_is_direct_conv) |
| 112 | { |
| 113 | auto w41_tensor = std::unique_ptr<TensorType>(new TensorType()); |
| 114 | auto w42_tensor = std::unique_ptr<TensorType>(new TensorType()); |
| 115 | auto w51_tensor = std::unique_ptr<TensorType>(new TensorType()); |
| 116 | auto w52_tensor = std::unique_ptr<TensorType>(new TensorType()); |
| 117 | w41_tensor->allocator()->init(TensorInfo(TensorShape(1920U * data_type_size, 192U / data_type_size), 1, _data_type, _fixed_point_position)); |
| 118 | w42_tensor->allocator()->init(TensorInfo(TensorShape(1920U * data_type_size, 192U / data_type_size), 1, _data_type, _fixed_point_position)); |
| 119 | w51_tensor->allocator()->init(TensorInfo(TensorShape(1920U * data_type_size, 128U / data_type_size), 1, _data_type, _fixed_point_position)); |
| 120 | w52_tensor->allocator()->init(TensorInfo(TensorShape(1920U * data_type_size, 128U / data_type_size), 1, _data_type, _fixed_point_position)); |
| 121 | w[2].allocator()->init(TensorInfo(TensorShape(2560U * data_type_size, 384U / data_type_size), 1, _data_type, _fixed_point_position)); |
| 122 | w41 = std::move(w41_tensor); |
| 123 | w42 = std::move(w42_tensor); |
| 124 | w51 = std::move(w51_tensor); |
| 125 | w52 = std::move(w52_tensor); |
| 126 | } |
| 127 | else |
| 128 | { |
| 129 | w[2].allocator()->init(TensorInfo(TensorShape(3U, 3U, 256U, 384U), 1, _data_type, _fixed_point_position)); |
| 130 | b[2].allocator()->init(TensorInfo(TensorShape(384U), 1, _data_type, _fixed_point_position)); |
| 131 | w[3].allocator()->init(TensorInfo(TensorShape(3U, 3U, 192U, 384U), 1, _data_type, _fixed_point_position)); |
| 132 | b[3].allocator()->init(TensorInfo(TensorShape(384U), 1, _data_type, _fixed_point_position)); |
| 133 | w[4].allocator()->init(TensorInfo(TensorShape(3U, 3U, 192U, 256U), 1, _data_type, _fixed_point_position)); |
| 134 | b[4].allocator()->init(TensorInfo(TensorShape(256U), 1, _data_type, _fixed_point_position)); |
| 135 | w41 = std::unique_ptr<SubTensorType>(new SubTensorType(&w[3], TensorShape(3U, 3U, 192U, 192U), Coordinates())); |
| 136 | w42 = std::unique_ptr<SubTensorType>(new SubTensorType(&w[3], TensorShape(3U, 3U, 192U, 192U), Coordinates(0, 0, 0, 192))); |
| 137 | b41 = std::unique_ptr<SubTensorType>(new SubTensorType(&b[3], TensorShape(192U), Coordinates())); |
| 138 | b42 = std::unique_ptr<SubTensorType>(new SubTensorType(&b[3], TensorShape(192U), Coordinates(192))); |
| 139 | |
| 140 | w51 = std::unique_ptr<SubTensorType>(new SubTensorType(&w[4], TensorShape(3U, 3U, 192U, 128U), Coordinates())); |
| 141 | w52 = std::unique_ptr<SubTensorType>(new SubTensorType(&w[4], TensorShape(3U, 3U, 192U, 128U), Coordinates(0, 0, 0, 128))); |
| 142 | b51 = std::unique_ptr<SubTensorType>(new SubTensorType(&b[4], TensorShape(128U), Coordinates())); |
| 143 | b52 = std::unique_ptr<SubTensorType>(new SubTensorType(&b[4], TensorShape(128U), Coordinates(128))); |
| 144 | } |
| 145 | |
| 146 | b[5].allocator()->init(TensorInfo(TensorShape(4096U), 1, _data_type, _fixed_point_position)); |
| 147 | b[6].allocator()->init(TensorInfo(TensorShape(4096U), 1, _data_type, _fixed_point_position)); |
| 148 | b[7].allocator()->init(TensorInfo(TensorShape(1000U), 1, _data_type, _fixed_point_position)); |
| 149 | |
| 150 | if(_batches > 1) |
