Narumol Prangnawarat | 0be4338 | 2019-05-27 11:29:59 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2017 Arm Ltd. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | #pragma once |
| 6 | |
| 7 | #include <ResolveType.hpp> |
| 8 | |
| 9 | #include <armnn/INetwork.hpp> |
| 10 | |
| 11 | #include <backendsCommon/test/CommonTestUtils.hpp> |
| 12 | |
| 13 | #include <boost/test/unit_test.hpp> |
| 14 | |
| 15 | #include <vector> |
| 16 | |
| 17 | namespace |
| 18 | { |
| 19 | |
| 20 | template<typename armnn::DataType DataType> |
| 21 | INetworkPtr CreateSplitterNetwork(const TensorShape& inputShape, |
| 22 | const std::vector<TensorShape>& outputShapes, |
| 23 | unsigned int splitAxis, |
| 24 | unsigned int numSplit, |
| 25 | const float qScale = 1.0f, |
| 26 | const int32_t qOffset = 0) |
| 27 | { |
| 28 | using namespace armnn; |
| 29 | // Builds up the structure of the network. |
| 30 | INetworkPtr net(INetwork::Create()); |
| 31 | |
| 32 | TensorInfo inputTensorInfo(inputShape, DataType, qScale, qOffset); |
| 33 | |
| 34 | std::vector<unsigned int> splitterDimSizes(inputShape.GetNumDimensions()); |
| 35 | |
| 36 | // Add current input shape to splitterDimSizes |
| 37 | for (unsigned int i = 0; i < inputShape.GetNumDimensions(); ++i) |
| 38 | { |
| 39 | splitterDimSizes[i] = inputTensorInfo.GetShape()[i]; |
| 40 | } |
| 41 | |
| 42 | if (splitterDimSizes[splitAxis] % numSplit != 0) |
| 43 | { |
| 44 | throw ParseException("Number of splits must evenly divide the dimension"); |
| 45 | } |
| 46 | splitterDimSizes[splitAxis] /= numSplit; |
| 47 | |
Narumol Prangnawarat | 0f072ab | 2019-05-29 14:12:46 +0100 | [diff] [blame] | 48 | SplitterDescriptor splitDesc(numSplit, inputShape.GetNumDimensions()); |
Narumol Prangnawarat | 0be4338 | 2019-05-27 11:29:59 +0100 | [diff] [blame] | 49 | for (unsigned int g = 0; g < numSplit; ++g) |
| 50 | { |
| 51 | // Set the size of the views. |
| 52 | for (unsigned int dimIdx = 0; dimIdx < splitterDimSizes.size(); ++dimIdx) |
| 53 | { |
| 54 | splitDesc.SetViewSize(g, dimIdx, splitterDimSizes[dimIdx]); |
| 55 | } |
| 56 | splitDesc.SetViewOriginCoord(g, splitAxis, splitterDimSizes[splitAxis] * g); |
| 57 | } |
| 58 | |
| 59 | IConnectableLayer* splitter = net->AddSplitterLayer(splitDesc, "splitter"); |
| 60 | IConnectableLayer* input = net->AddInputLayer(0, "input"); |
| 61 | Connect(input, splitter, inputTensorInfo, 0, 0); |
| 62 | |
| 63 | for (unsigned int i = 0; i < outputShapes.size(); ++i) |
| 64 | { |
| 65 | TensorInfo outputTensorInfo(outputShapes[i], DataType, qScale, qOffset); |
| 66 | IConnectableLayer* output = net->AddOutputLayer(boost::numeric_cast<LayerBindingId>(i)); |
| 67 | Connect(splitter, output, outputTensorInfo, i, 0); |
| 68 | } |
| 69 | |
| 70 | return net; |
| 71 | } |
| 72 | |
| 73 | template<armnn::DataType ArmnnType> |
Narumol Prangnawarat | 0f072ab | 2019-05-29 14:12:46 +0100 | [diff] [blame] | 74 | void Splitter1dEndToEnd(const std::vector<BackendId>& backends) |
