RModelParser_ONNX.cxx

Go to the documentation of this file.

#include "Byteswap.h"
#include "TMVA/RModelParser_ONNX.hxx"
// The operator base class is a private header: RModelParser_ONNX.hxx only
// forward-declares it, but this translation unit manages ROperator instances
// through std::unique_ptr and needs the complete type.
#include "TMVA/ROperator.hxx"
#include "onnx.hxx"
 
#include <algorithm>
#include <stdexcept>
#include <string>
#include <cstring>
#include <memory>
#include <cassert>
#include <iostream>
#include <unordered_map>
#include <functional>
#include "TMVA/SOFIE_common.hxx"
 
namespace TMVA {
namespace Experimental {
namespace SOFIE {
 
// Declaration of operators
// Unary operators
void RegisterBasicUnaryParsers(RModelParser_ONNX &parser);
// Binary operators
void RegisterBasicBinaryParsers(RModelParser_ONNX &parser);
// Nary operators
void RegisterBasicNaryParsers(RModelParser_ONNX &parser);
//Comparision Operators
void RegisterComparisionParsers(RModelParser_ONNX &parser);
//Is Operators
void RegisterBasicIsParsers(RModelParser_ONNX &parser);
extern ParserFuncSignature ParseNot;
// Reduce operators
void RegisterReduceParsers(RModelParser_ONNX &parser);
// Others
extern ParserFuncSignature ParseBatchNormalization;
extern ParserFuncSignature ParseConstant;
extern ParserFuncSignature ParseTranspose;
extern ParserFuncSignature ParseRelu;
extern ParserFuncSignature ParseTanh;
extern ParserFuncSignature ParseConv;
extern ParserFuncSignature ParseConvTranspose;
extern ParserFuncSignature ParseLeakyRelu;
extern ParserFuncSignature ParseGelu;
extern ParserFuncSignature ParseSelu;
extern ParserFuncSignature ParseSigmoid;
extern ParserFuncSignature ParseSwish;
extern ParserFuncSignature ParseGemm;
extern ParserFuncSignature ParseRNN;
extern ParserFuncSignature ParseLSTM;
extern ParserFuncSignature ParsePool;
extern ParserFuncSignature ParseReshape;
extern ParserFuncSignature ParseSlice;
extern ParserFuncSignature ParseGRU;
extern ParserFuncSignature ParseIdentity;
extern ParserFuncSignature ParseSoftmax;
extern ParserFuncSignature ParseConcat;
extern ParserFuncSignature ParseCast;
extern ParserFuncSignature ParseExpand;
extern ParserFuncSignature ParseShape;
extern ParserFuncSignature ParseMatMul;
extern ParserFuncSignature ParseLayerNormalization;
extern ParserFuncSignature ParseInstanceNormalization;
extern ParserFuncSignature ParseGather;
extern ParserFuncSignature ParseGatherND;
extern ParserFuncSignature ParseErf;
extern ParserFuncSignature ParseElu;
extern ParserFuncSignature ParseHardSigmoid;
extern ParserFuncSignature ParseHardSwish;
extern ParserFuncSignature ParseEyeLike;
extern ParserFuncSignature ParseRange;
extern ParserFuncSignature ParseTopK;
extern ParserFuncSignature ParseTile;
extern ParserFuncSignature ParseSplit;
extern ParserFuncSignature ParseIf;
extern ParserFuncSignature ParsePad;
extern ParserFuncSignature ParseWhere;
extern ParserFuncSignature ParseEinsum;
extern ParserFuncSignature ParseRandom;
extern ParserFuncSignature ParseScatterElements;
extern ParserFuncSignature ParseScatterND;
extern ParserFuncSignature ParseNonZero;
extern ParserFuncSignature ParseClip;
extern ParserFuncSignature ParseArgMax;
extern ParserFuncSignature ParsePRelu;
// Declaration of fused operators
extern ParserFuseFuncSignature ParseFuseConvAdd;
extern ParserFuseFuncSignature ParseFuseGemmRelu;
extern ParserFuseFuncSignature ParseFuseBatchnormRelu;
extern ParserFuseFuncSignature ParseFuseConvTransposeAdd;
extern ParserFuseFuncSignature ParseFuseMatMulAdd;
 
// Definition of  RModelParser_ONNX::OperatorsMap
struct RModelParser_ONNX::OperatorsMapImpl {
   // Registered operators
   std::unordered_map<std::string, ParserFuncSignature> fOperatorsMap;
};
 