| 151 | { |
| 152 | w[5].allocator()->init(TensorInfo(TensorShape(9216U * data_type_size, 4096U / data_type_size), 1, _data_type, _fixed_point_position)); |
| 153 | w[6].allocator()->init(TensorInfo(TensorShape(4096U * data_type_size, 4096U / data_type_size), 1, _data_type, _fixed_point_position)); |
| 154 | w[7].allocator()->init(TensorInfo(TensorShape(4096U * data_type_size, 1000U / data_type_size), 1, _data_type, _fixed_point_position)); |
| 155 | } |
| 156 | else |
| 157 | { |
| 158 | w[5].allocator()->init(TensorInfo(TensorShape(4096U, 9216U), 1, _data_type, _fixed_point_position)); |
| 159 | w[6].allocator()->init(TensorInfo(TensorShape(4096U, 4096U), 1, _data_type, _fixed_point_position)); |
| 160 | w[7].allocator()->init(TensorInfo(TensorShape(1000U, 4096U), 1, _data_type, _fixed_point_position)); |
| 161 | } |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 162 | } |
| 163 | } |
| 164 | |
| 165 | void build() |
| 166 | { |
| 167 | input.allocator()->init(TensorInfo(TensorShape(227U, 227U, 3U, _batches), 1, _data_type, _fixed_point_position)); |
| 168 | output.allocator()->init(TensorInfo(TensorShape(1000U, _batches), 1, _data_type, _fixed_point_position)); |
| 169 | |
| 170 | // Initialize intermediate tensors |
| 171 | // Layer 1 |
| 172 | conv1_out.allocator()->init(TensorInfo(TensorShape(55U, 55U, 96U, _batches), 1, _data_type, _fixed_point_position)); |
| 173 | act1_out.allocator()->init(TensorInfo(TensorShape(55U, 55U, 96U, _batches), 1, _data_type, _fixed_point_position)); |
| 174 | norm1_out.allocator()->init(TensorInfo(TensorShape(55U, 55U, 96U, _batches), 1, _data_type, _fixed_point_position)); |
| 175 | pool1_out.allocator()->init(TensorInfo(TensorShape(27U, 27U, 96U, _batches), 1, _data_type, _fixed_point_position)); |
| 176 | pool11_out = std::unique_ptr<SubTensorType>(new SubTensorType(&pool1_out, TensorShape(27U, 27U, 48U, _batches), Coordinates())); |
| 177 | pool12_out = std::unique_ptr<SubTensorType>(new SubTensorType(&pool1_out, TensorShape(27U, 27U, 48U, _batches), Coordinates(0, 0, 48))); |
| 178 | // Layer 2 |
| 179 | conv2_out.allocator()->init(TensorInfo(TensorShape(27U, 27U, 256U, _batches), 1, _data_type, _fixed_point_position)); |
| 180 | conv21_out = std::unique_ptr<SubTensorType>(new SubTensorType(&conv2_out, TensorShape(27U, 27U, 128U, _batches), Coordinates())); |
| 181 | conv22_out = std::unique_ptr<SubTensorType>(new SubTensorType(&conv2_out, TensorShape(27U, 27U, 128U, _batches), Coordinates(0, 0, 128))); |
| 182 | act2_out.allocator()->init(TensorInfo(TensorShape(27U, 27U, 256U, _batches), 1, _data_type, _fixed_point_position)); |
| 183 | norm2_out.allocator()->init(TensorInfo(TensorShape(27U, 27U, 256U, _batches), 1, _data_type, _fixed_point_position)); |
| 184 | pool2_out.allocator()->init(TensorInfo(TensorShape(13U, 13U, 256U, _batches), 1, _data_type, _fixed_point_position)); |
| 185 | // Layer 3 |
| 186 | conv3_out.allocator()->init(TensorInfo(TensorShape(13U, 13U, 384U, _batches), 1, _data_type, _fixed_point_position)); |
| 187 | act3_out.allocator()->init(TensorInfo(TensorShape(13U, 13U, 384U, _batches), 1, _data_type, _fixed_point_position)); |
| 188 | act31_out = std::unique_ptr<SubTensorType>(new SubTensorType(&act3_out, TensorShape(13U, 13U, 192U, _batches), Coordinates())); |
