【MNN学习一】基于Linux的MNN编译安装，模型转换以及benchmark测试

目录

一. MNN编译安装 (MNN编译)

二. 编译模型转换工具MNNConvert (MNNConvert)

三. 下载模型并使用MNNConvert (模型下载&转换)

四. 测试benchmark

---

一. MNN编译安装 (MNN编译)

1. 编译选项

cmake_minimum_required(VERSION 2.8)

project(MNN)

# complier options
set(CMAKE_C_STANDARD 99)
set(CMAKE_CXX_STANDARD 11)
enable_language(ASM)
# set(CMAKE_C_COMPILER gcc)
# set(CMAKE_CXX_COMPILER g++)

option(MNN_USE_CPP11 "Enable MNN use c++11" ON)

if (NOT MSVC)
    if(MNN_USE_CPP11)
        set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -std=gnu99")
        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11")
    else()
        set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -std=gnu99")
        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++0x")
    endif()
endif()

# build options
option(MNN_BUILD_HARD "Build -mfloat-abi=hard or not" OFF)
option(MNN_BUILD_SHARED_LIBS "MNN build shared or static lib" ON)
option(MNN_FORBID_MULTI_THREAD "Disable Multi Thread" OFF)
option(MNN_OPENMP "Enable Multiple Thread Linux|Android" ON)
option(MNN_USE_THREAD_POOL "Use Multiple Thread by Self ThreadPool" OFF)
option(MNN_SUPPORT_TRAIN "Enable Train Ops" ON)
option(MNN_BUILD_DEMO "Build demo/exec or not" ON)
option(MNN_BUILD_QUANTOOLS "Build Quantized Tools or not" ON)
option(MNN_USE_INT8_FAST "Build int8 or not" ON)

if (MNN_USE_THREAD_POOL)
    set(MNN_OPENMP OFF)
    add_definitions(-DMNN_USE_THREAD_POOL)
endif()

if(MNN_FORBID_MULTI_THREAD)
    add_definitions(-DMNN_FORBIT_MULTI_THREADS)
endif()
if(MNN_USE_INT8_FAST)
    add_definitions(-DMNN_USE_INT8_FAST)
endif()

# debug options
option(MNN_DEBUG "Enable MNN DEBUG" OFF)
if(CMAKE_BUILD_TYPE MATCHES "Debug")
    set(MNN_DEBUG ON)
endif()
option(MNN_DEBUG_MEMORY "MNN Debug Memory Access" OFF)
option(MNN_DEBUG_TENSOR_SIZE "Enable Tensor Size" OFF )
option(MNN_GPU_TRACE "Enable MNN Gpu Debug" OFF)

if(MNN_DEBUG_MEMORY)
    add_definitions(-DMNN_DEBUG_MEMORY)
endif()
if(MNN_DEBUG_TENSOR_SIZE)
    add_definitions(-DMNN_DEBUG_TENSOR_SIZE)
endif()
if(MNN_GPU_TRACE)
    add_definitions(-DMNN_GPU_FORCE_FINISH)
endif()

# backend options
option(MNN_METAL "Enable Metal" OFF)
option(MNN_OPENCL "Enable OpenCL" ON)
option(MNN_OPENGL "Enable OpenGL" OFF)
option(MNN_VULKAN "Enable Vulkan" OFF)
option(MNN_ARM82 "Enable ARM82" ON)

# codegen register ops
if (MNN_METAL)
    add_definitions(-DMNN_CODEGEN_REGISTER)
endif()

# target options
option(MNN_BUILD_BENCHMARK "Build benchmark or not" ON)
option(MNN_BUILD_TEST "Build tests or not" ON)
option(MNN_BUILD_FOR_ANDROID_COMMAND "Build from command" OFF)
option(MNN_BUILD_FOR_IOS "Build for ios" OFF)
set (MNN_HIDDEN FALSE)
if (NOT MNN_BUILD_TEST)
    if (NOT MNN_DEBUG)
        set (MNN_HIDDEN TRUE)
    endif()
endif()

include(cmake/macros.cmake)

message(STATUS ">>>>>>>>>>>>>")
message(STATUS "MNN BUILD INFO:")
message(STATUS "\tSystem: ${CMAKE_SYSTEM_NAME}")
message(STATUS "\tProcessor: ${CMAKE_SYSTEM_PROCESSOR}")
message(STATUS "\tDEBUG: ${MNN_DEBUG}")
message(STATUS "\tMetal: ${MNN_METAL}")
message(STATUS "\tOpenCL: ${MNN_OPENCL}")
message(STATUS "\tOpenGL: ${MNN_OPENGL}")
message(STATUS "\tVulkan: ${MNN_VULKAN}")
message(STATUS "\tOpenMP: ${MNN_OPENMP}")
message(STATUS "\tHideen: ${MNN_HIDDEN}")

.
.
.
.