// helper function to get initialized tensor data
template<typename T>
struct ExtractDataFromTP {
};
// trait function to extract data from TensorProto
template<>
struct ExtractDataFromTP<float> {
   static void Copy(onnx::TensorProto * tensor, void * data, int length) {
      if (tensor->float_data_size() != length)
         throw std::runtime_error("TMVA::SOFIE - Failed to read float initialized tensor - actual size is " + std::to_string(tensor->float_data_size()));
      const auto &src = tensor->float_data();
      std::copy(src.begin(), src.end(), static_cast<float *>(data));
   }
};
template<>
struct ExtractDataFromTP<double> {
   static void Copy(onnx::TensorProto * tensor, void * data, int length) {
      if (tensor->double_data_size() != length)
         throw std::runtime_error("TMVA::SOFIE - Failed to read double initialized tensor - actual size is " + std::to_string(tensor->double_data_size()));
      const auto &src = tensor->double_data();
      std::copy(src.begin(), src.end(), static_cast<double *>(data));
   }
};
template<>
struct ExtractDataFromTP<int32_t> {
   static void Copy(onnx::TensorProto * tensor, void * data, int length) {
      if (tensor->int32_data_size() != length)
         throw std::runtime_error("TMVA::SOFIE - Failed to read int32 initialized tensor - actual size is " + std::to_string(tensor->int32_data_size()));
      const auto &src = tensor->int32_data();
      std::copy(src.begin(), src.end(), static_cast<int32_t *>(data));
   }
};
template<>
struct ExtractDataFromTP<int64_t> {
   static void Copy(onnx::TensorProto * tensor, void * data, int length) {
      if (tensor->int64_data_size() != length)
         throw std::runtime_error("TMVA::SOFIE - Failed to read int64 initialized tensor - actual size is " + std::to_string(tensor->int64_data_size()));
      const auto &src = tensor->int64_data();
      std::copy(src.begin(), src.end(), static_cast<int64_t *>(data));
   }
};
 
#ifndef R__BYTESWAP
namespace {
 
// Copy nbytes from source to dest, byte-swapping each N-byte element. The
// temporary avoids misaligned loads from the protobuf string buffer and makes
// in-place swapping (dest == source) valid.
template <std::size_t N>
void CopyBswap(void *dest, const void *source, std::size_t nbytes)
{
   using value_type = typename RByteSwap<N>::value_type;
   auto dst = static_cast<unsigned char *>(dest);
   auto src = static_cast<const unsigned char *>(source);
   for (std::size_t k = 0; k < nbytes; k += N) {
      value_type v;
      std::memcpy(&v, src + k, N);
      v = RByteSwap<N>::bswap(v);
      std::memcpy(dst + k, &v, N);
   }
}
 
// Copy a buffer of little-endian tensor elements to host (big-endian) byte order
void CopyLEToHost(void *dest, const void *source, std::size_t nbytes, ETensorType tensor_type)
{
   switch (GetTypeSize(tensor_type)) {
   case 1:
      if (dest != source)
         std::memcpy(dest, source, nbytes);
      break;
   case 2: CopyBswap<2>(dest, source, nbytes); break;
   case 4: CopyBswap<4>(dest, source, nbytes); break;
   case 8: CopyBswap<8>(dest, source, nbytes); break;
   default:
      throw std::runtime_error("Data type " + ConvertTypeToString(tensor_type) + " in tensor is not supported!\n");
   }
}
 
} // anonymous namespace
#endif
 
std::shared_ptr<void> RModelParser_ONNX::GetInitializedTensorData(onnx::TensorProto *tensorproto, size_t tensor_size, ETensorType tensor_type)
{
 
   std::shared_ptr<void> data(malloc(tensor_size), free);
 
   // check if initialized tensors are stored internally
   if (tensorproto->data_location() != onnx::TensorProto::EXTERNAL) {
      if (tensorproto->raw_data().size() > 0) {
         if (tensorproto->raw_data().size() != tensor_size)
            throw std::runtime_error("TMVA::SOFIE - Failed to read raw data of initialized tensor - actual raw size is " +
                                 std::to_string(tensorproto->raw_data().size()));
 
#ifdef R__BYTESWAP
         // R__BYTESWAP is defined for little-endian architectures (most common ones)
         std::memcpy(data.get(), tensorproto->raw_data().c_str(), tensor_size);
#else
         // big-endian architectures - need to swap bytes
         CopyLEToHost(data.get(), tensorproto->raw_data().c_str(), tensor_size, tensor_type);
#endif
      } else {
         // case tensor data are stored as specific types and not in raw_data
         switch (tensor_type) {
            case ETensorType::FLOAT: {
               ExtractDataFromTP<float>::Copy(tensorproto, data.get(), tensor_size/ 4);
               break;
            }
            case ETensorType::DOUBLE: {
               ExtractDataFromTP<double>::Copy(tensorproto, data.get(), tensor_size/ 8);
               break;
            }
            case ETensorType::INT32: {
               ExtractDataFromTP<int32_t>::Copy(tensorproto, data.get(), tensor_size/ 4);
               break;
            }
            case ETensorType::INT64: {
               ExtractDataFromTP<int64_t>::Copy(tensorproto, data.get(), tensor_size/ 8);
               break;
            }
            case ETensorType::BOOL: {
               throw std::runtime_error("TMVA::SOFIE - ExtractData from TP in BOOL not supported");
               break;
            }
            case ETensorType::UINT8: {
               throw std::runtime_error("TMVA::SOFIE - ExtractData from TP in UINT8 not supported");
               break;
            }
            default:
               throw std::runtime_error("Data type " + ConvertTypeToString(tensor_type) + " in weight tensor is not supported!\n");
         }
      }
 
   }  else {
      // case of external data
 
      // read now tensor from file
      std::string location;
      size_t offset = 0, buffer_size = 0;
 
      for (const auto &kv : tensorproto->external_data()) {
         if (kv.key() == "location")  location = kv.value();
         else if (kv.key() == "offset") offset = std::stoull(kv.value());
         else if (kv.key() == "length") buffer_size = std::stoull(kv.value());
      }
 
