高考报名时间2022年11月1日还剩 天 距离2023年6月7日高考还有 天
栏目分类:
子分类:
返回
名师互学网用户登录
名师互学网
快速导航关闭
当前搜索
当前分类
前沿技术 软件开发 系统运维 产品运营 生活办公 面试经验 考试题库
子分类
人工智能 大数据 云计算 区块链 物联网 深度学习 机器学习 NLP 计算机视觉 语音识别 其他 大数据系统 数据可视化 数据挖掘与分析 其他 Docker/k8s 虚拟化 云平台 其他 基本原理 数字货币 智能合约 EOS应用 其他 通讯技术 嵌入式开发 单片机 物联网应用 HarmonyOS 其他 后端开发 Web开发 移动开发 游戏开发 Python Java 架构设计 C/C++/C# PHP .Net Go语言 R语言
实用工具
学习工具 小学数学练习 字帖生成 在线画板 函数绘制 拼音字母表 在线词典 黄历查询 亲戚关系计算 安全期计算 中国历史 Excel函数 模拟请求 json格式化 浏览器指纹
热门搜索
路由器设置 木托盘 宝塔面板 儿童python教程 心情低落 朋友圈 vim 双一流学科 专升本 我的学校 日记学校 西点培训学校 汽修学校 情书 化妆学校 塔沟武校 异形模板 西南大学排名 最精辟人生短句 6步教你追回被骗的钱 南昌大学排名 清朝十二帝 北京印刷学院排名 北方工业大学排名 北京航空航天大学排名 首都经济贸易大学排名 中国传媒大学排名 首都师范大学排名 中国地质大学(北京)排名 北京信息科技大学排名
名师互学网 > IT > 软件开发 > 后端开发 > Python

Tensorflowlite 部署到 arm开发板

Python 更新时间： 发布时间： IT归档 最新发布 模块sitemap 名妆网 法律咨询 聚返吧 英语巴士网 伯小乐 网商动力

Tensorflowlite 部署到 arm开发板

Tensorflowlite 部署到 arm开发板
  • 一 先在本机上操作
    • 1 下载TensorFlow
      • 下载依赖
    • 2 准备ARM的交叉编译环境
      • 2.1 下载安装包
      • 2.2 解压安装包
      • 2.3 配置环境变量
      • 2.4 查看编译器版本
    • 3 交叉编译生成静态库
    • 4 官方Demo ：label_image的编译
      • 4.1 整理头文件
      • 4.2 cmake ,make 进行编译
  • 二 在开发板上操作

一 先在本机上操作

本机：Ubuntu18.04

1 下载TensorFlow

我选择的版本是：tensorflow-2.6.0
下载网址：

https://github.com/tensorflow/tensorflow
点master,选择tag，v2.6.0

(注：我是直接下载文件夹，进行解压。也可以 git clone,由于网速不行，clone不下来，就直接下载文件了，结果都一样）

下载依赖 
#切换目录，切换到解压的tensorflow-2.6.0目录
cd  tensorflow-2.6.0
#执行download_dependencies.sh脚本，下载依赖，下载时，资源可能因网速下载失败，多试几次，增量下载。或者，哪一个下载不下来，会提示网址，手动下载，然后在sh脚本里将其下载代码注释掉即可。
./tensorflow/lite/tools/make/download_dependencies.sh

上述命令，会在make目录下生成downloads目录，目录下依赖目录如下：

在make目录下：ls

absl     eigen     fft2d        fp16      googletest  ruy
cpuinfo  farmhash  flatbuffers  gemmlowp  neon_2_sse
2 准备ARM的交叉编译环境

交叉编译选择的是：arm-linux-gnueabihf-gcc 7.5.0

2.1 下载安装包

位置：

https://releases.linaro.org/components/toolchain/binaries/7.5-2019.12/arm-linux-gnueabihf/

gcc-linaro-7.5.0-2019.12-x86_64_arm-linux-gnueabihf.tar.xz

2.2 解压安装包 2.3 配置环境变量

这里本文配置用户私有环境变量

vim .bashrc
#末尾添加：
export PATH=$PATH:/解压位置/gcc-linaro-arm-linux-gnueabihf-4.7-2013.03-20130313_linux/bin
#配置生效
source .bashrc
2.4 查看编译器版本

直接在命令窗口输入命令：

arm-linux-gnueabihf-gcc --version
3 交叉编译生成静态库 
#所在目录:tensorflow/lite/tools/make
#执行下面脚本
./build_rpi_lib.sh

主要目的：生成libtensorflow-lite.a静态库，交叉编译的时候用到
结果展示：

#所在目录:tensorflow/lite/tools/make
cd cd gen/rpi_armv7l
ls
bin  lib  obj
cd lib
benchmark-lib.a  libtensorflow-lite.a
4 官方Demo ：label_image的编译 4.1 整理头文件 
cd tensorflow-2.6.0/tensorflow/lite
find  -name "*.h" | tar -cf headers.tar -T -#这一步，是将lite头文件压缩打包

#找另外任意一个地方（离开tensorflow-2.6.0目录），新建一个目录，将我们需要进行编译相关的文件放在一起
mkdir test
cd test
mkdir -p include/tensorflow/lite#这一步
tar xvf headers.tar -C include/tensorflow/lite#将前面压缩打包的文件拷贝过来，解压

将absl flatbuffers gmock gtest tensorflow头文件也移动include目录下：
演示：flatbuffers，其他同理

#在test目录下操作
mkdir -p include/flatbuffers
cp tensorflow-2.6.0/tensorflow/lite/tools/make/downloads/flatbuffers/include/flatbuffers 

#include "tensorflow/lite/examples/label_image/label_image.h"

#include       // NOLINT(build/include_order)
#include      // NOLINT(build/include_order)
#include    // NOLINT(build/include_order)
#include   // NOLINT(build/include_order)
#include     // NOLINT(build/include_order)
#include      // NOLINT(build/include_order)

#include "tensorflow/lite/examples/label_image/label_image.h"

#include       // NOLINT(build/include_order)
#include      // NOLINT(build/include_order)
#include    // NOLINT(build/include_order)
#include   // NOLINT(build/include_order)
#include     // NOLINT(build/include_order)
#include      // NOLINT(build/include_order)

#include 
#include 
#include 
#include 
#include 
#include 
#include 
#include 
#include 
#include 
#include 
#include 

#include "tensorflow/lite/tools/make/downloads/absl/absl/memory/memory.h"
#include "tensorflow/lite/examples/label_image/bitmap_helpers.h"
#include "tensorflow/lite/examples/label_image/get_top_n.h"
#include "tensorflow/lite/examples/label_image/log.h"
#include "tensorflow/lite/kernels/register.h"
#include "tensorflow/lite/optional_debug_tools.h"
#include "tensorflow/lite/profiling/profiler.h"
#include "tensorflow/lite/string_util.h"
#include "tensorflow/lite/tools/command_line_flags.h"
#include "tensorflow/lite/tools/delegates/delegate_provider.h"

namespace tflite {
namespace label_image {

double get_us(struct timeval t) { return (t.tv_sec * 1000000 + t.tv_usec); }

using TfLiteDelegatePtr = tflite::Interpreter::TfLiteDelegatePtr;
using ProvidedDelegateList = tflite::tools::ProvidedDelegateList;
#if 0
template 
void resize(T* out, uint8_t* in, int image_height, int image_width,
            int image_channels, int wanted_height, int wanted_width,
            int wanted_channels, Settings* s) {
  cout<< "11111111111"< interpreter(new Interpreter);

  int base_index = 0;

  LOG(INFO)<< "11111111111"

  // two inputs: input and new_sizes
  interpreter->AddTensors(2, &base_index);
  // one output
  interpreter->AddTensors(1, &base_index);
  // set input and output tensors
  interpreter->SetInputs({0, 1});
  interpreter->SetOutputs({2});

  LOG(INFO)<< "11111111111"
  // set parameters of tensors
  TfLiteQuantizationParams quant;
  interpreter->SetTensorParametersReadWrite(
      0, kTfLiteFloat32, "input",
      {1, image_height, image_width, image_channels}, quant);
  interpreter->SetTensorParametersReadWrite(1, kTfLiteInt32, "new_size", {2},
                                            quant);
  LOG(INFO)<< "11111111111"
  interpreter->SetTensorParametersReadWrite(
      2, kTfLiteFloat32, "output",
      {1, wanted_height, wanted_width, wanted_channels}, quant);

  LOG(INFO)<< "11111111111"
  ops::builtin::BuiltinOpResolver resolver;
  const TfLiteRegistration* resize_op =
      resolver.FindOp(BuiltinOperator_RESIZE_BILINEAR, 1);
  auto* params = reinterpret_cast(
      malloc(sizeof(TfLiteResizeBilinearParams)));
  params->align_corners = false;
  params->half_pixel_centers = false;
  interpreter->AddNodeWithParameters({0, 1}, {2}, nullptr, 0, params, resize_op,
                                     nullptr);
  LOG(INFO)<< "interpreter->AllocateTensors()"

  interpreter->AllocateTensors();

  // fill input image
  // in[] are integers, cannot do memcpy() directly
  auto input = interpreter->typed_tensor(0);
  for (int i = 0; i < number_of_pixels; i++) {
    input[i] = in[i];
  }

  // fill new_sizes
  interpreter->typed_tensor(1)[0] = wanted_height;
  interpreter->typed_tensor(1)[1] = wanted_width;
  LOG(INFO)<< "interpreter->Invoke()"
  interpreter->Invoke();

  auto output = interpreter->typed_tensor(2);
  auto output_number_of_pixels = wanted_height * wanted_width * wanted_channels;

  for (int i = 0; i < output_number_of_pixels; i++) {
    switch (s->input_type) {
      case kTfLiteFloat32:
        out[i] = (output[i] - s->input_mean) / s->input_std;
        break;
      case kTfLiteInt8:
        out[i] = static_cast(output[i] - 128);
        break;
      case kTfLiteUInt8:
        out[i] = static_cast(output[i]);
        break;
      default:
        break;
    }
  }
}
#endif
// Takes a file name, and loads a list of labels from it, one per line, and
// returns a vector of the strings. It pads with empty strings so the length
// of the result is a multiple of 16, because our model expects that.
TfLiteStatus ReadLabelsFile(const string& file_name,
                            std::vector* result,
                            size_t* found_label_count) {
  std::ifstream file(file_name);
  if (!file) {
    LOG(ERROR) << "Labels file " << file_name << " not found";
    return kTfLiteError;
  }
  result->clear();
  string line;
  while (std::getline(file, line)) {
    result->push_back(line);
  }
  *found_label_count = result->size();
  const int padding = 16;
  while (result->size() % padding) {
    result->emplace_back();
  }
  return kTfLiteOk;
}

void PrintProfilingInfo(const profiling::ProfileEvent* e,
                        uint32_t subgraph_index, uint32_t op_index,
                        TfLiteRegistration registration) {
  // output something like
  // time (ms) , Node xxx, OpCode xxx, symbolic name
  //      5.352, Node   5, OpCode   4, DEPTHWISE_CONV_2D

  LOG(INFO) << std::fixed << std::setw(10) << std::setprecision(3)
            << (e->end_timestamp_us - e->begin_timestamp_us) / 1000.0
            << ", Subgraph " << std::setw(3) << std::setprecision(3)
            << subgraph_index << ", Node " << std::setw(3)
            << std::setprecision(3) << op_index << ", OpCode " << std::setw(3)
            << std::setprecision(3) << registration.builtin_code << ", "
            << EnumNameBuiltinOperator(
                   static_cast(registration.builtin_code));
}

/// 使用的是mobilenet_v2_1.0_224_quant.tflite 量化后模型
void RunInference(Settings* settings)
{
  std::unique_ptr model;
  std::unique_ptr interpreter;
  model = tflite::FlatBufferModel::BuildFromFile(settings->model_name.c_str());
  if (!model) {
    LOG(ERROR) << "Failed to mmap model " << settings->model_name;
    exit(-1);
  }
//  settings->model = model.get();
  LOG(INFO) << "Loaded model " << settings->model_name;
  //model->error_reporter();
  //LOG(INFO) << "resolved reporter";

  tflite::ops::builtin::BuiltinOpResolver resolver;

  tflite::InterpreterBuilder(*model, resolver)(&interpreter);
  if (!interpreter) {
    LOG(ERROR) << "Failed to construct interpreter";
    exit(-1);
  }

  interpreter->SetAllowFp16PrecisionForFp32(settings->allow_fp16);



  if (settings->verbose) {
    LOG(INFO) << "tensors size: " << interpreter->tensors_size();
    LOG(INFO) << "nodes size: " << interpreter->nodes_size();
    LOG(INFO) << "inputs: " << interpreter->inputs().size();
    LOG(INFO) << "input(0) name: " << interpreter->GetInputName(0);

    int t_size = interpreter->tensors_size();
    for (int i = 0; i < t_size; i++) {
      if (interpreter->tensor(i)->name)
        LOG(INFO) << i << ": " << interpreter->tensor(i)->name << ", "
                  << interpreter->tensor(i)->bytes << ", "
                  << interpreter->tensor(i)->type << ", "
                  << interpreter->tensor(i)->params.scale << ", "
                  << interpreter->tensor(i)->params.zero_point;
    }
  }

  if (settings->number_of_threads != -1) {
    interpreter->SetNumThreads(settings->number_of_threads);
  }
  LOG(INFO) << "load picture ";
  LOG(INFO) << "load picture ";
  int image_width = 224;
  int image_height = 224;
  int image_channels = 3;
  std::vector in = read_bmp(settings->input_bmp_name, &image_width,
                                     &image_height, &image_channels, settings);

  int input = interpreter->inputs()[0];
  if (settings->verbose) LOG(INFO) << "input: " << input;

  const std::vector inputs = interpreter->inputs();
  const std::vector outputs = interpreter->outputs();

  if (settings->verbose) {
    LOG(INFO) << "number of inputs: " << inputs.size();
    LOG(INFO) << "number of outputs: " << outputs.size();
  }


  if (interpreter->AllocateTensors() != kTfLiteOk) {
    LOG(ERROR) << "Failed to allocate tensors!";
    exit(-1);
  }



  if (settings->verbose) PrintInterpreterState(interpreter.get());

