xref: /aosp_15_r20/external/tensorflow/tensorflow/lite/toco/tflite/export.cc (revision b6fb3261f9314811a0f4371741dbb8839866f948)

/* Copyright 2019 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/tflite/export.h"

#include <string>

#include "flatbuffers/flexbuffers.h"
#include "absl/strings/str_join.h"
#include "tensorflow/core/lib/core/errors.h"
#include "tensorflow/lite/context.h"
#include "tensorflow/lite/schema/schema_conversion_utils.h"
#include "tensorflow/lite/schema/schema_generated.h"
#include "tensorflow/lite/toco/tflite/operator.h"
#include "tensorflow/lite/toco/tflite/types.h"
#include "tensorflow/lite/toco/tooling_util.h"
#include "tensorflow/lite/tools/optimize/quantize_weights.h"
#include "tensorflow/lite/tools/versioning/runtime_version.h"
#include "tensorflow/lite/version.h"

namespace toco {

namespace tflite {

using flatbuffers::FlatBufferBuilder;
using flatbuffers::Offset;
using flatbuffers::Vector;
using ::tflite::Buffer;
using ::tflite::BuiltinOperator;
using ::tflite::BuiltinOperator_CUSTOM;
using ::tflite::BuiltinOperator_MAX;
using ::tflite::BuiltinOperator_MIN;
using ::tflite::CreateBuffer;
using ::tflite::CreateMetadata;
using ::tflite::CreateModel;
using ::tflite::CreateOperator;
using ::tflite::CreateTensor;
using ::tflite::Metadata;
using ::tflite::Operator;
using ::tflite::OperatorCode;
using ::tflite::SubGraph;
using ::tflite::Tensor;

namespace {

// Check if a TensorFlow Op is a control flow op by its name.
bool IsControlFlowOp(const std::string& tensorflow_op) {
  // Technically this is equivalent to `::tensorflow::Node::IsControlFlow()`.
  // It requires to construct a `::tensorflow::Graph` to use that helper
  // function, so we simply hardcode the list of control flow ops here.
  if (tensorflow_op == "Switch" || tensorflow_op == "RefSwitch" ||
      tensorflow_op == "Merge" || tensorflow_op == "RefMerge" ||
      tensorflow_op == "Enter" || tensorflow_op == "RefEnter" ||
      tensorflow_op == "Exit" || tensorflow_op == "RefExit" ||
      tensorflow_op == "NextIteration" || tensorflow_op == "RefNextIteration") {
    return true;
  }
  // TODO(ycling): Also check how to handle Variable ops and Assign ops.
  return false;
}

// Check if a TensorFlow Op is unsupported by the Flex runtime.
bool IsUnsupportedFlexOp(const std::string& tensorflow_op) {
  if (IsControlFlowOp(tensorflow_op)) {
    return true;
  }
  // `HashTableV2` isn't supported for now since it requires an additional
  // initialization step.
  // TODO(b/117651199): Support `HashTableV2` with Flex runtime.
  if (tensorflow_op == "HashTableV2") {
    return true;
  }
  return false;
}

// Map from operator name to TF Lite enum value, for all builtins.
const std::map<std::string, BuiltinOperator>& GetBuiltinOpsMap() {
  static std::map<std::string, BuiltinOperator>* builtin_ops = nullptr;
  if (builtin_ops == nullptr) {
    builtin_ops = new std::map<std::string, BuiltinOperator>();

    for (int i = BuiltinOperator_MIN; i <= BuiltinOperator_MAX; ++i) {
      BuiltinOperator op = static_cast<BuiltinOperator>(i);
      std::string name = EnumNameBuiltinOperator(op);
      if (op != BuiltinOperator_CUSTOM && !name.empty()) {
        (*builtin_ops)[name] = op;
      }
    }
  }
  return *builtin_ops;
}

void WriteModelToString(const flatbuffers::FlatBufferBuilder& builder,
                        std::string* file_contents) {
  const uint8_t* buffer = builder.GetBufferPointer();
  int size = builder.GetSize();
  *file_contents = std::string(reinterpret_cast<const char*>(buffer), size);
}

