surmeh01 | bceff2f | 2018-03-29 16:29:27 +0100 | [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 "TfParser.hpp" |
| 6 | |
| 7 | #include <armnn/INetwork.hpp> |
| 8 | #include <armnn/Utils.hpp> |
| 9 | #include <armnn/TypesUtils.hpp> |
| 10 | #include <armnn/Exceptions.hpp> |
| 11 | #include <armnn/Descriptors.hpp> |
| 12 | |
| 13 | #include <GraphTopologicalSort.hpp> |
| 14 | #include <Permute.hpp> |
| 15 | |
| 16 | #include <google/protobuf/io/zero_copy_stream_impl.h> |
| 17 | #include <google/protobuf/text_format.h> |
| 18 | |
| 19 | #include "tensorflow/core/framework/graph.pb.h" |
| 20 | #include "tensorflow/core/framework/node_def.pb.h" |
| 21 | #include "tensorflow/core/framework/types.pb.h" |
| 22 | #include "tensorflow/core/framework/tensor.pb.h" |
| 23 | #include "tensorflow/core/framework/tensor_shape.pb.h" |
| 24 | |
| 25 | #include <boost/assert.hpp> |
| 26 | #include <boost/format.hpp> |
| 27 | #include <boost/core/ignore_unused.hpp> |
| 28 | #include <boost/log/trivial.hpp> |
| 29 | #include <boost/numeric/conversion/cast.hpp> |
| 30 | #include <boost/polymorphic_cast.hpp> |
| 31 | |
| 32 | #include <memory> |
| 33 | #include <sstream> |
| 34 | #include <numeric> |
| 35 | #include <functional> |
| 36 | |
| 37 | using namespace armnn; |
| 38 | |
| 39 | namespace armnnTfParser |
| 40 | { |
| 41 | namespace |
| 42 | { |
| 43 | |
| 44 | const PermutationVector NHWCToArmNN = { 0, 2, 3, 1 }; |
| 45 | const PermutationVector ArmNNToNHWC = { 0, 3, 1, 2 }; |
| 46 | |
| 47 | IConnectableLayer* AddSwizzleLayer(INetwork& network, IOutputSlot& input, const PermutationVector& mapping, |
| 48 | const std::string& name) |
| 49 | { |
| 50 | // Add swizzle layer |
| 51 | IConnectableLayer* const layer = network.AddPermuteLayer(mapping, name.c_str()); |
| 52 | |
| 53 | // Connect intput to swizzle layer |
| 54 | input.Connect(layer->GetInputSlot(0)); |
| 55 | |
| 56 | // Setup swizzled output |
| 57 | const TensorInfo outInfo = armnnUtils::Permuted(input.GetTensorInfo(), mapping); |
| 58 | layer->GetOutputSlot(0).SetTensorInfo(outInfo); |
| 59 | |
| 60 | return layer; |
| 61 | } |
| 62 | |
| 63 | IConnectableLayer* SwizzleInDeswizzleOut(INetwork& network, IOutputSlot& input, IConnectableLayer& layer, |
| 64 | const std::string& name) |
| 65 | { |
| 66 | // Add swizzle layer |
| 67 | IConnectableLayer* const swizzleLayer = AddSwizzleLayer(network, input, NHWCToArmNN, "swizzle_for-" + name); |
| 68 | |
| 69 | // Connect swizzledInput to layer |
| 70 | swizzleLayer->GetOutputSlot(0).Connect(layer.GetInputSlot(0)); |
| 71 | |
| 72 | // Add deswizzle layer |
| 73 | IConnectableLayer* const deswizzleLayer = AddSwizzleLayer(network, layer.GetOutputSlot(0), ArmNNToNHWC, |
| 74 | "deswizzle_for-" + name); |
| 75 | |
| 76 | return deswizzleLayer; |
| 77 | } |
| 78 | |
| 79 | template <typename Callable> |
| 80 | void ReadMandatoryNodeAttributeImpl(const tensorflow::NodeDef& nodeDef, |
| 81 | const std::string& attribName, |
| 82 | tensorflow::AttrValue::ValueCase expectedValueCase, |
| 83 | Callable callable) |
| 84 | { |
| 85 | auto iter = nodeDef.attr().find(attribName); |
| 86 | if (iter != nodeDef.attr().end()) |
| 87 | { |
| 88 | const auto& attrValue = iter->second; |
| 89 | if (attrValue.value_case() == expectedValueCase) |
| 90 | { |
| 91 | callable(attrValue); |
| 92 | } |
| 93 | else |
| 94 | { |
| 95 | throw ParseException(boost::str(boost::format( |
| 96 | "Attribute %1% of node %2% expected to have %3% as tensorflow::AttrValue::ValueCase, " |
| 97 | "but found %4% instead") |
| 98 | % attribName |
| 99 | % nodeDef.name() |
| 100 | % static_cast<int>(expectedValueCase) |
| 101 | % static_cast<int>(attrValue.value_case()))); |
| 102 | } |
| 103 | } |
| 104 | else |
| 105 | { |
| 106 | throw ParseException(boost::str(boost::format("Could not find required attribute %1% in node %2%") |
| 107 | % attribName % nodeDef.name())); |
| 108 | } |
| 109 | } |
| 110 | |
| 111 | template <typename Callable> |
| 112 | void ReadOptionalNodeAttributeImpl(const tensorflow::NodeDef& nodeDef, |
| 113 | const std::string& attribName, |
| 114 | tensorflow::AttrValue::ValueCase expectedValueCase, |
| 115 | Callable callable) |
| 116 | { |
| 117 | auto iter = nodeDef.attr().find(attribName); |
| 118 | if (iter != nodeDef.attr().end()) |
| 119 | { |
| 120 | const auto& attrValue = iter->second; |
| 121 | if (attrValue.value_case() == expectedValueCase) |
| 122 | { |
| 123 | callable(attrValue); |
| 124 | } |
| 125 | else |
| 126 | { |
| 127 | throw ParseException(boost::str(boost::format( |
| 128 | "Attribute %1% of node %2% expected to have %3% as tensorflow::AttrValue::ValueCase, " |
| 129 | "but found %4% instead") |
| 130 | % attribName |
| 131 | % nodeDef.name() |
| 132 | % static_cast<int>(expectedValueCase) |
| 133 | % static_cast<int>(attrValue.value_case()))); |
| 134 | } |
| 135 | } |
| 136 | } |
| 137 | |
| 138 | float ReadMandatoryNodeFloatAttribute(const tensorflow::NodeDef& nodeDef, const std::string& name) |
| 139 | { |
| 140 | float attribValue = 0.0f; |
| 141 | ReadMandatoryNodeAttributeImpl(nodeDef, name, tensorflow::AttrValue::kF, |
| 142 | [&attribValue](const tensorflow::AttrValue& attrValue) |
| 143 | { |
| 144 | attribValue = attrValue.f(); |
| 145 | }); |
| 146 | return attribValue; |
| 147 | } |
| 148 | |
| 149 | uint32_t ReadMandatoryNodeUint32Attribute(const tensorflow::NodeDef& nodeDef, const std::string& name) |
| 150 | { |
| 151 | uint32_t attribValue = 0u; |
| 152 | ReadMandatoryNodeAttributeImpl(nodeDef, name, tensorflow::AttrValue::kI, |
| 153 | [&attribValue](const tensorflow::AttrValue& attrValue) |
| 154 | { |
| 155 | attribValue = static_cast<uint32_t>(attrValue.i()); |
| 156 | }); |
| 157 | return attribValue; |
| 158 | } |
| 159 | |
| 160 | std::string ReadMandatoryNodeStringAttribute(const tensorflow::NodeDef& nodeDef, const std::string& name) |
| 161 | { |
| 162 | std::string attribValue = ""; |
| 163 | ReadMandatoryNodeAttributeImpl(nodeDef, name, tensorflow::AttrValue::kS, |
| 164 | [&attribValue](const tensorflow::AttrValue& attrValue) |
| 165 | { |
| 166 | attribValue = attrValue.s(); |
| 167 | }); |
| 168 | return attribValue; |
| 169 | } |
| 170 | |
| 171 | std::vector<uint32_t> ReadMandatoryNodeUint32ListAttribute(const tensorflow::NodeDef& nodeDef, |
| 172 | const std::string& name) |
| 173 | { |
| 174 | std::vector<uint32_t> attriList; |
| 175 | ReadMandatoryNodeAttributeImpl(nodeDef, name, tensorflow::AttrValue::kList, |
| 176 | [&attriList](const tensorflow::AttrValue& attrValue) |
| 177 | { |
| 178 | for (int attriNum = 0; attriNum < attrValue.list().i_size(); ++attriNum) |
| 179 | { |
| 180 | attriList.push_back(static_cast<uint32_t>(attrValue.list().i(attriNum))); |
| 181 | } |
| 182 | }); |
| 183 | |
| 184 | return attriList; |
| 185 | } |
| 186 | |
| 187 | std::vector<uint32_t> ReadOptionalNodeUint32ListAttribute(const tensorflow::NodeDef& nodeDef, |
| 188 | const std::string& name) |
| 189 | { |
| 190 | std::vector<uint32_t> attriList; |
| 191 | ReadOptionalNodeAttributeImpl(nodeDef, name, tensorflow::AttrValue::kList, |
| 192 | [&attriList](const tensorflow::AttrValue& attrValue) |
| 193 | { |
| 194 | for (int attriNum = 0; attriNum < attrValue.list().i_size(); ++attriNum) |
| 195 | { |
| 196 | attriList.push_back(static_cast<uint32_t>(attrValue.list().i(attriNum))); |
| 197 | } |
| 198 | }); |
| 199 | |
| 200 | return attriList; |
| 201 | } |
| 202 | |
| 203 | bool ReadOptionalNodeBoolAttribute(const tensorflow::NodeDef& nodeDef, |
| 204 | const std::string& name, |
| 205 | bool defaultValue = false) |
| 206 | { |
| 207 | bool attribValue = defaultValue; |
| 208 | ReadOptionalNodeAttributeImpl(nodeDef, name, tensorflow::AttrValue::kB, |
| 209 | [&attribValue](const tensorflow::AttrValue& attrValue) |
| 210 | { |
| 211 | attribValue = attrValue.b(); |
| 212 | }); |
| 213 | return attribValue; |
| 214 | } |
| 215 | |
| 216 | tensorflow::DataType ReadMandatoryNodeTypeAttribute(const tensorflow::NodeDef& nodeDef, const std::string& name) |
| 217 | { |
| 218 | tensorflow::DataType attribValue = tensorflow::DT_INVALID; |
| 219 | ReadMandatoryNodeAttributeImpl(nodeDef, name, tensorflow::AttrValue::kType, |
| 220 | [&attribValue](const tensorflow::AttrValue& attrValue) |
| 221 | { |
| 222 | attribValue = attrValue.type(); |
| 223 | }); |
| 224 | return attribValue; |
| 225 | } |
| 226 | |
| 227 | TensorInfo PrepareReshape(const TensorInfo& input, const std::vector<int32_t>& targetDims) |
| 228 | { |
| 229 | std::vector<unsigned int> outDims(targetDims.begin(), targetDims.end()); |
| 230 | const auto stretchDim = std::find(targetDims.begin(), targetDims.end(), -1); |
| 231 | |
| 232 | if (stretchDim != targetDims.end()) |
| 233 | { |
| 234 | if (std::find(std::next(stretchDim), targetDims.end(), -1) != targetDims.end()) |
| 235 | { |
| 236 | throw ParseException("At most one component of shape can be -1"); |
| 237 | } |
| 238 | |
| 239 | auto targetNumElements = boost::numeric_cast<unsigned int>(std::accumulate(targetDims.begin(), targetDims.end(), |
| 240 | -1, std::multiplies<int32_t>())); |
| 241 | auto stretchIndex = static_cast<size_t>(std::distance(targetDims.begin(), stretchDim)); |
| 242 | outDims[stretchIndex] = input.GetNumElements() / targetNumElements; |
| 243 | } |
| 244 | |
| 245 | TensorInfo reshapeInfo = input; |
| 246 | reshapeInfo.SetShape(TensorShape{ static_cast<unsigned int>(outDims.size()), outDims.data() }); |
| 247 | |
| 248 | return reshapeInfo; |
| 249 | } |
| 250 | |
| 251 | // We need the input0Slot to guide the reshape for input1Slot |
| 252 | IOutputSlot* BroadcastForAddandMul(IOutputSlot* input0Slot, IOutputSlot* input1Slot, bool isNHWC, INetwork& m_Network, |
| 253 | const tensorflow::NodeDef& nodeDef) |
| 254 | { |
| 255 | const TensorInfo& input1Info = input1Slot->GetTensorInfo(); |
| 256 | const TensorInfo inputTensorInfo = input0Slot->GetTensorInfo(); |
| 257 | const unsigned int matchDim = inputTensorInfo.GetNumDimensions() - (isNHWC ? 1 : 3); |
| 258 | std::array<unsigned int, MaxNumOfTensorDimensions> reshapedDimensions; |
| 259 | std::fill_n(reshapedDimensions.begin(), inputTensorInfo.GetNumDimensions(), 1); |
| 260 | reshapedDimensions[matchDim] = input1Info.GetShape()[0]; |
| 261 | |
| 262 | armnn::TensorInfo reshapedInfo = input1Info; |
| 263 | reshapedInfo.SetShape(TensorShape{ inputTensorInfo.GetNumDimensions(), reshapedDimensions.data() }); |
| 264 | |
| 265 | const std::string reshapeLayerName = "reshape_for-" + nodeDef.name(); |
| 266 | ReshapeDescriptor reshapeDesc; |
| 267 | reshapeDesc.m_TargetShape = reshapedInfo.GetShape(); |
| 268 | IConnectableLayer* const reshapeLayer = m_Network.AddReshapeLayer(reshapeDesc, reshapeLayerName.c_str()); |
| 269 | |
| 270 | input1Slot->Connect(reshapeLayer->GetInputSlot(0)); |
| 271 | reshapeLayer->GetOutputSlot(0).SetTensorInfo(reshapedInfo); |
| 272 | |
| 273 | input1Slot = &reshapeLayer->GetOutputSlot(0); |
| 274 | |
| 275 | return input1Slot; |
| 276 | } |
| 277 | |
| 278 | OutputId ParseOutputId(const std::string & name) |
| 279 | { |
| 280 | unsigned int outputNum = 0; |
| 281 | size_t colonPos = name.find_last_of(":"); |
| 282 | if (colonPos != std::string::npos) |
| 283 | { |
| 284 | int n = std::stoi(name.substr(colonPos+1)); |
| 285 | if (n<0 || n>100) |
| 286 | { |
| 287 | throw ParseException("Output tensor id is out of range for "+name); |
| 288 | } |
| 289 | outputNum = static_cast<unsigned int>(n); |
| 290 | } |
| 291 | return OutputId(name.substr(0,colonPos),outputNum); |
| 292 | } |
| 293 | |
| 294 | } // namespace |
| 295 | |
| 296 | const std::map<std::string, TfParser::OperationParsingFunction> TfParser::ms_OperationNameToParsingFunctions = { |
| 297 | { "Const", &TfParser::ParseConst }, |