| 75 | { |
| 76 | using namespace armnn; |
| 77 | using T = ResolveType<ArmnnType>; |
| 78 | |
| 79 | unsigned int splitAxis = 0; |
| 80 | unsigned int numSplit = 2; |
| 81 | const TensorShape& inputShape = { 4 }; |
| 82 | const std::vector<TensorShape> outputShapes{{ 2 }, { 2 }}; |
| 83 | |
| 84 | // Builds up the structure of the network |
| 85 | INetworkPtr net = CreateSplitterNetwork<ArmnnType>(inputShape, outputShapes, splitAxis, numSplit); |
| 86 | |
| 87 | BOOST_TEST_CHECKPOINT("create a network"); |
| 88 | |
| 89 | // Creates structures for input & output. |
| 90 | std::vector<T> inputData{ 1, 2, 3, 4 }; |
| 91 | |
| 92 | std::vector<T> expectedOutput0{ 1, 2 }; |
| 93 | std::vector<T> expectedOutput1{ 3, 4 }; |
| 94 | |
| 95 | std::map<int, std::vector<T>> inputTensorData = { { 0, inputData } }; |
| 96 | std::map<int, std::vector<T>> expectedOutputData = { { 0, expectedOutput0 }, {1, expectedOutput1} }; |
| 97 | |
| 98 | EndToEndLayerTestImpl<ArmnnType, ArmnnType>(move(net), inputTensorData, expectedOutputData, backends); |
| 99 | } |
| 100 | |
| 101 | template<armnn::DataType ArmnnType> |
| 102 | void Splitter2dDim0EndToEnd(const std::vector<BackendId>& backends) |
| 103 | { |
| 104 | using namespace armnn; |
| 105 | using T = ResolveType<ArmnnType>; |
| 106 | |
| 107 | unsigned int splitAxis = 0; |
| 108 | unsigned int numSplit = 2; |
| 109 | const TensorShape& inputShape = { 4, 3 }; |
| 110 | const std::vector<TensorShape> outputShapes{{ 2, 3 }, { 2, 3 }}; |
| 111 | |
| 112 | // Builds up the structure of the network |
| 113 | INetworkPtr net = CreateSplitterNetwork<ArmnnType>(inputShape, outputShapes, splitAxis, numSplit); |
| 114 | |
| 115 | BOOST_TEST_CHECKPOINT("create a network"); |
| 116 | |
| 117 | // Creates structures for input & output. |
| 118 | std::vector<T> inputData{ |
| 119 | 1, 2, |
| 120 | 3, 4, |
| 121 | 5, 6, |
| 122 | 7, 8, |
| 123 | 9, 10, |
| 124 | 11, 12 |
| 125 | }; |
| 126 | |
| 127 | std::vector<T> expectedOutput0{ 1, 2, 3, 4, 5, 6 }; |
| 128 | std::vector<T> expectedOutput1{ 7, 8, 9, 10, 11, 12 }; |
| 129 | |
| 130 | std::map<int, std::vector<T>> inputTensorData = { { 0, inputData } }; |
| 131 | std::map<int, std::vector<T>> expectedOutputData = { { 0, expectedOutput0 }, {1, expectedOutput1} }; |
| 132 | |
| 133 | EndToEndLayerTestImpl<ArmnnType, ArmnnType>(move(net), inputTensorData, expectedOutputData, backends); |
| 134 | } |
| 135 | |
| 136 | template<armnn::DataType ArmnnType> |
| 137 | void Splitter2dDim1EndToEnd(const std::vector<BackendId>& backends) |
| 138 | { |
| 139 | using namespace armnn; |
| 140 | using T = ResolveType<ArmnnType>; |
| 141 | |
| 142 | unsigned int splitAxis = 1; |
| 143 | unsigned int numSplit = 3; |
| 144 | const TensorShape& inputShape = { 4, 3 }; |
| 145 | const std::vector<TensorShape> outputShapes{{ 4, 1 }, { 4, 1 }, { 4, 1 }}; |
| 146 | |