}  // Anonymous namespace.

namespace details {

OperatorKey::OperatorKey(
    const ::toco::OperatorSignature& op_signature,
    const std::map<OperatorType, std::unique_ptr<BaseOperator>>& ops_by_type,
    bool enable_select_tf_ops) {
  // Get the op name (by Toco definition).
  const ::toco::Operator& op = *op_signature.op;
  std::string name = HelpfulOperatorTypeName(op);

  bool is_builtin = false;
  const auto& builtin_ops = GetBuiltinOpsMap();
  if (ops_by_type.count(op.type) != 0) {
    version_ = ops_by_type.at(op.type)->GetVersion(op_signature);
    name = ops_by_type.at(op.type)->name();
    is_builtin = (builtin_ops.count(name) > 0);
  }

  if (is_builtin) {
    // For TFLite supported builtin ops, find out its BuiltinOperator enum used
    // in FlatBuffer.
    type_ = builtin_ops.at(name);
    return;
  }
  // The logic below is all for custom ops or Flex ops.
  is_custom_op_ = true;
  type_ = BuiltinOperator_CUSTOM;

  if (op.type == OperatorType::kUnsupported) {
    const TensorFlowUnsupportedOperator& unsupported_op =
        static_cast<const TensorFlowUnsupportedOperator&>(op);
    const auto tensorflow_op = unsupported_op.tensorflow_op;

    if (ShouldExportAsFlexOp(enable_select_tf_ops,
                             unsupported_op.tensorflow_op)) {
      is_custom_op_ = false;
      is_flex_op_ = true;
      flex_tensorflow_op_ = tensorflow_op;
      custom_code_ =
          std::string(::tflite::kFlexCustomCodePrefix) + flex_tensorflow_op_;
    } else {
      custom_code_ = tensorflow_op;
    }
  } else if (enable_select_tf_ops && !op.tensorflow_node_def.empty()) {
    // For Toco-supported/TFLite-unsupported ops, if the TensorFlow NodeDef
    // is retained in the Toco Operator, we produce a Flex op if Flex mode
    // is enabled.
    is_custom_op_ = false;
    is_flex_op_ = true;
    flex_tensorflow_op_ = name;
    custom_code_ =
        std::string(::tflite::kFlexCustomCodePrefix) + flex_tensorflow_op_;
  } else {
    // If Flex is disabled or the original TensorFlow NodeDef isn't available,
    // we produce a custom op. This gives developers a chance to implement
    // custom ops.
    custom_code_ = name;
  }

  if (is_flex_op_) {
    if (IsUnsupportedFlexOp(flex_tensorflow_op_)) {
      is_unsupported_flex_op_ = true;
    }
  }
}

void LoadTensorsMap(const Model& model, TensorsMap* tensors_map) {
  // First find a list of unique array names.
  std::set<std::string> names;
  for (const auto& array_pair : model.GetArrayMap()) {
    names.insert(array_pair.first);
  }

  // Now assign indices to them and fill in the map.
  int index = 0;
  for (const auto& name : names) {
    (*tensors_map)[name] = index;
    ++index;
  }
}

void LoadOperatorsMap(
    const Model& model, OperatorsMap* operators_map,
    const std::map<OperatorType, std::unique_ptr<BaseOperator>>& ops_by_type,
    bool enable_select_tf_ops) {
  // First find a list of unique operator types.
  std::set<OperatorKey> keys;
  for (const auto& op : model.operators) {
    const toco::OperatorSignature op_signature = {op.get(), &model};
    keys.insert(OperatorKey(op_signature, ops_by_type, enable_select_tf_ops));
  }
  // Now assign indices to them and fill in the map.
  int index = 0;
  for (const auto& key : keys) {
    (*operators_map)[key] = index;
    ++index;
  }
}

