Tensorflowlite移植ARM平台iMX6

一、LINUX环境下操作：

1.安装交叉编译SDK （仅针对该型号：i.MX6,不同芯片需要对应的交叉编译SDK）

编译方法参考：手动编译用于i.MX6系列的交叉编译SDK

2.下载Tensorflow

git clone https://github.com/tensorflow/tensorflow.git

cd tensorflow

git checkout r1.10

Tensorflow与Bazel编译器（及CUDA，CUDNN）之间需要对应，否则会有兼容性问题。

tensorflowr1.10     python 2.7,3.6   Bazel:0.18.0-0.19.2

tensorflowr1.12     python 2.7,3.6   Bazel:0.18.0-0.19.2

tensorflowr1.14     python 2.7,3.6   Bazel:0.24.0 - 0.25.2

3、下载并安装编译工具Bazel

安装依赖包：

sudo apt-get install pkg-config zip g++ zilb1g-dev unzip

下载Bazel包：

wget https://github.com/bazelbuild/bazel/releases/download/0.18.1/bazel-0.181-installer-linux-86_64.sh

安装Bazel：

chmod +x bazel-0.18.1-installer-linux-86_64.sh

./bazel-0.18.1-installer-linux-86_64.sh --user

设置环境变量：

sudo vi ~/.bashrc，在文件最后添加：export PATH=$PATH":~/bin"

source ~/.bashrc

（如果仅仅是测试DEMO在ARM板上使用，可直接跳过4，5，6，7，8步，直接进行第9步）

4、编译配置：

在Tensorflow源码根目录运行：

./configure     （编译LINUX平台时使用默认设置：-march=native，编译ARM平台时需设置成相应值：-march=armv7-a）

5、编译pip:

bazel build --config=opt //tensorflow/tools/pip_package:build_pip_package

7、编译包：

./bazel-bin/tensorflow/tools/pip__package/build_pip_package /tmp/tensorflow_pkg

8、安装包：

pip install /tmp/tensorflow_pkg/tensorflow-version-tags.whl

9、下载依赖库：

./tensorflow/contrib/lite/tools/make/download_dependencies.sh(不同版本，位置略有不同，本文路径为r1.10版本)

10、编译Tensorflow Lite:

方法1：（参考网上的，但我调试了几天，并未生成该库，可行性待验证！）

目前Tensorflow仅支持树莓派： ./tensorflow/contrib/lite/tools/make/build_rpi_lib.sh，该脚本的目标编译平台是ARMv7，即使目标平台是ARMv8，也不要更改。因为设置编译平台为ARMv7可以优化编译，提高运行速度。

所以，需要针对iMX6，制作一份build_imx6_lib.sh

 set -e 

 SCRIPT_DIR="$(cd "$(dirname "$(BASH_SOURCE[0]}")" && pwd)"
 cd "$SCRIPT_DIR/../../.."

 CC_PREFIX=arm-poky-linux-gnueabi- make -j 3 -f tensorflow/contrib/lite/Makefile TARGET-imx6 TARGET_ARCH=armv7-a

生成静态库位置为：

./tensorflow/contrib/lite/toos/make/gen/rpi_arm7l/lib/libtensorflow-lite.a静态库。

方法2：（使用ARM（iMX6）交叉编译包，编译出来的包可以用于iMX6平台）

1.打开Tensorflow/contrib/lite/kernels/internal/BUILD

 config_setting(
        name = "armv7a",
        values = {"cpu":"armv7a",},
 )