| 147 | // Builds up the structure of the network |
| 148 | INetworkPtr net = CreateSplitterNetwork<ArmnnType>(inputShape, outputShapes, splitAxis, numSplit); |
| 149 | |
| 150 | BOOST_TEST_CHECKPOINT("create a network"); |
| 151 | |
| 152 | // Creates structures for input & output. |
| 153 | std::vector<T> inputData{ |
| 154 | 1, 2, |
| 155 | 3, 4, |
| 156 | 5, 6, |
| 157 | 7, 8, |
| 158 | 9, 10, |
| 159 | 11, 12 |
| 160 | }; |
| 161 | |
| 162 | std::vector<T> expectedOutput0{ 1, 4, 7, 10 }; |
| 163 | std::vector<T> expectedOutput1{ 2, 5, 8, 11 }; |
| 164 | std::vector<T> expectedOutput2{ 3, 6, 9, 12 }; |
| 165 | |
| 166 | std::map<int, std::vector<T>> inputTensorData = { { 0, inputData } }; |
| 167 | std::map<int, std::vector<T>> expectedOutputData = { { 0, expectedOutput0 }, |
| 168 | { 1, expectedOutput1 }, |
| 169 | { 2, expectedOutput2 } }; |
| 170 | |
| 171 | EndToEndLayerTestImpl<ArmnnType, ArmnnType>(move(net), inputTensorData, expectedOutputData, backends); |
| 172 | } |
| 173 | |
| 174 | template<armnn::DataType ArmnnType> |
| 175 | void Splitter3dDim0EndToEnd(const std::vector<BackendId>& backends) |
| 176 | { |
| 177 | using namespace armnn; |
| 178 | using T = ResolveType<ArmnnType>; |
| 179 | |
| 180 | unsigned int splitAxis = 0; |
| 181 | unsigned int numSplit = 2; |
| 182 | const TensorShape& inputShape = { 2, 4, 3 }; |
| 183 | const std::vector<TensorShape> outputShapes{{ 1, 4, 3 }, { 1, 4, 3 }}; |
| 184 | |
| 185 | // Builds up the structure of the network |
| 186 | INetworkPtr net = CreateSplitterNetwork<ArmnnType>(inputShape, outputShapes, splitAxis, numSplit); |
| 187 | |
| 188 | BOOST_TEST_CHECKPOINT("create a network"); |
| 189 | |
| 190 | // Creates structures for input & output. |
| 191 | std::vector<T> inputData{ |
| 192 | 1, 2, 3, |
| 193 | 4, 5, 6, |
| 194 | 7, 8, 9, |
| 195 | 10, 11, 12, |
| 196 | 13, 14, 15, |
| 197 | 16, 17, 18, |
| 198 | 19, 20, 21, |
| 199 | 22, 23, 24 |
| 200 | }; |
| 201 | |
| 202 | std::vector<T> expectedOutput0{ |
| 203 | 1, 2, 3, |
| 204 | 4, 5, 6, |
| 205 | 7, 8, 9, |
| 206 | 10, 11, 12 |
| 207 | }; |
| 208 | std::vector<T> expectedOutput1{ |
| 209 | 13, 14, 15, |
| 210 | 16, 17, 18, |
| 211 | 19, 20, 21, |
| 212 | 22, 23, 24 |
| 213 | }; |
| 214 | |
| 215 | std::map<int, std::vector<T>> inputTensorData = { { 0, inputData } }; |
| 216 | std::map<int, std::vector<T>> expectedOutputData = { { 0, expectedOutput0 }, |
| 217 | { 1, expectedOutput1 } }; |
| 218 | |
| 219 | EndToEndLayerTestImpl<ArmnnType, ArmnnType>(move(net), inputTensorData, expectedOutputData, backends); |
| 220 | } |
| 221 | |
| 222 | template<armnn::DataType ArmnnType> |
| 223 | void Splitter3dDim1EndToEnd(const std::vector<BackendId>& backends) |
| 224 | { |
| 225 | using namespace armnn; |
| 226 | using T = ResolveType<ArmnnType>; |
| 227 | |
| 228 | unsigned int splitAxis = 1; |
| 229 | unsigned int numSplit = 2; |
