path: root/tensorflow/contrib/lite/toco/graph_transformations/quantize.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 <algorithm>
#include <cmath>
#include <limits>
#include <memory>
#include <string>
#include <unordered_map>
#include <vector>

#include "tensorflow/contrib/lite/toco/graph_transformations/graph_transformations.h"
#include "tensorflow/contrib/lite/toco/graph_transformations/quantization_util.h"
#include "tensorflow/contrib/lite/toco/model.h"
#include "tensorflow/contrib/lite/toco/model_flags.pb.h"
#include "tensorflow/contrib/lite/toco/tooling_util.h"
#include "tensorflow/core/platform/logging.h"

namespace toco {

namespace {

bool SupportsQuantization(const Operator& op) {
  auto type = op.type;
  if (type == OperatorType::kUnsupported) {
    auto* unsupported = static_cast<const TensorFlowUnsupportedOperator*>(&op);
    return unsupported->quantized;
  }
  return type == OperatorType::kConv || type == OperatorType::kDepthwiseConv ||
         type == OperatorType::kFullyConnected ||
         type == OperatorType::kConcatenation ||
         type == OperatorType::kL2Normalization || type == OperatorType::kAdd ||
         type == OperatorType::kAveragePool || type == OperatorType::kMaxPool ||
         type == OperatorType::kMinimum || type == OperatorType::kMaximum ||
         type == OperatorType::kLogistic || type == OperatorType::kSoftmax ||
         type == OperatorType::kLogSoftmax || type == OperatorType::kSlice ||
         type == OperatorType::kResizeBilinear ||
         type == OperatorType::kSplit || type == OperatorType::kSub ||
         type == OperatorType::kSqueeze || type == OperatorType::kPad ||
         type == OperatorType::kPadV2 || type == OperatorType::kReshape ||
         type == OperatorType::kTanh || type == OperatorType::kMul ||
         type == OperatorType::kBatchToSpaceND || type == OperatorType::kSum ||
         type == OperatorType::kSpaceToBatchND ||
         type == OperatorType::kSpaceToDepth ||
         type == OperatorType::kStridedSlice ||
         type == OperatorType::kDepthToSpace ||
         type == OperatorType::kLstmCell || type == OperatorType::kGather ||
         type == OperatorType::kTranspose || type == OperatorType::kMean ||
         type == OperatorType::kGreater ||
         type == OperatorType::kGreaterEqual || type == OperatorType::kLess ||
         type == OperatorType::kLessEqual || type == OperatorType::kSelect ||
         type == OperatorType::kArgMax || type == OperatorType::kRelu ||
         type == OperatorType::kRelu1 || type == OperatorType::kRelu6 ||
         type == OperatorType::kShape || type == OperatorType::kExpandDims ||
         type == OperatorType::kPack || type == OperatorType::kTopK_V2;
}

// The quantized op allows output arrays of type float using
// the attribute support_output_type_float_in_quantized_op
bool SupportOutputTypeFloatInQuantizedOp(const Operator& op) {
  auto type = op.type;
  if (type == OperatorType::kUnsupported) {
    auto* unsupported = static_cast<const TensorFlowUnsupportedOperator*>(&op);
    return unsupported->support_output_type_float_in_quantized_op;
  }
  return false;
}
const MinMax& GetOrComputeMinMax(Model* model, const string& array_name) {
  auto& array = model->GetArray(array_name);
  // Normally we should have a MinMax recorded on this Array,
  // so we just use it.
  if (array.minmax != nullptr) {
    return *array.minmax;
  }