 "armv7a":[
     "-O3",
     "-mfpu=vfpv3",
     "-mfloat-abi=hard",
 ],

 add code above to NEON_FLAGS_IF_APPLICABLE

 ":armv6":[

     ":neon_tensor_utils",
     ],
 ":armv7a":[
     ":neon_tensor_utils",
 ],

 add code above to cc_library  tensor_utils select options

2.在根目录下创建一个文件：build_armv7_tflite.sh

 #!/bin/bash

 bazel build --copt="-fPIC" --copt="-march=armv6" \
      --copt="-mfloat-abi=hard" \
      --copt="-mfpu=vfpv3" \
     --copt="-funsafe-math-optimizations" \
      --copt="-Wno-unused-function" \
      --copt="-Wno-sign-compare" \
      --copt="-ftree-vectorize" \
      --copt="-fomit-frame-pointer" \
      --cxxopt='--std=c++11' \
      --cxxopt="-Wno-maybe-uninitialized" \
      --cxxopt="-Wno-narrowing" \
      --cxxopt="-Wno-unused" \
      --cxxopt="-Wno-comment" \
      --cxxopt="-Wno-unused-function" \
      --cxxopt="-Wno-sign-compare" \
      --cxxopt="-funsafe-math-optimizations" \
      --linkopt="-lstdc++" \
      --linkopt="-mfloat-abi=hard" \
      --linkopt="-mfpu=vfpv3" \
      --linkopt="-funsafe-math-optimizations" \
      --verbose_failures \
      --strip=always \
      --crosstool_top=//armv6-compiler:toolchain --cpu=armv7 --config=opt \
 33     tensorflow/contrib/lite/examples/label_image

3.编译该文件build_armv7_tflite.sh，会碰到一个错误：.../.../read-ld:unrecognized options : --icf=all

解决方法：找到文件./build_def.bzl ,打开，去除所有--icf=all标识的信息即可。

编译后，会在bazel_bin下生成指定的文件label_image。

4.生成头和库文件：

libbuiltin_ops.a libframework.a libneon_tensor_utils.a libquantization_util.a libtensor_utils.a

libcontext.a libfarmhash.a libgemm_support.a libportable_tensor_utils.a  libstring_util.a

11、编译模型：

默认情况下label_image并未编译进去，需要修改Makefile，可参考minimal APK,主要修改以下三部分内容：

LABELIMAGE_SRCS

LABELIMAGE_BINARY

LABEL_OBJS

 # Find where we're running from, so we can store generated files here.
 ifeq ($(origin MAKEFILE_DIR), undefined)
     MAKEFILE_DIR := $(shell dirname $(realpath $(lastword $(MAKEFILE_LIST))))
 endif

 # Try to figure out the host system
 HOST_OS :=
 ifeq ($(OS),Windows_NT)
     HOST_OS = WINDOWS
 else
     UNAME_S := $(shell uname -s)
     ifeq ($(UNAME_S),Linux)
             HOST_OS := LINUX
     endif
     ifeq ($(UNAME_S),Darwin)
         HOST_OS := OSX
     endif
 endif

 #HOST_ARCH := $(shell if [[ $(shell uname -m) =~ i[345678]86 ]]; then echo x86_32; else echo $(shell uname -m); fi)

 # Self-hosting
 TARGET_ARCH := ${HOST_ARCH}
 CROSS := imx6
 $(warning "CROSS :$(CROSS) HOST_ARCH:$(HOST_ARCH) TARGET_ARCH:$(TARGET_ARCH) TARGET_TOOLCHAIN_PREFIX:$(TARGET_TOOLCHAIN_PREFIX)")
 # Cross compiling
 ifeq ($(CROSS),imx6)
   TARGET_ARCH := armv7-a
   TARGET_TOOLCHAIN_PREFIX := arm-poky-linux-gnueabi-
 endif

 ifeq ($(CROSS),rpi)
   TARGET_ARCH := armv7l
   TARGET_TOOLCHAIN_PREFIX := arm-linux-gnueabihf-
 endif

 ifeq ($(CROSS),riscv)
   TARGET_ARCH := riscv
   TARGET_TOOLCHAIN_PREFIX := riscv32-unknown-elf-
 endif
 ifeq ($(CROSS),stm32f7)
   TARGET_ARCH := armf7
   TARGET_TOOLCHAIN_PREFIX := arm-none-eabi-
 endif
 ifeq ($(CROSS),stm32f1)
   TARGET_ARCH := armm1
   TARGET_TOOLCHAIN_PREFIX := arm-none-eabi-
 endif

