xref: /external/tensorflow/tensorflow/lite/toco/import_tensorflow.cc

/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/
#include "tensorflow/lite/toco/import_tensorflow.h"

#include <memory>
#include <string>
#include <utility>
#include <vector>

#include "google/protobuf/map.h"
#include "google/protobuf/text_format.h"
#include "absl/memory/memory.h"
#include "absl/strings/match.h"
#include "absl/strings/numbers.h"
#include "absl/strings/str_cat.h"
#include "absl/strings/str_split.h"
#include "absl/strings/strip.h"
#include "tensorflow/core/common_runtime/device_factory.h"
#include "tensorflow/core/common_runtime/function.h"
#include "tensorflow/core/common_runtime/graph_constructor.h"
#include "tensorflow/core/common_runtime/process_function_library_runtime.h"
#include "tensorflow/core/framework/attr_value.pb.h"
#include "tensorflow/core/framework/function.pb.h"
#include "tensorflow/core/framework/graph.pb.h"
#include "tensorflow/core/framework/node_def.pb.h"
#include "tensorflow/core/framework/tensor.pb.h"
#include "tensorflow/core/framework/tensor_shape.pb.h"
#include "tensorflow/core/framework/types.pb.h"
#include "tensorflow/core/lib/core/errors.h"
#include "tensorflow/core/lib/core/status.h"
#include "tensorflow/core/platform/logging.h"
#include "tensorflow/core/public/session_options.h"
#include "tensorflow/core/public/version.h"
#include "tensorflow/lite/toco/model.h"
#include "tensorflow/lite/toco/model_flags.pb.h"
#include "tensorflow/lite/toco/tensorflow_graph_matching/resolve_cluster.h"
#include "tensorflow/lite/toco/tensorflow_util.h"
#include "tensorflow/lite/toco/tooling_util.h"

using tensorflow::AttrValue;
using tensorflow::DT_BOOL;
using tensorflow::DT_COMPLEX64;
using tensorflow::DT_FLOAT;
using tensorflow::DT_INT32;
using tensorflow::DT_INT64;
using tensorflow::DT_QUINT8;
using tensorflow::DT_STRING;
using tensorflow::DT_UINT32;
using tensorflow::DT_UINT8;
using tensorflow::GraphDef;
using tensorflow::NodeDef;
using tensorflow::OpRegistry;
using tensorflow::TensorProto;
using tensorflow::TensorShapeProto;

namespace toco {

namespace {
bool HasAttr(const NodeDef& node, const std::string& attr_name) {
  return node.attr().count(attr_name) > 0;
}

bool HasWildcardDimension(const TensorShapeProto& shape) {
  for (const auto& dim : shape.dim()) {
    if (dim.size() == -1) return true;
  }
  return false;
}

const std::string& GetStringAttr(const NodeDef& node,
                                 const std::string& attr_name) {
  CHECK(HasAttr(node, attr_name));
  const auto& attr = node.attr().at(attr_name);
  CHECK_EQ(attr.value_case(), AttrValue::kS);
  return attr.s();
}

int64 GetIntAttr(const NodeDef& node, const std::string& attr_name) {
  CHECK(HasAttr(node, attr_name)) << attr_name << " not found in:\n"
                                  << node.DebugString();
  const auto& attr = node.attr().at(attr_name);
  CHECK_EQ(attr.value_case(), AttrValue::kI);
  return attr.i();
}

float GetFloatAttr(const NodeDef& node, const std::string& attr_name) {
  CHECK(HasAttr(node, attr_name));
  const auto& attr = node.attr().at(attr_name);
  CHECK_EQ(attr.value_case(), AttrValue::kF);
  return attr.f();
}

bool GetBoolAttr(const NodeDef& node, const std::string& attr_name) {
  CHECK(HasAttr(node, attr_name));
  const auto& attr = node.attr().at(attr_name);
  CHECK_EQ(attr.value_case(), AttrValue::kB);
  return attr.b();
}

tensorflow::DataType GetDataTypeAttr(const NodeDef& node,
                                     const std::string& attr_name) {
  CHECK(HasAttr(node, attr_name));
  const auto& attr = node.attr().at(attr_name);
  CHECK_EQ(attr.value_case(), AttrValue::kType);
  return attr.type();
}

const TensorShapeProto& GetShapeAttr(const NodeDef& node,
                                     const std::string& attr_name) {
  CHECK(HasAttr(node, attr_name));
  const auto& attr = node.attr().at(attr_name);
  CHECK_EQ(attr.value_case(), AttrValue::kShape);
  return attr.shape();
}

const TensorProto& GetTensorAttr(const NodeDef& node,
                                 const std::string& attr_name) {
  CHECK(HasAttr(node, attr_name)) << "No attr named '" << attr_name << "'";
  const auto& attr = node.attr().at(attr_name);
  CHECK_EQ(attr.value_case(), AttrValue::kTensor);
  return attr.tensor();
}

const AttrValue::ListValue& GetListAttr(const NodeDef& node,
                                        const std::string& attr_name) {
  CHECK(HasAttr(node, attr_name));
  const auto& attr = node.attr().at(attr_name);
  CHECK_EQ(attr.value_case(), AttrValue::kList);
  return attr.list();
}

tensorflow::Status CheckOptionalAttr(const NodeDef& node,
                                     const std::string& attr_name,
                                     const std::string& expected_value) {
  if (HasAttr(node, attr_name)) {
    const std::string& value = GetStringAttr(node, attr_name);
    if (value != expected_value) {
      return tensorflow::errors::InvalidArgument(
          "Unexpected value for attribute '" + attr_name + "'. Expected '" +
          expected_value + "'");
    }
  }
  return tensorflow::Status::OK();
}

tensorflow::Status CheckOptionalAttr(
    const NodeDef& node, const std::string& attr_name,
    const tensorflow::DataType& expected_value) {
  if (HasAttr(node, attr_name)) {
    const tensorflow::DataType& value = GetDataTypeAttr(node, attr_name);
    if (value != expected_value) {
      return tensorflow::errors::InvalidArgument(
          "Unexpected value for attribute '" + attr_name + "'. Expected '" +
          tensorflow::DataType_Name(expected_value) + "'");
    }
  }
  return tensorflow::Status::OK();
}

template <typename T1, typename T2>
tensorflow::Status ExpectValue(const T1& v1, const T2& v2,
                               const std::string& description) {
  if (v1 == v2) return tensorflow::Status::OK();
  return tensorflow::errors::InvalidArgument(absl::StrCat(
      "Unexpected ", description, ": got ", v1, ", expected ", v2));
}

ArrayDataType ConvertDataType(tensorflow::DataType dtype) {
  if (dtype == DT_UINT8)
    return ArrayDataType::kUint8;
  else if (dtype == DT_FLOAT)
    return ArrayDataType::kFloat;
  else if (dtype == DT_BOOL)
    return ArrayDataType::kBool;
  else if (dtype == DT_INT32)
    return ArrayDataType::kInt32;
  else if (dtype == DT_UINT32)
    return ArrayDataType::kUint32;
  else if (dtype == DT_INT64)
    return ArrayDataType::kInt64;
  else if (dtype == DT_STRING)
    return ArrayDataType::kString;
  else if (dtype == DT_COMPLEX64)
    return ArrayDataType::kComplex64;
  else
    LOG(INFO) << "Unsupported data type in placeholder op: " << dtype;
  return ArrayDataType::kNone;
}

tensorflow::Status ImportShape(
    const TFLITE_PROTO_NS::RepeatedPtrField<tensorflow::TensorShapeProto_Dim>&
        input_dims,
    int* input_flat_size, Shape* shape) {
  std::vector<int> input_dims_only_sizes;
  bool zero_sized_shape = false;
  for (auto& d : input_dims) {
    // TensorFlow's shapes use int64s, while TOCO uses ints.
    if (d.size() > std::numeric_limits<int>::max()) {
      return tensorflow::errors::InvalidArgument("Shape element overflows");
    }
    if (d.size() == 0) {
      zero_sized_shape = true;
    }
    input_dims_only_sizes.push_back(d.size());
  }

  // Note that up to this point we were OK with the input shape containing
  // elements valued -1 or 0, which are perfectly legal in tensorflow. However
  // our CheckValidShapeDimensions() insists on them being >= 1, with the
  // exception of the "scalar" shape [0]. The main issue with zero-values shape
  // elements is that the corresponding arrays don't contain any data and the
  // allocation code gets a bit confused. It seems that the code expects an
  // empty shape for zero-sized shapes, so we will do just that, except for the
  // [0] case.
  // TODO(b/119325030): In order to correctly import the "scalar" shapes the
  // following test must include "&& input_dims_only_sizes.size() > 1", but
  // that seems to slow everything down a lot.
  if (zero_sized_shape) {
    shape->mutable_dims()->clear();
    if (input_flat_size != nullptr) *input_flat_size = 0;
    return tensorflow::Status::OK();
  }

  *shape->mutable_dims() = input_dims_only_sizes;

  if (input_flat_size == nullptr) return tensorflow::Status::OK();

  return NumElements(input_dims_only_sizes, input_flat_size);
}

// Define ways to retrieve data from tensors of different types.
// TODO(b/80208043): simply use tensorflow::Tensor::FromProto() instead.
template <typename T>
struct TensorTraits;

template <>
struct TensorTraits<float> {
  static int size(const TensorProto& p) { return p.float_val_size(); }
  static float get(const TensorProto& p, int i) { return p.float_val(i); }
  static std::string accessor_name() { return "float_val"; }
  static std::string type_name() { return "float"; }
  static void CopyFromContent(const TensorProto& p, std::vector<float>* data) {
    toco::port::CopyToBuffer(p.tensor_content(),
                             reinterpret_cast<char*>(data->data()));
  }
};

template <>
struct TensorTraits<uint8_t> {
  static int size(const TensorProto& p) { return p.int_val_size(); }
  static uint8_t get(const TensorProto& p, int i) { return p.int_val(i); }
  static std::string accessor_name() { return "int_val"; }
  static std::string type_name() { return "uint8"; }
  static void CopyFromContent(const TensorProto& p,
                              std::vector<uint8_t>* data) {
    toco::port::CopyToBuffer(p.tensor_content(),
                             reinterpret_cast<char*>(data->data()));
  }
};

template <>
struct TensorTraits<std::complex<float>> {
  static int size(const TensorProto& p) { return p.scomplex_val_size() / 2; }
  static std::complex<float> get(const TensorProto& p, int i) {
    return std::complex<float>(p.scomplex_val(2 * i),
                               p.scomplex_val(2 * i + 1));
  }
  static std::string accessor_name() { return "scomplex_val"; }
  static std::string type_name() { return "complex64"; }
  static void CopyFromContent(const TensorProto& p,
                              std::vector<std::complex<float>>* data) {
    toco::port::CopyToBuffer(p.tensor_content(),
                             reinterpret_cast<char*>(data->data()));
  }
};

template <>
struct TensorTraits<int32> {
  static int size(const TensorProto& p) { return p.int_val_size(); }
  static int32 get(const TensorProto& p, int i) { return p.int_val(i); }
  static std::string accessor_name() { return "int_val"; }
  static std::string type_name() { return "int32"; }
  static void CopyFromContent(const TensorProto& p, std::vector<int32>* data) {
    toco::port::CopyToBuffer(p.tensor_content(),
                             reinterpret_cast<char*>(data->data()));
  }
};

template <>
struct TensorTraits<uint32> {
  static int size(const TensorProto& p) { return p.uint32_val_size(); }
  static int32 get(const TensorProto& p, int i) { return p.uint32_val(i); }
  static std::string accessor_name() { return "uint32_val"; }
  static std::string type_name() { return "uint32"; }
  static void CopyFromContent(const TensorProto& p, std::vector<uint32>* data) {
    toco::port::CopyToBuffer(p.tensor_content(),
                             reinterpret_cast<char*>(data->data()));
  }
};

template <>
struct TensorTraits<int64> {
  static int size(const TensorProto& p) { return p.int64_val_size(); }
  static int64 get(const TensorProto& p, int i) { return p.int64_val(i); }
  static std::string accessor_name() { return "int64_val"; }
  static std::string type_name() { return "int64"; }
  static void CopyFromContent(const TensorProto& p, std::vector<int64>* data) {
    toco::port::CopyToBuffer(p.tensor_content(),
                             reinterpret_cast<char*>(data->data()));
  }
};

template <>
struct TensorTraits<bool> {
  static int size(const TensorProto& p) { return p.bool_val_size(); }
  static bool get(const TensorProto& p, int i) { return p.bool_val(i); }
  static std::string accessor_name() { return "bool_val"; }
  static std::string type_name() { return "bool"; }
  static void CopyFromContent(const TensorProto& p, std::vector<bool>* data) {
    std::vector<char> buf(p.tensor_content().size());
    toco::port::CopyToBuffer(p.tensor_content(), buf.data());
    for (int i = 0; i < p.tensor_content().size(); i++) {
      (*data)[i] = static_cast<bool>(buf[i]);
    }
  }
};

