telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2017 Arm Ltd. All rights reserved. |
David Beck | ecb56cd | 2018-09-05 12:52:57 +0100 | [diff] [blame] | 3 | // SPDX-License-Identifier: MIT |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 4 | // |
| 5 | #include "OnnxParser.hpp" |
| 6 | |
| 7 | #include <armnn/ArmNN.hpp> |
| 8 | #include <armnn/Utils.hpp> |
| 9 | #include <VerificationHelpers.hpp> |
| 10 | |
Aron Virginas-Tar | d4f0fea | 2019-04-09 14:08:06 +0100 | [diff] [blame] | 11 | #include <boost/format.hpp> |
| 12 | #include <boost/numeric/conversion/cast.hpp> |
| 13 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 14 | #include <google/protobuf/text_format.h> |
| 15 | #include <google/protobuf/io/zero_copy_stream_impl.h> |
| 16 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 17 | #include <numeric> |
| 18 | |
| 19 | using namespace armnn; |
| 20 | |
| 21 | namespace armnnOnnxParser |
| 22 | { |
| 23 | namespace |
| 24 | { |
| 25 | void CheckValidDataType(std::initializer_list<onnx::TensorProto::DataType> validInputTypes, |
| 26 | const onnx::TensorProto::DataType actualValue, |
| 27 | const char* validExpr, |
| 28 | std::string nodeName, |
| 29 | std::string tensorName, |
| 30 | const armnn::CheckLocation& location) |
| 31 | { |
| 32 | bool isValid = std::any_of(validInputTypes.begin(), |
| 33 | validInputTypes.end(), |
| 34 | [&actualValue](onnx::TensorProto::DataType x) { return x == actualValue; } ); |
| 35 | if (!isValid) |
| 36 | { |
| 37 | throw ParseException( |
| 38 | boost::str( |
| 39 | boost::format("Datatype %1% is not valid for tensor '%2%' of node '%3%', not in {%4%}. %5%") % |
| 40 | onnx::TensorProto::DataType_Name(actualValue) % |
| 41 | tensorName % |
| 42 | nodeName % |
| 43 | validExpr % |
| 44 | location.AsString())); |
| 45 | } |
| 46 | } |
| 47 | |
| 48 | #define CHECK_VALID_DATATYPE(NODE, TENSOR, ACTUAL, ...) \ |
| 49 | CheckValidDataType({__VA_ARGS__}, ACTUAL, #__VA_ARGS__, NODE, TENSOR, CHECK_LOCATION()) |
| 50 | |
| 51 | using StrTypeListPair = std::pair<const char*, std::initializer_list<onnx::TensorProto::DataType>>; |
| 52 | #define STR_LIST(...) StrTypeListPair(#__VA_ARGS__, {__VA_ARGS__}) |
| 53 | |
| 54 | template <typename Callable> |
| 55 | void ReadMandatoryNodeAttributeImpl(const onnx::NodeProto& node, |
| 56 | const std::string& attribName, |
| 57 | onnx::AttributeProto::AttributeType expectedType, |
| 58 | Callable callable) |
| 59 | { |
| 60 | auto attribs = node.attribute(); |
| 61 | int attriNum = 0; |
| 62 | while (attriNum < node.attribute_size()) |
| 63 | { |
| 64 | if (attribs.Get(attriNum).name() == attribName) |
| 65 | { |
| 66 | if (attribs.Get(attriNum).type() == expectedType) |
| 67 | { |
| 68 | callable(attribs.Get(attriNum)); |
| 69 | } |
| 70 | else |
| 71 | { |
| 72 | throw ParseException(boost::str(boost::format( |
| 73 | "Attribute %1% of node %2% expected to have %3% as onnx::AttributeProto::AttributeType, " |
| 74 | "but found %4% instead %5%") |
| 75 | % attribName |
| 76 | % node.name() |
| 77 | % onnx::AttributeProto::AttributeType_Name(expectedType) |
| 78 | % onnx::AttributeProto::AttributeType_Name(attribs.Get(attriNum).type()) |
| 79 | % CHECK_LOCATION().AsString())); |
| 80 | } |
| 81 | break; |
| 82 | } |
| 83 | ++attriNum; |
| 84 | } |
| 85 | if (attriNum == node.attribute_size()) |
| 86 | { |
| 87 | throw ParseException(boost::str(boost::format("Could not find required attribute %1% in node %2% %3%") |
| 88 | % attribName % node.name() % CHECK_LOCATION().AsString())); |
| 89 | } |
| 90 | } |
| 91 | |
| 92 | template <typename Callable> |
| 93 | void ReadOptionalNodeAttributeImpl(const onnx::NodeProto& node, |
| 94 | const std::string& attribName, |
| 95 | onnx::AttributeProto::AttributeType expectedType, |
| 96 | Callable callable) |
| 97 | { |
| 98 | auto attribs = node.attribute(); |
| 99 | for (int attriNum = 0; attriNum < node.attribute_size(); ++attriNum) |
| 100 | { |
| 101 | if (attribs.Get(attriNum).name() == attribName) |
| 102 | { |
| 103 | if (attribs.Get(attriNum).type() == expectedType) |
| 104 | { |
| 105 | callable(attribs.Get(attriNum)); |
| 106 | } |
| 107 | else |
| 108 | { |
| 109 | throw ParseException(boost::str(boost::format( |
| 110 | "Attribute %1% of node %2% expected to have %3% as onnx::AttributeProto::AttributeType, " |
| 111 | "but found %4% instead %5%") |
| 112 | % attribName |
| 113 | % node.name() |
| 114 | % onnx::AttributeProto::AttributeType_Name(expectedType) |
| 115 | % onnx::AttributeProto::AttributeType_Name(attribs.Get(attriNum).type()) |
| 116 | % CHECK_LOCATION().AsString())); |
| 117 | } |
| 118 | } |
| 119 | } |
| 120 | } |
| 121 | |
| 122 | std::vector<uint32_t> ReadMandatoryNodeUint32ListAttribute(const onnx::NodeProto& node, |
| 123 | const std::string& name) |
| 124 | { |
| 125 | std::vector<uint32_t> attriList; |
| 126 | ReadMandatoryNodeAttributeImpl(node, name, onnx::AttributeProto::INTS, |
| 127 | [&attriList](const onnx::AttributeProto& attrValue) |
| 128 | { |
| 129 | for (int attriNum = 0; attriNum < attrValue.ints_size(); ++attriNum) |
| 130 | { |
| 131 | attriList.push_back(CHECKED_NON_NEGATIVE(CHECKED_INT32(attrValue.ints().Get(attriNum)))); |
| 132 | } |
| 133 | }); |
| 134 | return attriList; |
| 135 | } |
| 136 | |
| 137 | uint32_t ReadOptionalNodeUint32Attribute(const onnx::NodeProto& node, |
| 138 | const std::string& name, |
| 139 | const uint32_t defaultVal = 0u) |
| 140 | { |
| 141 | uint32_t attribValue = defaultVal; |
| 142 | ReadOptionalNodeAttributeImpl(node, name, onnx::AttributeProto::INT, |
| 143 | [&attribValue](const onnx::AttributeProto& attrValue) |
| 144 | { |
| 145 | attribValue = CHECKED_NON_NEGATIVE(CHECKED_INT32((attrValue.i()))); |
| 146 | }); |
| 147 | return attribValue; |
| 148 | } |
| 149 | |
| 150 | std::vector<uint32_t> ReadOptionalNodeUint32ListAttribute(const onnx::NodeProto& node, |
| 151 | const std::string& name) |
| 152 | { |
| 153 | std::vector<uint32_t> attriList; |
| 154 | ReadOptionalNodeAttributeImpl(node, name, onnx::AttributeProto::INTS, |
| 155 | [&attriList](const onnx::AttributeProto& attrValue) |
| 156 | { |
| 157 | for (int attriNum = 0; attriNum < attrValue.ints_size(); ++attriNum) |
| 158 | { |
| 159 | attriList.push_back(CHECKED_NON_NEGATIVE(CHECKED_INT32(attrValue.ints().Get(attriNum)))); |
| 160 | } |
| 161 | }); |
| 162 | |
| 163 | return attriList; |
| 164 | } |
| 165 | |
| 166 | float ReadOptionalNodeFloatAttribute(const onnx::NodeProto& node, |
| 167 | const std::string& name, |
| 168 | const float defaultValue = 0.0f) |
| 169 | { |
| 170 | float attribValue = defaultValue; |
| 171 | ReadOptionalNodeAttributeImpl(node, name, onnx::AttributeProto::FLOAT, |
| 172 | [&attribValue](const onnx::AttributeProto& attrValue) |
| 173 | { |
| 174 | attribValue = attrValue.f(); |
| 175 | }); |
| 176 | return attribValue; |
| 177 | } |
| 178 | |
| 179 | std::string ReadOptionalNodeStringAttribute(const onnx::NodeProto& node, const std::string& name) |
| 180 | { |
| 181 | std::string attribValue = ""; |
| 182 | ReadOptionalNodeAttributeImpl(node, name, onnx::AttributeProto::STRING, |
| 183 | [&attribValue](const onnx::AttributeProto& attrValue) |
| 184 | { |
| 185 | attribValue = attrValue.s(); |
| 186 | }); |
| 187 | return attribValue; |
| 188 | } |
| 189 | |
| 190 | armnn::TensorInfo ToTensorInfo(const onnx::ValueInfoProto& info) |
| 191 | { |
| 192 | const onnx::TensorShapeProto onnxShape = info.type().tensor_type().shape(); |
| 193 | std::vector<unsigned int> shapeDims; |
| 194 | for (int i = 0; i < onnxShape.dim_size(); ++i) |
| 195 | { |
| 196 | shapeDims.push_back(CHECKED_NON_NEGATIVE(CHECKED_INT32(onnxShape.dim(i).dim_value()))); |
| 197 | } |
| 198 | DataType type; |
| 199 | switch(info.type().tensor_type().elem_type()) |
| 200 | { |
| 201 | case onnx::TensorProto::FLOAT: |
| 202 | { |
| 203 | type = DataType::Float32; |
| 204 | break; |
| 205 | } |
| 206 | case onnx::TensorProto::INT32: |
| 207 | case onnx::TensorProto::INT64: |
| 208 | { |
| 209 | type = DataType::Signed32; |
| 210 | break; |
| 211 | } |
| 212 | default: |
| 213 | { |
| 214 | throw ParseException( |
| 215 | boost::str( |
| 216 | boost::format("'%1%' is not a currently supported datatype for tensor %2%." |
| 217 | " Supported dataTypes are FLOAT, INT32 and INT64. %3%") % |
Matteo Martincigh | e355dc2 | 2018-12-10 13:45:27 +0000 | [diff] [blame] | 218 | onnx::TensorProto::DataType_Name( |
| 219 | static_cast<onnx::TensorProto::DataType>(info.type().tensor_type().elem_type())) % |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 220 | info.name() % |
| 221 | CHECK_LOCATION().AsString() )); |
| 222 | } |
| 223 | |
| 224 | } |