| 230 | const TensorShape& inputShape = { 2, 4, 3 }; |
| 231 | const std::vector<TensorShape> outputShapes{{ 2, 2, 3 }, { 2, 2, 3 }}; |
| 232 | |
| 233 | // Builds up the structure of the network |
| 234 | INetworkPtr net = CreateSplitterNetwork<ArmnnType>(inputShape, outputShapes, splitAxis, numSplit); |
| 235 | |
| 236 | BOOST_TEST_CHECKPOINT("create a network"); |
| 237 | |
| 238 | // Creates structures for input & output. |
| 239 | std::vector<T> inputData{ |
| 240 | 1, 2, 3, |
| 241 | 4, 5, 6, |
| 242 | 7, 8, 9, |
| 243 | 10, 11, 12, |
| 244 | 13, 14, 15, |
| 245 | 16, 17, 18, |
| 246 | 19, 20, 21, |
| 247 | 22, 23, 24 |
| 248 | }; |
| 249 | |
| 250 | std::vector<T> expectedOutput0{ |
| 251 | 1, 2, 3, |
| 252 | 4, 5, 6, |
| 253 | 13, 14, 15, |
| 254 | 16, 17, 18 |
| 255 | }; |
| 256 | std::vector<T> expectedOutput1{ |
| 257 | 7, 8, 9, |
| 258 | 10, 11, 12, |
| 259 | 19, 20, 21, |
| 260 | 22, 23, 24 |
| 261 | }; |
| 262 | |
| 263 | std::map<int, std::vector<T>> inputTensorData = { { 0, inputData } }; |
| 264 | std::map<int, std::vector<T>> expectedOutputData = { { 0, expectedOutput0 }, |
| 265 | { 1, expectedOutput1 } }; |
| 266 | |
| 267 | EndToEndLayerTestImpl<ArmnnType, ArmnnType>(move(net), inputTensorData, expectedOutputData, backends); |
| 268 | } |
| 269 | |
| 270 | template<armnn::DataType ArmnnType> |
| 271 | void Splitter3dDim2EndToEnd(const std::vector<BackendId>& backends) |
| 272 | { |
| 273 | using namespace armnn; |
| 274 | using T = ResolveType<ArmnnType>; |
| 275 | |
| 276 | unsigned int splitAxis = 2; |
| 277 | unsigned int numSplit = 3; |
| 278 | const TensorShape& inputShape = { 2, 4, 3 }; |
| 279 | const std::vector<TensorShape> outputShapes{{ 2, 4, 1 }, { 2, 4, 1 }, { 2, 4, 1 }}; |
| 280 | |
| 281 | // Builds up the structure of the network |
| 282 | INetworkPtr net = CreateSplitterNetwork<ArmnnType>(inputShape, outputShapes, splitAxis, numSplit); |
| 283 | |
| 284 | BOOST_TEST_CHECKPOINT("create a network"); |
| 285 | |
| 286 | // Creates structures for input & output. |
| 287 | std::vector<T> inputData{ |
| 288 | 1, 2, 3, |
| 289 | 4, 5, 6, |
| 290 | 7, 8, 9, |
| 291 | 10, 11, 12, |
| 292 | 13, 14, 15, |
| 293 | 16, 17, 18, |
| 294 | 19, 20, 21, |
| 295 | 22, 23, 24 |
| 296 | }; |
| 297 | |
| 298 | std::vector<T> expectedOutput0{ 1, 4, 7, 10, 13, 16, 19, 22 }; |
| 299 | std::vector<T> expectedOutput1{ 2, 5, 8, 11, 14, 17, 20, 23 }; |
| 300 | std::vector<T> expectedOutput2{ 3, 6, 9, 12, 15, 18, 21, 24 }; |
| 301 | |
| 302 | std::map<int, std::vector<T>> inputTensorData = { { 0, inputData } }; |
| 303 | std::map<int, std::vector<T>> expectedOutputData = { { 0, expectedOutput0 }, |
| 304 | { 1, expectedOutput1 }, |
| 305 | { 2, expectedOutput2 } }; |
| 306 | |
| 307 | EndToEndLayerTestImpl<ArmnnType, ArmnnType>(move(net), inputTensorData, expectedOutputData, backends); |
| 308 | } |
| 309 | |
| 310 | template<armnn::DataType ArmnnType> |