| 298 | { "Add", &TfParser::ParseAdd }, |
| 299 | { "BiasAdd", &TfParser::ParseBiasAdd }, |
| 300 | { "Identity", &TfParser::ParseIdentity }, |
| 301 | { "Conv2D", &TfParser::ParseConv2D }, |
| 302 | { "DepthwiseConv2dNative", &TfParser::ParseDepthwiseConv2D }, |
| 303 | { "FusedBatchNorm", &TfParser::ParseFusedBatchNorm }, |
| 304 | { "ConcatV2", &TfParser::ParseConcat }, |
| 305 | { "LRN", &TfParser::ParseLrn }, |
| 306 | { "MatMul", &TfParser::ParseMatMul }, |
| 307 | { "Mul", &TfParser::ParseMul }, |
| 308 | { "Placeholder", &TfParser::ParsePlaceholder }, |
| 309 | { "Relu", &TfParser::ParseRelu }, |
| 310 | { "Relu6", &TfParser::ParseRelu6 }, |
| 311 | { "Reshape", &TfParser::ParseReshape }, |
| 312 | { "ResizeBilinear", &TfParser::ParseResizeBilinear }, |
| 313 | { "Shape", &TfParser::ParseShape }, |
| 314 | { "Squeeze", &TfParser::ParseSqueeze }, |
| 315 | { "Sigmoid", &TfParser::ParseSigmoid }, |
| 316 | { "Softmax", &TfParser::ParseSoftmax }, |
| 317 | { "Softplus", &TfParser::ParseSoftplus }, |
| 318 | { "Tanh", &TfParser::ParseTanh }, |
| 319 | { "MaxPool", &TfParser::ParseMaxPool }, |
| 320 | { "AvgPool", &TfParser::ParseAvgPool }, |
| 321 | }; |
| 322 | |
| 323 | ITfParser* ITfParser::CreateRaw() |
| 324 | { |
| 325 | return new TfParser(); |
| 326 | } |
| 327 | |
| 328 | ITfParserPtr ITfParser::Create() |
| 329 | { |
| 330 | return ITfParserPtr(CreateRaw(), &ITfParser::Destroy); |
| 331 | } |
| 332 | |
| 333 | void ITfParser::Destroy(ITfParser* parser) |
| 334 | { |
| 335 | delete parser; |
| 336 | } |
| 337 | |
| 338 | inline void CalculateSamePadding(uint32_t inputSize, uint32_t stride, |
| 339 | uint32_t filterSize, bool samePadding, |
| 340 | uint32_t* paddingFront, uint32_t* paddingBack) { |
| 341 | *paddingFront = 0; |
| 342 | *paddingBack = 0; |
| 343 | |
| 344 | if (samePadding) { |
| 345 | uint32_t outputSize = (inputSize + stride - 1) / stride; |
| 346 | uint32_t temp = (outputSize - 1) * stride + filterSize; |
| 347 | if (temp > inputSize) { |
| 348 | *paddingFront = (temp - inputSize) / 2; |
| 349 | *paddingBack = (temp - inputSize) - *paddingFront; |
| 350 | } |
| 351 | } |
| 352 | } |
| 353 | |
| 354 | void CalcPadding(uint32_t input, uint32_t kernel, uint32_t stride, uint32_t& outPadHead, uint32_t& outPadTail, |
| 355 | bool samePadding) |
| 356 | { |
| 357 | CalculateSamePadding(input, stride, kernel, samePadding, &outPadHead, &outPadTail); |
| 358 | } |
| 359 | |
| 360 | /// An Abstract base class which represents a single tensorflow operation (node) |
| 361 | /// that has been (potentially partially) converted to Armnn. |
| 362 | /// It may not yet have been fully converted into actual Armnn layers. |
| 363 | class ParsedTfOperation |
| 364 | { |
| 365 | public: |
| 366 | ParsedTfOperation(TfParser* parser, const tensorflow::NodeDef& node) |
| 367 | : m_Parser(parser) |
| 368 | , m_Node(node) |
| 369 | { |
| 370 | } |
| 371 | |
| 372 | virtual ~ParsedTfOperation() {}; |
| 373 | |
| 374 | const tensorflow::NodeDef& GetNode() const { return m_Node; } |
| 375 | |
| 376 | /// Gets the ArmNN IOutputSlot corresponding to the given output index of the Tensorflow operation. |
| 377 | /// This may result in the creation of Armnn layers if this was deferred (e.g. see ParsedConstTfOperation). |
| 378 | virtual IOutputSlot& ResolveArmnnOutputSlot(unsigned int tfOutputIndex) = 0; |
| 379 | |
| 380 | /// If this operation is an Identity then this will follow return the 'parent' operation (recursively). |
| 381 | virtual ParsedTfOperation* ResolveIdentityOperations() |
| 382 | { |
| 383 | return this; |
| 384 | } |
| 385 | |
| 386 | protected: |
| 387 | TfParser* m_Parser; |
| 388 | const tensorflow::NodeDef& m_Node; |
| 389 | }; |
| 390 | |
| 391 | /// An ParsedTfOperation where the Armnn equivalent is a single layer, |
| 392 | /// with output slots that correspond directly to the Tf node outputs. |
| 393 | class SingleLayerParsedTfOperation : public ParsedTfOperation |
| 394 | { |
| 395 | public: |
| 396 | SingleLayerParsedTfOperation(TfParser* parser, const tensorflow::NodeDef& node, IConnectableLayer* layer) |
| 397 | : ParsedTfOperation(parser, node) |
| 398 | , m_Layer(layer) |
| 399 | { |
| 400 | } |
| 401 | |
| 402 | IOutputSlot& ResolveArmnnOutputSlot(unsigned int tfOutputIndex) override |
| 403 | { |
| 404 | BOOST_ASSERT(m_Layer); |
| 405 | // Assume one-to-one mapping between Tf and armnn output slots. |
| 406 | unsigned int armnnOutputSlotIdx = tfOutputIndex; |
| 407 | if (armnnOutputSlotIdx >= m_Layer->GetNumOutputSlots()) |
| 408 | { |
| 409 | throw ParseException( |
| 410 | boost::str(boost::format("The requested output slot #%1% " |
| 411 | "for %2% does not exist") % armnnOutputSlotIdx % m_Layer->GetName())); |
| 412 | } |
| 413 | return m_Layer->GetOutputSlot(armnnOutputSlotIdx); |
| 414 | } |
| 415 | |
| 416 | protected: |
| 417 | IConnectableLayer* m_Layer; |
| 418 | }; |
| 419 | |
| 420 | /// A SingleLayerParsedTfOperation for deferred layer creation |
| 421 | class DeferredSingleLayerParsedTfOperation : public SingleLayerParsedTfOperation |
| 422 | { |
| 423 | public: |
| 424 | DeferredSingleLayerParsedTfOperation(TfParser* parser, const tensorflow::NodeDef& node) |
| 425 | : SingleLayerParsedTfOperation(parser, node, nullptr) |
| 426 | { |
| 427 | } |
| 428 | |
| 429 | IOutputSlot& ResolveArmnnOutputSlot(unsigned int tfOutputIndex) override |
| 430 | { |
| 431 | if (!m_Layer) |
| 432 | { |
| 433 | CreateLayerDeferred(); |
| 434 | } |
| 435 | return SingleLayerParsedTfOperation::ResolveArmnnOutputSlot(tfOutputIndex); |
| 436 | } |
| 437 | |
| 438 | private: |
| 439 | virtual void CreateLayerDeferred() = 0; |
| 440 | }; |
| 441 | |
| 442 | |
| 443 | TfParser::TfParser() |
| 444 | : m_Network(nullptr, nullptr) |
| 445 | { |
| 446 | } |
| 447 | |
| 448 | |
| 449 | const tensorflow::NodeDef* TfParser::ResolveIdentityNode(const tensorflow::NodeDef* nodeDef) |
| 450 | { |
| 451 | if (nodeDef->op() != "Identity") |
| 452 | { |
| 453 | return nodeDef; |
| 454 | } |
| 455 | |
| 456 | if (nodeDef->input_size() != 1) |
| 457 | { |
| 458 | throw ParseException("Identity node does not have correct amount of inputs!"); |
| 459 | } |
| 460 | |
| 461 | auto it = m_NodesByName.find(nodeDef->input(0)); |
| 462 | if (it != m_NodesByName.end()) |
| 463 | { |
| 464 | const tensorflow::NodeDef* inputNode = it->second; |
| 465 | return ResolveIdentityNode(inputNode); |
| 466 | } |
| 467 | else |
| 468 | { |
| 469 | throw ParseException("Cannot find what the Identity node is linked to!"); |
| 470 | } |
| 471 | } |
| 472 | |
| 473 | std::vector<OutputOfConstNodeDef> |
| 474 | TfParser::GetTfInputNodes(const tensorflow::NodeDef& nodeDef) const |
| 475 | { |
| 476 | std::vector<OutputOfConstNodeDef> ret; |
| 477 | |
| 478 | ret.reserve(boost::numeric_cast<size_t>(nodeDef.input_size())); |
| 479 | for (int j = 0; j < nodeDef.input_size(); ++j) |
| 480 | { |
| 481 | OutputId outputId = ParseOutputId(nodeDef.input(j)); |
| 482 | auto inputIt = m_NodesByName.find(outputId.m_IndexedValue); |
| 483 | if (inputIt == m_NodesByName.end()) |
| 484 | { |
| 485 | throw ParseException( |
| 486 | "Can't find node '" + nodeDef.input(j) + |
| 487 | "', which is listed as an input of '" + nodeDef.name() + "'"); |
| 488 | } |
| 489 | ret.push_back(OutputOfConstNodeDef(inputIt->second,outputId.m_Index)); |
| 490 | } |
| 491 | |
| 492 | return ret; |
| 493 | } |
| 494 | |
| 495 | std::vector<OutputOfParsedTfOperation> |
| 496 | TfParser::GetInputParsedTfOperationsChecked(const tensorflow::NodeDef& nodeDef, |
| 497 | std::size_t expectedNumInputs) |
| 498 | { |
| 499 | // Fetch the tensorflow nodes connected as inputs and validate the size. |
| 500 | std::vector<OutputOfConstNodeDef> nodes = GetTfInputNodes(nodeDef); |
| 501 | const std::size_t numInputs = nodes.size(); |
| 502 | if (numInputs != expectedNumInputs) |
| 503 | { |
| 504 | throw ParseException(boost::str(boost::format("Unexpected number of inputs for node %1%. " |
| 505 | "Expected %2%, found %3%") % nodeDef.name() % expectedNumInputs % numInputs)); |
| 506 | } |
| 507 | // Fetch the corresponding ParsedTfOperation operations |
| 508 | std::vector<OutputOfParsedTfOperation> result; |
| 509 | for (auto&& node : nodes) |
| 510 | { |
| 511 | auto it = m_ParsedTfOperations.find(node.m_IndexedValue->name()); |
| 512 | if (it == m_ParsedTfOperations.end()) |
| 513 | { |
| 514 | throw ParseException("Node with name '" + node.m_IndexedValue->name() + "' has not been parsed"); |
| 515 | } |
| 516 | ParsedTfOperation* parsedOp = it->second.get(); |
| 517 | // Transparently 'skip' any Identity operations. This simplifies the logic inside the ParseXXX() functions. |
| 518 | parsedOp = parsedOp->ResolveIdentityOperations(); |
| 519 | result.push_back(OutputOfParsedTfOperation(parsedOp,node.m_Index)); |
| 520 | } |
| 521 | return result; |
| 522 | } |
| 523 | |
| 524 | ParsedTfOperationPtr TfParser::ParseAdd(const tensorflow::NodeDef& nodeDef, const tensorflow::GraphDef& graphDef) |
| 525 | { |
| 526 | std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 2); |
| 527 | |
| 528 | // If one of the inputs is a MatMul and the other is a const, then we handle both nodes together as FullyConnected |
| 529 | if (inputs[0].m_IndexedValue->GetNode().op() == "MatMul" && |
| 530 | HasParsedConstTensor<float>(inputs[1].m_IndexedValue->GetNode().name())) |
| 531 | { |
| 532 | IConnectableLayer* layer = |
| 533 | AddFullyConnectedLayer(inputs[0].m_IndexedValue->GetNode(), |
| 534 | &nodeDef,nodeDef.name().c_str()); |
| 535 | return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer); |
| 536 | } |
| 537 | else if (HasParsedConstTensor<float>(inputs[0].m_IndexedValue->GetNode().name()) && |
| 538 | inputs[1].m_IndexedValue->GetNode().op() == "MatMul") |
| 539 | { |
| 540 | IConnectableLayer* layer = |
| 541 | AddFullyConnectedLayer(inputs[1].m_IndexedValue->GetNode(), |
| 542 | &nodeDef,nodeDef.name().c_str()); |
| 543 | return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer); |
| 544 | } |
| 545 | else |
| 546 | { |
| 547 | // Otherwise it's just a regular addition |
| 548 | return AddAdditionLayer(nodeDef); |
| 549 | } |
| 550 | } |
| 551 | |
| 552 | ParsedTfOperationPtr TfParser::ParseBiasAdd(const tensorflow::NodeDef& nodeDef, const tensorflow::GraphDef& graphDef) |
| 553 | { |
| 554 | return AddAdditionLayer(nodeDef, true); |
| 555 | } |
| 556 | |
| 557 | /// An ParsedTfOperation which forwards to another (used for Identity nodes). |
| 558 | class ParsedIdentityTfOperation : public ParsedTfOperation |
| 559 | { |
| 560 | public: |
| 561 | ParsedIdentityTfOperation(TfParser* parser, const tensorflow::NodeDef& node, ParsedTfOperation* representative) |
| 562 | : ParsedTfOperation(parser, node) |
| 563 | , m_Representative(representative) |
| 564 | { |
| 565 | } |
| 566 | |
| 567 | virtual IOutputSlot& ResolveArmnnOutputSlot(unsigned int tfOutputIndex) override |
| 568 | { |
| 569 | BOOST_ASSERT(m_Representative); |
| 570 | return m_Representative->ResolveArmnnOutputSlot(tfOutputIndex); |
| 571 | } |
| 572 | |
| 573 | virtual ParsedTfOperation* ResolveIdentityOperations() override |
| 574 | { |
| 575 | return m_Representative->ResolveIdentityOperations(); |
| 576 | } |
| 577 | |
| 578 | private: |
| 579 | ParsedTfOperation* m_Representative; |
| 580 | }; |
| 581 | |
| 582 | ParsedTfOperationPtr TfParser::ParseIdentity(const tensorflow::NodeDef& nodeDef, const tensorflow::GraphDef& graphDef) |
| 583 | { |
| 584 | std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 1); |
| 585 | // Any requests for the output slots of this node should be forwarded to the node connected as input. |
| 586 | return std::make_unique<ParsedIdentityTfOperation>(this, nodeDef, inputs[0].m_IndexedValue); |