| 225 | return TensorInfo(TensorShape(static_cast<unsigned int>(shapeDims.size()), shapeDims.data()), type); |
| 226 | } |
| 227 | |
| 228 | std::string TensorInfoAsString(const TensorInfo& info, |
| 229 | const std::string& name, |
| 230 | const onnx::TensorProto::DataType& type) |
| 231 | { |
| 232 | const TensorShape shape = info.GetShape(); |
| 233 | std::stringstream ss; |
| 234 | ss << "tensor '" << name << "' contains " |
| 235 | << onnx::TensorProto::DataType_Name(type) |
| 236 | << " and has shape ["; |
| 237 | |
| 238 | for (uint32_t i = 0; i < shape.GetNumDimensions() - 1; ++i) |
| 239 | { |
| 240 | ss << shape[i] << ", "; |
| 241 | } |
| 242 | ss << shape[shape.GetNumDimensions() - 1] << "]"; |
| 243 | return ss.str(); |
| 244 | } |
| 245 | |
| 246 | void CalcPadding(uint32_t inputSize, uint32_t filterSize, uint32_t stride, uint32_t* paddingFront, |
| 247 | uint32_t* paddingBack, bool isUpper) |
| 248 | { |
| 249 | uint32_t outputSize = (inputSize + stride - 1) / stride; |
| 250 | uint32_t temp = (outputSize - 1) * stride + filterSize; |
| 251 | *paddingFront = (temp - inputSize) / 2; |
| 252 | *paddingBack = *paddingFront; |
| 253 | if((temp - inputSize) % 2 == 1) |
| 254 | { |
| 255 | if (isUpper) |
| 256 | { |
| 257 | *paddingBack += 1; |
| 258 | } |
| 259 | else |
| 260 | { |
| 261 | *paddingFront += 1; |
| 262 | } |
| 263 | } |
| 264 | } |
| 265 | |
| 266 | TensorInfo ComputeReshapeInfo(const onnx::TensorProto& targetShapeTensor, |
| 267 | const TensorShape& inShape, |
| 268 | const std::string& outName) |
| 269 | { |
| 270 | std::vector<int> targetDims; |
| 271 | for(int i = 0; i < targetShapeTensor.int64_data_size(); ++i) |
| 272 | { |
| 273 | int val = CHECKED_INT32(targetShapeTensor.int64_data(i)); |
| 274 | if(val == 0) |
| 275 | { |
| 276 | targetDims.push_back(static_cast<int>(inShape[static_cast<uint>(i)])); |
| 277 | } |
| 278 | else |
| 279 | { |
| 280 | targetDims.push_back(val); |
| 281 | } |
| 282 | } |
| 283 | |
| 284 | std::vector<unsigned int> outDims(targetDims.begin(), targetDims.end()); |
| 285 | const auto stretchDim = std::find(targetDims.begin(), targetDims.end(), -1); |
| 286 | if (stretchDim != targetDims.end()) |
| 287 | { |
| 288 | if (std::find(std::next(stretchDim), targetDims.end(), -1) != targetDims.end()) |
| 289 | { |
| 290 | std::stringstream ss; |
| 291 | ss << "[ "; |
| 292 | for(uint i = 0; i < targetDims.size() - 1; ++i) |
| 293 | { |
| 294 | ss << targetDims[i] << ", "; |
| 295 | } |
| 296 | ss << targetDims[targetDims.size() - 1] << " ]"; |
| 297 | |
| 298 | throw ParseException(boost::str( |
| 299 | boost::format("Error during creation of reshaped tensor '%1%'. At most one component of shape can be " |
| 300 | " -1 and here, shape is %2% %3%") |
| 301 | % outName |
| 302 | % ss.str() |
| 303 | % CHECK_LOCATION().AsString())); |
| 304 | } |
| 305 | |
| 306 | auto targetNumElements = boost::numeric_cast<unsigned int>(std::accumulate(targetDims.begin(), targetDims.end(), |
| 307 | -1, std::multiplies<int32_t>())); |
| 308 | auto stretchIndex = static_cast<size_t>(std::distance(targetDims.begin(), stretchDim)); |
| 309 | outDims[stretchIndex] = inShape.GetNumElements() / targetNumElements; |
| 310 | } |
| 311 | TensorShape outShape = TensorShape{static_cast<unsigned int>(outDims.size()), outDims.data()}; |
| 312 | return TensorInfo(outShape, DataType::Float32); |
| 313 | } |
| 314 | |
| 315 | } //namespace |
| 316 | |
| 317 | const std::map<std::string, OnnxParser::OperationParsingFunction> OnnxParser::m_ParserFunctions = { |
| 318 | { "BatchNormalization", &OnnxParser::ParseBatchNormalization}, |
| 319 | { "GlobalAveragePool", &OnnxParser::ParseGlobalAveragePool}, |
| 320 | { "AveragePool", &OnnxParser::ParseAveragePool }, |
| 321 | { "Constant", &OnnxParser::ParseConstant }, |
| 322 | { "MaxPool", &OnnxParser::ParseMaxPool }, |
| 323 | { "Reshape", &OnnxParser::ParseReshape }, |
| 324 | { "Relu", &OnnxParser::ParseRelu }, |
| 325 | { "Conv", &OnnxParser::ParseConv }, |
| 326 | { "Add", &OnnxParser::ParseAdd }, |
| 327 | }; |
| 328 | |
| 329 | template<typename TypePair, typename Location> |
| 330 | void OnnxParser::ValidateInputs(const onnx::NodeProto& node, |
| 331 | TypePair validInputs, |
| 332 | const Location& location) |
| 333 | { |
| 334 | for(auto input : node.input()) |
| 335 | { |
| 336 | CheckValidDataType(validInputs.second, |
| 337 | m_TensorsInfo[input].m_dtype, |
| 338 | validInputs.first, |
| 339 | node.name(), |
| 340 | input, |
| 341 | location); |
| 342 | } |
| 343 | } |
| 344 | |
| 345 | #define VALID_INPUTS(NODE, VALID_INPUTS) \ |
| 346 | OnnxParser::ValidateInputs(NODE, \ |
| 347 | VALID_INPUTS, \ |
| 348 | CHECK_LOCATION()) |
| 349 | |
| 350 | std::vector<TensorInfo> OnnxParser::ComputeOutputInfo(std::vector<std::string> outNames, |
| 351 | const IConnectableLayer* layer, |
| 352 | std::vector<TensorShape> inputShapes) |
| 353 | { |
| 354 | BOOST_ASSERT(! outNames.empty()); |
| 355 | bool needCompute = std::any_of(outNames.begin(), |
| 356 | outNames.end(), |
| 357 | [this](std::string name) |
| 358 | { |
| 359 | return (m_TensorsInfo.count(name) == 0 || m_TensorsInfo[name].m_info == nullptr); |
| 360 | }); |
| 361 | std::vector<TensorInfo> outInfo; |
| 362 | //if the output info(s) are not here, we need to compute them |
| 363 | std::vector<TensorShape> inferredShapes; |
| 364 | if(needCompute) |
| 365 | { |
| 366 | inferredShapes = layer->InferOutputShapes(inputShapes); |
| 367 | BOOST_ASSERT(inferredShapes.size() == outNames.size()); |
| 368 | } |
| 369 | for (uint i = 0; i < outNames.size(); ++i) |
| 370 | { |
| 371 | if(needCompute) |
| 372 | { |
| 373 | m_TensorsInfo[outNames[i]] = OnnxTensor(); |
| 374 | m_TensorsInfo[outNames[i]].m_info = std::make_unique<TensorInfo>( |
| 375 | TensorInfo(inferredShapes[i], DataType::Float32)); |
| 376 | } |
| 377 | outInfo.push_back(*m_TensorsInfo[outNames[i]].m_info); |
| 378 | } |
| 379 | return outInfo; |
| 380 | } |
| 381 | |
| 382 | IOnnxParser* IOnnxParser::CreateRaw() |
| 383 | { |
| 384 | return new OnnxParser(); |
| 385 | } |
| 386 | |
| 387 | IOnnxParserPtr IOnnxParser::Create() |
| 388 | { |
| 389 | return IOnnxParserPtr(CreateRaw(), &IOnnxParser::Destroy); |
| 390 | } |
| 391 | |
| 392 | void IOnnxParser::Destroy(IOnnxParser* parser) |
| 393 | { |
| 394 | delete parser; |
| 395 | } |
| 396 | |
| 397 | OnnxParser::OnnxParser() |
| 398 | : m_Network(nullptr, nullptr) |
| 399 | { |
| 400 | } |
| 401 | |
| 402 | void OnnxParser::ResetParser() |
| 403 | { |
| 404 | m_Network = armnn::INetworkPtr(nullptr, nullptr); |
| 405 | m_Graph = nullptr; |
| 406 | } |
| 407 | |
| 408 | void OnnxParser::Cleanup() |
| 409 | { |
| 410 | m_TensorConnections.clear(); |
| 411 | m_TensorsInfo.clear(); |
| 412 | m_OutputsMap.clear(); |
| 413 | m_OutputsFusedAndUsed.clear(); |
| 414 | } |
| 415 | |
| 416 | std::pair<ConstTensor, std::unique_ptr<float[]>> OnnxParser::CreateConstTensor(const std::string name) |
| 417 | { |
| 418 | const TensorInfo tensorInfo = *m_TensorsInfo[name].m_info; |
| 419 | onnx::TensorProto onnxTensor = *m_TensorsInfo[name].m_tensor; |
| 420 | |
| 421 | auto srcData = onnxTensor.float_data().data(); |
Pablo Tello | 3dcc1c6 | 2019-04-24 14:20:21 +0100 | [diff] [blame] | 422 | std::unique_ptr<float[]> tensorData(new float[tensorInfo.GetNumElements()]); |
| 423 | const size_t tensorSizeInBytes = tensorInfo.GetNumBytes(); |
| 424 | // Copy the value list entries into the destination |
| 425 | if (!onnxTensor.has_raw_data()) |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 426 | { |
Pablo Tello | 3dcc1c6 | 2019-04-24 14:20:21 +0100 | [diff] [blame] | 427 | if(tensorInfo.GetNumElements() != static_cast<uint>(onnxTensor.float_data_size())) |
| 428 | { |
| 429 | throw ParseException(boost::str( |
| 430 | boost::format("The number of data provided (%1%) does not match the tensor '%2%' number of elements" |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 431 | " (%3%) %4%") |
| 432 | % onnxTensor.float_data_size() |
| 433 | % name |
| 434 | % tensorInfo.GetNumElements() |
| 435 | % CHECK_LOCATION().AsString())); |
Pablo Tello | 3dcc1c6 | 2019-04-24 14:20:21 +0100 | [diff] [blame] | 436 | } |
| 437 | ::memcpy(tensorData.get(), srcData, tensorSizeInBytes); |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 438 | } |
Pablo Tello | 3dcc1c6 | 2019-04-24 14:20:21 +0100 | [diff] [blame] | 439 | else |
| 440 | { |
| 441 | ::memcpy(tensorData.get(), onnxTensor.raw_data().c_str(), tensorSizeInBytes); |
| 442 | } |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 443 | |
| 444 | // Const tensors requires at least a list of values |
| 445 | if (tensorInfo.GetNumElements() == 0) |
| 446 | { |
| 447 | throw ParseException(boost::str( |
