telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame^] | 1 | // |
| 2 | // Copyright © 2017 Arm Ltd. All rights reserved. |
| 3 | // See LICENSE file in the project root for full license information. |
| 4 | // |
| 5 | #include "CaffeParser.hpp" |
| 6 | |
| 7 | #include "armnn/Descriptors.hpp" |
| 8 | #include "armnn/INetwork.hpp" |
| 9 | #include "armnn/Utils.hpp" |
| 10 | #include "armnn/Exceptions.hpp" |
| 11 | |
| 12 | #include "GraphTopologicalSort.hpp" |
| 13 | |
| 14 | #include <boost/numeric/conversion/cast.hpp> |
| 15 | #include <boost/assert.hpp> |
| 16 | #include <boost/format.hpp> |
| 17 | #include <boost/log/trivial.hpp> |
| 18 | |
| 19 | // Caffe |
| 20 | #include "caffe/proto/caffe.pb.h" |
| 21 | |
| 22 | // ProtoBuf |
| 23 | #include <google/protobuf/io/coded_stream.h> |
| 24 | #include <google/protobuf/io/zero_copy_stream.h> |
| 25 | #include <google/protobuf/io/zero_copy_stream_impl.h> |
| 26 | #include <google/protobuf/text_format.h> |
| 27 | #include <google/protobuf/stubs/common.h> |
| 28 | #include <google/protobuf/stubs/once.h> |
| 29 | #include <google/protobuf/io/coded_stream.h> |
| 30 | #include <google/protobuf/wire_format_lite_inl.h> |
| 31 | #include <google/protobuf/descriptor.h> |
| 32 | #include <google/protobuf/generated_message_reflection.h> |
| 33 | #include <google/protobuf/reflection_ops.h> |
| 34 | #include <google/protobuf/wire_format.h> |
| 35 | |
| 36 | #include <cmath> |
| 37 | #include <sstream> |
| 38 | #include <queue> |
| 39 | #include <fcntl.h> |
| 40 | |
| 41 | /// Caffe networks are loaded from protobuf files (binary or text) using the protobuf library and the generated |
| 42 | /// code from caffe.pb.h. This gives us a caffe::NetParameter which is an in-memory version of the file. |
| 43 | /// This contains a flat list of Caffe 'layers' (e.g. convolution, pooling etc.). |
| 44 | /// Each layer has inputs (called "bottoms") and outputs (called "tops"). Data flows from bottom to top. |
| 45 | /// The bottoms of a layer refer to the tops of other layers, not their names. |
| 46 | /// The names of layers seem to be arbitrary (you could rename a layer and the network wouldn't need any other changes). |
| 47 | /// |
| 48 | /// Some layers (e.g. Relu) can be configured so that their top and bottom are both the same. This is called an |
| 49 | /// "in-place" layer and is a Caffe runtime feature used to reduce memory usage by modifying tensors in-place. |
| 50 | /// This isn't relevant to the parser and so we preprocess these layers to convert them to regular layers, to result |
| 51 | /// in a consistent graph structure. |
| 52 | |
| 53 | namespace armnnCaffeParser |
| 54 | { |
| 55 | |
| 56 | using namespace armnn; |
| 57 | using namespace caffe; |
| 58 | using namespace std; |
| 59 | using namespace google::protobuf::io; |
| 60 | |
| 61 | const std::map<std::string, CaffeParser::OperationParsingFunction> CaffeParser::ms_CaffeLayerNameToParsingFunctions = { |
| 62 | { "Input", &CaffeParser::ParseInputLayer }, |
| 63 | { "Convolution", &CaffeParser::ParseConvLayer }, |
| 64 | { "Pooling", &CaffeParser::ParsePoolingLayer }, |
| 65 | { "ReLU", &CaffeParser::ParseReluLayer }, |
| 66 | { "LRN", &CaffeParser::ParseLRNLayer }, |
| 67 | { "InnerProduct", &CaffeParser::ParseInnerProductLayer }, |
| 68 | { "Softmax", &CaffeParser::ParseSoftmaxLayer }, |
| 69 | { "Eltwise", &CaffeParser::ParseEltwiseLayer }, |
| 70 | { "Concat", &CaffeParser::ParseConcatLayer }, |
| 71 | { "BatchNorm", &CaffeParser::ParseBatchNormLayer }, |
| 72 | { "Scale", &CaffeParser::ParseScaleLayer }, |
| 73 | { "Split", &CaffeParser::ParseSplitLayer }, |
| 74 | { "Dropout", &CaffeParser::ParseDropoutLayer}, |
| 75 | }; |
| 76 | |
| 77 | ICaffeParser* ICaffeParser::CreateRaw() |
| 78 | { |
| 79 | return new CaffeParser(); |
| 80 | } |
| 81 | |
| 82 | ICaffeParserPtr ICaffeParser::Create() |
| 83 | { |
| 84 | return ICaffeParserPtr(CreateRaw(), &ICaffeParser::Destroy); |
| 85 | } |
| 86 | |
| 87 | void ICaffeParser::Destroy(ICaffeParser* parser) |
| 88 | { |
| 89 | delete parser; |
| 90 | } |
| 91 | |
| 92 | CaffeParser::CaffeParser() |
| 93 | : m_Network(nullptr, nullptr) |
| 94 | { |
| 95 | |
| 96 | } |
| 97 | |
| 98 | void GetDataFromBlob(const LayerParameter& layerParam, vector<float>& outData, unsigned int blobIndex) |
| 99 | { |
| 100 | if (blobIndex >= boost::numeric_cast<unsigned int>(layerParam.blobs_size())) |
| 101 | { |
| 102 | throw ParseException(boost::str(boost::format("Expected data blob at index %1% in layer %2% not found") |
| 103 | % blobIndex % layerParam.name())); |
| 104 | } |
| 105 | |
| 106 | const BlobProto& blob = layerParam.blobs(boost::numeric_cast<int>(blobIndex)); |
| 107 | |
| 108 | if (boost::numeric_cast<size_t>(blob.data_size()) != outData.size()) |
| 109 | { |
| 110 | throw ParseException(boost::str(boost::format( |
| 111 | "Data blob at index %1% in layer %2% has an unexpected size. Expected %3% elements but got %4% elements") |
| 112 | % blobIndex % layerParam.name() % outData.size() % blob.data_size())); |
| 113 | } |
| 114 | |
| 115 | for (unsigned int i = 0; i < outData.size(); ++i) |
| 116 | { |
| 117 | outData[i] = blob.data(boost::numeric_cast<int>(i)); |
| 118 | } |
| 119 | } |
| 120 | |
| 121 | bool IsInRange(unsigned int value, unsigned int min, unsigned int max) |
| 122 | { |
| 123 | return (value >= min && value <= max) ? true : false; |
| 124 | } |
| 125 | |
| 126 | template <typename T> |
| 127 | size_t SizeOfVectorData(const vector<T>& vec) |
| 128 | { |
| 129 | return vec.size() * sizeof(T); |
| 130 | } |
| 131 | |
| 132 | void ValidateNumInputsOutputs(const caffe::LayerParameter& layerParameter, |
| 133 | unsigned int numInputs, |
| 134 | unsigned int numOutputs) |
| 135 | { |
| 136 | int numInputsActual = layerParameter.bottom_size(); |
| 137 | if (numInputs != boost::numeric_cast<unsigned int>(numInputsActual)) |
| 138 | { |
| 139 | throw ParseException("Loading layer: invalid number of inputs"); |
| 140 | } |
| 141 | |
| 142 | int numOutputsActual = layerParameter.top_size(); |
| 143 | if (numOutputs != boost::numeric_cast<unsigned int>(numOutputsActual)) |
| 144 | { |
| 145 | throw ParseException("Loading layer: invalid number of outputs"); |
| 146 | } |
| 147 | } |
| 148 | |
| 149 | BindingPointInfo CaffeParser::GetNetworkInputBindingInfo(const std::string& name) const |
| 150 | { |
| 151 | return GetBindingInfo(name, "input", m_NetworkInputsBindingInfo); |
| 152 | } |
| 153 | |
| 154 | BindingPointInfo CaffeParser::GetNetworkOutputBindingInfo(const std::string& name) const |
| 155 | { |
| 156 | return GetBindingInfo(name, "output", m_NetworkOutputsBindingInfo); |
| 157 | } |
| 158 | |
| 159 | std::pair<armnn::LayerBindingId, armnn::TensorInfo> CaffeParser::GetBindingInfo(const std::string& layerName, |
| 160 | const char* bindingPointDesc, |
| 161 | const std::unordered_map<std::string, BindingPointInfo>& nameToBindingInfo) |
| 162 | { |
| 163 | auto it = nameToBindingInfo.find(layerName); |
| 164 | if (it == nameToBindingInfo.end()) |
| 165 | { |
| 166 | throw InvalidArgumentException(boost::str(boost::format("Unknown %1% '%2%'") % bindingPointDesc % layerName)); |
| 167 | } |
| 168 | return it->second; |
| 169 | } |
| 170 | |
| 171 | TensorInfo CaffeParser::BlobShapeToTensorInfo(const caffe::BlobShape& blobShape) const |
| 172 | { |