  // We don't have a MinMax. That's bad news: we need
  // the graph to provide MinMax info for all arrays in order
  // for inference to reproduce faithfully the same quantization
  // error as the training process had.
  //
  // But we still want to support a fallback for constant arrays,
  // just using the plain min and max computed from array elements.
  // We should hopefully never rely on that in production, as that
  // will not give very good accuracy as that typically won't be
  // exactly what the training process used. But it will be useful
  // to allow easily trying out quantization even if the graph
  // lacks some minmax information.
  if (array.buffer != nullptr) {
    CHECK(array.buffer->type == ArrayDataType::kFloat);
    const auto& data = array.GetBuffer<ArrayDataType::kFloat>().data;
    // We always want [min, max] to contain 0.
    float min = 0.f;
    float max = 0.f;
    for (auto val : data) {
      min = std::min(min, val);
      max = std::max(max, val);
    }
    if (min == 0.f && max == 0.f) {
      // Prevent downstream anger from quantized math that expects min and max
      // to not be equal.
      max = 1.f;
    }
    // No need to warn about accuracy if all array values are equal to either
    // min or max:
    // in that case, quantization is exact, and such arrays are not learned
    // weights arrays for which fake-quantization would make sense, rather
    // they tend to be hardcoded arrays of zeros or ones used in some graphs.
    bool is_quantization_trivially_exact = true;
    for (auto val : data) {
      is_quantization_trivially_exact &= (val == min || val == max);
    }
    if (!is_quantization_trivially_exact) {
      LOG(WARNING)
          << "Constant array " << array_name
          << " lacks MinMax information. To make up for that, we will now "
             "compute"
          << " the MinMax from actual array elements. That will result in"
          << " quantization parameters that probably do not match whichever "
             "arithmetic"
          << " was used during training, and thus will probably be a cause of "
             "poor"
          << " inference accuracy.";
    }
    auto& minmax = array.GetOrCreateMinMax();
    minmax.min = min;
    minmax.max = max;
    return minmax;
  }

  LOG(FATAL) << "Array " << array_name
             << " does not have MinMax information, "
                "and is not a constant array. Cannot "
                "proceed with quantization.";
}

struct QuantizationPoints {
  int64 min_value;
  int64 max_value;
  int64 central_value;
};

template <ArrayDataType A>
QuantizationPoints GetQuantizationPoints() {
  QuantizationPoints qp;
  using Integer = DataType<A>;
  qp.min_value = std::numeric_limits<Integer>::min();
  qp.max_value = std::numeric_limits<Integer>::max();
  // eg [-128,127]...
  qp.central_value = (qp.min_value / 2 +        // -128 -> -64.
                      (qp.max_value - 1) / 2 +  // 127 -> 63.
                      1);
  return qp;
}

QuantizationPoints GetQuantizationPoints(ArrayDataType data_type) {
  switch (data_type) {
    case ArrayDataType::kUint8:
      return GetQuantizationPoints<ArrayDataType::kUint8>();
    case ArrayDataType::kInt16:
      return GetQuantizationPoints<ArrayDataType::kInt16>();
    case ArrayDataType::kInt32:
      return GetQuantizationPoints<ArrayDataType::kInt32>();
    default:
      LOG(FATAL) << "Unhandled case.";
  }
}

bool ChooseQuantizationForOperatorInput(
    GraphTransformation* transformation, Model* model, const Operator& op,
    std::size_t input_index, ArrayDataType* quantized_data_type,
    QuantizationParams* quantization_params) {
  const auto& input = op.inputs[input_index];
  auto& array = model->GetArray(input);
  if (array.data_type != ArrayDataType::kFloat) {
    return false;
  }