}  // namespace details

Offset<Vector<Offset<Tensor>>> ExportTensors(
    const Model& model, const details::TensorsMap& tensors_map,
    FlatBufferBuilder* builder,
    std::vector<Offset<Vector<uint8_t>>>* buffers_to_write,
    const std::set<int32_t>& variable_tensor_indices) {
  // In the end we will need to produce a vector sorted by the indices of the
  // tensors in the tensors_map.
  std::map<int, Offset<Tensor>> ordered_tensors;

  for (const auto& array_pair : model.GetArrayMap()) {
    const std::string& tensor_name = array_pair.first;
    const toco::Array& array = *array_pair.second;

    int buffer_index = buffers_to_write->size();
    auto type = DataType::Serialize(array.data_type);
    buffers_to_write->push_back(DataBuffer::Serialize(array, builder));

    std::vector<int> shape;
    if (array.has_shape()) {
      shape.reserve(array.shape().dims().size());
      for (const auto& d : array.shape().dims()) {
        shape.push_back(d);
      }
    }

    Offset<Vector<float>> min;
    Offset<Vector<float>> max;
    Offset<Vector<float>> scale;
    Offset<Vector<int64_t>> zero_point;
    if (array.minmax) {
      min = builder->CreateVector(
          std::vector<float>{static_cast<float>(array.minmax->min)});
      max = builder->CreateVector(
          std::vector<float>{static_cast<float>(array.minmax->max)});
    }
    if (array.quantization_params) {
      scale = builder->CreateVector(std::vector<float>{
          static_cast<float>(array.quantization_params->scale)});
      zero_point = builder->CreateVector(
          std::vector<int64_t>{array.quantization_params->zero_point});
    }
    auto q_param = ::tflite::CreateQuantizationParameters(*builder, min, max,
                                                          scale, zero_point);

    int index = tensors_map.at(tensor_name);
    bool is_variable =
        variable_tensor_indices.find(index) != variable_tensor_indices.end();
    ordered_tensors[index] =
        CreateTensor(*builder, builder->CreateVector(shape), type, buffer_index,
                     builder->CreateString(tensor_name), q_param, is_variable);
  }

  std::vector<Offset<Tensor>> tensor_vector;
  tensor_vector.reserve(ordered_tensors.size());
  for (const auto& tensor : ordered_tensors) {
    tensor_vector.push_back(tensor.second);
  }

  return builder->CreateVector(tensor_vector);
}

Offset<Vector<int32_t>> ExportInputTensors(
    const Model& model, const details::TensorsMap& tensors_map,
    FlatBufferBuilder* builder) {
  std::vector<int32_t> inputs;
  for (const auto& input : model.flags.input_arrays()) {
    inputs.push_back(tensors_map.at(input.name()));
  }
  return builder->CreateVector<int32_t>(inputs);
}

Offset<Vector<int32_t>> ExportOutputTensors(
    const Model& model, const details::TensorsMap& tensors_map,
    FlatBufferBuilder* builder) {
  std::vector<int32_t> outputs;
  for (const std::string& output : model.flags.output_arrays()) {
    outputs.push_back(tensors_map.at(output));
  }
  return builder->CreateVector<int32_t>(outputs);
}

Offset<Vector<Offset<OperatorCode>>> ExportOperatorCodes(
    const Model& model,
    const std::map<OperatorType, std::unique_ptr<BaseOperator>>& ops_by_type,
    const details::OperatorsMap& operators_map, FlatBufferBuilder* builder,
    const ExportParams& params) {
  // Map from operator name to TF Lite enum value, for all builtins.
  std::map<std::string, BuiltinOperator> builtin_ops;
  for (int i = BuiltinOperator_MIN; i <= BuiltinOperator_MAX; ++i) {
    BuiltinOperator op = static_cast<BuiltinOperator>(i);
    std::string name = EnumNameBuiltinOperator(op);
    if (op != BuiltinOperator_CUSTOM && !name.empty()) {
      builtin_ops[name] = op;
    }
  }