| 587 | } |
| 588 | |
| 589 | /// An ParsedTfOperation for a Const node. |
| 590 | /// Creation of the armnn ConstLayer is deferred until it is actually needed, because Const nodes are mostly used |
| 591 | /// for weight inputs to MatMul/Conv2D nodes and in these cases armnn doesn't need a ConstLayer. |
| 592 | template <typename T> |
| 593 | class ParsedConstTfOperation : public DeferredSingleLayerParsedTfOperation |
| 594 | { |
| 595 | public: |
| 596 | ParsedConstTfOperation(TfParser* parser, const tensorflow::NodeDef& node, |
| 597 | const T* tensorData, const TensorInfo& tensorInfo) |
| 598 | : DeferredSingleLayerParsedTfOperation(parser, node), |
| 599 | m_Storage(tensorData, tensorData + tensorInfo.GetNumElements()), |
| 600 | m_TensorInfo(tensorInfo) |
| 601 | { |
| 602 | BOOST_ASSERT(tensorInfo.GetDataType() == GetDataType<T>()); |
| 603 | } |
| 604 | |
| 605 | void CreateLayerDeferred() override |
| 606 | { |
| 607 | BOOST_ASSERT(m_Layer == nullptr); |
| 608 | m_Layer = m_Parser->m_Network->AddConstantLayer(ConstTensor(m_TensorInfo, m_Storage), m_Node.name().c_str()); |
| 609 | m_Layer->GetOutputSlot(0).SetTensorInfo(m_TensorInfo); |
| 610 | } |
| 611 | |
| 612 | ConstTensor GetConstTensor(bool swizzleForConvolutionWeights, std::vector<T>& outputTensorData) const |
| 613 | { |
| 614 | // Mappings from TensorFlow filter tensors to the ArmNN filter tensors. |
| 615 | // Tensorflow weights are [H, W, In, Out] |
| 616 | // ArmNN weights are [Out, In, H, W] |
| 617 | static const PermutationVector HWIOToOIHW = {2, 3, 1, 0}; |
| 618 | |
| 619 | const TensorInfo outInfo = swizzleForConvolutionWeights |
| 620 | ? armnnUtils::Permuted(m_TensorInfo, HWIOToOIHW) |
| 621 | : m_TensorInfo; |
| 622 | |
| 623 | outputTensorData.resize(m_TensorInfo.GetNumElements()); |
| 624 | |
| 625 | // Copy or swizzle from the permanent storage into the storage the caller provided. |
| 626 | if (swizzleForConvolutionWeights) |
| 627 | { |
| 628 | armnnUtils::Permute(outInfo.GetShape(), HWIOToOIHW, m_Storage.data(), outputTensorData.data()); |
| 629 | } |
| 630 | else |
| 631 | { |
| 632 | memcpy(outputTensorData.data(), m_Storage.data(), m_TensorInfo.GetNumBytes()); |
| 633 | } |
| 634 | // Update the result to point to the user provided storage |
| 635 | ConstTensor constTensor(outInfo, outputTensorData); |
| 636 | return constTensor; |
| 637 | } |
| 638 | |
| 639 | private: |
| 640 | ///< Manages the lifetime of the tensor data. |
| 641 | std::vector<T> m_Storage; |
| 642 | ///< Describes the layout of the tensor and points to the data in m_Storage. |
| 643 | TensorInfo m_TensorInfo; |
| 644 | }; |
| 645 | |
| 646 | DataType ConvertTfTensorDataType(const tensorflow::DataType tfDataType) |
| 647 | { |
| 648 | switch (tfDataType) |
| 649 | { |
| 650 | case tensorflow::DT_FLOAT: |
| 651 | return DataType::Float32; |
| 652 | break; |
| 653 | case tensorflow::DT_INT32: |
| 654 | return DataType::Signed32; |
| 655 | break; |
| 656 | default: |
| 657 | throw ParseException(boost::str( |
| 658 | boost::format("Unknown DataType %1% for node") |
| 659 | % tensorflow::DataType_Name(tfDataType))); |
| 660 | } |
| 661 | } |
| 662 | |
| 663 | struct ParseTfTensorValueList |
| 664 | { |
| 665 | template<typename DataType> |
| 666 | static void Parse( |
| 667 | const tensorflow::TensorProto& tfTensor, |
| 668 | unsigned int dstElements, |
| 669 | std::vector<int8_t>& outputData); |
| 670 | |
| 671 | template <typename DataType> |
| 672 | static void ReadData(const void* srcData, unsigned int numSrcElements, |
| 673 | std::vector<int8_t>& dstData, unsigned int numDstElements) |
| 674 | { |
| 675 | // If there are no entries in the list, perform no action |
| 676 | if (numSrcElements == 0) |
| 677 | { |
| 678 | return; |
| 679 | } |
| 680 | |
| 681 | // If no size was provided, use the length of the value list |
| 682 | if (numDstElements == 0) |
| 683 | { |
| 684 | numDstElements = numSrcElements; |
| 685 | } |
| 686 | |
| 687 | // Allocate memory |
| 688 | dstData.resize(std::max(numSrcElements, numDstElements) * sizeof(DataType)); |
| 689 | |
| 690 | const DataType* srcTensor = reinterpret_cast<const DataType*>(srcData); |
| 691 | DataType* dstTensor = reinterpret_cast<DataType*>(dstData.data()); |
| 692 | |
| 693 | // Copy the value list entries into the destination |
| 694 | std::copy(srcTensor, srcTensor + numSrcElements, dstTensor); |
| 695 | |
| 696 | if (numDstElements > numSrcElements) |
| 697 | { |
| 698 | // Use the last element in the list to fill the remaining entries |
| 699 | std::fill(dstTensor + numSrcElements, dstTensor + numDstElements, srcTensor[numSrcElements - 1]); |
| 700 | } |
| 701 | } |
| 702 | |
| 703 | }; |
| 704 | |
| 705 | template <> |
| 706 | void ParseTfTensorValueList::Parse<float>(const tensorflow::TensorProto& tfTensor, |
| 707 | unsigned int dstElements, std::vector<int8_t>& outputData) |
| 708 | { |
| 709 | ReadData<float>(tfTensor.float_val().data(), static_cast<unsigned int>(tfTensor.float_val_size()), |
| 710 | outputData, dstElements); |
| 711 | } |
| 712 | |
| 713 | template <> |
| 714 | void ParseTfTensorValueList::Parse<int32_t>(const tensorflow::TensorProto& tfTensor, |
| 715 | unsigned int dstElements, std::vector<int8_t>& outputData) |
| 716 | { |
| 717 | ReadData<int32_t>(tfTensor.int_val().data(), static_cast<unsigned int>(tfTensor.int_val_size()), |
| 718 | outputData, dstElements); |
| 719 | } |
| 720 | |
| 721 | template <template<typename> class OperatorType, typename T = int8_t> |
| 722 | struct MakeTfOperation |
| 723 | { |
| 724 | template<typename DataType, class... Args> |
| 725 | inline static std::unique_ptr<OperatorType<DataType>> Parse(TfParser* parser, const tensorflow::NodeDef& node, |
| 726 | Args&&... args) |
| 727 | { |
| 728 | return std::make_unique<OperatorType<DataType>>(parser, node, std::forward<Args>(args)...); |
| 729 | } |
| 730 | }; |
| 731 | |
| 732 | template <> |
| 733 | struct MakeTfOperation<ParsedConstTfOperation> |
| 734 | { |
| 735 | template<typename DataType, class... Args> |
| 736 | inline static std::unique_ptr<ParsedConstTfOperation<DataType>> Parse(TfParser* parser, |
| 737 | const tensorflow::NodeDef& node, const std::vector<int8_t>& tensorData, const TensorInfo& tensorInfo) |
| 738 | { |
| 739 | return std::make_unique<ParsedConstTfOperation<DataType>>(parser, node, |
| 740 | reinterpret_cast<const DataType*>(tensorData.data()), tensorInfo); |
| 741 | } |
| 742 | }; |
| 743 | |
| 744 | template <class FuncType> |
| 745 | struct InvokeParseFunction |
| 746 | { |
| 747 | template<class ResType, class... Args> |
| 748 | inline static ResType Result(DataType dataType, Args&&... args) |
| 749 | { |
| 750 | if (dataType == DataType::Float32) |
| 751 | { |
| 752 | return FuncType::template Parse<float>(std::forward<Args>(args)...); |
| 753 | } |
| 754 | else if (dataType == DataType::Signed32) |
| 755 | { |
| 756 | return FuncType::template Parse<int32_t>(std::forward<Args>(args)...); |
| 757 | } |
| 758 | |
| 759 | return ResType(); |
| 760 | } |
| 761 | |
| 762 | template<class... Args> |
| 763 | inline static void Result(DataType dataType, Args&&... args) |
| 764 | { |
| 765 | if (dataType == DataType::Float32) |
| 766 | { |
| 767 | FuncType::template Parse<float>(std::forward<Args>(args)...); |
| 768 | } |
| 769 | else if (dataType == DataType::Signed32) |
| 770 | { |
| 771 | FuncType::template Parse<int32_t>(std::forward<Args>(args)...); |
| 772 | } |
| 773 | } |
| 774 | }; |
| 775 | |
| 776 | ParsedTfOperationPtr TfParser::ParseConst(const tensorflow::NodeDef& nodeDef, const tensorflow::GraphDef& graphDef) |
| 777 | { |
| 778 | BOOST_ASSERT(nodeDef.op() == "Const"); |
| 779 | |
| 780 | if (nodeDef.attr().count("value") == 0) |
| 781 | { |
| 782 | throw ParseException(boost::str( |
| 783 | boost::format("Value not found for Const node - %1%") |
| 784 | % nodeDef.name())); |
| 785 | } |
| 786 | |
| 787 | const tensorflow::TensorProto& tfTensor = nodeDef.attr().at("value").tensor(); |
| 788 | const tensorflow::TensorShapeProto& tfTensorShape = tfTensor.tensor_shape(); |
| 789 | const tensorflow::DataType tfDataType = ReadMandatoryNodeTypeAttribute(nodeDef, "dtype"); |
| 790 | |
| 791 | const auto GetDimensionSize = [](auto& d) { return d.size(); }; |
| 792 | |
| 793 | std::vector<unsigned int> dimensionSizes; |
| 794 | std::transform(tfTensorShape.dim().begin(), tfTensorShape.dim().end(), |
| 795 | std::back_inserter(dimensionSizes), GetDimensionSize); |
| 796 | |
| 797 | // Calculate number of elements |
| 798 | const DataType dataType = ConvertTfTensorDataType(tfDataType); |
| 799 | unsigned int numElements = 0U; |
| 800 | |
| 801 | if (!dimensionSizes.empty()) |
| 802 | { |
| 803 | numElements = std::accumulate(dimensionSizes.begin(), dimensionSizes.end(), |
| 804 | 1U, std::multiplies<unsigned int>()); |
| 805 | } |
| 806 | |
| 807 | std::vector<int8_t> tensorData; |
| 808 | |
| 809 | // Get tensor data from the list of values attribute |
| 810 | if (tfTensor.tensor_content().empty()) |
| 811 | { |
| 812 | InvokeParseFunction<ParseTfTensorValueList>::Result<void>(dataType, tfTensor, numElements, tensorData); |
| 813 | |
| 814 | // If the tensor shape is not defined, but there is a value list, then interpret the data as a 1D |
| 815 | // tensor of the provided number of elements |
| 816 | if (numElements == 0) |
| 817 | { |
| 818 | const unsigned int tfNumElements = static_cast<unsigned int>(tensorData.size()) / GetDataTypeSize(dataType); |
| 819 | dimensionSizes.push_back(tfNumElements); |
| 820 | } |
| 821 | } |
| 822 | // Get tensor data from tensor content attribute |
| 823 | else |
| 824 | { |
| 825 | tensorData.assign(tfTensor.tensor_content().begin(), tfTensor.tensor_content().end()); |
| 826 | |
| 827 | // Check if a tensor shape is defined for the tensor content |
| 828 | if (numElements == 0) |
| 829 | { |
| 830 | throw ParseException(boost::str( |
| 831 | boost::format("No tensor shape found for Const node - %1%") |
| 832 | % nodeDef.name())); |
| 833 | } |
| 834 | } |
| 835 | |
| 836 | // Const node requires at least a list of values or a content attribute |
| 837 | if (tensorData.empty()) |
| 838 | { |
| 839 | throw ParseException(boost::str( |
| 840 | boost::format("No tensor data found for Const node - %1%") |
| 841 | % nodeDef.name())); |
| 842 | } |
| 843 | |
| 844 | const TensorInfo tensorInfo(static_cast<unsigned int>(dimensionSizes.size()), dimensionSizes.data(), dataType); |
| 845 | |
| 846 | // If we have a list of values, then the length of the list must be |
| 847 | // less than or equal to the number of elements implied by the shape argument |
| 848 | if (tensorData.size() > tensorInfo.GetNumBytes()) |
| 849 | { |
| 850 | throw ParseException(boost::str( |
| 851 | boost::format("Number of elements (%1%) should be less than or equal \ |
| 852 | to the number of elements implied by the shape argument (%2%) for Const node - %3%") |
| 853 | % (tensorData.size() / GetDataTypeSize(dataType)) |
| 854 | % tensorInfo.GetNumElements() |
| 855 | % nodeDef.name())); |
| 856 | } |
| 857 | |
| 858 | return InvokeParseFunction<MakeTfOperation<ParsedConstTfOperation>>::Result<ParsedTfOperationPtr>( |
| 859 | dataType, this, nodeDef, tensorData, tensorInfo); |
| 860 | } |
| 861 | |
| 862 | template<typename Type> |
| 863 | bool TfParser::HasParsedConstTensor(const std::string & nodeName) const |
| 864 | { |
| 865 | auto it = m_ParsedTfOperations.find(nodeName); |
| 866 | if (it == m_ParsedTfOperations.end() || |
| 867 | dynamic_cast<ParsedConstTfOperation<Type>*>(it->second.get()) == nullptr) |
| 868 | { |
| 869 | return false; |
| 870 | } |
| 871 | else |
| 872 | { |
| 873 | return true; |
| 874 | } |
| 875 | } |
| 876 | |