| 189 | act32_out = std::unique_ptr<SubTensorType>(new SubTensorType(&act3_out, TensorShape(13U, 13U, 192U, _batches), Coordinates(0, 0, 192))); |
| 190 | // Layer 4 |
| 191 | conv4_out.allocator()->init(TensorInfo(TensorShape(13U, 13U, 384U, _batches), 1, _data_type, _fixed_point_position)); |
| 192 | conv41_out = std::unique_ptr<SubTensorType>(new SubTensorType(&conv4_out, TensorShape(13U, 13U, 192U, _batches), Coordinates())); |
| 193 | conv42_out = std::unique_ptr<SubTensorType>(new SubTensorType(&conv4_out, TensorShape(13U, 13U, 192U, _batches), Coordinates(0, 0, 192))); |
| 194 | act4_out.allocator()->init(TensorInfo(TensorShape(13U, 13U, 384U, _batches), 1, _data_type, _fixed_point_position)); |
| 195 | act41_out = std::unique_ptr<SubTensorType>(new SubTensorType(&act4_out, TensorShape(13U, 13U, 192U, _batches), Coordinates())); |
| 196 | act42_out = std::unique_ptr<SubTensorType>(new SubTensorType(&act4_out, TensorShape(13U, 13U, 192U, _batches), Coordinates(0, 0, 192))); |
| 197 | // Layer 5 |
| 198 | conv5_out.allocator()->init(TensorInfo(TensorShape(13U, 13U, 256U, _batches), 1, _data_type, _fixed_point_position)); |
| 199 | conv51_out = std::unique_ptr<SubTensorType>(new SubTensorType(&conv5_out, TensorShape(13U, 13U, 128U, _batches), Coordinates())); |
| 200 | conv52_out = std::unique_ptr<SubTensorType>(new SubTensorType(&conv5_out, TensorShape(13U, 13U, 128U, _batches), Coordinates(0, 0, 128))); |
| 201 | act5_out.allocator()->init(TensorInfo(TensorShape(13U, 13U, 256U, _batches), 1, _data_type, _fixed_point_position)); |
| 202 | pool5_out.allocator()->init(TensorInfo(TensorShape(6U, 6U, 256U, _batches), 1, _data_type, _fixed_point_position)); |
| 203 | // Layer 6 |
| 204 | fc6_out.allocator()->init(TensorInfo(TensorShape(4096U, _batches), 1, _data_type, _fixed_point_position)); |
| 205 | act6_out.allocator()->init(TensorInfo(TensorShape(4096U, _batches), 1, _data_type, _fixed_point_position)); |
| 206 | // Layer 7 |
| 207 | fc7_out.allocator()->init(TensorInfo(TensorShape(4096U, _batches), 1, _data_type, _fixed_point_position)); |
| 208 | act7_out.allocator()->init(TensorInfo(TensorShape(4096U, _batches), 1, _data_type, _fixed_point_position)); |
| 209 | // Layer 8 |
| 210 | fc8_out.allocator()->init(TensorInfo(TensorShape(1000U, _batches), 1, _data_type, _fixed_point_position)); |
| 211 | |
| 212 | // Configure Layers |
| 213 | // Layer 1 |
| 214 | TensorType *b0 = _reshaped_weights ? nullptr : &b[0]; |
| 215 | conv1.configure(&input, &w[0], b0, &conv1_out, PadStrideInfo(4, 4, 0, 0), WeightsInfo(_reshaped_weights, 11U, 11U)); |
| 216 | act1.configure(&conv1_out, &act1_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); |
| 217 | norm1.configure(&act1_out, &norm1_out, NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f)); |
| 218 | pool1.configure(&norm1_out, &pool1_out, PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0))); |
| 219 | // Layer 2 |
| 220 | conv21.configure(pool11_out.get(), w21.get(), b21.get(), conv21_out.get(), PadStrideInfo(1, 1, 2, 2), WeightsInfo(_reshaped_weights, 5U, 5U)); |
| 221 | conv22.configure(pool12_out.get(), w22.get(), b22.get(), conv22_out.get(), PadStrideInfo(1, 1, 2, 2), WeightsInfo(_reshaped_weights, 5U, 5U)); |