      // an explicitly set data file (SetExternalDataFile) takes precedence;
      // otherwise use the location stored in the model, which is a path
      // relative to the model directory, and as a last resort the
      // conventional <model file>.data
      std::string dataFileName = fDataFileName;
      if (dataFileName.empty())
         dataFileName = location.empty() ? fDefaultDataFileName : fModelDirectory + location;
      if (dataFileName.empty())
         throw std::runtime_error("TMVA::SOFIE ONNX : tensor " + tensorproto->name() +
                                  " has external data but no data file location is available");
 
      if (fVerbose)
         std::cout << "Initialized data are stored externally in file " << dataFileName
                   << " at location " << location << " offset " << offset << " and with length " << buffer_size << std::endl;
 
      if (buffer_size != tensor_size)
         throw std::runtime_error("TMVA::SOFIE ONNX : invalid stored data size vs tensor size");
 
      // open the data file if needed (a previous tensor may have opened a different one)
      if (fDataFile.is_open() && fOpenedDataFileName != dataFileName)
         fDataFile.close();
      if (!fDataFile.is_open()) {
         fDataFile.open(dataFileName, std::ios::binary);
         if (!fDataFile.is_open())
            throw std::runtime_error("TMVA::SOFIE ONNX:  error reading external weight ONNX data file " + dataFileName);
         fOpenedDataFileName = dataFileName;
      }
 
      fDataFile.seekg(offset);
      fDataFile.read(reinterpret_cast<char *>(data.get()), buffer_size);
#ifndef R__BYTESWAP
      // external data is stored little-endian like raw_data - swap in place
      CopyLEToHost(data.get(), data.get(), buffer_size, tensor_type);
#endif
   }
 
   return data;
}
 
 
// Constructor of the parser
RModelParser_ONNX::RModelParser_ONNX() noexcept : fOperatorsMapImpl(std::make_unique<OperatorsMapImpl>()) {
   // Register operators
   // Unary operators
   RegisterBasicUnaryParsers(*this);
   // Binary operators
   RegisterBasicBinaryParsers(*this);
   // Nary operators
   RegisterBasicNaryParsers(*this);
   //Comparision Operators
   RegisterComparisionParsers(*this);
   // Is If operators
   RegisterBasicIsParsers(*this);
   RegisterOperator("Not", ParseNot);
   // Reduce operators
   RegisterReduceParsers(*this);
   // Others
   RegisterOperator("BatchNormalization", ParseBatchNormalization);
   RegisterOperator("Constant", ParseConstant);
   RegisterOperator("ConstantOfShape", ParseConstant);
   RegisterOperator("Cast", ParseCast);
   RegisterOperator("Concat", ParseConcat);
   RegisterOperator("Conv", ParseConv);
   RegisterOperator("ConvTranspose", ParseConvTranspose);
   RegisterOperator("Gemm", ParseGemm);
   RegisterOperator("GRU", ParseGRU);
   RegisterOperator("Identity", ParseIdentity);
   RegisterOperator("LeakyRelu", ParseLeakyRelu);
   RegisterOperator("LSTM", ParseLSTM);
   RegisterOperator("AveragePool", ParsePool);
   RegisterOperator("GlobalAveragePool", ParsePool);
   RegisterOperator("MaxPool", ParsePool);
   RegisterOperator("Relu", ParseRelu);
   RegisterOperator("Reshape", ParseReshape);
   RegisterOperator("Flatten", ParseReshape);
   RegisterOperator("Squeeze", ParseReshape);
   RegisterOperator("Unsqueeze", ParseReshape);
   RegisterOperator("RNN", ParseRNN);
   RegisterOperator("Gelu", ParseGelu);
   RegisterOperator("Selu", ParseSelu);
   RegisterOperator("Shape", ParseShape);
   RegisterOperator("Sigmoid", ParseSigmoid);
   RegisterOperator("Swish", ParseSwish);
   RegisterOperator("Slice", ParseSlice);
   RegisterOperator("Softmax", ParseSoftmax);
   RegisterOperator("LogSoftmax", ParseSoftmax);
   RegisterOperator("Tanh", ParseTanh);
   RegisterOperator("Transpose", ParseTranspose);
   RegisterOperator("MatMul", ParseMatMul);
   RegisterOperator("LayerNormalization", ParseLayerNormalization);
   RegisterOperator("Expand", ParseExpand);
   RegisterOperator("Gather", ParseGather);
   RegisterOperator("GatherND", ParseGatherND);
   RegisterOperator("Erf", ParseErf);
   RegisterOperator("Elu", ParseElu);
   RegisterOperator("HardSigmoid", ParseHardSigmoid);
   RegisterOperator("HardSwish", ParseHardSwish);
   RegisterOperator("EyeLike", ParseEyeLike);
   RegisterOperator("Range", ParseRange);
   RegisterOperator("TopK", ParseTopK);
   RegisterOperator("Tile", ParseTile);
   RegisterOperator("Split", ParseSplit);
   RegisterOperator("If", ParseIf);
   RegisterOperator("InstanceNormalization", ParseInstanceNormalization);
   RegisterOperator("Pad", ParsePad);
   RegisterOperator("Where", ParseWhere);
   RegisterOperator("Einsum", ParseEinsum);
   RegisterOperator("RandomNormal", ParseRandom);
   RegisterOperator("RandomNormalLike", ParseRandom);
   RegisterOperator("RandomUniform", ParseRandom);
   RegisterOperator("RandomUniformLike", ParseRandom);
   RegisterOperator("ScatterElements", ParseScatterElements);
   RegisterOperator("ScatterND", ParseScatterND);
   RegisterOperator("NonZero", ParseNonZero);
   RegisterOperator("Clip", ParseClip);
   RegisterOperator("ArgMax", ParseArgMax);
   RegisterOperator("PRelu", ParsePRelu);
}
 