2. 本地编译

    在这一步中先忽略第四小步，在MNNConvert安装以后再./tools/script/get_model.sh

二. 编译模型转换工具MNNConvert (MNNConvert)

1. 安装protobuf(3.0以上) (Protocol Buffers - Google's data interchange format)

sudo apt-get install autoconf automake libtool curl make g++ unzip

git clone https://github.com/protocolbuffers/protobuf.git
cd protobuf
git submodule update --init --recursive
./autogen.sh

./configure
make
make check
sudo make install
sudo ldconfig

2. 编译MNNConvert

以下两种方法可用：

    (1) 执行以下命令：

cd MNN/tools/converter
./generate_schema.sh
mkdir build
cd build && cmake .. && make -j4

    (2) 在本地(rk3399)编译时,先修改build_tool.sh中的 make -j16

#!/bin/bash

if [ -d "build" ]; then
  rm -rf build
fi
./generate_schema.sh

mkdir build
cd build

cmake ..
make clean
make -j4        # # #

3. 测试MNNConvert (模型转换的使用)

cd MNN/tools/converter/build
./MNNConvert -h

4. 测试MNNDump2Json (MNNDump2Json)

    将MNN模型bin文件 dump 成可读的类json格式文件，以方便对比原始模型参数

cd MNN/tools/converter/build
./MNNDump2Json.out ../../../benchmark/models/mobilenet_v1.caffe.mnn mobilenet_v1.caffe.json

三. 下载模型并使用MNNConvert (模型下载&转换)

1. 下载并使用第二步得到的MNNConvert来转换模型

cd MNN
./tools/script/get_model.sh

    下载的模型在/MNN/resource/model/ (应该没记错^_^)

2. 使用MNNConvert

    (1) Caffe2MNN

cd MNN/resource/build
./MNNConvert -f CAFFE --modelFile mobilenet_v1.caffe.caffemodel --prototxt mobilenet_v1.caffe.prototxt --MNNModel mobilenet_v1.caffe.mnn --bizCode MNN

    (2) TFLite2MNN

cd MNN/resource/build
./MNNConvert -f TFLITE --modelFile mobilenet_v2_1.0_224_quant.tflite --MNNModel mobilenet_v2_1.0_224_quant.tflite.mnn --bizCode MNN

    将mobilenet_v1.caffe.mnn，mobilenet_v2_1.0_224_quant.tflite.mnn复制到/MNN/benchmark/models/目录下用于下面的测试。

四. 测试benchmark

cd MNN/build
./benchmark.out ../benchmark/models/ 10 0


# ------------------ output --------------------- #
MNN benchmark
Forward type: **CPU** thread=4** precision=2
--------> Benchmarking... loop = 10
[ - ] MobileNetV2_224.mnn         max =   92.516ms  min =   88.733ms  avg =   90.227ms
[ - ] mobilenet_v2_1.0_224_quant.tflite.mnn    max =   87.738ms  min =   84.411ms  avg =   85.817ms
[ - ] mobilenet_v1_quant.caffe.mnn    max =   75.278ms  min =   74.261ms  avg =   74.763ms
[ - ] inception-v3.mnn            max =  996.025ms  min =  807.543ms  avg =  863.333ms
[ - ] resnet-v2-50.mnn            max =  531.352ms  min =  445.743ms  avg =  473.887ms
[ - ] mobilenet-v1-1.0.mnn        max =  103.705ms  min =   93.981ms  avg =   96.573ms
[ - ] SqueezeNetV1.0.mnn          max =  193.273ms  min =  115.615ms  avg =  127.908ms
[ - ] mobilenet_v1.caffe.mnn      max =  107.799ms  min =   94.063ms  avg =   98.914ms

cd MNN/build
./benchmark.out ../benchmark/models/ 10 0


# ------------------ output --------------------- #
MNN benchmark
Forward type: **OpenCL** thread=4** precision=2
--------> Benchmarking... loop = 10
Can't Find type=3 backend, use 0 instead
[ - ] MobileNetV2_224.mnn         max =   92.939ms  min =   88.632ms  avg =   89.630ms
Can't Find type=3 backend, use 0 instead
[ - ] mobilenet_v2_1.0_224_quant.tflite.mnn    max =   89.197ms  min =   84.408ms  avg =   86.093ms
Can't Find type=3 backend, use 0 instead
[ - ] inception-v3.mnn            max =  870.511ms  min =  817.133ms  avg =  839.259ms
Can't Find type=3 backend, use 0 instead
[ - ] resnet-v2-50.mnn            max =  501.031ms  min =  448.233ms  avg =  462.990ms
Can't Find type=3 backend, use 0 instead
[ - ] mobilenet-v1-1.0.mnn        max =  107.135ms  min =   94.491ms  avg =   98.316ms
Can't Find type=3 backend, use 0 instead
[ - ] SqueezeNetV1.0.mnn          max =  124.856ms  min =  113.669ms  avg =  118.974ms
Can't Find type=3 backend, use 0 instead
[ - ] mobilenet_v1.caffe.mnn      max =  107.162ms  min =   94.707ms  avg =   98.140ms

    编译MNN时，选择option(MNN_OPENCL "Enable OpenCL" ON)，为啥出现不能用OpenCL！！！

---

参考链接：https://github.com/alibaba/MNN