  // We will need to produce a vector of codes in the same order as they
  // appear in the operators_map.
  std::map<int, Offset<OperatorCode>> ordered_opcodes;

  for (const auto& op : model.operators) {
    const toco::OperatorSignature op_signature = {op.get(), &model};
    const details::OperatorKey operator_key = details::OperatorKey(
        op_signature, ops_by_type, params.enable_select_tf_ops);
    int op_index = operators_map.at(operator_key);

    flatbuffers::Offset<flatbuffers::String> custom_code = 0;
    if (!operator_key.custom_code().empty()) {
      custom_code = builder->CreateString(operator_key.custom_code());
    }

    ordered_opcodes[op_index] = CreateOperatorCode(
        *builder, operator_key.type(), custom_code, operator_key.version());
  }

  std::vector<Offset<OperatorCode>> opcode_vector;
  opcode_vector.reserve(ordered_opcodes.size());
  for (const auto& opcode : ordered_opcodes) {
    opcode_vector.push_back(opcode.second);
  }

  return builder->CreateVector(opcode_vector);
}

Offset<Vector<Offset<Operator>>> ExportOperators(
    const Model& model,
    const std::map<OperatorType, std::unique_ptr<BaseOperator>>& ops_by_type,
    const details::OperatorsMap& operators_map,
    const details::TensorsMap& tensors_map, FlatBufferBuilder* builder,
    std::set<int32_t>* variable_tensor_indices, const ExportParams& params) {
  variable_tensor_indices->clear();

  auto is_tflite_builtin = [](const BaseOperator* op) {
    const auto& tflite_builtins = GetBuiltinOpsMap();
    return (op && tflite_builtins.find(op->name()) != tflite_builtins.end());
  };

  // The operators are in execution order, so we just follow tf.mini order.
  std::vector<Offset<Operator>> op_vector;
  for (const auto& op : model.operators) {
    std::vector<int32_t> inputs;
    for (const std::string& input : op->inputs) {
      // -1 is the ID for optional tensor in TFLite output
      int id = model.IsOptionalArray(input) ? -1 : tensors_map.at(input);
      inputs.push_back(id);
    }
    std::vector<int32_t> outputs;
    for (const std::string& output : op->outputs) {
      outputs.push_back(tensors_map.at(output));
    }
    const toco::OperatorSignature op_signature = {op.get(), &model};
    const auto key = details::OperatorKey(op_signature, ops_by_type,
                                          params.enable_select_tf_ops);
    int op_index = operators_map.at(key);

    auto tflite_op_it = ops_by_type.find(op->type);
    BaseOperator* tflite_op = tflite_op_it == ops_by_type.end()
                                  ? nullptr
                                  : tflite_op_it->second.get();

    // This is a custom op unless we can find it in ops_by_type, and even then
    // it could be a custom op (such as kUnsupported).
    auto options = Options::Custom(0);

    std::vector<bool> mutating_input_variables;

    // It is conceivable that an op is exportable via Serialize() but does not
    // have a corresponding TFLITE builtin. In that case, when flex mode is
    // enabled we should export it as a flex op, not as a native.
    bool export_as_flex_op = !is_tflite_builtin(tflite_op) &&
                             key.is_flex_op() &&
                             !op->tensorflow_node_def.empty();
    if (export_as_flex_op) {
      auto fbb = WriteFlexOpOptions(op->tensorflow_node_def);
      if (fbb) {
        options = Options::Custom(builder->CreateVector(fbb->GetBuffer()));
      }
    } else if (tflite_op) {
      options = tflite_op->Serialize(*op, builder);
      mutating_input_variables = tflite_op->GetMutatingInputVariables(*op);

      if (!mutating_input_variables.empty()) {
        for (uint32_t i = 0; i < op->inputs.size(); ++i) {
          if (!mutating_input_variables[i]) {
            continue;
          }
          int32_t variable_tensor_index = tensors_map.at(op->inputs[i]);
          variable_tensor_indices->insert(variable_tensor_index);
        }
      }
    } else {
      // We don't know much about this op. It doesn't have a serializer and
      // it is not supposed to be exported as a flex op. We will treat it as
      // a regular custom op: we will still create an operator for it, but it
      // will not have any 'options'.
    }