| 222 | act2.configure(&conv2_out, &act2_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); |
| 223 | norm2.configure(&act2_out, &norm2_out, NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f)); |
| 224 | pool2.configure(&norm2_out, &pool2_out, PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0))); |
| 225 | // Layer 3 |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame^] | 226 | TensorType *b2 = (_reshaped_weights && !_is_direct_conv) ? nullptr : &b[2]; |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 227 | conv3.configure(&pool2_out, &w[2], b2, &conv3_out, PadStrideInfo(1, 1, 1, 1), WeightsInfo(_reshaped_weights, 3U, 3U)); |
| 228 | act3.configure(&conv3_out, &act3_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); |
| 229 | // Layer 4 |
| 230 | conv41.configure(act31_out.get(), w41.get(), b41.get(), conv41_out.get(), PadStrideInfo(1, 1, 1, 1), WeightsInfo(_reshaped_weights, 3U, 3U)); |
| 231 | conv42.configure(act32_out.get(), w42.get(), b42.get(), conv42_out.get(), PadStrideInfo(1, 1, 1, 1), WeightsInfo(_reshaped_weights, 3U, 3U)); |
| 232 | act4.configure(&conv4_out, &act4_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); |
| 233 | // Layer 5 |
| 234 | conv51.configure(act41_out.get(), w51.get(), b51.get(), conv51_out.get(), PadStrideInfo(1, 1, 1, 1), WeightsInfo(_reshaped_weights, 3U, 3U)); |
| 235 | conv52.configure(act42_out.get(), w52.get(), b52.get(), conv52_out.get(), PadStrideInfo(1, 1, 1, 1), WeightsInfo(_reshaped_weights, 3U, 3U)); |
| 236 | act5.configure(&conv5_out, &act5_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); |
| 237 | pool5.configure(&act5_out, &pool5_out, PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0))); |
| 238 | // Layer 6 |
| 239 | fc6.configure(&pool5_out, &w[5], &b[5], &fc6_out, true, _reshaped_weights); |
| 240 | act6.configure(&fc6_out, &act6_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); |
| 241 | // Layer 7 |
| 242 | fc7.configure(&act6_out, &w[6], &b[6], &fc7_out, true, _reshaped_weights); |
| 243 | act7.configure(&fc7_out, &act7_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); |
| 244 | // Layer 8 |
| 245 | fc8.configure(&act7_out, &w[7], &b[7], &fc8_out, true, _reshaped_weights); |
| 246 | // Softmax |
| 247 | smx.configure(&fc8_out, &output); |
| 248 | } |
| 249 | |
| 250 | void allocate() |
| 251 | { |
| 252 | input.allocator()->allocate(); |
| 253 | output.allocator()->allocate(); |
| 254 | |
| 255 | if(!_reshaped_weights) |
| 256 | { |
| 257 | for(auto &wi : w) |
| 258 | { |
| 259 | wi.allocator()->allocate(); |
| 260 | } |
| 261 | |
| 262 | for(auto &bi : b) |
| 263 | { |
| 264 | bi.allocator()->allocate(); |
| 265 | } |
| 266 | } |
| 267 | else |
| 268 | { |
| 269 | w[0].allocator()->allocate(); |
| 270 | w[2].allocator()->allocate(); |
| 271 | w[5].allocator()->allocate(); |
| 272 | w[6].allocator()->allocate(); |
| 273 | w[7].allocator()->allocate(); |
| 274 | |
| 275 | b[5].allocator()->allocate(); |
| 276 | b[6].allocator()->allocate(); |
| 277 | b[7].allocator()->allocate(); |
| 278 | |
| 279 | dynamic_cast<TensorType *>(w21.get())->allocator()->allocate(); |
| 280 | dynamic_cast<TensorType *>(w22.get())->allocator()->allocate(); |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame^] | 281 | if(!_is_direct_conv) |