| 448 | boost::format("No tensor data found for Const tensor '%1%' %2%") |
| 449 | % name |
| 450 | % CHECK_LOCATION().AsString())); |
| 451 | } |
| 452 | return std::make_pair(ConstTensor(tensorInfo, tensorData.get()), std::move(tensorData)); |
| 453 | } |
| 454 | |
| 455 | ModelPtr OnnxParser::LoadModelFromTextFile(const char* graphFile) |
| 456 | { |
| 457 | FILE* fd = fopen(graphFile, "r"); |
| 458 | |
| 459 | if (fd == nullptr) |
| 460 | { |
| 461 | throw FileNotFoundException(boost::str( |
| 462 | boost::format("Invalid (null) filename %1%") % CHECK_LOCATION().AsString())); |
| 463 | } |
| 464 | |
| 465 | // Parse the file into a message |
| 466 | ModelPtr modelProto = std::make_unique<onnx::ModelProto>(); |
| 467 | using google::protobuf::io::FileInputStream; |
| 468 | std::unique_ptr<FileInputStream> input = std::make_unique<FileInputStream>(fileno(fd)); |
| 469 | bool success = google::protobuf::TextFormat::Parse(input.get(), modelProto.get()); |
| 470 | fclose(fd); |
| 471 | |
| 472 | if (!success) |
| 473 | { |
| 474 | std::stringstream error; |
| 475 | error << "Failed to parse graph file"; |
| 476 | throw ParseException(boost::str( |
| 477 | boost::format("%1% %2%") % error.str() % CHECK_LOCATION().AsString())); |
| 478 | } |
| 479 | return modelProto; |
| 480 | } |
| 481 | |
| 482 | INetworkPtr OnnxParser::CreateNetworkFromTextFile(const char* graphFile) |
| 483 | { |
| 484 | ResetParser(); |
| 485 | ModelPtr modelProto = LoadModelFromTextFile(graphFile); |
| 486 | return CreateNetworkFromModel(*modelProto); |
| 487 | } |
| 488 | |
| 489 | |
| 490 | ModelPtr OnnxParser::LoadModelFromBinaryFile(const char* graphFile) |
| 491 | { |
| 492 | FILE* fd = fopen(graphFile, "rb"); |
| 493 | |
| 494 | if (fd == nullptr) |
| 495 | { |
| 496 | throw FileNotFoundException(boost::str( |
| 497 | boost::format("Invalid (null) filename %1%") % CHECK_LOCATION().AsString())); |
| 498 | } |
| 499 | |
| 500 | // Parse the file into a message |
| 501 | ModelPtr modelProto = std::make_unique<onnx::ModelProto>(); |
| 502 | |
| 503 | google::protobuf::io::FileInputStream inStream(fileno(fd)); |
| 504 | google::protobuf::io::CodedInputStream codedStream(&inStream); |
| 505 | codedStream.SetTotalBytesLimit(INT_MAX, INT_MAX); |
| 506 | bool success = modelProto.get()->ParseFromCodedStream(&codedStream); |
| 507 | fclose(fd); |
| 508 | |
| 509 | if (!success) |
| 510 | { |
| 511 | std::stringstream error; |
| 512 | error << "Failed to parse graph file"; |
| 513 | throw ParseException(boost::str( |
| 514 | boost::format("%1% %2%") % error.str() % CHECK_LOCATION().AsString())); |
| 515 | } |
| 516 | return modelProto; |
| 517 | |
| 518 | } |
| 519 | |
| 520 | INetworkPtr OnnxParser::CreateNetworkFromBinaryFile(const char* graphFile) |
| 521 | { |
| 522 | ResetParser(); |
| 523 | ModelPtr modelProto = LoadModelFromBinaryFile(graphFile); |
| 524 | return CreateNetworkFromModel(*modelProto); |
| 525 | } |
| 526 | |
| 527 | ModelPtr OnnxParser::LoadModelFromString(const std::string& protoText) |
| 528 | { |
| 529 | if (protoText == "") |
| 530 | { |
| 531 | throw InvalidArgumentException(boost::str( |
| 532 | boost::format("Invalid (empty) string for model parameter %1%") % CHECK_LOCATION().AsString())); |
| 533 | } |
| 534 | // Parse the string into a message |
| 535 | ModelPtr modelProto = std::make_unique<onnx::ModelProto>(); |
| 536 | bool success = google::protobuf::TextFormat::ParseFromString(protoText, modelProto.get()); |
| 537 | if (!success) |
| 538 | { |
| 539 | std::stringstream error; |
| 540 | error << "Failed to parse graph file"; |
| 541 | throw ParseException(boost::str( |
| 542 | boost::format("%1% %2%") % error.str() % CHECK_LOCATION().AsString())); |
| 543 | } |
| 544 | return modelProto; |
| 545 | } |
| 546 | |
| 547 | INetworkPtr OnnxParser::CreateNetworkFromString(const std::string& protoText) |
| 548 | { |
| 549 | ResetParser(); |
| 550 | ModelPtr modelProto = LoadModelFromString(protoText); |
| 551 | return CreateNetworkFromModel(*modelProto); |
| 552 | } |
| 553 | |
| 554 | INetworkPtr OnnxParser::CreateNetworkFromModel(onnx::ModelProto& model) |
| 555 | { |
| 556 | m_Network = INetwork::Create(); |
| 557 | try |
| 558 | { |
| 559 | m_Graph = std::make_unique<onnx::GraphProto>(*model.mutable_graph()); |
| 560 | LoadGraph(); |
| 561 | } |
| 562 | catch (const ParseException& e) |
| 563 | { |
| 564 | Cleanup(); |
| 565 | throw e; |
| 566 | } |
| 567 | Cleanup(); |
| 568 | return std::move(m_Network); |
| 569 | } |
| 570 | |
| 571 | void OnnxParser::LoadGraph() |
| 572 | { |
| 573 | BOOST_ASSERT(m_Graph.get() != nullptr); |
| 574 | |
| 575 | //Fill m_TensorsInfo with the shapes and value of every tensor |
| 576 | SetupInfo(m_Graph->mutable_output()); |
| 577 | SetupInfo(m_Graph->mutable_input()); |
| 578 | SetupInfo(m_Graph->mutable_value_info()); |
| 579 | |
| 580 | for (auto tensor : m_Graph->initializer()) |
| 581 | { |
| 582 | m_TensorsInfo[tensor.name()].m_tensor = std::make_unique<const onnx::TensorProto>(tensor); |
| 583 | } |
| 584 | |
| 585 | SetupInputLayers(); |
| 586 | SetupOutputLayers(); |
| 587 | |
| 588 | //Detect FullyConnected layers with bias and update the FusedAndUsed map acccordingly |
| 589 | DetectFullyConnected(); |
| 590 | |
| 591 | //Parsing the graph |
| 592 | for(size_t nodeIndex = 0; nodeIndex < static_cast<size_t>(m_Graph->node_size()); nodeIndex++) |
| 593 | { |
| 594 | auto node = m_Graph->node(static_cast<int>(nodeIndex)); |
| 595 | const std::string& operation = node.op_type(); |
| 596 | |
| 597 | // check which layers we handled already (add and matmul fused as FC) |
| 598 | if(operation == "MatMul" ) |
| 599 | { |
| 600 | if(m_OutputsFusedAndUsed[nodeIndex].inputForNodes != m_OutputsFusedAndUsed[nodeIndex].fusedWithNodes.size()) |
| 601 | { |
| 602 | //Node which can not be fused as a FullyConnected layer (used in layers as a simple matmul output) |
| 603 | AddFullyConnected(node); |
| 604 | } |
| 605 | } |
| 606 | else if (!(m_OutputsFusedAndUsed[nodeIndex].fusedWithNodes.empty()) && operation == "Add") |
| 607 | { |
| 608 | int matmulIndex = static_cast<int> (m_OutputsFusedAndUsed[nodeIndex].fusedWithNodes[0]); |
| 609 | AddFullyConnected(m_Graph->node(matmulIndex), &node); |
| 610 | } |
| 611 | else if (m_OutputsFusedAndUsed[nodeIndex].fusedWithNodes.empty()) //node is not part of a fused layer |
| 612 | { |
| 613 | auto it = m_ParserFunctions.find(operation); |
| 614 | if (it != m_ParserFunctions.end()) |
| 615 | { |
| 616 | auto func = it->second; |
| 617 | (this->*func)(node); |
| 618 | } |
| 619 | else |
| 620 | { |
| 621 | throw ParseException(boost::str( |
| 622 | boost::format("Unsupported operation %1% for node '%2%' %3%") |
| 623 | % operation |
| 624 | % node.name() |
| 625 | % CHECK_LOCATION().AsString())); |
| 626 | } |
| 627 | } |
| 628 | } |
| 629 | |
| 630 | //Making the connections between outputs and inputs of each layers |
| 631 | for (const auto& tensorCon : m_TensorConnections) |
| 632 | { |
| 633 | if (tensorCon.second.outputSlot != nullptr) |
| 634 | { |
| 635 | for (size_t inputSlotIdx = 0; inputSlotIdx < tensorCon.second.inputSlots.size(); ++inputSlotIdx) |
| 636 | { |
| 637 | tensorCon.second.outputSlot->Connect(*(tensorCon.second.inputSlots[inputSlotIdx])); |
| 638 | } |
| 639 | } |
| 640 | } |
| 641 | } |
| 642 | |
| 643 | void OnnxParser::SetupInfo(const google::protobuf::RepeatedPtrField<onnx::ValueInfoProto >* list) |
| 644 | { |
| 645 | for (auto tensor : *list) |
| 646 | { |
| 647 | m_TensorsInfo[tensor.name()] = OnnxTensor(); |
| 648 | m_TensorsInfo[tensor.name()].m_info = std::make_unique<TensorInfo>(ToTensorInfo(tensor)); |
Matteo Martincigh | e355dc2 | 2018-12-10 13:45:27 +0000 | [diff] [blame] | 649 | m_TensorsInfo[tensor.name()].m_dtype = |
| 650 | static_cast<onnx::TensorProto::DataType>(tensor.type().tensor_type().elem_type()); |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 651 | } |
| 652 | } |
| 653 | |
| 654 | void OnnxParser::DetectFullyConnected() |
| 655 | { |
| 656 | m_OutputsFusedAndUsed = std::vector<UsageSummary> (static_cast<size_t>(m_Graph->node_size()), UsageSummary()); |
| 657 | auto matmulAndConstant = [&](const std::string& constInput, |
| 658 | const std::string& matmulInput, |
| 659 | int& nodeIndex) |
| 660 | { |
| 661 | auto matmulIt = m_OutputsMap.find(matmulInput); |
| 662 | if(matmulIt != m_OutputsMap.end() && matmulIt->second.first->op_type() == "MatMul" |
| 663 | && m_TensorsInfo[constInput].isConstant()) |
| 664 | { |
| 665 | nodeIndex = matmulIt->second.second; |
| 666 | return true; |
| 667 | } |
| 668 | return false; |
| 669 | }; |
| 670 | |
| 671 | for(int nodeIndex = 0; nodeIndex < m_Graph->node_size(); nodeIndex++) |
| 672 | { |