  // get input dimension from the input tensor metadata
  // assuming one input only
  TfLiteIntArray* dims = interpreter->tensor(input)->dims;
  int wanted_height = dims->data[1];
  int wanted_width = dims->data[2];
  int wanted_channels = dims->data[3];



  settings->input_type = interpreter->tensor(input)->type;
    LOG(INFO) << "wanted width, height, channels: " < interpreter1(new Interpreter);


  int base_index = 0;



  // two inputs: input and new_sizes
  interpreter1->AddTensors(2, &base_index);
  // one output
  interpreter1->AddTensors(1, &base_index);
  // set input and output tensors
  interpreter1->SetInputs({0, 1});
  interpreter1->SetOutputs({2});


    // set parameters of tensors
  TfLiteQuantizationParams quant;
  interpreter1->SetTensorParametersReadWrite(
      0, kTfLiteFloat32, "input",
      {1, image_height, image_width, image_channels}, quant);
  interpreter1->SetTensorParametersReadWrite(1, kTfLiteInt32, "new_size", {2},
                                            quant);

  interpreter1->SetTensorParametersReadWrite(
      2, kTfLiteFloat32, "output",
      {1, wanted_height, wanted_width, wanted_channels}, quant);



  ops::builtin::BuiltinOpResolver resolver1;
  const TfLiteRegistration* resize_op =
      resolver1.FindOp(BuiltinOperator_RESIZE_BILINEAR, 1);
  auto* params = reinterpret_cast(
      malloc(sizeof(TfLiteResizeBilinearParams)));
  params->align_corners = false;
  params->half_pixel_centers = false;
  interpreter1->AddNodeWithParameters({0, 1}, {2}, nullptr, 0, params, resize_op,
                                     nullptr);

  interpreter1->AllocateTensors();

  // fill input image
  // in[] are integers, cannot do memcpy() directly
  auto input1 = interpreter1->typed_tensor(0);
  for (int i = 0; i < number_of_pixels; i++) {
    input1[i] = in.data()[i];
  }

  // fill new_sizes
  interpreter1->typed_tensor(1)[0] = wanted_height;
  interpreter1->typed_tensor(1)[1] = wanted_width;

  interpreter1->Invoke();


  auto outputxx = interpreter1->typed_tensor(2);
  auto output_number_of_pixels = wanted_height * wanted_width * wanted_channels;
  LOG(INFO)<< "output_number_of_pixels@@@@@@@@@@@@@@@@@@@:" << output_number_of_pixels  ;

  	  printf("pinput11^^^^^^^^^^^^^^^^^^^^^^^");


	printf("typed_tensor(0):%pn",interpreter->typed_tensor(0));
    printf("typed_tensor(171):%pn",interpreter->typed_tensor(171));

    auto input11 = interpreter->typed_tensor(171);


	///将修改好尺寸的图像数据输入进模型



    for (int i = 0; i < output_number_of_pixels; i++) {


	    input11[i] = static_cast(outputxx[i]);

  }
  printf("pinput11  赋值完毕^^^^^^^^^^^^^^^^^^^^^^^");
 
#if 1
  auto profiler = absl::make_unique(
      settings->max_profiling_buffer_entries);
  interpreter->SetProfiler(profiler.get());

  if (settings->profiling) profiler->StartProfiling();
  if (settings->loop_count > 1) {
    for (int i = 0; i < settings->number_of_warmup_runs; i++) {
      if (interpreter->Invoke() != kTfLiteOk) {
        LOG(ERROR) << "Failed to invoke tflite!";
        exit(-1);
      }
    }
  }

#endif



  struct timeval start_time, stop_time;
  gettimeofday(&start_time, nullptr);
  for (int i = 0; i < settings->loop_count; i++) {
    if (interpreter->Invoke() != kTfLiteOk) {
      LOG(ERROR) << "Failed to invoke tflite!";
      exit(-1);
    }
  }
  gettimeofday(&stop_time, nullptr);
  LOG(INFO) << "invoked";
  LOG(INFO) << "average time: "
            << (get_us(stop_time) - get_us(start_time)) /
                   (settings->loop_count * 1000)
            << " ms";
#if 1
  if (settings->profiling) {
    profiler->StopProfiling();
    auto profile_events = profiler->GetProfileEvents();
    for (int i = 0; i < profile_events.size(); i++) {
      auto subgraph_index = profile_events[i]->extra_event_metadata;
      auto op_index = profile_events[i]->event_metadata;
      const auto subgraph = interpreter->subgraph(subgraph_index);
      const auto node_and_registration =
          subgraph->node_and_registration(op_index);
      const TfLiteRegistration registration = node_and_registration->second;
      PrintProfilingInfo(profile_events[i], subgraph_index, op_index,
                         registration);
    }
  }
#endif

   const float threshold = 0.001f;

  std::vector> top_results;

  int output = interpreter->outputs()[0];
  TfLiteIntArray* output_dims = interpreter->tensor(output)->dims;
  // assume output dims to be something like (1, 1, ... ,size)
  auto output_size = output_dims->data[output_dims->size - 1];
  switch (interpreter->tensor(output)->type) {
    case kTfLiteFloat32:
      get_top_n(interpreter->typed_output_tensor(0), output_size,
                       settings->number_of_results, threshold, &top_results,
                       settings->input_type);
      break;
    case kTfLiteInt8:
      get_top_n(interpreter->typed_output_tensor(0),
                        output_size, settings->number_of_results, threshold,
                        &top_results, settings->input_type);
      break;
    case kTfLiteUInt8:
      get_top_n(interpreter->typed_output_tensor(0),
                         output_size, settings->number_of_results, threshold,
                         &top_results, settings->input_type);
      break;
    default:
      LOG(ERROR) << "cannot handle output type "
                 << interpreter->tensor(output)->type << " yet";
      exit(-1);
  }

  std::vector labels;
  size_t label_count;

  if (ReadLabelsFile(settings->labels_file_name, &labels, &label_count) !=
      kTfLiteOk)
    exit(-1);

  for (const auto& result : top_results) {
    const float confidence = result.first;
    const int index = result.second;
    LOG(INFO) << confidence << ": " << index << " " << labels[index];
  }
}

void display_usage() {
  LOG(INFO)
      << "label_imagen"
      << "--accelerated, -a: [0|1], use Android NNAPI or notn"
      << "--allow_fp16, -f: [0|1], allow running fp32 models with fp16 or notn"
      << "--count, -c: loop interpreter->Invoke() for certain timesn"
      << "--gl_backend, -g: [0|1]: use GL GPU Delegate on Androidn"
      << "--hexagon_delegate, -j: [0|1]: use Hexagon Delegate on Androidn"
      << "--input_mean, -b: input meann"
      << "--input_std, -s: input standard deviationn"
      << "--image, -i: image_name.bmpn"
      << "--labels, -l: labels for the modeln"
      << "--tflite_model, -m: model_name.tfliten"
      << "--profiling, -p: [0|1], profiling or notn"
      << "--num_results, -r: number of results to shown"
      << "--threads, -t: number of threadsn"
      << "--verbose, -v: [0|1] print more informationn"
      << "--warmup_runs, -w: number of warmup runsn"
      << "--xnnpack_delegate, -x [0:1]: xnnpack delegaten";
}

int Main(int argc, char** argv) {

  Settings s;

  int c;
  while (true) {
    static struct option long_options[] = {
        {"accelerated", required_argument, nullptr, 'a'},
        {"allow_fp16", required_argument, nullptr, 'f'},
        {"count", required_argument, nullptr, 'c'},
        {"verbose", required_argument, nullptr, 'v'},
        {"image", required_argument, nullptr, 'i'},
        {"labels", required_argument, nullptr, 'l'},
        {"tflite_model", required_argument, nullptr, 'm'},
        {"profiling", required_argument, nullptr, 'p'},
        {"threads", required_argument, nullptr, 't'},
        {"input_mean", required_argument, nullptr, 'b'},
        {"input_std", required_argument, nullptr, 's'},
        {"num_results", required_argument, nullptr, 'r'},
        {"max_profiling_buffer_entries", required_argument, nullptr, 'e'},
        {"warmup_runs", required_argument, nullptr, 'w'},
        {"gl_backend", required_argument, nullptr, 'g'},
        {"hexagon_delegate", required_argument, nullptr, 'j'},
        {"xnnpack_delegate", required_argument, nullptr, 'x'},
        {nullptr, 0, nullptr, 0}};

    
    int option_index = 0;

    c = getopt_long(argc, argv,
                    "a:b:c:d:e:f:g:i:j:l:m:p:r:s:t:v:w:x:", long_options,
                    &option_index);

    
    if (c == -1) break;

    switch (c) {
      case 'a':
        s.accel = strtol(optarg, nullptr, 10);  // NOLINT(runtime/deprecated_fn)
        break;
      case 'b':
        s.input_mean = strtod(optarg, nullptr);
        break;
      case 'c':
        s.loop_count =
            strtol(optarg, nullptr, 10);  // NOLINT(runtime/deprecated_fn)
        break;
      case 'e':
        s.max_profiling_buffer_entries =
            strtol(optarg, nullptr, 10);  // NOLINT(runtime/deprecated_fn)
        break;
      case 'f':
        s.allow_fp16 =
            strtol(optarg, nullptr, 10);  // NOLINT(runtime/deprecated_fn)
        break;
      case 'g':
        s.gl_backend =
            strtol(optarg, nullptr, 10);  // NOLINT(runtime/deprecated_fn)
        break;
      case 'i':
        s.input_bmp_name = optarg;
        break;
      case 'j':
        s.hexagon_delegate = optarg;
        break;
      case 'l':
        s.labels_file_name = optarg;
        break;
      case 'm':
        s.model_name = optarg;
        break;
      case 'p':
        s.profiling =
            strtol(optarg, nullptr, 10);  // NOLINT(runtime/deprecated_fn)
        break;
      case 'r':
        s.number_of_results =
            strtol(optarg, nullptr, 10);  // NOLINT(runtime/deprecated_fn)
        break;
      case 's':
        s.input_std = strtod(optarg, nullptr);
        break;
      case 't':
        s.number_of_threads = strtol(  // NOLINT(runtime/deprecated_fn)
            optarg, nullptr, 10);
        break;
      case 'v':
        s.verbose =
            strtol(optarg, nullptr, 10);  // NOLINT(runtime/deprecated_fn)
        break;
      case 'w':
        s.number_of_warmup_runs =
            strtol(optarg, nullptr, 10);  // NOLINT(runtime/deprecated_fn)
        break;
      case 'x':
        s.xnnpack_delegate =
            strtol(optarg, nullptr, 10);  // NOLINT(runtime/deprecated_fn)
        break;
      case 'h':
      case '?':
        
        display_usage();
        exit(-1);
      default:
        exit(-1);
    }
  }

  //delegate_providers.MergeSettingsIntoParams(s);
  RunInference(&s);
  return 0;
}

}  // namespace label_image
}  // namespace tflite

int main(int argc, char** argv) {
  return tflite::label_image::Main(argc, argv);
}



#include 
#include 
#include 
#include 
#include 
#include 
#include 
#include 
#include 
#include 
#include 
#include 