template <typename T>
tensorflow::Status ImportTensorData(const TensorProto& input_tensor,
                                    int input_flat_size,
                                    std::vector<T>* output_data) {
  CHECK_GE(output_data->size(), input_flat_size);
  int num_elements_in_tensor = TensorTraits<T>::size(input_tensor);
  if (num_elements_in_tensor == input_flat_size) {
    for (int i = 0; i < num_elements_in_tensor; i++) {
      (*output_data)[i] = TensorTraits<T>::get(input_tensor, i);
    }
  } else if (input_tensor.tensor_content().size() ==
             input_flat_size * sizeof(T)) {
    TensorTraits<T>::CopyFromContent(input_tensor, output_data);
  } else if (num_elements_in_tensor > 0 &&
             num_elements_in_tensor < input_flat_size) {
    // TODO(b/80208043): use tensorflow::Tensor::FromProto() which is the
    // official way to import tensor data. This particular else-if handles a
    // grappler optimization where the last few elements in a tensor are
    // omitted if they are repeated.
    int i = 0;
    for (; i < num_elements_in_tensor; ++i) {
      (*output_data)[i] = TensorTraits<T>::get(input_tensor, i);
    }
    auto last = (*output_data)[i - 1];
    for (; i < input_flat_size; ++i) {
      (*output_data)[i] = last;
    }
  } else {
    std::string accessor_name = TensorTraits<T>::accessor_name();
    std::string type_name = TensorTraits<T>::type_name();
    return tensorflow::errors::InvalidArgument(
        absl::StrCat("Neither input_content (",
                     input_tensor.tensor_content().size() / sizeof(T), ") nor ",
                     accessor_name, " (", num_elements_in_tensor,
                     ") have the right dimensions (", input_flat_size,
                     ") for this ", type_name, " tensor"));
  }
  return tensorflow::Status::OK();
}

tensorflow::Status ImportFloatArray(const TensorProto& input_tensor,
                                    Array* output_array) {
  CHECK_EQ(input_tensor.dtype(), DT_FLOAT);
  const auto& input_shape = input_tensor.tensor_shape();
  CHECK_LE(input_shape.dim_size(), 6);
  int input_flat_size;
  auto status = ImportShape(input_shape.dim(), &input_flat_size,
                            output_array->mutable_shape());
  if (!status.ok()) return status;

  auto& output_float_data =
      output_array->GetMutableBuffer<ArrayDataType::kFloat>().data;
  output_float_data.resize(RequiredBufferSizeForShape(output_array->shape()),
                           0.f);
  return ImportTensorData<float>(input_tensor, input_flat_size,
                                 &output_float_data);
}

tensorflow::Status ImportComplex64Array(const TensorProto& input_tensor,
                                        Array* output_array) {
  CHECK_EQ(input_tensor.dtype(), DT_COMPLEX64);
  const auto& input_shape = input_tensor.tensor_shape();
  CHECK_LE(input_shape.dim_size(), 4);
  int input_flat_size;
  auto status = ImportShape(input_shape.dim(), &input_flat_size,
                            output_array->mutable_shape());
  if (!status.ok()) return status;

  auto& output_complex_data =
      output_array->GetMutableBuffer<ArrayDataType::kComplex64>().data;
  output_complex_data.resize(RequiredBufferSizeForShape(output_array->shape()),
                             std::complex<float>(0.f, 0.f));
  return ImportTensorData<std::complex<float>>(input_tensor, input_flat_size,
                                               &output_complex_data);
}

tensorflow::Status ImportQuint8Array(const TensorProto& input_tensor,
                                     Array* output_array) {
  CHECK_EQ(input_tensor.dtype(), DT_QUINT8);
  const auto& input_shape = input_tensor.tensor_shape();
  CHECK_LE(input_shape.dim_size(), 6);
  int input_flat_size;
  auto status = ImportShape(input_shape.dim(), &input_flat_size,
                            output_array->mutable_shape());
  if (!status.ok()) return status;

  auto& output_int_data =
      output_array->GetMutableBuffer<ArrayDataType::kUint8>().data;
  output_int_data.resize(RequiredBufferSizeForShape(output_array->shape()), 0);
  return ImportTensorData<uint8_t>(input_tensor, input_flat_size,
                                   &output_int_data);
}

tensorflow::Status ImportInt32Array(const TensorProto& input_tensor,
                                    Array* output_array) {
  CHECK_EQ(input_tensor.dtype(), DT_INT32);
  const auto& input_shape = input_tensor.tensor_shape();
  CHECK_LE(input_shape.dim_size(), 6);
  int input_flat_size;
  auto status = ImportShape(input_shape.dim(), &input_flat_size,
                            output_array->mutable_shape());
  if (!status.ok()) return status;

  auto& output_int_data =
      output_array->GetMutableBuffer<ArrayDataType::kInt32>().data;
  output_int_data.resize(RequiredBufferSizeForShape(output_array->shape()), 0);
  return ImportTensorData<int32>(input_tensor, input_flat_size,
                                 &output_int_data);
}

tensorflow::Status ImportUint32Array(const TensorProto& input_tensor,
                                     Array* output_array) {
  CHECK_EQ(input_tensor.dtype(), DT_UINT32);
  const auto& input_shape = input_tensor.tensor_shape();
  CHECK_LE(input_shape.dim_size(), 6);
  int input_flat_size;
  auto status = ImportShape(input_shape.dim(), &input_flat_size,
                            output_array->mutable_shape());
  if (!status.ok()) return status;

  auto& output_int_data =
      output_array->GetMutableBuffer<ArrayDataType::kUint32>().data;
  output_int_data.resize(RequiredBufferSizeForShape(output_array->shape()), 0);
  return ImportTensorData<uint32>(input_tensor, input_flat_size,
                                  &output_int_data);
}

tensorflow::Status ImportInt64Array(const TensorProto& input_tensor,
                                    Array* output_array) {
  CHECK_EQ(input_tensor.dtype(), DT_INT64);
  const auto& input_shape = input_tensor.tensor_shape();
  CHECK_LE(input_shape.dim_size(), 6);
  int input_flat_size;
  auto status = ImportShape(input_shape.dim(), &input_flat_size,
                            output_array->mutable_shape());
  if (!status.ok()) return status;

  auto& output_int_data =
      output_array->GetMutableBuffer<ArrayDataType::kInt64>().data;
  output_int_data.resize(RequiredBufferSizeForShape(output_array->shape()), 0);
  return ImportTensorData<int64>(input_tensor, input_flat_size,
                                 &output_int_data);
}

tensorflow::Status ImportBoolArray(const TensorProto& input_tensor,
                                   Array* output_array) {
  CHECK_EQ(input_tensor.dtype(), DT_BOOL);
  const auto& input_shape = input_tensor.tensor_shape();
  CHECK_LE(input_shape.dim_size(), 6);
  int input_flat_size;
  auto status = ImportShape(input_shape.dim(), &input_flat_size,
                            output_array->mutable_shape());
  if (!status.ok()) return status;

  auto& output_bool_data =
      output_array->GetMutableBuffer<ArrayDataType::kBool>().data;
  output_bool_data.resize(RequiredBufferSizeForShape(output_array->shape()),
                          false);
  status =
      ImportTensorData<bool>(input_tensor, input_flat_size, &output_bool_data);
  if (!status.ok() && output_bool_data.size() == 1) {
    // Some graphs have bool const nodes without actual value...
    // assuming that 'false' is implied.
    // So far only encountered that in an array with 1 entry, let's
    // require that until we encounter a graph where that's not the case.
    output_bool_data[0] = false;
    return tensorflow::Status::OK();
  }
  return status;
}

tensorflow::Status ImportStringArray(const TensorProto& input_tensor,
                                     Array* output_array) {
  CHECK_EQ(input_tensor.dtype(), DT_STRING);
  const auto& input_shape = input_tensor.tensor_shape();
  CHECK_LE(input_shape.dim_size(), 6);
  int input_flat_size;
  auto status = ImportShape(input_shape.dim(), &input_flat_size,
                            output_array->mutable_shape());
  if (!status.ok()) return status;

  if (input_flat_size != input_tensor.string_val_size()) {
    return tensorflow::errors::InvalidArgument(
        "Input_content string_val doesn't have the right dimensions "
        "for this string tensor");
  }

  auto& output_string_data =
      output_array->GetMutableBuffer<ArrayDataType::kString>().data;
  output_string_data.resize(RequiredBufferSizeForShape(output_array->shape()));
  CHECK_GE(output_string_data.size(), input_flat_size);
  for (int i = 0; i < input_flat_size; ++i) {
    output_string_data[i] = input_tensor.string_val(i);
  }
  return tensorflow::Status::OK();
}

// Count the number of inputs of a given node. If
// `tf_import_flags.drop_control_dependency` is true, count the number of
// non-control-dependency inputs.
int GetInputsCount(const NodeDef& node,
                   const TensorFlowImportFlags& tf_import_flags) {
  if (tf_import_flags.drop_control_dependency) {
    for (size_t i = 0; i < node.input_size(); ++i) {
      if (node.input(i)[0] == '^') {
        return i;
      }
    }
  }
  return node.input_size();
}

tensorflow::Status CheckInputsCount(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    int expected_input_count) {
  if (GetInputsCount(node, tf_import_flags) != expected_input_count) {
    return tensorflow::errors::FailedPrecondition(
        node.op(), " node expects ", expected_input_count,
        " input(s) other than control dependencies: ", node.DebugString());
  }
  return tensorflow::Status::OK();
}

template <ArrayDataType T>
std::string CreateConstArray(
    Model* model, std::string const& name,
    std::vector<typename toco::DataType<T>> const& data) {
  // Utility function to create a const 1D array, useful for input parameters.
  std::string array_name = toco::AvailableArrayName(*model, name);
  auto& array = model->GetOrCreateArray(array_name);
  array.data_type = T;
  array.mutable_shape()->mutable_dims()->emplace_back(
      static_cast<int>(data.size()));
  array.GetMutableBuffer<T>().data = data;
  return array_name;
}

// Retain TensorFlow NodeDef in Toco Operator.
//
// If an op is supported by Toco but not supported by TFLite, TFLite exporter
// will use the retained NodeDef to populate a Flex op when Flex mode is
// enabled.
//
// This can't be easily applied to all operations, because a TensorFlow node
// may become multiple Toco operators. Thus we need to call this function in
// operator conversion functions one by one whenever feasible.
//
// This may cause problems if a graph transformation rule changes parameters
// of the node. When calling this function, please check if any existing
// graph transformation rule will change an existing operator with the same
// type.
//
// This provides a route to handle Toco-supported & TFLite-unsupported ops
// in Flex mode. However it's not a solid solution. Eventually we should
// get rid of this.
// TODO(b/117327937): Implement all Toco-supported ops in TFLite, and remove
// this function.
void RetainTensorFlowNodeDef(const NodeDef& node, Operator* op) {
  node.SerializeToString(&op->tensorflow_node_def);
}

void GetOutputNamesFromNodeDef(const NodeDef& node,
                               const tensorflow::OpDef& op_def,
                               TensorFlowUnsupportedOperator* op) {
  int next_output = 0;
  auto add_output = [&node, &next_output, op]() {
    if (next_output == 0) {
      op->outputs.push_back(node.name());  // Implicit :0.
    } else {
      op->outputs.push_back(absl::StrCat(node.name(), ":", next_output));
    }
    ++next_output;
  };
  for (int i = 0; i < op_def.output_arg_size(); ++i) {
    std::string multiples = op_def.output_arg(i).number_attr();
    if (!multiples.empty()) {
      CHECK(HasAttr(node, multiples)) << "No attr named " << multiples;
      int num_outputs = GetIntAttr(node, multiples);
      for (int j = 0; j < num_outputs; ++j) {
        add_output();
      }
    } else {
      std::string list = op_def.output_arg(i).type_list_attr();
      if (!list.empty()) {
        CHECK(HasAttr(node, list)) << "No attr named " << list;
        const AttrValue::ListValue& list_value = GetListAttr(node, list);
        for (int j = 0; j < list_value.type_size(); ++j) {
          add_output();
        }
      } else {
        add_output();
      }
    }
  }
}

void GetOutputTypesFromNodeDef(const NodeDef& node,
                               const tensorflow::OpDef& op_def,
                               TensorFlowUnsupportedOperator* op) {
  // The given type to the op, or clear the types if invalid.
  auto add_type = [&node, op](tensorflow::DataType type) {
    if (type == tensorflow::DT_INVALID) {
      LOG(WARNING) << "Op node missing output type attribute: " << node.name();
      op->output_data_types.clear();
    } else {
      op->output_data_types.push_back(ConvertDataType(type));
    }
  };

  // Retrieve the data type according to the OpDef definition: either the
  // "type" or "type_attr" field will be set.
  auto get_type = [&node](const tensorflow::OpDef::ArgDef& a) {
    if (a.type() != tensorflow::DT_INVALID) {
      return a.type();
    } else if (HasAttr(node, a.type_attr())) {
      return GetDataTypeAttr(node, a.type_attr());
    } else {
      return tensorflow::DT_INVALID;
    }
  };

  for (int i = 0; i < op_def.output_arg_size(); ++i) {
    std::string multiples = op_def.output_arg(i).number_attr();
    if (!multiples.empty()) {
      CHECK(HasAttr(node, multiples)) << "No attr named " << multiples;
      int num_outputs = GetIntAttr(node, multiples);
      auto type = get_type(op_def.output_arg(i));
      for (int j = 0; j < num_outputs; ++j) {
        add_type(type);
      }
    } else {
      std::string list = op_def.output_arg(i).type_list_attr();
      if (!list.empty()) {
        CHECK(HasAttr(node, list)) << "No attr named " << list;
        const AttrValue::ListValue& list_value = GetListAttr(node, list);
        for (int j = 0; j < list_value.type_size(); ++j) {
          add_type(list_value.type(j));
        }
      } else {
        add_type(get_type(op_def.output_arg(i)));
      }
    }
  }
}

tensorflow::Status ConvertUnsupportedOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  // Names of special attributes in TF graph that are used by Toco.
  static constexpr char kAttrOutputQuantized[] = "_output_quantized";
  static constexpr char kAttrOutputTypes[] = "_output_types";
  static constexpr char kAttrOutputShapes[] = "_output_shapes";
  static constexpr char kAttrSupportOutputTypeFloatInQuantizedOp[] =
      "_support_output_type_float_in_quantized_op";