| 673 | const onnx::NodeProto* node = &m_Graph->node(nodeIndex); |
| 674 | for (const std::string& output : node->output()) |
| 675 | { |
| 676 | m_OutputsMap[output] = std::make_pair(node, nodeIndex); |
| 677 | } |
| 678 | |
| 679 | for (const std::string& input : node->input()) //count how many time a node is used as input |
| 680 | { |
| 681 | auto matmulIt = m_OutputsMap.find(input); |
| 682 | if(matmulIt != m_OutputsMap.end()){ |
| 683 | ++m_OutputsFusedAndUsed[static_cast<size_t>(matmulIt->second.second)].inputForNodes; //node used |
| 684 | } |
| 685 | } |
| 686 | |
| 687 | if (node->op_type() == "Add") |
| 688 | { |
| 689 | int matmulIndex = 0; |
| 690 | if (matmulAndConstant(node->input(0), node->input(1), matmulIndex) || |
| 691 | matmulAndConstant(node->input(1), node->input(0), matmulIndex)) |
| 692 | { |
| 693 | //matmul and add were fused |
| 694 | m_OutputsFusedAndUsed[static_cast<size_t>(matmulIndex)].fusedWithNodes |
| 695 | .push_back(static_cast<size_t>(nodeIndex)); |
| 696 | |
| 697 | m_OutputsFusedAndUsed[static_cast<size_t>(nodeIndex)].fusedWithNodes |
| 698 | .push_back(static_cast<size_t>(matmulIndex)); |
| 699 | } |
| 700 | } |
| 701 | } |
| 702 | |
| 703 | for (auto output: m_Graph->output()) { //Add usages as output of the graph in count of usages |
| 704 | auto matmulIt = m_OutputsMap.find(output.name()); |
| 705 | if(matmulIt != m_OutputsMap.end()){ |
| 706 | ++m_OutputsFusedAndUsed[static_cast<size_t>(matmulIt->second.second)].inputForNodes; |
| 707 | } |
| 708 | } |
| 709 | } |
| 710 | |
| 711 | template<typename Location> |
| 712 | void OnnxParser::GetInputAndParam(const onnx::NodeProto& node, |
| 713 | std::string* inputName, |
| 714 | std::string* constName, |
| 715 | const Location& location) |
| 716 | { |
| 717 | int cstIndex; |
| 718 | if (m_TensorsInfo[node.input(0)].isConstant()) |
| 719 | { |
| 720 | cstIndex = 0; |
| 721 | } |
| 722 | else if (m_TensorsInfo[node.input(1)].isConstant()) |
| 723 | { |
| 724 | cstIndex = 1; |
| 725 | } |
| 726 | else |
| 727 | { |
| 728 | throw ParseException(boost::str( |
| 729 | boost::format("One of the input tensors ('%1%' or '%2%') should be constant in node '%3%' %4%") |
| 730 | % node.input(0) |
| 731 | % node.input(1) |
| 732 | % node.name() |
| 733 | % location.AsString())); |
| 734 | } |
| 735 | if(constName) |
| 736 | { |
| 737 | *constName = node.input(cstIndex); |
| 738 | } |
| 739 | if(inputName) |
| 740 | { |
| 741 | *inputName = node.input(!cstIndex); |
| 742 | } |
| 743 | } |
| 744 | |
| 745 | template<typename Location> |
| 746 | void OnnxParser::To1DTensor(const std::string& name, const Location& location) |
| 747 | { |
| 748 | TensorShape shape = m_TensorsInfo[name].m_info->GetShape(); |
| 749 | std::vector<uint32_t> newShape; |
| 750 | for(uint i = 0; i < shape.GetNumDimensions() - 1; ++i) |
| 751 | { |
| 752 | if(shape[i] != 1) |
| 753 | { |
| 754 | throw ParseException(boost::str( |
| 755 | boost::format("Only tensors with shape [1, ..., 1, X] can be converted to 1D and %1% %2%") |
| 756 | % TensorInfoAsString(*m_TensorsInfo[name].m_info, name, m_TensorsInfo[name].m_dtype) |
| 757 | % location.AsString())); |
| 758 | } |
| 759 | } |
| 760 | newShape.push_back(shape[shape.GetNumDimensions() - 1]); |
| 761 | |
| 762 | m_TensorsInfo[name].m_info->SetShape(TensorShape(static_cast<unsigned int>(newShape.size()), newShape.data())); |
| 763 | } |
| 764 | |
| 765 | void OnnxParser::AddFullyConnected(const onnx::NodeProto& matmulNode, const onnx::NodeProto* addNode) |
| 766 | { |
| 767 | |
| 768 | // find matmul inputs |
| 769 | std::string weightName; |
| 770 | std::string inputName; |
| 771 | CHECK_VALID_SIZE(static_cast<size_t>(matmulNode.input_size()), 2); |
| 772 | CHECK_VALID_SIZE(static_cast<size_t>(matmulNode.output_size()), 1); |
| 773 | VALID_INPUTS(matmulNode, STR_LIST(onnx::TensorProto::FLOAT)); |
| 774 | |
| 775 | GetInputAndParam(matmulNode, &inputName, &weightName, CHECK_LOCATION()); |
| 776 | |
| 777 | FullyConnectedDescriptor desc; |
| 778 | desc.m_BiasEnabled = addNode != nullptr; |
| 779 | |
| 780 | IConnectableLayer* layer = nullptr; |
| 781 | if(desc.m_BiasEnabled) |
| 782 | { |
| 783 | // find bias const |
| 784 | std::string biasName; |
| 785 | CHECK_VALID_SIZE(static_cast<size_t>(addNode->input_size()), 2); |
| 786 | CHECK_VALID_SIZE(static_cast<size_t>(addNode->output_size()), 1); |
| 787 | VALID_INPUTS(*addNode, STR_LIST(onnx::TensorProto::FLOAT)); |
| 788 | |
| 789 | GetInputAndParam(*addNode, nullptr, &biasName, CHECK_LOCATION()); |
| 790 | |
| 791 | //Output shape is [1, weights[1]] and 1d vec in ONNX can be [1,X] so we convert biases to "armnn" 1D |
| 792 | To1DTensor(biasName, CHECK_LOCATION()); |
| 793 | TensorInfo weightInfo = *m_TensorsInfo[weightName].m_info; |
| 794 | TensorInfo biasInfo = *m_TensorsInfo[biasName].m_info; |
| 795 | |
| 796 | if (weightInfo.GetShape()[1] != biasInfo.GetShape()[0]) |
| 797 | { |
| 798 | throw ParseException(boost::str( |
| 799 | boost::format("Shape of weights '%1%' and bias of following Add node '%2%' do not match : %3%" |
| 800 | " and %4% ( /!\\ bias should be a 1D tensor) %5%") |
| 801 | % weightName |
| 802 | % addNode->name() |
| 803 | % TensorInfoAsString(*m_TensorsInfo[weightName].m_info, |
| 804 | weightName, |
| 805 | m_TensorsInfo[weightName].m_dtype) |
| 806 | % TensorInfoAsString(*m_TensorsInfo[biasName].m_info, biasName, |
| 807 | m_TensorsInfo[biasName].m_dtype ) |
| 808 | % CHECK_LOCATION().AsString())); |
| 809 | } |
| 810 | layer = m_Network->AddFullyConnectedLayer(desc, |
| 811 | CreateConstTensor(weightName).first, |
Matteo Martincigh | fc598e1 | 2019-05-14 10:36:13 +0100 | [diff] [blame] | 812 | Optional<ConstTensor>(CreateConstTensor(biasName).first), |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 813 | matmulNode.name().c_str()); |
| 814 | BOOST_ASSERT(layer != nullptr); |
| 815 | |
| 816 | auto outputInfo = ComputeOutputInfo({addNode->output(0)}, layer, |
| 817 | {m_TensorsInfo[inputName].m_info->GetShape(), |
| 818 | m_TensorsInfo[weightName].m_info->GetShape()}); |
| 819 | |
| 820 | layer->GetOutputSlot(0).SetTensorInfo(outputInfo[0]); |
| 821 | |
| 822 | RegisterInputSlots(layer, {inputName}); |
| 823 | RegisterOutputSlots(layer, {addNode->output(0)}); |
| 824 | } |
| 825 | else |
| 826 | { |
Matteo Martincigh | fc598e1 | 2019-05-14 10:36:13 +0100 | [diff] [blame] | 827 | layer = m_Network->AddFullyConnectedLayer(desc, |
| 828 | CreateConstTensor(weightName).first, |
| 829 | EmptyOptional(), |
| 830 | matmulNode.name().c_str()); |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 831 | BOOST_ASSERT(layer != nullptr); |
| 832 | |
| 833 | auto outputInfo = ComputeOutputInfo({matmulNode.output(0)}, layer, |
| 834 | {m_TensorsInfo[inputName].m_info->GetShape(), |
| 835 | m_TensorsInfo[weightName].m_info->GetShape()}); |
| 836 | layer->GetOutputSlot(0).SetTensorInfo(outputInfo[0]); |
| 837 | |
| 838 | RegisterInputSlots(layer, {inputName}); |
| 839 | RegisterOutputSlots(layer, {matmulNode.output(0)}); |
| 840 | } |
| 841 | } |
| 842 | |
| 843 | void OnnxParser::CreateConstantLayer(const std::string& tensorName, const std::string& layerName) |
| 844 | { |
| 845 | auto armnnTensor = CreateConstTensor(tensorName); |
| 846 | |
| 847 | IConnectableLayer* layer = m_Network->AddConstantLayer(armnnTensor.first, layerName.c_str()); |
| 848 | layer->GetOutputSlot(0).SetTensorInfo(armnnTensor.first.GetInfo()); |
| 849 | RegisterOutputSlots(layer, {tensorName}); |
| 850 | } |
| 851 | |
| 852 | void OnnxParser::ParseConstant(const onnx::NodeProto& node) |
| 853 | { |
| 854 | CHECK_VALID_SIZE(static_cast<size_t>(node.attribute_size()), 1); |
| 855 | |
| 856 | if (!node.attribute(0).has_t()) |
| 857 | { |
| 858 | throw ParseException(boost::str( |
| 859 | boost::format("Value not found for Constant node '%1%' %2%") |
| 860 | % node.name() |
| 861 | % CHECK_LOCATION().AsString())); |
| 862 | } |
| 863 | const onnx::TensorProto& onnxTensor = node.attribute(0).t(); |
| 864 | |
| 865 | //ONNX can have Float16 and double constant nodes but ArmNN only supports float32 |
Matteo Martincigh | e355dc2 | 2018-12-10 13:45:27 +0000 | [diff] [blame] | 866 | CHECK_VALID_DATATYPE(node.name(), onnxTensor.name(), |
| 867 | static_cast<onnx::TensorProto::DataType>(onnxTensor.data_type()), onnx::TensorProto::FLOAT); |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 868 | |
| 869 | //Register this as a m_ConstParam so we know we can use it as a constant param in future layers. |
| 870 | m_TensorsInfo[node.output(0)].m_tensor = std::make_unique<const onnx::TensorProto>(onnxTensor); |
| 871 | |
| 872 | CreateConstantLayer(node.output(0), node.name()); |
| 873 | |
| 874 | } |
| 875 | |