| 173 | std::vector<unsigned int> shape; |
| 174 | for (int j = 0; j < blobShape.dim_size(); ++j) |
| 175 | { |
| 176 | shape.push_back(static_cast<unsigned int>(blobShape.dim(j))); |
| 177 | } |
| 178 | |
| 179 | return TensorInfo(boost::numeric_cast<unsigned int>(shape.size()), shape.data(), DataType::Float32); |
| 180 | } |
| 181 | |
| 182 | BlobShape TensorDescToBlobShape(const TensorInfo& desc) |
| 183 | { |
| 184 | BlobShape ret; |
| 185 | for (unsigned int i = 0; i < desc.GetNumDimensions(); ++i) |
| 186 | { |
| 187 | ret.add_dim(i); |
| 188 | ret.set_dim(boost::numeric_cast<int>(i), desc.GetShape()[i]); |
| 189 | } |
| 190 | |
| 191 | return ret; |
| 192 | } |
| 193 | |
| 194 | vector<const LayerParameter*> CaffeParser::GetInputs(const LayerParameter& layerParam) |
| 195 | { |
| 196 | std::vector<const caffe::LayerParameter*> ret; |
| 197 | ret.reserve(boost::numeric_cast<size_t>(layerParam.bottom_size())); |
| 198 | for (int j = 0; j < layerParam.bottom_size(); ++j) |
| 199 | { |
| 200 | std::string inputName = layerParam.bottom(j); |
| 201 | auto inputIt = m_CaffeLayersByTopName.find(inputName); |
| 202 | if (inputIt == m_CaffeLayersByTopName.end()) |
| 203 | { |
| 204 | throw ParseException( |
| 205 | "Can't find Caffe layer with top called '" + inputName + "', which is listed as an input of '" + |
| 206 | layerParam.name() + "'"); |
| 207 | } |
| 208 | ret.push_back(inputIt->second); |
| 209 | } |
| 210 | |
| 211 | return ret; |
| 212 | } |
| 213 | |
| 214 | void CaffeParser::ParseInputLayer(const LayerParameter& layerParam) |
| 215 | { |
| 216 | BOOST_ASSERT(layerParam.type() == "Input"); |
| 217 | ValidateNumInputsOutputs(layerParam, 0, 1); |
| 218 | |
| 219 | const InputParameter& param = layerParam.input_param(); |
| 220 | |
| 221 | const armnn::LayerBindingId inputId = boost::numeric_cast<armnn::LayerBindingId>(m_NetworkInputsBindingInfo.size()); |
| 222 | armnn::IConnectableLayer* const inputLayer = m_Network->AddInputLayer(inputId, layerParam.name().c_str()); |
| 223 | |
| 224 | // Decide on the tensor info for this input. This can be specified in the Caffe network but can also |
| 225 | // be overriden by user input (m_inputShapes). |
| 226 | armnn::TensorInfo inputTensorInfo; |
| 227 | |
| 228 | const BlobShape* originalShape = param.shape_size() > 0 && param.shape(0).dim_size() > 0 ? |
| 229 | ¶m.shape(0) : nullptr; |
| 230 | if (originalShape) |
| 231 | { |
| 232 | inputTensorInfo = BlobShapeToTensorInfo(*originalShape); |
| 233 | } |
| 234 | |
| 235 | auto overrideIt = m_InputShapes.find(layerParam.name()); |
| 236 | if (overrideIt != m_InputShapes.end()) |
| 237 | { |
| 238 | const TensorShape& overrideShape = overrideIt->second; |
| 239 | if (originalShape && |
| 240 | ( originalShape->dim(1) != overrideShape[1] |
| 241 | || originalShape->dim(2) != overrideShape[2] |
| 242 | || originalShape->dim(3) != overrideShape[3])) |
| 243 | { |
| 244 | throw ParseException("Parsed input shape for '" + layerParam.name() + |
| 245 | "' is incompatible with the override provided"); |
| 246 | } |
| 247 | inputTensorInfo.SetShape(overrideShape); |
| 248 | } |
| 249 | else if (!originalShape) |
| 250 | { |
| 251 | throw ParseException("No input descriptor given for '" + layerParam.name() + |
| 252 | "' and no input shape found in caffe model"); |
| 253 | } |
| 254 | |
| 255 | TrackInputBinding(inputLayer, inputId, inputTensorInfo); |
| 256 | inputLayer->GetOutputSlot(0).SetTensorInfo(inputTensorInfo); |
| 257 | SetArmnnOutputSlotForCaffeTop(layerParam.top(0), inputLayer->GetOutputSlot(0)); |
| 258 | } |
| 259 | |
| 260 | void CaffeParser::ParseConvLayer(const LayerParameter& layerParam) |
| 261 | { |
| 262 | BOOST_ASSERT(layerParam.type() == "Convolution"); |
| 263 | ValidateNumInputsOutputs(layerParam, 1, 1); |
| 264 | |
| 265 | ConvolutionParameter convParam = layerParam.convolution_param(); |
| 266 | BlobShape inputShape = TensorDescToBlobShape(GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).GetTensorInfo()); |
| 267 | |
| 268 | unsigned int kernelH = 0; |
| 269 | unsigned int kernelW = 0; |
| 270 | if (convParam.has_kernel_h() && convParam.has_kernel_w()) |
| 271 | { |
| 272 | kernelH = convParam.kernel_h(); |
| 273 | kernelW = convParam.kernel_w(); |
| 274 | } |
| 275 | else if (convParam.kernel_size_size() > 0) |
| 276 | { |
| 277 | kernelH = (convParam.kernel_size()).Get(0); |
| 278 | kernelW = (convParam.kernel_size()).Get(0); |
| 279 | } |
| 280 | else |
| 281 | { |
| 282 | throw ParseException("Loading Convolution Layer: Kernel Size defined Illegally"); |
| 283 | } |
| 284 | |
| 285 | if (!IsInRange(kernelH, 0, 11) || !IsInRange(kernelW, 0, 11) || (kernelH != kernelW)) |
| 286 | { |
| 287 | throw ParseException("Loading Convolution Layer: Kernel has invalid size"); |
| 288 | } |
| 289 | |
| 290 | unsigned int strideH = 0; |
| 291 | unsigned int strideW = 0; |
| 292 | |
| 293 | if (convParam.has_stride_h() && convParam.has_stride_w()) |
| 294 | { |
| 295 | strideH = convParam.stride_h(); |
| 296 | strideW = convParam.stride_w(); |
| 297 | } |
| 298 | else if (convParam.stride_size() > 0) |
| 299 | { |
| 300 | strideH = (convParam.stride()).Get(0); |
| 301 | strideW = (convParam.stride()).Get(0); |
| 302 | } |
| 303 | else |
| 304 | { |
| 305 | // Caffe stride default is 1 |
| 306 | strideH = strideW = 1; |
| 307 | } |
| 308 | |
| 309 | if (!IsInRange(strideH, 0, 11) || !IsInRange(strideW, 0, 11) || (strideH != strideW)) |
| 310 | { |
| 311 | throw ParseException("Loading Convolution Layer: stride has invalid size"); |
| 312 | } |
| 313 | |
| 314 | unsigned int padH = 0; |
| 315 | unsigned int padW = 0; |
| 316 | |
| 317 | if (convParam.has_pad_h() && convParam.has_pad_w()) |
| 318 | { |
| 319 | padH = convParam.pad_h(); |
| 320 | padW = convParam.pad_w(); |
| 321 | } |
| 322 | else if (convParam.pad_size() > 0) |
| 323 | { |
| 324 | padH = (convParam.pad()).Get(0); |
| 325 | padW = (convParam.pad()).Get(0); |
| 326 | } |
| 327 | else |
| 328 | { |
| 329 | padH = 0; |
| 330 | padW = 0; |
| 331 | } |
| 332 | |
| 333 | if (!IsInRange(padH, 0, 11) || !IsInRange(padW, 0, 11) || (padH != padW)) |
| 334 | { |
| 335 | throw ParseException("Loading Convolution Layer: pad has invalid size"); |
| 336 | } |
| 337 | |
| 338 | // Handle grouping |
| 339 | const unsigned int numGroups = convParam.has_group() ? convParam.group() : 1; |
| 340 | armnn::IOutputSlot& inputConnection = GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)); |
| 341 | |
| 342 | vector<string> convLayerNames(numGroups); |
| 343 | vector<armnn::IConnectableLayer*> convLayers(numGroups); |
| 344 | convLayerNames[0] = layerParam.name(); |
| 345 | |
| 346 | armnn::IConnectableLayer* splitterLayer = nullptr; |
| 347 | if (numGroups > 1) |
| 348 | { |
| 349 | // This convolution is to be applied to chunks of the input data so add a splitter layer |
| 350 | |
| 351 | // Redirect the convolution input to the splitter |
| 352 | unsigned int splitterDimSizes[4] = {static_cast<unsigned int>(inputShape.dim(0)), |
| 353 | static_cast<unsigned int>(inputShape.dim(1)), |
| 354 | static_cast<unsigned int>(inputShape.dim(2)), |
| 355 | static_cast<unsigned int>(inputShape.dim(3))}; |
| 356 | |
| 357 | // Split dimension 1 of the splitter output shape and conv input shapes |
| 358 | // according to the number of groups |