  LOG(INFO) << "Converting unsupported operation: " << node.op();

  auto* op = new TensorFlowUnsupportedOperator;
  op->tensorflow_op = node.op();

  // For Flex mode. Please read the comments of the function.
  RetainTensorFlowNodeDef(node, op);

  model->operators.emplace_back(op);

  // Parse inputs.
  const int num_inputs = GetInputsCount(node, tf_import_flags);
  for (int i = 0; i < num_inputs; ++i) {
    op->inputs.push_back(node.input(i));
  }

  // Parse outputs. Name them after the node's name, plus an ordinal suffix.
  // Note that some outputs are to be multiplied by a named attribute.
  const tensorflow::OpDef* op_def = nullptr;
  if (tensorflow::OpRegistry::Global()->LookUpOpDef(node.op(), &op_def).ok()) {
    GetOutputNamesFromNodeDef(node, *op_def, op);
  } else {
    op->outputs.push_back(node.name());  // Implicit :0.
  }

  // Parse if the op supports quantization
  if (HasAttr(node, kAttrOutputQuantized)) {
    op->quantized = GetBoolAttr(node, kAttrOutputQuantized);
  }
  // Parse if the quantized op allows output arrays of type float
  if (HasAttr(node, kAttrSupportOutputTypeFloatInQuantizedOp)) {
    op->support_output_type_float_in_quantized_op =
        GetBoolAttr(node, kAttrSupportOutputTypeFloatInQuantizedOp);
  }

  // Parse output type(s).
  if (HasAttr(node, kAttrOutputTypes)) {
    const auto& output_types = GetListAttr(node, kAttrOutputTypes);
    for (int i = 0; i < output_types.type_size(); ++i) {
      op->output_data_types.push_back(ConvertDataType(output_types.type(i)));
    }
  } else if (HasAttr(node, "Tout")) {
    const auto& output_type = GetDataTypeAttr(node, "Tout");
    op->output_data_types.push_back(ConvertDataType(output_type));
  } else if (op_def != nullptr) {
    GetOutputTypesFromNodeDef(node, *op_def, op);
  } else {
    // TODO(b/113613439): Figure out how to propagate types for custom ops
    // that have no OpDef.
    LOG(INFO) << "Unable to determine output type for op: " << node.op();
  }

  // Parse output shape(s).
  if (HasAttr(node, kAttrOutputShapes)) {
    const auto& output_shapes = GetListAttr(node, kAttrOutputShapes);
    Shape output_shape;
    for (int i = 0; i < output_shapes.shape_size(); ++i) {
      const auto& shape = output_shapes.shape(i);
      // TOCO doesn't yet properly handle shapes with wildcard dimensions.
      // TODO(b/113613439): Handle shape inference for unsupported ops that have
      // shapes with wildcard dimensions.
      if (HasWildcardDimension(shape)) {
        LOG(INFO) << "Skipping wildcard output shape(s) for node: "
                  << node.name();
        op->output_shapes.clear();
        break;
      }
      const auto status =
          ImportShape(shape.dim(), /*input_flat_size=*/nullptr, &output_shape);
      if (!status.ok()) {
        return status;
      }
      op->output_shapes.push_back(output_shape);
    }
  }
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertConstOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "Const");
  const auto& tensor = GetTensorAttr(node, "value");
  const auto dtype = GetDataTypeAttr(node, "dtype");

  tensorflow::Status status = tensorflow::Status::OK();

  auto& array = model->GetOrCreateArray(node.name());
  switch (dtype) {
    case DT_FLOAT:
      array.data_type = ArrayDataType::kFloat;
      status = ImportFloatArray(tensor, &array);
      break;
    case DT_INT32:
      array.data_type = ArrayDataType::kInt32;
      status = ImportInt32Array(tensor, &array);
      break;
    case DT_UINT32:
      array.data_type = ArrayDataType::kUint32;
      status = ImportUint32Array(tensor, &array);
      break;
    case DT_QUINT8:
      array.data_type = ArrayDataType::kUint8;
      status = ImportQuint8Array(tensor, &array);
      break;
    case DT_INT64:
      array.data_type = ArrayDataType::kInt64;
      status = ImportInt64Array(tensor, &array);
      break;
    case DT_STRING:
      array.data_type = ArrayDataType::kString;
      status = ImportStringArray(tensor, &array);
      break;
    case DT_BOOL:
      array.data_type = ArrayDataType::kBool;
      status = ImportBoolArray(tensor, &array);
      break;
    case DT_COMPLEX64:
      array.data_type = ArrayDataType::kComplex64;
      status = ImportComplex64Array(tensor, &array);
      break;
    default:
      array.data_type = ArrayDataType::kNone;
      // do nothing, silently ignore the Const data.
      // We just make a dummy buffer to indicate that
      // this array does not rely on external input.
      array.GetMutableBuffer<ArrayDataType::kNone>();
      break;
  }
  TF_RETURN_WITH_CONTEXT_IF_ERROR(
      status, " (while processing node '" + node.name() + "')");
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertConvOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "Conv2D");
  TF_RETURN_IF_ERROR(CheckInputsCount(node, tf_import_flags, 2));

  // We only support NHWC, which is the default data_format.
  // So if data_format is not defined, we're all good.
  TF_RETURN_IF_ERROR(CheckOptionalAttr(node, "data_format", "NHWC"));
  TF_RETURN_IF_ERROR(CheckOptionalAttr(node, "T", DT_FLOAT));

  const auto& input_name = node.input(0);
  const auto& weights_name = node.input(1);
  const auto& reordered_weights_name =
      AvailableArrayName(*model, weights_name + "_reordered");
  // Check if a ReorderAxesOperator was already created for these weights
  // (that happens when multiple layers share the same weights).
  const Operator* existing_reorder =
      GetOpWithOutput(*model, reordered_weights_name);
  if (existing_reorder) {
    // Check that it is safe to rely on the _reordered naming of the output
    // array!
    CHECK(existing_reorder->type == OperatorType::kReorderAxes);
  } else {
    // Create a new ReorderAxesOperator
    auto* reorder = new ReorderAxesOperator;
    reorder->inputs = {weights_name};
    reorder->outputs = {reordered_weights_name};
    reorder->input_axes_order = AxesOrder::kHWIO;
    reorder->output_axes_order = AxesOrder::kOHWI;
    model->operators.emplace_back(reorder);
  }
  if (!HasAttr(node, "strides")) {
    return tensorflow::errors::InvalidArgument("Missing attribute 'strides'");
  }
  const auto& strides = GetListAttr(node, "strides");
  TF_RETURN_IF_ERROR(ExpectValue(strides.i_size(), 4, "number of strides"));
  TF_RETURN_IF_ERROR(ExpectValue(strides.i(0), 1, "strides(0)"));
  TF_RETURN_IF_ERROR(ExpectValue(strides.i(3), 1, "strides(3)"));
  int dilation_height_factor;
  int dilation_width_factor;
  if (HasAttr(node, "dilations")) {
    const auto& dilations = GetListAttr(node, "dilations");
    TF_RETURN_IF_ERROR(
        ExpectValue(dilations.i_size(), 4, "number of dilations"));
    if (dilations.i(0) != 1 || dilations.i(3) != 1) {
      return tensorflow::errors::InvalidArgument(absl::StrCat(
          "Can only import Conv ops with dilation along the height "
          "(1st) or width (2nd) axis. TensorFlow op \"",
          node.name(), "\" had dilations:[ ", dilations.i(0), ", ",
          dilations.i(1), ", ", dilations.i(2), ", ", dilations.i(3), "]."));
    }
    dilation_height_factor = dilations.i(1);
    dilation_width_factor = dilations.i(2);
  } else {
    dilation_height_factor = 1;
    dilation_width_factor = 1;
  }
  const auto& padding = GetStringAttr(node, "padding");
  PaddingType padding_type;
  if (padding == "SAME") {
    padding_type = PaddingType::kSame;
  } else if (padding == "VALID") {
    padding_type = PaddingType::kValid;
  } else {
    return tensorflow::errors::InvalidArgument(
        "Bad padding (only SAME and VALID are supported)");
  }
  auto* conv = new ConvOperator;
  conv->inputs = {input_name, reordered_weights_name};
  conv->outputs = {node.name()};
  conv->stride_height = strides.i(1);
  conv->stride_width = strides.i(2);
  conv->dilation_height_factor = dilation_height_factor;
  conv->dilation_width_factor = dilation_width_factor;
  conv->padding.type = padding_type;
  model->operators.emplace_back(conv);

  return tensorflow::Status::OK();
}

tensorflow::Status ConvertDepthwiseConvOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "DepthwiseConv2dNative");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 2));

  // We only support NHWC, which is the default data_format.
  // So if data_format is not defined, we're all good.
  if (HasAttr(node, "data_format")) {
    CHECK_EQ(GetStringAttr(node, "data_format"), "NHWC");
  }
  CHECK_EQ(GetDataTypeAttr(node, "T"), DT_FLOAT);

  const auto& input_name = node.input(0);
  const auto& weights_name = node.input(1);
  const auto& reordered_weights_name = weights_name + "_reordered";
  // Check if a ReorderAxesOperator was already created for these weights
  // (that happens when multiple layers share the same weights).
  const Operator* existing_reorder =
      GetOpWithOutput(*model, reordered_weights_name);
  if (existing_reorder) {
    // Check that it is safe to rely on the _reordered naming of the output
    // array!
    CHECK(existing_reorder->type == OperatorType::kReorderAxes);
  } else {
    // Create a new ReorderAxesOperator
    auto* reorder = new ReorderAxesOperator;
    reorder->inputs = {weights_name};
    reorder->outputs = {reordered_weights_name};
    reorder->input_axes_order = AxesOrder::kHWIM;
    reorder->output_axes_order = AxesOrder::k1HWO;
    model->operators.emplace_back(reorder);
  }
  const auto& strides = GetListAttr(node, "strides");
  TF_RETURN_IF_ERROR(ExpectValue(strides.i_size(), 4, "number of strides"));
  TF_RETURN_IF_ERROR(ExpectValue(strides.i(0), 1, "strides(0)"));
  TF_RETURN_IF_ERROR(ExpectValue(strides.i(3), 1, "strides(3)"));
  int dilation_height_factor;
  int dilation_width_factor;
  if (HasAttr(node, "dilations")) {
    const auto& dilations = GetListAttr(node, "dilations");
    TF_RETURN_IF_ERROR(
        ExpectValue(dilations.i_size(), 4, "number of dilations"));
    if (dilations.i(0) != 1 || dilations.i(3) != 1) {
      return tensorflow::errors::InvalidArgument(absl::StrCat(
          "Can only import Conv ops with dilation along the height "
          "(1st) or width (2nd) axis. TensorFlow op \"",
          node.name(), "\" had dilations:[ ", dilations.i(0), ", ",
          dilations.i(1), ", ", dilations.i(2), ", ", dilations.i(3), "]."));
    }
    dilation_height_factor = dilations.i(1);
    dilation_width_factor = dilations.i(2);
  } else {
    dilation_height_factor = 1;
    dilation_width_factor = 1;
  }
  const auto& padding = GetStringAttr(node, "padding");
  PaddingType padding_type;
  if (padding == "SAME") {
    padding_type = PaddingType::kSame;
  } else if (padding == "VALID") {
    padding_type = PaddingType::kValid;
  } else {
    return tensorflow::errors::InvalidArgument(
        "Bad padding (only SAME and VALID are supported)");
  }
  auto* conv = new DepthwiseConvOperator;
  conv->inputs = {input_name, reordered_weights_name};
  conv->outputs = {node.name()};
  conv->stride_height = strides.i(1);
  conv->stride_width = strides.i(2);
  conv->dilation_height_factor = dilation_height_factor;
  conv->dilation_width_factor = dilation_width_factor;
  conv->padding.type = padding_type;
  model->operators.emplace_back(conv);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertDepthToSpaceOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "DepthToSpace");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 1));

  tensorflow::DataType dtype = GetDataTypeAttr(node, "T");
  if (dtype != DT_FLOAT && dtype != DT_UINT8 && dtype != DT_INT32 &&
      dtype != DT_INT64) {
    const auto* enum_descriptor = tensorflow::DataType_descriptor();
    LOG(FATAL) << "TFLite does not support DepthToSpace with type T:"
               << enum_descriptor->FindValueByNumber(dtype)->name() << ". "
               << "T must be one of {DT_FLOAT, DT_UINT8, DT_INT32, DT_INT64}.";
  }
  auto* op = new DepthToSpaceOperator;
  op->inputs.push_back(node.input(0));
  op->outputs.push_back(node.name());
  op->block_size = GetIntAttr(node, "block_size");
  QCHECK_GE(op->block_size, 2);
  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertSpaceToDepthOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "SpaceToDepth");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 1));

  tensorflow::DataType dtype = GetDataTypeAttr(node, "T");
  if (dtype != DT_FLOAT && dtype != DT_UINT8 && dtype != DT_INT32 &&
      dtype != DT_INT64) {
    const auto* enum_descriptor = tensorflow::DataType_descriptor();
    LOG(FATAL) << "TFLite does not support SpaceToDepth with type T:"
               << enum_descriptor->FindValueByNumber(dtype)->name() << ". "
               << "T must be one of {DT_FLOAT, DT_UINT8, DT_INT32, DT_INT64}.";
  }
  auto* op = new SpaceToDepthOperator;
  op->inputs.push_back(node.input(0));
  op->outputs.push_back(node.name());
  op->block_size = GetIntAttr(node, "block_size");
  QCHECK_GE(op->block_size, 2);
  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertBiasAddOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "BiasAdd");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 2));

  const auto& input_name = node.input(0);
  const auto& bias_name = node.input(1);
  CHECK_EQ(GetDataTypeAttr(node, "T"), DT_FLOAT);
  auto* biasadd = new AddOperator;
  biasadd->inputs.push_back(input_name);
  biasadd->inputs.push_back(bias_name);
  biasadd->outputs.push_back(node.name());
  model->operators.emplace_back(biasadd);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertRandomUniform(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "RandomUniform");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 1));