  // Quantization of bias vectors
  bool is_bias_vector = false;
  int activations_input_index;
  int weights_input_index;
  if (op.type == OperatorType::kConv ||
      op.type == OperatorType::kDepthwiseConv ||
      op.type == OperatorType::kFullyConnected) {
    if (input_index == 2) {
      is_bias_vector = true;
      activations_input_index = 0;
      weights_input_index = 1;
    }
  }
  if (op.type == OperatorType::kLstmCell) {
    if (input_index == LstmCellOperator::BIASES_INPUT) {
      is_bias_vector = true;
      activations_input_index = LstmCellOperator::DATA_INPUT;
      weights_input_index = LstmCellOperator::WEIGHTS_INPUT;
    }
  }
  if (is_bias_vector) {
    // Quantization of bias vector.
    // We need both of the mandatory inputs (input activations and weights) to
    // have been already quantized.
    const auto& input_activations =
        model->GetArray(op.inputs[activations_input_index]);
    const auto& input_weights = model->GetArray(op.inputs[weights_input_index]);
    if (!input_activations.quantization_params ||
        !input_weights.quantization_params) {
      transformation->AddMessageF(
          "Input array %s is a bias vector but has no qparams", input);
      return false;
    }
    const auto input_activations_scale =
        input_activations.quantization_params->scale;
    const auto input_weights_scale = input_weights.quantization_params->scale;
    quantization_params->scale = input_activations_scale * input_weights_scale;
    quantization_params->zero_point = 0;
    *quantized_data_type = GetQuantizedDataType(array, ArrayDataType::kInt32);
    transformation->AddMessageF(
        "Input array %s is a bias vector. Choosing quantization params "
        "accordingly.",
        input);
    return true;
  }

  const MinMax& minmax = GetOrComputeMinMax(model, input);

  if (op.type == OperatorType::kLstmCell) {
    if (input_index == LstmCellOperator::PREV_STATE_INPUT) {
      *quantized_data_type = ArrayDataType::kInt16;
      ChooseQuantizationParamsForArrayAndQuantizedDataType(
          array, *quantized_data_type, quantization_params);
      return true;
    }
  }

  *quantized_data_type = GetQuantizedDataType(array, ArrayDataType::kUint8);
  ChooseQuantizationParamsForArrayAndQuantizedDataType(
      array, *quantized_data_type, quantization_params);
  transformation->AddMessageF(
      "For input array %s with min=%g, max=%g, chose to quantize as %s (f=%s) "
      "with zero_point=%d, scale=%g",
      input, minmax.min, minmax.max, ArrayDataTypeName(*quantized_data_type),
      ArrayDataTypeName(array.final_data_type), quantization_params->zero_point,
      quantization_params->scale);
  return true;
}

bool IsExactlyRepresentable(double real_value, ArrayDataType data_type,
                            const QuantizationParams& quantization_params) {
  const double scaled_value =
      quantization_params.zero_point + real_value / quantization_params.scale;
  const double fractional_scaled_value =
      scaled_value - std::round(scaled_value);
  if (std::abs(fractional_scaled_value) > 1e-12) {
    return false;
  }
  const double rounded_scaled_value = std::round(scaled_value);
  if (data_type == ArrayDataType::kUint8) {
    if (rounded_scaled_value < 0 || rounded_scaled_value > 255) {
      return false;
    }
  }
  return true;
}