 # Where compiled objects are stored.
 OBJDIR := $(MAKEFILE_DIR)/gen/obj/
 BINDIR := $(MAKEFILE_DIR)/gen/bin/
 LIBDIR := $(MAKEFILE_DIR)/gen/lib/
 GENDIR := $(MAKEFILE_DIR)/gen/obj/

 LIBS :=
 ifeq ($(TARGET_ARCH),x86_64)
         CXXFLAGS += -fPIC -DGEMMLOWP_ALLOW_SLOW_SCALAR_FALLBACK -pthread # -msse4.2
 endif

 ifeq ($(TARGET_ARCH),armv7-a)
         CXXFLAGS += -pthread -fPIC
     LIBS += -ldl
 endif

 ifeq ($(TARGET_ARCH),armv7l)
         CXXFLAGS += -mfpu=neon -pthread -fPIC
     LIBS += -ldl
 endif

 ifeq ($(TARGET_ARCH),riscv)
 #        CXXFLAGS += -march=gap8
         CXXFLAGS += -DTFLITE_MCU
     LIBS += -ldl
     BUILD_TYPE := micro
 endif

 ifeq ($(TARGET_ARCH),armf7)
         CXXFLAGS += -DGEMMLOWP_ALLOW_SLOW_SCALAR_FALLBACK -DTFLITE_MCU
         CXXFLAGS += -fno-rtti -fmessage-length=0 -fno-exceptions -fno-builtin -ffunction-sections -fdata-sections
         CXXFLAGS += -funsigned-char -MMD
         CXXFLAGS += -mcpu=cortex-m7 -mthumb -mfpu=fpv5-sp-d16 -mfloat-abi=softfp
         CXXFLAGS += '-std=gnu++11' '-fno-rtti' '-Wvla' '-c' '-Wall' '-Wextra' '-Wno-unused-parameter' '-Wno-missing-field-initializers' '-fmessage-length=0' '-fno-exceptions' '-fno-builtin' '-ffunction-sections' '-fdata-sections' '-funsigned-char' '-MMD' '-fno-delete-null-pointer-checks' '-fomit-frame-pointer' '-Os'
     LIBS += -ldl
     BUILD_TYPE := micro
 endif
 ifeq ($(TARGET_ARCH),armm1)
         CXXFLAGS += -DGEMMLOWP_ALLOW_SLOW_SCALAR_FALLBACK -mcpu=cortex-m1 -mthumb -DTFLITE_MCU
         CXXFLAGS += -fno-rtti -fmessage-length=0 -fno-exceptions -fno-builtin -ffunction-sections -fdata-sections
         CXXFLAGS += -funsigned-char -MMD
     LIBS += -ldl
 endif

 # Settings for the host compiler.
 #CXX := $(CC_PREFIX) ${TARGET_TOOLCHAIN_PREFIX}g++
 #CXX := ${TARGET_TOOLCHAIN_PREFIX}g++ -march=armv7-a -mfpu=neon  -mfloat-abi=hard -mcpu=cortex-a9 --sysroot=/opt/fsl-imx-x11/4.1.15-2.1.0/sysroots/x86_64-pokysdk-linux/usr/bin/arm-poky-linux-gnueabi
 CXXFLAGS += -O3 -DNDEBUG
 CCFLAGS := ${CXXFLAGS}
 #CXXFLAGS += --std=c++11
 CXXFLAGS += --std=c++0x     #wly
 #CC := ${TARGET_TOOLCHAIN_PREFIX}gcc
 #AR := ${TARGET_TOOLCHAIN_PREFIX}ar
 CFLAGS :=
 LDOPTS :=
 LDOPTS += -L/usr/local/lib
 ARFLAGS := -r