| 876 | void OnnxParser::ParseMaxPool(const onnx::NodeProto& node) |
| 877 | { |
| 878 | Pooling2dDescriptor desc; |
| 879 | desc.m_PoolType = PoolingAlgorithm::Max; |
| 880 | desc.m_PaddingMethod = PaddingMethod::Exclude; |
| 881 | AddPoolingLayer(node, desc); |
| 882 | } |
| 883 | |
| 884 | void OnnxParser::ParseGlobalAveragePool(const onnx::NodeProto& node) |
| 885 | { |
| 886 | Pooling2dDescriptor desc = Pooling2dDescriptor(); |
| 887 | desc.m_PoolType = PoolingAlgorithm::Average; |
| 888 | |
| 889 | //kernel size is the same as input |
| 890 | TensorShape inputShape = m_TensorsInfo[node.input(0)].m_info->GetShape(); |
| 891 | desc.m_PoolWidth = inputShape[3]; |
| 892 | desc.m_PoolHeight = inputShape[2]; |
| 893 | |
| 894 | IConnectableLayer* layer = m_Network->AddPooling2dLayer(desc, node.name().c_str()); |
| 895 | BOOST_ASSERT(layer != nullptr); |
| 896 | |
| 897 | auto outputInfo = ComputeOutputInfo({node.output(0)}, layer, {inputShape}); |
| 898 | layer->GetOutputSlot(0).SetTensorInfo(outputInfo[0]); |
| 899 | |
| 900 | // register the input connection slots for the layer, connections are made after all layers have been created |
| 901 | // only the tensors for the inputs are relevant, exclude the const tensors |
| 902 | RegisterInputSlots(layer, {node.input(0)}); |
| 903 | |
| 904 | // register the output connection slots for the layer, connections are made after all layers have been created |
| 905 | RegisterOutputSlots(layer, {node.output(0)}); |
| 906 | } |
| 907 | |
| 908 | void OnnxParser::ParseAveragePool(const onnx::NodeProto& node) |
| 909 | { |
| 910 | Pooling2dDescriptor desc; |
| 911 | desc.m_PoolType = PoolingAlgorithm::Average; |
| 912 | |
| 913 | uint32_t count_include_pad = 0; |
| 914 | count_include_pad = ReadOptionalNodeUint32Attribute(node, "count_include_pad"); |
| 915 | if(count_include_pad) { |
| 916 | desc.m_PaddingMethod = PaddingMethod::IgnoreValue; |
| 917 | } |
| 918 | AddPoolingLayer(node, desc); |
| 919 | } |
| 920 | |
| 921 | void OnnxParser::AddPoolingLayer(const onnx::NodeProto& node, Pooling2dDescriptor& desc) |
| 922 | { |
| 923 | |
| 924 | CHECK_VALID_SIZE(static_cast<size_t>(node.input_size()), 1); |
| 925 | CHECK_VALID_SIZE(static_cast<size_t>(node.output_size()), 1); |
| 926 | |
| 927 | VALID_INPUTS(node, STR_LIST(onnx::TensorProto::FLOAT)); |
| 928 | |
| 929 | std::vector<uint32_t> kernel_shape = ReadMandatoryNodeUint32ListAttribute(node, "kernel_shape"); //size of pool win |
| 930 | std::vector<uint32_t> strides = ReadOptionalNodeUint32ListAttribute(node, "strides"); |
| 931 | std::vector<uint32_t> pads = ReadOptionalNodeUint32ListAttribute(node, "pads"); |
| 932 | |
| 933 | desc.m_OutputShapeRounding = OutputShapeRounding::Floor; |
| 934 | desc.m_PoolWidth = kernel_shape[1]; |
| 935 | desc.m_PoolHeight = kernel_shape[0]; |
| 936 | |
| 937 | if(strides.empty()) |
| 938 | { |
| 939 | desc.m_StrideX = 1; |
| 940 | desc.m_StrideY = 1; |
| 941 | } |
| 942 | else |
| 943 | { |
| 944 | desc.m_StrideX = strides[1]; |
| 945 | desc.m_StrideY = strides[0]; |
| 946 | } |
| 947 | |
| 948 | //Check new padding version first |
| 949 | if(pads.empty()) |
| 950 | { |
| 951 | //Check deprecated version |
| 952 | std::string paddingString = ReadOptionalNodeStringAttribute(node, "auto_pad"); |
| 953 | if(paddingString != "VALID" && paddingString != "" && paddingString != "NOTSET") |
| 954 | { |
| 955 | bool isUpper; |
| 956 | if( paddingString == "SAME_LOWER") |
| 957 | { |
| 958 | isUpper = false; |
| 959 | } |
| 960 | else if (paddingString == "SAME_UPPER") |
| 961 | { |
| 962 | isUpper = true; |
| 963 | } |
| 964 | else |
| 965 | { |
| 966 | throw ParseException(boost::str( |
| 967 | boost::format("Invalid auto_pad attribute for node %1%. " |
| 968 | "Only SAME_UPPER, SAME_LOWER or VALID supported and found %2% %3%") |
| 969 | % node.name() |
| 970 | % paddingString |
| 971 | % CHECK_LOCATION().AsString())); |
| 972 | } |
| 973 | auto inputInfo = *m_TensorsInfo[node.input(0)].m_info; |
| 974 | uint32_t inputHeight = inputInfo.GetShape()[2]; |
| 975 | uint32_t inputWidth = inputInfo.GetShape()[3]; |
| 976 | CalcPadding(inputHeight, desc.m_PoolHeight, desc.m_StrideY, &desc.m_PadTop, &desc.m_PadBottom, isUpper); |
| 977 | CalcPadding(inputWidth, desc.m_PoolWidth, desc.m_StrideX, &desc.m_PadLeft, &desc.m_PadRight, isUpper); |
| 978 | } |
| 979 | } |
| 980 | else |
| 981 | { |
| 982 | desc.m_PadTop = pads[0]; |
| 983 | desc.m_PadLeft = pads[1]; |
| 984 | desc.m_PadBottom = pads[2]; |
| 985 | desc.m_PadRight = pads[3]; |
| 986 | } |
| 987 | |
| 988 | IConnectableLayer* layer = m_Network->AddPooling2dLayer(desc, node.name().c_str()); |
| 989 | BOOST_ASSERT(layer != nullptr); |
| 990 | |
| 991 | auto outputInfo = ComputeOutputInfo({node.output(0)}, layer, {m_TensorsInfo[node.input(0)].m_info->GetShape()}); |
| 992 | layer->GetOutputSlot(0).SetTensorInfo(outputInfo[0]); |
| 993 | |
| 994 | // register the input connection slots for the layer, connections are made after all layers have been created |
| 995 | // only the tensors for the inputs are relevant, exclude the const tensors |
| 996 | RegisterInputSlots(layer, {node.input(0)}); |
| 997 | |
| 998 | // register the output connection slots for the layer, connections are made after all layers have been created |
| 999 | RegisterOutputSlots(layer, {node.output(0)}); |
| 1000 | } |
| 1001 | |
| 1002 | void OnnxParser::CreateReshapeLayer(const std::string& inputName, |
| 1003 | const std::string& outputName, |
| 1004 | const std::string& layerName) |
| 1005 | { |
| 1006 | const TensorInfo outputTensorInfo = *m_TensorsInfo[outputName].m_info; |
| 1007 | ReshapeDescriptor reshapeDesc; |
| 1008 | reshapeDesc.m_TargetShape = outputTensorInfo.GetShape(); |
| 1009 | |
| 1010 | IConnectableLayer* layer = m_Network->AddReshapeLayer(reshapeDesc, layerName.c_str()); |
| 1011 | BOOST_ASSERT(layer != nullptr); |
| 1012 | layer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo); |
| 1013 | |
| 1014 | // register the input connection slots for the layer, connections are made after all layers have been created |
| 1015 | // only the tensors for the inputs are relevant, exclude the const tensors |
| 1016 | RegisterInputSlots(layer, {inputName}); |
| 1017 | |
| 1018 | // register the output connection slots for the layer, connections are made after all layers have been created |
| 1019 | RegisterOutputSlots(layer, {outputName}); |
| 1020 | } |
| 1021 | |
| 1022 | void OnnxParser::ParseReshape(const onnx::NodeProto& node) |
| 1023 | { |
| 1024 | CHECK_VALID_SIZE(static_cast<size_t>(node.input_size()), 2); |
| 1025 | CHECK_VALID_SIZE(static_cast<size_t>(node.output_size()), 1); |
| 1026 | |
| 1027 | CHECK_VALID_DATATYPE(node.name(), node.input(0), |
| 1028 | m_TensorsInfo[node.input(0)].m_dtype, |
| 1029 | onnx::TensorProto::FLOAT); //input |
| 1030 | CHECK_VALID_DATATYPE(node.name(), node.input(1), |
| 1031 | m_TensorsInfo[node.input(1)].m_dtype, |
| 1032 | onnx::TensorProto::INT64); //shape |
| 1033 | |
| 1034 | if(!m_TensorsInfo[node.input(1)].isConstant()) |
| 1035 | { |
| 1036 | throw ParseException(boost::str( |
| 1037 | boost::format("Shape '%1%' should be constant in Reshape layer '%2%' %3%") |
| 1038 | % node.input(1) |
| 1039 | % node.name() |
| 1040 | % CHECK_LOCATION().AsString())); |
| 1041 | } |
| 1042 | |
| 1043 | if(m_TensorsInfo[node.input(0)].isConstant()) |
| 1044 | { |
| 1045 | //make a new cst tensor -> move the data to the output tensor (the shape is already good in the output tensor) |
| 1046 | if(m_TensorsInfo.count(node.output(0)) == 0) |
| 1047 | { |
| 1048 | m_TensorsInfo[node.output(0)] = OnnxTensor(); |
| 1049 | } |
| 1050 | m_TensorsInfo[node.output(0)].m_tensor = |
| 1051 | std::make_unique<onnx::TensorProto>(*m_TensorsInfo[node.input(0)].m_tensor); |
| 1052 | } |
| 1053 | else |
| 1054 | { |
| 1055 | TensorShape inputShape = m_TensorsInfo[node.input(0)].m_info->GetShape(); |
| 1056 | |
| 1057 | if(m_TensorsInfo.count(node.output(0)) == 0 || m_TensorsInfo[node.output(0)].m_info == nullptr) |
| 1058 | { |
| 1059 | auto outInfo = ComputeReshapeInfo(*m_TensorsInfo[node.input(1)].m_tensor, inputShape, node.output(0)); |
| 1060 | m_TensorsInfo[node.output(0)].m_info = std::make_unique<TensorInfo>(outInfo); |
| 1061 | } |
| 1062 | |
| 1063 | CreateReshapeLayer(node.input(0), node.output(0), node.name()); |
| 1064 | } |
| 1065 | } |
| 1066 | |
| 1067 | void OnnxParser::ParseRelu(const onnx::NodeProto& node) |
| 1068 | { |
| 1069 | CHECK_VALID_SIZE(static_cast<size_t>(node.input_size()), 1); |
| 1070 | CHECK_VALID_SIZE(static_cast<size_t>(node.output_size()), 1); |
| 1071 | |
| 1072 | VALID_INPUTS(node, STR_LIST(onnx::TensorProto::FLOAT)); |
| 1073 | |
| 1074 | ActivationDescriptor desc; |