| 311 | void Splitter4dDim0EndToEnd(const std::vector<BackendId>& backends) |
Narumol Prangnawarat | 0be4338 | 2019-05-27 11:29:59 +0100 | [diff] [blame] | 312 | { |
| 313 | using namespace armnn; |
| 314 | using T = ResolveType<ArmnnType>; |
| 315 | |
| 316 | unsigned int splitAxis = 0; |
| 317 | unsigned int numSplit = 2; |
| 318 | const TensorShape& inputShape = { 4, 3, 2, 2 }; |
| 319 | const std::vector<TensorShape> outputShapes{{ 2, 3, 2, 2 }, { 2, 3, 2, 2 }}; |
| 320 | |
| 321 | // Builds up the structure of the network |
| 322 | INetworkPtr net = CreateSplitterNetwork<ArmnnType>(inputShape, outputShapes, splitAxis, numSplit); |
| 323 | |
| 324 | BOOST_TEST_CHECKPOINT("create a network"); |
| 325 | |
| 326 | // Creates structures for input & output. |
| 327 | std::vector<T> inputData{ |
| 328 | 1, 2, |
| 329 | 3, 4, |
| 330 | 5, 6, |
| 331 | 7, 8, |
| 332 | 9, 10, |
| 333 | 11, 12, |
| 334 | 13, 14, |
| 335 | 15, 16, |
| 336 | 17, 18, |
| 337 | 19, 20, |
| 338 | 21, 22, |
| 339 | 23, 24, |
| 340 | 25, 26, |
| 341 | 27, 28, |
| 342 | 29, 30, |
| 343 | 31, 32, |
| 344 | 33, 34, |
| 345 | 35, 36, |
| 346 | 37, 38, |
| 347 | 39, 40, |
| 348 | 41, 42, |
| 349 | 43, 44, |
| 350 | 45, 46, |
| 351 | 47, 48 |
| 352 | }; |
| 353 | |
| 354 | std::vector<T> expectedOutput0{ |
| 355 | 1, 2, |
| 356 | 3, 4, |
| 357 | 5, 6, |
| 358 | 7, 8, |
| 359 | 9, 10, |
| 360 | 11, 12, |
| 361 | 13, 14, |
| 362 | 15, 16, |
| 363 | 17, 18, |
| 364 | 19, 20, |
| 365 | 21, 22, |
| 366 | 23, 24 |
| 367 | }; |
| 368 | |
| 369 | std::vector<T> expectedOutput1{ |
| 370 | 25, 26, |
| 371 | 27, 28, |
| 372 | 29, 30, |
| 373 | 31, 32, |
| 374 | 33, 34, |
| 375 | 35, 36, |
| 376 | 37, 38, |
| 377 | 39, 40, |
| 378 | 41, 42, |
| 379 | 43, 44, |
| 380 | 45, 46, |
| 381 | 47, 48 |
| 382 | }; |
| 383 | |
| 384 | std::map<int, std::vector<T>> inputTensorData = {{ 0,inputData }}; |
| 385 | std::map<int, std::vector<T>> expectedOutputData = {{ 0, expectedOutput0 }, { 1, expectedOutput1 }}; |
| 386 | |
| 387 | EndToEndLayerTestImpl<ArmnnType, ArmnnType>(move(net), inputTensorData, expectedOutputData, backends); |
| 388 | } |
| 389 | |
| 390 | template<armnn::DataType ArmnnType> |
Narumol Prangnawarat | 0f072ab | 2019-05-29 14:12:46 +0100 | [diff] [blame] | 391 | void Splitter4dDim1EndToEnd(const std::vector<BackendId>& backends) |
Narumol Prangnawarat | 0be4338 | 2019-05-27 11:29:59 +0100 | [diff] [blame] | 392 | { |
| 393 | using namespace armnn; |
| 394 | using T = ResolveType<ArmnnType>; |
| 395 | |
| 396 | unsigned int splitAxis = 1; |
| 397 | unsigned int numSplit = 2; |
| 398 | const TensorShape& inputShape = { 2, 6, 2, 2 }; |
| 399 | const std::vector<TensorShape> outputShapes{{ 2, 3, 2, 2 }, { 2, 3, 2, 2 }}; |
| 400 | |
| 401 | // Builds up the structure of the network |
| 402 | INetworkPtr net = CreateSplitterNetwork<ArmnnType>(inputShape, outputShapes, splitAxis, numSplit); |
| 403 | |
| 404 | BOOST_TEST_CHECKPOINT("create a network"); |