| 282 | { |
| 283 | dynamic_cast<TensorType *>(w41.get())->allocator()->allocate(); |
| 284 | dynamic_cast<TensorType *>(w42.get())->allocator()->allocate(); |
| 285 | dynamic_cast<TensorType *>(w51.get())->allocator()->allocate(); |
| 286 | dynamic_cast<TensorType *>(w52.get())->allocator()->allocate(); |
| 287 | } |
| 288 | else |
| 289 | { |
| 290 | b[2].allocator()->allocate(); |
| 291 | b[3].allocator()->allocate(); |
| 292 | b[4].allocator()->allocate(); |
| 293 | w[3].allocator()->allocate(); |
| 294 | w[4].allocator()->allocate(); |
| 295 | } |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 296 | } |
| 297 | |
| 298 | conv1_out.allocator()->allocate(); |
| 299 | act1_out.allocator()->allocate(); |
| 300 | norm1_out.allocator()->allocate(); |
| 301 | pool1_out.allocator()->allocate(); |
| 302 | conv2_out.allocator()->allocate(); |
| 303 | act2_out.allocator()->allocate(); |
| 304 | norm2_out.allocator()->allocate(); |
| 305 | pool2_out.allocator()->allocate(); |
| 306 | conv3_out.allocator()->allocate(); |
| 307 | act3_out.allocator()->allocate(); |
| 308 | conv4_out.allocator()->allocate(); |
| 309 | act4_out.allocator()->allocate(); |
| 310 | conv5_out.allocator()->allocate(); |
| 311 | act5_out.allocator()->allocate(); |
| 312 | pool5_out.allocator()->allocate(); |
| 313 | fc6_out.allocator()->allocate(); |
| 314 | act6_out.allocator()->allocate(); |
| 315 | fc7_out.allocator()->allocate(); |
| 316 | act7_out.allocator()->allocate(); |
| 317 | fc8_out.allocator()->allocate(); |
| 318 | } |
| 319 | |
| 320 | /** Fills the trainable parameters and input with random data. */ |
| 321 | void fill_random() |
| 322 | { |
| 323 | library->fill_tensor_uniform(Accessor(input), 0); |
| 324 | |
| 325 | if(!_reshaped_weights) |
| 326 | { |
| 327 | for(unsigned int i = 0; i < w.size(); ++i) |
| 328 | { |
| 329 | library->fill_tensor_uniform(Accessor(w[i]), i + 1); |
| 330 | library->fill_tensor_uniform(Accessor(b[i]), i + 10); |
| 331 | } |
| 332 | } |
| 333 | else |
| 334 | { |
| 335 | library->fill_tensor_uniform(Accessor(w[0]), 1); |
| 336 | library->fill_tensor_uniform(Accessor(w[2]), 2); |
| 337 | |
| 338 | library->fill_tensor_uniform(Accessor(w[5]), 3); |
| 339 | library->fill_tensor_uniform(Accessor(b[5]), 4); |
| 340 | library->fill_tensor_uniform(Accessor(w[6]), 5); |
| 341 | library->fill_tensor_uniform(Accessor(b[6]), 6); |
| 342 | library->fill_tensor_uniform(Accessor(w[7]), 7); |
| 343 | library->fill_tensor_uniform(Accessor(b[7]), 8); |
| 344 | |
| 345 | library->fill_tensor_uniform(Accessor(*dynamic_cast<TensorType *>(w21.get())), 9); |
| 346 | library->fill_tensor_uniform(Accessor(*dynamic_cast<TensorType *>(w22.get())), 10); |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame^] | 347 | |
| 348 | if(!_is_direct_conv) |
| 349 | { |
| 350 | library->fill_tensor_uniform(Accessor(*dynamic_cast<TensorType *>(w41.get())), 11); |
| 351 | library->fill_tensor_uniform(Accessor(*dynamic_cast<TensorType *>(w42.get())), 12); |
| 352 | library->fill_tensor_uniform(Accessor(*dynamic_cast<TensorType *>(w51.get())), 13); |
| 353 | library->fill_tensor_uniform(Accessor(*dynamic_cast<TensorType *>(w52.get())), 14); |