| 1075 | desc.m_Function = ActivationFunction::ReLu; |
| 1076 | |
| 1077 | IConnectableLayer* const layer = m_Network->AddActivationLayer(desc, node.name().c_str()); |
| 1078 | BOOST_ASSERT(layer != nullptr); |
| 1079 | |
| 1080 | auto outputInfo = ComputeOutputInfo({ node.output(0)}, layer, {m_TensorsInfo[node.input(0)].m_info->GetShape()}); |
| 1081 | layer->GetOutputSlot(0).SetTensorInfo(outputInfo[0]); |
| 1082 | |
| 1083 | // register the input connection slots for the layer, connections are made after all layers have been created |
| 1084 | // only the tensors for the inputs are relevant, exclude the const tensors |
| 1085 | RegisterInputSlots(layer, {node.input(0)}); |
| 1086 | |
| 1087 | // register the output connection slots for the layer, connections are made after all layers have been created |
| 1088 | RegisterOutputSlots(layer, {node.output(0)}); |
| 1089 | } |
| 1090 | |
| 1091 | |
| 1092 | void OnnxParser::AddConvLayerWithDepthwiseConv(const onnx::NodeProto& node, const Convolution2dDescriptor& convDesc) |
| 1093 | { |
| 1094 | BOOST_ASSERT(node.op_type() == "Conv"); |
| 1095 | |
| 1096 | DepthwiseConvolution2dDescriptor desc; |
| 1097 | desc.m_PadLeft = convDesc.m_PadLeft; |
| 1098 | desc.m_PadRight = convDesc.m_PadRight; |
| 1099 | desc.m_PadTop = convDesc.m_PadTop; |
| 1100 | desc.m_PadBottom = convDesc.m_PadBottom; |
| 1101 | desc.m_StrideX = convDesc.m_StrideX; |
| 1102 | desc.m_StrideY = convDesc.m_StrideY; |
| 1103 | desc.m_BiasEnabled = convDesc.m_BiasEnabled; |
| 1104 | |
| 1105 | armnn::IConnectableLayer* layer; |
| 1106 | auto weightTensor = CreateConstTensor(node.input(1)); |
| 1107 | TensorShape& weightShape = weightTensor.first.GetShape(); |
| 1108 | weightShape[1] = weightShape[0]; |
| 1109 | weightShape[0] = 1; |
| 1110 | m_TensorsInfo[node.input(1)].m_info->SetShape(weightShape); |
| 1111 | |
| 1112 | if (node.input_size() == 3) |
| 1113 | { |
| 1114 | if(!m_TensorsInfo[node.input(2)].isConstant()) |
| 1115 | { |
| 1116 | throw ParseException(boost::str( |
| 1117 | boost::format("Bias '%1%' should be constant in Conv layer '%2%' %3%") |
| 1118 | % node.input(2) |
| 1119 | % node.name() |
| 1120 | % CHECK_LOCATION().AsString())); |
| 1121 | } |
| 1122 | desc.m_BiasEnabled = true; |
| 1123 | auto biasTensor = CreateConstTensor(node.input(2)); |
| 1124 | layer = m_Network->AddDepthwiseConvolution2dLayer(desc, |
| 1125 | weightTensor.first, |
Matteo Martincigh | fc598e1 | 2019-05-14 10:36:13 +0100 | [diff] [blame] | 1126 | Optional<ConstTensor>(biasTensor.first), |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1127 | node.name().c_str()); |
| 1128 | } |
| 1129 | else |
| 1130 | { |
| 1131 | layer = m_Network->AddDepthwiseConvolution2dLayer(desc, |
| 1132 | weightTensor.first, |
Matteo Martincigh | fc598e1 | 2019-05-14 10:36:13 +0100 | [diff] [blame] | 1133 | EmptyOptional(), |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1134 | node.name().c_str()); |
| 1135 | } |
| 1136 | BOOST_ASSERT(layer != nullptr); |
| 1137 | |
| 1138 | auto outputInfo = ComputeOutputInfo({ node.output(0) }, layer, |
| 1139 | { m_TensorsInfo[node.input(0)].m_info->GetShape(), |
| 1140 | m_TensorsInfo[node.input(1)].m_info->GetShape() }); |
| 1141 | |
| 1142 | layer->GetOutputSlot(0).SetTensorInfo(outputInfo[0]); |
| 1143 | |
| 1144 | // register the input connection slots for the layer, connections are made after all layers have been created |
| 1145 | // only the tensors for the inputs are relevant, exclude the const tensors |
| 1146 | RegisterInputSlots(layer, {node.input(0)}); |
| 1147 | |
| 1148 | // register the output connection slots for the layer, connections are made after all layers have been created |
| 1149 | RegisterOutputSlots(layer, {node.output(0)}); |
| 1150 | } |
| 1151 | |
| 1152 | void OnnxParser::ParseConv(const onnx::NodeProto& node) |
| 1153 | { |
| 1154 | CHECK_VALID_SIZE(static_cast<size_t>(node.input_size()), 2, 3); //input, weight, (bias) |
| 1155 | CHECK_VALID_SIZE(static_cast<size_t>(node.output_size()), 1); |
| 1156 | |
| 1157 | VALID_INPUTS(node, STR_LIST(onnx::TensorProto::FLOAT)); |
| 1158 | |
| 1159 | if(m_TensorsInfo[node.input(0)].m_info->GetNumDimensions() != 4) |
| 1160 | { |
| 1161 | throw ParseException(boost::str( |
| 1162 | boost::format("ArmNN only supports 2D convolution and Conv layer '%1%' input %2% %3%") |
| 1163 | % node.name() |
| 1164 | % TensorInfoAsString(*m_TensorsInfo[node.input(0)].m_info, node.input(0), |
| 1165 | m_TensorsInfo[node.input(0)].m_dtype) |
| 1166 | % CHECK_LOCATION().AsString())); |
| 1167 | } |
| 1168 | |
| 1169 | if(!m_TensorsInfo[node.input(1)].isConstant()) |
| 1170 | { |
| 1171 | throw ParseException(boost::str( |
| 1172 | boost::format("Weights '%1%' should be constant in Conv layer '%2%' %3%") |
| 1173 | % node.input(1) |
| 1174 | % node.name() |
| 1175 | % CHECK_LOCATION().AsString())); |
| 1176 | } |
| 1177 | |
| 1178 | auto inputInfo = *m_TensorsInfo[node.input(0)].m_info; |
| 1179 | |
| 1180 | std::vector<uint32_t> dilations = ReadOptionalNodeUint32ListAttribute(node, "dilations"); |
| 1181 | if (!dilations.empty()) |
| 1182 | { |
| 1183 | std::stringstream ss; |
| 1184 | ss << "[ "; |
| 1185 | for (auto dilation : dilations) |
| 1186 | { |
| 1187 | ss << dilation << ", "; |
| 1188 | if (dilation != 1u) |
| 1189 | { |
| 1190 | ss << "... ]"; |
| 1191 | throw ParseException(boost::str( |
| 1192 | boost::format("ArmNN only supports Convolution layers with dilations [1,1], and node '%1%' " |
| 1193 | "has dilatation %2% %3%") |
| 1194 | % node.name() |
| 1195 | % ss.str() |
| 1196 | % CHECK_LOCATION().AsString())); |
| 1197 | } |
| 1198 | } |
| 1199 | } |
| 1200 | |
| 1201 | Convolution2dDescriptor desc; |
| 1202 | desc.m_BiasEnabled = false; |
| 1203 | |
| 1204 | std::vector<uint32_t> strides = ReadOptionalNodeUint32ListAttribute(node, "strides"); |
| 1205 | if(strides.empty()) |
| 1206 | { |
| 1207 | desc.m_StrideX = 1; |
| 1208 | desc.m_StrideY = 1; |
| 1209 | } |
| 1210 | else |
| 1211 | { |
| 1212 | desc.m_StrideX = strides[1]; |
| 1213 | desc.m_StrideY = strides[0]; |
| 1214 | } |
| 1215 | |
| 1216 | std::vector<uint32_t> pads = ReadOptionalNodeUint32ListAttribute(node, "pads"); |
| 1217 | //Check new padding version first |
| 1218 | if(pads.empty()) |
| 1219 | { |
| 1220 | //Check deprecated version |
| 1221 | std::string paddingString = ReadOptionalNodeStringAttribute(node, "auto_pad"); |
| 1222 | if(paddingString != "VALID" && paddingString != "" && paddingString != "NOTSET") |
| 1223 | { |
| 1224 | bool isUpper; |
| 1225 | if( paddingString == "SAME_LOWER") |
| 1226 | { |
| 1227 | isUpper = false; |
| 1228 | } |
| 1229 | else if (paddingString == "SAME_UPPER") |
| 1230 | { |
| 1231 | isUpper = true; |
| 1232 | } |
| 1233 | else |
| 1234 | { |
| 1235 | throw ParseException(boost::str( |
| 1236 | boost::format("Invalid auto_pad attribute for node %1%. " |
| 1237 | "Only SAME_UPPER, SAME_LOWER or VALID supported and found %2% %3%") |
| 1238 | % node.name() |
| 1239 | % paddingString |
| 1240 | % CHECK_LOCATION().AsString())); |
| 1241 | } |
| 1242 | uint32_t inputHeight = inputInfo.GetShape()[2]; |
| 1243 | uint32_t inputWidth = inputInfo.GetShape()[3]; |
| 1244 | |
| 1245 | uint32_t weightHeight; |
| 1246 | uint32_t weightWidth; |
| 1247 | std::vector<uint32_t> kernel_shape = ReadOptionalNodeUint32ListAttribute(node, "kernel_shape"); |
| 1248 | if (kernel_shape.empty()) |
| 1249 | { |
| 1250 | const TensorInfo weightTensorInfo = *m_TensorsInfo[node.input(1)].m_info; |
| 1251 | weightHeight = weightTensorInfo.GetShape()[2]; |
| 1252 | weightWidth = weightTensorInfo.GetShape()[3]; |
| 1253 | } |
| 1254 | else |
| 1255 | { |
| 1256 | weightHeight = kernel_shape[0]; |
| 1257 | weightWidth = kernel_shape[1]; |
| 1258 | } |
| 1259 | CalcPadding(inputHeight, weightHeight, desc.m_StrideY, &desc.m_PadTop, &desc.m_PadBottom, isUpper); |
| 1260 | CalcPadding(inputWidth, weightWidth, desc.m_StrideX, &desc.m_PadLeft, &desc.m_PadRight, isUpper); |
| 1261 | } |
| 1262 | } |
| 1263 | else |
| 1264 | { |
| 1265 | desc.m_PadTop = pads[0]; |
| 1266 | desc.m_PadLeft = pads[1]; |
| 1267 | desc.m_PadBottom = pads[2]; |
| 1268 | desc.m_PadRight = pads[3]; |
| 1269 | } |
| 1270 | |
| 1271 | uint32_t group = ReadOptionalNodeUint32Attribute(node, "group", 1); |
| 1272 | if(group > 1) |
| 1273 | { |
| 1274 | if (group > inputInfo.GetShape()[1]) |
| 1275 | { |
| 1276 | throw ParseException( |
| 1277 | boost::str( |
| 1278 | boost::format( |
| 1279 | "Error parsing Convolution node: %1%. " |
| 1280 | "The 'group'=%2% parameter cannot be larger than the " |
| 1281 | "channel of the input shape=%3% (in NCHW format). %4%") % |
| 1282 | node.name() % |
| 1283 | group % |