| 359 | splitterDimSizes[1] /= numGroups; |
| 360 | inputShape.set_dim(1, splitterDimSizes[1]); |
| 361 | |
| 362 | // This is used to describe how the input is to be split |
| 363 | ViewsDescriptor splitterDesc(numGroups); |
| 364 | |
| 365 | // Create an output node for each group, giving each a unique name |
| 366 | for (unsigned int g = 0; g < numGroups; ++g) |
| 367 | { |
| 368 | // Work out the names of the splitter layers child convolutions |
| 369 | stringstream ss; |
| 370 | ss << layerParam.name() << "_" << g; |
| 371 | convLayerNames[g] = ss.str(); |
| 372 | |
| 373 | splitterDesc.SetViewOriginCoord(g, 1, splitterDimSizes[1] * g); |
| 374 | |
| 375 | // Set the size of the views. |
| 376 | for (unsigned int dimIdx=0; dimIdx < 4; dimIdx++) |
| 377 | { |
| 378 | splitterDesc.SetViewSize(g, dimIdx, splitterDimSizes[dimIdx]); |
| 379 | } |
| 380 | } |
| 381 | |
| 382 | const std::string splitterLayerName = std::string("splitter_") + layerParam.bottom(0); |
| 383 | |
| 384 | // Add the splitter layer |
| 385 | splitterLayer = m_Network->AddSplitterLayer(splitterDesc, |
| 386 | splitterLayerName.c_str()); |
| 387 | |
| 388 | inputConnection.Connect(splitterLayer->GetInputSlot(0)); |
| 389 | for (unsigned int i = 0; i < splitterLayer->GetNumOutputSlots(); i++) |
| 390 | { |
| 391 | splitterLayer->GetOutputSlot(i).SetTensorInfo(BlobShapeToTensorInfo(inputShape)); |
| 392 | } |
| 393 | } |
| 394 | |
| 395 | // Ignored Caffe Parameters |
| 396 | // * Dilation Size |
| 397 | // * Weight Filler |
| 398 | // * Bias Filler |
| 399 | // * Engine |
| 400 | // * Force nd_im2col |
| 401 | // * Axis |
| 402 | |
| 403 | // Not Available ArmNN Interface Parameters |
| 404 | // * Rounding policy; |
| 405 | |
| 406 | Convolution2dDescriptor convolution2dDescriptor; |
| 407 | convolution2dDescriptor.m_PadLeft = padW; |
| 408 | convolution2dDescriptor.m_PadRight = padW; |
| 409 | convolution2dDescriptor.m_PadTop = padH; |
| 410 | convolution2dDescriptor.m_PadBottom = padH; |
| 411 | convolution2dDescriptor.m_StrideX = strideW; |
| 412 | convolution2dDescriptor.m_StrideY = strideH; |
| 413 | |
| 414 | unsigned int numFilters = convParam.num_output(); |
| 415 | |
| 416 | // Populate convolution output tensor descriptor dimensions |
| 417 | BlobShape outputShape; |
| 418 | outputShape.add_dim(0); |
| 419 | outputShape.set_dim(0, inputShape.dim(0)); |
| 420 | outputShape.add_dim(1); |
| 421 | // Ensure that dimension 1 of the convolution output is split according to the number of groups. |
| 422 | outputShape.set_dim(1, numFilters / numGroups); |
| 423 | outputShape.add_dim(2); |
| 424 | outputShape.set_dim( |
| 425 | 2, (static_cast<int>(static_cast<float>(inputShape.dim(2) + 2 * padH - kernelH) / |
| 426 | boost::numeric_cast<float>(strideH)) + 1)); |
| 427 | outputShape.add_dim(3); |
| 428 | outputShape.set_dim( |
| 429 | 3, (static_cast<int>(static_cast<float>(inputShape.dim(3) + 2 * padW - kernelW) / |
| 430 | boost::numeric_cast<float>(strideW)) + 1)); |
| 431 | |
| 432 | // Load the weight data for ALL groups |
| 433 | vector<float> weightData(boost::numeric_cast<size_t>(numGroups * inputShape.dim(1) * outputShape.dim(1) * |
| 434 | kernelH * kernelW)); |
| 435 | GetDataFromBlob(layerParam, weightData, 0); |
| 436 | |
| 437 | const unsigned int weightDimSizes[4] = { |
| 438 | static_cast<unsigned int>(outputShape.dim(1)), static_cast<unsigned int>(inputShape.dim(1)), kernelH, kernelW}; |
| 439 | |
| 440 | // Bias data - This defaults to true in Caffe |
| 441 | TensorInfo biasInfo; |
| 442 | vector<float> biasData; |
| 443 | convolution2dDescriptor.m_BiasEnabled = convParam.has_bias_term() ? convParam.bias_term() : true; |
| 444 | if (convolution2dDescriptor.m_BiasEnabled) |
| 445 | { |
| 446 | biasData.resize(boost::numeric_cast<size_t>(numGroups * outputShape.dim(1)), 1.f); |
| 447 | GetDataFromBlob(layerParam, biasData, 1); |
| 448 | |
| 449 | const unsigned int biasDimSizes[1] = {static_cast<unsigned int>(outputShape.dim(1))}; |
| 450 | biasInfo = TensorInfo(1, biasDimSizes, DataType::Float32); |
| 451 | } |
| 452 | |
| 453 | const unsigned int numWeightsPerGroup = boost::numeric_cast<unsigned int>(weightData.size()) / numGroups; |
| 454 | const unsigned int numBiasesPerGroup = boost::numeric_cast<unsigned int>(biasData.size()) / numGroups; |
| 455 | |
| 456 | armnn::IConnectableLayer* returnLayer = nullptr; |
| 457 | |
| 458 | for (unsigned int g = 0; g < numGroups; ++g) |
| 459 | { |
| 460 | // set the slot index, group 0 should be connected to the 0th output of the splitter |
| 461 | // group 1 should be connected to the 1st output of the splitter |
| 462 | |
| 463 | // Pull out the weights for this group from that loaded from the model file earlier |
| 464 | ConstTensor weights(TensorInfo(4, weightDimSizes, DataType::Float32), |
| 465 | weightData.data() + numWeightsPerGroup * g); |
| 466 | |
| 467 | IConnectableLayer* convLayer = nullptr; |
| 468 | if (convolution2dDescriptor.m_BiasEnabled) |
| 469 | { |
| 470 | // Pull out the biases for this group from that loaded from the model file earlier |
| 471 | ConstTensor biases(biasInfo, biasData.data() + numBiasesPerGroup * g); |
| 472 | |
| 473 | convLayer = m_Network->AddConvolution2dLayer(convolution2dDescriptor, |
| 474 | weights, biases, convLayerNames[g].c_str()); |
| 475 | } |
| 476 | else |
| 477 | { |
| 478 | convLayer = m_Network->AddConvolution2dLayer(convolution2dDescriptor, |
| 479 | weights, convLayerNames[g].c_str()); |
| 480 | } |
| 481 | convLayers[g] = convLayer; |
| 482 | |
| 483 | // If we have more than one group then the input to the nth convolution the splitter layer's nth output, |
| 484 | // otherwise it's the regular input to this layer. |
| 485 | armnn::IOutputSlot& splitterInputConnection = splitterLayer ? splitterLayer->GetOutputSlot(g) : inputConnection; |
| 486 | splitterInputConnection.Connect(convLayer->GetInputSlot(0)); |
| 487 | convLayer->GetOutputSlot(0).SetTensorInfo(BlobShapeToTensorInfo(outputShape)); |
| 488 | |
| 489 | returnLayer = convLayer; |
| 490 | } |
| 491 | |
| 492 | if (numGroups > 1) |
| 493 | { |
| 494 | // If the convolution was performed in chunks, add a layer to merge the results |
| 495 | |
| 496 | // The merge input shape matches that of the convolution output |
| 497 | unsigned int mergeDimSizes[4] = {static_cast<unsigned int>(outputShape.dim(0)), |
| 498 | static_cast<unsigned int>(outputShape.dim(1)), |
| 499 | static_cast<unsigned int>(outputShape.dim(2)), |
| 500 | static_cast<unsigned int>(outputShape.dim(3))}; |
| 501 | |
| 502 | // This is used to describe how the input is to be merged |
| 503 | OriginsDescriptor mergeDesc(numGroups); |
| 504 | |
| 505 | // Now create an input node for each group, using the name from |
| 506 | // the output of the corresponding convolution |
| 507 | for (unsigned int g = 0; g < numGroups; ++g) |
| 508 | { |
| 509 | mergeDesc.SetViewOriginCoord(g, 1, mergeDimSizes[1] * g); |
| 510 | } |
| 511 | |
| 512 | // Make sure the output from the merge is the correct size to hold the data for all groups |
| 513 | mergeDimSizes[1] *= numGroups; |
| 514 | outputShape.set_dim(1, mergeDimSizes[1]); |
| 515 | |
| 516 | // The merge layer just assumes the name of the original convolution |
| 517 | // layer so the following layer connection "just works" |
| 518 | const string mergeOutputName = layerParam.name(); |