| 354 | } |
| 355 | else |
| 356 | { |
| 357 | library->fill_tensor_uniform(Accessor(w[3]), 11); |
| 358 | library->fill_tensor_uniform(Accessor(b[3]), 12); |
| 359 | library->fill_tensor_uniform(Accessor(w[4]), 13); |
| 360 | library->fill_tensor_uniform(Accessor(b[4]), 14); |
| 361 | } |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 362 | } |
| 363 | } |
| 364 | |
| 365 | #ifdef INTERNAL_ONLY |
| 366 | /** Fills the trainable parameters from binary files |
| 367 | * |
| 368 | * @param weights Files names containing the weights data |
| 369 | * @param biases Files names containing the bias data |
| 370 | */ |
| 371 | void fill(std::vector<std::string> weights, std::vector<std::string> biases) |
| 372 | { |
| 373 | ARM_COMPUTE_ERROR_ON(weights.size() != w.size()); |
| 374 | ARM_COMPUTE_ERROR_ON(biases.size() != b.size()); |
| 375 | ARM_COMPUTE_ERROR_ON(_reshaped_weights); |
| 376 | |
| 377 | for(unsigned int i = 0; i < weights.size(); ++i) |
| 378 | { |
| 379 | library->fill_layer_data(Accessor(w[i]), weights[i]); |
| 380 | library->fill_layer_data(Accessor(b[i]), biases[i]); |
| 381 | } |
| 382 | } |
| 383 | |
| 384 | /** Feed input to network from file. |
| 385 | * |
| 386 | * @param name File name of containing the input data. |
| 387 | */ |
| 388 | void feed(std::string name) |
| 389 | { |
| 390 | library->fill_layer_data(Accessor(input), name); |
| 391 | } |
| 392 | #endif /* INTERNAL_ONLY */ |
| 393 | |
| 394 | /** Get the classification results. |
| 395 | * |
| 396 | * @return Vector containing the classified labels |
| 397 | */ |
| 398 | std::vector<unsigned int> get_classifications() |
| 399 | { |
| 400 | std::vector<unsigned int> classified_labels; |
| 401 | Accessor output_accessor(output); |
| 402 | |
| 403 | Window window; |
| 404 | window.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 405 | for(unsigned int d = 1; d < output_accessor.shape().num_dimensions(); ++d) |
| 406 | { |
| 407 | window.set(d, Window::Dimension(0, output_accessor.shape()[d], 1)); |
| 408 | } |
| 409 | |
| 410 | execute_window_loop(window, [&](const Coordinates & id) |
| 411 | { |
| 412 | int max_idx = 0; |
| 413 | float val = 0; |
| 414 | const void *const out_ptr = output_accessor(id); |
| 415 | for(unsigned int l = 0; l < output_accessor.shape().x(); ++l) |
| 416 | { |
| 417 | float curr_val = reinterpret_cast<const float *>(out_ptr)[l]; |
| 418 | if(curr_val > val) |
| 419 | { |
| 420 | max_idx = l; |
| 421 | val = curr_val; |
| 422 | } |
| 423 | } |
| 424 | classified_labels.push_back(max_idx); |
| 425 | }); |
| 426 | return classified_labels; |
| 427 | } |
| 428 | |
| 429 | /** Clear all allocated memory from the tensor objects */ |
| 430 | void clear() |
| 431 | { |
| 432 | // Free allocations |
| 433 | input.allocator()->free(); |
| 434 | output.allocator()->free(); |
| 435 | |
| 436 | if(!_reshaped_weights) |
| 437 | { |
| 438 | for(auto &wi : w) |
| 439 | { |
| 440 | wi.allocator()->free(); |
| 441 | } |
| 442 | |
| 443 | for(auto &bi : b) |
| 444 | { |
| 445 | bi.allocator()->free(); |
| 446 | } |
| 447 | } |
| 448 | else |
| 449 | { |
| 450 | w[0].allocator()->free(); |
| 451 | w[2].allocator()->free(); |
| 452 | w[5].allocator()->free(); |
| 453 | w[6].allocator()->free(); |