// Quantized data type is preset to the type of the input before this function.
bool ChooseHardcodedQuantizationForOperatorOutput(
    const Operator& op, const Array& array, ArrayDataType* quantized_data_type,
    QuantizationParams* quantization_params) {
  if (op.type == OperatorType::kL2Normalization) {
    // L2Normalization has range: [-1, 1].
    // 0 should be exactly representable, as values will typically be centered
    // around 0, with many values near 0.
    *quantized_data_type = GetQuantizedDataType(array, *quantized_data_type);
    const QuantizationPoints qp = GetQuantizationPoints(*quantized_data_type);
    quantization_params->zero_point = qp.central_value;
    quantization_params->scale = 1. / (qp.central_value - qp.min_value);
    CHECK(
        IsExactlyRepresentable(0., *quantized_data_type, *quantization_params));
    return true;
  }
  if (op.type == OperatorType::kLogistic || op.type == OperatorType::kSoftmax) {
    // Logistic and Softmax have range: [0, 1].
    //
    // For Logistic, 0.5 should be exactly representable, as implementations
    // will typically exploit the symmetry logistic(-x) = 1 - logistic(x), and
    // the glueing of the two halves of the graph will only be seamless if we
    // are accurately representing logistic(0) == 0.5.
    *quantized_data_type = GetQuantizedDataType(array, *quantized_data_type);
    const QuantizationPoints qp = GetQuantizationPoints(*quantized_data_type);
    quantization_params->zero_point = 0;
    quantization_params->scale = 1. / (qp.max_value + 1);
    CHECK(IsExactlyRepresentable(0.5, *quantized_data_type,
                                 *quantization_params));
    return true;
  }
  if (op.type == OperatorType::kLogSoftmax) {
    // LogSoftmax has range: [LogSoftmaxOperator::kOutputRangeMin, 0].
    *quantized_data_type = GetQuantizedDataType(array, *quantized_data_type);
    const QuantizationPoints qp = GetQuantizationPoints(*quantized_data_type);
    quantization_params->zero_point = qp.max_value;
    quantization_params->scale =
        -LogSoftmaxOperator::kOutputRangeMin / (qp.max_value + 1);
    // While not strictly necessary, it is easier to interpret output data and
    // quantization if the scale is similar to others (such as power of 2).
    CHECK(IsExactlyRepresentable(LogSoftmaxOperator::kOutputRangeMin / 2,
                                 *quantized_data_type, *quantization_params));
    return true;
  }
  if (op.type == OperatorType::kTanh) {
    // Tanh has the range: [-1, 1].
    *quantized_data_type = GetQuantizedDataType(array, *quantized_data_type);
    const QuantizationPoints qp = GetQuantizationPoints(*quantized_data_type);
    quantization_params->zero_point = qp.central_value;
    quantization_params->scale = 1. / (qp.central_value - qp.min_value);
    // 0 should be exactly representable, as values will typically be centered
    // around 0, with many values near 0.
    CHECK(
        IsExactlyRepresentable(0., *quantized_data_type, *quantization_params));
    return true;
  }
  return false;
}

bool ChooseQuantizationForOperatorOutput(
    GraphTransformation* transformation, Model* model, const Operator& op,
    std::size_t output_index, ArrayDataType* quantized_data_type,
    QuantizationParams* quantization_params) {
  const auto& output = op.outputs[output_index];
  auto& array = model->GetArray(output);
  if (array.data_type != ArrayDataType::kFloat) {
    transformation->AddMessageF("Array data type already set to %s, final=%s",
                                ArrayDataTypeName(array.data_type),
                                ArrayDataTypeName(array.final_data_type));
    return false;
  }
  *quantized_data_type = model->GetArray(op.inputs[0]).data_type;
  if (ChooseHardcodedQuantizationForOperatorOutput(
          op, array, quantized_data_type, quantization_params)) {
    transformation->AddMessageF(
        "Output array %s is produced by a %s operator. Choosing fixed "
        "quantization params accordingly.",
        output, OperatorTypeName(op.type));
    return true;
  }
  if ((op.type == OperatorType::kConcatenation &&
       model->flags.change_concat_input_ranges()) ||
      op.type == OperatorType::kDepthToSpace ||
      op.type == OperatorType::kSpaceToDepth ||
      op.type == OperatorType::kReshape || op.type == OperatorType::kSplit ||
      op.type == OperatorType::kRelu || op.type == OperatorType::kRelu1 ||
      op.type == OperatorType::kRelu6) {
    int data_input_index = 0;
    if (op.type == OperatorType::kSplit) {
      data_input_index = 1;
    }
    // Copying and rearrangement ops should preserve the quantization parameters
    // of the input array.
    const auto& input_array = model->GetArray(op.inputs[data_input_index]);
    const auto& input_quantization_params = input_array.GetQuantizationParams();
    *quantized_data_type =
        GetQuantizedDataType(input_array, ArrayDataType::kUint8);
    *quantized_data_type = GetQuantizedDataType(array, *quantized_data_type);
    quantization_params->zero_point = input_quantization_params.zero_point;
    quantization_params->scale = input_quantization_params.scale;