  CHECK_EQ(GetDataTypeAttr(node, "T"), DT_INT32);
  auto op = absl::make_unique<RandomUniformOperator>();
  op->inputs.push_back(node.input(0));
  op->outputs.push_back(node.name());
  op->dtype = ConvertDataType(GetDataTypeAttr(node, "dtype"));
  op->seed = GetIntAttr(node, "seed");
  op->seed2 = GetIntAttr(node, "seed2");
  CHECK(model != nullptr);
  model->operators.emplace_back(std::move(op));
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertIdentityOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK(node.op() == "Identity" || node.op() == "CheckNumerics" ||
        node.op() == "PlaceholderWithDefault" || node.op() == "StopGradient" ||
        node.op() == "Snapshot");
  auto* op = new TensorFlowIdentityOperator;
  // Amazingly, some TensorFlow graphs (at least rajeev_lstm.pb) have
  // identity nodes with multiple inputs, but the other inputs seem
  // to be gratuitous (in the case of rajeev_lstm.pb, these are
  // enumerating the LSTM state arrays). We will just ignore extra
  // inputs beyond the first input.
  QCHECK_GE(node.input_size(), 1)
      << node.op()
      << " node expects at least 1 input other than control dependencies: "
      << node.DebugString();
  const auto& input_name = node.input(0);
  op->inputs.push_back(input_name);
  op->outputs.push_back(node.name());
  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertIdentityNOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "IdentityN");
  for (int i = 0; i < node.input_size(); ++i) {
    auto* op = new TensorFlowIdentityOperator;
    const auto& input_name = node.input(i);
    std::string output_name = node.name();
    if (i > 0) {
      output_name = output_name + ":" + std::to_string(i);
    }
    op->inputs.push_back(input_name);
    op->outputs.push_back(output_name);
    model->operators.emplace_back(op);
  }
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertFakeQuantWithMinMaxArgs(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "FakeQuantWithMinMaxArgs");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 1));
  auto* op = new FakeQuantOperator;
  op->inputs.push_back(node.input(0));
  op->minmax.reset(new MinMax);
  auto& minmax = *op->minmax;
  minmax.min = GetFloatAttr(node, "min");
  minmax.max = GetFloatAttr(node, "max");
  op->outputs.push_back(node.name());
  // tf.fake_quant_with_min_max_args num_bits defaults to 8.
  op->num_bits = HasAttr(node, "num_bits") ? GetIntAttr(node, "num_bits") : 8;
  if (HasAttr(node, "narrow_range")) {
    op->narrow_range = GetBoolAttr(node, "narrow_range");
  }
  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertFakeQuantWithMinMaxVars(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "FakeQuantWithMinMaxVars");
  const int num_inputs = GetInputsCount(node, tf_import_flags);
  QCHECK(num_inputs == 3 || num_inputs == 4)
      << "FakeQuantWithMinMaxVars node expects 3 or 4 inputs other than "
         "control dependencies: "
      << node.DebugString();
  auto* op = new FakeQuantOperator;
  for (int i = 0; i < 3; i++) {
    op->inputs.push_back(node.input(i));
  }
  op->outputs.push_back(node.name());
  op->num_bits = HasAttr(node, "num_bits") ? GetIntAttr(node, "num_bits") : 8;
  if (HasAttr(node, "narrow_range")) {
    op->narrow_range = GetBoolAttr(node, "narrow_range");
  }
  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertSqueezeOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "Squeeze");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 1));
  auto* op = new SqueezeOperator;
  op->inputs.push_back(node.input(0));
  op->outputs.push_back(node.name());

  // When omitted we are to squeeze all dimensions == 1.
  if (HasAttr(node, "squeeze_dims")) {
    const auto& squeeze_dims = GetListAttr(node, "squeeze_dims");
    for (int i = 0; i < squeeze_dims.i_size(); ++i) {
      op->squeeze_dims.push_back(squeeze_dims.i(i));
    }
  }

  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertSplitOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "Split");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 2));
  auto* op = new TensorFlowSplitOperator;
  op->inputs.push_back(node.input(0));
  op->inputs.push_back(node.input(1));
  const int num_split = GetIntAttr(node, "num_split");
  op->outputs.push_back(node.name());
  for (int i = 1; i < num_split; i++) {
    op->outputs.push_back(absl::StrCat(node.name(), ":", i));
  }
  op->num_split = num_split;
  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertSplitVOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "SplitV");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 3));
  auto* op = new TensorFlowSplitVOperator;
  op->inputs.push_back(node.input(0));
  op->inputs.push_back(node.input(1));
  op->inputs.push_back(node.input(2));
  const int num_split = GetIntAttr(node, "num_split");
  op->outputs.push_back(node.name());
  for (int i = 1; i < num_split; i++) {
    op->outputs.push_back(absl::StrCat(node.name(), ":", i));
  }
  op->num_split = num_split;
  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertSwitchOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "Switch");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 2));
  auto* op = new TensorFlowSwitchOperator;
  op->inputs.push_back(node.input(0));
  op->inputs.push_back(node.input(1));
  op->outputs.push_back(node.name());
  // Switch operators have two outputs: "name" and "name:1".
  op->outputs.push_back(node.name() + ":1");
  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertSoftmaxOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "Softmax");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 1));
  const auto& input_name = node.input(0);
  auto* softmax = new SoftmaxOperator;
  softmax->inputs.push_back(input_name);
  softmax->outputs.push_back(node.name());
  // TensorFlow's Softmax doesn't seem to admit a 'beta' parameter.
  CHECK(!node.attr().count("beta"));  // Stab in the dark, just in case.
  if (node.attr().count("_softmax_beta")) {
    softmax->beta = GetFloatAttr(node, "_softmax_beta");
  } else {
    softmax->beta = 1.f;
  }
  model->operators.emplace_back(softmax);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertLRNOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "LRN");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 1));
  const auto& input_name = node.input(0);
  auto* lrn = new LocalResponseNormalizationOperator;
  lrn->inputs.push_back(input_name);
  lrn->outputs.push_back(node.name());
  lrn->range = GetIntAttr(node, "depth_radius");
  lrn->bias = GetFloatAttr(node, "bias");
  lrn->alpha = GetFloatAttr(node, "alpha");
  lrn->beta = GetFloatAttr(node, "beta");
  model->operators.emplace_back(lrn);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertMaxPoolOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "MaxPool");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 1));
  const auto& input_name = node.input(0);
  // We only support NHWC, which is the default data_format.
  // So if data_format is not defined, we're all good.
  if (node.attr().count("data_format")) {
    CHECK_EQ(GetStringAttr(node, "data_format"), "NHWC");
  }
  if (HasAttr(node, "T")) {
    CHECK_EQ(GetDataTypeAttr(node, "T"), DT_FLOAT);
  } else {
    LOG(WARNING) << "Found MaxPool operator missing 'T' attribute";
  }
  auto* maxpool = new MaxPoolOperator;
  maxpool->inputs.push_back(input_name);
  maxpool->outputs.push_back(node.name());
  const auto& strides = GetListAttr(node, "strides");
  CHECK_EQ(strides.i_size(), 4);
  CHECK_EQ(strides.i(0), 1);
  CHECK_EQ(strides.i(3), 1);
  maxpool->stride_height = strides.i(1);
  maxpool->stride_width = strides.i(2);
  const auto& ksize = GetListAttr(node, "ksize");
  CHECK_EQ(ksize.i_size(), 4);
  CHECK_EQ(ksize.i(0), 1);
  CHECK_EQ(ksize.i(3), 1);
  maxpool->kheight = ksize.i(1);
  maxpool->kwidth = ksize.i(2);
  const auto& padding = GetStringAttr(node, "padding");
  if (padding == "SAME") {
    maxpool->padding.type = PaddingType::kSame;
  } else if (padding == "VALID") {
    maxpool->padding.type = PaddingType::kValid;
  } else {
    LOG(FATAL) << "Bad padding (only SAME and VALID are supported)";
  }
  model->operators.emplace_back(maxpool);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertAvgPoolOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "AvgPool");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 1));
  const auto& input_name = node.input(0);
  // We only support NHWC, which is the default data_format.
  // So if data_format is not defined, we're all good.
  if (node.attr().count("data_format")) {
    CHECK_EQ(GetStringAttr(node, "data_format"), "NHWC");
  }
  CHECK_EQ(GetDataTypeAttr(node, "T"), DT_FLOAT);
  auto* avgpool = new AveragePoolOperator;
  avgpool->inputs.push_back(input_name);
  avgpool->outputs.push_back(node.name());
  const auto& strides = GetListAttr(node, "strides");
  CHECK_EQ(strides.i_size(), 4);
  CHECK_EQ(strides.i(0), 1);
  CHECK_EQ(strides.i(3), 1);
  avgpool->stride_height = strides.i(1);
  avgpool->stride_width = strides.i(2);
  const auto& ksize = GetListAttr(node, "ksize");
  CHECK_EQ(ksize.i_size(), 4);
  CHECK_EQ(ksize.i(0), 1);
  CHECK_EQ(ksize.i(3), 1);
  avgpool->kheight = ksize.i(1);
  avgpool->kwidth = ksize.i(2);
  const auto& padding = GetStringAttr(node, "padding");
  if (padding == "SAME") {
    avgpool->padding.type = PaddingType::kSame;
  } else if (padding == "VALID") {
    avgpool->padding.type = PaddingType::kValid;
  } else {
    LOG(FATAL) << "Bad padding (only SAME and VALID are supported)";
  }
  model->operators.emplace_back(avgpool);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertBatchMatMulOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 2));

  auto* batch_matmul = new BatchMatMulOperator;
  // https://www.tensorflow.org/versions/r0.12/api_docs/python/math_ops/matrix_math_functions
  if (HasAttr(node, "adj_x")) {
    batch_matmul->adj_x = GetBoolAttr(node, "adj_x");
  }
  if (HasAttr(node, "adj_y")) {
    batch_matmul->adj_y = GetBoolAttr(node, "adj_y");
  }
  batch_matmul->inputs = {node.input(0), node.input(1)};
  batch_matmul->outputs = {node.name()};

  // For Flex mode. Please read the comments of the function.
  RetainTensorFlowNodeDef(node, batch_matmul);

  model->operators.emplace_back(batch_matmul);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertMatMulOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 2));

  CHECK(!HasAttr(node, "adjoint_a") ||
        (GetBoolAttr(node, "adjoint_a") == false));
  CHECK(!HasAttr(node, "adjoint_b") ||
        (GetBoolAttr(node, "adjoint_b") == false));

  auto* matmul = new TensorFlowMatMulOperator;
  if (HasAttr(node, "transpose_a")) {
    matmul->transpose_a = GetBoolAttr(node, "transpose_a");
  }
  if (HasAttr(node, "transpose_b")) {
    matmul->transpose_b = GetBoolAttr(node, "transpose_b");
  }

  matmul->inputs = {node.input(0), node.input(1)};
  matmul->outputs = {node.name()};
  model->operators.emplace_back(matmul);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertConcatOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  Operator* op = nullptr;
  if (node.op() == "Concat") {
    op = new TensorFlowConcatOperator;
  } else if (node.op() == "ConcatV2") {
    op = new TensorFlowConcatV2Operator;
  } else {
    LOG(FATAL) << "Expected Concat or ConcatV2";
  }
  const int num_inputs = GetInputsCount(node, tf_import_flags);
  QCHECK_GE(num_inputs, 2)
      << node.op()
      << " node expects at least 2 inputs other than control dependencies: "
      << node.DebugString();
  CHECK_EQ(num_inputs, 1 + GetIntAttr(node, "N"));
  for (int i = 0; i < num_inputs; ++i) {
    op->inputs.push_back(node.input(i));
  }
  op->outputs.push_back(node.name());
  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertMirrorPadOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  if (node.op() != "MirrorPad") {
    LOG(FATAL) << "Expected MirrorPad.";
  }
  const int num_inputs = GetInputsCount(node, tf_import_flags);
  CHECK_EQ(num_inputs, 2);
  auto* op = new MirrorPadOperator;
  for (int i = 0; i < num_inputs; ++i) {
    op->inputs.push_back(node.input(i));
  }
  op->outputs.push_back(node.name());
  const auto mode = GetStringAttr(node, "mode");
  if (mode == "REFLECT") {
    op->mode = toco::MirrorPadMode::kReflect;
  } else if (mode == "SYMMETRIC") {
    op->mode = toco::MirrorPadMode::kSymmetric;
  }

  model->operators.emplace_back(op);

  return tensorflow::Status::OK();
}

static constexpr int kAnyNumInputs = -1;

enum FlexSupport { kFlexOk, kFlexNotOk };