    // The only supported CustomOptionFormat is FLEXBUFFERS now.
    op_vector.push_back(CreateOperator(
        *builder, op_index, builder->CreateVector(inputs),
        builder->CreateVector(outputs), options.type, options.builtin,
        options.custom, ::tflite::CustomOptionsFormat_FLEXBUFFERS,
        builder->CreateVector(mutating_input_variables)));
  }

  return builder->CreateVector(op_vector);
}

Offset<Vector<Offset<Buffer>>> ExportBuffers(
    const Model& model,
    const std::vector<Offset<Vector<uint8_t>>>& buffers_to_write,
    FlatBufferBuilder* builder) {
  std::vector<Offset<Buffer>> buffer_vector;
  buffer_vector.reserve(buffers_to_write.size());
  for (const auto buffer : buffers_to_write) {
    buffer_vector.push_back(CreateBuffer(*builder, buffer));
  }
  return builder->CreateVector(buffer_vector);
}

tensorflow::Status Export(const Model& model, std::string* output_file_contents,
                          const ExportParams& params) {
  const auto ops_by_type = BuildOperatorByTypeMap(params.enable_select_tf_ops);
  return Export(model, output_file_contents, params, ops_by_type);
}

void ParseControlFlowErrors(std::set<std::string>* custom_ops,
                            std::vector<std::string>* error_msgs) {
  std::set<std::string> unsupported_control_flow_ops;
  // Check if unsupported ops contains control flow ops. It's impossible
  // to implement these ops as custom ops at the moment.
  for (const auto& op : *custom_ops) {
    if (IsControlFlowOp(op)) {
      unsupported_control_flow_ops.insert(op);
    }
  }
  if (!unsupported_control_flow_ops.empty()) {
    error_msgs->push_back(absl::StrCat(
        "TensorFlow Lite currently doesn't support control flow ops: ",
        absl::StrJoin(unsupported_control_flow_ops, ", "), ".",
        " We are working on supporting control flow ops, please see github "
        "issue at "
        "https://github.com/tensorflow/tensorflow/issues/28485."));
  }
  // Remove control flow ops from `custom_ops` set so that they won't be
  // reported again in later messages.
  for (const auto& op : unsupported_control_flow_ops) {
    custom_ops->erase(op);
  }
}

// Exports appropriate model metadata.
// This consists of a 16-byte placeholder string buffer that will be populated
// with the model's minimum required runtime version after the model is
// finalized. We choose 16 bytes as it's the alignment of flatbuffer buffers
// and prevents wasted space if the final string is shorter than 16 bytes.
// Note: This logic matches that used in the MLIR flatbuffer export path.
flatbuffers::Offset<tflite::Metadata> ExportMetadata(
    const Model& model, std::vector<Offset<Vector<uint8_t>>>* buffers_to_write,
    FlatBufferBuilder* builder) {
  auto metadata =
      CreateMetadata(*builder, builder->CreateString("min_runtime_version"),
                     buffers_to_write->size());
  const std::string min_runtime_placeholder = std::string(16, '\0');
  buffers_to_write->push_back(builder->CreateVector(
      reinterpret_cast<const uint8_t*>(min_runtime_placeholder.data()),
      min_runtime_placeholder.size()));
  return metadata;
}

tensorflow::Status Export(
    const Model& model, std::string* output_file_contents,
    const ExportParams& params,
    const std::map<OperatorType, std::unique_ptr<BaseOperator>>& ops_by_type) {
  for (const std::string& input_array : model.GetInvalidInputArrays()) {
    if (model.HasArray(input_array)) {
      return tensorflow::errors::InvalidArgument(
          absl::StrCat("Placeholder ", input_array,
                       " should be specified by "
                       "input_arrays."));
    }
  }

  flatbuffers::FlatBufferBuilder builder(/*initial_size=*/10240);

  details::TensorsMap tensors_map;
  details::LoadTensorsMap(model, &tensors_map);

  details::OperatorsMap operators_map;
  details::LoadOperatorsMap(model, &operators_map, ops_by_type,
                            params.enable_select_tf_ops);