#include "tensorflow/lite/tools/make/downloads/absl/absl/memory/memory.h"
#include "tensorflow/lite/examples/label_image/bitmap_helpers.h"
#include "tensorflow/lite/examples/label_image/get_top_n.h"
#include "tensorflow/lite/examples/label_image/log.h"
#include "tensorflow/lite/kernels/register.h"
#include "tensorflow/lite/optional_debug_tools.h"
#include "tensorflow/lite/profiling/profiler.h"
#include "tensorflow/lite/string_util.h"
#include "tensorflow/lite/tools/command_line_flags.h"
#include "tensorflow/lite/tools/delegates/delegate_provider.h"

namespace tflite {
namespace label_image {

double get_us(struct timeval t) { return (t.tv_sec * 1000000 + t.tv_usec); }

using TfLiteDelegatePtr = tflite::Interpreter::TfLiteDelegatePtr;
using ProvidedDelegateList = tflite::tools::ProvidedDelegateList;
#if 0
template 
void resize(T* out, uint8_t* in, int image_height, int image_width,
            int image_channels, int wanted_height, int wanted_width,
            int wanted_channels, Settings* s) {
  cout<< "11111111111"< interpreter(new Interpreter);

  int base_index = 0;

  LOG(INFO)<< "11111111111"

  // two inputs: input and new_sizes
  interpreter->AddTensors(2, &base_index);
  // one output
  interpreter->AddTensors(1, &base_index);
  // set input and output tensors
  interpreter->SetInputs({0, 1});
  interpreter->SetOutputs({2});

  LOG(INFO)<< "11111111111"
  // set parameters of tensors
  TfLiteQuantizationParams quant;
  interpreter->SetTensorParametersReadWrite(
      0, kTfLiteFloat32, "input",
      {1, image_height, image_width, image_channels}, quant);
  interpreter->SetTensorParametersReadWrite(1, kTfLiteInt32, "new_size", {2},
                                            quant);
  LOG(INFO)<< "11111111111"
  interpreter->SetTensorParametersReadWrite(
      2, kTfLiteFloat32, "output",
      {1, wanted_height, wanted_width, wanted_channels}, quant);

  LOG(INFO)<< "11111111111"
  ops::builtin::BuiltinOpResolver resolver;
  const TfLiteRegistration* resize_op =
      resolver.FindOp(BuiltinOperator_RESIZE_BILINEAR, 1);
  auto* params = reinterpret_cast(
      malloc(sizeof(TfLiteResizeBilinearParams)));
  params->align_corners = false;
  params->half_pixel_centers = false;
  interpreter->AddNodeWithParameters({0, 1}, {2}, nullptr, 0, params, resize_op,
                                     nullptr);
  LOG(INFO)<< "interpreter->AllocateTensors()"

  interpreter->AllocateTensors();

  // fill input image
  // in[] are integers, cannot do memcpy() directly
  auto input = interpreter->typed_tensor(0);
  for (int i = 0; i < number_of_pixels; i++) {
    input[i] = in[i];
  }

  // fill new_sizes
  interpreter->typed_tensor(1)[0] = wanted_height;
  interpreter->typed_tensor(1)[1] = wanted_width;
  LOG(INFO)<< "interpreter->Invoke()"
  interpreter->Invoke();

  auto output = interpreter->typed_tensor(2);
  auto output_number_of_pixels = wanted_height * wanted_width * wanted_channels;

  for (int i = 0; i < output_number_of_pixels; i++) {
    switch (s->input_type) {
      case kTfLiteFloat32:
        out[i] = (output[i] - s->input_mean) / s->input_std;
        break;
      case kTfLiteInt8:
        out[i] = static_cast(output[i] - 128);
        break;
      case kTfLiteUInt8:
        out[i] = static_cast(output[i]);
        break;
      default:
        break;
    }
  }
}
#endif
// Takes a file name, and loads a list of labels from it, one per line, and
// returns a vector of the strings. It pads with empty strings so the length
// of the result is a multiple of 16, because our model expects that.
TfLiteStatus ReadLabelsFile(const string& file_name,
                            std::vector* result,
                            size_t* found_label_count) {
  std::ifstream file(file_name);
  if (!file) {
    LOG(ERROR) << "Labels file " << file_name << " not found";
    return kTfLiteError;
  }
  result->clear();
  string line;
  while (std::getline(file, line)) {
    result->push_back(line);
  }
  *found_label_count = result->size();
  const int padding = 16;
  while (result->size() % padding) {
    result->emplace_back();
  }
  return kTfLiteOk;
}

void PrintProfilingInfo(const profiling::ProfileEvent* e,
                        uint32_t subgraph_index, uint32_t op_index,
                        TfLiteRegistration registration) {
  // output something like
  // time (ms) , Node xxx, OpCode xxx, symbolic name
  //      5.352, Node   5, OpCode   4, DEPTHWISE_CONV_2D

  LOG(INFO) << std::fixed << std::setw(10) << std::setprecision(3)
            << (e->end_timestamp_us - e->begin_timestamp_us) / 1000.0
            << ", Subgraph " << std::setw(3) << std::setprecision(3)
            << subgraph_index << ", Node " << std::setw(3)
            << std::setprecision(3) << op_index << ", OpCode " << std::setw(3)
            << std::setprecision(3) << registration.builtin_code << ", "
            << EnumNameBuiltinOperator(
                   static_cast(registration.builtin_code));
}

/// 使用的是mobilenet_v2_1.0_224_quant.tflite 量化后模型
void RunInference(Settings* settings)
{
  std::unique_ptr model;
  std::unique_ptr interpreter;
  model = tflite::FlatBufferModel::BuildFromFile(settings->model_name.c_str());
  if (!model) {
    LOG(ERROR) << "Failed to mmap model " << settings->model_name;
    exit(-1);
  }
//  settings->model = model.get();
  LOG(INFO) << "Loaded model " << settings->model_name;
  //model->error_reporter();
  //LOG(INFO) << "resolved reporter";

  tflite::ops::builtin::BuiltinOpResolver resolver;

  tflite::InterpreterBuilder(*model, resolver)(&interpreter);
  if (!interpreter) {
    LOG(ERROR) << "Failed to construct interpreter";
    exit(-1);
  }

  interpreter->SetAllowFp16PrecisionForFp32(settings->allow_fp16);



  if (settings->verbose) {
    LOG(INFO) << "tensors size: " << interpreter->tensors_size();
    LOG(INFO) << "nodes size: " << interpreter->nodes_size();
    LOG(INFO) << "inputs: " << interpreter->inputs().size();
    LOG(INFO) << "input(0) name: " << interpreter->GetInputName(0);

    int t_size = interpreter->tensors_size();
    for (int i = 0; i < t_size; i++) {
      if (interpreter->tensor(i)->name)
        LOG(INFO) << i << ": " << interpreter->tensor(i)->name << ", "
                  << interpreter->tensor(i)->bytes << ", "
                  << interpreter->tensor(i)->type << ", "
                  << interpreter->tensor(i)->params.scale << ", "
                  << interpreter->tensor(i)->params.zero_point;
    }
  }

  if (settings->number_of_threads != -1) {
    interpreter->SetNumThreads(settings->number_of_threads);
  }
  LOG(INFO) << "load picture ";
  LOG(INFO) << "load picture ";
  int image_width = 224;
  int image_height = 224;
  int image_channels = 3;
  std::vector in = read_bmp(settings->input_bmp_name, &image_width,
                                     &image_height, &image_channels, settings);

  int input = interpreter->inputs()[0];
  if (settings->verbose) LOG(INFO) << "input: " << input;

  const std::vector inputs = interpreter->inputs();
  const std::vector outputs = interpreter->outputs();

  if (settings->verbose) {
    LOG(INFO) << "number of inputs: " << inputs.size();
    LOG(INFO) << "number of outputs: " << outputs.size();
  }


  if (interpreter->AllocateTensors() != kTfLiteOk) {
    LOG(ERROR) << "Failed to allocate tensors!";
    exit(-1);
  }



  if (settings->verbose) PrintInterpreterState(interpreter.get());

  // get input dimension from the input tensor metadata
  // assuming one input only
  TfLiteIntArray* dims = interpreter->tensor(input)->dims;
  int wanted_height = dims->data[1];
  int wanted_width = dims->data[2];
  int wanted_channels = dims->data[3];



  settings->input_type = interpreter->tensor(input)->type;
    LOG(INFO) << "wanted width, height, channels: " < interpreter1(new Interpreter);


  int base_index = 0;



  // two inputs: input and new_sizes
  interpreter1->AddTensors(2, &base_index);
  // one output
  interpreter1->AddTensors(1, &base_index);
  // set input and output tensors
  interpreter1->SetInputs({0, 1});
  interpreter1->SetOutputs({2});


    // set parameters of tensors
  TfLiteQuantizationParams quant;
  interpreter1->SetTensorParametersReadWrite(
      0, kTfLiteFloat32, "input",
      {1, image_height, image_width, image_channels}, quant);
  interpreter1->SetTensorParametersReadWrite(1, kTfLiteInt32, "new_size", {2},
                                            quant);

  interpreter1->SetTensorParametersReadWrite(
      2, kTfLiteFloat32, "output",
      {1, wanted_height, wanted_width, wanted_channels}, quant);



  ops::builtin::BuiltinOpResolver resolver1;
  const TfLiteRegistration* resize_op =
      resolver1.FindOp(BuiltinOperator_RESIZE_BILINEAR, 1);
  auto* params = reinterpret_cast(
      malloc(sizeof(TfLiteResizeBilinearParams)));
  params->align_corners = false;
  params->half_pixel_centers = false;
  interpreter1->AddNodeWithParameters({0, 1}, {2}, nullptr, 0, params, resize_op,
                                     nullptr);

  interpreter1->AllocateTensors();

  // fill input image
  // in[] are integers, cannot do memcpy() directly
  auto input1 = interpreter1->typed_tensor(0);
  for (int i = 0; i < number_of_pixels; i++) {
    input1[i] = in.data()[i];
  }

  // fill new_sizes
  interpreter1->typed_tensor(1)[0] = wanted_height;
  interpreter1->typed_tensor(1)[1] = wanted_width;

  interpreter1->Invoke();


  auto outputxx = interpreter1->typed_tensor(2);
  auto output_number_of_pixels = wanted_height * wanted_width * wanted_channels;
  LOG(INFO)<< "output_number_of_pixels@@@@@@@@@@@@@@@@@@@:" << output_number_of_pixels  ;

  	  printf("pinput11^^^^^^^^^^^^^^^^^^^^^^^");


	printf("typed_tensor(0):%pn",interpreter->typed_tensor(0));
    printf("typed_tensor(171):%pn",interpreter->typed_tensor(171));

    auto input11 = interpreter->typed_tensor(171);


	///将修改好尺寸的图像数据输入进模型



    for (int i = 0; i < output_number_of_pixels; i++) {


	    input11[i] = static_cast(outputxx[i]);

  }
  printf("pinput11  赋值完毕^^^^^^^^^^^^^^^^^^^^^^^");
 
#if 1
  auto profiler = absl::make_unique(
      settings->max_profiling_buffer_entries);
  interpreter->SetProfiler(profiler.get());

  if (settings->profiling) profiler->StartProfiling();
  if (settings->loop_count > 1) {
    for (int i = 0; i < settings->number_of_warmup_runs; i++) {
      if (interpreter->Invoke() != kTfLiteOk) {
        LOG(ERROR) << "Failed to invoke tflite!";
        exit(-1);
      }
    }
  }

#endif



  struct timeval start_time, stop_time;
  gettimeofday(&start_time, nullptr);
  for (int i = 0; i < settings->loop_count; i++) {
    if (interpreter->Invoke() != kTfLiteOk) {
      LOG(ERROR) << "Failed to invoke tflite!";
      exit(-1);
    }
  }
  gettimeofday(&stop_time, nullptr);
  LOG(INFO) << "invoked";
  LOG(INFO) << "average time: "
            << (get_us(stop_time) - get_us(start_time)) /
                   (settings->loop_count * 1000)
            << " ms";
#if 1
  if (settings->profiling) {
    profiler->StopProfiling();
    auto profile_events = profiler->GetProfileEvents();
    for (int i = 0; i < profile_events.size(); i++) {
      auto subgraph_index = profile_events[i]->extra_event_metadata;
      auto op_index = profile_events[i]->event_metadata;
      const auto subgraph = interpreter->subgraph(subgraph_index);
      const auto node_and_registration =
          subgraph->node_and_registration(op_index);
      const TfLiteRegistration registration = node_and_registration->second;
      PrintProfilingInfo(profile_events[i], subgraph_index, op_index,
                         registration);
    }
  }
#endif

   const float threshold = 0.001f;

  std::vector> top_results;

  int output = interpreter->outputs()[0];
  TfLiteIntArray* output_dims = interpreter->tensor(output)->dims;
  // assume output dims to be something like (1, 1, ... ,size)
  auto output_size = output_dims->data[output_dims->size - 1];
  switch (interpreter->tensor(output)->type) {
    case kTfLiteFloat32:
      get_top_n(interpreter->typed_output_tensor(0), output_size,
                       settings->number_of_results, threshold, &top_results,
                       settings->input_type);
      break;
    case kTfLiteInt8:
      get_top_n(interpreter->typed_output_tensor(0),
                        output_size, settings->number_of_results, threshold,
                        &top_results, settings->input_type);
      break;
    case kTfLiteUInt8:
      get_top_n(interpreter->typed_output_tensor(0),
                         output_size, settings->number_of_results, threshold,
                         &top_results, settings->input_type);
      break;
    default:
      LOG(ERROR) << "cannot handle output type "
                 << interpreter->tensor(output)->type << " yet";
      exit(-1);
  }

