tensorflow/lite/tools/optimize/quantize_model_test.cc

/* Copyright 2018 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/tools/optimize/quantize_model.h"

#include <cstddef>
#include <cstdint>
#include <memory>

#include <gmock/gmock.h>
#include <gtest/gtest.h>
#include "flatbuffers/flatbuffers.h"  // TF:flatbuffers
#include "flatbuffers/flexbuffers.h"  // TF:flatbuffers
#include "tensorflow/core/lib/io/path.h"
#include "tensorflow/core/platform/init_main.h"
#include "tensorflow/core/util/command_line_flags.h"
#include "tensorflow/lite/model.h"
#include "tensorflow/lite/schema/schema_generated.h"
#include "tensorflow/lite/tools/optimize/test_util.h"

// Note: More rigorous model tests can be found in subgraph_quantizer_test.cc

namespace {
tensorflow::string* g_test_model_dir = nullptr;
}  // namespace

namespace tflite {
namespace optimize {
namespace {

std::unique_ptr<FlatBufferModel> ReadModel(const string& model_name) {
  auto model_path = tensorflow::io::JoinPath(*g_test_model_dir, model_name);
  return FlatBufferModel::BuildFromFile(model_path.c_str());
}

template <typename T>
std::vector<T> GetAsVector(const flatbuffers::Vector<T>* vec) {
  return std::vector<T>(vec->begin(), vec->end());
}

void VerifyAsymmetricQuantizationScale(
    const QuantizationParameters& float_quant_params,
    const QuantizationParametersT& quantized_quant_params) {
  const float eps = 1e-7;
  ASSERT_EQ(float_quant_params.min()->size(), 1);
  ASSERT_EQ(float_quant_params.max()->size(), 1);
  float float_min = std::min(0.f, float_quant_params.min()->Get(0));
  float float_max = std::max(0.f, float_quant_params.max()->Get(0));

  ASSERT_EQ(quantized_quant_params.scale.size(), 1);
  ASSERT_EQ(quantized_quant_params.zero_point.size(), 1);

  float scale = (float_max - float_min) / 255;
  EXPECT_NEAR(scale, quantized_quant_params.scale[0], eps);
}

class QuantizeModelTest : public testing::Test {
 protected:
  QuantizeModelTest() {
    input_model_ = ReadModel(internal::kConvModelWith0Plus10Weights);
    readonly_model_ = input_model_->GetModel();
    readonly_model_->UnPackTo(&model_);
  }

  std::unique_ptr<FlatBufferModel> input_model_;
  const Model* readonly_model_;
  tflite::ModelT model_;
  flatbuffers::FlatBufferBuilder builder_;
  internal::FailOnErrorReporter error_reporter_;
};

class QuantizeConvModelTest : public QuantizeModelTest {
 protected:
  QuantizeConvModelTest() {
    input_model_ = ReadModel(internal::kConvModelWith0Plus10Weights);
    readonly_model_ = input_model_->GetModel();
    readonly_model_->UnPackTo(&model_);
  }
};

TEST_F(QuantizeConvModelTest, QuantizationSucceeds) {
  auto status = QuantizeModel(&builder_, &model_, TensorType_INT8,
                              TensorType_INT8, &error_reporter_);
  EXPECT_EQ(status, kTfLiteOk);
  const uint8_t* buffer = builder_.GetBufferPointer();
  const Model* output_model = GetModel(buffer);
  ASSERT_TRUE(output_model);
}

TEST_F(QuantizeConvModelTest, SkipUnspecifiedLayer) {
  auto status =
      QuantizeModel(&builder_, &model_, TensorType_FLOAT32, TensorType_FLOAT32,
                    /*allow_float=*/true, {}, &error_reporter_);
  EXPECT_EQ(status, kTfLiteOk);
  ASSERT_EQ(model_.subgraphs.size(), readonly_model_->subgraphs()->size());
  // The resulting model should be the same.
  ASSERT_EQ(model_.subgraphs.size(), readonly_model_->subgraphs()->size());
  for (size_t subgraph_idx = 0; subgraph_idx < model_.subgraphs.size();
       subgraph_idx++) {
    const auto quantized_graph = model_.subgraphs[subgraph_idx].get();
    const auto float_graph = readonly_model_->subgraphs()->Get(subgraph_idx);
    ASSERT_EQ(quantized_graph->tensors.size(), float_graph->tensors()->size());
    for (size_t i = 0; i < quantized_graph->tensors.size(); i++) {
      const auto quant_tensor = quantized_graph->tensors[i].get();
      const auto float_tensor = float_graph->tensors()->Get(i);
      EXPECT_EQ(quant_tensor->buffer, float_tensor->buffer());
      EXPECT_EQ(quant_tensor->is_variable, float_tensor->is_variable());
      EXPECT_EQ(quant_tensor->shape, GetAsVector(float_tensor->shape()));
      EXPECT_EQ(quant_tensor->name, float_tensor->name()->str());
      EXPECT_EQ(quant_tensor->type, float_tensor->type());
    }
  }
}

TEST_F(QuantizeConvModelTest, TensorShapesAndStructureIsUnchanged) {
  auto status = QuantizeModel(&builder_, &model_, TensorType_INT8,
                              TensorType_INT8, &error_reporter_);
  EXPECT_EQ(status, kTfLiteOk);
  ASSERT_EQ(model_.subgraphs.size(), readonly_model_->subgraphs()->size());
  for (size_t subgraph_idx = 0; subgraph_idx < model_.subgraphs.size();
       subgraph_idx++) {
    const auto quantized_graph = model_.subgraphs[subgraph_idx].get();
    const auto float_graph = readonly_model_->subgraphs()->Get(subgraph_idx);
    ASSERT_EQ(quantized_graph->tensors.size(), float_graph->tensors()->size());
    for (size_t i = 0; i < quantized_graph->tensors.size(); i++) {
      const auto quant_tensor = quantized_graph->tensors[i].get();
      const auto float_tensor = float_graph->tensors()->Get(i);
      EXPECT_EQ(quant_tensor->buffer, float_tensor->buffer());
      EXPECT_EQ(quant_tensor->is_variable, float_tensor->is_variable());
      EXPECT_EQ(quant_tensor->shape, GetAsVector(float_tensor->shape()));
      EXPECT_EQ(quant_tensor->name, float_tensor->name()->str());
    }
  }
  // check op and versioning.
  EXPECT_EQ(model_.operator_codes.size(), 1);
  EXPECT_EQ(model_.operator_codes[0]->builtin_code, BuiltinOperator_CONV_2D);
  EXPECT_EQ(model_.operator_codes[0]->version, 3);
}