// This method supports simple operators without additional attributes.
// Converts a simple operator that takes no attributes. The list of inputs is
// taken from the given NodeDef, and its number must match NumInputs, unless
// kAnyNumInputs is passed in. If kFlexOk is passed in the resulting operator
// will be eligible for being exported as a flex op.
template <typename Op, int NumInputs, int NumOutputs, FlexSupport flex>
tensorflow::Status ConvertSimpleOperatorGeneric(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  if (NumInputs != kAnyNumInputs) {
    TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, NumInputs));
  }
  auto* op = new Op;
  const int num_inputs = GetInputsCount(node, tf_import_flags);
  for (int i = 0; i < num_inputs; ++i) {
    op->inputs.push_back(node.input(i));
  }
  op->outputs.push_back(node.name());
  if (NumOutputs > 1) {
    for (int i = 1; i < NumOutputs; ++i) {
      op->outputs.push_back(node.name() + ":" + std::to_string(i));
    }
  }

  if (flex == kFlexOk) {
    RetainTensorFlowNodeDef(node, op);
  }

  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

// Convert a simple operator which is not valid as a flex op.
template <typename Op, int NumInputs, int NumOutputs>
tensorflow::Status ConvertSimpleOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  return ConvertSimpleOperatorGeneric<Op, NumInputs, NumOutputs, kFlexNotOk>(
      node, tf_import_flags, model_flags, model);
}

// Convert a simple operator which is valid as a flex op.
template <typename Op, int NumInputs, int NumOutputs>
tensorflow::Status ConvertSimpleOperatorFlexOk(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  return ConvertSimpleOperatorGeneric<Op, NumInputs, NumOutputs, kFlexOk>(
      node, tf_import_flags, model_flags, model);
}

// Same as ConvertConstOperator, but revert to ConvertUnsupportedOperator if
// the types are not supported. Converting Const operators here avoids
// expensive copies of the protocol buffers downstream in the flex delegate.
tensorflow::Status ConditionallyConvertConstOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  // We avoid incomplete and zero shapes because the resulting arrays
  // are not completely compatible with Eager/TensorFlow.
  const auto& tensor = GetTensorAttr(node, "value");
  const auto& shape = tensor.tensor_shape();
  for (const auto& dim : shape.dim()) {
    if (dim.size() <= 0) {
      return ConvertUnsupportedOperator(node, tf_import_flags, model_flags,
                                        model);
    }
  }
  switch (GetDataTypeAttr(node, "dtype")) {
    case DT_FLOAT:
    case DT_INT32:
    case DT_QUINT8:
    case DT_INT64:
    case DT_STRING:
    case DT_BOOL:
    case DT_COMPLEX64:
      return ConvertConstOperator(node, tf_import_flags, model_flags, model);
    default:
      return ConvertUnsupportedOperator(node, tf_import_flags, model_flags,
                                        model);
  }
}

tensorflow::Status ConvertStridedSliceOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "StridedSlice");
  // TODO(soroosh): The 4th input (strides) should be e optional, to be
  // consistent with TF.
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 4));

  auto* op = new StridedSliceOperator;
  for (const auto& input : node.input()) {
    op->inputs.push_back(input);
  }
  op->outputs.push_back(node.name());

  op->begin_mask =
      HasAttr(node, "begin_mask") ? GetIntAttr(node, "begin_mask") : 0;
  op->ellipsis_mask =
      HasAttr(node, "ellipsis_mask") ? GetIntAttr(node, "ellipsis_mask") : 0;
  op->end_mask = HasAttr(node, "end_mask") ? GetIntAttr(node, "end_mask") : 0;
  op->new_axis_mask =
      HasAttr(node, "new_axis_mask") ? GetIntAttr(node, "new_axis_mask") : 0;
  op->shrink_axis_mask = HasAttr(node, "shrink_axis_mask")
                             ? GetIntAttr(node, "shrink_axis_mask")
                             : 0;

  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertPlaceholderOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK(node.op() == "Placeholder" || node.op() == "LegacyFedInput");
  if (node.op() == "Placeholder") {
    TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 0));
  }

  bool inside_input_arrays = false;
  for (const auto& input_array : model_flags.input_arrays()) {
    if (node.name() == input_array.name()) {
      inside_input_arrays = true;
      break;
    }
  }

  if (!inside_input_arrays) {
    model->AddInvalidInputArray(node.name());
  }

  auto& array = model->GetOrCreateArray(node.name());
  if (node.attr().count("dtype")) {
    array.data_type = ConvertDataType(GetDataTypeAttr(node, "dtype"));
  }
  if (node.attr().count("shape")) {
    const auto& shape = GetShapeAttr(node, "shape");
    auto num_dims = shape.dim_size();
    // TODO(b/62716978): This logic needs to be revisited.  During dims
    // refactoring it is an interim fix.
    if (num_dims > 0 && !HasWildcardDimension(shape)) {
      auto& dst_array_dims = *array.mutable_shape()->mutable_dims();
      dst_array_dims.resize(num_dims);
      for (std::size_t i = 0; i < num_dims; i++) {
        dst_array_dims[i] = shape.dim(i).size();
      }
    }
  }
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertNoOpOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertCastOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "Cast");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 1));
  const auto tf_src_dtype = GetDataTypeAttr(node, "SrcT");
  const auto tf_dst_dtype = GetDataTypeAttr(node, "DstT");
  auto* op = new CastOperator;
  op->src_data_type = ConvertDataType(tf_src_dtype);
  op->dst_data_type = ConvertDataType(tf_dst_dtype);
  op->inputs.push_back(node.input(0));
  op->outputs.push_back(node.name());
  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertFloorOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "Floor");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 1));
  const auto data_type = GetDataTypeAttr(node, "T");
  CHECK(data_type == DT_FLOAT);
  auto* op = new FloorOperator;
  op->inputs.push_back(node.input(0));
  op->outputs.push_back(node.name());
  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertCeilOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "Ceil");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 1));
  const auto data_type = GetDataTypeAttr(node, "T");
  CHECK(data_type == DT_FLOAT);
  auto* op = new CeilOperator;
  op->inputs.push_back(node.input(0));
  op->outputs.push_back(node.name());
  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertRoundOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "Round");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 1));
  const auto data_type = GetDataTypeAttr(node, "T");
  CHECK(data_type == DT_FLOAT);
  auto* op = new RoundOperator;
  op->inputs.push_back(node.input(0));
  op->outputs.push_back(node.name());
  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertGatherOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK(node.op() == "Gather" || node.op() == "GatherV2");
  if (node.op() == "Gather")
    TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 2));
  if (node.op() == "GatherV2")
    TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 3));
  const auto indices_data_type = GetDataTypeAttr(node, "Tindices");
  CHECK(indices_data_type == DT_INT32 || indices_data_type == DT_INT64);
  auto* op = new GatherOperator;
  op->inputs.push_back(node.input(0));
  op->inputs.push_back(node.input(1));
  if (node.input_size() >= 3) {
    // GatherV2 form where we are provided an axis. It may be either a constant
    // or runtime defined value, so we just wire up the array and let
    // ResolveGatherAttributes take care of it later on.
    const auto axis_data_type = GetDataTypeAttr(node, "Taxis");
    CHECK(axis_data_type == DT_INT32 || axis_data_type == DT_INT64);
    op->inputs.push_back(node.input(2));
  } else {
    // Gather form that assumes axis=0.
    op->axis = {0};
  }
  op->outputs.push_back(node.name());
  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertGatherNdOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "GatherNd");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 2));
  const auto indices_data_type = GetDataTypeAttr(node, "Tindices");
  CHECK(indices_data_type == DT_INT32 || indices_data_type == DT_INT64);
  auto* op = new GatherNdOperator;
  op->inputs.push_back(node.input(0));
  op->inputs.push_back(node.input(1));
  op->outputs.push_back(node.name());
  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

template <typename Op>
tensorflow::Status ConvertArgMinMaxOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 2));
  const auto axis_data_type =
      HasAttr(node, "Tidx") ? GetDataTypeAttr(node, "Tidx") : DT_INT32;
  const auto output_type = HasAttr(node, "output_type")
                               ? GetDataTypeAttr(node, "output_type")
                               : DT_INT64;
  CHECK(axis_data_type == DT_INT64 || axis_data_type == DT_INT32);
  CHECK(output_type == DT_INT64 || output_type == DT_INT32);
  auto* op = new Op;
  op->output_data_type = ConvertDataType(output_type);
  op->inputs.push_back(node.input(0));
  op->inputs.push_back(node.input(1));
  op->outputs.push_back(node.name());
  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertArgMaxOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "ArgMax");
  return ConvertArgMinMaxOperator<ArgMaxOperator>(node, tf_import_flags,
                                                  model_flags, model);
}

tensorflow::Status ConvertArgMinOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "ArgMin");
  return ConvertArgMinMaxOperator<ArgMinOperator>(node, tf_import_flags,
                                                  model_flags, model);
}

tensorflow::Status ConvertResizeBilinearOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "ResizeBilinear");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 2));
  auto* op = new ResizeBilinearOperator;

  op->align_corners = false;
  op->half_pixel_centers = false;
  if (HasAttr(node, "align_corners")) {
    op->align_corners = GetBoolAttr(node, "align_corners");
  }
  if (HasAttr(node, "half_pixel_centers")) {
    op->half_pixel_centers = GetBoolAttr(node, "half_pixel_centers");
  }

  op->inputs.push_back(node.input(0));
  op->inputs.push_back(node.input(1));
  op->outputs.push_back(node.name());
  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertResizeNearestNeighborOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "ResizeNearestNeighbor");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 2));
  auto* op = new ResizeNearestNeighborOperator;

  op->align_corners = false;
  op->half_pixel_centers = false;
  if (HasAttr(node, "align_corners")) {
    op->align_corners = GetBoolAttr(node, "align_corners");
  }
  if (HasAttr(node, "half_pixel_centers")) {
    op->half_pixel_centers = GetBoolAttr(node, "half_pixel_centers");
  }

  op->inputs.push_back(node.input(0));
  op->inputs.push_back(node.input(1));
  op->outputs.push_back(node.name());
  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertBatchNormWithGlobalNormalizationOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "BatchNormWithGlobalNormalization");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 5));

  // TODO(ahentz): to really match tensorflow we need to add variance_epsilon
  // to the input, before feeding it into TensorFlowRsqrtOperator.
  // CHECK_EQ(GetFloatAttr(node, "variance_epsilon"), 0.001f);

  std::string multiplier = node.name() + "_mul";
  if (GetBoolAttr(node, "scale_after_normalization")) {
    // Create graph:
    //   v -> RSQRT ->
    //                 MUL  -> multiplier
    //   gamma  ----->
    std::string rsqrt = node.name() + "_rsqrt";

    auto* rsqrt_op = new TensorFlowRsqrtOperator;
    rsqrt_op->inputs.push_back(node.input(2));
    rsqrt_op->outputs.push_back(rsqrt);
    model->operators.emplace_back(rsqrt_op);

    auto* mul_op = new MulOperator;
    mul_op->inputs.push_back(rsqrt);
    mul_op->inputs.push_back(node.input(4));
    mul_op->outputs.push_back(multiplier);
    model->operators.emplace_back(mul_op);
  } else {
    // Create graph:
    //   v -> RSQRT -> multiplier
    auto* rsqrt_op = new TensorFlowRsqrtOperator;
    rsqrt_op->inputs.push_back(node.input(2));
    rsqrt_op->outputs.push_back(multiplier);
    model->operators.emplace_back(rsqrt_op);
  }

  auto* op = new BatchNormalizationOperator;
  op->global_normalization = true;

  op->inputs.push_back(node.input(0));
  op->inputs.push_back(node.input(1));
  op->inputs.push_back(multiplier);
  op->inputs.push_back(node.input(3));
  op->outputs.push_back(node.name());

  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertFusedBatchNormOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK((node.op() == "FusedBatchNorm") || (node.op() == "FusedBatchNormV3"));
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 5));

  // Declare shortcuts for the inputs.
  const std::string& gamma_input = node.input(1);
  const std::string& beta_input = node.input(2);
  const std::string& moving_mean_input = node.input(3);
  const std::string& moving_variance_input = node.input(4);

  // Create an array holding the epsilon value (typically, 0.001).
  const std::string epsilon_array_name =
      CreateConstArray<ArrayDataType::kFloat>(model,
                                              node.name() + "_epsilon_array",
                                              {GetFloatAttr(node, "epsilon")});

  // Add epsilon to the moving variance.
  const std::string epsilon_add_op_name = node.name() + "_epsilon";
  auto* epsilon_add_op = new AddOperator;
  epsilon_add_op->inputs.push_back(moving_variance_input);
  epsilon_add_op->inputs.push_back(epsilon_array_name);
  epsilon_add_op->outputs.push_back(epsilon_add_op_name);
  model->operators.emplace_back(epsilon_add_op);