| 1284 | inputInfo.GetShape()[1] % |
| 1285 | CHECK_LOCATION().AsString())); |
| 1286 | } |
| 1287 | else if (group == inputInfo.GetShape()[1]) |
| 1288 | { |
| 1289 | // we use a depthwise convolution here, because the number of groups equals to the |
| 1290 | // input channels |
| 1291 | AddConvLayerWithDepthwiseConv(node, desc); |
| 1292 | return; |
| 1293 | } |
| 1294 | else |
| 1295 | { |
| 1296 | // TODO: split the input by channels into channels/groups separate convolutions |
Jim Flynn | e242f2d | 2019-05-22 14:24:13 +0100 | [diff] [blame] | 1297 | // and concatenate the results afterwards |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1298 | throw ParseException(boost::str( |
| 1299 | boost::format("Error parsing Convolution node: %1%. " |
| 1300 | "The 'group'=%2% parameter should be 1 or be equal to the " |
| 1301 | "channel of the input shape=%3% (in NCHW format). %4%") % |
| 1302 | node.name() % |
| 1303 | group % |
| 1304 | inputInfo.GetShape()[1] % |
| 1305 | CHECK_LOCATION().AsString())); |
| 1306 | } |
| 1307 | } |
| 1308 | |
| 1309 | armnn::IConnectableLayer* layer; |
| 1310 | auto weightTensor = CreateConstTensor(node.input(1)); |
| 1311 | |
| 1312 | if (node.input_size() == 3) |
| 1313 | { |
| 1314 | if(!m_TensorsInfo[node.input(2)].isConstant()) |
| 1315 | { |
| 1316 | throw ParseException(boost::str( |
| 1317 | boost::format("Bias '%1%' should be constant in Conv layer '%2%' %3%") |
| 1318 | % node.input(2) |
| 1319 | % node.name() |
| 1320 | % CHECK_LOCATION().AsString())); |
| 1321 | } |
| 1322 | desc.m_BiasEnabled = true; |
| 1323 | auto biasTensor = CreateConstTensor(node.input(2)); |
| 1324 | layer = m_Network->AddConvolution2dLayer(desc, |
| 1325 | weightTensor.first, |
Matteo Martincigh | fc598e1 | 2019-05-14 10:36:13 +0100 | [diff] [blame] | 1326 | Optional<ConstTensor>(biasTensor.first), |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1327 | node.name().c_str()); |
| 1328 | } |
| 1329 | else |
| 1330 | { |
| 1331 | layer = m_Network->AddConvolution2dLayer(desc, |
| 1332 | weightTensor.first, |
Matteo Martincigh | fc598e1 | 2019-05-14 10:36:13 +0100 | [diff] [blame] | 1333 | EmptyOptional(), |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1334 | node.name().c_str()); |
| 1335 | } |
| 1336 | BOOST_ASSERT(layer != nullptr); |
| 1337 | |
| 1338 | auto outputInfo = ComputeOutputInfo({ node.output(0) }, layer, |
| 1339 | { m_TensorsInfo[node.input(0)].m_info->GetShape(), |
| 1340 | m_TensorsInfo[node.input(1)].m_info->GetShape() }); |
| 1341 | layer->GetOutputSlot(0).SetTensorInfo(outputInfo[0]); |
| 1342 | |
| 1343 | // register the input connection slots for the layer, connections are made after all layers have been created |
| 1344 | // only the tensors for the inputs are relevant, exclude the const tensors |
| 1345 | RegisterInputSlots(layer, {node.input(0)}); |
| 1346 | |
| 1347 | // register the output connection slots for the layer, connections are made after all layers have been created |
| 1348 | RegisterOutputSlots(layer, {node.output(0)}); |
| 1349 | } |
| 1350 | |
| 1351 | void OnnxParser::PrependForBroadcast(const std::string& outputName, |
| 1352 | const std::string& input0, |
| 1353 | const std::string& input1) |
| 1354 | { |
| 1355 | //input0 should be reshaped to have same number of dim as input1 |
| 1356 | TensorInfo outputTensorInfo = TensorInfo(*m_TensorsInfo[input0].m_info); |
| 1357 | |
| 1358 | TensorShape input0Shape = m_TensorsInfo[input0].m_info->GetShape(); |
| 1359 | TensorShape input1Shape = m_TensorsInfo[input1].m_info->GetShape(); |
| 1360 | |
| 1361 | uint32_t diff = input1Shape.GetNumDimensions() - input0Shape.GetNumDimensions(); |
| 1362 | std::vector<uint32_t> newShape; |
| 1363 | while(diff > 0) |
| 1364 | { |
| 1365 | newShape.push_back(1); |
| 1366 | diff--; |
| 1367 | } |
| 1368 | for (uint dim = 0; dim < input0Shape.GetNumDimensions(); ++dim) |
| 1369 | { |
| 1370 | newShape.push_back(input0Shape[dim]); |
| 1371 | } |
| 1372 | outputTensorInfo.SetShape(TensorShape(static_cast<unsigned int>(newShape.size()), newShape.data())); |
| 1373 | |
| 1374 | //add the new tensor to m_TensorsInfo |
| 1375 | m_TensorsInfo[outputName] = OnnxTensor(); |
| 1376 | m_TensorsInfo[outputName].m_info = std::make_unique<TensorInfo>(outputTensorInfo); |
| 1377 | |
| 1378 | //add reshape layer if the parent was not constant... |
| 1379 | if( ! m_TensorsInfo[input0].isConstant()) |
| 1380 | { |
| 1381 | CreateReshapeLayer(input0, outputName, boost::str(boost::format("Add:reshapeOf%1%") % input0)); |
| 1382 | } |
| 1383 | else //make it constant and it will be create in Add |
| 1384 | { |
| 1385 | m_TensorsInfo[outputName].m_tensor = std::make_unique<onnx::TensorProto>(*m_TensorsInfo[input0].m_tensor); |
| 1386 | |
| 1387 | } |
| 1388 | } |
| 1389 | |
| 1390 | std::pair<std::string, std::string> OnnxParser::AddPrepareBroadcast(const std::string& input0, |
| 1391 | const std::string& input1) |
| 1392 | { |
| 1393 | std::pair<std::string, std::string> inputs = std::make_pair(input0, input1); |
| 1394 | |
| 1395 | TensorShape input0Shape = m_TensorsInfo[input0].m_info->GetShape(); |
| 1396 | TensorShape input1Shape = m_TensorsInfo[input1].m_info->GetShape(); |
| 1397 | |
| 1398 | if(input1Shape.GetNumDimensions() < input0Shape.GetNumDimensions()) |
| 1399 | { |
| 1400 | auto outputName = boost::str(boost::format("reshape_output_%1%") % input1); |
| 1401 | PrependForBroadcast(outputName, input1, input0); |
| 1402 | inputs.second = outputName; |
| 1403 | } |
| 1404 | else if(input0Shape.GetNumDimensions() < input1Shape.GetNumDimensions()) |
| 1405 | { |
| 1406 | auto outputName = boost::str(boost::format("reshape_output_%1%") % input0); |
| 1407 | PrependForBroadcast(outputName, input0, input1); |
| 1408 | inputs.first = outputName; |
| 1409 | } |
| 1410 | return inputs; |
| 1411 | } |
| 1412 | |
| 1413 | void OnnxParser::ParseAdd(const onnx::NodeProto& node) |
| 1414 | { |
| 1415 | CHECK_VALID_SIZE(static_cast<size_t>(node.input_size()), 2); |
| 1416 | CHECK_VALID_SIZE(static_cast<size_t>(node.output_size()), 1); |
| 1417 | |
| 1418 | VALID_INPUTS(node, STR_LIST(onnx::TensorProto::FLOAT)); |
| 1419 | |
| 1420 | // TODO: unify broadcast validation code across layers |
| 1421 | // tracked by: IVGCVSW-1576 |
| 1422 | |
| 1423 | // Checking broadcast compatibility : only scalar or 1D tensors |
| 1424 | auto inputs = AddPrepareBroadcast(node.input(0), node.input(1)); |
| 1425 | auto input0 = *m_TensorsInfo[inputs.first].m_info; |
| 1426 | auto input1 = *m_TensorsInfo[inputs.second].m_info; |
| 1427 | BOOST_ASSERT(input0.GetNumDimensions() == input1.GetNumDimensions()); |
| 1428 | |
| 1429 | unsigned int numDims = input0.GetNumDimensions(); |
| 1430 | for (unsigned int i = 0; i < numDims; i++) |
| 1431 | { |
| 1432 | unsigned int dim0 = input0.GetShape()[i]; |
| 1433 | unsigned int dim1 = input1.GetShape()[i]; |
| 1434 | if (dim0 != dim1 && dim0 != 1 && dim1 != 1) |
| 1435 | { |
| 1436 | throw ParseException(boost::str( |
| 1437 | boost::format("Broadcast is only supported for scalar or 1D tensors in Add node '%1%'. " |
| 1438 | "Input dimensions should either match or one should be of size 1 and here, " |
| 1439 | "%2% and %3% %4%") |
| 1440 | % node.name() |
| 1441 | % TensorInfoAsString(*m_TensorsInfo[inputs.first].m_info, inputs.first, |
| 1442 | m_TensorsInfo[inputs.first].m_dtype) |
| 1443 | % TensorInfoAsString(*m_TensorsInfo[inputs.second].m_info, inputs.second, |
| 1444 | m_TensorsInfo[inputs.second].m_dtype) |
| 1445 | % CHECK_LOCATION().AsString())); |
| 1446 | } |
| 1447 | } |
| 1448 | |
| 1449 | |
| 1450 | IConnectableLayer* layer = m_Network->AddAdditionLayer(node.name().c_str()); |
| 1451 | BOOST_ASSERT(layer != nullptr); |
| 1452 | |
| 1453 | auto outputInfo = ComputeOutputInfo({ node.output(0) }, layer, |
| 1454 | { m_TensorsInfo[inputs.first].m_info->GetShape(), |
| 1455 | m_TensorsInfo[inputs.second].m_info->GetShape() }); |
| 1456 | layer->GetOutputSlot(0).SetTensorInfo(outputInfo[0]); |
| 1457 | |
| 1458 | // register the input connection -> for constant inputs, we need to make a newDim constant layer |
| 1459 | if(m_TensorsInfo[inputs.first].isConstant()) { |
| 1460 | |
| 1461 | CreateConstantLayer(inputs.first, boost::str(boost::format("Add:constant_of_%1%") % node.input(0))); |
| 1462 | } |
| 1463 | if(m_TensorsInfo[inputs.second].isConstant()) { |
| 1464 | |
| 1465 | CreateConstantLayer(inputs.second, boost::str(boost::format("Add:constant_of_%1%") % node.input(1))); |
| 1466 | } |
| 1467 | RegisterInputSlots(layer, {inputs.first, inputs.second}); |
| 1468 | |
| 1469 | // register the output connection |
| 1470 | RegisterOutputSlots(layer, {node.output(0)}); |
| 1471 | } |
| 1472 | |