TEST_F(QuantizeConvModelTest, OperatorsAreUnchanged) {
  auto status = QuantizeModel(&builder_, &model_, TensorType_INT8,
                              TensorType_INT8, &error_reporter_);
  EXPECT_EQ(status, kTfLiteOk);
  ASSERT_EQ(model_.operator_codes.size(),
            readonly_model_->operator_codes()->size());
  for (size_t i = 0; i < model_.operator_codes.size(); i++) {
    const auto float_model_op = readonly_model_->operator_codes()->Get(i);
    EXPECT_EQ(model_.operator_codes[i]->builtin_code,
              float_model_op->builtin_code());
    if (model_.operator_codes[i]->builtin_code == BuiltinOperator_CONV_2D) {
      EXPECT_EQ(model_.operator_codes[i]->version, 3);
    } else {
      EXPECT_EQ(model_.operator_codes[i]->version, 2);
    }
  }

  ASSERT_EQ(model_.subgraphs.size(), readonly_model_->subgraphs()->size());
  for (size_t subgraph_idx = 0; subgraph_idx < model_.subgraphs.size();
       subgraph_idx++) {
    const auto quantized_graph = model_.subgraphs[subgraph_idx].get();
    const auto float_graph = readonly_model_->subgraphs()->Get(subgraph_idx);
    ASSERT_EQ(quantized_graph->operators.size(),
              float_graph->operators()->size());
    for (size_t i = 0; i < quantized_graph->operators.size(); i++) {
      const auto quant_op = quantized_graph->operators[i].get();
      const auto float_op = float_graph->operators()->Get(i);
      EXPECT_EQ(quant_op->inputs, GetAsVector(float_op->inputs()));
      EXPECT_EQ(quant_op->outputs, GetAsVector(float_op->outputs()));
      EXPECT_EQ(quant_op->opcode_index, float_op->opcode_index());
    }
  }
}

TEST_F(QuantizeConvModelTest, GraphIsFullyQuantized) {
  auto status = QuantizeModel(&builder_, &model_, TensorType_INT8,
                              TensorType_INT8, &error_reporter_);
  EXPECT_EQ(status, kTfLiteOk);
  for (const auto& subgraph : model_.subgraphs) {
    for (const auto& tensor : subgraph->tensors) {
      EXPECT_TRUE(tensor->type == TensorType_INT32 ||
                  tensor->type == TensorType_INT8);
    }
  }
}

TEST_F(QuantizeConvModelTest, FloatInputAndOutput) {
  auto status = QuantizeModel(&builder_, &model_, &error_reporter_);
  EXPECT_EQ(status, kTfLiteOk);

  for (int32_t subgraph_idx = 0; subgraph_idx < model_.subgraphs.size();
       ++subgraph_idx) {
    const auto& subgraph = model_.subgraphs[subgraph_idx];
    const auto& readonly_subgraph =
        readonly_model_->subgraphs()->Get(subgraph_idx);
    // The model has one input and output, so the converted model should have
    // two extra ops, a Quantize and Dequantize.
    EXPECT_EQ(subgraph->operators.size(),
              readonly_subgraph->operators()->size() + 2);
    // Check that the first op is Quantize and the last is Dequant.
    const auto& quant_op = subgraph->operators[0];
    const auto& dequant_op =
        subgraph->operators[subgraph->operators.size() - 1];
    const int32_t quant_idx = quant_op->opcode_index;
    const int32_t dequant_idx = dequant_op->opcode_index;
    EXPECT_EQ(model_.operator_codes[quant_idx]->builtin_code,
              BuiltinOperator_QUANTIZE);
    EXPECT_EQ(model_.operator_codes[dequant_idx]->builtin_code,
              BuiltinOperator_DEQUANTIZE);
    // The model should only have one input and output.
    EXPECT_EQ(subgraph->inputs.size(), 1);
    EXPECT_EQ(subgraph->outputs.size(), 1);
    const int32_t input_idx = subgraph->inputs[0];
    const int32_t output_idx = subgraph->outputs[0];
    // Ensure: new input -> Quant -> old input.
    EXPECT_EQ(quant_op->inputs[0], input_idx);
    EXPECT_EQ(quant_op->outputs[0], readonly_subgraph->inputs()->Get(0));
    // Ensure: old output -> dequant -> new output.
    EXPECT_EQ(dequant_op->inputs[0], readonly_subgraph->outputs()->Get(0));
    EXPECT_EQ(dequant_op->outputs[0], output_idx);
    // The input and output types should be float.
    EXPECT_EQ(subgraph->tensors[input_idx]->type, TensorType_FLOAT32);
    EXPECT_EQ(subgraph->tensors[input_idx]->name, "input");
    EXPECT_EQ(subgraph->tensors[output_idx]->type, TensorType_FLOAT32);
    EXPECT_EQ(subgraph->tensors[output_idx]->name, "output");
    // The original input and output has been renamed.
    EXPECT_EQ(subgraph->tensors[quant_op->outputs[0]]->name, "input_int8");
    EXPECT_EQ(subgraph->tensors[dequant_op->inputs[0]]->name, "output_int8");
    for (int tensor_idx = 0; tensor_idx < subgraph->tensors.size();
         ++tensor_idx) {
      const auto& tensor = subgraph->tensors[tensor_idx];
      if (input_idx != tensor_idx && output_idx != tensor_idx) {
        EXPECT_TRUE(tensor->type == TensorType_INT32 ||
                    tensor->type == TensorType_INT8);
      }
    }
  }
}

TEST_F(QuantizeConvModelTest, Uint8InputAndOutput) {
  auto status = QuantizeModel(&builder_, &model_, TensorType_UINT8,
                              TensorType_UINT8, &error_reporter_);
  EXPECT_EQ(status, kTfLiteOk);