| 454 | w[7].allocator()->free(); |
| 455 | |
| 456 | b[5].allocator()->free(); |
| 457 | b[6].allocator()->free(); |
| 458 | b[7].allocator()->free(); |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame^] | 459 | |
| 460 | if(_is_direct_conv) |
| 461 | { |
| 462 | w[3].allocator()->free(); |
| 463 | w[4].allocator()->free(); |
| 464 | b[2].allocator()->free(); |
| 465 | b[3].allocator()->free(); |
| 466 | b[4].allocator()->free(); |
| 467 | } |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 468 | } |
| 469 | |
| 470 | w21.reset(); |
| 471 | w22.reset(); |
| 472 | b21.reset(); |
| 473 | b21.reset(); |
| 474 | w41.reset(); |
| 475 | w42.reset(); |
| 476 | b41.reset(); |
| 477 | b42.reset(); |
| 478 | w51.reset(); |
| 479 | w52.reset(); |
| 480 | b51.reset(); |
| 481 | b52.reset(); |
| 482 | |
| 483 | conv1_out.allocator()->free(); |
| 484 | act1_out.allocator()->free(); |
| 485 | norm1_out.allocator()->free(); |
| 486 | pool1_out.allocator()->free(); |
| 487 | conv2_out.allocator()->free(); |
| 488 | act2_out.allocator()->free(); |
| 489 | norm2_out.allocator()->free(); |
| 490 | pool2_out.allocator()->free(); |
| 491 | conv3_out.allocator()->free(); |
| 492 | act3_out.allocator()->free(); |
| 493 | conv4_out.allocator()->free(); |
| 494 | act4_out.allocator()->free(); |
| 495 | conv5_out.allocator()->free(); |
| 496 | act5_out.allocator()->free(); |
| 497 | pool5_out.allocator()->free(); |
| 498 | fc6_out.allocator()->free(); |
| 499 | act6_out.allocator()->free(); |
| 500 | fc7_out.allocator()->free(); |
| 501 | act7_out.allocator()->free(); |
| 502 | fc8_out.allocator()->free(); |
| 503 | } |
| 504 | |
| 505 | /** Runs the model */ |
| 506 | void run() |
| 507 | { |
| 508 | // Layer 1 |
| 509 | conv1.run(); |
| 510 | act1.run(); |
| 511 | norm1.run(); |
| 512 | pool1.run(); |
| 513 | // Layer 2 |
| 514 | conv21.run(); |
| 515 | conv22.run(); |
| 516 | act2.run(); |
| 517 | norm2.run(); |
| 518 | pool2.run(); |
| 519 | // Layer 3 |
| 520 | conv3.run(); |
| 521 | act3.run(); |
| 522 | // Layer 4 |
| 523 | conv41.run(); |
| 524 | conv42.run(); |
| 525 | act4.run(); |
| 526 | // Layer 5 |
| 527 | conv51.run(); |
| 528 | conv52.run(); |
| 529 | act5.run(); |
| 530 | pool5.run(); |
| 531 | // Layer 6 |
| 532 | fc6.run(); |
| 533 | act6.run(); |
| 534 | // Layer 7 |
| 535 | fc7.run(); |
| 536 | act7.run(); |
| 537 | // Layer 8 |
| 538 | fc8.run(); |
| 539 | // Softmax |
| 540 | smx.run(); |
| 541 | } |
| 542 | |
| 543 | private: |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame^] | 544 | struct DirectConv |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 545 | { |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame^] | 546 | template <typename ConvolutionLayerFunction1 = ConvolutionLayerFunction, typename DirectConvolutionLayerFunction1 = DirectConvolutionLayerFunction> |
| 547 | typename std::enable_if < !std::is_same<ConvolutionLayerFunction1, DirectConvolutionLayerFunction1>::value, void >::type |
| 548 | 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] | 549 | { |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame^] | 550 | _func.configure(input, weights, biases, output, conv_info); |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 551 | } |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame^] | 552 | |