| 877 | ParsedTfOperationPtr TfParser::ParseConv2D(const tensorflow::NodeDef& nodeDef, |
| 878 | const tensorflow::GraphDef& graphDef) |
| 879 | { |
| 880 | std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 2); |
| 881 | IOutputSlot& inputSlot = inputs[0].m_IndexedValue->ResolveArmnnOutputSlot(inputs[0].m_Index); |
| 882 | TensorInfo inputTensorInfo = inputSlot.GetTensorInfo(); |
| 883 | |
| 884 | if (!HasParsedConstTensor<float>(inputs[1].m_IndexedValue->GetNode().name())) |
| 885 | { |
| 886 | throw ParseException("ArmNN only supports Convolution layers with constant weights"); |
| 887 | } |
| 888 | ParsedConstTfOperation<float>* weightNode = |
| 889 | boost::polymorphic_downcast<ParsedConstTfOperation<float> *>(inputs[1].m_IndexedValue); |
| 890 | |
| 891 | std::string paddingString = ReadMandatoryNodeStringAttribute(nodeDef, "padding"); |
| 892 | std::string dataFormat = ReadMandatoryNodeStringAttribute(nodeDef, "data_format"); |
| 893 | std::vector<uint32_t> strides = ReadMandatoryNodeUint32ListAttribute(nodeDef, "strides"); |
| 894 | |
| 895 | // read the dilations, if present - only [1,1,1,1] (the default) is supported |
| 896 | std::vector<uint32_t> dilations = ReadOptionalNodeUint32ListAttribute(nodeDef, "dilations"); |
| 897 | if (!dilations.empty()) |
| 898 | { |
| 899 | for (auto dilation : dilations) |
| 900 | { |
| 901 | if (dilation != 1u) |
| 902 | { |
| 903 | throw ParseException("ArmNN only supports Convolution layers with dilations [1,1,1,1]"); |
| 904 | } |
| 905 | } |
| 906 | } |
| 907 | |
| 908 | Convolution2dDescriptor desc; |
| 909 | desc.m_BiasEnabled = false; |
| 910 | |
| 911 | if (dataFormat == "NHWC") |
| 912 | { |
| 913 | desc.m_StrideX = strides[2]; |
| 914 | desc.m_StrideY = strides[1]; |
| 915 | // Swizzle input to supported memory layout |
| 916 | inputTensorInfo = armnnUtils::Permuted(inputSlot.GetTensorInfo(), NHWCToArmNN); |
| 917 | } |
| 918 | else if (dataFormat == "NCHW") |
| 919 | { |
| 920 | desc.m_StrideX = strides[3]; |
| 921 | desc.m_StrideY = strides[2]; |
| 922 | } |
| 923 | else |
| 924 | { |
| 925 | throw ParseException("Unsupported data format passed for Conv2D. Only NHWC and NCHW supported"); |
| 926 | } |
| 927 | |
| 928 | uint32_t inputHeight = inputTensorInfo.GetShape()[2]; |
| 929 | uint32_t inputWidth = inputTensorInfo.GetShape()[3]; |
| 930 | |
| 931 | std::vector<float> outputTensorData; |
| 932 | |
| 933 | ConstTensor weightTensor = weightNode->GetConstTensor(true, outputTensorData); |
| 934 | |
| 935 | uint32_t weightHeight = weightTensor.GetShape()[2]; |
| 936 | uint32_t weightWidth = weightTensor.GetShape()[3]; |
| 937 | |
| 938 | bool padding = false; |
| 939 | TensorInfo outputInfo; |
| 940 | if (paddingString == "SAME") |
| 941 | { |
| 942 | padding = true; |
| 943 | outputInfo = TensorInfo({ inputTensorInfo.GetShape()[0], |
| 944 | weightTensor.GetShape()[0], |
| 945 | static_cast<uint32_t>(ceil( |
| 946 | static_cast<float>(inputHeight) / |
| 947 | static_cast<float>(desc.m_StrideY))), |
| 948 | static_cast<uint32_t>(ceil( |
| 949 | static_cast<float>(inputWidth) / |
| 950 | static_cast<float>(desc.m_StrideX))) |
| 951 | }, DataType::Float32); |
| 952 | } |
| 953 | else if (paddingString == "VALID") |
| 954 | { |
| 955 | padding = false; |
| 956 | outputInfo = TensorInfo({ inputTensorInfo.GetShape()[0], |
| 957 | weightTensor.GetShape()[0], |
| 958 | static_cast<uint32_t>(ceil( |
| 959 | static_cast<float>(inputHeight - weightHeight + 1) / |
| 960 | static_cast<float>(desc.m_StrideY))), |
| 961 | static_cast<uint32_t>(ceil( |
| 962 | static_cast<float>(inputWidth - weightWidth + 1) / |
| 963 | static_cast<float>(desc.m_StrideX))) |
| 964 | }, DataType::Float32); |
| 965 | } |
| 966 | else |
| 967 | { |
| 968 | throw ParseException("Only 'SAME' and 'VALID' padding supported"); |
| 969 | } |
| 970 | |
| 971 | CalcPadding(inputHeight, weightHeight, desc.m_StrideY, desc.m_PadTop, desc.m_PadBottom, padding); |
| 972 | CalcPadding(inputWidth, weightWidth, desc.m_StrideX, desc.m_PadLeft, desc.m_PadRight, padding); |
| 973 | |
| 974 | IConnectableLayer* layer = m_Network->AddConvolution2dLayer(desc, weightTensor, nodeDef.name().c_str()); |
| 975 | layer->GetOutputSlot(0).SetTensorInfo(outputInfo); |
| 976 | |
| 977 | if (dataFormat == "NHWC") |
| 978 | { |
| 979 | layer = SwizzleInDeswizzleOut(*m_Network, inputSlot, *layer, nodeDef.name()); |
| 980 | } |
| 981 | else |
| 982 | { |
| 983 | inputSlot.Connect(layer->GetInputSlot(0)); |
| 984 | } |
| 985 | |
| 986 | return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer); |
| 987 | } |
| 988 | |
| 989 | ParsedTfOperationPtr TfParser::ParseDepthwiseConv2D(const tensorflow::NodeDef& nodeDef, |
| 990 | const tensorflow::GraphDef& graphDef) |
| 991 | { |
| 992 | std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 2); |
| 993 | IOutputSlot& inputSlot = inputs[0].m_IndexedValue->ResolveArmnnOutputSlot(inputs[0].m_Index); |
| 994 | TensorInfo inputTensorInfo = inputSlot.GetTensorInfo(); |
| 995 | |
| 996 | if (!HasParsedConstTensor<float>(inputs[1].m_IndexedValue->GetNode().name())) |
| 997 | { |
| 998 | throw ParseException("ArmNN only supports Depthwise Convolution layers with constant weights"); |
| 999 | } |
| 1000 | ParsedConstTfOperation<float>* weightNode = |
| 1001 | boost::polymorphic_downcast<ParsedConstTfOperation<float> *>(inputs[1].m_IndexedValue); |
| 1002 | |
| 1003 | |
| 1004 | std::string paddingString = ReadMandatoryNodeStringAttribute(nodeDef, "padding"); |
| 1005 | std::string dataFormat = ReadMandatoryNodeStringAttribute(nodeDef, "data_format"); |
| 1006 | std::vector<uint32_t> strides = ReadMandatoryNodeUint32ListAttribute(nodeDef, "strides"); |
| 1007 | |
| 1008 | DepthwiseConvolution2dDescriptor desc; |
| 1009 | desc.m_BiasEnabled = false; |
| 1010 | |
| 1011 | if (dataFormat == "NHWC") |
| 1012 | { |
| 1013 | desc.m_StrideX = strides[2]; |
| 1014 | desc.m_StrideY = strides[1]; |
| 1015 | // Swizzle input to supported memory layout |
| 1016 | inputTensorInfo = armnnUtils::Permuted(inputSlot.GetTensorInfo(), NHWCToArmNN); |
| 1017 | } |
| 1018 | else if (dataFormat == "NCHW") |
| 1019 | { |
| 1020 | desc.m_StrideX = strides[3]; |
| 1021 | desc.m_StrideY = strides[2]; |
| 1022 | } |
| 1023 | else |
| 1024 | { |
| 1025 | throw ParseException("Unsupported data format passed for DepthwiseConv2dNative. Only NHWC and NCHW supported"); |
| 1026 | } |
| 1027 | |
| 1028 | uint32_t inputHeight = inputTensorInfo.GetShape()[2]; |
| 1029 | uint32_t inputWidth = inputTensorInfo.GetShape()[3]; |
| 1030 | |
| 1031 | std::vector<float> outputTensorData; |
| 1032 | |
| 1033 | ConstTensor weightTensor = weightNode->GetConstTensor(true, outputTensorData); |
| 1034 | |
| 1035 | uint32_t weightHeight = weightTensor.GetShape()[2]; |
| 1036 | uint32_t weightWidth = weightTensor.GetShape()[3]; |
| 1037 | |
| 1038 | bool padding = false; |
| 1039 | TensorInfo outputInfo; |
| 1040 | if (paddingString == "SAME") |
| 1041 | { |
| 1042 | padding = true; |
| 1043 | outputInfo = TensorInfo({ inputTensorInfo.GetShape()[0], |
| 1044 | weightTensor.GetShape()[0] * weightTensor.GetShape()[1], |
| 1045 | static_cast<uint32_t>(ceil( |
| 1046 | static_cast<float>(inputHeight) / |
| 1047 | static_cast<float>(desc.m_StrideY))), |
| 1048 | static_cast<uint32_t>(ceil( |
| 1049 | static_cast<float>(inputWidth) / |
| 1050 | static_cast<float>(desc.m_StrideX))) |
| 1051 | }, DataType::Float32); |
| 1052 | } |
| 1053 | else if (paddingString == "VALID") |
| 1054 | { |
| 1055 | padding = false; |
| 1056 | outputInfo = TensorInfo({ inputTensorInfo.GetShape()[0], |
| 1057 | weightTensor.GetShape()[0] * weightTensor.GetShape()[1], |
| 1058 | static_cast<uint32_t>(ceil( |
| 1059 | static_cast<float>(inputHeight - weightHeight + 1) / |
| 1060 | static_cast<float>(desc.m_StrideY))), |
| 1061 | static_cast<uint32_t>(ceil( |
| 1062 | static_cast<float>(inputWidth - weightWidth + 1) / |
| 1063 | static_cast<float>(desc.m_StrideX))) |
| 1064 | }, DataType::Float32); |
| 1065 | } |
| 1066 | else |
| 1067 | { |
| 1068 | throw ParseException("Only 'SAME' and 'VALID' padding supported"); |
| 1069 | } |
| 1070 | |
| 1071 | CalcPadding(inputHeight, weightHeight, desc.m_StrideY, desc.m_PadTop, desc.m_PadBottom, padding); |
| 1072 | CalcPadding(inputWidth, weightWidth, desc.m_StrideX, desc.m_PadLeft, desc.m_PadRight, padding); |
| 1073 | |
| 1074 | IConnectableLayer* layer = m_Network->AddDepthwiseConvolution2dLayer(desc, weightTensor, nodeDef.name().c_str()); |
| 1075 | layer->GetOutputSlot(0).SetTensorInfo(outputInfo); |
| 1076 | |
| 1077 | if (dataFormat == "NHWC") |
| 1078 | { |
| 1079 | layer = SwizzleInDeswizzleOut(*m_Network, inputSlot, *layer, nodeDef.name()); |
| 1080 | } |
| 1081 | else |
| 1082 | { |
| 1083 | inputSlot.Connect(layer->GetInputSlot(0)); |
| 1084 | } |
| 1085 | |
| 1086 | return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer); |
| 1087 | } |
| 1088 | |
| 1089 | ParsedTfOperationPtr TfParser::ParseFusedBatchNorm(const tensorflow::NodeDef& nodeDef, |
| 1090 | const tensorflow::GraphDef& graphDef) |
| 1091 | { |
| 1092 | std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 5); |
| 1093 | |
| 1094 | if (!HasParsedConstTensor<float>(inputs[1].m_IndexedValue->GetNode().name())) |
| 1095 | { |
| 1096 | throw ParseException("ArmNN only supports FusedBatchNormalization layers with constant scale"); |
| 1097 | } |
| 1098 | ParsedConstTfOperation<float>* scaleNode = |
| 1099 | boost::polymorphic_downcast<ParsedConstTfOperation<float> *>(inputs[1].m_IndexedValue); |
| 1100 | |
| 1101 | if (!HasParsedConstTensor<float>(inputs[2].m_IndexedValue->GetNode().name())) |
| 1102 | { |
| 1103 | throw ParseException("ArmNN only supports FusedBatchNormalization layers with constant offset"); |
| 1104 | } |
| 1105 | ParsedConstTfOperation<float>* offsetNode = |
| 1106 | boost::polymorphic_downcast<ParsedConstTfOperation<float> *>(inputs[2].m_IndexedValue); |
| 1107 | |
| 1108 | if (!HasParsedConstTensor<float>(inputs[3].m_IndexedValue->GetNode().name())) |
| 1109 | { |
| 1110 | throw ParseException("ArmNN only supports FusedBatchNormalization layers with constant mean"); |
| 1111 | } |
| 1112 | ParsedConstTfOperation<float>* meanNode = |
| 1113 | boost::polymorphic_downcast<ParsedConstTfOperation<float> *>(inputs[3].m_IndexedValue); |
| 1114 | |
| 1115 | if (!HasParsedConstTensor<float>(inputs[4].m_IndexedValue->GetNode().name())) |
| 1116 | { |
| 1117 | throw ParseException("ArmNN only supports FusedBatchNormalization layers with constant variance"); |
| 1118 | } |
| 1119 | ParsedConstTfOperation<float>* varianceNode = |
| 1120 | boost::polymorphic_downcast<ParsedConstTfOperation<float> *>(inputs[4].m_IndexedValue); |
| 1121 | |
| 1122 | // The descriptor only has the epsilon attribute |
| 1123 | BatchNormalizationDescriptor desc; |
| 1124 | desc.m_Eps = ReadMandatoryNodeFloatAttribute(nodeDef, "epsilon"); |
| 1125 | |
| 1126 | // data for the parsed tensor args (scale, offset, mean, variance) must be stored locally until the layer is added |
| 1127 | std::vector<float> scaleTensorData; |
| 1128 | ConstTensor scaleTensor = scaleNode->GetConstTensor(false, scaleTensorData); |
| 1129 | |
| 1130 | std::vector<float> offsetTensorData; |
| 1131 | ConstTensor offsetTensor = offsetNode->GetConstTensor(false, offsetTensorData); |
| 1132 | |
| 1133 | std::vector<float> meanTensorData; |
| 1134 | ConstTensor meanTensor = meanNode->GetConstTensor(false, meanTensorData); |
| 1135 | |
| 1136 | std::vector<float> varianceTensorData; |
| 1137 | ConstTensor varianceTensor = varianceNode->GetConstTensor(false, varianceTensorData); |
| 1138 | |
| 1139 | IConnectableLayer* layer = m_Network->AddBatchNormalizationLayer(desc, |
| 1140 | meanTensor, |
| 1141 | varianceTensor, |
| 1142 | offsetTensor, |