  for (int32_t subgraph_idx = 0; subgraph_idx < model_.subgraphs.size();
       ++subgraph_idx) {
    const auto& subgraph = model_.subgraphs[subgraph_idx];
    const auto& readonly_subgraph =
        readonly_model_->subgraphs()->Get(subgraph_idx);
    // The model has one input and output, so the converted model should have
    // two extra ops, a Quantize and Dequantize.
    EXPECT_EQ(subgraph->operators.size(),
              readonly_subgraph->operators()->size() + 2);
    // Check that the first op is Quantize and the last is Dequant.
    const auto& quant_op_uint8_int8 = subgraph->operators[0];
    const auto& quant_op_int8_uint8 =
        subgraph->operators[subgraph->operators.size() - 1];
    const int32_t quant_op_uint8_int8_idx = quant_op_uint8_int8->opcode_index;
    const int32_t quant_op_int8_uint8_idx = quant_op_int8_uint8->opcode_index;
    EXPECT_EQ(model_.operator_codes[quant_op_uint8_int8_idx]->builtin_code,
              BuiltinOperator_QUANTIZE);
    EXPECT_EQ(model_.operator_codes[quant_op_int8_uint8_idx]->builtin_code,
              BuiltinOperator_QUANTIZE);
    // The model should only have one input and output.
    EXPECT_EQ(subgraph->inputs.size(), 1);
    EXPECT_EQ(subgraph->outputs.size(), 1);
    const int32_t input_idx = subgraph->inputs[0];
    const int32_t output_idx = subgraph->outputs[0];
    // Ensure: new input -> Quant -> old input.
    EXPECT_EQ(quant_op_uint8_int8->inputs[0], input_idx);
    EXPECT_EQ(quant_op_uint8_int8->outputs[0],
              readonly_subgraph->inputs()->Get(0));
    // Ensure: old output -> dequant -> new output.
    EXPECT_EQ(quant_op_int8_uint8->inputs[0],
              readonly_subgraph->outputs()->Get(0));
    EXPECT_EQ(quant_op_int8_uint8->outputs[0], output_idx);
    // The input and output types should be uint8.
    EXPECT_EQ(subgraph->tensors[input_idx]->type, TensorType_UINT8);
    EXPECT_EQ(subgraph->tensors[input_idx]->name, "input");
    EXPECT_EQ(subgraph->tensors[input_idx]->quantization->scale.size(), 1);
    EXPECT_FLOAT_EQ(subgraph->tensors[input_idx]->quantization->scale[0],
                    0.0392156877);
    EXPECT_EQ(subgraph->tensors[input_idx]->quantization->zero_point.size(), 1);
    EXPECT_EQ(subgraph->tensors[input_idx]->quantization->zero_point[0], 0);
    EXPECT_EQ(subgraph->tensors[output_idx]->type, TensorType_UINT8);
    EXPECT_EQ(subgraph->tensors[output_idx]->name, "output");
    EXPECT_EQ(subgraph->tensors[output_idx]->quantization->scale.size(), 1);
    EXPECT_FLOAT_EQ(subgraph->tensors[output_idx]->quantization->scale[0],
                    0.0392156877);
    EXPECT_EQ(subgraph->tensors[output_idx]->quantization->zero_point.size(),
              1);
    EXPECT_EQ(subgraph->tensors[output_idx]->quantization->zero_point[0], 0);
    // The original input and output has been renamed.
    EXPECT_EQ(subgraph->tensors[quant_op_uint8_int8->outputs[0]]->name,
              "input_int8");
    EXPECT_EQ(subgraph->tensors[quant_op_int8_uint8->inputs[0]]->name,
              "output_int8");
    for (int tensor_idx = 0; tensor_idx < subgraph->tensors.size();
         ++tensor_idx) {
      const auto& tensor = subgraph->tensors[tensor_idx];
      if (input_idx != tensor_idx && output_idx != tensor_idx) {
        EXPECT_TRUE(tensor->type == TensorType_INT32 ||
                    tensor->type == TensorType_INT8);
      }
    }
  }
}

class QuantizeConcatModelTest : public QuantizeModelTest {
 protected:
  QuantizeConcatModelTest() {
    input_model_ = ReadModel(internal::kFloatConcatMax5Max10Max10);
    readonly_model_ = input_model_->GetModel();
    readonly_model_->UnPackTo(&model_);
  }
};

// There are two inputs for concat, "input0" and "input1". "input0" has [0, 5]
// as min/max and "input1" has [0, 10] as min/max. The output "output" for
// concat has [0, 10] as min/max.
// After applyging QuantizeModel(), "input0" will have a requant op added, along
// with a tensor "input0_reqaunt" that has [0, 10] as min/max. So the topology
// becomes:
// input0 -> requant -> input0_requant \
//                                       concat - output
//                              input1 /
TEST_F(QuantizeConcatModelTest, AddRequantBeforeConcat) {
  auto status = QuantizeModel(&builder_, &model_, TensorType_INT8,
                              TensorType_INT8, &error_reporter_);
  EXPECT_EQ(status, kTfLiteOk);

  // There is only one subgraph.
  const int32_t subgraph_idx = 0;
  const auto& subgraph = model_.subgraphs[subgraph_idx];
  const auto& readonly_subgraph =
      readonly_model_->subgraphs()->Get(subgraph_idx);

  // There should be two ops: quant and concat.
  EXPECT_EQ(readonly_subgraph->operators()->size(), 1);
  EXPECT_EQ(subgraph->operators.size(), 2);
  const auto& requant = subgraph->operators[0];
  const auto& concat = subgraph->operators[1];
  EXPECT_EQ(model_.operator_codes[requant->opcode_index]->builtin_code,
            BuiltinOperator_QUANTIZE);
  EXPECT_EQ(model_.operator_codes[concat->opcode_index]->builtin_code,
            BuiltinOperator_CONCATENATION);

  // There should be 4 tensors: input0, input1, input0_requantized, output.
  EXPECT_EQ(subgraph->tensors.size(), 4);
  EXPECT_EQ(subgraph->tensors[0]->type, TensorType_INT8);
  EXPECT_EQ(subgraph->tensors[0]->name, "input0");
  EXPECT_EQ(subgraph->tensors[0]->quantization->scale.size(), 1);
  EXPECT_EQ(subgraph->tensors[0]->quantization->zero_point.size(), 1);
  EXPECT_FLOAT_EQ(subgraph->tensors[0]->quantization->scale[0], 0.019607844);
  EXPECT_FLOAT_EQ(subgraph->tensors[0]->quantization->zero_point[0], -128);
  EXPECT_EQ(subgraph->tensors[1]->type, TensorType_INT8);
  EXPECT_EQ(subgraph->tensors[1]->name, "input1");
  EXPECT_EQ(subgraph->tensors[1]->quantization->scale.size(), 1);
  EXPECT_EQ(subgraph->tensors[1]->quantization->zero_point.size(), 1);
  EXPECT_FLOAT_EQ(subgraph->tensors[1]->quantization->scale[0], 0.039215688);
  EXPECT_FLOAT_EQ(subgraph->tensors[1]->quantization->zero_point[0], -128);
  EXPECT_EQ(subgraph->tensors[2]->type, TensorType_INT8);
  EXPECT_EQ(subgraph->tensors[2]->name, "output");
  EXPECT_EQ(subgraph->tensors[2]->quantization->scale.size(), 1);
  EXPECT_EQ(subgraph->tensors[2]->quantization->zero_point.size(), 1);
  EXPECT_FLOAT_EQ(subgraph->tensors[2]->quantization->scale[0], 0.039215688);
  EXPECT_FLOAT_EQ(subgraph->tensors[2]->quantization->zero_point[0], -128);
  EXPECT_EQ(subgraph->tensors[3]->type, TensorType_INT8);
  EXPECT_EQ(subgraph->tensors[3]->name, "input0_requantized");
  EXPECT_EQ(subgraph->tensors[3]->quantization->scale.size(), 1);
  EXPECT_EQ(subgraph->tensors[3]->quantization->zero_point.size(), 1);
  EXPECT_FLOAT_EQ(subgraph->tensors[3]->quantization->scale[0], 0.039215688);
  EXPECT_FLOAT_EQ(subgraph->tensors[3]->quantization->zero_point[0], -128);

  // The connection should be what is described in the comment.
  EXPECT_EQ(requant->inputs.size(), 1);
  EXPECT_EQ(requant->outputs.size(), 1);
  EXPECT_EQ(requant->inputs[0], 0);
  EXPECT_EQ(requant->outputs[0], 3);
  EXPECT_EQ(concat->inputs.size(), 2);
  EXPECT_EQ(concat->outputs.size(), 1);
  EXPECT_EQ(concat->inputs[0], 3);
  EXPECT_EQ(concat->inputs[1], 1);
  EXPECT_EQ(concat->outputs[0], 2);