  // Take the inverse square root of the (variance + epsilon).
  const std::string rsqrt_op_name = node.name() + "_rsqrt";
  auto* rsqrt_op = new TensorFlowRsqrtOperator;
  rsqrt_op->inputs.push_back(epsilon_add_op_name);
  rsqrt_op->outputs.push_back(rsqrt_op_name);
  model->operators.emplace_back(rsqrt_op);

  // Multiply the result by gamma.
  const std::string multiplier = node.name() + "_mul";
  auto* mul_op = new MulOperator;
  mul_op->inputs.push_back(rsqrt_op_name);
  mul_op->inputs.push_back(gamma_input);
  mul_op->outputs.push_back(multiplier);
  model->operators.emplace_back(mul_op);

  // Now we have all required inputs for the BatchNormalizationOperator.
  auto* op = new BatchNormalizationOperator;
  op->global_normalization = true;

  op->inputs.push_back(node.input(0));
  op->inputs.push_back(moving_mean_input);
  op->inputs.push_back(multiplier);
  op->inputs.push_back(beta_input);
  op->outputs.push_back(node.name());

  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertSpaceToBatchNDOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "SpaceToBatchND");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 3));
  CHECK_EQ(GetDataTypeAttr(node, "Tblock_shape"), DT_INT32);
  CHECK_EQ(GetDataTypeAttr(node, "Tpaddings"), DT_INT32);
  auto* op = new SpaceToBatchNDOperator;
  op->inputs.push_back(node.input(0));
  op->inputs.push_back(node.input(1));
  op->inputs.push_back(node.input(2));
  op->outputs.push_back(node.name());
  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertBatchToSpaceNDOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "BatchToSpaceND");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 3));
  CHECK_EQ(GetDataTypeAttr(node, "Tblock_shape"), DT_INT32);
  CHECK_EQ(GetDataTypeAttr(node, "Tcrops"), DT_INT32);
  auto* op = new BatchToSpaceNDOperator;
  op->inputs.push_back(node.input(0));
  op->inputs.push_back(node.input(1));
  op->inputs.push_back(node.input(2));
  op->outputs.push_back(node.name());
  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

template <typename T>
tensorflow::Status ConvertReduceOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 2));
  auto* op = new T;
  op->inputs.push_back(node.input(0));
  op->inputs.push_back(node.input(1));
  op->outputs.push_back(node.name());
  model->operators.emplace_back(op);
  if (HasAttr(node, "keepdims")) {
    op->keep_dims = GetBoolAttr(node, "keepdims");
  } else if (HasAttr(node, "keep_dims")) {
    op->keep_dims = GetBoolAttr(node, "keep_dims");
  }
  return tensorflow::Status::OK();
}

// TODO(b/139320642): Add test when fused op is supported.
tensorflow::Status ConvertSvdfOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "Svdf");
  const int input_size = GetInputsCount(node, tf_import_flags);
  QCHECK(input_size == 4 || input_size == 5)
      << "Svdf node expects 3 or 4 inputs other than control dependencies: "
      << node.DebugString();
  bool has_bias = (input_size == 5);
  auto* op = new SvdfOperator;
  int index = 0;
  op->inputs.push_back(node.input(index++));
  op->inputs.push_back(node.input(index++));
  op->inputs.push_back(node.input(index++));
  if (has_bias) {
    op->inputs.push_back(node.input(index++));
  }
  op->inputs.push_back(node.input(index));
  op->outputs.push_back(node.name());
  if (node.attr().at("ActivationFunction").s() == "Relu") {
    op->fused_activation_function = FusedActivationFunctionType::kRelu;
  } else {
    op->fused_activation_function = FusedActivationFunctionType::kNone;
  }
  op->rank = node.attr().at("Rank").i();
  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

// This is just bare bones support to get the shapes to propagate.
tensorflow::Status ConvertTransposeConvOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "Conv2DBackpropInput");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 3));
  auto* op = new TransposeConvOperator;
  op->inputs.push_back(node.input(0));
  op->inputs.push_back(node.input(1));
  op->inputs.push_back(node.input(2));
  op->outputs.push_back(node.name());
  const auto& strides = GetListAttr(node, "strides");
  op->stride_height = strides.i(1);
  op->stride_width = strides.i(2);
  CHECK_EQ(strides.i_size(), 4)
      << "Can only import TransposeConv ops with 4D strides. TensorFlow op \""
      << node.name() << "\" has " << strides.i_size() << "D strides.";
  CHECK((strides.i(0) == 1) && (strides.i(3) == 1))
      << "Can only import TransposeConv ops with striding along the height "
         "(1st) or width (2nd) axis. TensorFlow op \""
      << node.name() << "\" had strides:[ " << strides.i(0) << ", "
      << strides.i(1) << ", " << strides.i(2) << ", " << strides.i(3) << "].";
  op->stride_height = strides.i(1);
  op->stride_width = strides.i(2);
  if (HasAttr(node, "dilations")) {
    const auto& dilations = GetListAttr(node, "dilations");
    CHECK_EQ(dilations.i_size(), 4)
        << "Dilation unsupported in TransposeConv. TensorFlow op \""
        << node.name() << "\" had dilations";
    CHECK((dilations.i(0) == 1) && (dilations.i(1) == 1) &&
          (dilations.i(2) == 1) && (dilations.i(3) == 1))
        << "Dilation unsupported in TransposeConv. TensorFlow op \""
        << node.name() << "\" had dilations:[ " << dilations.i(0) << ", "
        << dilations.i(1) << ", " << dilations.i(2) << ", " << dilations.i(3)
        << "].";
  }

  const std::string& weights_name = node.input(TransposeConvOperator::WEIGHTS);
  const std::string& transposed_weights_name = weights_name + "_transposed";
  // Check if a TransposeOperator was already created for these weights
  // (can happen when multiple layers share the same weights).
  const Operator* existing_transpose =
      GetOpWithOutput(*model, transposed_weights_name);
  if (existing_transpose) {
    CHECK(existing_transpose->type == OperatorType::kTranspose);
  } else {
    // Transpose weights from HWOI order to OHWI order, which is more efficient
    // for computation. (Note that TensorFlow considers the order as HWIO
    // because they consider this a backward conv, inverting the sense of
    // input/output.)
    TransposeOperator* transpose = new TransposeOperator;
    std::string perm_array = CreateConstArray<ArrayDataType::kInt32>(
        model, node.name() + "_transpose_perm", {2, 0, 1, 3});
    transpose->inputs = {weights_name, perm_array};
    transpose->outputs = {transposed_weights_name};
    model->operators.emplace_back(transpose);
  }
  op->inputs[1] = transposed_weights_name;

  auto const& padding = GetStringAttr(node, "padding");
  if (padding == "SAME") {
    op->padding.type = PaddingType::kSame;
  } else if (padding == "VALID") {
    op->padding.type = PaddingType::kValid;
  } else {
    LOG(FATAL) << "Only SAME and VALID padding supported on "
                  "Conv2DBackpropInput nodes.";
  }
  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertRangeOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "Range");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 3));
  auto* op = new RangeOperator;
  if (HasAttr(node, "Tidx")) {
    const auto dtype = toco::GetDataTypeAttr(node, "Tidx");
    CHECK(dtype == DT_UINT8 || dtype == DT_INT32 || dtype == DT_INT64 ||
          dtype == DT_FLOAT);
    op->dtype = ConvertDataType(dtype);
  }
  op->inputs.push_back(node.input(0));
  op->inputs.push_back(node.input(1));
  op->inputs.push_back(node.input(2));
  op->outputs.push_back(node.name());

  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

// Note that it's easy to confuse/conflate "Stack" and "Pack" operators, but
// they aren't the same thing.  tf.stack results in a "Pack" operator.  "Stack"
// operators also exist, but involve manipulating the TF runtime stack, and are
// not directly related to tf.stack() usage.
tensorflow::Status ConvertPackOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "Pack");
  auto op = absl::make_unique<PackOperator>();
  const int num_inputs = GetInputsCount(node, tf_import_flags);
  QCHECK_GE(num_inputs, 1)
      << node.op()
      << " node expects at least 1 input other than control dependencies: "
      << node.DebugString();
  CHECK_EQ(num_inputs, GetIntAttr(node, "N"));
  for (int i = 0; i < num_inputs; ++i) {
    op->inputs.push_back(node.input(i));
  }
  op->values_count = HasAttr(node, "N") ? GetIntAttr(node, "N") : num_inputs;
  op->axis = HasAttr(node, "axis") ? GetIntAttr(node, "axis") : 0;
  op->dtype = ConvertDataType(toco::GetDataTypeAttr(node, "T"));
  op->outputs.push_back(node.name());
  model->operators.emplace_back(std::move(op));
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertUnpackOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "Unpack");
  auto op = absl::make_unique<UnpackOperator>();
  const int num_inputs = GetInputsCount(node, tf_import_flags);
  QCHECK_EQ(num_inputs, 1);
  op->inputs.push_back(node.input(0));
  op->num = GetIntAttr(node, "num");
  op->axis = HasAttr(node, "axis") ? GetIntAttr(node, "axis") : 0;
  op->dtype = ConvertDataType(toco::GetDataTypeAttr(node, "T"));

  op->outputs.push_back(node.name());  // Implicit :0.
  for (int i = 1; i < op->num; ++i) {
    op->outputs.push_back(node.name() + ":" + std::to_string(i));
  }
  model->operators.emplace_back(std::move(op));
  return tensorflow::Status::OK();
}

// Some TensorFlow ops only occur in graph cycles, representing
// control flow. We do not currently support control flow, so we wouldn't
// be able to fully support such graphs, including performing inference,
// anyway. However, rather than erroring out early on graphs being cyclic,
// it helps to at least support these just enough to allow getting a
// graph visualization. This is not trivial, as we require graphs to be
// acyclic aside from RNN back-edges. The solution is to special-case
// such ops as RNN back-edges, which is technically incorrect (does not
// allow representing the op's semantics) but good enough to get a
// graph visualization.
tensorflow::Status ConvertOperatorSpecialCasedAsRNNBackEdge(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  // At the moment, the only type of operator special-cased in this way is
  // NextIteration, occurring only in control-flow cycles.
  CHECK_EQ(node.op(), "NextIteration");
  CHECK_EQ(node.input_size(), 1);
  auto* rnn_state = model->flags.add_rnn_states();
  // This RNN state is not explicitly created by the user, so it's
  // OK for some later graph transformation to discard it.
  rnn_state->set_discardable(true);
  rnn_state->set_state_array(node.name());
  rnn_state->set_back_edge_source_array(node.input(0));
  // TODO(tianjuny): Temporary set the size to 1 to avoid transient array
  // allocation crash. The real value should depend on the hidden_size of RNN.
  rnn_state->set_size(1);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertShapeOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "Shape");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 1));
  const auto out_type =
      HasAttr(node, "out_type") ? GetDataTypeAttr(node, "out_type") : DT_INT32;
  CHECK(out_type == DT_INT64 || out_type == DT_INT32);
  auto op = absl::make_unique<TensorFlowShapeOperator>();
  op->output_data_type = ConvertDataType(out_type);
  op->inputs.push_back(node.input(0));
  op->outputs.push_back(node.name());
  model->operators.push_back(std::move(op));
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertReverseSequenceOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "ReverseSequence");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 2));
  auto op = absl::make_unique<ReverseSequenceOperator>();
  if (HasAttr(node, "seq_dim")) {
    op->seq_dim = GetIntAttr(node, "seq_dim");
  }
  // In tf.reverse_sequence, batch_dim defaults to 0.
  op->batch_dim =
      HasAttr(node, "batch_dim") ? GetIntAttr(node, "batch_dim") : 0;
  const int num_inputs = GetInputsCount(node, tf_import_flags);
  for (int i = 0; i < num_inputs; ++i) {
    op->inputs.push_back(node.input(i));
  }
  op->outputs.push_back(node.name());
  model->operators.push_back(std::move(op));
  return tensorflow::Status::OK();
}

void StripCaretFromArrayNames(Model* model) {
  for (auto& op : model->operators) {
    for (auto& input : op->inputs) {
      input = std::string(absl::StripPrefix(input, "^"));
    }
    for (auto& output : op->outputs) {
      output = std::string(absl::StripPrefix(output, "^"));
    }
  }
  for (auto& array : model->GetArrayMap()) {
    if (absl::StartsWith(array.first, "^")) {
      LOG(FATAL) << "What?";
    }
  }
}

void StripZeroOutputIndexFromInputs(NodeDef* node) {
  for (auto& input : *node->mutable_input()) {
    input = std::string(absl::StripSuffix(input, ":0"));
  }
}