| 405 | |
| 406 | // Creates structures for input & output. |
| 407 | std::vector<T> inputData{ |
| 408 | 1, 2, |
| 409 | 3, 4, |
| 410 | 5, 6, |
| 411 | 7, 8, |
| 412 | 9, 10, |
| 413 | 11, 12, |
| 414 | 13, 14, |
| 415 | 15, 16, |
| 416 | 17, 18, |
| 417 | 19, 20, |
| 418 | 21, 22, |
| 419 | 23, 24, |
| 420 | 25, 26, |
| 421 | 27, 28, |
| 422 | 29, 30, |
| 423 | 31, 32, |
| 424 | 33, 34, |
| 425 | 35, 36, |
| 426 | 37, 38, |
| 427 | 39, 40, |
| 428 | 41, 42, |
| 429 | 43, 44, |
| 430 | 45, 46, |
| 431 | 47, 48 |
| 432 | }; |
| 433 | |
| 434 | std::vector<T> expectedOutput0{ |
| 435 | 1, 2, |
| 436 | 3, 4, |
| 437 | 5, 6, |
| 438 | 7, 8, |
| 439 | 9, 10, |
| 440 | 11, 12, |
| 441 | 25, 26, |
| 442 | 27, 28, |
| 443 | 29, 30, |
| 444 | 31, 32, |
| 445 | 33, 34, |
| 446 | 35, 36 |
| 447 | }; |
| 448 | |
| 449 | std::vector<T> expectedOutput1{ |
| 450 | 13, 14, |
| 451 | 15, 16, |
| 452 | 17, 18, |
| 453 | 19, 20, |
| 454 | 21, 22, |
| 455 | 23, 24, |
| 456 | 37, 38, |
| 457 | 39, 40, |
| 458 | 41, 42, |
| 459 | 43, 44, |
| 460 | 45, 46, |
| 461 | 47, 48 |
| 462 | }; |
| 463 | |
| 464 | std::map<int, std::vector<T>> inputTensorData = {{ 0,inputData }}; |
| 465 | std::map<int, std::vector<T>> expectedOutputData = {{ 0, expectedOutput0 }, { 1, expectedOutput1 }}; |
| 466 | |
| 467 | EndToEndLayerTestImpl<ArmnnType, ArmnnType>(move(net), inputTensorData, expectedOutputData, backends); |
| 468 | } |
| 469 | |
| 470 | template<armnn::DataType ArmnnType> |
Narumol Prangnawarat | 0f072ab | 2019-05-29 14:12:46 +0100 | [diff] [blame] | 471 | void Splitter4dDim2EndToEnd(const std::vector<BackendId>& backends) |
Narumol Prangnawarat | 0be4338 | 2019-05-27 11:29:59 +0100 | [diff] [blame] | 472 | { |
| 473 | using namespace armnn; |
| 474 | using T = ResolveType<ArmnnType>; |
| 475 | |
| 476 | unsigned int splitAxis = 2; |
| 477 | unsigned int numSplit = 2; |
| 478 | const TensorShape& inputShape = { 2, 3, 4, 2 }; |
| 479 | const std::vector<TensorShape> outputShapes{{ 2, 3, 2, 2 }, { 2, 3, 2, 2 }}; |
| 480 | |
| 481 | // Builds up the structure of the network |
| 482 | INetworkPtr net = CreateSplitterNetwork<ArmnnType>(inputShape, outputShapes, splitAxis, numSplit); |
| 483 | |
| 484 | BOOST_TEST_CHECKPOINT("create a network"); |
| 485 | |
| 486 | // Creates structures for input & output. |
| 487 | std::vector<T> inputData{ |
| 488 | 1, 2, |
| 489 | 3, 4, |
| 490 | 5, 6, |
| 491 | 7, 8, |
| 492 | 9, 10, |
| 493 | 11, 12, |
| 494 | 13, 14, |
| 495 | 15, 16, |
| 496 | 17, 18, |
| 497 | 19, 20, |
| 498 | 21, 22, |
| 499 | 23, 24, |
| 500 | 25, 26, |
| 501 | 27, 28, |
| 502 | 29, 30, |
| 503 | 31, 32, |
| 504 | 33, 34, |
| 505 | 35, 36, |
| 506 | 37, 38, |
| 507 | 39, 40, |
| 508 | 41, 42, |
| 509 | 43, 44, |
| 510 | 45, 46, |
| 511 | 47, 48 |
| 512 | }; |
| 513 | |
| 514 | std::vector<T> expectedOutput0{ |
| 515 | 1, 2, |
| 516 | 3, 4, |
| 517 | 9, 10, |
| 518 | 11, 12, |