  std::vector labels;
  size_t label_count;

  if (ReadLabelsFile(settings->labels_file_name, &labels, &label_count) !=
      kTfLiteOk)
    exit(-1);

  for (const auto& result : top_results) {
    const float confidence = result.first;
    const int index = result.second;
    LOG(INFO) << confidence << ": " << index << " " << labels[index];
  }
}

void display_usage() {
  LOG(INFO)
      << "label_imagen"
      << "--accelerated, -a: [0|1], use Android NNAPI or notn"
      << "--allow_fp16, -f: [0|1], allow running fp32 models with fp16 or notn"
      << "--count, -c: loop interpreter->Invoke() for certain timesn"
      << "--gl_backend, -g: [0|1]: use GL GPU Delegate on Androidn"
      << "--hexagon_delegate, -j: [0|1]: use Hexagon Delegate on Androidn"
      << "--input_mean, -b: input meann"
      << "--input_std, -s: input standard deviationn"
      << "--image, -i: image_name.bmpn"
      << "--labels, -l: labels for the modeln"
      << "--tflite_model, -m: model_name.tfliten"
      << "--profiling, -p: [0|1], profiling or notn"
      << "--num_results, -r: number of results to shown"
      << "--threads, -t: number of threadsn"
      << "--verbose, -v: [0|1] print more informationn"
      << "--warmup_runs, -w: number of warmup runsn"
      << "--xnnpack_delegate, -x [0:1]: xnnpack delegaten";
}

int Main(int argc, char** argv) {

  Settings s;

  int c;
  while (true) {
    static struct option long_options[] = {
        {"accelerated", required_argument, nullptr, 'a'},
        {"allow_fp16", required_argument, nullptr, 'f'},
        {"count", required_argument, nullptr, 'c'},
        {"verbose", required_argument, nullptr, 'v'},
        {"image", required_argument, nullptr, 'i'},
        {"labels", required_argument, nullptr, 'l'},
        {"tflite_model", required_argument, nullptr, 'm'},
        {"profiling", required_argument, nullptr, 'p'},
        {"threads", required_argument, nullptr, 't'},
        {"input_mean", required_argument, nullptr, 'b'},
        {"input_std", required_argument, nullptr, 's'},
        {"num_results", required_argument, nullptr, 'r'},
        {"max_profiling_buffer_entries", required_argument, nullptr, 'e'},
        {"warmup_runs", required_argument, nullptr, 'w'},
        {"gl_backend", required_argument, nullptr, 'g'},
        {"hexagon_delegate", required_argument, nullptr, 'j'},
        {"xnnpack_delegate", required_argument, nullptr, 'x'},
        {nullptr, 0, nullptr, 0}};

    
    int option_index = 0;

    c = getopt_long(argc, argv,
                    "a:b:c:d:e:f:g:i:j:l:m:p:r:s:t:v:w:x:", long_options,
                    &option_index);

    
    if (c == -1) break;

    switch (c) {
      case 'a':
        s.accel = strtol(optarg, nullptr, 10);  // NOLINT(runtime/deprecated_fn)
        break;
      case 'b':
        s.input_mean = strtod(optarg, nullptr);
        break;
      case 'c':
        s.loop_count =
            strtol(optarg, nullptr, 10);  // NOLINT(runtime/deprecated_fn)
        break;
      case 'e':
        s.max_profiling_buffer_entries =
            strtol(optarg, nullptr, 10);  // NOLINT(runtime/deprecated_fn)
        break;
      case 'f':
        s.allow_fp16 =
            strtol(optarg, nullptr, 10);  // NOLINT(runtime/deprecated_fn)
        break;
      case 'g':
        s.gl_backend =
            strtol(optarg, nullptr, 10);  // NOLINT(runtime/deprecated_fn)
        break;
      case 'i':
        s.input_bmp_name = optarg;
        break;
      case 'j':
        s.hexagon_delegate = optarg;
        break;
      case 'l':
        s.labels_file_name = optarg;
        break;
      case 'm':
        s.model_name = optarg;
        break;
      case 'p':
        s.profiling =
            strtol(optarg, nullptr, 10);  // NOLINT(runtime/deprecated_fn)
        break;
      case 'r':
        s.number_of_results =
            strtol(optarg, nullptr, 10);  // NOLINT(runtime/deprecated_fn)
        break;
      case 's':
        s.input_std = strtod(optarg, nullptr);
        break;
      case 't':
        s.number_of_threads = strtol(  // NOLINT(runtime/deprecated_fn)
            optarg, nullptr, 10);
        break;
      case 'v':
        s.verbose =
            strtol(optarg, nullptr, 10);  // NOLINT(runtime/deprecated_fn)
        break;
      case 'w':
        s.number_of_warmup_runs =
            strtol(optarg, nullptr, 10);  // NOLINT(runtime/deprecated_fn)
        break;
      case 'x':
        s.xnnpack_delegate =
            strtol(optarg, nullptr, 10);  // NOLINT(runtime/deprecated_fn)
        break;
      case 'h':
      case '?':
        
        display_usage();
        exit(-1);
      default:
        exit(-1);
    }
  }

  //delegate_providers.MergeSettingsIntoParams(s);
  RunInference(&s);
  return 0;
}

}  // namespace label_image
}  // namespace tflite

int main(int argc, char** argv) {
  return tflite::label_image::Main(argc, argv);
}



 cmake CMakeLists.txt 
 make

生成的label_image可执行文件使我们需要的。

需要传到开发板上的数据（共4个文件），目录结构如下：

.
├── grace_hopper.bmp
├── label_image
├── labels_mobilenet_quant_v1_224.txt
└── model
    └── mobilenet_v2_1.0_224_quant.tflite

grace_hopper.bmp:从test/label_image/testdata目录下获取。测试图片

label_image:是上面我们编译出来的可执行文件。
labels_mobilenet_quant_v1_224.txt：获取网址：

https://tensorflow.google.cn/lite/examples/image_classification/overview?hl=zh-cn