 INCLUDES := \
 -I. \
 -I$(MAKEFILE_DIR)/../../../ \
 -I$(MAKEFILE_DIR)/downloads/ \
 -I$(MAKEFILE_DIR)/downloads/eigen \
 -I$(MAKEFILE_DIR)/downloads/gemmlowp \
 -I$(MAKEFILE_DIR)/downloads/neon_2_sse \
 -I$(MAKEFILE_DIR)/downloads/farmhash/src \
 -I$(MAKEFILE_DIR)/downloads/flatbuffers/include \
 -I$(GENDIR)
 # This is at the end so any globally-installed frameworks like protobuf don't
 # override local versions in the source tree.
 #INCLUDES += -I/usr/local/include
 INCLUDES += -I/usr/include
 LIBS += \
 -lstdc++ \
 -lpthread \
 -lm \
 -lz

 # If we're on Linux, also link in the dl library.
 ifeq ($(HOST_OS),LINUX)
     LIBS += -ldl
 endif

 include $(MAKEFILE_DIR)/ios_makefile.inc
 ifeq ($(CROSS),rpi)
 #include $(MAKEFILE_DIR)/rpi_makefile.inc
 endif
 ifeq ($(CROSS),imx6)
 include $(MAKEFILE_DIR)/imx6_makefile.inc
 endif
 # This library is the main target for this makefile. It will contain a minimal
 # runtime that can be linked in to other programs.
 LIB_NAME := libtensorflow-lite.a
 LIB_PATH := $(LIBDIR)$(LIB_NAME)

 # A small example program that shows how to link against the library.
 MINIMAL_PATH := $(BINDIR)minimal
 LABEL_IMAGE_PATH :=$(BINDIR)label_image

 MINIMAL_SRCS := \
 tensorflow/contrib/lite/examples/minimal/minimal.cc
 MINIMAL_OBJS := $(addprefix $(OBJDIR), \
 $(patsubst %.cc,%.o,$(patsubst %.c,%.o,$(MINIMAL_SRCS))))

 LABEL_IMAGE_SRCS := \
 tensorflow/contrib/lite/examples/label_image/label_image.cc \
 tensorflow/contrib/lite/examples/label_image/bitmap_helpers.cc
 LABEL_IMAGE_OBJS := $(addprefix $(OBJDIR), \
 $(patsubst %.cc,%.o,$(patsubst %.c,%.o,$(LABEL_IMAGE_SRCS))))

 # What sources we want to compile, must be kept in sync with the main Bazel
 # build files.

 PROFILER_SRCS := \
     tensorflow/contrib/lite/profiling/time.cc
 PROFILE_SUMMARIZER_SRCS := \
     tensorflow/contrib/lite/profiling/profile_summarizer.cc \
     tensorflow/core/util/stats_calculator.cc

 CORE_CC_ALL_SRCS := \
 $(wildcard tensorflow/contrib/lite/*.cc) \
 $(wildcard tensorflow/contrib/lite/*.c)
 ifneq ($(BUILD_TYPE),micro)
 CORE_CC_ALL_SRCS += \
 $(wildcard tensorflow/contrib/lite/kernels/*.cc) \
 $(wildcard tensorflow/contrib/lite/kernels/internal/*.cc) \
 $(wildcard tensorflow/contrib/lite/kernels/internal/optimized/*.cc) \
 $(wildcard tensorflow/contrib/lite/kernels/internal/reference/*.cc) \
 $(PROFILER_SRCS) \
 $(wildcard tensorflow/contrib/lite/kernels/*.c) \
 $(wildcard tensorflow/contrib/lite/kernels/internal/*.c) \
 $(wildcard tensorflow/contrib/lite/kernels/internal/optimized/*.c) \
 $(wildcard tensorflow/contrib/lite/kernels/internal/reference/*.c) \
 $(wildcard tensorflow/contrib/lite/downloads/farmhash/src/farmhash.cc) \
 $(wildcard tensorflow/contrib/lite/downloads/fft2d/fftsg.c)
 endif
 # Remove any duplicates.
 CORE_CC_ALL_SRCS := $(sort $(CORE_CC_ALL_SRCS))
 CORE_CC_EXCLUDE_SRCS := \
 $(wildcard tensorflow/contrib/lite/*test.cc) \
 $(wildcard tensorflow/contrib/lite/*/*test.cc) \
 $(wildcard tensorflow/contrib/lite/*/*/*test.cc) \
 $(wildcard tensorflow/contrib/lite/*/*/*/*test.cc) \
 $(wildcard tensorflow/contrib/lite/kernels/test_util.cc) \
 $(MINIMAL_SRCS) \
 $(LABEL_IMAGE_SRCS)
 ifeq ($(BUILD_TYPE),micro)
 CORE_CC_EXCLUDE_SRCS += \
 tensorflow/contrib/lite/model.cc \
 tensorflow/contrib/lite/nnapi_delegate.cc
 endif
 # Filter out all the excluded files.
 TF_LITE_CC_SRCS := $(filter-out $(CORE_CC_EXCLUDE_SRCS), $(CORE_CC_ALL_SRCS))
 # File names of the intermediate files target compilation generates.
 TF_LITE_CC_OBJS := $(addprefix $(OBJDIR), \
 $(patsubst %.cc,%.o,$(patsubst %.c,%.o,$(TF_LITE_CC_SRCS))))
 LIB_OBJS := $(TF_LITE_CC_OBJS)