// Destructor of the parser
RModelParser_ONNX::~RModelParser_ONNX() = default;
 
void RModelParser_ONNX::RegisterOperator(const std::string &name, ParserFuncSignature func)
{
   fOperatorsMapImpl->fOperatorsMap[name] = func;
}
 
bool RModelParser_ONNX::IsRegisteredOperator(const std::string &name)
{
   return fOperatorsMapImpl->fOperatorsMap.find(name) != fOperatorsMapImpl->fOperatorsMap.end();
}
 
std::vector<std::string> RModelParser_ONNX::GetRegisteredOperators()
{
   std::vector<std::string> ops;
   ops.reserve(fOperatorsMapImpl->fOperatorsMap.size());
   for (auto &it : fOperatorsMapImpl->fOperatorsMap) {
      ops.emplace_back(it.first);
   }
   // return sorted list in alphabetical order
   std::sort(ops.begin(), ops.end());
   return ops;
}
 
void RModelParser_ONNX::RegisterTensorType(const std::string &name, ETensorType type)
{
   fTensorTypeMap[UTILITY::Clean_name(name)] = type;
}
 
bool RModelParser_ONNX::IsRegisteredTensorType(const std::string &name)
{
   return fTensorTypeMap.find(UTILITY::Clean_name(name)) != fTensorTypeMap.end();
}
 
ETensorType RModelParser_ONNX::GetTensorType(const std::string &name)
{
   return fTensorTypeMap[UTILITY::Clean_name(name)];
}
 
namespace {
 
/// Is the Add following a Conv / ConvTranspose really that convolution's bias?
///
/// Only if the convolution has no bias yet and the added tensor is a rank-1 initializer, one
/// value per output channel. Anything else - a residual connection, an operand computed at
/// run time - is a genuine addition.
bool IsConvBiasAdd(const onnx::GraphProto &graph, const onnx::NodeProto &convnode, const onnx::NodeProto &addnode)
{
   if (convnode.input_size() > 2 || addnode.input_size() != 2)
      return false;
   const std::string &added = (addnode.input(0) == convnode.output(0)) ? addnode.input(1) : addnode.input(0);
   for (int i = 0; i < graph.initializer_size(); i++) {
      if (graph.initializer(i).name() == added)
         return graph.initializer(i).dims_size() == 1;
   }
   return false;
}
 
} // namespace
 
// Parse an operator
std::unique_ptr<ROperator>
RModelParser_ONNX::ParseOperator(const size_t i, const onnx::GraphProto &graphproto, const std::vector<size_t> &nodes, const std::vector<int> & children)
{
   if (i >= nodes.size())
      throw std::runtime_error("TMVA::SOFIE - Error in parsing ordered operators " + std::to_string(i) + " is >=  " + std::to_string(nodes.size()));
   int idx = nodes[i];
   const auto &nodeproto = graphproto.node(idx);
   const std::string op_type = nodeproto.op_type();
   if (fVerbose)
      std::cout << "Parsing operator " << op_type << std::endl;
 
   // perform the fusion of  operators
   if (fFusedOperators.count(idx) == 1) {
      int idx1 = fFusedOperators[idx].second;
      if (fVerbose) {
         std::cout << "\tFusing operators " << graphproto.node(idx1).name()
                   << " with  " <<  graphproto.node(idx1).name() << std::endl;
      }
      if (fFusedOperators[idx].first == EFusedOp::kMatMulAdd) {
         return ParseFuseMatMulAdd(*this, graphproto.node(idx1), graphproto.node(idx));
      } else if (fFusedOperators[idx].first == EFusedOp::kConvAdd) {
         return ParseFuseConvAdd(*this, graphproto.node(idx1), graphproto.node(idx));
      } else if (fFusedOperators[idx].first == EFusedOp::kConvTransAdd) {
         return ParseFuseConvTransposeAdd(*this, graphproto.node(idx1), graphproto.node(idx));
      } else if (fFusedOperators[idx].first == EFusedOp::kGemmRelu) {
         return ParseFuseGemmRelu(*this, graphproto.node(idx1), graphproto.node(idx));
      } else if (fFusedOperators[idx].first == EFusedOp::kBatchnormRelu) {
         return ParseFuseBatchnormRelu(*this, graphproto.node(idx1), graphproto.node(idx));
      }
   }
 