| 519 | |
| 520 | // Finally add the merge layer |
| 521 | IConnectableLayer* layer = m_Network->AddMergerLayer(mergeDesc, mergeOutputName.c_str()); |
| 522 | |
| 523 | for (unsigned int g = 0; g < numGroups; ++g) |
| 524 | { |
| 525 | convLayers[g]->GetOutputSlot(0).Connect(layer->GetInputSlot(g)); |
| 526 | } |
| 527 | layer->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo(4, mergeDimSizes, DataType::Float32)); |
| 528 | |
| 529 | returnLayer = layer; |
| 530 | } |
| 531 | |
| 532 | BOOST_ASSERT(returnLayer); |
| 533 | SetArmnnOutputSlotForCaffeTop(layerParam.top(0), returnLayer->GetOutputSlot(0)); |
| 534 | } |
| 535 | |
| 536 | void CaffeParser::ParsePoolingLayer(const LayerParameter& layerParam) |
| 537 | { |
| 538 | ValidateNumInputsOutputs(layerParam, 1, 1); |
| 539 | |
| 540 | PoolingParameter param = layerParam.pooling_param(); |
| 541 | |
| 542 | const TensorInfo& inputInfo = GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).GetTensorInfo(); |
| 543 | |
| 544 | // Kernel size |
| 545 | unsigned int kernel_h = 0; |
| 546 | unsigned int kernel_w = 0; |
| 547 | if (param.has_kernel_h() && param.has_kernel_w()) |
| 548 | { |
| 549 | kernel_h = param.kernel_h(); |
| 550 | kernel_w = param.kernel_w(); |
| 551 | } |
| 552 | else if (param.kernel_size() > 0) |
| 553 | { |
| 554 | kernel_h = param.kernel_size(); |
| 555 | kernel_w = param.kernel_size(); |
| 556 | } |
| 557 | else if (param.has_global_pooling()) |
| 558 | { |
| 559 | kernel_h = inputInfo.GetShape()[2]; |
| 560 | kernel_w = inputInfo.GetShape()[3]; |
| 561 | } |
| 562 | else |
| 563 | { |
| 564 | throw ParseException("Loading Pooling Layer: Kernel Size defined Illegally"); |
| 565 | } |
| 566 | |
| 567 | if (!IsInRange(kernel_h, 0, 11) || !IsInRange(kernel_w, 0, 11) || (kernel_h != kernel_w)) |
| 568 | { |
| 569 | throw ParseException(boost::str( |
| 570 | boost::format("Loading Pooling Layer: kernel has invalid size: %1% x %2%") % kernel_h % kernel_w)); |
| 571 | } |
| 572 | |
| 573 | // Strides |
| 574 | // Default to a valid value for the case of global pooling (where the strides don't have to be explicitly set) |
| 575 | unsigned int stride_h = 1; |
| 576 | unsigned int stride_w = 1; |
| 577 | if (param.has_stride_h() && param.has_stride_w()) |
| 578 | { |
| 579 | stride_h = param.stride_h(); |
| 580 | stride_w = param.stride_w(); |
| 581 | } |
| 582 | else if (param.has_stride()) |
| 583 | { |
| 584 | stride_h = param.stride(); |
| 585 | stride_w = param.stride(); |
| 586 | } |
| 587 | else if (!param.has_global_pooling()) |
| 588 | { |
| 589 | throw ParseException("Loading Pooling Layer: Stride Size defined Illegally"); |
| 590 | } |
| 591 | |
| 592 | if (!IsInRange(stride_h, 0, 11) || !IsInRange(stride_w, 0, 11) || (stride_h != stride_w)) |
| 593 | { |
| 594 | throw ParseException("Loading Pooling Layer: stride has invalid size"); |
| 595 | } |
| 596 | |
| 597 | // Padding |
| 598 | unsigned int pad_h = 0; |
| 599 | unsigned int pad_w = 0; |
| 600 | if (param.has_pad_h() && param.has_pad_w()) |
| 601 | { |
| 602 | pad_h = param.pad_h(); |
| 603 | pad_w = param.pad_w(); |
| 604 | } |
| 605 | else if (param.has_pad()) |
| 606 | { |
| 607 | pad_h = param.pad(); |
| 608 | pad_w = param.pad(); |
| 609 | } |
| 610 | else |
| 611 | { |
| 612 | pad_h = 0; |
| 613 | pad_w = 0; |
| 614 | } |
| 615 | |
| 616 | if (!IsInRange(pad_h, 0, 11) || !IsInRange(pad_w, 0, 11) || (pad_h != pad_w)) |
| 617 | { |
| 618 | throw ParseException("Loading Pooling Layer: pad has invalid size"); |
| 619 | } |
| 620 | |
| 621 | // Ignored Caffe Parameters |
| 622 | // Stochastic Pooling |
| 623 | // Engine |
| 624 | |
| 625 | // Populate Weight and Bias Filter Descriptor |
| 626 | Pooling2dDescriptor pooling2dDescriptor; |
| 627 | if (param.has_pool()) |
| 628 | { |
| 629 | PoolingParameter_PoolMethod p = param.pool(); |
| 630 | switch (p) |
| 631 | { |
| 632 | case PoolingParameter_PoolMethod_MAX: |
| 633 | { |
| 634 | pooling2dDescriptor.m_PoolType = PoolingAlgorithm::Max; |
| 635 | break; |
| 636 | } |
| 637 | case PoolingParameter_PoolMethod_AVE: |
| 638 | { |
| 639 | pooling2dDescriptor.m_PoolType = PoolingAlgorithm::Average; |
| 640 | break; |
| 641 | } |
| 642 | case PoolingParameter_PoolMethod_STOCHASTIC: |
| 643 | { |
| 644 | throw ParseException("Loading Pooling Layer: Stochastic Pooling Not Supported"); |
| 645 | } |
| 646 | default: |
| 647 | { |
| 648 | throw ParseException("Loading Pooling Layer: Mode Not Supported"); |
| 649 | } |
| 650 | } |
| 651 | } |
| 652 | else |
| 653 | { |
| 654 | throw ParseException("Loading Pooling Layer: No Pooling Method Defined"); |
| 655 | } |
| 656 | |
| 657 | pooling2dDescriptor.m_PadLeft = pad_w; |
| 658 | pooling2dDescriptor.m_PadRight = pad_w; |
| 659 | pooling2dDescriptor.m_PadTop = pad_h; |
| 660 | pooling2dDescriptor.m_PadBottom = pad_h; |
| 661 | pooling2dDescriptor.m_StrideX = stride_w; |
| 662 | pooling2dDescriptor.m_StrideY = stride_h; |
| 663 | pooling2dDescriptor.m_PoolWidth = kernel_w; |
| 664 | pooling2dDescriptor.m_PoolHeight = kernel_h; |
| 665 | |
| 666 | pooling2dDescriptor.m_OutputShapeRounding = OutputShapeRounding::Ceiling; |
| 667 | pooling2dDescriptor.m_PaddingMethod = PaddingMethod::IgnoreValue; |
| 668 | |
| 669 | armnn::IConnectableLayer* poolingLayer = m_Network->AddPooling2dLayer(pooling2dDescriptor, |
| 670 | layerParam.name().c_str()); |
| 671 | |
| 672 | |
| 673 | TensorInfo outputInfo( |
| 674 | { inputInfo.GetShape()[0], |
| 675 | inputInfo.GetShape()[1], |
| 676 | static_cast<unsigned int>(ceil( |
| 677 | static_cast<float>(inputInfo.GetShape()[2] + 2 * pad_h - kernel_h) / |
| 678 | boost::numeric_cast<float>(stride_h))) + 1, |
| 679 | static_cast<unsigned int>(ceil( |
| 680 | static_cast<float>(inputInfo.GetShape()[3] + 2 * pad_w - kernel_w) / |
| 681 | boost::numeric_cast<float>(stride_w))) + 1 }, |
| 682 | DataType::Float32); |
| 683 | |
| 684 | GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).Connect(poolingLayer->GetInputSlot(0)); |
| 685 | poolingLayer->GetOutputSlot(0).SetTensorInfo(outputInfo); |
| 686 | SetArmnnOutputSlotForCaffeTop(layerParam.top(0), poolingLayer->GetOutputSlot(0)); |
| 687 | } |
| 688 | |
| 689 | void CaffeParser::ParseReluLayer(const LayerParameter& layerParam) |
| 690 | { |
| 691 | ValidateNumInputsOutputs(layerParam, 1, 1); |
| 692 | |
| 693 | const string& name = layerParam.name(); |
| 694 | const ReLUParameter& param = layerParam.relu_param(); |
| 695 | |
| 696 | ActivationDescriptor activationDescriptor; |
| 697 | const float negativeSlope = param.negative_slope(); |
| 698 | if (negativeSlope == 0.0f) |
| 699 | { |
| 700 | activationDescriptor.m_Function = ActivationFunction::ReLu; |
| 701 | } |
| 702 | else |
| 703 | { |
| 704 | activationDescriptor.m_Function = ActivationFunction::LeakyReLu; |
| 705 | activationDescriptor.m_A = negativeSlope; |
| 706 | } |
| 707 | |
| 708 | const TensorInfo& inputInfo = GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).GetTensorInfo(); |
| 709 | IConnectableLayer* const activationLayer = m_Network->AddActivationLayer(activationDescriptor, name.c_str()); |
| 710 | GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).Connect(activationLayer->GetInputSlot(0)); |
| 711 | activationLayer->GetOutputSlot(0).SetTensorInfo(inputInfo); |