| 1143 | scaleTensor, |
| 1144 | nodeDef.name().c_str()); |
| 1145 | |
| 1146 | IOutputSlot& inputSlot = inputs[0].m_IndexedValue->ResolveArmnnOutputSlot(inputs[0].m_Index); |
| 1147 | |
| 1148 | const std::string dataFormat = ReadMandatoryNodeStringAttribute(nodeDef, "data_format"); |
| 1149 | |
| 1150 | if (dataFormat == "NHWC") |
| 1151 | { |
| 1152 | const TensorInfo outputTensorInfo = armnnUtils::Permuted(inputSlot.GetTensorInfo(), NHWCToArmNN); |
| 1153 | layer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo); |
| 1154 | layer = SwizzleInDeswizzleOut(*m_Network, inputSlot, *layer, nodeDef.name()); |
| 1155 | } |
| 1156 | else |
| 1157 | { |
| 1158 | layer->GetOutputSlot(0).SetTensorInfo(inputSlot.GetTensorInfo()); |
| 1159 | inputSlot.Connect(layer->GetInputSlot(0)); |
| 1160 | } |
| 1161 | |
| 1162 | return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer); |
| 1163 | } |
| 1164 | |
| 1165 | ParsedTfOperationPtr TfParser::ParseConcat(const tensorflow::NodeDef& nodeDef, |
| 1166 | const tensorflow::GraphDef& graphDef) |
| 1167 | { |
| 1168 | std::vector<OutputOfConstNodeDef> nodes = GetTfInputNodes(nodeDef); |
| 1169 | // In tensorflow, we have the last input of the Concat layer as the axis for concatenation |
| 1170 | unsigned int numInputs = static_cast<unsigned int>(nodes.size()); |
| 1171 | unsigned int numConcatView = numInputs - 1; |
| 1172 | |
| 1173 | OriginsDescriptor concatDescriptor(static_cast<uint32_t>(numConcatView), MaxNumOfTensorDimensions); |
| 1174 | std::vector<unsigned int>mergeDimSizes(MaxNumOfTensorDimensions, 0u); |
| 1175 | |
| 1176 | unsigned int mergeDim = 0; |
| 1177 | std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, numInputs); |
| 1178 | |
| 1179 | // The last input is the axis for concatenation |
| 1180 | if (!HasParsedConstTensor<int32_t>(inputs[numInputs - 1].m_IndexedValue->GetNode().name())) |
| 1181 | { |
| 1182 | throw ParseException("ArmNN only supports Concat with constant axis"); |
| 1183 | } |
| 1184 | ParsedConstTfOperation<int32_t>* shapeNode = |
| 1185 | boost::polymorphic_downcast<ParsedConstTfOperation<int32_t>*>(inputs[numInputs - 1].m_IndexedValue); |
| 1186 | |
| 1187 | std::vector<int32_t> axisTensorData; |
| 1188 | ConstTensor axisTensor = shapeNode->GetConstTensor(false, axisTensorData); |
| 1189 | |
| 1190 | // This concatDim indicates the data format: 3 is the NHWC, 1 is the NCHW |
| 1191 | const unsigned int concatDimInput = static_cast<unsigned int>(axisTensorData[0]); |
| 1192 | |
| 1193 | // Armnn supports concatenation along the channel dimension for data format NHWC and NCHW |
| 1194 | if (concatDimInput == 0 || concatDimInput == 2) |
| 1195 | { |
| 1196 | throw ParseException("The dimension for concatenation is not supported by Armnn"); |
| 1197 | } |
| 1198 | |
| 1199 | // This is the only concatDim we support in Armnn |
| 1200 | const unsigned int concatDim = 1; |
| 1201 | for (unsigned int viewIndex = 0; viewIndex < numConcatView; ++viewIndex) |
| 1202 | { |
| 1203 | // need to double check whether it should be |
| 1204 | IOutputSlot& inputSlot = |
| 1205 | inputs[viewIndex].m_IndexedValue->ResolveArmnnOutputSlot(inputs[viewIndex].m_Index); |
| 1206 | TensorInfo inputTensorInfo = inputSlot.GetTensorInfo(); |
| 1207 | |
| 1208 | if (inputTensorInfo.GetNumDimensions() != MaxNumOfTensorDimensions) |
| 1209 | { |
| 1210 | throw ParseException("The number of dimensions for input tensors of the concatenation op should be 4"); |
| 1211 | } |
| 1212 | |
| 1213 | if (concatDimInput == 3) |
| 1214 | { |
| 1215 | inputTensorInfo = armnnUtils::Permuted(inputTensorInfo, NHWCToArmNN); |
| 1216 | } |
| 1217 | |
| 1218 | for (unsigned int dim = 0; dim < MaxNumOfTensorDimensions; ++dim) |
| 1219 | { |
| 1220 | mergeDimSizes[dim] = inputTensorInfo.GetShape()[dim]; |
| 1221 | } |
| 1222 | |
| 1223 | for (unsigned int j = 0; j < concatDim; ++j) |
| 1224 | { |
| 1225 | concatDescriptor.SetViewOriginCoord(viewIndex, j, 0); |
| 1226 | } |
| 1227 | |
| 1228 | concatDescriptor.SetViewOriginCoord(viewIndex, concatDim, mergeDim); |
| 1229 | mergeDim += mergeDimSizes[concatDim]; |
| 1230 | |
| 1231 | for (unsigned int j = concatDim+1; j < MaxNumOfTensorDimensions; ++j) |
| 1232 | { |
| 1233 | concatDescriptor.SetViewOriginCoord(viewIndex, j, 0); |
| 1234 | } |
| 1235 | } |
| 1236 | |
| 1237 | mergeDimSizes[concatDim] = mergeDim; |
| 1238 | armnn::IConnectableLayer *layer = m_Network->AddMergerLayer(concatDescriptor, nodeDef.name().c_str()); |
| 1239 | |
| 1240 | layer->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo(MaxNumOfTensorDimensions, mergeDimSizes.data(), |
| 1241 | DataType::Float32)); |
| 1242 | |
| 1243 | for (unsigned int v = 0; v < numConcatView; ++v) |
| 1244 | { |
| 1245 | IOutputSlot& inputSlot = inputs[v].m_IndexedValue->ResolveArmnnOutputSlot(inputs[v].m_Index); |
| 1246 | if (concatDimInput == 3) |
| 1247 | { |
| 1248 | IConnectableLayer* const swizzleLayer = AddSwizzleLayer(*m_Network, inputSlot, NHWCToArmNN, |
| 1249 | "swizzle_for-" + nodeDef.name()); |
| 1250 | swizzleLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(v)); |
| 1251 | } |
| 1252 | else |
| 1253 | { |
| 1254 | inputSlot.Connect(layer->GetInputSlot(v)); |
| 1255 | } |
| 1256 | } |
| 1257 | |
| 1258 | if (concatDimInput == 3) |
| 1259 | { |
| 1260 | IConnectableLayer* const deswizzleLayer = AddSwizzleLayer(*m_Network, layer->GetOutputSlot(0), ArmNNToNHWC, |
| 1261 | "deswizzle_for-" + nodeDef.name()); |
| 1262 | layer = deswizzleLayer; |
| 1263 | } |
| 1264 | |
| 1265 | return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer); |
| 1266 | } |
| 1267 | |
| 1268 | ParsedTfOperationPtr TfParser::ParseShape(const tensorflow::NodeDef& nodeDef, |
| 1269 | const tensorflow::GraphDef& graphDef) |
| 1270 | { |
| 1271 | // Note: The Shape layer is handled in a special way, because: |
| 1272 | // 1. ARMNN doesn't support int32 tensors which it outputs |
| 1273 | // 2. ARMNN works with statically shaped tensors which are known at parse time |
| 1274 | // 3. because of 1. and 2. we treat the output of Shape as a temporary const int32 |
| 1275 | // tensor which may be used as an input to other ops, most likely a Reshape |
| 1276 | |
| 1277 | const tensorflow::DataType tfDataType = ReadMandatoryNodeTypeAttribute(nodeDef, "out_type"); |
| 1278 | if (tfDataType != tensorflow::DT_INT32) |
| 1279 | { |
| 1280 | throw ParseException("Armnn only supports DT_INT32 as out_type"); |
| 1281 | } |
| 1282 | |
| 1283 | const std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 1); |
| 1284 | IOutputSlot& prevLayerOutputSlot = inputs[0].m_IndexedValue->ResolveArmnnOutputSlot(inputs[0].m_Index); |
| 1285 | const TensorInfo& prevLayerTensorInfo = prevLayerOutputSlot.GetTensorInfo(); |
| 1286 | unsigned int prevLayerDimensions = prevLayerTensorInfo.GetNumDimensions(); |
| 1287 | |
| 1288 | std::vector<int32_t> shapeTensorData; |
| 1289 | shapeTensorData.reserve(prevLayerDimensions); |
| 1290 | |
| 1291 | for (unsigned int i=0; i<prevLayerDimensions; ++i) |
| 1292 | { |
| 1293 | shapeTensorData.push_back(static_cast<int32_t>(prevLayerTensorInfo.GetShape()[i])); |
| 1294 | } |
| 1295 | |
| 1296 | TensorInfo shapeTensorInfo(1, &prevLayerDimensions, DataType::Signed32); |
| 1297 | |
| 1298 | return std::make_unique<ParsedConstTfOperation<int32_t>>(this, |
| 1299 | nodeDef, |
| 1300 | &shapeTensorData[0], |
| 1301 | shapeTensorInfo); |
| 1302 | } |
| 1303 | |
| 1304 | ParsedTfOperationPtr TfParser::ParseReshape(const tensorflow::NodeDef& nodeDef, |
| 1305 | const tensorflow::GraphDef& graphDef) |
| 1306 | { |
| 1307 | std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 2); |
| 1308 | ParsedTfOperation* inputNode = inputs[0].m_IndexedValue; |
| 1309 | |
| 1310 | if (!HasParsedConstTensor<int32_t>(inputs[1].m_IndexedValue->GetNode().name())) |
| 1311 | { |
| 1312 | throw ParseException("ArmNN only supports Reshape layers with constant shapes"); |
| 1313 | } |
| 1314 | ParsedConstTfOperation<int32_t>* shapeNode = |
| 1315 | boost::polymorphic_downcast<ParsedConstTfOperation<int32_t>*>(inputs[1].m_IndexedValue); |
| 1316 | |
| 1317 | armnn::IOutputSlot& prevLayerOutputSlot = inputNode->ResolveArmnnOutputSlot(inputs[0].m_Index); |
| 1318 | TensorInfo inputTensorInfo = prevLayerOutputSlot.GetTensorInfo(); |
| 1319 | |
| 1320 | std::vector<int32_t> shapeTensorData; |
| 1321 | ConstTensor shapeTensor = shapeNode->GetConstTensor(false, shapeTensorData); |
| 1322 | const TensorInfo outputTensorInfo = PrepareReshape(inputTensorInfo, shapeTensorData); |
| 1323 | |
| 1324 | TensorShape targetShape = outputTensorInfo.GetShape(); |
| 1325 | ReshapeDescriptor reshapeDesc; |
| 1326 | reshapeDesc.m_TargetShape = targetShape; |
| 1327 | |
| 1328 | IConnectableLayer* layer = m_Network->AddReshapeLayer(reshapeDesc, nodeDef.name().c_str()); |
| 1329 | prevLayerOutputSlot.Connect(layer->GetInputSlot(0)); |
| 1330 | layer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo); |
| 1331 | |
| 1332 | return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer); |
| 1333 | } |
| 1334 | |
| 1335 | ParsedTfOperationPtr TfParser::ParseResizeBilinear(const tensorflow::NodeDef& nodeDef, |
| 1336 | const tensorflow::GraphDef& graphDef) |
| 1337 | { |
| 1338 | std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 2); |
| 1339 | |
| 1340 | if (!HasParsedConstTensor<int32_t>(inputs[1].m_IndexedValue->GetNode().name())) |
| 1341 | { |
| 1342 | throw ParseException("ArmNN only supports ResizeBilinear layers with constant sizes"); |
| 1343 | } |
| 1344 | ParsedConstTfOperation<int32_t>* sizeNode = |
| 1345 | boost::polymorphic_downcast<ParsedConstTfOperation<int32_t>*>(inputs[1].m_IndexedValue); |
| 1346 | |
| 1347 | // Check the align_corners attribute is not set |
| 1348 | if (ReadOptionalNodeBoolAttribute(nodeDef, "align_corners", false)) |
| 1349 | { |
| 1350 | throw ParseException("ArmNN only supports ResizeBilinear layers with align_corners set to false"); |
| 1351 | } |
| 1352 | |
| 1353 | // data for the parsed tensor args (size) must be stored locally |
| 1354 | std::vector<int32_t> sizeTensorData; |
| 1355 | ConstTensor sizeTensor = sizeNode->GetConstTensor(false, sizeTensorData); |
| 1356 | |
| 1357 | // The descriptor only has target height and width attributes, which we get from the size tensor |
| 1358 | ResizeBilinearDescriptor desc; |
| 1359 | desc.m_TargetHeight = static_cast<uint32_t> (sizeTensorData[0]); |
| 1360 | desc.m_TargetWidth = static_cast<uint32_t> (sizeTensorData[1]); |
| 1361 | |
| 1362 | IConnectableLayer* layer = m_Network->AddResizeBilinearLayer(desc, nodeDef.name().c_str()); |
| 1363 | |
| 1364 | IOutputSlot& inputSlot = inputs[0].m_IndexedValue->ResolveArmnnOutputSlot(inputs[0].m_Index); |
| 1365 | TensorInfo inputTensorInfo = inputSlot.GetTensorInfo(); |
| 1366 | // the input shape is always in BHWC format, this will be swizzled below; for now, |
| 1367 | // get the batch and channels to make up the ArmNN output shape with the target size |
| 1368 | unsigned int outBatch = inputTensorInfo.GetShape()[0]; |
| 1369 | unsigned int outChannels = inputTensorInfo.GetShape()[3]; |
| 1370 | unsigned int outHeight = desc.m_TargetHeight; |
| 1371 | unsigned int outWidth = desc.m_TargetWidth; |
| 1372 | TensorShape outShape({outBatch, outChannels, outHeight, outWidth}); |
| 1373 | // The output DataType is always Float32, regardless of the input DataType |
| 1374 | const TensorInfo outputTensorInfo(outShape, armnn::DataType::Float32); |
| 1375 | layer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo); |
| 1376 | |
| 1377 | // TensorFlow ResizeBilinear input is always in BHWC format, so add swizzle and deswizzle layers |
| 1378 | layer = SwizzleInDeswizzleOut(*m_Network, inputSlot, *layer, nodeDef.name()); |