  // check op and versioning.
  EXPECT_EQ(model_.operator_codes.size(), 2);
  EXPECT_EQ(model_.operator_codes[0]->builtin_code,
            BuiltinOperator_CONCATENATION);
  EXPECT_EQ(model_.operator_codes[0]->version, 2);
  EXPECT_EQ(model_.operator_codes[1]->builtin_code, BuiltinOperator_QUANTIZE);
  EXPECT_EQ(model_.operator_codes[1]->version, 2);
}

class QuantizeSplitModelTest : public QuantizeModelTest {
 protected:
  QuantizeSplitModelTest() {
    input_model_ = ReadModel(internal::kModelSplit);
    readonly_model_ = input_model_->GetModel();
    readonly_model_->UnPackTo(&model_);
  }
};

// There are two outputs for split with different scales, the resulting model
// should have the scales be hardcodes to the input scale value.
TEST_F(QuantizeSplitModelTest, QuantizeSplit) {
  auto status = QuantizeModel(&builder_, &model_, TensorType_INT8,
                              TensorType_INT8, &error_reporter_);
  EXPECT_EQ(status, kTfLiteOk);

  // There is only one subgraph.
  const int32_t subgraph_idx = 0;
  const auto& subgraph = model_.subgraphs[subgraph_idx];
  const auto& readonly_subgraph =
      readonly_model_->subgraphs()->Get(subgraph_idx);

  // There should be two ops: the split and add in the original model.
  EXPECT_EQ(readonly_subgraph->operators()->size(), 2);
  EXPECT_EQ(subgraph->operators.size(), 2);
  const auto& split = subgraph->operators[0];
  const auto& add = subgraph->operators[1];
  EXPECT_EQ(model_.operator_codes[split->opcode_index]->builtin_code,
            BuiltinOperator_SPLIT);
  EXPECT_EQ(model_.operator_codes[add->opcode_index]->builtin_code,
            BuiltinOperator_ADD);

  // There should be 5 tensors: input, output, split, split/split_dim, split:1.
  EXPECT_EQ(subgraph->tensors.size(), 5);

  EXPECT_EQ(subgraph->tensors[0]->type, TensorType_INT8);
  EXPECT_EQ(subgraph->tensors[0]->name, "input");
  EXPECT_EQ(subgraph->tensors[0]->quantization->scale.size(), 1);
  EXPECT_EQ(subgraph->tensors[0]->quantization->zero_point.size(), 1);
  EXPECT_FLOAT_EQ(subgraph->tensors[0]->quantization->scale[0], 1.0);
  EXPECT_FLOAT_EQ(subgraph->tensors[0]->quantization->zero_point[0], -128);
  EXPECT_EQ(subgraph->tensors[1]->type, TensorType_INT8);
  EXPECT_EQ(subgraph->tensors[1]->name, "output");
  EXPECT_EQ(subgraph->tensors[1]->quantization->scale.size(), 1);
  EXPECT_EQ(subgraph->tensors[1]->quantization->zero_point.size(), 1);
  EXPECT_FLOAT_EQ(subgraph->tensors[1]->quantization->scale[0], 1.0);
  EXPECT_FLOAT_EQ(subgraph->tensors[1]->quantization->zero_point[0], -128);
  EXPECT_EQ(subgraph->tensors[2]->type, TensorType_INT8);
  EXPECT_EQ(subgraph->tensors[2]->name, "split");
  EXPECT_EQ(subgraph->tensors[2]->quantization->scale.size(), 1);
  EXPECT_EQ(subgraph->tensors[2]->quantization->zero_point.size(), 1);
  EXPECT_FLOAT_EQ(subgraph->tensors[2]->quantization->scale[0], 1.0);
  EXPECT_FLOAT_EQ(subgraph->tensors[2]->quantization->zero_point[0], -128);
  EXPECT_EQ(subgraph->tensors[4]->type, TensorType_INT8);
  EXPECT_EQ(subgraph->tensors[4]->name, "split:1");
  EXPECT_EQ(subgraph->tensors[4]->quantization->scale.size(), 1);
  EXPECT_EQ(subgraph->tensors[4]->quantization->zero_point.size(), 1);
  EXPECT_FLOAT_EQ(subgraph->tensors[4]->quantization->scale[0], 1.0);
  EXPECT_FLOAT_EQ(subgraph->tensors[4]->quantization->zero_point[0], -128);

  // check op and versioning.
  EXPECT_EQ(model_.operator_codes.size(), 2);
  EXPECT_EQ(model_.operator_codes[1]->builtin_code, BuiltinOperator_SPLIT);
  EXPECT_EQ(model_.operator_codes[0]->version, 2);
}

class QuantizeConvModel1Test : public QuantizeModelTest {
 protected:
  QuantizeConvModel1Test() {
    input_model_ = ReadModel(internal::kConvModelWithMinus128Plus127Weights);
    readonly_model_ = input_model_->GetModel();
    readonly_model_->UnPackTo(&model_);
  }
};

TEST_F(QuantizeConvModel1Test, VerifyConvQuantizationWithUnitScale) {
  auto status = QuantizeModel(&builder_, &model_, TensorType_INT8,
                              TensorType_INT8, &error_reporter_);
  EXPECT_EQ(status, kTfLiteOk);
  const auto& subgraph = model_.subgraphs[0];

  auto conv_op = subgraph->operators[0].get();
  const int input_tensor_idx = 0;
  const int weights_tensor_idx = 1;
  const int bias_tensor_index = 2;
  const int output_tensor_idx = 0;
  const auto bias_tensor =
      subgraph->tensors[conv_op->inputs[bias_tensor_index]].get();
  const auto input_tensor =
      subgraph->tensors[conv_op->inputs[input_tensor_idx]].get();
  const auto weights_tensor =
      subgraph->tensors[conv_op->inputs[weights_tensor_idx]].get();
  const auto output_tensor =
      subgraph->tensors[conv_op->outputs[output_tensor_idx]].get();

  EXPECT_EQ(bias_tensor->type, TensorType_INT32);
  EXPECT_EQ(input_tensor->type, TensorType_INT8);
  EXPECT_EQ(weights_tensor->type, TensorType_INT8);

  ASSERT_TRUE(weights_tensor->quantization);
  const int out_channel_size = weights_tensor->shape[0];
  ASSERT_TRUE(bias_tensor->quantization);
  ASSERT_TRUE(weights_tensor->quantization);
  const std::vector<float>& bias_scales = bias_tensor->quantization->scale;
  const std::vector<float>& weights_scales =
      weights_tensor->quantization->scale;

  const std::vector<int64_t>& weights_zero_points =
      weights_tensor->quantization->zero_point;