   // try to fuse with following operator in case it is not last one and having only a single child
   if (children.size() == 1) {
      int idx2 = children.front();
      if (op_type == "MatMul") {
        // Fuse MatMul and Add
         if (idx2 < graphproto.node_size() && graphproto.node(idx2).op_type() == "Add") {
            fFusedOperators[idx2] = {EFusedOp::kMatMulAdd, idx};
            return nullptr;
         }
      } else if (nodeproto.op_type() == "Conv" || nodeproto.op_type() == "ConvTranspose") {
         // Fuse Conv or ConvTranspose without bias and Add, when the Add really is the bias
         if (idx2 < graphproto.node_size() && graphproto.node(idx2).op_type() == "Add" &&
             IsConvBiasAdd(graphproto, nodeproto, graphproto.node(idx2))) {
            if (nodeproto.op_type() == "Conv") {
               fFusedOperators[idx2] = { EFusedOp::kConvAdd, idx};
               return nullptr;
            } else {
               fFusedOperators[idx2] = { EFusedOp::kConvTransAdd, idx};
               return nullptr;
            }
         }
      } else if (nodeproto.op_type() == "Gemm") {
         // Fuse Gemm with activation operators
         if (idx2 < graphproto.node_size() && graphproto.node(idx2).op_type() == "Relu") {
            fFusedOperators[idx2] = {EFusedOp::kGemmRelu, idx};
            return nullptr;
         }
      } else if (nodeproto.op_type() == "BatchNormalization") {
         if (idx2 < graphproto.node_size() && graphproto.node(idx2).op_type() == "Relu") {
            fFusedOperators[idx2] = {EFusedOp::kBatchnormRelu, idx};
            return nullptr;
         }
      }
   }
 
   auto it = fOperatorsMapImpl->fOperatorsMap.find(op_type);
   if (it == fOperatorsMapImpl->fOperatorsMap.end()) {
      std::cout << "operator " << op_type << " is not supported" << std::endl;
      throw std::runtime_error("TMVA::SOFIE Operator type " + op_type + " is not yet supported");
   }
   if (fVerbose) {
      std::cout << "\tCreating operator " << op_type << std::endl;
   }
   return it->second(*this, nodeproto);
}
 
// Parse a model
RModel RModelParser_ONNX::Parse(std::string const &filename, bool verbose)
{
   fVerbose = verbose;
 
   fTensorTypeMap.clear();
 
   auto model = LoadModel(filename);
   if (!model)
      throw std::runtime_error("TMVA::SOFIE - Failed to load onnx file " + filename);
 
   const onnx::GraphProto &graph = model->graph(); // not a memory leak. model freed automatically at the end.
 
 
   std::time_t ttime = std::time(0);
   std::tm *gmt_time = std::gmtime(&ttime);
   std::string parsetime(std::asctime(gmt_time));
 
   // get name of model (filename without directory name)
   char sep = '/';
#ifdef _WIN32
   sep = '\\';
#endif
   size_t isep = filename.rfind(sep, filename.length());
   std::string filename_nodir = filename;
   if (isep != std::string::npos) {
      filename_nodir = (filename.substr(isep + 1, filename.length() - isep));
   }
 
   fModelDirectory = (isep != std::string::npos) ? filename.substr(0, isep + 1) : "";
   fDefaultDataFileName = filename + ".data";
 
   RModel rmodel(filename_nodir, parsetime);
   ParseONNXGraph(rmodel, graph, filename_nodir);
   ResetExternalDataState();
   return rmodel;
}
 
RModel RModelParser_ONNX::Parse(std::istream &input, std::string const &name, bool verbose)
{
   fVerbose = verbose;
 
   fTensorTypeMap.clear();
 
   auto model = LoadModel(input);
   if (!model)
      throw std::runtime_error("TMVA::SOFIE - Failed to parse ONNX model from input stream");
 
   const onnx::GraphProto &graph = model->graph(); // not a memory leak. model freed automatically at the end.
 
   std::time_t ttime = std::time(0);
   std::tm *gmt_time = std::gmtime(&ttime);
   std::string parsetime(std::asctime(gmt_time));
 
   RModel rmodel(name, parsetime);
   ParseONNXGraph(rmodel, graph, name);
   ResetExternalDataState();
   return rmodel;
}
 
// Reset the state used to read external weight data, so that the next Parse
// call does not pick up the data file of a previously parsed model. The
// file name set with SetExternalDataFile is valid for a single Parse call.
void RModelParser_ONNX::ResetExternalDataState()
{
   fDataFileName.clear();
   fModelDirectory.clear();
   fDefaultDataFileName.clear();
   fOpenedDataFileName.clear();
   if (fDataFile.is_open())
      fDataFile.close();
}
 
std::unique_ptr<onnx::ModelProto> RModelParser_ONNX::LoadModel(const std::string &filename) {
   std::fstream input(filename, std::ios::in | std::ios::binary);
   if (!input) {
      std::cerr << "TMVA::SOFIE - Failed to open onnx file " << filename << std::endl;
      return {};
   }
 
   return LoadModel(input);
}
 
std::unique_ptr<onnx::ModelProto> RModelParser_ONNX::LoadModel(std::istream &input)
{
   auto model = std::make_unique<onnx::ModelProto>();
 
   if (!model->ParseFromIstream(&input)) {
      std::cerr << "TMVA::SOFIE - Failed to parse ONNX model from input stream" << std::endl;
      return {};
   }
 