| 1473 | void OnnxParser::ParseBatchNormalization(const onnx::NodeProto& node) |
| 1474 | { |
| 1475 | //IGNORE momentum parameter and spatial parameters |
| 1476 | |
| 1477 | CHECK_VALID_SIZE(static_cast<size_t>(node.input_size()), 5); |
| 1478 | CHECK_VALID_SIZE(static_cast<size_t>(node.output_size()), 1); |
| 1479 | |
| 1480 | VALID_INPUTS(node, STR_LIST(onnx::TensorProto::FLOAT)); |
| 1481 | for(int ind = 1; ind < node.input_size(); ++ind) |
| 1482 | { |
| 1483 | auto tensor = node.input(ind); |
| 1484 | if(! m_TensorsInfo[tensor].isConstant()) |
| 1485 | { |
| 1486 | throw ParseException(boost::str( |
| 1487 | boost::format("Input tensor '%1%' should be constant in BatchNormalization node '%2%' %3%") |
| 1488 | % tensor |
| 1489 | % node.name() |
| 1490 | % CHECK_LOCATION().AsString())); |
| 1491 | } |
| 1492 | } |
| 1493 | |
| 1494 | float epsilon = ReadOptionalNodeFloatAttribute(node, "epsilon", 1e-5f); |
| 1495 | BatchNormalizationDescriptor desc; |
| 1496 | desc.m_Eps = epsilon; |
| 1497 | |
| 1498 | auto scaleTensor = CreateConstTensor(node.input(1)); |
| 1499 | auto biasTensor = CreateConstTensor(node.input(2)); |
| 1500 | auto meanTensor = CreateConstTensor(node.input(3)); |
| 1501 | auto varTensor = CreateConstTensor(node.input(4)); |
| 1502 | |
| 1503 | IConnectableLayer* layer = m_Network->AddBatchNormalizationLayer(desc, |
| 1504 | meanTensor.first, |
| 1505 | varTensor.first, |
| 1506 | biasTensor.first, |
| 1507 | scaleTensor.first, |
| 1508 | node.name().c_str()); |
| 1509 | BOOST_ASSERT(layer != nullptr); |
| 1510 | |
| 1511 | auto outputInfo = ComputeOutputInfo({node.output(0)}, layer, {m_TensorsInfo[node.input(0)].m_info->GetShape()}); |
| 1512 | layer->GetOutputSlot(0).SetTensorInfo(outputInfo[0]); |
| 1513 | |
| 1514 | RegisterInputSlots(layer, {node.input(0)}); //don't register constant inputs |
| 1515 | |
| 1516 | // register the output connection |
| 1517 | RegisterOutputSlots(layer, {node.output(0)}); |
| 1518 | } |
| 1519 | |
| 1520 | void OnnxParser::SetupInputLayers() |
| 1521 | { |
| 1522 | //Find user input and add their layers |
| 1523 | for(int inputIndex = 0; inputIndex < m_Graph->input_size(); ++inputIndex) |
| 1524 | { |
| 1525 | auto input = m_Graph->input(inputIndex); |
| 1526 | if (! m_TensorsInfo[input.name()].isConstant()) |
| 1527 | { |
| 1528 | IConnectableLayer* layer = |
| 1529 | m_Network->AddInputLayer(static_cast<armnn::LayerBindingId>(inputIndex), input.name().c_str()); |
| 1530 | auto tensorInfo = ToTensorInfo(input); |
| 1531 | layer->GetOutputSlot(0).SetTensorInfo(tensorInfo); |
| 1532 | |
| 1533 | RegisterOutputSlots(layer,{ input.name() }); |
| 1534 | } |
| 1535 | } |
| 1536 | } |
| 1537 | |
| 1538 | void OnnxParser::SetupOutputLayers() |
| 1539 | { |
| 1540 | if(m_Graph->output_size() == 0) |
| 1541 | { |
| 1542 | throw ParseException(boost::str(boost::format("The given model does not have any outputs %1%") |
| 1543 | % CHECK_LOCATION().AsString())); |
| 1544 | } |
| 1545 | |
| 1546 | for(int outputIndex = 0; outputIndex < m_Graph->output_size(); ++outputIndex) |
| 1547 | { |
| 1548 | IConnectableLayer* layer = |
| 1549 | m_Network->AddOutputLayer(static_cast<armnn::LayerBindingId>(outputIndex), |
| 1550 | m_Graph->output(outputIndex).name().c_str()); |
| 1551 | |
| 1552 | RegisterInputSlots(layer, { m_Graph->output(outputIndex).name() }); |
| 1553 | } |
| 1554 | } |
| 1555 | |
| 1556 | void OnnxParser::RegisterInputSlots(IConnectableLayer* layer, const std::vector<std::string>& tensorIds) |
| 1557 | { |
| 1558 | BOOST_ASSERT(layer != nullptr); |
| 1559 | if (tensorIds.size() != layer->GetNumInputSlots()) |
| 1560 | { |
| 1561 | throw ParseException( |
| 1562 | boost::str(boost::format("The number of tensor inputs (%1%) does not match the number expected (%2%) %3%") % |
| 1563 | tensorIds.size() % |
| 1564 | layer->GetNumInputSlots() % |
| 1565 | CHECK_LOCATION().AsString())); |
| 1566 | } |
| 1567 | for (unsigned int slotIndex = 0; slotIndex < layer->GetNumInputSlots(); ++slotIndex) |
| 1568 | { |
| 1569 | std::string tensorId = tensorIds[slotIndex]; |
| 1570 | armnn::IInputSlot* slot = &(layer->GetInputSlot(slotIndex)); |
| 1571 | |
| 1572 | auto it = m_TensorConnections.find(tensorId); |
| 1573 | |
| 1574 | if (it == m_TensorConnections.end()) |
| 1575 | { |
| 1576 | //First time seing this tensor, we need to map it |
| 1577 | m_TensorConnections[tensorId] = TensorSlots(); |
| 1578 | } |
| 1579 | m_TensorConnections[tensorId].inputSlots.push_back(slot); |
| 1580 | } |
| 1581 | } |
| 1582 | |
| 1583 | void OnnxParser::RegisterOutputSlots(IConnectableLayer* layer, const std::vector<std::string>& tensorIds) |
| 1584 | { |
| 1585 | BOOST_ASSERT(layer != nullptr); |
| 1586 | if (tensorIds.size() != layer->GetNumOutputSlots()) |
| 1587 | { |
| 1588 | throw ParseException( |
| 1589 | boost::str(boost::format("The number of tensor outputs (%1%) does not match the number expected (%2%) %3% ") |
| 1590 | % tensorIds.size() |
| 1591 | % layer->GetNumOutputSlots() |
| 1592 | % CHECK_LOCATION().AsString())); |
| 1593 | } |
| 1594 | |
| 1595 | for (unsigned int slotIndex = 0; slotIndex < layer->GetNumOutputSlots(); ++slotIndex) |
| 1596 | { |
| 1597 | std::string tensorId = tensorIds[slotIndex]; |
| 1598 | armnn::IOutputSlot* slot = &(layer->GetOutputSlot(slotIndex)); |
| 1599 | |
| 1600 | auto it = m_TensorConnections.find(tensorId); |
| 1601 | |
| 1602 | if (it == m_TensorConnections.end()) |
| 1603 | { |
| 1604 | //First time seing this tensor, we need to map it |
| 1605 | m_TensorConnections[tensorId] = TensorSlots(); |
| 1606 | } |
| 1607 | |
| 1608 | TensorSlots & tensorSlots = m_TensorConnections[tensorId]; |
| 1609 | |
| 1610 | // assuming there is only one producer for that tensor |
| 1611 | if (tensorSlots.outputSlot != nullptr) |
| 1612 | { |
| 1613 | throw ParseException(boost::str( |
| 1614 | boost::format("Another layer has already registered itself as the producer of " |
| 1615 | "tensor:%2% %3%") % |
| 1616 | tensorId % |
| 1617 | CHECK_LOCATION().AsString())); |
| 1618 | } |
| 1619 | tensorSlots.outputSlot = slot; |
| 1620 | } |
| 1621 | } |
| 1622 | |
| 1623 | BindingPointInfo OnnxParser::GetNetworkInputBindingInfo(const std::string& name) const |
| 1624 | { |
| 1625 | for(int i = 0; i < m_Graph->input_size(); ++i) |
| 1626 | { |
| 1627 | auto input = m_Graph->input(i); |
| 1628 | if(input.name() == name) |
| 1629 | { |
| 1630 | return std::make_pair(static_cast<armnn::LayerBindingId>(i), ToTensorInfo(input)); |
| 1631 | } |
| 1632 | } |
| 1633 | throw InvalidArgumentException(boost::str(boost::format("The input layer '%1%' does not exist %2%") |
| 1634 | % name % CHECK_LOCATION().AsString())); |
| 1635 | } |
| 1636 | |
| 1637 | BindingPointInfo OnnxParser::GetNetworkOutputBindingInfo(const std::string& name) const |
| 1638 | { |
| 1639 | for(int i = 0; i < m_Graph->output_size(); ++i) |
| 1640 | { |
| 1641 | auto output = m_Graph->output(i); |
| 1642 | if(output.name() == name) |
| 1643 | { |
| 1644 | return std::make_pair(static_cast<armnn::LayerBindingId>(i), ToTensorInfo(output)); |
| 1645 | } |
| 1646 | } |
| 1647 | throw InvalidArgumentException(boost::str(boost::format("The output layer '%1%' does not exist %2%") |
| 1648 | % name % CHECK_LOCATION().AsString())); |
| 1649 | } |
| 1650 | |
| 1651 | std::vector<std::string> OnnxParser::GetInputs(ModelPtr& model) |
| 1652 | { |
| 1653 | if(model == nullptr) { |
| 1654 | throw InvalidArgumentException(boost::str( |
| 1655 | boost::format("The given model cannot be null %1%") |
| 1656 | % CHECK_LOCATION().AsString())); |
| 1657 | } |
| 1658 | |
| 1659 | std::vector<std::string> inputNames; |
| 1660 | std::map<std::string, bool> isConstant; |
| 1661 | for(auto tensor : model->graph().initializer()) |
| 1662 | { |
| 1663 | isConstant[tensor.name()] = true; |
| 1664 | } |
| 1665 | for(auto input : model->graph().input()) |
| 1666 | { |
| 1667 | auto it = isConstant.find(input.name()); |
| 1668 | if(it == isConstant.end()) |
| 1669 | { |
| 1670 | inputNames.push_back(input.name()); |
| 1671 | } |
| 1672 | } |
| 1673 | return inputNames; |
| 1674 | } |
| 1675 | |
| 1676 | std::vector<std::string> OnnxParser::GetOutputs(ModelPtr& model) |
| 1677 | { |
| 1678 | if(model == nullptr) { |
| 1679 | throw InvalidArgumentException(boost::str( |
| 1680 | boost::format("The given model cannot be null %1%") |
| 1681 | % CHECK_LOCATION().AsString())); |
| 1682 | } |
| 1683 | |
| 1684 | std::vector<std::string> outputNames; |
| 1685 | for(auto output : model->graph().output()) |
| 1686 | { |
| 1687 | outputNames.push_back(output.name()); |
| 1688 | } |
| 1689 | return outputNames; |
| 1690 | } |
| 1691 | |
| 1692 | } // namespace armnnOnnxParser |