里面有个下载新手图像分类及变迁，只取里面的标签文件即可。

mobilenet_v2_1.0_224_quant.tflite：获取网址：

https://tensorflow.google.cn/lite/guide/hosted_models
二 在开发板上操作

登录开发板，切换到传过来的项目目录里，执行下面第一行命令（最后 0.717647: 653 military uniform概率最大，预测效果还可以）：

Node  50 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[62,70,68] -> 165280B (0.16MB)
  1 Output Tensors:[69] -> 47040B (0.04MB)
Node  51 Operator Builtin Code   4 DEPTHWISE_CONV_2D (not delegated)
  3 Input Tensors:[69,67,66] -> 59520B (0.06MB)
  1 Output Tensors:[65] -> 47040B (0.04MB)
Node  52 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[65,73,71] -> 201280B (0.19MB)
  1 Output Tensors:[72] -> 7840B (0.01MB)
Node  53 Operator Builtin Code   0 ADD (not delegated)
  2 Input Tensors:[72,62] -> 15680B (0.01MB)
  1 Output Tensors:[64] -> 7840B (0.01MB)
Node  54 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[64,80,78] -> 165280B (0.16MB)
  1 Output Tensors:[79] -> 47040B (0.04MB)
Node  55 Operator Builtin Code   4 DEPTHWISE_CONV_2D (not delegated)
  3 Input Tensors:[79,77,76] -> 59520B (0.06MB)
  1 Output Tensors:[75] -> 47040B (0.04MB)
Node  56 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[75,83,81] -> 201280B (0.19MB)
  1 Output Tensors:[82] -> 7840B (0.01MB)
./label_image -v 1 -m ./model/mobilenet_v2_1.0_224_quant.tflite -i ./grace_hopper.bmp -l ./labels_mobilenet_quant_v1_224.txt -t 1
INFO: Loaded model ./model/mobilenet_v2_1.0_224_quant.tflite
INFO: tensors size: 173
INFO: nodes size: 65
INFO: inputs: 1
INFO: input(0) name: input
INFO: 0: MobilenetV2/Conv/Conv2D_Fold_bias, 128, 2, 0.000265382, 0
INFO: 1: MobilenetV2/Conv/Relu6, 401408, 3, 0.0235285, 0
INFO: 2: MobilenetV2/Conv/weights_quant/FakeQuantWithMinMaxVars, 864, 3, 0.0339689, 122
INFO: 3: MobilenetV2/Conv_1/Conv2D_Fold_bias, 5120, 2, 0.000604263, 0
INFO: 4: MobilenetV2/Conv_1/Relu6, 62720, 3, 0.0235285, 0
INFO: 5: MobilenetV2/Conv_1/weights_quant/FakeQuantWithMinMaxVars, 409600, 3, 0.00516707, 125
INFO: 6: MobilenetV2/Logits/AvgPool, 1280, 3, 0.0235285, 0
INFO: 7: MobilenetV2/Logits/Conv2d_1c_1x1/BiasAdd, 1001, 3, 0.0988925, 58
INFO: 8: MobilenetV2/Logits/Conv2d_1c_1x1/Conv2D_bias, 4004, 2, 3.97886e-05, 0
INFO: 9: MobilenetV2/Logits/Conv2d_1c_1x1/weights_quant/FakeQuantWithMinMaxVars, 1281280, 3, 0.00169108, 113
INFO: 10: MobilenetV2/Logits/Squeeze_shape, 8, 2, 0, 0
INFO: 11: MobilenetV2/expanded_conv/depthwise/Relu6, 401408, 3, 0.0235285, 0
INFO: 12: MobilenetV2/expanded_conv/depthwise/depthwise_Fold_bias, 128, 2, 0.00808663, 0
INFO: 13: MobilenetV2/expanded_conv/depthwise/weights_quant/FakeQuantWithMinMaxVars, 288, 3, 0.343696, 165
INFO: 14: MobilenetV2/expanded_conv/project/Conv2D_Fold_bias, 64, 2, 0.0008793, 0
INFO: 15: MobilenetV2/expanded_conv/project/add_fold, 200704, 3, 0.354413, 129
INFO: 16: MobilenetV2/expanded_conv/project/weights_quant/FakeQuantWithMinMaxVars, 512, 3, 0.0373718, 140
INFO: 17: MobilenetV2/expanded_conv_1/depthwise/Relu6, 301056, 3, 0.0235285, 0
INFO: 18: MobilenetV2/expanded_conv_1/depthwise/depthwise_Fold_bias, 384, 2, 0.000493372, 0
INFO: 19: MobilenetV2/expanded_conv_1/depthwise/weights_quant/FakeQuantWithMinMaxVars, 864, 3, 0.0209691, 109
INFO: 20: MobilenetV2/expanded_conv_1/expand/Conv2D_Fold_bias, 384, 2, 0.00345855, 0
INFO: 21: MobilenetV2/expanded_conv_1/expand/Relu6, 1204224, 3, 0.0235285, 0
INFO: 22: MobilenetV2/expanded_conv_1/expand/weights_quant/FakeQuantWithMinMaxVars, 1536, 3, 0.00975851, 127
INFO: 23: MobilenetV2/expanded_conv_1/project/Conv2D_Fold_bias, 96, 2, 0.000530238, 0
INFO: 24: MobilenetV2/expanded_conv_1/project/add_fold, 75264, 3, 0.275834, 119
INFO: 25: MobilenetV2/expanded_conv_1/project/weights_quant/FakeQuantWithMinMaxVars, 2304, 3, 0.022536, 156
INFO: 26: MobilenetV2/expanded_conv_10/depthwise/Relu6, 75264, 3, 0.0235285, 0
INFO: 27: MobilenetV2/expanded_conv_10/depthwise/depthwise_Fold_bias, 1536, 2, 0.00073192, 0
INFO: 28: MobilenetV2/expanded_conv_10/depthwise/weights_quant/FakeQuantWithMinMaxVars, 3456, 3, 0.0311078, 143
INFO: 29: MobilenetV2/expanded_conv_10/expand/Conv2D_Fold_bias, 1536, 2, 0.00035866, 0
INFO: 30: MobilenetV2/expanded_conv_10/expand/Relu6, 75264, 3, 0.0235285, 0
INFO: 31: MobilenetV2/expanded_conv_10/expand/weights_quant/FakeQuantWithMinMaxVars, 24576, 3, 0.00162825, 131
INFO: 32: MobilenetV2/expanded_conv_10/project/Conv2D_Fold_bias, 384, 2, 0.000174965, 0
INFO: 33: MobilenetV2/expanded_conv_10/project/add_fold, 18816, 3, 0.170611, 129
INFO: 34: MobilenetV2/expanded_conv_10/project/weights_quant/FakeQuantWithMinMaxVars, 36864, 3, 0.00743631, 129
INFO: 35: MobilenetV2/expanded_conv_11/add, 18816, 3, 0.176158, 127
INFO: 36: MobilenetV2/expanded_conv_11/depthwise/Relu6, 112896, 3, 0.0235285, 0
INFO: 37: MobilenetV2/expanded_conv_11/depthwise/depthwise_Fold_bias, 2304, 2, 0.00166601, 0
INFO: 38: MobilenetV2/expanded_conv_11/depthwise/weights_quant/FakeQuantWithMinMaxVars, 5184, 3, 0.0708081, 66
INFO: 39: MobilenetV2/expanded_conv_11/expand/Conv2D_Fold_bias, 2304, 2, 0.000278264, 0
INFO: 40: MobilenetV2/expanded_conv_11/expand/Relu6, 112896, 3, 0.0235285, 0
INFO: 41: MobilenetV2/expanded_conv_11/expand/weights_quant/FakeQuantWithMinMaxVars, 55296, 3, 0.00163099, 134
INFO: 42: MobilenetV2/expanded_conv_11/project/Conv2D_Fold_bias, 384, 2, 0.000197221, 0
INFO: 43: MobilenetV2/expanded_conv_11/project/add_fold, 18816, 3, 0.123328, 127
INFO: 44: MobilenetV2/expanded_conv_11/project/weights_quant/FakeQuantWithMinMaxVars, 55296, 3, 0.00838223, 136
INFO: 45: MobilenetV2/expanded_conv_12/add, 18816, 3, 0.233401, 126
INFO: 46: MobilenetV2/expanded_conv_12/depthwise/Relu6, 112896, 3, 0.0235285, 0
INFO: 47: MobilenetV2/expanded_conv_12/depthwise/depthwise_Fold_bias, 2304, 2, 0.00175259, 0
INFO: 48: MobilenetV2/expanded_conv_12/depthwise/weights_quant/FakeQuantWithMinMaxVars, 5184, 3, 0.0744879, 159
INFO: 49: MobilenetV2/expanded_conv_12/expand/Conv2D_Fold_bias, 2304, 2, 0.000321643, 0
INFO: 50: MobilenetV2/expanded_conv_12/expand/Relu6, 112896, 3, 0.0235285, 0
INFO: 51: MobilenetV2/expanded_conv_12/expand/weights_quant/FakeQuantWithMinMaxVars, 55296, 3, 0.00182588, 138
INFO: 52: MobilenetV2/expanded_conv_12/project/Conv2D_Fold_bias, 384, 2, 0.000564274, 0
INFO: 53: MobilenetV2/expanded_conv_12/project/add_fold, 18816, 3, 0.186196, 127
INFO: 54: MobilenetV2/expanded_conv_12/project/weights_quant/FakeQuantWithMinMaxVars, 55296, 3, 0.0239826, 154
INFO: 55: MobilenetV2/expanded_conv_13/depthwise/Relu6, 28224, 3, 0.0235285, 0
INFO: 56: MobilenetV2/expanded_conv_13/depthwise/depthwise_Fold_bias, 2304, 2, 0.000358996, 0
INFO: 57: MobilenetV2/expanded_conv_13/depthwise/weights_quant/FakeQuantWithMinMaxVars, 5184, 3, 0.0152579, 92
INFO: 58: MobilenetV2/expanded_conv_13/expand/Conv2D_Fold_bias, 2304, 2, 0.000322747, 0
INFO: 59: MobilenetV2/expanded_conv_13/expand/Relu6, 112896, 3, 0.0235285, 0
INFO: 60: MobilenetV2/expanded_conv_13/expand/weights_quant/FakeQuantWithMinMaxVars, 55296, 3, 0.0013828, 123
INFO: 61: MobilenetV2/expanded_conv_13/project/Conv2D_Fold_bias, 640, 2, 0.000222296, 0
INFO: 62: MobilenetV2/expanded_conv_13/project/add_fold, 7840, 3, 0.132378, 132
INFO: 63: MobilenetV2/expanded_conv_13/project/weights_quant/FakeQuantWithMinMaxVars, 92160, 3, 0.00944795, 140
INFO: 64: MobilenetV2/expanded_conv_14/add, 7840, 3, 0.15071, 134
INFO: 65: MobilenetV2/expanded_conv_14/depthwise/Relu6, 47040, 3, 0.0235285, 0
INFO: 66: MobilenetV2/expanded_conv_14/depthwise/depthwise_Fold_bias, 3840, 2, 0.000980373, 0
INFO: 67: MobilenetV2/expanded_conv_14/depthwise/weights_quant/FakeQuantWithMinMaxVars, 8640, 3, 0.0416675, 147
INFO: 68: MobilenetV2/expanded_conv_14/expand/Conv2D_Fold_bias, 3840, 2, 0.000267696, 0
INFO: 69: MobilenetV2/expanded_conv_14/expand/Relu6, 47040, 3, 0.0235285, 0
INFO: 70: MobilenetV2/expanded_conv_14/expand/weights_quant/FakeQuantWithMinMaxVars, 153600, 3, 0.00202221, 135
INFO: 71: MobilenetV2/expanded_conv_14/project/Conv2D_Fold_bias, 640, 2, 0.000185844, 0
INFO: 72: MobilenetV2/expanded_conv_14/project/add_fold, 7840, 3, 0.100457, 129
INFO: 73: MobilenetV2/expanded_conv_14/project/weights_quant/FakeQuantWithMinMaxVars, 153600, 3, 0.0078987, 139
INFO: 74: MobilenetV2/expanded_conv_15/add, 7840, 3, 0.210051, 131
INFO: 75: MobilenetV2/expanded_conv_15/depthwise/Relu6, 47040, 3, 0.0235285, 0
INFO: 76: MobilenetV2/expanded_conv_15/depthwise/depthwise_Fold_bias, 3840, 2, 0.00100747, 0
INFO: 77: MobilenetV2/expanded_conv_15/depthwise/weights_quant/FakeQuantWithMinMaxVars, 8640, 3, 0.0428194, 102
INFO: 78: MobilenetV2/expanded_conv_15/expand/Conv2D_Fold_bias, 3840, 2, 0.000240298, 0
INFO: 79: MobilenetV2/expanded_conv_15/expand/Relu6, 47040, 3, 0.0235285, 0