// In TensorFlow GraphDef, when a node has multiple outputs, they are named
// name:0, name:1, ...
// where 'name' is the node's name(). Just 'name' is an equivalent shorthand
// form for name:0.
// A TensorFlow GraphDef does not explicitly list all the outputs of each node
// (unlike inputs), it being implied by the node's name and operator type
// (the latter implies the number of outputs).
// This makes it non-trivial for us to reconstruct the list of all arrays
// present in the graph and, for each operator, the list of its outputs.
// We do that by taking advantage of the fact that
// at least each node lists explicitly its inputs, so after we've loaded
// all nodes, we can use that information.
void AddExtraOutputs(Model* model) {
  // Construct the list of all arrays consumed by anything in the graph.
  std::vector<std::string> consumed_arrays;
  // Add arrays consumed by an op.
  for (const auto& consumer_op : model->operators) {
    for (const std::string& input : consumer_op->inputs) {
      consumed_arrays.push_back(input);
    }
  }
  // Add global outputs of the model.
  for (const std::string& output_array : model->flags.output_arrays()) {
    consumed_arrays.push_back(output_array);
  }
  // Add arrays consumed by a RNN back-edge.
  for (const auto& rnn_state : model->flags.rnn_states()) {
    consumed_arrays.push_back(rnn_state.back_edge_source_array());
  }
  // Now add operator outputs so that all arrays that are consumed,
  // are produced.
  for (const std::string& consumed_array : consumed_arrays) {
    // Test if consumed_array is already the output of some op.
    // This has occurred in a model where separate nodes had names of the form
    // foo:$i with the same base name foo.
    if (GetOpWithOutput(*model, consumed_array)) {
      continue;
    }
    // Split the consumed array name into the form name:output_index.
    const std::vector<std::string>& split = absl::StrSplit(consumed_array, ':');
    // If not of the form name:output_index, then this is not an additional
    // output of a node with multiple outputs, so nothing to do here.
    if (split.size() != 2) {
      continue;
    }
    int output_index = 0;
    if (!absl::SimpleAtoi(split[1], &output_index)) {
      continue;
    }
    // Each op is initially recorded as producing at least the array that
    // has its name. We use that to identify the producer node.
    auto* producer_op = GetOpWithOutput(*model, split[0]);
    if (!producer_op) {
      continue;
    }
    // Add extra outputs to that producer node, all the way to the
    // output_index.
    while (producer_op->outputs.size() <= output_index) {
      using toco::port::StringF;
      producer_op->outputs.push_back(
          StringF("%s:%d", split[0], producer_op->outputs.size()));
    }
  }
}

bool InlineAllFunctions(GraphDef* graphdef) {
  if (graphdef->library().function().empty()) {
    VLOG(kLogLevelModelUnchanged) << "No functions to inline.";
    return false;
  }

  // Override "_noinline" attribute on all functions
  GraphDef graphdef_copy(*graphdef);
  for (auto& function :
       (*graphdef_copy.mutable_library()->mutable_function())) {
    auto* attributes = function.mutable_attr();
    if (attributes->count(tensorflow::kNoInlineAttr) != 0) {
      (*attributes)[tensorflow::kNoInlineAttr].set_b(false);
    }
  }

  // Construct minimum resources needed to use ExpandInlineFunctions().
  tensorflow::SessionOptions options;
  auto* device_count = options.config.mutable_device_count();
  device_count->insert({"CPU", 1});
  std::vector<std::unique_ptr<tensorflow::Device>> devices;
  TF_CHECK_OK(tensorflow::DeviceFactory::AddDevices(
      options, "/job:localhost/replica:0/task:0", &devices));

  tensorflow::FunctionLibraryDefinition fld(tensorflow::OpRegistry::Global(),
                                            graphdef_copy.library());
  tensorflow::StaticDeviceMgr device_mgr(std::move(devices));
  tensorflow::ProcessFunctionLibraryRuntime pflr(
      &device_mgr, tensorflow::Env::Default(), &options.config,
      TF_GRAPH_DEF_VERSION, &fld,
      options.config.graph_options().optimizer_options(), nullptr);
  tensorflow::FunctionLibraryRuntime* flr;
  flr = pflr.GetFLR("/job:localhost/replica:0/task:0/cpu:0");

  tensorflow::Graph graph(fld);
  tensorflow::ImportGraphDefOptions gc_opts;
  gc_opts.validate_shape = false;
  const auto& tf_convert_status = tensorflow::ImportGraphDef(
      gc_opts, graphdef_copy, &graph, nullptr, nullptr);
  if (!tf_convert_status.ok()) {
    LOG(ERROR) << "tensorflow::ImportGraphDef failed with status: "
               << tf_convert_status.ToString();
    return false;
  }

  // Iterate over the graph until there are no more nodes to be inlined.
  bool graph_modified = false;
  while (tensorflow::ExpandInlineFunctions(flr, &graph)) {
    graph_modified = true;
  }

  // Output inlined graph
  if (graph_modified) {
    LOG(INFO) << "Found and inlined TensorFlow functions.";
    graph.ToGraphDef(graphdef);
  }
  return graph_modified;
}

tensorflow::Status ConvertTopKV2Operator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK((node.op() == "TopK") || (node.op() == "TopKV2"));
  auto op = absl::make_unique<TopKV2Operator>();
  op->inputs.push_back(node.input(0));
  // K can be encoded as attr (TopK) convert it to a const.
  if (HasAttr(node, "k")) {
    std::string k_array = CreateConstArray<ArrayDataType::kInt32>(
        model, node.name() + "k", {static_cast<int32>(GetIntAttr(node, "k"))});
    op->inputs.push_back(k_array);
  } else {
    TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 2));
    op->inputs.push_back(node.input(1));
  }
  // The op has two outputs.
  op->outputs.push_back(node.name());
  op->outputs.push_back(node.name() + ":1");
  model->operators.emplace_back(op.release());
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertDynamicPartitionOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  auto op = absl::make_unique<DynamicPartitionOperator>();
  CHECK(HasAttr(node, "num_partitions"));
  op->num_partitions = GetIntAttr(node, "num_partitions");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 2));
  op->inputs.push_back(node.input(0));
  op->inputs.push_back(node.input(1));
  CHECK_GT(op->num_partitions, 1);
  op->outputs.push_back(node.name());  // Implicit :0.
  for (int i = 1; i < op->num_partitions; ++i) {
    op->outputs.push_back(node.name() + ":" + std::to_string(i));
  }
  model->operators.emplace_back(op.release());
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertDynamicStitchOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  // The parallel and non-parallel variants are the same besides whether they
  // have a parallel loop; there are no behavioral differences.
  CHECK(node.op() == "DynamicStitch" || node.op() == "ParallelDynamicStitch");
  auto op = absl::make_unique<DynamicStitchOperator>();
  CHECK(HasAttr(node, "N"));
  op->num_partitions = GetIntAttr(node, "N");
  // Expect all ID partitions + all value partitions.
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, op->num_partitions * 2));
  for (int i = 0; i < op->num_partitions * 2; ++i) {
    op->inputs.push_back(node.input(i));
  }
  op->outputs.push_back(node.name());
  model->operators.emplace_back(op.release());
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertSparseToDenseOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "SparseToDense");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 4));

  auto* op = new SparseToDenseOperator;
  for (const std::string& input : node.input()) {
    op->inputs.push_back(input);
  }
  op->outputs.push_back(node.name());

  op->validate_indices = HasAttr(node, "validate_indices")
                             ? GetBoolAttr(node, "validate_indices")
                             : true;
  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertOneHotOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "OneHot");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 4));

  const auto dtype = GetDataTypeAttr(node, "T");
  // TODO(b/111744875): Support DT_UINT8 and quantization.
  CHECK(dtype == DT_INT32 || dtype == DT_INT64 || dtype == DT_FLOAT ||
        dtype == DT_BOOL);

  auto op = absl::make_unique<OneHotOperator>();
  op->axis = HasAttr(node, "axis") ? GetIntAttr(node, "axis") : -1;
  for (const std::string& input : node.input()) {
    op->inputs.push_back(input);
  }
  op->outputs.push_back(node.name());
  model->operators.emplace_back(op.release());
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertCTCBeamSearchDecoderOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "CTCBeamSearchDecoder");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 2));

  auto* op = new CTCBeamSearchDecoderOperator;
  for (const std::string& input : node.input()) {
    op->inputs.push_back(input);
  }

  op->beam_width =
      HasAttr(node, "beam_width") ? GetIntAttr(node, "beam_width") : 1;
  op->top_paths =
      HasAttr(node, "top_paths") ? GetIntAttr(node, "top_paths") : 1;
  op->merge_repeated = HasAttr(node, "merge_repeated")
                           ? GetBoolAttr(node, "merge_repeated")
                           : true;

  // There are top_paths + 1 outputs.
  op->outputs.push_back(node.name());  // Implicit :0.
  for (int i = 0; i < op->top_paths; ++i) {
    op->outputs.push_back(node.name() + ":" + std::to_string(i + 1));
  }
  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

// This isn't a TensorFlow builtin op. Currently this node can only be generated
// with TfLite OpHint API.
tensorflow::Status ConvertUnidirectionalSequenceLstm(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  DCHECK_EQ(node.op(), "UnidirectionalSequenceLstm");

  const auto& indices = GetListAttr(node, "_tflite_input_indices");

  auto* op = new UnidirectionalSequenceLstmOperator();

  // The input size needs to be the same as the TfLite UniDirectionalSequence
  // Lstm implementation.
  const int kInputsSize = 20;

  op->inputs.resize(kInputsSize);

  if (indices.i_size() != node.input().size()) {
    // New version, the optional inputs are filled with constant nodes.
    int count = 0;
    for (int idx = 0; idx < kInputsSize; ++idx) {
      if (count < indices.i_size() && indices.i(count) == idx) {
        // Specified input.
        op->inputs[idx] = node.input(idx);
        count++;
      } else {
        // Optional input.
        std::string optional_name = node.name() + "_" + std::to_string(idx);
        model->CreateOptionalArray(optional_name);
        op->inputs[idx] = optional_name;
      }
    }
  } else {  // Legacy version.
    std::vector<bool> done(kInputsSize);
    int idx = 0;
    for (const std::string& input : node.input()) {
      int real_index = indices.i(idx);
      op->inputs[real_index] = (input);
      done[real_index] = true;
      idx++;
    }

    for (int idx = 0; idx < done.size(); idx++) {
      if (!done[idx]) {
        std::string optional_name = node.name() + "_" + std::to_string(idx);
        model->CreateOptionalArray(optional_name);
        op->inputs[idx] = optional_name;
      }
    }
  }

  // There're three outputs, only the last one is required.
  op->outputs.push_back(node.name() + ":2");
  model->operators.emplace_back(op);

  return tensorflow::Status::OK();
}

tensorflow::Status ConvertLeakyReluOperator(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  CHECK_EQ(node.op(), "LeakyRelu");
  TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 1));
  CHECK_EQ(GetDataTypeAttr(node, "T"), DT_FLOAT);
  const auto& input_name = node.input(0);
  auto* op = new LeakyReluOperator;
  op->inputs.push_back(input_name);
  op->outputs.push_back(node.name());
  op->alpha = GetFloatAttr(node, "alpha");
  model->operators.emplace_back(op);
  return tensorflow::Status::OK();
}

tensorflow::Status ConvertUnidirectionalSequenceRnn(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model) {
  DCHECK_EQ(node.op(), "UnidirectionalSequenceRnn");

  const auto& indices = GetListAttr(node, "_tflite_input_indices");
  if (indices.i_size() != node.input().size()) {
    return tensorflow::errors::InvalidArgument("Input size does not match.");
  }

  auto* op = new UnidirectionalSequenceRnnOperator();
  for (const std::string& input : node.input()) {
    op->inputs.push_back(input);
  }
  // Only use the last one as input.
  op->outputs.push_back(node.name() + ":1");
  model->operators.emplace_back(op);

  return tensorflow::Status::OK();
}

}  // namespace

namespace internal {

using ConverterType = tensorflow::Status (*)(
    const NodeDef& node, const TensorFlowImportFlags& tf_import_flags,
    const ModelFlags& model_flags, Model* model);
using ConverterMapType = std::unordered_map<std::string, ConverterType>;

ConverterMapType GetTensorFlowNodeConverterMapForFlex() {
  return std::unordered_map<std::string, ConverterType>({
      // We need to let TOCO convert Placeholder information into
      // array data, so that the data types are correct.
      {"LegacyFedInput", ConvertPlaceholderOperator},
      {"Placeholder", ConvertPlaceholderOperator},
      {"Const", ConditionallyConvertConstOperator},
  });
}