  ASSERT_EQ(bias_scales.size(), out_channel_size);
  ASSERT_EQ(weights_scales.size(), out_channel_size);
  ASSERT_EQ(weights_zero_points.size(), out_channel_size);
  ASSERT_EQ(input_tensor->quantization->scale.size(), 1);
  ASSERT_EQ(output_tensor->quantization->scale.size(), 1);

  for (size_t i = 0; i < out_channel_size; i++) {
    EXPECT_EQ(weights_scales[i], 1);
    EXPECT_EQ(bias_scales[i], 1);
    EXPECT_EQ(weights_zero_points[i], 0);
  }

  EXPECT_EQ(input_tensor->quantization->scale[0], 1);
  EXPECT_EQ(output_tensor->quantization->scale[0], 1);

  const auto bias_buffer = model_.buffers[bias_tensor->buffer].get();
  ASSERT_EQ(bias_buffer->data.size(), sizeof(int32_t) * bias_tensor->shape[0]);
  const int32_t* bias_values =
      reinterpret_cast<int32_t*>(bias_buffer->data.data());
  const auto original_bias_buffer =
      readonly_model_->buffers()->Get(bias_tensor->buffer);
  const float* bias_float_buffer =
      reinterpret_cast<const float*>(original_bias_buffer->data()->data());

  const float eps = 1e-7;
  for (size_t i = 0; i < bias_tensor->shape[0]; i++) {
    const float bias_scale =
        input_tensor->quantization->scale[0] * weights_scales[i];
    auto dequantized_value = bias_values[i] * bias_scale;
    EXPECT_NEAR(dequantized_value, bias_float_buffer[i], eps);
  }

  const auto weights_buffer = model_.buffers[weights_tensor->buffer].get();
  const auto original_weights_buffer =
      readonly_model_->buffers()->Get(weights_tensor->buffer);
  const int8_t* weight_values =
      reinterpret_cast<int8_t*>(weights_buffer->data.data());
  const float* weights_float_buffer =
      reinterpret_cast<const float*>(original_weights_buffer->data()->data());
  ASSERT_EQ(sizeof(float) * weights_buffer->data.size(),
            original_weights_buffer->data()->size());
  int num_values_in_channel = weights_buffer->data.size() / out_channel_size;
  for (size_t channel_idx = 0; channel_idx < out_channel_size; channel_idx++) {
    for (size_t j = 0; j < num_values_in_channel; j++) {
      size_t element_idx = channel_idx * out_channel_size + j;
      auto dequantized_value =
          weight_values[element_idx] * weights_scales[channel_idx];
      EXPECT_NEAR(dequantized_value, weights_float_buffer[element_idx], eps);
    }
  }

  // check op and versioning.
  EXPECT_EQ(model_.operator_codes.size(), 1);
  EXPECT_EQ(model_.operator_codes[0]->builtin_code, BuiltinOperator_CONV_2D);
  EXPECT_EQ(model_.operator_codes[0]->version, 3);
}

class QuantizeConvModel2Test : public QuantizeModelTest {
 protected:
  QuantizeConvModel2Test() {
    input_model_ = ReadModel(internal::kConvModelWith0Plus10Weights);
    readonly_model_ = input_model_->GetModel();
    readonly_model_->UnPackTo(&model_);
  }
};

TEST_F(QuantizeConvModel2Test, VerifyConvQuantization) {
  auto status = QuantizeModel(&builder_, &model_, TensorType_INT8,
                              TensorType_INT8, &error_reporter_);
  ASSERT_EQ(kTfLiteOk, status);
  const auto& subgraph = model_.subgraphs[0];
  auto conv_op = subgraph->operators[0].get();
  const int input_tensor_idx = 0;
  const int weights_tensor_idx = 1;
  const int bias_tensor_index = 2;
  const int output_tensor_idx = 0;
  const auto bias_tensor =
      subgraph->tensors[conv_op->inputs[bias_tensor_index]].get();
  const auto input_tensor =
      subgraph->tensors[conv_op->inputs[input_tensor_idx]].get();
  const auto weights_tensor =
      subgraph->tensors[conv_op->inputs[weights_tensor_idx]].get();
  const auto output_tensor =
      subgraph->tensors[conv_op->outputs[output_tensor_idx]].get();

  EXPECT_EQ(bias_tensor->type, TensorType_INT32);
  EXPECT_EQ(input_tensor->type, TensorType_INT8);
  EXPECT_EQ(weights_tensor->type, TensorType_INT8);

  ASSERT_TRUE(weights_tensor->quantization);
  const int out_channel_size = weights_tensor->shape[0];
  ASSERT_TRUE(bias_tensor->quantization);
  ASSERT_TRUE(weights_tensor->quantization);
  const std::vector<float>& bias_scales = bias_tensor->quantization->scale;
  const std::vector<float>& weights_scales =
      weights_tensor->quantization->scale;
  const std::vector<int64_t>& weights_zero_points =
      weights_tensor->quantization->zero_point;

  ASSERT_EQ(bias_scales.size(), out_channel_size);
  ASSERT_EQ(weights_scales.size(), out_channel_size);
  ASSERT_EQ(weights_zero_points.size(), out_channel_size);
  ASSERT_EQ(input_tensor->quantization->scale.size(), 1);
  ASSERT_EQ(output_tensor->quantization->scale.size(), 1);

  const float eps = 1e-7;

  // Bias scale should be input * per_channel_weight_scale.
  for (size_t i = 0; i < out_channel_size; i++) {
    EXPECT_NEAR(bias_scales[i],
                input_tensor->quantization->scale[0] * weights_scales[i], eps);
  }

  const auto bias_buffer = model_.buffers[bias_tensor->buffer].get();
  ASSERT_EQ(bias_buffer->data.size(), sizeof(int32_t) * bias_tensor->shape[0]);
  const int32_t* bias_values =
      reinterpret_cast<int32_t*>(bias_buffer->data.data());
  const auto original_bias_buffer =
      readonly_model_->buffers()->Get(bias_tensor->buffer);
  const float* bias_float_buffer =
      reinterpret_cast<const float*>(original_bias_buffer->data()->data());

  for (size_t i = 0; i < out_channel_size; i++) {
    auto dequantized_value = bias_values[i] * bias_scales[i];
    EXPECT_NEAR(dequantized_value, bias_float_buffer[i], bias_scales[i] / 2);
  }

  const auto weights_buffer = model_.buffers[weights_tensor->buffer].get();
  const auto original_weights_buffer =
      readonly_model_->buffers()->Get(weights_tensor->buffer);
  const int8_t* weight_values =
      reinterpret_cast<int8_t*>(weights_buffer->data.data());
  const float* weights_float_buffer =
      reinterpret_cast<const float*>(original_weights_buffer->data()->data());
  ASSERT_EQ(sizeof(float) * weights_buffer->data.size(),
            original_weights_buffer->data()->size());
  int num_values_in_channel = weights_buffer->data.size() / out_channel_size;
  for (size_t channel_idx = 0; channel_idx < out_channel_size; channel_idx++) {
    for (size_t j = 0; j < num_values_in_channel; j++) {
      size_t element_idx = channel_idx * out_channel_size + j;
      auto scale = weights_scales[channel_idx];
      auto zero_point = weights_zero_points[channel_idx];
      auto dequantized_value = weight_values[element_idx] * scale;
      EXPECT_NEAR(dequantized_value, weights_float_buffer[element_idx],
                  scale / 2);
      EXPECT_EQ(zero_point, 0);
    }
  }