| 553 | template <typename ConvolutionLayerFunction1 = ConvolutionLayerFunction, typename DirectConvolutionLayerFunction1 = DirectConvolutionLayerFunction> |
| 554 | typename std::enable_if<std::is_same<ConvolutionLayerFunction1, DirectConvolutionLayerFunction1>::value, void>::type |
| 555 | 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] | 556 | { |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame^] | 557 | _func.configure(input, weights, biases, output, conv_info, weights_info); |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 558 | } |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame^] | 559 | |
| 560 | void run() |
| 561 | { |
| 562 | _func.run(); |
| 563 | } |
| 564 | |
| 565 | DirectConvolutionLayerFunction _func{}; |
| 566 | }; |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 567 | |
| 568 | DataType _data_type{ DataType::UNKNOWN }; |
| 569 | int _fixed_point_position{ 0 }; |
| 570 | unsigned int _batches{ 0 }; |
| 571 | bool _reshaped_weights{ false }; |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame^] | 572 | bool _is_direct_conv{ !std::is_same<ConvolutionLayerFunction, DirectConvolutionLayerFunction>::value }; |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 573 | |
| 574 | ActivationLayerFunction act1{}, act2{}, act3{}, act4{}, act5{}, act6{}, act7{}; |
steniu01 | a629da1 | 2017-07-28 14:40:58 +0100 | [diff] [blame^] | 575 | ConvolutionLayerFunction conv1{}, conv21{}, conv22{}; |
| 576 | DirectConv conv3{}, conv41{}, conv42{}, conv51{}, conv52{}; |
Moritz Pflanzer | ee493ae | 2017-07-05 10:52:21 +0100 | [diff] [blame] | 577 | FullyConnectedLayerFunction fc6{}, fc7{}, fc8{}; |
| 578 | NormalizationLayerFunction norm1{}, norm2{}; |
| 579 | PoolingLayerFunction pool1{}, pool2{}, pool5{}; |
| 580 | SoftmaxLayerFunction smx{}; |
| 581 | |
| 582 | TensorType input{}, output{}; |
| 583 | std::array<TensorType, 8> w{ {} }, b{ {} }; |
| 584 | std::unique_ptr<ITensorType> w21{ nullptr }, w22{ nullptr }, b21{ nullptr }, b22{ nullptr }; |
| 585 | std::unique_ptr<ITensorType> w41{ nullptr }, w42{ nullptr }, b41{ nullptr }, b42{ nullptr }; |
| 586 | std::unique_ptr<ITensorType> w51{ nullptr }, w52{ nullptr }, b51{ nullptr }, b52{ nullptr }; |
| 587 | |
| 588 | TensorType conv1_out{}, act1_out{}, norm1_out{}, pool1_out{}; |
| 589 | TensorType conv2_out{}, act2_out{}, pool2_out{}, norm2_out{}; |
| 590 | TensorType conv3_out{}, act3_out{}; |
| 591 | TensorType conv4_out{}, act4_out{}; |
| 592 | TensorType conv5_out{}, act5_out{}, pool5_out{}; |
| 593 | TensorType fc6_out{}, act6_out{}; |
| 594 | TensorType fc7_out{}, act7_out{}; |
| 595 | TensorType fc8_out{}; |
| 596 | |
| 597 | std::unique_ptr<SubTensorType> pool11_out{}, pool12_out{}; |
| 598 | std::unique_ptr<SubTensorType> conv21_out{}, conv22_out{}; |
| 599 | std::unique_ptr<SubTensorType> act31_out{}, act32_out{}; |
| 600 | std::unique_ptr<SubTensorType> conv41_out{}, conv42_out{}, act41_out{}, act42_out{}; |
| 601 | std::unique_ptr<SubTensorType> conv51_out{}, conv52_out{}; |
| 602 | }; |
| 603 | } // namespace networks |
| 604 | } // namespace test |
| 605 | } // namespace arm_compute |
| 606 | #endif //__ARM_COMPUTE_TEST_MODEL_OBJECTS_ALEXNET_H__ |