INFO: 80: MobilenetV2/expanded_conv_15/expand/weights_quant/FakeQuantWithMinMaxVars, 153600, 3, 0.00159444, 127
INFO: 81: MobilenetV2/expanded_conv_15/project/Conv2D_Fold_bias, 640, 2, 0.000869944, 0
INFO: 82: MobilenetV2/expanded_conv_15/project/add_fold, 7840, 3, 0.169606, 133
INFO: 83: MobilenetV2/expanded_conv_15/project/weights_quant/FakeQuantWithMinMaxVars, 153600, 3, 0.0369741, 131
INFO: 84: MobilenetV2/expanded_conv_16/depthwise/Relu6, 47040, 3, 0.0235285, 0
INFO: 85: MobilenetV2/expanded_conv_16/depthwise/depthwise_Fold_bias, 3840, 2, 0.00387191, 0
INFO: 86: MobilenetV2/expanded_conv_16/depthwise/weights_quant/FakeQuantWithMinMaxVars, 8640, 3, 0.164563, 201
INFO: 87: MobilenetV2/expanded_conv_16/expand/Conv2D_Fold_bias, 3840, 2, 0.000429939, 0
INFO: 88: MobilenetV2/expanded_conv_16/expand/Relu6, 47040, 3, 0.0235285, 0
INFO: 89: MobilenetV2/expanded_conv_16/expand/weights_quant/FakeQuantWithMinMaxVars, 153600, 3, 0.00204683, 135
INFO: 90: MobilenetV2/expanded_conv_16/project/Conv2D_Fold_bias, 1280, 2, 0.000188446, 0
INFO: 91: MobilenetV2/expanded_conv_16/project/add_fold, 15680, 3, 0.116945, 130
INFO: 92: MobilenetV2/expanded_conv_16/project/weights_quant/FakeQuantWithMinMaxVars, 307200, 3, 0.00800929, 111
INFO: 93: MobilenetV2/expanded_conv_2/add, 75264, 3, 0.432169, 133
INFO: 94: MobilenetV2/expanded_conv_2/depthwise/Relu6, 451584, 3, 0.0235285, 0
INFO: 95: MobilenetV2/expanded_conv_2/depthwise/depthwise_Fold_bias, 576, 2, 0.00399559, 0
INFO: 96: MobilenetV2/expanded_conv_2/depthwise/weights_quant/FakeQuantWithMinMaxVars, 1296, 3, 0.169819, 52
INFO: 97: MobilenetV2/expanded_conv_2/expand/Conv2D_Fold_bias, 576, 2, 0.00100837, 0
INFO: 98: MobilenetV2/expanded_conv_2/expand/Relu6, 451584, 3, 0.0235285, 0
INFO: 99: MobilenetV2/expanded_conv_2/expand/weights_quant/FakeQuantWithMinMaxVars, 3456, 3, 0.0036557, 144
INFO: 100: MobilenetV2/expanded_conv_2/project/Conv2D_Fold_bias, 96, 2, 0.000644889, 0
INFO: 101: MobilenetV2/expanded_conv_2/project/add_fold, 75264, 3, 0.401493, 136
INFO: 102: MobilenetV2/expanded_conv_2/project/weights_quant/FakeQuantWithMinMaxVars, 3456, 3, 0.0274089, 122
INFO: 103: MobilenetV2/expanded_conv_3/depthwise/Relu6, 112896, 3, 0.0235285, 0
INFO: 104: MobilenetV2/expanded_conv_3/depthwise/depthwise_Fold_bias, 576, 2, 0.000404757, 0
INFO: 105: MobilenetV2/expanded_conv_3/depthwise/weights_quant/FakeQuantWithMinMaxVars, 1296, 3, 0.0172029, 143
INFO: 106: MobilenetV2/expanded_conv_3/expand/Conv2D_Fold_bias, 576, 2, 0.00129602, 0
INFO: 107: MobilenetV2/expanded_conv_3/expand/Relu6, 451584, 3, 0.0235285, 0
INFO: 108: MobilenetV2/expanded_conv_3/expand/weights_quant/FakeQuantWithMinMaxVars, 3456, 3, 0.00299887, 104
INFO: 109: MobilenetV2/expanded_conv_3/project/Conv2D_Fold_bias, 128, 2, 0.00039633, 0
INFO: 110: MobilenetV2/expanded_conv_3/project/add_fold, 25088, 3, 0.218362, 127
INFO: 111: MobilenetV2/expanded_conv_3/project/weights_quant/FakeQuantWithMinMaxVars, 4608, 3, 0.0168447, 111
INFO: 112: MobilenetV2/expanded_conv_4/add, 25088, 3, 0.25969, 130
INFO: 113: MobilenetV2/expanded_conv_4/depthwise/Relu6, 150528, 3, 0.0235285, 0
INFO: 114: MobilenetV2/expanded_conv_4/depthwise/depthwise_Fold_bias, 768, 2, 0.00153525, 0
INFO: 115: MobilenetV2/expanded_conv_4/depthwise/weights_quant/FakeQuantWithMinMaxVars, 1728, 3, 0.0652507, 118
INFO: 116: MobilenetV2/expanded_conv_4/expand/Conv2D_Fold_bias, 768, 2, 0.000420222, 0
INFO: 117: MobilenetV2/expanded_conv_4/expand/Relu6, 150528, 3, 0.0235285, 0
INFO: 118: MobilenetV2/expanded_conv_4/expand/weights_quant/FakeQuantWithMinMaxVars, 6144, 3, 0.00192442, 128
INFO: 119: MobilenetV2/expanded_conv_4/project/Conv2D_Fold_bias, 128, 2, 0.000448521, 0
INFO: 120: MobilenetV2/expanded_conv_4/project/add_fold, 25088, 3, 0.227942, 121
INFO: 121: MobilenetV2/expanded_conv_4/project/weights_quant/FakeQuantWithMinMaxVars, 6144, 3, 0.0190629, 146
INFO: 122: MobilenetV2/expanded_conv_5/add, 25088, 3, 0.331715, 124
INFO: 123: MobilenetV2/expanded_conv_5/depthwise/Relu6, 150528, 3, 0.0235285, 0
INFO: 124: MobilenetV2/expanded_conv_5/depthwise/depthwise_Fold_bias, 768, 2, 0.00186105, 0
INFO: 125: MobilenetV2/expanded_conv_5/depthwise/weights_quant/FakeQuantWithMinMaxVars, 1728, 3, 0.0790978, 95
INFO: 126: MobilenetV2/expanded_conv_5/expand/Conv2D_Fold_bias, 768, 2, 0.000354455, 0
INFO: 127: MobilenetV2/expanded_conv_5/expand/Relu6, 150528, 3, 0.0235285, 0
INFO: 128: MobilenetV2/expanded_conv_5/expand/weights_quant/FakeQuantWithMinMaxVars, 6144, 3, 0.00136492, 135
INFO: 129: MobilenetV2/expanded_conv_5/project/Conv2D_Fold_bias, 128, 2, 0.000430409, 0
INFO: 130: MobilenetV2/expanded_conv_5/project/add_fold, 25088, 3, 0.257749, 124
INFO: 131: MobilenetV2/expanded_conv_5/project/weights_quant/FakeQuantWithMinMaxVars, 6144, 3, 0.0182931, 128
INFO: 132: MobilenetV2/expanded_conv_6/depthwise/Relu6, 37632, 3, 0.0235285, 0
INFO: 133: MobilenetV2/expanded_conv_6/depthwise/depthwise_Fold_bias, 768, 2, 0.000237353, 0
INFO: 134: MobilenetV2/expanded_conv_6/depthwise/weights_quant/FakeQuantWithMinMaxVars, 1728, 3, 0.0100879, 127
INFO: 135: MobilenetV2/expanded_conv_6/expand/Conv2D_Fold_bias, 768, 2, 0.000635912, 0
INFO: 136: MobilenetV2/expanded_conv_6/expand/Relu6, 150528, 3, 0.0235285, 0
INFO: 137: MobilenetV2/expanded_conv_6/expand/weights_quant/FakeQuantWithMinMaxVars, 6144, 3, 0.00191704, 127
INFO: 138: MobilenetV2/expanded_conv_6/project/Conv2D_Fold_bias, 256, 2, 0.000343546, 0
INFO: 139: MobilenetV2/expanded_conv_6/project/add_fold, 12544, 3, 0.185405, 126
INFO: 140: MobilenetV2/expanded_conv_6/project/weights_quant/FakeQuantWithMinMaxVars, 12288, 3, 0.0146013, 147
INFO: 141: MobilenetV2/expanded_conv_7/add, 12544, 3, 0.18911, 122
INFO: 142: MobilenetV2/expanded_conv_7/depthwise/Relu6, 75264, 3, 0.0235285, 0
INFO: 143: MobilenetV2/expanded_conv_7/depthwise/depthwise_Fold_bias, 1536, 2, 0.00143352, 0
INFO: 144: MobilenetV2/expanded_conv_7/depthwise/weights_quant/FakeQuantWithMinMaxVars, 3456, 3, 0.0609271, 110
INFO: 145: MobilenetV2/expanded_conv_7/expand/Conv2D_Fold_bias, 1536, 2, 0.0002881, 0
INFO: 146: MobilenetV2/expanded_conv_7/expand/Relu6, 75264, 3, 0.0235285, 0
INFO: 147: MobilenetV2/expanded_conv_7/expand/weights_quant/FakeQuantWithMinMaxVars, 24576, 3, 0.00155389, 125
INFO: 148: MobilenetV2/expanded_conv_7/project/Conv2D_Fold_bias, 256, 2, 0.000394877, 0
INFO: 149: MobilenetV2/expanded_conv_7/project/add_fold, 12544, 3, 0.172635, 109
INFO: 150: MobilenetV2/expanded_conv_7/project/weights_quant/FakeQuantWithMinMaxVars, 24576, 3, 0.0167829, 124
INFO: 151: MobilenetV2/expanded_conv_8/add, 12544, 3, 0.199681, 124
INFO: 152: MobilenetV2/expanded_conv_8/depthwise/Relu6, 75264, 3, 0.0235285, 0
INFO: 153: MobilenetV2/expanded_conv_8/depthwise/depthwise_Fold_bias, 1536, 2, 0.00123308, 0
INFO: 154: MobilenetV2/expanded_conv_8/depthwise/weights_quant/FakeQuantWithMinMaxVars, 3456, 3, 0.0524078, 133
INFO: 155: MobilenetV2/expanded_conv_8/expand/Conv2D_Fold_bias, 1536, 2, 0.000278048, 0
INFO: 156: MobilenetV2/expanded_conv_8/expand/Relu6, 75264, 3, 0.0235285, 0
INFO: 157: MobilenetV2/expanded_conv_8/expand/weights_quant/FakeQuantWithMinMaxVars, 24576, 3, 0.0014703, 134
INFO: 158: MobilenetV2/expanded_conv_8/project/Conv2D_Fold_bias, 256, 2, 0.000303476, 0
INFO: 159: MobilenetV2/expanded_conv_8/project/add_fold, 12544, 3, 0.147155, 123
INFO: 160: MobilenetV2/expanded_conv_8/project/weights_quant/FakeQuantWithMinMaxVars, 24576, 3, 0.0128983, 125
INFO: 161: MobilenetV2/expanded_conv_9/add, 12544, 3, 0.220273, 120
INFO: 162: MobilenetV2/expanded_conv_9/depthwise/Relu6, 75264, 3, 0.0235285, 0
INFO: 163: MobilenetV2/expanded_conv_9/depthwise/depthwise_Fold_bias, 1536, 2, 0.000959465, 0
INFO: 164: MobilenetV2/expanded_conv_9/depthwise/weights_quant/FakeQuantWithMinMaxVars, 3456, 3, 0.0407789, 155
INFO: 165: MobilenetV2/expanded_conv_9/expand/Conv2D_Fold_bias, 1536, 2, 0.000274232, 0
INFO: 166: MobilenetV2/expanded_conv_9/expand/Relu6, 75264, 3, 0.0235285, 0
INFO: 167: MobilenetV2/expanded_conv_9/expand/weights_quant/FakeQuantWithMinMaxVars, 24576, 3, 0.00137335, 127
INFO: 168: MobilenetV2/expanded_conv_9/project/Conv2D_Fold_bias, 256, 2, 0.000460252, 0
INFO: 169: MobilenetV2/expanded_conv_9/project/add_fold, 12544, 3, 0.156276, 122
INFO: 170: MobilenetV2/expanded_conv_9/project/weights_quant/FakeQuantWithMinMaxVars, 24576, 3, 0.0195615, 144
INFO: 171: input, 150528, 3, 0.0078125, 128
INFO: 172: output, 1001, 3, 0.0988925, 58
INFO: load picture 
INFO: load picture 
INFO: len: 940650
INFO: width, height, channels: 517, 606, 3
INFO: input: 171
INFO: number of inputs: 1
INFO: number of outputs: 1
Interpreter has 1 subgraphs.