  // check op and versioning.
  EXPECT_EQ(model_.operator_codes.size(), 1);
  EXPECT_EQ(model_.operator_codes[0]->builtin_code, BuiltinOperator_CONV_2D);
  EXPECT_EQ(model_.operator_codes[0]->version, 3);
}

class QuantizeSoftmaxTest : public QuantizeModelTest {
 protected:
  QuantizeSoftmaxTest() {
    input_model_ = ReadModel(internal::kSingleSoftmaxModelMinMinus5MaxPlus5);
    readonly_model_ = input_model_->GetModel();
    readonly_model_->UnPackTo(&model_);
  }
};

TEST_F(QuantizeSoftmaxTest, VerifySoftmaxQuantization) {
  auto status = QuantizeModel(&builder_, &model_, TensorType_INT8,
                              TensorType_INT8, &error_reporter_);
  ASSERT_EQ(kTfLiteOk, status);

  const auto& subgraph = model_.subgraphs[0];
  auto op = subgraph->operators[0].get();
  // Model has a single softmax op.
  ASSERT_EQ(op->opcode_index, 0);
  ASSERT_EQ(model_.operator_codes[0].get()->builtin_code,
            BuiltinOperator_SOFTMAX);

  ASSERT_EQ(op->inputs.size(), 1);
  ASSERT_EQ(op->outputs.size(), 1);
  auto float_graph = readonly_model_->subgraphs()->Get(0);

  // Verify input.
  ASSERT_EQ(float_graph->tensors()->Get(op->inputs[0])->type(),
            TensorType_FLOAT32);
  ASSERT_EQ(float_graph->tensors()->Get(op->outputs[0])->type(),
            TensorType_FLOAT32);

  EXPECT_EQ(subgraph->tensors[op->inputs[0]].get()->type, TensorType_INT8);
  EXPECT_EQ(subgraph->tensors[op->outputs[0]].get()->type, TensorType_INT8);

  auto float_input_quant_params =
      float_graph->tensors()->Get(op->inputs[0])->quantization();
  auto input_quant_params =
      subgraph->tensors[op->inputs[0]]->quantization.get();
  VerifyAsymmetricQuantizationScale(*float_input_quant_params,
                                    *input_quant_params);

  // Verify output.
  auto float_output_quant_params =
      float_graph->tensors()->Get(op->outputs[0])->quantization();
  auto output_quant_params =
      subgraph->tensors[op->outputs[0]]->quantization.get();
  ASSERT_EQ(float_output_quant_params->min()->size(), 1);
  ASSERT_EQ(float_output_quant_params->max()->size(), 1);

  ASSERT_EQ(output_quant_params->scale.size(), 1);
  ASSERT_EQ(output_quant_params->zero_point.size(), 1);
  ASSERT_EQ(1.0f / 256.0f, output_quant_params->scale[0]);
  ASSERT_EQ(-128, output_quant_params->zero_point[0]);

  // check op and versioning.
  EXPECT_EQ(model_.operator_codes.size(), 1);
  EXPECT_EQ(model_.operator_codes[0]->builtin_code, BuiltinOperator_SOFTMAX);
  EXPECT_EQ(model_.operator_codes[0]->version, 2);
}

class QuantizeAvgPoolTest : public QuantizeModelTest {
 protected:
  QuantizeAvgPoolTest() {
    input_model_ = ReadModel(internal::kSingleAvgPoolModelMinMinus5MaxPlus5);
    readonly_model_ = input_model_->GetModel();
    readonly_model_->UnPackTo(&model_);
  }
};

TEST_F(QuantizeAvgPoolTest, VerifyAvgPoolQuantization) {
  auto status = QuantizeModel(&builder_, &model_, TensorType_INT8,
                              TensorType_INT8, &error_reporter_);
  ASSERT_EQ(kTfLiteOk, status);

  const auto& subgraph = model_.subgraphs[0];
  auto op = subgraph->operators[0].get();
  // Model has a single AveragePool op.
  ASSERT_EQ(op->opcode_index, 0);
  ASSERT_EQ(model_.operator_codes[0].get()->builtin_code,
            BuiltinOperator_AVERAGE_POOL_2D);

  ASSERT_EQ(op->inputs.size(), 1);
  ASSERT_EQ(op->outputs.size(), 1);

  auto float_graph = readonly_model_->subgraphs()->Get(0);
  ASSERT_EQ(float_graph->tensors()->Get(op->inputs[0])->type(),
            TensorType_FLOAT32);
  ASSERT_EQ(float_graph->tensors()->Get(op->outputs[0])->type(),
            TensorType_FLOAT32);

  EXPECT_EQ(subgraph->tensors[op->inputs[0]].get()->type, TensorType_INT8);
  EXPECT_EQ(subgraph->tensors[op->outputs[0]].get()->type, TensorType_INT8);

  auto float_input_quant_params =
      float_graph->tensors()->Get(op->inputs[0])->quantization();
  auto input_quant_params =
      subgraph->tensors[op->inputs[0]]->quantization.get();
  VerifyAsymmetricQuantizationScale(*float_input_quant_params,
                                    *input_quant_params);

  auto float_output_quant_params =
      float_graph->tensors()->Get(op->outputs[0])->quantization();
  auto output_quant_params =
      subgraph->tensors[op->outputs[0]]->quantization.get();
  ASSERT_EQ(float_output_quant_params->min()->size(), 1);
  ASSERT_EQ(float_output_quant_params->max()->size(), 1);
  ASSERT_EQ(output_quant_params->min.size(), 1);
  ASSERT_EQ(output_quant_params->max.size(), 1);

  // Make sure the input min/maxes are propagated to outputs.
  EXPECT_EQ(input_quant_params->min[0], output_quant_params->min[0]);
  EXPECT_EQ(input_quant_params->max[0], output_quant_params->max[0]);
  EXPECT_EQ(input_quant_params->scale[0], output_quant_params->scale[0]);

  // check op and versioning.
  EXPECT_EQ(model_.operator_codes.size(), 1);
  EXPECT_EQ(model_.operator_codes[0]->builtin_code,
            BuiltinOperator_AVERAGE_POOL_2D);
  EXPECT_EQ(model_.operator_codes[0]->version, 2);
}

class QuantizeMultiInputAddWithReshapeTest : public QuantizeModelTest {
 protected:
  QuantizeMultiInputAddWithReshapeTest() {
    input_model_ = ReadModel(internal::kMultiInputAddWithReshape);
    readonly_model_ = input_model_->GetModel();
    readonly_model_->UnPackTo(&model_);
  }
};