    transformation->AddMessageF(
        "Output array %s is produced by a %s operator. Copying quantization "
        "params from input array.",
        output, OperatorTypeName(op.type));
    return true;
  }
  const MinMax& minmax = GetOrComputeMinMax(model, output);
  if (op.type == OperatorType::kLstmCell) {
    if (output_index == LstmCellOperator::STATE_OUTPUT ||
        output_index == LstmCellOperator::ACTIV_TEMP) {
      *quantized_data_type = ArrayDataType::kInt16;
      ChooseQuantizationParamsForArrayAndQuantizedDataType(
          array, *quantized_data_type, quantization_params);
      return true;
    }
  }
  *quantized_data_type = GetQuantizedDataType(array, ArrayDataType::kUint8);
  ChooseQuantizationParamsForArrayAndQuantizedDataType(
      array, *quantized_data_type, quantization_params);
  transformation->AddMessageF(
      "For output array %s with min=%g, max=%g"
      ", chose to quantize as %s with zero_point=%d"
      ", scale=%g",
      output, minmax.min, minmax.max, ArrayDataTypeName(*quantized_data_type),
      quantization_params->zero_point, quantization_params->scale);

  return true;
}

// Fixes array minmax info to match the quantization parameters.
// This is required for when quantization parameters change for an array during
// quantization (such as ChooseQuantizationForOperatorOutput).
void FixMinMaxPostQuantization(GraphTransformation* transformation,
                               ArrayDataType quantized_data_type,
                               const QuantizationParams& quantization_params,
                               MinMax* minmax) {
  double quantized_min, quantized_max;
  if (!GetQuantizedDataTypeNumericalRange(quantized_data_type, &quantized_min,
                                          &quantized_max)) {
    // Not quantized - no update required.
    return;
  }

  // Compute new minmax values.
  double min = (quantized_min - quantization_params.zero_point) *
               quantization_params.scale;
  double max = (quantized_max - quantization_params.zero_point) *
               quantization_params.scale;

  // If we are close to the existing minmax values don't bother changing them.
  // This prevents propagating small floating point precision errors.
  constexpr double kMinMaxThreshold = 1e-5;
  const double width = max - min;
  if (std::abs(min - minmax->min) > kMinMaxThreshold * width ||
      std::abs(max - minmax->max) > kMinMaxThreshold * width) {
    transformation->AddMessageF(
        "Adjusting min/max from %g,%g to %g,%g to match quantization params",
        minmax->min, minmax->max, min, max);
    minmax->min = min;
    minmax->max = max;
  }
}

}  // namespace

::tensorflow::Status Quantize::Run(Model* model, std::size_t op_index,
                                   bool* modified) {
  *modified = false;
  // Our general "quantization" graph transformation consists in replacing
  //   QuantizedInputArrays[] ->
  //     DequantizeOperators[] ->
  //       FloatInputArrays[] ->
  //         Operator ->
  //           FloatOutputArray
  // by
  //   QuantizedInputArrays[] ->
  //     Operator ->
  //       QuantizedOutputArray ->
  //         DequantizeOperator ->
  //           FloatOutputArray
  //
  // In other words, this is pushing Dequantize operators to the right of
  // other operators.
  //

  auto& op = *model->operators[op_index];
  if (op.type == OperatorType::kDequantize ||
      op.type == OperatorType::kFakeQuant) {
    return ::tensorflow::Status::OK();
  }