  std::vector<Offset<Vector<uint8_t>>> buffers_to_write;
  // Insert an empty buffer to the beginning of the list.
  buffers_to_write.push_back(0);

  auto op_codes =
      ExportOperatorCodes(model, ops_by_type, operators_map, &builder, params);

  for (const auto& op : model.operators) {
    if (op->type == OperatorType::kFakeQuant) {
      LOG(WARNING) << "FAKE_QUANT operation " << LogName(*op)
                   << " was not converted. If running quantized make sure you "
                      "are passing --inference_type=QUANTIZED_UINT8 and values "
                      "for --std_values and --mean_values.";
    }
  }

  // The set of used builtin ops.
  std::set<std::string> builtin_ops;
  // The set of custom ops (not including Flex ops).
  std::set<std::string> custom_ops;
  // The set of Flex ops which are not supported.
  std::set<std::string> unsupported_flex_ops;

  for (const auto& it : operators_map) {
    const details::OperatorKey& key = it.first;
    if (key.is_custom_op()) {
      custom_ops.insert(key.custom_code());
    }
    if (key.is_unsupported_flex_op()) {
      unsupported_flex_ops.insert(key.flex_tensorflow_op());
    }
    if (!key.is_custom_op() && !key.is_flex_op() &&
        !key.is_unsupported_flex_op()) {
      builtin_ops.insert(EnumNameBuiltinOperator(key.type()));
    }
  }

  if (!custom_ops.empty()) {
    if (!params.allow_custom_ops) {
      auto please_report_bug_message = []() {
        return "We are continually in the process of adding support to "
               "TensorFlow Lite for more ops. It would be helpful if you could "
               "inform us of how this conversion went by opening a github "
               "issue at "
               "https://github.com/tensorflow/tensorflow/issues/new?template="
               "40-tflite-op-request.md\n and pasting the following:\n\n";
      };

      std::vector<std::string> error_msgs;
      ParseControlFlowErrors(&custom_ops, &error_msgs);

      // Remove ExpandDims and ReorderAxes from unimplemented list unless they
      // compose the list. Both ops are removed during graph transformations.
      // However, if an op is unimplemented earlier in the model, the graph
      // transformation is unable to run because the output shape is not
      // defined. This causes unnecessary confusion during model conversion
      // time.
      std::set<std::string> custom_ops_final;
      for (const auto& op_type : custom_ops) {
        if (op_type != "ReorderAxes" && op_type != "ExpandDims") {
          custom_ops_final.insert(op_type);
        }
      }
      if (custom_ops_final.empty()) {
        custom_ops_final = custom_ops;
      }

      if (!custom_ops_final.empty()) {
        if (params.enable_select_tf_ops) {
          error_msgs.push_back(absl::StrCat(
              "Some of the operators in the model are not supported by "
              "the standard TensorFlow Lite runtime and are not recognized "
              "by "
              "TensorFlow. If you have a custom "
              "implementation for them you can disable this error with "
              "--allow_custom_ops, or by setting allow_custom_ops=True "
              "when calling tf.lite.TFLiteConverter(). Here is a list "
              "of builtin operators you are using: ",
              absl::StrJoin(builtin_ops, ", "),
              ". Here is a list "
              "of operators for which you will need custom implementations: ",
              absl::StrJoin(custom_ops_final, ", "), "."));
        } else {
          error_msgs.push_back(absl::StrCat(
              "Some of the operators in the model are not supported by "
              "the standard TensorFlow Lite runtime. If those are native "
              "TensorFlow operators, you might be able to use the extended "
              "runtime by passing --enable_select_tf_ops, or by setting "
              "target_ops=TFLITE_BUILTINS,SELECT_TF_OPS when calling "
              "tf.lite.TFLiteConverter(). Otherwise, if you have a "
              "custom implementation for them you can disable this error "
              "with "
              "--allow_custom_ops, or by setting allow_custom_ops=True "
              "when calling tf.lite.TFLiteConverter(). Here is a list "
              "of builtin operators you are using: ",
              absl::StrJoin(builtin_ops, ", "),
              ". Here is a list "
              "of operators for which you will need custom implementations: ",
              absl::StrJoin(custom_ops_final, ", "), "."));
        }
      }
      if (!error_msgs.empty()) {
        return tensorflow::errors::InvalidArgument(absl::StrCat(
            please_report_bug_message(), absl::StrJoin(error_msgs, " ")));
      }
    }
  }