TEST_F(QuantizeMultiInputAddWithReshapeTest, VerifyReshapeQuantization) {
  auto status = QuantizeModel(&builder_, &model_, TensorType_INT8,
                              TensorType_INT8, &error_reporter_);
  ASSERT_EQ(kTfLiteOk, status);

  // Verify Reshape is quantized.
  const auto& subgraph = model_.subgraphs[0];
  auto op = subgraph->operators[1].get();
  ASSERT_EQ(model_.operator_codes[op->opcode_index].get()->builtin_code,
            BuiltinOperator_RESHAPE);

  ASSERT_EQ(op->inputs.size(), 2);
  ASSERT_EQ(op->outputs.size(), 1);

  auto float_graph = readonly_model_->subgraphs()->Get(0);
  ASSERT_EQ(float_graph->tensors()->Get(op->inputs[0])->type(),
            TensorType_FLOAT32);
  ASSERT_EQ(float_graph->tensors()->Get(op->outputs[0])->type(),
            TensorType_FLOAT32);

  EXPECT_EQ(subgraph->tensors[op->inputs[0]].get()->type, TensorType_INT8);
  EXPECT_EQ(subgraph->tensors[op->outputs[0]].get()->type, TensorType_INT8);

  auto float_input_quant_params =
      float_graph->tensors()->Get(op->inputs[0])->quantization();
  auto input_quant_params =
      subgraph->tensors[op->inputs[0]]->quantization.get();
  VerifyAsymmetricQuantizationScale(*float_input_quant_params,
                                    *input_quant_params);

  auto float_output_quant_params =
      float_graph->tensors()->Get(op->outputs[0])->quantization();
  auto output_quant_params =
      subgraph->tensors[op->outputs[0]]->quantization.get();
  ASSERT_EQ(float_output_quant_params->min()->size(), 1);
  ASSERT_EQ(float_output_quant_params->max()->size(), 1);
  ASSERT_EQ(output_quant_params->min.size(), 1);
  ASSERT_EQ(output_quant_params->max.size(), 1);

  // check op and versioning.
  EXPECT_EQ(model_.operator_codes.size(), 2);
  EXPECT_EQ(model_.operator_codes[0]->builtin_code, BuiltinOperator_ADD);
  EXPECT_EQ(model_.operator_codes[0]->version, 2);
  EXPECT_EQ(model_.operator_codes[1]->builtin_code, BuiltinOperator_RESHAPE);
  EXPECT_EQ(model_.operator_codes[1]->version, 1);
}

TEST_F(QuantizeMultiInputAddWithReshapeTest, VerifyAddQuantization) {
  auto status = QuantizeModel(&builder_, &model_, TensorType_INT8,
                              TensorType_INT8, &error_reporter_);
  ASSERT_EQ(kTfLiteOk, status);

  // Verify ADD is quantized.
  const auto& subgraph = model_.subgraphs[0];
  auto op = subgraph->operators[0].get();
  ASSERT_EQ(model_.operator_codes[op->opcode_index].get()->builtin_code,
            BuiltinOperator_ADD);

  ASSERT_EQ(op->inputs.size(), 2);
  ASSERT_EQ(op->outputs.size(), 1);

  auto float_graph = readonly_model_->subgraphs()->Get(0);
  ASSERT_EQ(float_graph->tensors()->Get(op->inputs[0])->type(),
            TensorType_FLOAT32);
  ASSERT_EQ(float_graph->tensors()->Get(op->inputs[1])->type(),
            TensorType_FLOAT32);
  ASSERT_EQ(float_graph->tensors()->Get(op->outputs[0])->type(),
            TensorType_FLOAT32);

  for (size_t input_idx = 0; input_idx < 2; ++input_idx) {
    EXPECT_EQ(subgraph->tensors[op->inputs[input_idx]].get()->type,
              TensorType_INT8);
    auto float_input_quant_params =
        float_graph->tensors()->Get(op->inputs[input_idx])->quantization();
    auto input_quant_params =
        subgraph->tensors[op->inputs[input_idx]]->quantization.get();
    VerifyAsymmetricQuantizationScale(*float_input_quant_params,
                                      *input_quant_params);
  }

  EXPECT_EQ(subgraph->tensors[op->outputs[0]].get()->type, TensorType_INT8);
  auto float_output_quant_params =
      float_graph->tensors()->Get(op->outputs[0])->quantization();
  auto output_quant_params =
      subgraph->tensors[op->outputs[0]]->quantization.get();
  ASSERT_EQ(float_output_quant_params->min()->size(), 1);
  ASSERT_EQ(float_output_quant_params->max()->size(), 1);
  ASSERT_EQ(output_quant_params->min.size(), 1);
  ASSERT_EQ(output_quant_params->max.size(), 1);

  // check op and versioning.
  EXPECT_EQ(model_.operator_codes.size(), 2);
  EXPECT_EQ(model_.operator_codes[0]->builtin_code, BuiltinOperator_ADD);
  EXPECT_EQ(model_.operator_codes[0]->version, 2);
  EXPECT_EQ(model_.operator_codes[1]->builtin_code, BuiltinOperator_RESHAPE);
  EXPECT_EQ(model_.operator_codes[1]->version, 1);
}

class QuantizeConstInputTest : public QuantizeModelTest {
 protected:
  QuantizeConstInputTest() {
    input_model_ = ReadModel(internal::kConstInputAddModel);
    readonly_model_ = input_model_->GetModel();
    readonly_model_->UnPackTo(&model_);
  }
};

TEST_F(QuantizeConstInputTest, VerifyConstOpInput) {
  auto status = QuantizeModel(&builder_, &model_, TensorType_INT8,
                              TensorType_INT8, &error_reporter_);
  ASSERT_EQ(kTfLiteOk, status);

  // Verify ConstOp is quantized.
  const auto& subgraph = model_.subgraphs[0];
  auto op = subgraph->operators[0].get();
  ASSERT_EQ(model_.operator_codes[op->opcode_index].get()->builtin_code,
            BuiltinOperator_ADD);

  ASSERT_EQ(op->inputs.size(), 2);
  ASSERT_EQ(op->outputs.size(), 1);

  auto float_graph = readonly_model_->subgraphs()->Get(0);
  ASSERT_EQ(float_graph->tensors()->Get(op->inputs[0])->type(),
            TensorType_FLOAT32);
  ASSERT_EQ(float_graph->tensors()->Get(op->outputs[0])->type(),
            TensorType_FLOAT32);

  for (size_t input_idx = 0; input_idx < 2; ++input_idx) {
    EXPECT_EQ(subgraph->tensors[op->inputs[input_idx]].get()->type,
              TensorType_INT8);
  }

  EXPECT_EQ(subgraph->tensors[op->outputs[0]].get()->type, TensorType_INT8);

  // check op and versioning.
  EXPECT_EQ(model_.operator_codes.size(), 1);
  EXPECT_EQ(model_.operator_codes[0]->builtin_code, BuiltinOperator_ADD);
  EXPECT_EQ(model_.operator_codes[0]->version, 2);
}

class QuantizeArgMaxTest : public QuantizeModelTest {
 protected:
  QuantizeArgMaxTest() {
    input_model_ = ReadModel(internal::kModelWithArgMaxOp);
    readonly_model_ = input_model_->GetModel();
    readonly_model_->UnPackTo(&model_);
  }
};