  // Our assumption here is that the input arrays are already quantized -
  // that is typically the case in models operating on an input bitmap
  // image, and MakeInitialDequantizeOp should have already resolved
  // the handling of the input image as an initial Dequantize op.
  //
  // Thus we are building around the assumption that the graph always starts
  // with a quantized input array, and only after some Dequantize op do we have
  // float arrays. The problem of quantizing the graph thus becomes a problem of
  // pushing Dequantize ops to the right of other ops.
  //
  // Let us just guard this assumption by the following assertion:
  for (const auto& input : op.inputs) {
    const auto& input_array = model->GetArray(input);
    if (IsInputArray(*model, input) &&
        input_array.data_type == ArrayDataType::kFloat) {
      CHECK(input_array.quantization_params)
          << "Input array " << input << " is missing quantization_params";
    }
  }
  if (!SupportsQuantization(op)) {
    LOG(FATAL) << "Unimplemented: this graph contains an operator of type "
               << HelpfulOperatorTypeName(op)
               << " for which the quantized form is not yet implemented. "
                  "Sorry, and patches welcome (that's a relatively fun patch "
                  "to write, mostly providing the actual quantized arithmetic "
                  "code for this op).";
  }

  for (const auto& input : op.inputs) {
    const auto& array = model->GetArray(input);
    if (array.data_type == ArrayDataType::kFloat) {
      if (!array.minmax && !array.buffer) {
        LOG(ERROR) << "Can't quantize input array " << input
                   << " because it lacks min/max info";
        return ::tensorflow::Status::OK();
      }
      const auto* other_op = GetOpWithOutput(*model, input);
      if (other_op && other_op->type != OperatorType::kDequantize) {
        AddMessageF(
            "Not quantizing %s for now, because its input array %s is not "
            "produced by a Dequantize op, "
            "which means that we should yield and let other ops "
            "get quantized first",
            LogName(op), input);
        return ::tensorflow::Status::OK();
      }
    }
  }

  bool changed = false;

  // Quantize inputs, remove any Dequantize op on the inputs side
  for (std::size_t input_index = 0; input_index < op.inputs.size();
       input_index++) {
    ArrayDataType quantized_data_type;
    QuantizationParams quantization_params;
    if (ChooseQuantizationForOperatorInput(this, model, op, input_index,
                                           &quantized_data_type,
                                           &quantization_params)) {
      const auto& input = op.inputs[input_index];
      if (IsConstantParameterArray(*model, input)) {
        QuantizeArray(this, model, input, quantized_data_type,
                      quantization_params);
        changed = true;
      } else {
        auto dequantize_it = FindOpWithOutput(*model, input);
        if (dequantize_it != model->operators.end()) {
          auto* dequantize_op = dequantize_it->get();
          CHECK(dequantize_op->type == OperatorType::kDequantize);
          op.inputs[input_index] = dequantize_op->inputs[0];
          // Check if the output of that Dequantize op was not used by any
          // other operator. We will then erase that Dequantize op.
          if (!CountOpsWithInput(*model, dequantize_op->outputs[0])) {
            if (IsDiscardableArray(*model, dequantize_op->outputs[0])) {
              // Usual case: we can just discard the dequantize output.
              model->EraseArray(dequantize_op->outputs[0]);
            } else {
              // The dequantize output is not discardable. Special care needed.
              // If any of the model's output_arrays was pointing to the
              // Dequantize op's output, let it point to the Dequantize op's
              // input instead.
              for (int i = 0; i < model->flags.output_arrays_size(); i++) {
                if (model->flags.output_arrays(i) ==
                    dequantize_op->outputs[0]) {
                  // TODO(b/78013785): never rename output arrays.
                  if (IsInputArray(*model, dequantize_op->inputs[0])) {
                    // The op input is an input array and the output is an
                    // output array and we can't have an array be both. Insert a
                    // copy op to ensure the two arrays stay separate.
                    AddMessageF(
                        "Tried to rename output array %d while removing "
                        "dequant "
                        "op %s but array is also an input; inserting copy %s "
                        "-> %s",
                        i, LogName(*dequantize_op),
                        model->flags.output_arrays(i),
                        dequantize_op->inputs[0]);
                    InsertCopyOperator(model, dequantize_op->inputs[0],
                                       dequantize_op->outputs[0]);
                  } else {
                    // Op output is strictly used as an output array, so we can
                    // just rename the array and directly bypass the op.
                    AddMessageF(
                        "Renaming output array %d after removing dequant op "
                        "%s: "
                        "%s -> %s",
                        i, LogName(*dequantize_op),
                        model->flags.output_arrays(i),
                        dequantize_op->inputs[0]);
                    model->flags.set_output_arrays(i, dequantize_op->inputs[0]);
                    model->EraseArray(dequantize_op->outputs[0]);
                  }
                  break;
                }
              }
            }
            model->operators.erase(dequantize_it);
          }
          changed = true;
        } else {
          // This input array is not produced by a Dequantize op.
          // We have encountered this situation in RNN graphs, whose cyclic
          // nature defeats the basic assumption underlying the quantization
          // algorithm implemented here. For now, when we have seen this
          // happening, the array in question was a RNN state array itself,
          // so let us just implement this case here, and guard that assumption
          // with a CHECK. A more general fix would involve revisiting the
          // design of this whole Quantization transformation.
          bool is_rnn_state_array = false;
          for (const auto& rnn_state : model->flags.rnn_states()) {
            if (rnn_state.state_array() == input) {
              is_rnn_state_array = true;
              break;
            }
          }
          CHECK(is_rnn_state_array);
          QuantizeArray(this, model, input, quantized_data_type,
                        quantization_params);
          changed = true;
        }
      }
    }
  }