ConverterMapType GetTensorFlowNodeConverterMap() {
  return std::unordered_map<std::string, ConverterType>({
      {"Abs", ConvertSimpleOperator<AbsOperator, kAnyNumInputs, 1>},
      {"Add", ConvertSimpleOperator<AddOperator, 2, 1>},
      {"AddV2", ConvertSimpleOperator<AddOperator, 2, 1>},
      {"AddN", ConvertSimpleOperator<AddNOperator, kAnyNumInputs, 1>},
      {"All", ConvertSimpleOperator<TensorFlowAllOperator, kAnyNumInputs, 1>},
      {"Any", ConvertReduceOperator<TensorFlowAnyOperator>},
      {"ArgMax", ConvertArgMaxOperator},
      {"ArgMin", ConvertArgMinOperator},
      {"Assert",
       ConvertSimpleOperator<TensorFlowAssertOperator, kAnyNumInputs, 1>},
      {"AvgPool", ConvertAvgPoolOperator},
      {"BatchMatMul", ConvertBatchMatMulOperator},
      {"BatchMatMulV2", ConvertBatchMatMulOperator},
      {"BatchNormWithGlobalNormalization",
       ConvertBatchNormWithGlobalNormalizationOperator},
      {"BatchToSpaceND", ConvertBatchToSpaceNDOperator},
      {"BiasAdd", ConvertBiasAddOperator},
      {"Cast", ConvertCastOperator},
      {"Ceil", ConvertCeilOperator},
      {"CheckNumerics", ConvertIdentityOperator},
      {"Concat", ConvertConcatOperator},
      {"ConcatV2", ConvertConcatOperator},
      {"Const", ConvertConstOperator},
      {"Conv2D", ConvertConvOperator},
      {"Conv2DBackpropInput", ConvertTransposeConvOperator},
      {"Cos", ConvertSimpleOperator<CosOperator, 1, 1>},
      {"CTCBeamSearchDecoder", ConvertCTCBeamSearchDecoderOperator},
      {"DepthToSpace", ConvertDepthToSpaceOperator},
      {"DepthwiseConv2dNative", ConvertDepthwiseConvOperator},
      {"Div", ConvertSimpleOperator<DivOperator, 2, 1>},
      {"DynamicPartition", ConvertDynamicPartitionOperator},
      {"DynamicStitch", ConvertDynamicStitchOperator},
      {"Elu", ConvertSimpleOperator<EluOperator, 1, 1>},
      {"Equal", ConvertSimpleOperator<TensorFlowEqualOperator, 2, 1>},
      {"Exp", ConvertSimpleOperator<ExpOperator, 1, 1>},
      {"ExpandDims", ConvertSimpleOperator<ExpandDimsOperator, 2, 1>},
      {"FakeQuantWithMinMaxArgs", ConvertFakeQuantWithMinMaxArgs},
      {"FakeQuantWithMinMaxVars", ConvertFakeQuantWithMinMaxVars},
      {"Fill", ConvertSimpleOperator<FillOperator, 2, 1>},
      {"Floor", ConvertFloorOperator},
      {"FloorDiv", ConvertSimpleOperator<FloorDivOperator, 2, 1>},
      {"FloorMod", ConvertSimpleOperator<FloorModOperator, 2, 1>},
      {"FusedBatchNorm", ConvertFusedBatchNormOperator},
      {"FusedBatchNormV3", ConvertFusedBatchNormOperator},
      {"Gather", ConvertGatherOperator},
      {"GatherV2", ConvertGatherOperator},
      {"GatherNd", ConvertGatherNdOperator},
      {"Greater", ConvertSimpleOperator<TensorFlowGreaterOperator, 2, 1>},
      {"GreaterEqual",
       ConvertSimpleOperator<TensorFlowGreaterEqualOperator, 2, 1>},
      {"Identity", ConvertIdentityOperator},
      {"IdentityN", ConvertIdentityNOperator},
      {"LRN", ConvertLRNOperator},
      {"LeakyRelu", ConvertLeakyReluOperator},
      {"LegacyFedInput", ConvertPlaceholderOperator},
      {"Less", ConvertSimpleOperator<TensorFlowLessOperator, 2, 1>},
      {"LessEqual", ConvertSimpleOperator<TensorFlowLessEqualOperator, 2, 1>},
      {"Log", ConvertSimpleOperator<LogOperator, 1, 1>},
      {"LogicalAnd", ConvertSimpleOperator<LogicalAndOperator, 2, 1>},
      {"LogicalOr", ConvertSimpleOperator<LogicalOrOperator, 2, 1>},
      {"LogicalNot", ConvertSimpleOperator<LogicalNotOperator, 1, 1>},
      {"LogSoftmax", ConvertSimpleOperator<LogSoftmaxOperator, 1, 1>},
      {"MatMul", ConvertMatMulOperator},
      {"MatrixDiag", ConvertSimpleOperator<MatrixDiagOperator, 1, 1>},
      {"MatrixDiagV2", ConvertSimpleOperator<MatrixDiagV2Operator, 5, 1>},
      // `MatrixDiagV3` has an `align` attribute. However, Toco only converts
      // `MatrixDiagV3` to `MatrixDiag` with default `k, num_rows, num_cols,
      // padding_value` inputs. In this case, `align` can be ignored.
      {"MatrixDiagV3", ConvertSimpleOperator<MatrixDiagV3Operator, 5, 1>},
      {"MatrixSetDiag", ConvertSimpleOperator<MatrixSetDiagOperator, 2, 1>},
      {"MatrixSetDiagV2", ConvertSimpleOperator<MatrixSetDiagV2Operator, 3, 1>},
      // `MatrixSetDiagV3` has an `align` attribute. However, Toco only converts
      // `MatrixSetDiagV3` to `MatrixSetDiag` with default `k` inputs. In this
      // case, `align` can be ignored.
      {"MatrixSetDiagV3", ConvertSimpleOperator<MatrixSetDiagV3Operator, 3, 1>},
      {"Max", ConvertReduceOperator<TensorFlowMaxOperator>},
      {"MaxPool", ConvertMaxPoolOperator},
      {"Maximum", ConvertSimpleOperator<TensorFlowMaximumOperator, 2, 1>},
      {"Mean", ConvertReduceOperator<MeanOperator>},
      {"Merge",
       ConvertSimpleOperator<TensorFlowMergeOperator, kAnyNumInputs, 1>},
      {"Min", ConvertReduceOperator<TensorFlowMinOperator>},
      {"Minimum", ConvertSimpleOperator<TensorFlowMinimumOperator, 2, 1>},
      {"Mul", ConvertSimpleOperator<MulOperator, 2, 1>},
      {"Neg", ConvertSimpleOperator<NegOperator, 1, 1>},
      {"NextIteration", ConvertOperatorSpecialCasedAsRNNBackEdge},
      {"NoOp", ConvertNoOpOperator},
      {"NotEqual", ConvertSimpleOperator<TensorFlowNotEqualOperator, 2, 1>},
      {"OneHot", ConvertOneHotOperator},
      {"Pack", ConvertPackOperator},
      {"Pad", ConvertSimpleOperator<PadOperator, 2, 1>},
      {"PadV2", ConvertSimpleOperator<PadV2Operator, 3, 1>},
      {"ParallelDynamicStitch", ConvertDynamicStitchOperator},
      {"Placeholder", ConvertPlaceholderOperator},
      {"PlaceholderWithDefault", ConvertIdentityOperator},
      {"Pow", ConvertSimpleOperator<PowOperator, 2, 1>},
      {"Prod", ConvertReduceOperator<TensorFlowProdOperator>},
      {"RandomUniform", ConvertRandomUniform},
      {"Range", ConvertRangeOperator},
      {"Rank", ConvertSimpleOperator<TensorFlowRankOperator, 1, 1>},
      {"RealDiv", ConvertSimpleOperator<DivOperator, 2, 1>},
      {"Relu", ConvertSimpleOperator<ReluOperator, 1, 1>},
      {"Relu6", ConvertSimpleOperator<Relu6Operator, 1, 1>},
      {"Reshape", ConvertSimpleOperator<TensorFlowReshapeOperator, 2, 1>},
      {"ResizeBilinear", ConvertResizeBilinearOperator},
      {"ResizeNearestNeighbor", ConvertResizeNearestNeighborOperator},
      {"ReverseSequence", ConvertReverseSequenceOperator},
      {"ReverseV2", ConvertSimpleOperator<ReverseV2Operator, 2, 1>},
      {"Round", ConvertRoundOperator},
      {"Rsqrt", ConvertSimpleOperator<TensorFlowRsqrtOperator, 1, 1>},
      {"ScatterNd", ConvertSimpleOperator<ScatterNdOperator, 3, 1>},
      {"SegmentSum", ConvertSimpleOperator<SegmentSumOperator, 2, 1>},
      {"Select", ConvertSimpleOperator<SelectOperator, 3, 1>},
      {"SelectV2", ConvertSimpleOperator<SelectOperator, 3, 1>},
      {"Shape", ConvertShapeOperator},
      {"Sigmoid", ConvertSimpleOperator<LogisticOperator, 1, 1>},
      {"Sin", ConvertSimpleOperator<SinOperator, 1, 1>},
      {"Slice", ConvertSimpleOperator<SliceOperator, 3, 1>},
      {"Softmax", ConvertSoftmaxOperator},
      {"SpaceToBatchND", ConvertSpaceToBatchNDOperator},
      {"SpaceToDepth", ConvertSpaceToDepthOperator},
      {"SparseToDense", ConvertSparseToDenseOperator},
      {"Split", ConvertSplitOperator},
      {"SplitV", ConvertSplitVOperator},
      {"Sqrt", ConvertSimpleOperator<TensorFlowSqrtOperator, 1, 1>},
      {"Square", ConvertSimpleOperator<TensorFlowSquareOperator, 1, 1>},
      {"SquaredDifference",
       ConvertSimpleOperator<SquaredDifferenceOperator, 2, 1>},
      {"Snapshot", ConvertIdentityOperator},
      {"Squeeze", ConvertSqueezeOperator},
      {"StopGradient", ConvertIdentityOperator},
      {"StridedSlice", ConvertStridedSliceOperator},
      {"Sub", ConvertSimpleOperator<SubOperator, 2, 1>},
      {"Sum", ConvertReduceOperator<TensorFlowSumOperator>},
      {"Svdf", ConvertSvdfOperator},
      {"Switch", ConvertSwitchOperator},
      {"Tanh", ConvertSimpleOperator<TanhOperator, 1, 1>},
      {"Tile", ConvertSimpleOperator<TensorFlowTileOperator, 2, 1>},
      {"TopK", ConvertTopKV2Operator},
      {"TopKV2", ConvertTopKV2Operator},
      {"Transpose", ConvertSimpleOperator<TransposeOperator, 2, 1>},
      {"Unpack", ConvertUnpackOperator},
      {"ZerosLike", ConvertSimpleOperator<TensorFlowZerosLikeOperator, 1, 1>},
      {"UnidirectionalSequenceLstm", ConvertUnidirectionalSequenceLstm},
      {"UnidirectionalSequenceRnn", ConvertUnidirectionalSequenceRnn},
      {"MirrorPad", ConvertMirrorPadOperator},
      {"Unique", ConvertSimpleOperator<UniqueOperator, 1, 2>},
      {"Where", ConvertSimpleOperator<WhereOperator, 1, 1>},
  });
}

tensorflow::Status ImportTensorFlowNode(
    const tensorflow::NodeDef& node,
    const TensorFlowImportFlags& tf_import_flags, const ModelFlags& model_flags,
    Model* model, const ConverterMapType& converter_map) {
  auto converter = converter_map.find(node.op());
  if (converter == converter_map.end()) {
    return ConvertUnsupportedOperator(node, tf_import_flags, model_flags,
                                      model);
  } else {
    return converter->second(node, tf_import_flags, model_flags, model);
  }
}
}  // namespace internal

std::unique_ptr<Model> ImportTensorFlowGraphDef(
    const ModelFlags& model_flags, const TensorFlowImportFlags& tf_import_flags,
    const GraphDef& tf_graph) {
  LogDumpGraphDef(kLogLevelModelChanged, "AT IMPORT", tf_graph);

  GraphDef inlined_graph(tf_graph);
  if (InlineAllFunctions(&inlined_graph)) {
    LogDumpGraphDef(kLogLevelModelChanged, "AFTER INLINING", inlined_graph);
  }

  // Check input and output specification.
  for (const auto& specified_input_array : model_flags.input_arrays()) {
    CHECK(!absl::EndsWith(specified_input_array.name(), ":0"))
        << "Unsupported explicit zero output index: "
        << specified_input_array.name();
  }
  for (const std::string& specified_output_array :
       model_flags.output_arrays()) {
    CHECK(!absl::EndsWith(specified_output_array, ":0"))
        << "Unsupported explicit zero output index: " << specified_output_array;
  }

  Model* model = new Model;
  internal::ConverterMapType converter_map;

  // This is used for the TFLite "Full Flex Mode" conversion. All the ops are
  // imported as `TensorFlowUnsupportedOperator`, and later all these ops are
  // converted to TFLite Flex ops.
  if (!tf_import_flags.import_all_ops_as_unsupported) {
    converter_map = internal::GetTensorFlowNodeConverterMap();
  } else {
    converter_map = internal::GetTensorFlowNodeConverterMapForFlex();
  }

  for (auto node : inlined_graph.node()) {
    StripZeroOutputIndexFromInputs(&node);
    auto status = internal::ImportTensorFlowNode(
        node, tf_import_flags, model_flags, model, converter_map);
    CHECK(status.ok()) << status.error_message();
  }

  ResolveModelFlags(model_flags, model);

  StripCaretFromArrayNames(model);
  AddExtraOutputs(model);
  FixNoMissingArray(model);
  FixNoOrphanedArray(model);
  FixOperatorOrdering(model);
  CheckInvariants(*model);

  // if rnn state arrays are constant, make them transient
  for (const auto& rnn_state : model->flags.rnn_states()) {
    model->GetArray(rnn_state.state_array()).buffer = nullptr;
  }

  return std::unique_ptr<Model>(model);
}

std::unique_ptr<Model> ImportTensorFlowGraphDef(
    const ModelFlags& model_flags, const TensorFlowImportFlags& tf_import_flags,
    const std::string& input_file_contents) {
  std::unique_ptr<GraphDef> tf_graph(new GraphDef);
  CHECK(ParseFromStringEitherTextOrBinary(input_file_contents, tf_graph.get()));

  std::unique_ptr<GraphDef> pruned_graph =
      MaybeReplaceCompositeSubgraph(*tf_graph);
  if (pruned_graph) {
    tf_graph = std::move(pruned_graph);
  }
  return ImportTensorFlowGraphDef(model_flags, tf_import_flags, *tf_graph);
}

std::vector<std::string> GetPotentiallySupportedOps() {
  std::vector<std::string> supported_ops;
  const internal::ConverterMapType& converter_map =
      internal::GetTensorFlowNodeConverterMap();

  for (const auto& item : converter_map) {
    supported_ops.push_back(item.first);
  }
  return supported_ops;
}

}  // namespace toco