   // ONNX version is ir_version()  - model_version() returns 0
   if (fVerbose) {
      std::cout << "ONNX Version " << model->ir_version() << std::endl;
   }
   return model;
}
 
void RModelParser_ONNX::CheckGraph(const onnx::GraphProto & graph, int & level, std::map<std::string, int> & missingOperators) {
   if (fVerbose)
      std::cout << "\n" << graph.name() << " Graph operator list\n";
   for (int i = 0; i < graph.node_size(); i++) {
      const auto & node = graph.node(i);
      const std::string opType =  node.op_type();
      if (fVerbose) {
         std::cout << "\tOperator " << i << " : " << opType << " (" << node.name() << "), " << graph.node(i).input_size()
                      << " inputs : {";
            for (int j = 0; j < graph.node(i).input_size(); j++) {
               std::cout << graph.node(i).input(j);
               if (j < graph.node(i).input_size() - 1)
                  std::cout << ", ";
            }
         std::cout << " }" << std::endl;
      }
      // check if operator exists
      if (!IsRegisteredOperator(opType))
         missingOperators[opType] = level;
      // see if sub-graph exists as node attributes
      for (int j = 0; j < node.attribute_size(); j++) {
         const auto & attribute = node.attribute(j);
         if (attribute.has_g()) {
            const auto & subGraph = attribute.g();
            level += 1;
            CheckGraph(subGraph, level, missingOperators);
         }
      }
   }
}
 
bool RModelParser_ONNX::CheckModel(std::string filename, bool verbose) {
 
   fVerbose = verbose;
   auto model = LoadModel(filename);
   if (!model) return false;
 
   const onnx::GraphProto &graph = model->graph();
    // Initial operator order
   if (fVerbose)
      std::cout << "\nModel operator list " << model->producer_name() << "\n";
 
   std::map<std::string, int> missingOperators;
   int level = 1;
   CheckGraph(graph, level, missingOperators);
 
   if (!missingOperators.empty()) {
      std::cout << "List of missing operators for model loaded from file " << filename << std::endl;
      for (auto & op : missingOperators) {
         std::cout << op.first << "  " << op.second << std::endl;
      }
      return false;
   }
   std::cout << "All operators in the loaded model are supported!\n";
   return true;
}
 
void RModelParser_ONNX::ParseONNXGraph(RModel & rmodel, const onnx::GraphProto & graph, std::string  graphName)
{
   bool verbose = fVerbose;
 
   if (graphName.empty())
      graphName = graph.name();
 
   if (verbose)
      std::cout << "\nParsing Graph - " << graphName << std::endl;
 
   // fFusedOperators is keyed by node index, so it is only valid for the graph
   // being parsed: neither a second model parsed with the same parser nor a
   // subgraph (e.g. of the If operator) may inherit it.
   struct FusedOperatorsGuard {
      std::map<int, std::pair<EFusedOp, int>> &fMap;
      std::map<int, std::pair<EFusedOp, int>> fSaved;
      FusedOperatorsGuard(std::map<int, std::pair<EFusedOp, int>> &map) : fMap(map) { fSaved.swap(fMap); }
      ~FusedOperatorsGuard() { fMap.swap(fSaved); }
   } fusedOperatorsGuard{fFusedOperators};
 
   std::unordered_set<std::string> initializer_names;
   for (int i = 0; i < graph.initializer_size(); i++) {
      initializer_names.insert(graph.initializer(i).name());
   }
 
   if (verbose)
      std::cout << "Parsing model inputs...." << std::endl;
   /// Loop on model inputs
   for (int i = 0; i < graph.input_size(); i++) {
      RegisterTensorType(graph.input(i).name(),
                         static_cast<ETensorType>(graph.input(i).type().tensor_type().elem_type()));
 
      if (verbose)
         std::cout << "\tgraph input " << i << " name " << graph.input(i).name() << " type "
                   << graph.input(i).type().tensor_type().elem_type() << std::endl;
 
      if (initializer_names.find(graph.input(i).name()) != initializer_names.end())
         continue;
 
      // input data node is not a weight node (has no initializer)
      const onnx::ValueInfoProto &valueinfoproto = graph.input(i);
      std::string input_name = valueinfoproto.name();
 