  // Quantize outputs, add Dequantize ops as needed on the outputs side
  if (SupportOutputTypeFloatInQuantizedOp(op)) {
    LOG(WARNING)
        << HelpfulOperatorTypeName(op) << " is a quantized op"
        << "but it has a model flag that sets the output arrays to float.";
  } else {
    for (std::size_t output_index = 0; output_index < op.outputs.size();
         output_index++) {
      QuantizationParams quantization_params;
      ArrayDataType quantized_data_type;
      if (ChooseQuantizationForOperatorOutput(this, model, op, output_index,
                                              &quantized_data_type,
                                              &quantization_params)) {
        changed = true;
        const auto& output = op.outputs[output_index];
        auto& output_array = model->GetArray(output);

        // Fix up the min/max information on the output array to match the
        // chosen quantization parameters.
        CHECK(output_array.minmax)
            << "Output array named " << output << " lacks minmax";
        auto& output_minmax = output_array.GetMinMax();
        FixMinMaxPostQuantization(this, quantized_data_type,
                                  quantization_params, &output_minmax);

        QuantizeArray(this, model, output, quantized_data_type,
                      quantization_params);

        const auto& dequantized_output =
            AvailableArrayName(*model, output + "_dequantized");
        auto& dequantized_output_array =
            model->GetOrCreateArray(dequantized_output);
        dequantized_output_array.data_type = ArrayDataType::kFloat;
        dequantized_output_array.final_data_type = output_array.data_type;
        auto& dequantized_output_minmax =
            dequantized_output_array.GetOrCreateMinMax();
        dequantized_output_minmax.min = output_minmax.min;
        dequantized_output_minmax.max = output_minmax.max;
        for (const auto& other_op : model->operators) {
          for (auto& other_op_input : other_op->inputs) {
            if (other_op_input == output) {
              other_op_input = dequantized_output;
            }
          }
        }
        auto* dequantize_op = new DequantizeOperator;
        dequantize_op->inputs = {output};
        dequantize_op->outputs = {dequantized_output};
        for (int i = 0; i < model->flags.output_arrays_size(); i++) {
          if (model->flags.output_arrays(i) == output) {
            // TODO(b/78013785): never rename output arrays.
            AddMessageF(
                "Renaming output array %d after inserting dequant op %s: %s -> "
                "%s",
                i, LogName(*dequantize_op), model->flags.output_arrays(i),
                dequantized_output);
            model->flags.set_output_arrays(i, dequantized_output);
          }
        }
        const auto op_it = FindOp(*model, &op);
        model->operators.emplace(op_it + 1, dequantize_op);
      }
    }
  }

  *modified = changed;
  return ::tensorflow::Status::OK();
}

}  // namespace toco