TEST_F(QuantizeArgMaxTest, VerifyArgMax) {
  auto status = QuantizeModel(&builder_, &model_, TensorType_INT8,
                              TensorType_INT8, &error_reporter_);
  ASSERT_EQ(kTfLiteOk, status);

  const auto& subgraph = model_.subgraphs[0];
  auto op = subgraph->operators[0].get();
  ASSERT_EQ(model_.operator_codes[op->opcode_index].get()->builtin_code,
            BuiltinOperator_ARG_MAX);

  ASSERT_EQ(op->inputs.size(), 2);
  ASSERT_EQ(op->outputs.size(), 1);

  auto float_graph = readonly_model_->subgraphs()->Get(0);
  // Verify ArgMax input is quantized.
  ASSERT_EQ(float_graph->tensors()->Get(op->inputs[0])->type(),
            TensorType_FLOAT32);
  EXPECT_EQ(subgraph->tensors[op->inputs[0]].get()->type, TensorType_INT8);

  // Verify ArgMax input axis should still be the same type.
  ASSERT_EQ(float_graph->tensors()->Get(op->inputs[1])->type(),
            subgraph->tensors[op->inputs[1]].get()->type);

  // The output of ArgMax should still be the same type.
  ASSERT_EQ(float_graph->tensors()->Get(op->outputs[0])->type(),
            subgraph->tensors[op->outputs[0]].get()->type);

  // check op and versioning.
  EXPECT_EQ(model_.operator_codes.size(), 1);
  EXPECT_EQ(model_.operator_codes[0]->builtin_code, BuiltinOperator_ARG_MAX);
  EXPECT_EQ(model_.operator_codes[0]->version, 2);
}

class QuantizeFCTest : public QuantizeModelTest {
 protected:
  QuantizeFCTest() {
    input_model_ = ReadModel(internal::kModelWithFCOp);
    readonly_model_ = input_model_->GetModel();
    readonly_model_->UnPackTo(&model_);
  }
};

TEST_F(QuantizeFCTest, VerifyFC) {
  auto status = QuantizeModel(&builder_, &model_, TensorType_INT8,
                              TensorType_INT8, &error_reporter_);
  ASSERT_EQ(kTfLiteOk, status);

  const auto& subgraph = model_.subgraphs[0];
  auto op = subgraph->operators[0].get();
  ASSERT_EQ(model_.operator_codes[op->opcode_index].get()->builtin_code,
            BuiltinOperator_FULLY_CONNECTED);

  ASSERT_EQ(op->inputs.size(), 3);
  ASSERT_EQ(op->outputs.size(), 1);

  auto float_graph = readonly_model_->subgraphs()->Get(0);
  // Verify FC input and weight is quantized.
  ASSERT_EQ(float_graph->tensors()->Get(op->inputs[0])->type(),
            TensorType_FLOAT32);
  EXPECT_EQ(subgraph->tensors[op->inputs[0]].get()->type, TensorType_INT8);
  ASSERT_EQ(float_graph->tensors()->Get(op->inputs[1])->type(),
            TensorType_FLOAT32);
  EXPECT_EQ(subgraph->tensors[op->inputs[1]].get()->type, TensorType_INT8);

  // Verify FC bias should be int32 quantized.
  ASSERT_EQ(float_graph->tensors()->Get(op->inputs[2])->type(),
            TensorType_FLOAT32);
  EXPECT_EQ(subgraph->tensors[op->inputs[2]].get()->type, TensorType_INT32);

  // The output of FC should be quantized.
  ASSERT_EQ(float_graph->tensors()->Get(op->outputs[0])->type(),
            TensorType_FLOAT32);
  EXPECT_EQ(subgraph->tensors[op->outputs[0]].get()->type, TensorType_INT8);

  // check op and versioning.
  EXPECT_EQ(model_.operator_codes.size(), 2);
  EXPECT_EQ(model_.operator_codes[0]->builtin_code,
            BuiltinOperator_FULLY_CONNECTED);
  EXPECT_EQ(model_.operator_codes[0]->version, 4);
  EXPECT_EQ(model_.operator_codes[1]->builtin_code, BuiltinOperator_RESHAPE);
  EXPECT_EQ(model_.operator_codes[1]->version, 1);
}

class QuantizeCustomOpTest : public QuantizeModelTest {
 protected:
  QuantizeCustomOpTest() {
    input_model_ = ReadModel(internal::kModelMixed);
    readonly_model_ = input_model_->GetModel();
    readonly_model_->UnPackTo(&model_);
  }
};

TEST_F(QuantizeCustomOpTest, VerifyMixedQuantization) {
  auto status =
      QuantizeModel(&builder_, &model_, TensorType_INT8, TensorType_INT8,
                    /*allow_float=*/true, &error_reporter_);
  ASSERT_EQ(kTfLiteOk, status);
  const auto& subgraph = model_.subgraphs[0];
  auto float_graph = readonly_model_->subgraphs()->Get(0);
  // The original model reshape->custom->custom->squeeze.
  ASSERT_EQ(float_graph->operators()->size(), 4);
  // The resulting model should be:
  // reshape->dequantize->custom->custom->quantize->squeeze.
  ASSERT_EQ(subgraph->operators.size(), 6);
  const std::vector<BuiltinOperator> op_codes = {
      BuiltinOperator_RESHAPE,  BuiltinOperator_DEQUANTIZE,
      BuiltinOperator_CUSTOM,   BuiltinOperator_CUSTOM,
      BuiltinOperator_QUANTIZE, BuiltinOperator_SQUEEZE};
  const std::vector<TensorType> op_input_types = {
      TensorType_INT8,    TensorType_INT8,    TensorType_FLOAT32,
      TensorType_FLOAT32, TensorType_FLOAT32, TensorType_INT8};
  for (int i = 0; i < subgraph->operators.size(); ++i) {
    OperatorT* op = subgraph->operators[i].get();
    ASSERT_EQ(model_.operator_codes[op->opcode_index]->builtin_code,
              op_codes[i]);
    ASSERT_EQ(subgraph->tensors[op->inputs[0]]->type, op_input_types[i]);
  }
}

}  // namespace
}  // namespace optimize
}  // namespace tflite

int main(int argc, char** argv) {
  tensorflow::string model_file;
  const std::vector<tensorflow::Flag> flag_list = {
      tensorflow::Flag("test_model_file", &model_file,
                       "Path to test tflite model file."),
  };

  const bool parse_result = tensorflow::Flags::Parse(&argc, argv, flag_list);
  if (!parse_result) {
    std::cerr << "Required test_model_file\n";
    std::abort();
  }
  g_test_model_dir =
      new tensorflow::string(tensorflow::io::Dirname(model_file));
  ::tensorflow::port::InitMain(argv[0], &argc, &argv);
  return RUN_ALL_TESTS();
}