-----------Subgraph-0 has 174 tensors and 65 nodes------------
1 Inputs: [171] -> 150528B (0.14MB)
1 Outputs: [172] -> 1001B (0.00MB)

Tensor  ID Name                      Type            AllocType          Size (Bytes/MB)    Shape      MemAddr-Offset  
Tensor   0 MobilenetV2/Conv/Conv2... kTfLiteInt32    kTfLiteMmapRo      128      / 0.00 [32] [3477068, 3477196)
Tensor   1 MobilenetV2/Conv/Relu6    kTfLiteUInt8    kTfLiteArenaRw     401408   / 0.38 [1,112,112,32] [150528, 551936)
Tensor   2 MobilenetV2/Conv/weigh... kTfLiteUInt8    kTfLiteMmapRo      864      / 0.00 [32,3,3,3] [3476180, 3477044)
Tensor   3 MobilenetV2/Conv_1/Con... kTfLiteInt32    kTfLiteMmapRo      5120     / 0.00 [1280] [3535656, 3540776)
Tensor   4 MobilenetV2/Conv_1/Relu6  kTfLiteUInt8    kTfLiteArenaRw     62720    / 0.06 [1,7,7,1280] [150528, 213248)
Tensor   5 MobilenetV2/Conv_1/wei... kTfLiteUInt8    kTfLiteMmapRo      409600   / 0.39 [1280,1,1,320] [2696436, 3106036)
Tensor   6 MobilenetV2/Logits/Avg... kTfLiteUInt8    kTfLiteArenaRw     1280     / 0.00 [1,1,1,1280] [213248, 214528)
Tensor   7 MobilenetV2/Logits/Con... kTfLiteUInt8    kTfLiteArenaRw     1001     / 0.00 [1,1,1,1001] [150528, 151529)
Tensor   8 MobilenetV2/Logits/Con... kTfLiteInt32    kTfLiteMmapRo      4004     / 0.00 [1001] [3540788, 3544792)
Tensor   9 MobilenetV2/Logits/Con... kTfLiteUInt8    kTfLiteMmapRo      1281280  / 1.22 [1001,1,1,1280] [1275816, 2557096)
Tensor  10 MobilenetV2/Logits/Squ... kTfLiteInt32    kTfLiteMmapRo      8        / 0.00 [2] [3544804, 3544812)
Tensor  11 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     401408   / 0.38 [1,112,112,32] [551936, 953344)
Tensor  12 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      128      / 0.00 [32] [3477208, 3477336)
Tensor  13 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      288      / 0.00 [1,3,3,32] [1274980, 1275268)
Tensor  14 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      64       / 0.00 [16] [3477348, 3477412)
Tensor  15 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     200704   / 0.19 [1,112,112,16] [1354752, 1555456)
Tensor  16 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      512      / 0.00 [16,1,1,32] [1275292, 1275804)
Tensor  17 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     301056   / 0.29 [1,56,56,96] [1354752, 1655808)
Tensor  18 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      384      / 0.00 [96] [3477424, 3477808)
Tensor  19 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      864      / 0.00 [1,3,3,96] [831076, 831940)
Tensor  20 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      384      / 0.00 [96] [3475784, 3476168)
Tensor  21 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     1204224  / 1.15 [1,112,112,96] [150528, 1354752)
Tensor  22 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      1536     / 0.00 [96,1,1,16] [1273432, 1274968)
Tensor  23 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      96       / 0.00 [24] [3477820, 3477916)
Tensor  24 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     75264    / 0.07 [1,56,56,24] [1053696, 1128960)
Tensor  25 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      2304     / 0.00 [24,1,1,96] [3418160, 3420464)
Tensor  26 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     75264    / 0.07 [1,14,14,384] [225792, 301056)
Tensor  27 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      1536     / 0.00 [384] [2602780, 2604316)
Tensor  28 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      3456     / 0.00 [1,3,3,384] [2628920, 2632376)
Tensor  29 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      1536     / 0.00 [384] [3492664, 3494200)
Tensor  30 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     75264    / 0.07 [1,14,14,384] [150528, 225792)
Tensor  31 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      24576    / 0.02 [384,1,1,64] [1170072, 1194648)
Tensor  32 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      384      / 0.00 [96] [3494212, 3494596)
Tensor  33 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     18816    / 0.02 [1,14,14,96] [376320, 395136)
Tensor  34 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      36864    / 0.04 [96,1,1,384] [1236556, 1273420)
Tensor  35 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     18816    / 0.02 [1,14,14,96] [395136, 413952)
Tensor  36 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     112896   / 0.11 [1,14,14,576] [263424, 376320)
Tensor  37 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      2304     / 0.00 [576] [3496924, 3499228)
Tensor  38 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      5184     / 0.00 [1,3,3,576] [2660484, 2665668)
Tensor  39 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      2304     / 0.00 [576] [3494608, 3496912)
Tensor  40 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     112896   / 0.11 [1,14,14,576] [150528, 263424)
Tensor  41 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      55296    / 0.05 [576,1,1,96] [831952, 887248)
Tensor  42 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      384      / 0.00 [96] [3499240, 3499624)
Tensor  43 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     18816    / 0.02 [1,14,14,96] [150528, 169344)
Tensor  44 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      55296    / 0.05 [96,1,1,576] [3420476, 3475772)
Tensor  45 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     18816    / 0.02 [1,14,14,96] [263424, 282240)
Tensor  46 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     112896   / 0.11 [1,14,14,576] [263424, 376320)
Tensor  47 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      2304     / 0.00 [576] [3501952, 3504256)
Tensor  48 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      5184     / 0.00 [1,3,3,576] [887260, 892444)
Tensor  49 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      2304     / 0.00 [576] [3499636, 3501940)
Tensor  50 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     112896   / 0.11 [1,14,14,576] [150528, 263424)
Tensor  51 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      55296    / 0.05 [576,1,1,96] [1114764, 1170060)
Tensor  52 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      384      / 0.00 [96] [3504268, 3504652)
Tensor  53 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     18816    / 0.02 [1,14,14,96] [376320, 395136)
Tensor  54 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      55296    / 0.05 [96,1,1,576] [775768, 831064)
Tensor  55 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     28224    / 0.03 [1,7,7,576] [263424, 291648)
Tensor  56 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      2304     / 0.00 [576] [3506980, 3509284)
Tensor  57 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      5184     / 0.00 [1,3,3,576] [715260, 720444)
Tensor  58 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      2304     / 0.00 [576] [3504664, 3506968)
Tensor  59 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     112896   / 0.11 [1,14,14,576] [150528, 263424)
Tensor  60 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      55296    / 0.05 [576,1,1,96] [720460, 775756)
Tensor  61 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      640      / 0.00 [160] [3509296, 3509936)
Tensor  62 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     7840     / 0.01 [1,7,7,160] [244608, 252448)
Tensor  63 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      92160    / 0.09 [160,1,1,576] [623088, 715248)
Tensor  64 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     7840     / 0.01 [1,7,7,160] [252480, 260320)
Tensor  65 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     47040    / 0.04 [1,7,7,960] [197568, 244608)
Tensor  66 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      3840     / 0.00 [960] [3513800, 3517640)
Tensor  67 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      8640     / 0.01 [1,3,3,960] [1227904, 1236544)
Tensor  68 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      3840     / 0.00 [960] [3509948, 3513788)
Tensor  69 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     47040    / 0.04 [1,7,7,960] [150528, 197568)
Tensor  70 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      153600   / 0.15 [960,1,1,160] [469476, 623076)
Tensor  71 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      640      / 0.00 [160] [3517652, 3518292)
Tensor  72 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     7840     / 0.01 [1,7,7,160] [150528, 158368)
Tensor  73 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      153600   / 0.15 [160,1,1,960] [892456, 1046056)
Tensor  74 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     7840     / 0.01 [1,7,7,160] [197568, 205408)
Tensor  75 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     47040    / 0.04 [1,7,7,960] [197568, 244608)
Tensor  76 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      3840     / 0.00 [960] [3522156, 3525996)
Tensor  77 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      8640     / 0.01 [1,3,3,960] [2557116, 2565756)
Tensor  78 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      3840     / 0.00 [960] [3518304, 3522144)
Tensor  79 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     47040    / 0.04 [1,7,7,960] [150528, 197568)
Tensor  80 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      153600   / 0.15 [960,1,1,160] [3264524, 3418124)
Tensor  81 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      640      / 0.00 [160] [3526008, 3526648)
Tensor  82 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     7840     / 0.01 [1,7,7,160] [244608, 252448)
Tensor  83 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      153600   / 0.15 [160,1,1,960] [3110124, 3263724)
Tensor  84 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     47040    / 0.04 [1,7,7,960] [197568, 244608)
Tensor  85 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      3840     / 0.00 [960] [3530512, 3534352)
Tensor  86 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      8640     / 0.01 [1,3,3,960] [307212, 315852)
Tensor  87 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      3840     / 0.00 [960] [3526660, 3530500)
Tensor  88 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     47040    / 0.04 [1,7,7,960] [150528, 197568)
Tensor  89 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      153600   / 0.15 [960,1,1,160] [315864, 469464)
Tensor  90 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      1280     / 0.00 [320] [3534364, 3535644)
Tensor  91 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     15680    / 0.01 [1,7,7,320] [244608, 260288)
Tensor  92 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      307200   / 0.29 [320,1,1,960] [0, 307200)
Tensor  93 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     75264    / 0.07 [1,56,56,24] [602112, 677376)
Tensor  94 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     451584   / 0.43 [1,56,56,144] [602112, 1053696)
Tensor  95 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      576      / 0.00 [144] [3478516, 3479092)
Tensor  96 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      1296     / 0.00 [1,3,3,144] [1223128, 1224424)
Tensor  97 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      576      / 0.00 [144] [3477928, 3478504)
Tensor  98 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     451584   / 0.43 [1,56,56,144] [150528, 602112)
Tensor  99 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      3456     / 0.00 [144,1,1,24] [1224436, 1227892)
Tensor 100 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      96       / 0.00 [24] [3479104, 3479200)
Tensor 101 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     75264    / 0.07 [1,56,56,24] [150528, 225792)
Tensor 102 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      3456     / 0.00 [24,1,1,144] [1219660, 1223116)
Tensor 103 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     112896   / 0.11 [1,28,28,144] [602112, 715008)
Tensor 104 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      576      / 0.00 [144] [3479212, 3479788)
Tensor 105 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      1296     / 0.00 [1,3,3,144] [1218352, 1219648)
Tensor 106 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      576      / 0.00 [144] [3106060, 3106636)
Tensor 107 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     451584   / 0.43 [1,56,56,144] [150528, 602112)
Tensor 108 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      3456     / 0.00 [144,1,1,24] [1214880, 1218336)
Tensor 109 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      128      / 0.00 [32] [3479800, 3479928)
Tensor 110 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     25088    / 0.02 [1,28,28,32] [451584, 476672)
Tensor 111 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      4608     / 0.00 [32,1,1,144] [1210260, 1214868)
Tensor 112 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     25088    / 0.02 [1,28,28,32] [476672, 501760)
Tensor 113 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     150528   / 0.14 [1,28,28,192] [301056, 451584)
Tensor 114 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      768      / 0.00 [192] [3480720, 3481488)
Tensor 115 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      1728     / 0.00 [1,3,3,192] [1202364, 1204092)
Tensor 116 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      768      / 0.00 [192] [3479940, 3480708)
Tensor 117 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     150528   / 0.14 [1,28,28,192] [150528, 301056)
Tensor 118 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      6144     / 0.01 [192,1,1,32] [1204104, 1210248)
Tensor 119 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      128      / 0.00 [32] [3481500, 3481628)
Tensor 120 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     25088    / 0.02 [1,28,28,32] [150528, 175616)
Tensor 121 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      6144     / 0.01 [32,1,1,192] [2665684, 2671828)
Tensor 122 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     25088    / 0.02 [1,28,28,32] [301056, 326144)
Tensor 123 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     150528   / 0.14 [1,28,28,192] [301056, 451584)
Tensor 124 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      768      / 0.00 [192] [3481640, 3482408)
Tensor 125 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      1728     / 0.00 [1,3,3,192] [1113024, 1114752)
Tensor 126 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      768      / 0.00 [192] [3263736, 3264504)
Tensor 127 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     150528   / 0.14 [1,28,28,192] [150528, 301056)
Tensor 128 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      6144     / 0.01 [192,1,1,32] [1196208, 1202352)
Tensor 129 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      128      / 0.00 [32] [3482420, 3482548)
Tensor 130 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     25088    / 0.02 [1,28,28,32] [451584, 476672)
Tensor 131 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      6144     / 0.01 [32,1,1,192] [2572016, 2578160)
Tensor 132 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     37632    / 0.04 [1,14,14,192] [301056, 338688)
Tensor 133 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      768      / 0.00 [192] [3483340, 3484108)
Tensor 134 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      1728     / 0.00 [1,3,3,192] [1098980, 1100708)
Tensor 135 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      768      / 0.00 [192] [3482560, 3483328)
Tensor 136 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     150528   / 0.14 [1,28,28,192] [150528, 301056)
Tensor 137 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      6144     / 0.01 [192,1,1,32] [2565836, 2571980)
Tensor 138 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      256      / 0.00 [64] [1098712, 1098968)
Tensor 139 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     12544    / 0.01 [1,14,14,64] [338688, 351232)
Tensor 140 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      12288    / 0.01 [64,1,1,192] [1100720, 1113008)
Tensor 141 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     12544    / 0.01 [1,14,14,64] [313600, 326144)
Tensor 142 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     75264    / 0.07 [1,14,14,384] [225792, 301056)
Tensor 143 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      1536     / 0.00 [384] [1194660, 1196196)
Tensor 144 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      3456     / 0.00 [1,3,3,384] [1095244, 1098700)
Tensor 145 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      1536     / 0.00 [384] [3484120, 3485656)
Tensor 146 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     75264    / 0.07 [1,14,14,384] [150528, 225792)
Tensor 147 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      24576    / 0.02 [384,1,1,64] [2671844, 2696420)
Tensor 148 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      256      / 0.00 [64] [3485668, 3485924)
Tensor 149 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     12544    / 0.01 [1,14,14,64] [301056, 313600)
Tensor 150 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      24576    / 0.02 [64,1,1,384] [1070656, 1095232)
Tensor 151 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     12544    / 0.01 [1,14,14,64] [326144, 338688)
Tensor 152 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     75264    / 0.07 [1,14,14,384] [225792, 301056)
Tensor 153 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      1536     / 0.00 [384] [3487484, 3489020)
Tensor 154 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      3456     / 0.00 [1,3,3,384] [3106652, 3110108)
Tensor 155 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      1536     / 0.00 [384] [3485936, 3487472)
Tensor 156 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     75264    / 0.07 [1,14,14,384] [150528, 225792)
Tensor 157 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      24576    / 0.02 [384,1,1,64] [1046068, 1070644)
Tensor 158 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      256      / 0.00 [64] [3489032, 3489288)
Tensor 159 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     12544    / 0.01 [1,14,14,64] [301056, 313600)
Tensor 160 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      24576    / 0.02 [64,1,1,384] [2604328, 2628904)
Tensor 161 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     12544    / 0.01 [1,14,14,64] [313600, 326144)
Tensor 162 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     75264    / 0.07 [1,14,14,384] [225792, 301056)
Tensor 163 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      1536     / 0.00 [384] [3490848, 3492384)
Tensor 164 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      3456     / 0.00 [1,3,3,384] [2632392, 2635848)
Tensor 165 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      1536     / 0.00 [384] [3489300, 3490836)
Tensor 166 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     75264    / 0.07 [1,14,14,384] [150528, 225792)
Tensor 167 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      24576    / 0.02 [384,1,1,64] [2635868, 2660444)
Tensor 168 MobilenetV2/expanded_c... kTfLiteInt32    kTfLiteMmapRo      256      / 0.00 [64] [3492396, 3492652)
Tensor 169 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteArenaRw     12544    / 0.01 [1,14,14,64] [301056, 313600)
Tensor 170 MobilenetV2/expanded_c... kTfLiteUInt8    kTfLiteMmapRo      24576    / 0.02 [64,1,1,384] [2578188, 2602764)
Tensor 171 input                     kTfLiteUInt8    kTfLiteArenaRw     150528   / 0.14 [1,224,224,3] [0, 150528)
Tensor 172 output                    kTfLiteUInt8    kTfLiteArenaRw     1001     / 0.00 [1,1001] [151552, 152553)
Tensor 173 (nil)                     kTfLiteUInt8    kTfLiteArenaRw     338688   / 0.32 [1,112,112,27] [551936, 890624)

kTfLiteArenaRw Info: 
Tensor 21 has the max size 1204224 bytes (1.148 MB).
This memory arena is estimated as[0xf7107440, 0xf6f73040), taking 1655808 bytes (1.579 MB).
One possible set of tensors that have non-overlapping memory spaces with each other, and they take up the whole arena:
Tensor 171 -> 21 -> 17.

kTfLiteArenaRwPersistent Info: not holding any allocation.

kTfLiteMmapRo Info: 
Tensor 9 has the max size 1281280 bytes (1.222 MB).
This memory arena is estimated as[0xf7635a04, 0xf72d4318), taking 3544812 bytes (3.381 MB).
One possible set of tensors that have non-overlapping memory spaces with each other, and they take up the whole arena:
Tensor 92 -> 86 -> 89 -> 70 -> 63 -> 57 -> 60 -> 54 -> 19 -> 41 -> 48 -> 73 -> 157 -> 150 -> 144 -> 138 -> 134 -> 140 -> 125 -> 51 -> 31 -> 143 -> 128 -> 115 -> 118 -> 111 -> 108 -> 105 -> 102 -> 96 -> 99 -> 67 -> 34 -> 22 -> 13 -> 16 -> 9 -> 77 -> 137 -> 131 -> 170 -> 27 -> 160 -> 28 -> 164 -> 167 -> 38 -> 121 -> 147 -> 5 -> 106 -> 154 -> 83 -> 126 -> 80 -> 25 -> 44 -> 20 -> 2 -> 0 -> 12 -> 14 -> 18 -> 23 -> 97 -> 95 -> 100 -> 104 -> 109 -> 116 -> 114 -> 119 -> 124 -> 129 -> 135 -> 133 -> 145 -> 148 -> 155 -> 153 -> 158 -> 165 -> 163 -> 168 -> 29 -> 32 -> 39 -> 37 -> 42 -> 49 -> 47 -> 52 -> 58 -> 56 -> 61 -> 68 -> 66 -> 71 -> 78 -> 76 -> 81 -> 87 -> 85 -> 90 -> 3 -> 8 -> 10.