| 712 | SetArmnnOutputSlotForCaffeTop(layerParam.top(0), activationLayer->GetOutputSlot(0)); |
| 713 | } |
| 714 | |
| 715 | void CaffeParser::ParseLRNLayer(const LayerParameter& layerParam) |
| 716 | { |
| 717 | ValidateNumInputsOutputs(layerParam, 1, 1); |
| 718 | |
| 719 | LRNParameter param = layerParam.lrn_param(); |
| 720 | |
| 721 | const TensorInfo& inputInfo = GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).GetTensorInfo(); |
| 722 | |
| 723 | // Ignored BATCH NORMALIZATION Caffe Parameters |
| 724 | // Ignored MVN Caffe Parameters |
| 725 | // Ignored LRN Caffe Parameters |
| 726 | // Engine |
| 727 | |
| 728 | NormalizationDescriptor normalizationDescriptor; |
| 729 | if (param.has_norm_region()) |
| 730 | { |
| 731 | LRNParameter_NormRegion n = param.norm_region(); |
| 732 | switch (n) |
| 733 | { |
| 734 | case LRNParameter_NormRegion_ACROSS_CHANNELS: |
| 735 | { |
| 736 | normalizationDescriptor.m_NormChannelType = NormalizationAlgorithmChannel::Across; |
| 737 | break; |
| 738 | } |
| 739 | case LRNParameter_NormRegion_WITHIN_CHANNEL: |
| 740 | { |
| 741 | normalizationDescriptor.m_NormChannelType = NormalizationAlgorithmChannel::Within; |
| 742 | break; |
| 743 | } |
| 744 | default: |
| 745 | throw ParseException("Loading LRN Layer: Mode Not Supported"); |
| 746 | } |
| 747 | } |
| 748 | else |
| 749 | { |
| 750 | // Caffe defaults to normalization across channels |
| 751 | normalizationDescriptor.m_NormChannelType = NormalizationAlgorithmChannel::Across; |
| 752 | } |
| 753 | |
| 754 | normalizationDescriptor.m_NormMethodType = NormalizationAlgorithmMethod::LocalBrightness; |
| 755 | if (param.has_local_size()) |
| 756 | { |
| 757 | normalizationDescriptor.m_NormSize = param.local_size(); |
| 758 | } |
| 759 | else |
| 760 | { |
| 761 | throw ParseException("Loading LRN Layer: Local_size not defined"); |
| 762 | } |
| 763 | |
| 764 | if (param.has_alpha()) |
| 765 | { |
| 766 | normalizationDescriptor.m_Alpha = param.alpha(); |
| 767 | normalizationDescriptor.m_Alpha /= boost::numeric_cast<float>(param.local_size()); |
| 768 | } |
| 769 | else |
| 770 | { |
| 771 | throw ParseException("Loading LRN Layer: Alpha not defined"); |
| 772 | } |
| 773 | if (param.has_beta()) |
| 774 | { |
| 775 | normalizationDescriptor.m_Beta = param.beta(); |
| 776 | } |
| 777 | else |
| 778 | { |
| 779 | throw ParseException("Loading LRN Layer: Beta not defined"); |
| 780 | } |
| 781 | if (param.has_k()) |
| 782 | { |
| 783 | normalizationDescriptor.m_K = param.k(); |
| 784 | } |
| 785 | else |
| 786 | normalizationDescriptor.m_K = 1; |
| 787 | |
| 788 | IConnectableLayer* const normLayer = m_Network->AddNormalizationLayer(normalizationDescriptor, |
| 789 | layerParam.name().c_str()); |
| 790 | GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).Connect(normLayer->GetInputSlot(0)); |
| 791 | normLayer->GetOutputSlot(0).SetTensorInfo(inputInfo); |
| 792 | |
| 793 | SetArmnnOutputSlotForCaffeTop(layerParam.top(0), normLayer->GetOutputSlot(0)); |
| 794 | } |
| 795 | |
| 796 | void CaffeParser::ParseInnerProductLayer(const LayerParameter& layerParam) |
| 797 | { |
| 798 | InnerProductParameter param = layerParam.inner_product_param(); |
| 799 | |
| 800 | ValidateNumInputsOutputs(layerParam, 1, 1); |
| 801 | |
| 802 | unsigned int outputSize = param.num_output(); |
| 803 | |
| 804 | // Ignored Caffe Parameters |
| 805 | // Weight Filler |
| 806 | // Bias Filler |
| 807 | // Engine |
| 808 | // Axis |
| 809 | |
| 810 | FullyConnectedDescriptor tensorFullyConnectedDescriptor; |
| 811 | |
| 812 | if (param.has_transpose()) |
| 813 | { |
| 814 | // If true assume transposed weights |
| 815 | tensorFullyConnectedDescriptor.m_TransposeWeightMatrix = param.transpose(); |
| 816 | } |
| 817 | else |
| 818 | { |
| 819 | // caffe defaults to transposed |
| 820 | tensorFullyConnectedDescriptor.m_TransposeWeightMatrix = true; |
| 821 | } |
| 822 | |
| 823 | const TensorInfo& inputInfo = GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).GetTensorInfo(); |
| 824 | |
| 825 | TensorInfo weightInfo; |
| 826 | TensorInfo biasInfo; |
| 827 | |
| 828 | // allow implicit flattening of extra dimensions |
| 829 | unsigned int inputSize = inputInfo.GetShape()[1]; |
| 830 | for (unsigned int i = 2; i < inputInfo.GetNumDimensions(); ++i) |
| 831 | { |
| 832 | inputSize *= inputInfo.GetShape()[i]; |
| 833 | } |
| 834 | |
| 835 | vector<float> weightData(inputSize * outputSize); |
| 836 | |
| 837 | GetDataFromBlob(layerParam, weightData, 0); |
| 838 | const unsigned int swTD[2] = { outputSize, inputSize }; |
| 839 | ConstTensor weights(TensorInfo(2, swTD, DataType::Float32), weightData); |
| 840 | |
| 841 | tensorFullyConnectedDescriptor.m_BiasEnabled = true; |
| 842 | // Todo: check whether bias enabled |
| 843 | armnn::IConnectableLayer* fullyConnectedLayer = nullptr; |
| 844 | if (tensorFullyConnectedDescriptor.m_BiasEnabled) |
| 845 | { |
| 846 | // BIAS VALUE |
| 847 | vector<float> biasData(outputSize); |
| 848 | |
| 849 | GetDataFromBlob(layerParam, biasData, 1); |
| 850 | |
| 851 | const unsigned int sbTD[1] = { outputSize }; |
| 852 | |
| 853 | ConstTensor biases(TensorInfo(1, sbTD, DataType::Float32), biasData); |
| 854 | |
| 855 | fullyConnectedLayer = m_Network->AddFullyConnectedLayer(tensorFullyConnectedDescriptor, weights, biases, |
| 856 | layerParam.name().c_str()); |
| 857 | } |
| 858 | else |
| 859 | { |
| 860 | fullyConnectedLayer = m_Network->AddFullyConnectedLayer(tensorFullyConnectedDescriptor, weights, |
| 861 | layerParam.name().c_str()); |
| 862 | } |
| 863 | |
| 864 | TensorInfo outputInfo({ inputInfo.GetShape()[0], outputSize }, DataType::Float32); |
| 865 | GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).Connect(fullyConnectedLayer->GetInputSlot(0)); |
| 866 | fullyConnectedLayer->GetOutputSlot(0).SetTensorInfo(outputInfo); |
| 867 | SetArmnnOutputSlotForCaffeTop(layerParam.top(0), fullyConnectedLayer->GetOutputSlot(0)); |
| 868 | } |
| 869 | |
| 870 | void CaffeParser::ParseSoftmaxLayer(const LayerParameter& layerParam) |
| 871 | { |
| 872 | ValidateNumInputsOutputs(layerParam, 1, 1); |
| 873 | |
| 874 | SoftmaxParameter param = layerParam.softmax_param(); |
| 875 | |
| 876 | const TensorInfo& inputInfo = GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).GetTensorInfo(); |
| 877 | |
| 878 | // Ignored Caffe Parameters |
| 879 | // axis |
| 880 | // Engine |
| 881 | |
| 882 | armnn::SoftmaxDescriptor softmaxDescriptor; |
| 883 | armnn::IConnectableLayer* const softmaxLayer = m_Network->AddSoftmaxLayer( |
| 884 | softmaxDescriptor, |
| 885 | layerParam.name().c_str()); |
| 886 | GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).Connect(softmaxLayer->GetInputSlot(0)); |
| 887 | softmaxLayer->GetOutputSlot(0).SetTensorInfo(inputInfo); |
| 888 | SetArmnnOutputSlotForCaffeTop(layerParam.top(0), softmaxLayer->GetOutputSlot(0)); |
| 889 | } |
| 890 | |
| 891 | void CaffeParser::ParseEltwiseLayer(const LayerParameter& layerParam) |
| 892 | { |
| 893 | ValidateNumInputsOutputs(layerParam, 2, 1); |
| 894 | |
| 895 | const TensorInfo& inputInfo = GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).GetTensorInfo(); |
| 896 | |
| 897 | // Ignored Caffe Parameters |
| 898 | // coeff |
| 899 | |
| 900 | EltwiseParameter_EltwiseOp operation = EltwiseParameter_EltwiseOp_SUM; // default to sum as per caffe |
| 901 | |