| 519 | 17, 18, |
| 520 | 19, 20, |
| 521 | 25, 26, |
| 522 | 27, 28, |
| 523 | 33, 34, |
| 524 | 35, 36, |
| 525 | 41, 42, |
| 526 | 43, 44 |
| 527 | }; |
| 528 | |
| 529 | std::vector<T> expectedOutput1{ |
| 530 | 5, 6, |
| 531 | 7, 8, |
| 532 | 13, 14, |
| 533 | 15, 16, |
| 534 | 21, 22, |
| 535 | 23, 24, |
| 536 | 29, 30, |
| 537 | 31, 32, |
| 538 | 37, 38, |
| 539 | 39, 40, |
| 540 | 45, 46, |
| 541 | 47, 48 |
| 542 | }; |
| 543 | |
| 544 | std::map<int, std::vector<T>> inputTensorData = {{ 0,inputData }}; |
| 545 | std::map<int, std::vector<T>> expectedOutputData = {{ 0, expectedOutput0 }, { 1, expectedOutput1 }}; |
| 546 | |
| 547 | EndToEndLayerTestImpl<ArmnnType, ArmnnType>(move(net), inputTensorData, expectedOutputData, backends); |
| 548 | } |
| 549 | |
| 550 | template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>> |
Narumol Prangnawarat | 0f072ab | 2019-05-29 14:12:46 +0100 | [diff] [blame] | 551 | void Splitter4dDim3EndToEnd(const std::vector<BackendId>& backends) |
Narumol Prangnawarat | 0be4338 | 2019-05-27 11:29:59 +0100 | [diff] [blame] | 552 | { |
| 553 | using namespace armnn; |
| 554 | |
| 555 | unsigned int splitAxis = 3; |
| 556 | unsigned int numSplit = 2; |
| 557 | const TensorShape& inputShape = { 2, 3, 4, 2 }; |
| 558 | const std::vector<TensorShape> outputShapes{{ 2, 3, 4, 1 }, { 2, 3, 4, 1 }}; |
| 559 | |
| 560 | // Builds up the structure of the network |
| 561 | INetworkPtr net = CreateSplitterNetwork<ArmnnType>(inputShape, outputShapes, splitAxis, numSplit); |
| 562 | |
| 563 | BOOST_TEST_CHECKPOINT("create a network"); |
| 564 | |
| 565 | // Creates structures for input & output. |
| 566 | std::vector<T> inputData{ |
| 567 | 1, 2, |
| 568 | 3, 4, |
| 569 | 5, 6, |
| 570 | 7, 8, |
| 571 | 9, 10, |
| 572 | 11, 12, |
| 573 | 13, 14, |
| 574 | 15, 16, |
| 575 | 17, 18, |
| 576 | 19, 20, |
| 577 | 21, 22, |
| 578 | 23, 24, |
| 579 | 25, 26, |
| 580 | 27, 28, |
| 581 | 29, 30, |
| 582 | 31, 32, |
| 583 | 33, 34, |
| 584 | 35, 36, |
| 585 | 37, 38, |
| 586 | 39, 40, |
| 587 | 41, 42, |
| 588 | 43, 44, |
| 589 | 45, 46, |
| 590 | 47, 48 |
| 591 | }; |
| 592 | |
| 593 | std::vector<T> expectedOutput0{ |
| 594 | 1, 3, 5, 7, |
| 595 | 9, 11, 13, 15, |
| 596 | 17, 19, 21, 23, |
| 597 | 25, 27, 29, 31, |
| 598 | 33, 35, 37, 39, |
| 599 | 41, 43, 45, 47 |
| 600 | }; |
| 601 | |
| 602 | std::vector<T> expectedOutput1{ |
| 603 | 2, 4, 6, 8, |
| 604 | 10, 12, 14, 16, |
| 605 | 18, 20, 22, 24, |
| 606 | 26, 28, 30, 32, |
| 607 | 34, 36, 38, 40, |
| 608 | 42, 44, 46, 48 |
| 609 | }; |
| 610 | |
| 611 | std::map<int, std::vector<T>> inputTensorData = {{ 0,inputData }}; |
| 612 | std::map<int, std::vector<T>> expectedOutputData = {{ 0, expectedOutput0 }, { 1, expectedOutput1 }}; |
| 613 | |
| 614 | EndToEndLayerTestImpl<ArmnnType, ArmnnType>(move(net), inputTensorData, expectedOutputData, backends); |
| 615 | } |
| 616 | |
| 617 | } // anonymous namespace |