Node   0 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[171,2,0] -> 151520B (0.14MB)
  1 Output Tensors:[1] -> 401408B (0.38MB)
  1 Temporary Tensors:[173] -> 338688B (0.32MB)
Node   1 Operator Builtin Code   4 DEPTHWISE_CONV_2D (not delegated)
  3 Input Tensors:[1,13,12] -> 401824B (0.38MB)
  1 Output Tensors:[11] -> 401408B (0.38MB)
Node   2 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[11,16,14] -> 401984B (0.38MB)
  1 Output Tensors:[15] -> 200704B (0.19MB)
Node   3 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[15,22,20] -> 202624B (0.19MB)
  1 Output Tensors:[21] -> 1204224B (1.15MB)
Node   4 Operator Builtin Code   4 DEPTHWISE_CONV_2D (not delegated)
  3 Input Tensors:[21,19,18] -> 1205472B (1.15MB)
  1 Output Tensors:[17] -> 301056B (0.29MB)
Node   5 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[17,25,23] -> 303456B (0.29MB)
  1 Output Tensors:[24] -> 75264B (0.07MB)
Node   6 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[24,99,97] -> 79296B (0.08MB)
  1 Output Tensors:[98] -> 451584B (0.43MB)
Node   7 Operator Builtin Code   4 DEPTHWISE_CONV_2D (not delegated)
  3 Input Tensors:[98,96,95] -> 453456B (0.43MB)
  1 Output Tensors:[94] -> 451584B (0.43MB)
Node   8 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[94,102,100] -> 455136B (0.43MB)
  1 Output Tensors:[101] -> 75264B (0.07MB)
Node   9 Operator Builtin Code   0 ADD (not delegated)
  2 Input Tensors:[101,24] -> 150528B (0.14MB)
  1 Output Tensors:[93] -> 75264B (0.07MB)
Node  10 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[93,108,106] -> 79296B (0.08MB)
  1 Output Tensors:[107] -> 451584B (0.43MB)
Node  11 Operator Builtin Code   4 DEPTHWISE_CONV_2D (not delegated)
  3 Input Tensors:[107,105,104] -> 453456B (0.43MB)
  1 Output Tensors:[103] -> 112896B (0.11MB)
Node  12 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[103,111,109] -> 117632B (0.11MB)
  1 Output Tensors:[110] -> 25088B (0.02MB)
Node  13 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[110,118,116] -> 32000B (0.03MB)
  1 Output Tensors:[117] -> 150528B (0.14MB)
Node  14 Operator Builtin Code   4 DEPTHWISE_CONV_2D (not delegated)
  3 Input Tensors:[117,115,114] -> 153024B (0.15MB)
  1 Output Tensors:[113] -> 150528B (0.14MB)
Node  15 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[113,121,119] -> 156800B (0.15MB)
  1 Output Tensors:[120] -> 25088B (0.02MB)
Node  16 Operator Builtin Code   0 ADD (not delegated)
  2 Input Tensors:[120,110] -> 50176B (0.05MB)
  1 Output Tensors:[112] -> 25088B (0.02MB)
Node  17 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[112,128,126] -> 32000B (0.03MB)
  1 Output Tensors:[127] -> 150528B (0.14MB)
Node  18 Operator Builtin Code   4 DEPTHWISE_CONV_2D (not delegated)
  3 Input Tensors:[127,125,124] -> 153024B (0.15MB)
  1 Output Tensors:[123] -> 150528B (0.14MB)
Node  19 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[123,131,129] -> 156800B (0.15MB)
  1 Output Tensors:[130] -> 25088B (0.02MB)
Node  20 Operator Builtin Code   0 ADD (not delegated)
  2 Input Tensors:[130,112] -> 50176B (0.05MB)
  1 Output Tensors:[122] -> 25088B (0.02MB)
Node  21 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[122,137,135] -> 32000B (0.03MB)
  1 Output Tensors:[136] -> 150528B (0.14MB)
Node  22 Operator Builtin Code   4 DEPTHWISE_CONV_2D (not delegated)
  3 Input Tensors:[136,134,133] -> 153024B (0.15MB)
  1 Output Tensors:[132] -> 37632B (0.04MB)
Node  23 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[132,140,138] -> 50176B (0.05MB)
  1 Output Tensors:[139] -> 12544B (0.01MB)
Node  24 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[139,147,145] -> 38656B (0.04MB)
  1 Output Tensors:[146] -> 75264B (0.07MB)
Node  25 Operator Builtin Code   4 DEPTHWISE_CONV_2D (not delegated)
  3 Input Tensors:[146,144,143] -> 80256B (0.08MB)
  1 Output Tensors:[142] -> 75264B (0.07MB)
Node  26 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[142,150,148] -> 100096B (0.10MB)
  1 Output Tensors:[149] -> 12544B (0.01MB)
Node  27 Operator Builtin Code   0 ADD (not delegated)
  2 Input Tensors:[149,139] -> 25088B (0.02MB)
  1 Output Tensors:[141] -> 12544B (0.01MB)
Node  28 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[141,157,155] -> 38656B (0.04MB)
  1 Output Tensors:[156] -> 75264B (0.07MB)
Node  29 Operator Builtin Code   4 DEPTHWISE_CONV_2D (not delegated)
  3 Input Tensors:[156,154,153] -> 80256B (0.08MB)
  1 Output Tensors:[152] -> 75264B (0.07MB)
Node  30 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[152,160,158] -> 100096B (0.10MB)
  1 Output Tensors:[159] -> 12544B (0.01MB)
Node  31 Operator Builtin Code   0 ADD (not delegated)
  2 Input Tensors:[159,141] -> 25088B (0.02MB)
  1 Output Tensors:[151] -> 12544B (0.01MB)
Node  32 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[151,167,165] -> 38656B (0.04MB)
  1 Output Tensors:[166] -> 75264B (0.07MB)
Node  33 Operator Builtin Code   4 DEPTHWISE_CONV_2D (not delegated)
  3 Input Tensors:[166,164,163] -> 80256B (0.08MB)
  1 Output Tensors:[162] -> 75264B (0.07MB)
Node  34 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[162,170,168] -> 100096B (0.10MB)
  1 Output Tensors:[169] -> 12544B (0.01MB)
Node  35 Operator Builtin Code   0 ADD (not delegated)
  2 Input Tensors:[169,151] -> 25088B (0.02MB)
  1 Output Tensors:[161] -> 12544B (0.01MB)
Node  36 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[161,31,29] -> 38656B (0.04MB)
  1 Output Tensors:[30] -> 75264B (0.07MB)
Node  37 Operator Builtin Code   4 DEPTHWISE_CONV_2D (not delegated)
  3 Input Tensors:[30,28,27] -> 80256B (0.08MB)
  1 Output Tensors:[26] -> 75264B (0.07MB)
Node  38 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[26,34,32] -> 112512B (0.11MB)
  1 Output Tensors:[33] -> 18816B (0.02MB)
Node  39 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[33,41,39] -> 76416B (0.07MB)
  1 Output Tensors:[40] -> 112896B (0.11MB)
Node  40 Operator Builtin Code   4 DEPTHWISE_CONV_2D (not delegated)
  3 Input Tensors:[40,38,37] -> 120384B (0.11MB)
  1 Output Tensors:[36] -> 112896B (0.11MB)
Node  41 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[36,44,42] -> 168576B (0.16MB)
  1 Output Tensors:[43] -> 18816B (0.02MB)
Node  42 Operator Builtin Code   0 ADD (not delegated)
  2 Input Tensors:[43,33] -> 37632B (0.04MB)
  1 Output Tensors:[35] -> 18816B (0.02MB)
Node  43 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[35,51,49] -> 76416B (0.07MB)
  1 Output Tensors:[50] -> 112896B (0.11MB)
Node  44 Operator Builtin Code   4 DEPTHWISE_CONV_2D (not delegated)
  3 Input Tensors:[50,48,47] -> 120384B (0.11MB)
  1 Output Tensors:[46] -> 112896B (0.11MB)
Node  45 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[46,54,52] -> 168576B (0.16MB)
  1 Output Tensors:[53] -> 18816B (0.02MB)
Node  46 Operator Builtin Code   0 ADD (not delegated)
  2 Input Tensors:[53,35] -> 37632B (0.04MB)
  1 Output Tensors:[45] -> 18816B (0.02MB)
Node  47 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[45,60,58] -> 76416B (0.07MB)
  1 Output Tensors:[59] -> 112896B (0.11MB)
Node  48 Operator Builtin Code   4 DEPTHWISE_CONV_2D (not delegated)
  3 Input Tensors:[59,57,56] -> 120384B (0.11MB)
  1 Output Tensors:[55] -> 28224B (0.03MB)
Node  49 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[55,63,61] -> 121024B (0.12MB)
  1 Output Tensors:[62] -> 7840B (0.01MB)
Node  50 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[62,70,68] -> 165280B (0.16MB)
  1 Output Tensors:[69] -> 47040B (0.04MB)
Node  51 Operator Builtin Code   4 DEPTHWISE_CONV_2D (not delegated)
  3 Input Tensors:[69,67,66] -> 59520B (0.06MB)
  1 Output Tensors:[65] -> 47040B (0.04MB)
Node  52 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[65,73,71] -> 201280B (0.19MB)
  1 Output Tensors:[72] -> 7840B (0.01MB)
Node  53 Operator Builtin Code   0 ADD (not delegated)
  2 Input Tensors:[72,62] -> 15680B (0.01MB)
  1 Output Tensors:[64] -> 7840B (0.01MB)
Node  54 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[64,80,78] -> 165280B (0.16MB)
  1 Output Tensors:[79] -> 47040B (0.04MB)
Node  55 Operator Builtin Code   4 DEPTHWISE_CONV_2D (not delegated)
  3 Input Tensors:[79,77,76] -> 59520B (0.06MB)
  1 Output Tensors:[75] -> 47040B (0.04MB)
Node  56 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[75,83,81] -> 201280B (0.19MB)
  1 Output Tensors:[82] -> 7840B (0.01MB)
Node  57 Operator Builtin Code   0 ADD (not delegated)
  2 Input Tensors:[82,64] -> 15680B (0.01MB)
  1 Output Tensors:[74] -> 7840B (0.01MB)
Node  58 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[74,89,87] -> 165280B (0.16MB)
  1 Output Tensors:[88] -> 47040B (0.04MB)
Node  59 Operator Builtin Code   4 DEPTHWISE_CONV_2D (not delegated)
  3 Input Tensors:[88,86,85] -> 59520B (0.06MB)
  1 Output Tensors:[84] -> 47040B (0.04MB)
Node  60 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[84,92,90] -> 355520B (0.34MB)
  1 Output Tensors:[91] -> 15680B (0.01MB)
Node  61 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[91,5,3] -> 430400B (0.41MB)
  1 Output Tensors:[4] -> 62720B (0.06MB)
Node  62 Operator Builtin Code   1 AVERAGE_POOL_2D (not delegated)
  1 Input Tensors:[4] -> 62720B (0.06MB)
  1 Output Tensors:[6] -> 1280B (0.00MB)
Node  63 Operator Builtin Code   3 CONV_2D (not delegated)
  3 Input Tensors:[6,9,8] -> 1286564B (1.23MB)
  1 Output Tensors:[7] -> 1001B (0.00MB)
Node  64 Operator Builtin Code  22 RESHAPE (not delegated)
  2 Input Tensors:[7,10] -> 1009B (0.00MB)
  1 Output Tensors:[172] -> 1001B (0.00MB)

Execution plan as the list of 65 nodes invoked in-order: [0-64]
--------------Subgraph-0 dump has completed--------------

INFO: wanted width, height, channels: 224, 224, 3
INFO: output_number_of_pixels@@@@@@@@@@@@@@@@@@@:150528
pinput11^^^^^^^^^^^^^^^^^^^^^^^typed_tensor(0):(nil)
typed_tensor(171):0xf6f73040
pinput11  赋值完毕^^^^^^^^^^^^^^^^^^^^^^^INFO: invoked
INFO: average time: 84.954 ms
INFO: 0.717647: 653 military uniform
INFO: 0.564706: 835 suit
INFO: 0.533333: 458 bow tie
INFO: 0.52549: 907 Windsor tie
INFO: 0.513726: 753 racket

参考文章：
[1]:https://blog.csdn.net/m0_59824404/article/details/123328263
[2]: http://blog.cvosrobot.com/?post=460

转载请注明：文章转载自 www.mshxw.com
本文地址：https://www.mshxw.com/it/931874.html

Python相关栏目本月热门文章

热门相关搜索
路由器设置 木托盘 宝塔面板 儿童python教程 心情低落 朋友圈 vim 双一流学科 专升本 我的学校 日记学校 西点培训学校 汽修学校 情书 化妆学校 塔沟武校 异形模板 西南大学排名 最精辟人生短句 6步教你追回被骗的钱 南昌大学排名 清朝十二帝 北京印刷学院排名 北方工业大学排名 北京航空航天大学排名 首都经济贸易大学排名 中国传媒大学排名 首都师范大学排名 中国地质大学(北京)排名 北京信息科技大学排名 中央民族大学排名 北京舞蹈学院排名 北京电影学院排名 中国戏曲学院排名 河北政法职业学院排名 河北经贸大学排名 天津中德应用技术大学排名 天津医学高等专科学校排名 天津美术学院排名 天津音乐学院排名 天津工业大学排名 北京工业大学耿丹学院排名 北京警察学院排名 天津科技大学排名 北京邮电大学(宏福校区)排名 北京网络职业学院排名 北京大学医学部排名 河北科技大学排名 河北地质大学排名 河北体育学院排名
学习工具
代数计算器
三角函数
解析几何
立体几何
知识解答
教育知识
百科知识
生活知识
常识知识
写作必备
作文大全
作文素材
句子大全
实用范文
关于我们
关于我们
联系我们
网站地图
交流群

名师互学网交流群

名师互学网客服

名师互学网客服

名师互学网 版权所有 (c)2021-2022 ICP备案号：晋ICP备2021003244-6号

我们一直用心在做
关于我们 文章归档 网站地图 联系我们

版权所有 (c)2021-2022 MSHXW.COM

ICP备案号：晋ICP备2021003244-6号