| 902 | if (layerParam.has_eltwise_param() && layerParam.eltwise_param().has_operation()) |
| 903 | { |
| 904 | operation = layerParam.eltwise_param().operation(); |
| 905 | } |
| 906 | |
| 907 | armnn::IConnectableLayer* newLayer = nullptr; |
| 908 | switch (operation) |
| 909 | { |
| 910 | case EltwiseParameter_EltwiseOp_SUM: |
| 911 | { |
| 912 | newLayer = m_Network->AddAdditionLayer(layerParam.name().c_str()); |
| 913 | break; |
| 914 | } |
| 915 | case EltwiseParameter_EltwiseOp_PROD: |
| 916 | { |
| 917 | newLayer = m_Network->AddMultiplicationLayer(layerParam.name().c_str()); |
| 918 | break; |
| 919 | } |
| 920 | default: |
| 921 | { |
| 922 | throw ParseException("Unsupported operation in Eltwise layer"); |
| 923 | } |
| 924 | } |
| 925 | |
| 926 | GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).Connect(newLayer->GetInputSlot(0)); |
| 927 | GetArmnnOutputSlotForCaffeTop(layerParam.bottom(1)).Connect(newLayer->GetInputSlot(1)); |
| 928 | newLayer->GetOutputSlot(0).SetTensorInfo(inputInfo); |
| 929 | SetArmnnOutputSlotForCaffeTop(layerParam.top(0), newLayer->GetOutputSlot(0)); |
| 930 | } |
| 931 | |
| 932 | void CaffeParser::ParseConcatLayer(const LayerParameter& layerParam) |
| 933 | { |
| 934 | unsigned int numInputs = static_cast<unsigned int>(layerParam.bottom_size()); |
| 935 | // we assume concat happens along the channel dimension, which is 1 in (0, 1, 2, 3) |
| 936 | unsigned int concatDim = 1; |
| 937 | unsigned int numOfDims = 4; |
| 938 | |
| 939 | OriginsDescriptor concatDescriptor(static_cast<uint32_t>(numInputs), numOfDims);// we only consider 4-D tensor here |
| 940 | std::vector<unsigned int>mergeDimSizes(numOfDims, 0u); |
| 941 | |
| 942 | unsigned int mergeDim = 0; |
| 943 | for (unsigned int viewIndex = 0; viewIndex < numInputs; ++viewIndex) |
| 944 | { |
| 945 | const TensorInfo& inputInfo = GetArmnnOutputSlotForCaffeTop( |
| 946 | layerParam.bottom(boost::numeric_cast<int>(viewIndex))).GetTensorInfo(); |
| 947 | // Check whether the dimensions of the input tensors are actually 4 |
| 948 | if (inputInfo.GetNumDimensions()!=4) |
| 949 | { |
| 950 | throw ParseException("The number of dimensions for input tensors of the concatenation op should be 4."); |
| 951 | } |
| 952 | |
| 953 | mergeDimSizes[0] = inputInfo.GetShape()[0]; |
| 954 | mergeDimSizes[1] = inputInfo.GetShape()[1]; |
| 955 | mergeDimSizes[2] = inputInfo.GetShape()[2]; |
| 956 | mergeDimSizes[3] = inputInfo.GetShape()[3]; |
| 957 | |
| 958 | for (unsigned int j = 0; j < concatDim; ++j) |
| 959 | { |
| 960 | concatDescriptor.SetViewOriginCoord(viewIndex, j, 0); |
| 961 | } |
| 962 | |
| 963 | concatDescriptor.SetViewOriginCoord(viewIndex, concatDim, mergeDim); |
| 964 | mergeDim += mergeDimSizes[concatDim]; |
| 965 | |
| 966 | for (unsigned int j = concatDim+1; j < numOfDims; ++j) |
| 967 | { |
| 968 | concatDescriptor.SetViewOriginCoord(viewIndex, j, 0); |
| 969 | } |
| 970 | } |
| 971 | mergeDimSizes[concatDim] = mergeDim; |
| 972 | |
| 973 | armnn::IConnectableLayer *concatlayer = m_Network->AddMergerLayer(concatDescriptor, layerParam.name().c_str()); |
| 974 | for (unsigned int i = 0; i < numInputs; ++i) |
| 975 | { |
| 976 | armnn::IOutputSlot& outputSlot = GetArmnnOutputSlotForCaffeTop(layerParam.bottom(boost::numeric_cast<int>(i))); |
| 977 | outputSlot.Connect(concatlayer->GetInputSlot(i)); |
| 978 | } |
| 979 | |
| 980 | concatlayer->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo(numOfDims, mergeDimSizes.data(), DataType::Float32)); |
| 981 | SetArmnnOutputSlotForCaffeTop(layerParam.top(0), concatlayer->GetOutputSlot(0)); |
| 982 | } |
| 983 | |
| 984 | void CaffeParser::ParseBatchNormLayer(const LayerParameter& layerParam) |
| 985 | { |
| 986 | ValidateNumInputsOutputs(layerParam, 1, 1); |
| 987 | |
| 988 | const TensorInfo& inputInfo = GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).GetTensorInfo(); |
| 989 | |
| 990 | string name = layerParam.name(); |
| 991 | |
| 992 | BatchNormParameter param = layerParam.batch_norm_param(); |
| 993 | // If use_global_stats is not explicitly set in the model, assume it to be true (its default value |
| 994 | // when the network is in the testing phase). |
| 995 | if (param.has_use_global_stats()) |
| 996 | { |
| 997 | if (!param.use_global_stats()) |
| 998 | { |
| 999 | throw ParseException(boost::str(boost::format("Error parsing Batch Norm layer '%1%': " |
| 1000 | "Parameter 'use_global_stats' is set to false, which is unsupported (value used for training).") |
| 1001 | % name)); |
| 1002 | } |
| 1003 | } |
| 1004 | |
| 1005 | BatchNormalizationDescriptor desc; |
| 1006 | desc.m_Eps = param.eps(); |
| 1007 | |
| 1008 | unsigned int channels = inputInfo.GetShape()[1]; |
| 1009 | unsigned int shape[] = {channels}; |
| 1010 | |
| 1011 | vector<float> meanData(channels); |
| 1012 | GetDataFromBlob(layerParam, meanData, 0); |
| 1013 | |
| 1014 | vector<float> varianceData(channels); |
| 1015 | GetDataFromBlob(layerParam, varianceData, 1); |
| 1016 | |
| 1017 | // identity scale operation |
| 1018 | vector<float> betaData(channels, 0.0f); |
| 1019 | vector<float> gammaData(channels, 1.0f); |
| 1020 | |
| 1021 | ConstTensor mean(TensorInfo(1, shape, armnn::DataType::Float32), meanData); |
| 1022 | ConstTensor variance(TensorInfo(1, shape, armnn::DataType::Float32), varianceData); |
| 1023 | ConstTensor beta(TensorInfo(1, shape, armnn::DataType::Float32), betaData); |
| 1024 | ConstTensor gamma(TensorInfo(1, shape, armnn::DataType::Float32), gammaData); |
| 1025 | |
| 1026 | armnn::IConnectableLayer* const batchNormLayer = m_Network->AddBatchNormalizationLayer(desc, |
| 1027 | mean, variance, beta, gamma, name.c_str()); |
| 1028 | GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).Connect(batchNormLayer->GetInputSlot(0)); |
| 1029 | batchNormLayer->GetOutputSlot(0).SetTensorInfo(inputInfo); |
| 1030 | SetArmnnOutputSlotForCaffeTop(layerParam.top(0), batchNormLayer->GetOutputSlot(0)); |
| 1031 | } |
| 1032 | |
| 1033 | void CaffeParser::ParseScaleLayer(const LayerParameter& layerParam) |
| 1034 | { |
| 1035 | // current unoptimal solution: add a batchnormalization layer with 0 mean and 1 variance |
| 1036 | ValidateNumInputsOutputs(layerParam, 1, 1); |
| 1037 | |
| 1038 | const TensorInfo& inputInfo = GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).GetTensorInfo(); |
| 1039 | |
| 1040 | string name = layerParam.name(); |
| 1041 | |
| 1042 | ScaleParameter param = layerParam.scale_param(); |
| 1043 | if (param.axis() != 1) |
| 1044 | { |
| 1045 | // Would have to use something other than BatchNormalizationLayer in this case |
| 1046 | throw ParseException("Loading Scale Layer: Only axis 1 supported currently"); |
| 1047 | } |
| 1048 | |
| 1049 | unsigned int channels = inputInfo.GetShape()[1]; |
| 1050 | unsigned int shape[] = {channels}; |
| 1051 | |
| 1052 | BatchNormalizationDescriptor desc; |
| 1053 | desc.m_Eps = 0.0f; // don't need epsilon if variance is 1 |
| 1054 | vector<float> meanData(channels, 0.0f); |
| 1055 | vector<float> varianceData(channels, 1.0f); |
| 1056 | vector<float> betaData(channels, 0.0f); |
| 1057 | vector<float> gammaData(channels); |
| 1058 | |
| 1059 | GetDataFromBlob(layerParam, gammaData, 0); |
| 1060 | |
| 1061 | if(param.has_bias_term()) |
| 1062 | { |
| 1063 | GetDataFromBlob(layerParam, betaData, 1); |