| 1379 | |
| 1380 | return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer); |
| 1381 | } |
| 1382 | |
| 1383 | TensorInfo OutputShapeOfSqueeze(const tensorflow::NodeDef& nodeDef, TensorInfo inputTensorInfo) |
| 1384 | { |
| 1385 | BOOST_ASSERT(nodeDef.op() == "Squeeze"); |
| 1386 | tensorflow::DataType tfDataType = ReadMandatoryNodeTypeAttribute(nodeDef, "T"); |
| 1387 | |
| 1388 | DataType type; |
| 1389 | if (tfDataType == tensorflow::DT_FLOAT) |
| 1390 | { |
| 1391 | type = DataType::Float32; |
| 1392 | } |
| 1393 | else if (tfDataType == tensorflow::DT_INT32) |
| 1394 | { |
| 1395 | type = DataType::Signed32; |
| 1396 | } |
| 1397 | else |
| 1398 | { |
| 1399 | throw ParseException(boost::str( |
| 1400 | boost::format("Unsupported DataType %1% for Squeeze operation") |
| 1401 | % tensorflow::DataType_Name(tfDataType))); |
| 1402 | } |
| 1403 | |
| 1404 | std::vector<uint32_t> squeezeDims = ReadOptionalNodeUint32ListAttribute(nodeDef, "squeeze_dims"); |
| 1405 | if (squeezeDims.empty()) |
| 1406 | { |
| 1407 | for(unsigned int i = 0; i < inputTensorInfo.GetNumDimensions(); i++) |
| 1408 | { |
| 1409 | if (inputTensorInfo.GetShape()[i] == 1) |
| 1410 | { |
| 1411 | squeezeDims.push_back(i); |
| 1412 | } |
| 1413 | } |
| 1414 | } |
| 1415 | |
| 1416 | std::vector<uint32_t> outputDims; |
| 1417 | for(unsigned int i = 0; i < inputTensorInfo.GetNumDimensions(); i++) |
| 1418 | { |
| 1419 | bool includeDimension = (std::find(squeezeDims.begin(), squeezeDims.end(), i) == squeezeDims.end()); |
| 1420 | if (includeDimension) |
| 1421 | { |
| 1422 | outputDims.push_back(inputTensorInfo.GetShape()[i]); |
| 1423 | } |
| 1424 | } |
| 1425 | |
| 1426 | if (outputDims.size() > 4) |
| 1427 | { |
| 1428 | throw ParseException("Unsupported shape for Squeeze"); |
| 1429 | } |
| 1430 | |
| 1431 | TensorInfo outTensorInfo = TensorInfo(boost::numeric_cast<unsigned int>(outputDims.size()), |
| 1432 | outputDims.data(), |
| 1433 | type); |
| 1434 | |
| 1435 | return outTensorInfo; |
| 1436 | } |
| 1437 | |
| 1438 | ParsedTfOperationPtr TfParser::ParseSqueeze(const tensorflow::NodeDef& nodeDef, const tensorflow::GraphDef& graphDef) |
| 1439 | { |
| 1440 | std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 1); |
| 1441 | |
| 1442 | IOutputSlot& prevLayerOutputSlot = inputs[0].m_IndexedValue->ResolveArmnnOutputSlot(inputs[0].m_Index); |
| 1443 | TensorInfo inputTensorInfo = prevLayerOutputSlot.GetTensorInfo(); |
| 1444 | |
| 1445 | TensorInfo outputInfo; |
| 1446 | outputInfo = OutputShapeOfSqueeze(nodeDef, inputTensorInfo); |
| 1447 | |
| 1448 | ReshapeDescriptor reshapeDesc; |
| 1449 | reshapeDesc.m_TargetShape = outputInfo.GetShape(); |
| 1450 | IConnectableLayer* layer = m_Network->AddReshapeLayer(reshapeDesc, nodeDef.name().c_str()); |
| 1451 | prevLayerOutputSlot.Connect(layer->GetInputSlot(0)); |
| 1452 | layer->GetOutputSlot(0).SetTensorInfo(outputInfo); |
| 1453 | |
| 1454 | return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer); |
| 1455 | } |
| 1456 | |
| 1457 | ParsedTfOperationPtr TfParser::ParseLrn(const tensorflow::NodeDef& nodeDef, const tensorflow::GraphDef& graphDef) |
| 1458 | { |
| 1459 | std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 1); |
| 1460 | |
| 1461 | NormalizationDescriptor normalizationDescriptor; |
| 1462 | normalizationDescriptor.m_NormMethodType = NormalizationAlgorithmMethod::LocalBrightness; |
| 1463 | normalizationDescriptor.m_NormChannelType = NormalizationAlgorithmChannel::Across; |
| 1464 | normalizationDescriptor.m_Alpha = ReadMandatoryNodeFloatAttribute(nodeDef, "alpha"); |
| 1465 | normalizationDescriptor.m_Beta = ReadMandatoryNodeFloatAttribute(nodeDef, "beta"); |
| 1466 | normalizationDescriptor.m_K = ReadMandatoryNodeFloatAttribute(nodeDef, "bias"); |
| 1467 | normalizationDescriptor.m_NormSize = ReadMandatoryNodeUint32Attribute(nodeDef, "depth_radius"); |
| 1468 | |
| 1469 | // The window size must be an odd value. For a window size of (2 * n + 1), TensorFlow defines depth_radius = n. |
| 1470 | normalizationDescriptor.m_NormSize = normalizationDescriptor.m_NormSize * 2 + 1; |
| 1471 | |
| 1472 | IOutputSlot& prevLayerOutputSlot = inputs[0].m_IndexedValue->ResolveArmnnOutputSlot(inputs[0].m_Index); |
| 1473 | |
| 1474 | IConnectableLayer* layer = m_Network->AddNormalizationLayer(normalizationDescriptor, |
| 1475 | nodeDef.name().c_str()); |
| 1476 | |
| 1477 | const TensorInfo permutedInfo = armnnUtils::Permuted(prevLayerOutputSlot.GetTensorInfo(), NHWCToArmNN); |
| 1478 | layer->GetOutputSlot(0).SetTensorInfo(permutedInfo); |
| 1479 | |
| 1480 | layer = SwizzleInDeswizzleOut(*m_Network, prevLayerOutputSlot, *layer, nodeDef.name()); |
| 1481 | |
| 1482 | return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer); |
| 1483 | } |
| 1484 | |
| 1485 | /// An ParsedTfOperation for a MatMul node. |
| 1486 | /// Creation of the armnn FullyConnected layer is deferred until it is actually needed, because MatMul nodes are |
| 1487 | /// often used for the first part of a biased FullyConnected (MatMul followed by Add) and in these cases armnn doesn't |
| 1488 | /// need a separate layer for the MatMul. |
| 1489 | class ParsedMatMulTfOperation : public DeferredSingleLayerParsedTfOperation |
| 1490 | { |
| 1491 | public: |
| 1492 | ParsedMatMulTfOperation(TfParser* parser, const tensorflow::NodeDef& node) |
| 1493 | : DeferredSingleLayerParsedTfOperation(parser, node) |
| 1494 | { |
| 1495 | } |
| 1496 | |
| 1497 | void CreateLayerDeferred() override |
| 1498 | { |
| 1499 | BOOST_ASSERT(m_Layer == nullptr); |
| 1500 | m_Layer = m_Parser->AddFullyConnectedLayer(m_Node, nullptr, m_Node.name().c_str()); |
| 1501 | } |
| 1502 | }; |
| 1503 | |
| 1504 | ParsedTfOperationPtr TfParser::ParseMatMul(const tensorflow::NodeDef& nodeDef, const tensorflow::GraphDef& graphDef) |
| 1505 | { |
| 1506 | // Defer the creation of the layer (see ParsedMatMulTfOperation). |
| 1507 | return std::make_unique<ParsedMatMulTfOperation>(this, nodeDef); |
| 1508 | } |
| 1509 | |
| 1510 | ParsedTfOperationPtr TfParser::ParseMul(const tensorflow::NodeDef& nodeDef, const tensorflow::GraphDef& graphDef) |
| 1511 | { |
| 1512 | boost::ignore_unused(graphDef); |
| 1513 | |
| 1514 | std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 2); |
| 1515 | |
| 1516 | IConnectableLayer* const layer = m_Network->AddMultiplicationLayer(nodeDef.name().c_str()); |
| 1517 | IOutputSlot* input0Slot = &inputs[0].m_IndexedValue->ResolveArmnnOutputSlot(inputs[0].m_Index); |
| 1518 | IOutputSlot* input1Slot = &inputs[1].m_IndexedValue->ResolveArmnnOutputSlot(inputs[1].m_Index); |
| 1519 | |
| 1520 | auto const input0NumDims = input0Slot->GetTensorInfo().GetNumDimensions(); |
| 1521 | auto const input1NumDims = input1Slot->GetTensorInfo().GetNumDimensions(); |
| 1522 | |
| 1523 | if (input0NumDims < input1NumDims) |
| 1524 | { |
| 1525 | const bool isNHWC = true; |
| 1526 | input0Slot = BroadcastForAddandMul(input1Slot, input0Slot, isNHWC, *m_Network, nodeDef); |
| 1527 | } |
| 1528 | if (input1NumDims < input0NumDims) |
| 1529 | { |
| 1530 | const bool isNHWC = true; |
| 1531 | input1Slot = BroadcastForAddandMul(input0Slot, input1Slot, isNHWC, *m_Network, nodeDef); |
| 1532 | } |
| 1533 | |
| 1534 | input0Slot->Connect(layer->GetInputSlot(0)); |
| 1535 | input1Slot->Connect(layer->GetInputSlot(1)); |
| 1536 | |
| 1537 | if (input0NumDims < input1NumDims) |
| 1538 | { |
| 1539 | layer->GetOutputSlot(0).SetTensorInfo(input1Slot->GetTensorInfo()); |
| 1540 | } |
| 1541 | else |
| 1542 | { |
| 1543 | layer->GetOutputSlot(0).SetTensorInfo(input0Slot->GetTensorInfo()); |
| 1544 | } |
| 1545 | return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer); |
| 1546 | } |
| 1547 | |
| 1548 | ParsedTfOperationPtr TfParser::ParsePlaceholder(const tensorflow::NodeDef& nodeDef, |
| 1549 | const tensorflow::GraphDef& graphDef) |
| 1550 | { |
| 1551 | boost::ignore_unused(graphDef); |
| 1552 | |
| 1553 | std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 0); |
| 1554 | |
| 1555 | const LayerBindingId layerId = boost::numeric_cast<LayerBindingId>(m_NetworkInputsBindingInfo.size()); |
| 1556 | |
| 1557 | auto it = m_InputShapes.find(nodeDef.name()); |
| 1558 | if (it == m_InputShapes.end()) |
| 1559 | { |
| 1560 | throw ParseException("Missing input shape for Placeholder '" + nodeDef.name() + "'"); |
| 1561 | } |
| 1562 | TensorInfo tensorInfo(it->second, DataType::Float32); |
| 1563 | |
| 1564 | IConnectableLayer* const layer = m_Network->AddInputLayer(layerId, nodeDef.name().c_str()); |
| 1565 | |
| 1566 | layer->GetOutputSlot(0).SetTensorInfo(tensorInfo); |
| 1567 | |
| 1568 | TrackInputBinding(layer, layerId, tensorInfo); |
| 1569 | |
| 1570 | return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer); |
| 1571 | } |
| 1572 | |
| 1573 | ParsedTfOperationPtr TfParser::ParseRelu(const tensorflow::NodeDef& nodeDef, |
| 1574 | const tensorflow::GraphDef& graphDef) |
| 1575 | { |
| 1576 | boost::ignore_unused(graphDef); |
| 1577 | |
| 1578 | ActivationDescriptor activationDesc; |
| 1579 | activationDesc.m_Function = ActivationFunction::ReLu; |
| 1580 | return AddActivationLayer(nodeDef, activationDesc); |
| 1581 | } |
| 1582 | |
| 1583 | ParsedTfOperationPtr TfParser::ParseRelu6(const tensorflow::NodeDef& nodeDef, |
| 1584 | const tensorflow::GraphDef& graphDef) |
| 1585 | { |
| 1586 | boost::ignore_unused(graphDef); |
| 1587 | |
| 1588 | ActivationDescriptor activationDesc; |
| 1589 | activationDesc.m_Function = ActivationFunction::BoundedReLu; |
| 1590 | activationDesc.m_A = 6.0f; |
| 1591 | activationDesc.m_B = 0.0f; |
| 1592 | |
| 1593 | return AddActivationLayer(nodeDef, activationDesc); |
| 1594 | } |
| 1595 | |
| 1596 | ParsedTfOperationPtr TfParser::ParseSigmoid(const tensorflow::NodeDef& nodeDef, |
| 1597 | const tensorflow::GraphDef& graphDef) |
| 1598 | { |
| 1599 | boost::ignore_unused(graphDef); |
| 1600 | |
| 1601 | ActivationDescriptor activationDesc; |
| 1602 | activationDesc.m_Function = ActivationFunction::Sigmoid; |
| 1603 | |
| 1604 | return AddActivationLayer(nodeDef, activationDesc); |
| 1605 | } |
| 1606 | |
| 1607 | ParsedTfOperationPtr TfParser::ParseSoftmax(const tensorflow::NodeDef& nodeDef, |
| 1608 | const tensorflow::GraphDef& graphDef) |
| 1609 | { |
| 1610 | boost::ignore_unused(graphDef); |
| 1611 | |
| 1612 | std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 1); |
| 1613 | |
| 1614 | SoftmaxDescriptor softmaxDescriptor; |
| 1615 | IConnectableLayer* const layer = m_Network->AddSoftmaxLayer(softmaxDescriptor, nodeDef.name().c_str()); |
| 1616 | |
| 1617 | IOutputSlot& prevLayerSlot = inputs[0].m_IndexedValue->ResolveArmnnOutputSlot(inputs[0].m_Index); |
| 1618 | prevLayerSlot.Connect(layer->GetInputSlot(0)); |
| 1619 | layer->GetOutputSlot(0).SetTensorInfo(prevLayerSlot.GetTensorInfo()); |
| 1620 | |
| 1621 | return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer); |
| 1622 | } |
| 1623 | |
| 1624 | ParsedTfOperationPtr TfParser::ParseSoftplus(const tensorflow::NodeDef& nodeDef, |
| 1625 | const tensorflow::GraphDef& graphDef) |
| 1626 | { |
| 1627 | boost::ignore_unused(graphDef); |
| 1628 | |
| 1629 | ActivationDescriptor activationDesc; |
| 1630 | activationDesc.m_Function = ActivationFunction::SoftReLu; |
| 1631 | |
| 1632 | return AddActivationLayer(nodeDef, activationDesc); |
| 1633 | } |
| 1634 | |
| 1635 | ParsedTfOperationPtr TfParser::ParseTanh(const tensorflow::NodeDef& nodeDef, const tensorflow::GraphDef& graphDef) |
| 1636 | { |
| 1637 | boost::ignore_unused(graphDef); |
| 1638 | |
| 1639 | ActivationDescriptor activationDesc; |
| 1640 | activationDesc.m_Function = ActivationFunction::TanH; |
| 1641 | activationDesc.m_A = 1.0f; |
| 1642 | activationDesc.m_B = 1.0f; |