  if (!unsupported_flex_ops.empty()) {
    return tensorflow::errors::InvalidArgument(
        absl::StrCat("Some of the operators in the model are not supported by "
                     "TensorFlow Flex runtime: ",
                     absl::StrJoin(unsupported_flex_ops, ", "), "."));
  }

  std::set<int32_t> variable_tensor_indices;
  auto ops = ExportOperators(model, ops_by_type, operators_map, tensors_map,
                             &builder, &variable_tensor_indices, params);

  auto tensors = ExportTensors(model, tensors_map, &builder, &buffers_to_write,
                               variable_tensor_indices);
  auto inputs = ExportInputTensors(model, tensors_map, &builder);
  auto outputs = ExportOutputTensors(model, tensors_map, &builder);

  // TODO(aselle): add support to toco for multiple subgraphs.
  auto subgraph = CreateSubGraph(builder, tensors, inputs, outputs, ops,
                                 /* name */ 0);
  std::vector<flatbuffers::Offset<SubGraph>> subgraphs = {subgraph};

  // TODO(wangtz): offline memory planning for activation Tensors.
  if (!params.allow_dynamic_tensors) {
    return tensorflow::errors::Unimplemented(
        "Unsupported flag: allow_dynamic_tensors. Offline memory planning is "
        "not implemented yet.");
  }

  auto metadata = ExportMetadata(model, &buffers_to_write, &builder);
  std::vector<flatbuffers::Offset<Metadata>> metadatas = {metadata};

  auto buffers = ExportBuffers(model, buffers_to_write, &builder);
  auto description = builder.CreateString("TOCO Converted.");

  auto new_model_location =
      CreateModel(builder, TFLITE_SCHEMA_VERSION, op_codes,
                  builder.CreateVector(subgraphs), description, buffers,
                  /* metadata_buffer */ 0, builder.CreateVector(metadatas));
  ::tflite::FinishModelBuffer(builder, new_model_location);
  ::tflite::UpdateMinimumRuntimeVersionForModel(builder.GetBufferPointer());

  if (params.quantize_weights == QuantizedBufferType::NONE) {
    WriteModelToString(builder, output_file_contents);
  } else {
    // Call the quantize_weights tool.
    LOG(INFO) << "Quantizing TFLite model after conversion to flatbuffer. "
                 "dump_graphviz will only output the model before this "
                 "transformation. To visualize the output graph use "
                 "lite/tools/optimize.py.";
    flatbuffers::FlatBufferBuilder q_builder(/*initial_size=*/10240);
    const uint8_t* buffer = builder.GetBufferPointer();
    const ::tflite::Model* input_model = ::tflite::GetModel(buffer);
    ::tflite::optimize::BufferType quantized_type;
    if (params.quantize_weights == QuantizedBufferType::INT8) {
      quantized_type = ::tflite::optimize::BufferType::QUANTIZED_INT8;
    } else if (params.quantize_weights == QuantizedBufferType::FLOAT16) {
      quantized_type = ::tflite::optimize::BufferType::QUANTIZED_FLOAT16;
    } else {
      return tensorflow::errors::InvalidArgument(
          "Quantized type not recognized");
    }
    if (::tflite::optimize::QuantizeWeights(
            &q_builder, input_model, quantized_type,
            !params.disable_per_channel,
            ::tflite::optimize::QuantizerType::OLD_QUANTIZER) != kTfLiteOk) {
      return tensorflow::errors::InvalidArgument(
          "Quantize weights transformation failed.");
    }
    WriteModelToString(q_builder, output_file_contents);
  }

  return tensorflow::Status();
}

}  // namespace tflite

}  // namespace toco