| 1064 | } |
| 1065 | |
| 1066 | ConstTensor mean(TensorInfo(1, shape, armnn::DataType::Float32), meanData); |
| 1067 | ConstTensor variance(TensorInfo(1, shape, armnn::DataType::Float32), varianceData); |
| 1068 | ConstTensor beta(TensorInfo(1, shape, armnn::DataType::Float32), betaData); |
| 1069 | ConstTensor gamma(TensorInfo(1, shape, armnn::DataType::Float32), gammaData); |
| 1070 | |
| 1071 | armnn::IConnectableLayer* const batchNormLayer = m_Network->AddBatchNormalizationLayer(desc, |
| 1072 | mean, variance, beta, gamma, name.c_str()); |
| 1073 | GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).Connect(batchNormLayer->GetInputSlot(0)); |
| 1074 | batchNormLayer->GetOutputSlot(0).SetTensorInfo(inputInfo); |
| 1075 | SetArmnnOutputSlotForCaffeTop(layerParam.top(0), batchNormLayer->GetOutputSlot(0)); |
| 1076 | } |
| 1077 | |
| 1078 | void CaffeParser::ParseSplitLayer(const caffe::LayerParameter& layerParam) |
| 1079 | { |
| 1080 | // Used in caffe to duplicate memory - not necessary in armnn |
| 1081 | if (layerParam.bottom_size() != 1) |
| 1082 | { |
| 1083 | throw ParseException("Split layer '" + layerParam.name() + "' should have exactly 1 bottom"); |
| 1084 | } |
| 1085 | armnn::IOutputSlot& outputSlot = GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)); |
| 1086 | for (int i = 0; i < layerParam.top_size(); i++) |
| 1087 | { |
| 1088 | SetArmnnOutputSlotForCaffeTop(layerParam.top(i), outputSlot); |
| 1089 | } |
| 1090 | } |
| 1091 | |
| 1092 | void CaffeParser::ParseDropoutLayer(const caffe::LayerParameter& layerParam) |
| 1093 | { |
| 1094 | // Ignored for inference so patch the single input to its single output |
| 1095 | if (layerParam.bottom_size() != 1 || layerParam.top_size() != 1) |
| 1096 | { |
| 1097 | throw ParseException("Dropout layer '" + layerParam.name() + "' should have exactly 1 bottom and 1 top"); |
| 1098 | } |
| 1099 | SetArmnnOutputSlotForCaffeTop(layerParam.top(0), GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0))); |
| 1100 | } |
| 1101 | |
| 1102 | void CaffeParser::TrackInputBinding(armnn::IConnectableLayer* layer, |
| 1103 | armnn::LayerBindingId id, |
| 1104 | const armnn::TensorInfo& tensorInfo) |
| 1105 | { |
| 1106 | return TrackBindingPoint(layer, id, tensorInfo, layer->GetName(), m_NetworkInputsBindingInfo); |
| 1107 | } |
| 1108 | |
| 1109 | void CaffeParser::TrackOutputBinding(armnn::IConnectableLayer* layer, |
| 1110 | armnn::LayerBindingId id, |
| 1111 | const armnn::TensorInfo& tensorInfo) |
| 1112 | { |
| 1113 | return TrackBindingPoint(layer, id, tensorInfo, layer->GetName(), m_NetworkOutputsBindingInfo); |
| 1114 | } |
| 1115 | |
| 1116 | void CaffeParser::TrackBindingPoint(armnn::IConnectableLayer* layer, |
| 1117 | armnn::LayerBindingId id, |
| 1118 | const armnn::TensorInfo& tensorInfo, |
| 1119 | const char* bindingPointDesc, |
| 1120 | std::unordered_map<std::string, BindingPointInfo>& nameToBindingInfo) |
| 1121 | { |
| 1122 | const std::string layerName = layer->GetName(); |
| 1123 | auto it = nameToBindingInfo.find(layerName); |
| 1124 | if (it == nameToBindingInfo.end()) |
| 1125 | { |
| 1126 | nameToBindingInfo[layerName] = std::make_pair(id, tensorInfo); |
| 1127 | } |
| 1128 | else |
| 1129 | { |
| 1130 | throw ParseException(boost::str( |
| 1131 | boost::format("Id %1% used by more than one %2% layer") % id % bindingPointDesc)); |
| 1132 | } |
| 1133 | } |
| 1134 | |
| 1135 | armnn::IOutputSlot& CaffeParser::GetArmnnOutputSlotForCaffeTop(const std::string& caffeTopName) const |
| 1136 | { |
| 1137 | auto it = m_ArmnnOutputSlotForCaffeTop.find(caffeTopName); |
| 1138 | if (it != m_ArmnnOutputSlotForCaffeTop.end()) |
| 1139 | { |
| 1140 | return *it->second; |
| 1141 | } |
| 1142 | else |
| 1143 | { |
| 1144 | throw ParseException(boost::str(boost::format( |
| 1145 | "Could not find armnn output slot for Caffe top '%1%'") % caffeTopName)); |
| 1146 | } |
| 1147 | } |
| 1148 | |
| 1149 | void CaffeParser::SetArmnnOutputSlotForCaffeTop(const std::string& caffeTopName, armnn::IOutputSlot& armnnOutputSlot) |
| 1150 | { |
| 1151 | auto it = m_ArmnnOutputSlotForCaffeTop.find(caffeTopName); |
| 1152 | if (it == m_ArmnnOutputSlotForCaffeTop.end()) |
| 1153 | { |
| 1154 | m_ArmnnOutputSlotForCaffeTop[caffeTopName] = &armnnOutputSlot; |
| 1155 | } |
| 1156 | else |
| 1157 | { |
| 1158 | throw ParseException("Attempting to add duplicate entry for Caffe top '" + caffeTopName + "'"); |
| 1159 | } |
| 1160 | } |
| 1161 | |
| 1162 | void CaffeParser::ResolveInPlaceLayers(caffe::NetParameter& netParameter) |
| 1163 | { |
| 1164 | // Find layers with the same top |
| 1165 | std::map<std::string, std::vector<caffe::LayerParameter*>> layersByTop; |
| 1166 | for (int layerIdx = 0; layerIdx < netParameter.layer_size(); ++layerIdx) |
| 1167 | { |
| 1168 | caffe::LayerParameter& layer = *netParameter.mutable_layer(layerIdx); |
| 1169 | for (int i = 0; i < layer.top_size(); ++i) |
| 1170 | { |
| 1171 | layersByTop[layer.top(i)].push_back(&layer); |
| 1172 | } |
| 1173 | } |
| 1174 | |
| 1175 | // For each set of layers with the same top, resolve them to a linear chain rather than in-place layers. |
| 1176 | // Note that for 'regular' layers, there will be a single layer in each group and so this will be a no-op. |
| 1177 | for (auto layersWithSameTopIt : layersByTop) |
| 1178 | { |
| 1179 | const std::string& top = layersWithSameTopIt.first; |
| 1180 | const std::vector<caffe::LayerParameter*>& layersWithSameTop = layersWithSameTopIt.second; |
| 1181 | |
| 1182 | // Chain the layers together in the order that they are listed in the prototxt (hopefully this is correct). |
| 1183 | // Note that the last layer will not have its top modified so that other layers will continue to reference it. |
| 1184 | for (unsigned int layerIdx = 0; layerIdx < layersWithSameTop.size() - 1; ++layerIdx) |
| 1185 | { |
| 1186 | caffe::LayerParameter& layer1 = *layersWithSameTop[layerIdx]; |
| 1187 | caffe::LayerParameter& layer2 = *layersWithSameTop[layerIdx+1]; |
| 1188 | if (layer1.top_size() != 1) |
| 1189 | { |
| 1190 | throw ParseException("Node '" + layer1.name() + "' is an in-place layer but " |
| 1191 | "doesn't have exactly one top."); |
| 1192 | } |
| 1193 | std::string newTop = layer1.name() + "_top"; |
| 1194 | layer1.set_top(0, newTop); |
| 1195 | if (layer2.bottom_size() != 1 || layer2.bottom(0) != top) |
| 1196 | { |
| 1197 | throw ParseException("Node '" + layer2.name() + "' is an in-place layer but " |
| 1198 | " doesn't have exactly one bottom, or it doesn't match its top."); |
| 1199 | } |
| 1200 | layer2.set_bottom(0, newTop); |
| 1201 | } |
| 1202 | } |
| 1203 | } |
| 1204 | |
| 1205 | void CaffeParser::LoadNetParam(NetParameter& netParameter) |
| 1206 | { |
| 1207 | // caffe models sometimes have an implicit input layer. |
| 1208 | // in that case, add an explicit one |
| 1209 | if (netParameter.input_size() > 0) |
| 1210 | { |
| 1211 | LayerParameter* newLayer = netParameter.add_layer(); |
| 1212 | |
| 1213 | newLayer->set_type("Input"); |
| 1214 | newLayer->set_name(netParameter.input(0)); |
| 1215 | newLayer->add_top(netParameter.input(0)); |
| 1216 | |
| 1217 | InputParameter* inputParam = newLayer->mutable_input_param(); |
| 1218 | BlobShape* shape = inputParam->add_shape(); |
| 1219 | |
| 1220 | int dim_size = netParameter.input_dim_size(); |