| 1643 | |
| 1644 | return AddActivationLayer(nodeDef, activationDesc); |
| 1645 | } |
| 1646 | |
| 1647 | ParsedTfOperationPtr TfParser::AddActivationLayer(const tensorflow::NodeDef& nodeDef, |
| 1648 | ActivationDescriptor& activationDesc) |
| 1649 | { |
| 1650 | std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 1); |
| 1651 | |
| 1652 | IConnectableLayer* const layer = m_Network->AddActivationLayer(activationDesc, nodeDef.name().c_str()); |
| 1653 | |
| 1654 | IOutputSlot& prevLayerOutputSlot = inputs[0].m_IndexedValue->ResolveArmnnOutputSlot(inputs[0].m_Index); |
| 1655 | prevLayerOutputSlot.Connect(layer->GetInputSlot(0)); |
| 1656 | layer->GetOutputSlot(0).SetTensorInfo(prevLayerOutputSlot.GetTensorInfo()); |
| 1657 | return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer); |
| 1658 | } |
| 1659 | |
| 1660 | ParsedTfOperationPtr TfParser::ParseMaxPool(const tensorflow::NodeDef& nodeDef, |
| 1661 | const tensorflow::GraphDef& graphDef) |
| 1662 | { |
| 1663 | return ParsePooling2d(nodeDef, graphDef, PoolingAlgorithm::Max); |
| 1664 | } |
| 1665 | |
| 1666 | ParsedTfOperationPtr TfParser::ParseAvgPool(const tensorflow::NodeDef& nodeDef, |
| 1667 | const tensorflow::GraphDef& graphDef) |
| 1668 | { |
| 1669 | return ParsePooling2d(nodeDef, graphDef, PoolingAlgorithm::Average); |
| 1670 | } |
| 1671 | |
| 1672 | ParsedTfOperationPtr TfParser::ParsePooling2d(const tensorflow::NodeDef& nodeDef, |
| 1673 | const tensorflow::GraphDef& graphDef, PoolingAlgorithm pooltype) |
| 1674 | { |
| 1675 | std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 1); |
| 1676 | IOutputSlot& inputSlot = inputs[0].m_IndexedValue->ResolveArmnnOutputSlot(inputs[0].m_Index); |
| 1677 | TensorInfo inputTensorInfo = inputSlot.GetTensorInfo(); |
| 1678 | |
| 1679 | if (inputs.size() != 1) |
| 1680 | { |
| 1681 | throw ParseException("2D Pooling expects one input!"); |
| 1682 | } |
| 1683 | |
| 1684 | std::string paddingString = ReadMandatoryNodeStringAttribute(nodeDef, "padding"); |
| 1685 | std::string dataFormat = ReadMandatoryNodeStringAttribute(nodeDef, "data_format"); |
| 1686 | std::vector<uint32_t> strides = ReadMandatoryNodeUint32ListAttribute(nodeDef, "strides"); |
| 1687 | std::vector<uint32_t> ksize = ReadMandatoryNodeUint32ListAttribute(nodeDef, "ksize"); // size of pool windows |
| 1688 | |
| 1689 | Pooling2dDescriptor pooling2dDescriptor; |
| 1690 | pooling2dDescriptor.m_PoolType = pooltype; |
| 1691 | pooling2dDescriptor.m_PaddingMethod = PaddingMethod::Exclude; |
| 1692 | pooling2dDescriptor.m_OutputShapeRounding = OutputShapeRounding::Floor; |
| 1693 | |
| 1694 | if (dataFormat == "NHWC") |
| 1695 | { |
| 1696 | pooling2dDescriptor.m_StrideX = strides[2]; |
| 1697 | pooling2dDescriptor.m_StrideY = strides[1]; |
| 1698 | pooling2dDescriptor.m_PoolWidth = ksize[2]; |
| 1699 | pooling2dDescriptor.m_PoolHeight = ksize[1]; |
| 1700 | // Swizzle input to supported memory layout |
| 1701 | inputTensorInfo = armnnUtils::Permuted(inputSlot.GetTensorInfo(), NHWCToArmNN); |
| 1702 | } |
| 1703 | else if (dataFormat == "NCHW") |
| 1704 | { |
| 1705 | pooling2dDescriptor.m_StrideX = strides[3]; |
| 1706 | pooling2dDescriptor.m_StrideY = strides[2]; |
| 1707 | pooling2dDescriptor.m_PoolWidth = ksize[3]; |
| 1708 | pooling2dDescriptor.m_PoolHeight = ksize[2]; |
| 1709 | } |
| 1710 | else |
| 1711 | { |
| 1712 | throw ParseException("Only NHWC or NCHW supported for Pooling2d"); |
| 1713 | } |
| 1714 | |
| 1715 | uint32_t inputHeight = inputTensorInfo.GetShape()[2]; |
| 1716 | uint32_t inputWidth = inputTensorInfo.GetShape()[3]; |
| 1717 | |
| 1718 | bool padding = false; |
| 1719 | TensorInfo outputInfo; |
| 1720 | if (paddingString == "SAME") |
| 1721 | { |
| 1722 | padding = true; |
| 1723 | outputInfo = TensorInfo({ inputTensorInfo.GetShape()[0], |
| 1724 | inputTensorInfo.GetShape()[1], |
| 1725 | static_cast<uint32_t>(ceil( |
| 1726 | static_cast<float>(inputHeight) / |
| 1727 | static_cast<float>(pooling2dDescriptor.m_StrideY))), |
| 1728 | static_cast<uint32_t>(ceil( |
| 1729 | static_cast<float>(inputWidth) / |
| 1730 | static_cast<float>(pooling2dDescriptor.m_StrideX))) |
| 1731 | }, DataType::Float32); |
| 1732 | } |
| 1733 | else if (paddingString == "VALID") |
| 1734 | { |
| 1735 | padding = false; |
| 1736 | outputInfo = TensorInfo({ inputTensorInfo.GetShape()[0], |
| 1737 | inputTensorInfo.GetShape()[1], |
| 1738 | static_cast<uint32_t>(ceil( |
| 1739 | static_cast<float>(inputHeight - pooling2dDescriptor.m_PoolHeight + 1) / |
| 1740 | static_cast<float>(pooling2dDescriptor.m_StrideY))), |
| 1741 | static_cast<uint32_t>(ceil( |
| 1742 | static_cast<float>(inputWidth - pooling2dDescriptor.m_PoolWidth + 1) / |
| 1743 | static_cast<float>(pooling2dDescriptor.m_StrideX))) |
| 1744 | }, DataType::Float32); |
| 1745 | } |
| 1746 | else |
| 1747 | { |
| 1748 | throw ParseException("Only 'SAME' and 'VALID' padding supported"); |
| 1749 | } |
| 1750 | |
| 1751 | CalcPadding(inputWidth, pooling2dDescriptor.m_PoolWidth, pooling2dDescriptor.m_StrideX, |
| 1752 | pooling2dDescriptor.m_PadLeft, pooling2dDescriptor.m_PadRight, padding); |
| 1753 | CalcPadding(inputHeight, pooling2dDescriptor.m_PoolHeight, pooling2dDescriptor.m_StrideY, |
| 1754 | pooling2dDescriptor.m_PadTop, pooling2dDescriptor.m_PadBottom, padding); |
| 1755 | |
| 1756 | |
| 1757 | IConnectableLayer* layer = m_Network->AddPooling2dLayer(pooling2dDescriptor, nodeDef.name().c_str()); |
| 1758 | if (layer == nullptr) |
| 1759 | { |
| 1760 | throw ParseException("Failed to add pooling2d layer"); |
| 1761 | } |
| 1762 | |
| 1763 | layer->GetOutputSlot(0).SetTensorInfo(outputInfo); |
| 1764 | |
| 1765 | if (dataFormat == "NHWC") |
| 1766 | { |
| 1767 | layer = SwizzleInDeswizzleOut(*m_Network, inputSlot, *layer, nodeDef.name()); |
| 1768 | } |
| 1769 | else |
| 1770 | { |
| 1771 | inputSlot.Connect(layer->GetInputSlot(0)); |
| 1772 | } |
| 1773 | |
| 1774 | return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer); |
| 1775 | } |
| 1776 | |
| 1777 | ParsedTfOperationPtr TfParser::AddAdditionLayer(const tensorflow::NodeDef& nodeDef, bool isBiasAdd) |
| 1778 | { |
| 1779 | std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 2); |
| 1780 | |
| 1781 | IOutputSlot* input0Slot = &inputs[0].m_IndexedValue->ResolveArmnnOutputSlot(inputs[0].m_Index); |
| 1782 | IOutputSlot* input1Slot = &inputs[1].m_IndexedValue->ResolveArmnnOutputSlot(inputs[1].m_Index); |
| 1783 | |
| 1784 | const TensorInfo& input0Info = input0Slot->GetTensorInfo(); |
| 1785 | const TensorInfo& input1Info = input1Slot->GetTensorInfo(); |
| 1786 | |
| 1787 | if (isBiasAdd) |
| 1788 | { |
| 1789 | // BiasAdd takes bias as a 1D tensor. We need to add a reshape layer to create a 4D tensor |
| 1790 | // with the same data in the correct dimension for broadcast in addition. |
| 1791 | if(input1Info.GetNumDimensions() != 1) |
| 1792 | { |
| 1793 | throw ParseException("Unsupported bias for BiasAdd. It should be a 1D vector."); |
| 1794 | } |
| 1795 | |
| 1796 | const std::string dataFormat = ReadMandatoryNodeStringAttribute(nodeDef, "data_format"); |
| 1797 | const bool isNHWC = (dataFormat == "NHWC"); |
| 1798 | const bool isNCHW = (dataFormat == "NCHW"); |
| 1799 | |
| 1800 | if (!isNHWC && ! isNCHW) |
| 1801 | { |
| 1802 | throw ParseException("Only NHWC or NCHW supported for BiasAdd"); |
| 1803 | } |
| 1804 | |
| 1805 | input1Slot = BroadcastForAddandMul(input0Slot, input1Slot, isNHWC, *m_Network, nodeDef); |
| 1806 | } |
| 1807 | else |
| 1808 | { |
| 1809 | if (input0Info.GetNumDimensions() == 1) |
| 1810 | { |
| 1811 | const bool isNHWC = true; |
| 1812 | input0Slot = BroadcastForAddandMul(input1Slot, input0Slot, isNHWC, *m_Network, nodeDef); |
| 1813 | } |
| 1814 | |
| 1815 | if (input1Info.GetNumDimensions() == 1) |
| 1816 | { |
| 1817 | const bool isNHWC = true; |
| 1818 | input1Slot = BroadcastForAddandMul(input0Slot, input1Slot, isNHWC, *m_Network, nodeDef); |
| 1819 | } |
| 1820 | } |
| 1821 | |
| 1822 | IConnectableLayer* const layer = m_Network->AddAdditionLayer(nodeDef.name().c_str()); |
| 1823 | |
| 1824 | input0Slot->Connect(layer->GetInputSlot(0)); |
| 1825 | input1Slot->Connect(layer->GetInputSlot(1)); |
| 1826 | |
| 1827 | if (input0Info.GetNumDimensions() == 1 && isBiasAdd == false) |
| 1828 | { |
| 1829 | layer->GetOutputSlot(0).SetTensorInfo(input1Slot->GetTensorInfo()); |
| 1830 | } |
| 1831 | else |
| 1832 | { |
| 1833 | layer->GetOutputSlot(0).SetTensorInfo(input0Slot->GetTensorInfo()); |
| 1834 | } |
| 1835 | |
| 1836 | return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer); |
| 1837 | } |
| 1838 | |
| 1839 | IConnectableLayer* TfParser::AddFullyConnectedLayer(const tensorflow::NodeDef& matMulNodeDef, |
| 1840 | const tensorflow::NodeDef* addNodeDef, const char* armnnLayerName) |
| 1841 | { |
| 1842 | // find bias const (if applicable) |
| 1843 | ParsedConstTfOperation<float>* biasNode = nullptr; |
| 1844 | if (addNodeDef != nullptr) |
| 1845 | { |
| 1846 | std::vector<OutputOfParsedTfOperation> addInputs = GetInputParsedTfOperationsChecked(*addNodeDef, 2); |
| 1847 | // find our inputs |
| 1848 | if (HasParsedConstTensor<float>(addInputs[0].m_IndexedValue->GetNode().name())) |
| 1849 | { |
| 1850 | biasNode = boost::polymorphic_downcast<ParsedConstTfOperation<float>*>(addInputs[0].m_IndexedValue); |
| 1851 | } |
| 1852 | else if (HasParsedConstTensor<float>(addInputs[1].m_IndexedValue->GetNode().name())) |
| 1853 | { |
| 1854 | biasNode = boost::polymorphic_downcast<ParsedConstTfOperation<float>*>(addInputs[1].m_IndexedValue); |
| 1855 | } |
| 1856 | else |
| 1857 | { |
| 1858 | throw ParseException("ArmNN only supports fully connected layers with constant bias"); |
| 1859 | } |
| 1860 | } |
| 1861 | |
| 1862 | // find matmul inputs |
| 1863 | ParsedConstTfOperation<float>* weightNode = nullptr; |
| 1864 | ParsedTfOperation* inputNode = nullptr; |
| 1865 | unsigned int inputIdx = 0; |
| 1866 | std::vector<OutputOfParsedTfOperation> mulInputs = GetInputParsedTfOperationsChecked(matMulNodeDef, 2); |
| 1867 | if (HasParsedConstTensor<float>(mulInputs[0].m_IndexedValue->GetNode().name())) |
| 1868 | { |
| 1869 | weightNode = boost::polymorphic_downcast<ParsedConstTfOperation<float>*>(mulInputs[0].m_IndexedValue); |
| 1870 | inputNode = mulInputs[1].m_IndexedValue; |
| 1871 | inputIdx = mulInputs[1].m_Index; |
| 1872 | } |
| 1873 | else if (HasParsedConstTensor<float>(mulInputs[1].m_IndexedValue->GetNode().name())) |
| 1874 | { |
| 1875 | weightNode = boost::polymorphic_downcast<ParsedConstTfOperation<float>*>(mulInputs[1].m_IndexedValue); |
| 1876 | inputNode = mulInputs[0].m_IndexedValue; |
| 1877 | inputIdx = mulInputs[0].m_Index; |
| 1878 | } |
| 1879 | else |
| 1880 | { |
| 1881 | throw ParseException("ArmNN only supports fully connected layers with constant weights"); |
| 1882 | } |
| 1883 | |
| 1884 | std::vector<float> weightTensorData; |
| 1885 | // handle weight |
| 1886 | ConstTensor weights = weightNode->GetConstTensor(false, weightTensorData); |
| 1887 | |
| 1888 | FullyConnectedDescriptor desc; |
| 1889 | desc.m_BiasEnabled = addNodeDef != nullptr; |
| 1890 | |
| 1891 | IConnectableLayer* layer = nullptr; |
| 1892 | // make the layer |
| 1893 | if (addNodeDef != nullptr) |
| 1894 | { |
| 1895 | std::vector<float> biasTensorData; |
| 1896 | ConstTensor biases = biasNode->GetConstTensor(false, biasTensorData); |
| 1897 | |
| 1898 | if (weights.GetShape()[1] != biases.GetShape()[0]) |
| 1899 | { |
| 1900 | throw ParseException("shape of matmul and bias do not match"); |
| 1901 | } |
| 1902 | |
| 1903 | layer = m_Network->AddFullyConnectedLayer(desc, weights, biases, armnnLayerName); |
| 1904 | } |
| 1905 | else |
| 1906 | { |