      ETensorType type = static_cast<ETensorType>(valueinfoproto.type().tensor_type().elem_type());
 
      std::vector<Dim> fShape;
      bool existParam = false;
      if (!valueinfoproto.type().tensor_type().has_shape())
         throw std::runtime_error("TMVA::SOFIE data node with no shape restrictions is not supported yet");
      for (int j = 0; j < valueinfoproto.type().tensor_type().shape().dim_size(); j++) {
         Dim dim;
         if (valueinfoproto.type().tensor_type().shape().dim(j).value_case() ==
             onnx::TensorShapeProto_Dimension::ValueCase::kDimValue) {
             int dim_value = valueinfoproto.type().tensor_type().shape().dim(j).dim_value();
             dim.dim = dim_value;
             // case input dim is -1 - set a parametric shape
             if (dim_value < 0) {
               dim.isParam = true;
               existParam = true;
               dim.param = UTILITY::Clean_name(input_name) + "_size";
             }
         } else if (valueinfoproto.type().tensor_type().shape().dim(j).value_case() ==
                    onnx::TensorShapeProto_Dimension::ValueCase::kDimParam) {
            dim.isParam = true;
            existParam = true;
            dim.param = valueinfoproto.type().tensor_type().shape().dim(j).dim_param();
         } else {
            throw std::runtime_error("TMVA::SOFIE ONNX file error: Valueinfoproto " + input_name +
                                     " has neither dim_value nor dim_param! \n");
         }
         fShape.push_back(dim);
      }
      if (valueinfoproto.type().tensor_type().shape().dim_size() == 0) {
         Dim dim;
         dim.dim = 1;
         fShape.push_back(dim);
      } // in case this TensorShapeProto has no dimension message: ONNX IR defines this to be a scalar
 
      if (!existParam) {
         std::vector<size_t> fShape_sizet;
         for (auto &j : fShape) {
            fShape_sizet.push_back(j.dim);
         }
 
         rmodel.AddInputTensorInfo(input_name, type, fShape_sizet);
      } else {
         rmodel.AddInputTensorInfo(input_name, type, fShape);
      }
      rmodel.AddInputTensorName(input_name); // store also names in given order
   }
 
   std::map<std::string, int> allInitializedTensors;
 
   if (verbose)
      std::cout << "\nParsing graph initializer list and fill model initialized tensors" << std::endl;
 
   for (int i = 0; i < graph.initializer_size(); i++) {
      onnx::TensorProto *tensorproto = const_cast<onnx::TensorProto *>(&graph.initializer(i));
      std::vector<std::size_t> shape;
      std::size_t tensor_length = 1;
      for (int j = 0; j < tensorproto->dims_size(); j++) {
         shape.push_back(tensorproto->dims(j));
         tensor_length *= tensorproto->dims(j);
      }
      // in case of scalars keep an empty shape but with length =1
 
      std::string tensor_name = graph.initializer(i).name();
 
      if (verbose)
         std::cout << "\t initializer " << i << " name " << tensor_name << " type " << graph.initializer(i).data_type()
                   << " and length " << tensor_length << std::endl;
 
 
      // register also the initialized tensors
      auto tensor_type = static_cast<ETensorType>(graph.initializer(i).data_type());
      RegisterTensorType(tensor_name, tensor_type);
 
      std::shared_ptr<void> data = GetInitializedTensorData(tensorproto, tensor_length * GetTypeSize(tensor_type), tensor_type);
      rmodel.AddInitializedTensor(tensor_name, tensor_type, shape, data);
      allInitializedTensors[tensor_name] = i;
 
      if (verbose) {
         std::cout << "add initialized tensor " << tensor_name << "with shape " << ConvertShapeToString(shape) << "and  ";
         if (tensor_type == ETensorType::FLOAT) {
            std::cout << " float data: ";
            for (int j = 0; j < std::min(int(tensor_length),3); j++) std::cout << static_cast<float*>(data.get())[j] << "  ";
         }
         else if (tensor_type == ETensorType::INT64) {
            std::cout << " int64 data: ";
            for (int j = 0; j < std::min(int(tensor_length),3); j++) std::cout << static_cast<int64_t*>(data.get())[j] << "  ";
         }
         else if (tensor_type == ETensorType::UINT8) {
            std::cout << " uint8 data: ";
            for (int j = 0; j < std::min(int(tensor_length),3); j++) std::cout << static_cast<uint8_t*>(data.get())[j] << "  ";
         }
         else if (tensor_type == ETensorType::BOOL) {
            std::cout << " Boolean data: ";
            for (int j = 0; j < std::min(int(tensor_length),3); j++) std::cout << static_cast<bool*>(data.get())[j] << "  ";
         }
         std::cout << std::endl;
      }
   }   // end initializer list
 
   // Initial operator order
   if (verbose) {
      std::cout << "\nGraph operator list (ONNX order)\n";
      for (int i = 0; i < graph.node_size(); i++) {
         std::cout << "\tOperator " << i << " : " << graph.node(i).op_type() << " , " << graph.node(i).input_size()
                   << " inputs : {";
         for (int j = 0; j < graph.node(i).input_size(); j++) {
            std::cout << graph.node(i).input(j);
            if (j < graph.node(i).input_size() - 1)
               std::cout << ", ";
         }
         std::cout << " }" << std::endl;
      }
   }
 
   // make order of nodes:
   if (verbose)
      std::cout << "\n***********************\nRe-Order graph operator list\n*************************\n";
   std::vector<size_t> nodesOrder;
   nodesOrder.reserve(graph.node_size());
   std::vector<bool> foundNodes(graph.node_size());
 