| 1221 | for (int i = 0; i < dim_size; ++i) |
| 1222 | { |
| 1223 | shape->add_dim(netParameter.input_dim(i)); |
| 1224 | } |
| 1225 | } |
| 1226 | |
| 1227 | // Replace in-place layers with regular ones to make the rest of the parsing easier. |
| 1228 | ResolveInPlaceLayers(netParameter); |
| 1229 | |
| 1230 | // Create a lookup of Caffe layers by name |
| 1231 | for (int i = 0; i < netParameter.layer_size(); ++i) |
| 1232 | { |
| 1233 | const caffe::LayerParameter& layer = netParameter.layer(i); |
| 1234 | for (int i = 0; i < layer.top_size(); ++i) |
| 1235 | { |
| 1236 | m_CaffeLayersByTopName[layer.top(i)] = &layer; |
| 1237 | } |
| 1238 | } |
| 1239 | |
| 1240 | // Find the output layers the user requested |
| 1241 | std::vector<const caffe::LayerParameter*> targetLayers; |
| 1242 | for (const std::string& requestedOutputName : m_RequestedOutputs) |
| 1243 | { |
| 1244 | auto nodeIt = m_CaffeLayersByTopName.find(requestedOutputName); |
| 1245 | if (nodeIt == m_CaffeLayersByTopName.end()) |
| 1246 | { |
| 1247 | throw ParseException("Couldn't find requested output layer '" + requestedOutputName + "' in graph"); |
| 1248 | } |
| 1249 | targetLayers.push_back(nodeIt->second); |
| 1250 | } |
| 1251 | |
| 1252 | // Sort them into a linear ordering such that all inputs of a node are before the node itself |
| 1253 | std::vector<const caffe::LayerParameter*> sortedNodes; |
| 1254 | if (!armnnUtils::GraphTopologicalSort<const caffe::LayerParameter*>( |
| 1255 | targetLayers, |
| 1256 | [this](const caffe::LayerParameter* node) |
| 1257 | { |
| 1258 | return GetInputs(*node); |
| 1259 | }, |
| 1260 | sortedNodes)) |
| 1261 | { |
| 1262 | throw ParseException("Cycle detected in graph"); |
| 1263 | } |
| 1264 | |
| 1265 | // Parse each node in order, knowing that all inputs of a node will be processed before the node itself |
| 1266 | for (const caffe::LayerParameter* current : sortedNodes) |
| 1267 | { |
| 1268 | auto it = ms_CaffeLayerNameToParsingFunctions.find(current->type()); |
| 1269 | if (it == ms_CaffeLayerNameToParsingFunctions.end()) |
| 1270 | { |
| 1271 | throw ParseException("Unsupported layer type '" + current->type() + "'"); |
| 1272 | } |
| 1273 | auto func = it->second; |
| 1274 | (this->*func)(*current); |
| 1275 | } |
| 1276 | |
| 1277 | // Add ArmNN output layers connected to each requested output |
| 1278 | for (const std::string& requestedOutput : m_RequestedOutputs) |
| 1279 | { |
| 1280 | armnn::IOutputSlot& outputSlot = GetArmnnOutputSlotForCaffeTop(requestedOutput); |
| 1281 | |
| 1282 | const armnn::LayerBindingId outputId = boost::numeric_cast<armnn::LayerBindingId>( |
| 1283 | m_NetworkOutputsBindingInfo.size()); |
| 1284 | armnn::IConnectableLayer* const outputLayer = m_Network->AddOutputLayer(outputId, requestedOutput.c_str()); |
| 1285 | outputSlot.Connect(outputLayer->GetInputSlot(0)); |
| 1286 | |
| 1287 | TrackOutputBinding(outputLayer, outputId, outputLayer->GetInputSlot(0).GetConnection()->GetTensorInfo()); |
| 1288 | } |
| 1289 | } |
| 1290 | |
| 1291 | INetworkPtr CaffeParser::CreateNetworkFromTextFile(const char* graphFile, |
| 1292 | const std::map<std::string, armnn::TensorShape>& inputShapes, |
| 1293 | const std::vector<std::string>& requestedOutputs) |
| 1294 | { |
| 1295 | FILE* fd = fopen(graphFile, "r"); |
| 1296 | |
| 1297 | if (fd == nullptr) |
| 1298 | { |
| 1299 | std::stringstream error; |
| 1300 | error << "Graph file " << graphFile << " failed to open"; |
| 1301 | throw FileNotFoundException(error.str()); |
| 1302 | } |
| 1303 | |
| 1304 | // Parse the file into a message |
| 1305 | NetParameter netParam; |
| 1306 | auto input = new google::protobuf::io::FileInputStream(fileno(fd)); |
| 1307 | bool success = google::protobuf::TextFormat::Parse(input, &netParam); |
| 1308 | delete input; |
| 1309 | fclose(fd); |
| 1310 | |
| 1311 | if (!success) |
| 1312 | { |
| 1313 | std::stringstream error; |
| 1314 | error << "Failed to parse graph file"; |
| 1315 | throw ParseException(error.str()); |
| 1316 | } |
| 1317 | |
| 1318 | return CreateNetworkFromNetParameter(netParam, inputShapes, requestedOutputs); |
| 1319 | } |
| 1320 | |
| 1321 | INetworkPtr CaffeParser::CreateNetworkFromString(const char* protoText, |
| 1322 | const std::map<std::string, armnn::TensorShape>& inputShapes, |
| 1323 | const std::vector<std::string>& requestedOutputs) |
| 1324 | { |
| 1325 | // Parse the string into a message |
| 1326 | NetParameter netParam; |
| 1327 | bool success = google::protobuf::TextFormat::ParseFromString(protoText, &netParam); |
| 1328 | |
| 1329 | if (!success) |
| 1330 | { |
| 1331 | std::stringstream error; |
| 1332 | error << "Failed to parse graph string"; |
| 1333 | throw ParseException(error.str()); |
| 1334 | } |
| 1335 | |
| 1336 | return CreateNetworkFromNetParameter(netParam, inputShapes, requestedOutputs); |
| 1337 | } |
| 1338 | |
| 1339 | INetworkPtr CaffeParser::CreateNetworkFromBinaryFile(const char* graphFile, |
| 1340 | const std::map<std::string, armnn::TensorShape>& inputShapes, |
| 1341 | const std::vector<std::string>& requestedOutputs) |
| 1342 | { |
| 1343 | FILE* fd = fopen(graphFile, "rb"); |
| 1344 | |
| 1345 | if (fd == nullptr) |
| 1346 | { |
| 1347 | std::stringstream error; |
| 1348 | error << "Graph file " << graphFile << " failed to open"; |
| 1349 | throw FileNotFoundException(error.str()); |
| 1350 | } |
| 1351 | |
| 1352 | // Parse the file into a message |
| 1353 | NetParameter netParam; |
| 1354 | |
| 1355 | FileInputStream inStream(fileno(fd)); |
| 1356 | CodedInputStream codedStream(&inStream); |
| 1357 | codedStream.SetTotalBytesLimit(INT_MAX, INT_MAX); |
| 1358 | bool success = netParam.ParseFromCodedStream(&codedStream); |
| 1359 | fclose(fd); |
| 1360 | |
| 1361 | if (!success) |
| 1362 | { |
| 1363 | std::stringstream error; |
| 1364 | error << "Failed to parse protobuf file" << graphFile; |
| 1365 | throw ParseException(error.str()); |
| 1366 | } |
| 1367 | |
| 1368 | return CreateNetworkFromNetParameter(netParam, inputShapes, requestedOutputs); |
| 1369 | } |
| 1370 | |
| 1371 | INetworkPtr CaffeParser::CreateNetworkFromNetParameter(NetParameter& netParam, |
| 1372 | const std::map<std::string, armnn::TensorShape>& inputShapes, |
| 1373 | const std::vector<std::string>& requestedOutputs) |
| 1374 | { |
| 1375 | m_NetworkInputsBindingInfo.clear(); |
| 1376 | m_NetworkOutputsBindingInfo.clear(); |
| 1377 | |
| 1378 | m_Network = INetwork::Create(); |
| 1379 | |
| 1380 | m_InputShapes = inputShapes; |
| 1381 | if (requestedOutputs.size() == 0) |
| 1382 | { |
| 1383 | throw ParseException("requestedOutputs must have at least one entry"); |
| 1384 | } |
| 1385 | m_RequestedOutputs = requestedOutputs; |
| 1386 | |
| 1387 | try |
| 1388 | { |
| 1389 | LoadNetParam(netParam); |
| 1390 | } |
| 1391 | catch (const ParseException& e) |
| 1392 | { |
| 1393 | Cleanup(); |
| 1394 | throw e; |
| 1395 | } |
| 1396 | |
| 1397 | Cleanup(); |
| 1398 | |
| 1399 | return move(m_Network); |
| 1400 | } |
| 1401 | |
| 1402 | void CaffeParser::Cleanup() |
| 1403 | { |
| 1404 | // cleanup, in case we reuse this parser |
| 1405 | m_CaffeLayersByTopName.clear(); |
| 1406 | m_InputShapes.clear(); |
| 1407 | m_RequestedOutputs.clear(); |
| 1408 | m_ArmnnOutputSlotForCaffeTop.clear(); |
| 1409 | } |
| 1410 | |
| 1411 | } |
| 1412 | |
| 1413 | |