| 1907 | layer = m_Network->AddFullyConnectedLayer(desc, weights, armnnLayerName); |
| 1908 | } |
| 1909 | |
| 1910 | BOOST_ASSERT(layer != nullptr); |
| 1911 | |
| 1912 | inputNode->ResolveArmnnOutputSlot(inputIdx).Connect(layer->GetInputSlot(0)); |
| 1913 | unsigned int batches = inputNode->ResolveArmnnOutputSlot(inputIdx).GetTensorInfo().GetShape()[0]; |
| 1914 | |
| 1915 | // handle output |
| 1916 | TensorInfo outputInfo({ batches, weights.GetShape()[1] }, DataType::Float32); |
| 1917 | layer->GetOutputSlot(0).SetTensorInfo(outputInfo); |
| 1918 | return layer; |
| 1919 | } |
| 1920 | |
| 1921 | void TfParser::LoadNodeDef(const tensorflow::NodeDef& nodeDef, const tensorflow::GraphDef& graphDef) |
| 1922 | { |
| 1923 | // get the type of the node (assume float) |
| 1924 | tensorflow::DataType type = tensorflow::DT_FLOAT; |
| 1925 | if (nodeDef.attr().count("T") != 0) |
| 1926 | { |
| 1927 | auto attr = nodeDef.attr().at("T"); |
| 1928 | type = attr.type(); |
| 1929 | } |
| 1930 | else if (nodeDef.attr().count("dtype") != 0) |
| 1931 | { |
| 1932 | auto attr = nodeDef.attr().at("dtype"); |
| 1933 | type = attr.type(); |
| 1934 | } |
| 1935 | |
| 1936 | if (type != tensorflow::DT_FLOAT && nodeDef.op() != "Const") |
| 1937 | { |
| 1938 | throw ParseException("Currently only FLOAT is supported for tensorflow nodes (apart from Const)"); |
| 1939 | } |
| 1940 | |
| 1941 | const std::string& operation = nodeDef.op(); |
| 1942 | auto it = ms_OperationNameToParsingFunctions.find(operation); |
| 1943 | if (it != ms_OperationNameToParsingFunctions.end()) |
| 1944 | { |
| 1945 | auto func = it->second; |
| 1946 | ParsedTfOperationPtr parsedTfOperation = (this->*func)(nodeDef, graphDef); |
| 1947 | ParsedTfOperation* parsedTfOperationRaw = parsedTfOperation.get(); |
| 1948 | |
| 1949 | // Store the parsed operation so that dependent layers can connect to it |
| 1950 | auto it = m_ParsedTfOperations.find(nodeDef.name()); |
| 1951 | if (it != m_ParsedTfOperations.end()) |
| 1952 | { |
| 1953 | throw ParseException(boost::str(boost::format("Name %1% used by more than one node") % nodeDef.name())); |
| 1954 | } |
| 1955 | m_ParsedTfOperations[nodeDef.name()] = std::move(parsedTfOperation); |
| 1956 | |
| 1957 | // If this node was requested as an output from the network then add an ArmNN output layer |
| 1958 | if (std::find(m_RequestedOutputs.begin(), m_RequestedOutputs.end(), nodeDef.name()) != |
| 1959 | m_RequestedOutputs.end()) |
| 1960 | { |
| 1961 | auto outId = ParseOutputId(nodeDef.name()); |
| 1962 | const LayerBindingId layerId = boost::numeric_cast<LayerBindingId>(m_NetworkOutputsBindingInfo.size()); |
| 1963 | IOutputSlot& prevSlot = parsedTfOperationRaw->ResolveArmnnOutputSlot(outId.m_Index); |
| 1964 | |
| 1965 | TensorInfo tensorInfo = prevSlot.GetTensorInfo(); |
| 1966 | |
| 1967 | IConnectableLayer* outputLayer = m_Network->AddOutputLayer(layerId, nodeDef.name().c_str()); |
| 1968 | |
| 1969 | prevSlot.Connect(outputLayer->GetInputSlot(0)); |
| 1970 | |
| 1971 | TrackOutputBinding(outputLayer, layerId, tensorInfo); |
| 1972 | } |
| 1973 | } |
| 1974 | else |
| 1975 | { |
| 1976 | throw ParseException(boost::str( |
| 1977 | boost::format("Unsupported operation %1% in tensorflow::GraphDef") % operation)); |
| 1978 | } |
| 1979 | } |
| 1980 | |
| 1981 | void TfParser::LoadGraphDef(const tensorflow::GraphDef& graphDef) |
| 1982 | { |
| 1983 | // add all nodes to our map |
| 1984 | m_NodesByName.clear(); |
| 1985 | m_NetworkInputsBindingInfo.clear(); |
| 1986 | m_NetworkOutputsBindingInfo.clear(); |
| 1987 | |
| 1988 | for (int i = 0; i < graphDef.node_size(); ++i) |
| 1989 | { |
| 1990 | const tensorflow::NodeDef& node = graphDef.node(i); |
| 1991 | m_NodesByName[node.name()] = &node; |
| 1992 | } |
| 1993 | |
| 1994 | // Find the output nodes the user requested |
| 1995 | std::vector<const tensorflow::NodeDef*> targetNodes; |
| 1996 | for (const std::string& requestedOutputName : m_RequestedOutputs) |
| 1997 | { |
| 1998 | auto nodeIt = m_NodesByName.find(requestedOutputName); |
| 1999 | if (nodeIt == m_NodesByName.end()) |
| 2000 | { |
| 2001 | throw ParseException("Couldn't find requested output node '" + requestedOutputName + "' in graph"); |
| 2002 | } |
| 2003 | targetNodes.push_back(nodeIt->second); |
| 2004 | } |
| 2005 | |
| 2006 | // Sort them into a linear ordering such that all inputs of a node are before the node itself |
| 2007 | std::vector<const tensorflow::NodeDef*> sortedNodes; |
| 2008 | if (!armnnUtils::GraphTopologicalSort<const tensorflow::NodeDef*>( |
| 2009 | targetNodes, |
| 2010 | [this](const tensorflow::NodeDef* node) |
| 2011 | { |
| 2012 | auto outputs = GetTfInputNodes(*node); |
| 2013 | std::vector<const tensorflow::NodeDef*> nodesOnly; |
| 2014 | for (const auto & o : outputs) { |
| 2015 | nodesOnly.push_back(o.m_IndexedValue); |
| 2016 | } |
| 2017 | return nodesOnly; |
| 2018 | }, |
| 2019 | sortedNodes)) |
| 2020 | { |
| 2021 | throw ParseException("Cycle detected in graph"); |
| 2022 | } |
| 2023 | |
| 2024 | // Parse each node in order, knowing that all inputs of a node will be processed before the node itself |
| 2025 | for (const auto& it : sortedNodes) |
| 2026 | { |
| 2027 | const tensorflow::NodeDef& currentNode = *it; |
| 2028 | LoadNodeDef(currentNode, graphDef); |
| 2029 | } |
| 2030 | } |
| 2031 | |
| 2032 | INetworkPtr TfParser::CreateNetworkFromTextFile(const char* graphFile, |
| 2033 | const std::map<std::string, TensorShape>& inputShapes, |
| 2034 | const std::vector<std::string>& requestedOutputs) |
| 2035 | { |
| 2036 | FILE* fd = fopen(graphFile, "r"); |
| 2037 | |
| 2038 | if (fd == nullptr) |
| 2039 | { |
| 2040 | std::stringstream error; |
| 2041 | error << "Graph file " << graphFile << " failed to open"; |
| 2042 | throw FileNotFoundException(error.str()); |
| 2043 | } |
| 2044 | |
| 2045 | // Parse the file into a message |
| 2046 | tensorflow::GraphDef graphDef; |
| 2047 | auto input = new google::protobuf::io::FileInputStream(fileno(fd)); |
| 2048 | bool success = google::protobuf::TextFormat::Parse(input, &graphDef); |
| 2049 | delete input; |
| 2050 | fclose(fd); |
| 2051 | |
| 2052 | if (!success) |
| 2053 | { |
| 2054 | std::stringstream error; |
| 2055 | error << "Failed to parse graph file"; |
| 2056 | throw ParseException(error.str()); |
| 2057 | } |
| 2058 | |
| 2059 | return CreateNetworkFromGraphDef(graphDef, inputShapes, requestedOutputs); |
| 2060 | } |
| 2061 | |
| 2062 | INetworkPtr TfParser::CreateNetworkFromString(const char* protoText, |
| 2063 | const std::map<std::string, TensorShape>& inputShapes, |
| 2064 | const std::vector<std::string>& requestedOutputs) |
| 2065 | { |
| 2066 | // Parse the string into a message |
| 2067 | tensorflow::GraphDef graphDef; |
| 2068 | bool success = google::protobuf::TextFormat::ParseFromString(protoText, &graphDef); |
| 2069 | |
| 2070 | if (!success) |
| 2071 | { |
| 2072 | std::stringstream error; |
| 2073 | error << "Failed to parse graph file"; |
| 2074 | throw ParseException(error.str()); |
| 2075 | } |
| 2076 | |
| 2077 | return CreateNetworkFromGraphDef(graphDef, inputShapes, requestedOutputs); |
| 2078 | } |
| 2079 | |
| 2080 | INetworkPtr TfParser::CreateNetworkFromBinaryFile(const char* graphFile, |
| 2081 | const std::map<std::string, TensorShape>& inputShapes, |
| 2082 | const std::vector<std::string>& requestedOutputs) |
| 2083 | { |
| 2084 | FILE* fd = fopen(graphFile, "rb"); |
| 2085 | |
| 2086 | if (fd == nullptr) |
| 2087 | { |
| 2088 | std::stringstream error; |
| 2089 | error << "Graph file " << graphFile << " failed to open"; |
| 2090 | throw FileNotFoundException(error.str()); |
| 2091 | } |
| 2092 | |
| 2093 | // Parse the file into a message |
| 2094 | tensorflow::GraphDef graphDef; |
| 2095 | |
| 2096 | google::protobuf::io::FileInputStream inStream(fileno(fd)); |
| 2097 | google::protobuf::io::CodedInputStream codedStream(&inStream); |
| 2098 | codedStream.SetTotalBytesLimit(INT_MAX, INT_MAX); |
| 2099 | bool success = graphDef.ParseFromCodedStream(&codedStream); |
| 2100 | fclose(fd); |
| 2101 | |
| 2102 | if (!success) |
| 2103 | { |
| 2104 | std::stringstream error; |
| 2105 | error << "Failed to parse protobuf file" << graphFile; |
| 2106 | throw ParseException(error.str()); |
| 2107 | } |
| 2108 | |
| 2109 | return CreateNetworkFromGraphDef(graphDef, inputShapes, requestedOutputs); |
| 2110 | } |
| 2111 | |
| 2112 | INetworkPtr TfParser::CreateNetworkFromGraphDef(const tensorflow::GraphDef& graphDef, |
| 2113 | const std::map<std::string, TensorShape>& inputShapes, |
| 2114 | const std::vector<std::string>& requestedOutputs) |
| 2115 | { |
| 2116 | m_Network = INetwork::Create(); |
| 2117 | |
| 2118 | m_InputShapes = inputShapes; |
| 2119 | if (requestedOutputs.size() == 0) |
| 2120 | { |
| 2121 | throw ParseException("requestedOutputs must have at least one entry"); |
| 2122 | } |
| 2123 | m_RequestedOutputs = requestedOutputs; |
| 2124 | |
| 2125 | try |
| 2126 | { |
| 2127 | LoadGraphDef(graphDef); |
| 2128 | } |
| 2129 | catch (const ParseException& e) |
| 2130 | { |
| 2131 | Cleanup(); |
| 2132 | throw e; |
| 2133 | } |
| 2134 | |
| 2135 | Cleanup(); |
| 2136 | |
| 2137 | return std::move(m_Network); |
| 2138 | } |
| 2139 | |
| 2140 | void TfParser::Cleanup() |
| 2141 | { |
| 2142 | // cleanup, in case we reuse this parser |
| 2143 | m_InputShapes.clear(); |
| 2144 | m_RequestedOutputs.clear(); |
| 2145 | m_NodesByName.clear(); |
| 2146 | m_ParsedTfOperations.clear(); |
| 2147 | } |
| 2148 | |
| 2149 | BindingPointInfo TfParser::GetNetworkInputBindingInfo(const std::string& name) const |
| 2150 | { |
| 2151 | return GetBindingInfo(name, "input", m_NetworkInputsBindingInfo); |
| 2152 | } |
| 2153 | |
| 2154 | BindingPointInfo TfParser::GetNetworkOutputBindingInfo(const std::string& name) const |
| 2155 | { |
| 2156 | return GetBindingInfo(name, "output", m_NetworkOutputsBindingInfo); |
| 2157 | } |
| 2158 | |
| 2159 | std::pair<LayerBindingId, TensorInfo> TfParser::GetBindingInfo(const std::string& layerName, |
| 2160 | const char* bindingPointDesc, |
| 2161 | const std::unordered_map<std::string, BindingPointInfo>& nameToBindingInfo) |
| 2162 | { |
| 2163 | auto it = nameToBindingInfo.find(layerName); |
| 2164 | if (it == nameToBindingInfo.end()) |
| 2165 | { |
| 2166 | throw InvalidArgumentException(boost::str(boost::format("Unknown %1% '%2%'") % bindingPointDesc % layerName)); |
| 2167 | } |
| 2168 | return it->second; |
| 2169 | } |
| 2170 | |
| 2171 | void TfParser::TrackInputBinding(IConnectableLayer* layer, LayerBindingId id, const TensorInfo& tensorInfo) |
| 2172 | { |
| 2173 | return TrackBindingPoint(layer, id, tensorInfo, "input", m_NetworkInputsBindingInfo); |
| 2174 | } |
| 2175 | |
| 2176 | void TfParser::TrackOutputBinding(IConnectableLayer* layer, LayerBindingId id, const TensorInfo& tensorInfo) |
| 2177 | { |
| 2178 | return TrackBindingPoint(layer, id, tensorInfo, "output", m_NetworkOutputsBindingInfo); |
| 2179 | } |
| 2180 | |
| 2181 | void TfParser::TrackBindingPoint(IConnectableLayer* layer, |
| 2182 | LayerBindingId id, |
| 2183 | const TensorInfo& tensorInfo, |
| 2184 | const char* bindingPointDesc, |
| 2185 | std::unordered_map<std::string, BindingPointInfo>& nameToBindingInfo) |
| 2186 | { |
| 2187 | const std::string layerName = layer->GetName(); |
| 2188 | auto it = nameToBindingInfo.find(layerName); |
| 2189 | if (it == nameToBindingInfo.end()) |
| 2190 | { |
| 2191 | nameToBindingInfo[layerName] = std::make_pair(id, tensorInfo); |
| 2192 | } |
| 2193 | else |
| 2194 | { |
| 2195 | throw ParseException(boost::str( |
| 2196 | boost::format("Id %1% used by more than one %2% layer") % id % bindingPointDesc)); |
| 2197 | } |
| 2198 | } |
| 2199 | |
| 2200 | } // namespace armnnTfParser |