   // loop at graph inputs
   std::map<std::string, int> allInputs;
   for (int i = 0; i < graph.input_size(); i++) {
      allInputs[graph.input(i).name()] = -1;
   }
   do {
      auto psize = nodesOrder.size();
      for (int i = 0; i < graph.node_size(); i++) {
         if (foundNodes[i])
            continue;
         // check if all input exists add to list
         bool existInputs = true;
         int input_size = graph.node(i).input_size();
         // special case for Reshape where shape is input and not a weight tensor
         if (fVerbose)
            std::cout << "Checking input of  Node " << i << " : " << graph.node(i).name() << std::endl;
         for (int j = 0; j < input_size; j++) {
            std::string name = graph.node(i).input(j);
            // skip empty names
            if (!name.empty()) {
               existInputs &= (allInputs.find(name) != allInputs.end() ||
                               allInitializedTensors.find(name) != allInitializedTensors.end());
               if (fVerbose) {
                  std::cout << "\t\t input " << name << " "
                     << bool(allInputs.find(name) != allInputs.end()) << "  " <<
                     bool(allInitializedTensors.find(name) != allInitializedTensors.end()) << "  " <<
                     existInputs << std::endl;
               }
            }
         }
         if (!existInputs) {
            if (fVerbose) {
               std::cout << "skip node " << graph.node(i).op_type() << "  " << graph.node(i).name() << " inputs are not existing ";
               for (int j = 0; j < input_size; j++) {
                  std::cout << graph.node(i).input(j) << " ";
               }
               std::cout << std::endl;
            }
            continue;
         }
 
         // adding node to the currectly ordered list
         if (verbose)
            std::cout << "===> New node " << graph.node(i).op_type() << "  " << graph.node(i).name() << " order " << i << std::endl;
 
         nodesOrder.push_back(i);
         foundNodes[i] = true;
         // register the outputs
         for (int j = 0; j < graph.node(i).output_size(); j++) {
            if (fVerbose) std::cout << "\toutput : " << graph.node(i).output(j) << std::endl;
            allInputs[graph.node(i).output(j)] = i;
         }
      }
      // no increment in nodes - something wrong
      if (nodesOrder.size() == psize) {
         int ilast = nodesOrder.back();
         std::cout << "cannot find a new node after " << graph.node(ilast).op_type() << " " << graph.node(ilast).name() << std::endl;
         throw std::runtime_error("TMVA::SOFIE - cannot find a new node ");
      }
   } while ((int)nodesOrder.size() < graph.node_size());
 
 
   // find list of children for each operator (used for fusing oiperators)
   std::vector<std::vector<int>> nodesChildren(graph.node_size());
 
   for (int k = 0; k < graph.node_size(); k++) {
      int i = nodesOrder[k];
      // compute the number of output for the operators
      if (graph.node(i).output_size() > 0) nodesChildren[i].reserve(graph.node(i).output_size());
      for (const auto& output_name : graph.node(i).output()) {
         // loop on all nodes
         for (int l = k; l < graph.node_size(); l++) {
            int j = nodesOrder[l];
            for (const auto& input_name : graph.node(j).input()) {
               if (input_name == output_name)
                  nodesChildren[i].push_back(j);
            }
         }
      }
   }
 
   // print lit of order operators with list of inputs and list of children nodes
   if (verbose) {
      std::cout << "\nGraph operator list (re-ordered)\n";
      for (int k = 0; k < graph.node_size(); k++) {
         int i = nodesOrder[k];
         std::cout << "\tOperator " << i << " : " << graph.node(i).op_type() << " , " << graph.node(i).name() << " input tensors : {";
            for (int j = 0; j < graph.node(i).input_size(); j++) {
            std::cout << graph.node(i).input(j);
            if (j < graph.node(i).input_size() - 1)
               std::cout << ", ";
         }
         std::cout << " } ";
         std::cout << " children : {";
         for ( const auto & ichild : nodesChildren[i]) {
            std::cout << " [ " << ichild << " " << graph.node(ichild).op_type() << " , " << graph.node(ichild).name() << "]";
         }
         std::cout << "}" << std::endl;
      }
   }
 
   // fill model with operators
   if (verbose) {
      std::cout << "Fill RModel with operators...\n";
   }
 
   // we have to record order of node execution separately to
   // account for fused operators
   size_t node_order_exec = 0;
   for (int i = 0; i < graph.node_size(); i++) {
      std::string op_type = graph.node(nodesOrder[i]).op_type();
 
      if (verbose) {
         std::cout << "\t" << i << "  " << nodesOrder[i] << " parsing operator " << op_type << std::endl;
      }
 
      std::unique_ptr<ROperator> op = ParseOperator(i, graph, nodesOrder, nodesChildren[nodesOrder[i]]);
      if (!op) {
         if (verbose) {
            std::cout << "\t\tskipping operator since it is fused with previous one" << std::endl;
         }
         // for skipping the fused nodes like Add after MatMul
         continue;
      }
      rmodel.AddOperator(std::move(op), node_order_exec++);
   }
 
   std::vector<std::string> outputnames;
   if (verbose)
      std::cout << "\nParsing Graph output list\n";
   for (int i = 0; i < graph.output_size(); i++) {
      if (verbose)
         std::cout << "\toutput " << i << " name " << graph.output(i).name() << std::endl;
      outputnames.push_back(graph.output(i).name());
   }
   rmodel.AddOutputTensorNameList(outputnames);
 
   return;
}
 
} // namespace SOFIE
} // namespace Experimental
} // namespace TMVA