 # Benchmark sources
 BENCHMARK_SRCS_DIR := tensorflow/contrib/lite/tools/benchmark
 BENCHMARK_ALL_SRCS := $(TFLITE_CC_SRCS) \
     $(wildcard $(BENCHMARK_SRCS_DIR)/*.cc) \
     $(PROFILE_SUMMARIZER_SRCS)

 BENCHMARK_SRCS := $(filter-out \
     $(wildcard $(BENCHMARK_SRCS_DIR)/*_test.cc), \
     $(BENCHMARK_ALL_SRCS))

 BENCHMARK_OBJS := $(addprefix $(OBJDIR), \
 $(patsubst %.cc,%.o,$(patsubst %.c,%.o,$(BENCHMARK_SRCS))))

 # For normal manually-created TensorFlow C++ source files.
 $(OBJDIR)%.o: %.cc
     @mkdir -p $(dir $@)
     $(CXX) $(CXXFLAGS) $(INCLUDES) -c $< -o $@
 # For normal manually-created TensorFlow C++ source files.
 $(OBJDIR)%.o: %.c
     @mkdir -p $(dir $@)
     $(CC) $(CCFLAGS) $(INCLUDES) -c $< -o $@

 # The target that's compiled if there's no command-line arguments.
 all: $(LIB_PATH)  $(MINIMAL_PATH) $(LABEL_IMAGE_PATH) $(BENCHMARK_BINARY) 

 # The target that's compiled for micro-controllers
 micro: $(LIB_PATH)

 # Gathers together all the objects we've compiled into a single '.a' archive.
 $(LIB_PATH): $(LIB_OBJS)
     @mkdir -p $(dir $@)
     $(AR) $(ARFLAGS) $(LIB_PATH) $(LIB_OBJS)

 $(MINIMAL_PATH): $(MINIMAL_OBJS) $(LIB_PATH)
     @mkdir -p $(dir $@)
     $(CXX) $(CXXFLAGS) $(INCLUDES) \
     -o $(MINIMAL_PATH) $(MINIMAL_OBJS) \
     $(LIBFLAGS) $(LIB_PATH) $(LDFLAGS) $(LIBS)

 # $(LABEL_IMAGE_PATH): $(LABEL_IMAGE_OBJS) $(LIBS)

 $(LABEL_IMAGE_PATH) : $(LABEL_IMAGE_OBJS) $(LIB_PATH)
     @mkdir -p $(dir $@)
     $(CXX) $(CXXFLAGS) $(INCLUDES) \
     -o $(LABEL_IMAGE_PATH) $(LABEL_IMAGE_OBJS)\
     $(LIBFLAGS) $(LIB_PATH) $(LDFLAGS) $(LIBS)

 $(BENCHMARK_LIB) : $(LIB_PATH) $(BENCHMARK_OBJS)
     @mkdir -p $(dir $@)
     $(AR) $(ARFLAGS) $(BENCHMARK_LIB) $(LIB_OBJS) $(BENCHMARK_OBJS)

 benchmark_lib: $(BENCHMARK_LIB)
 $(info $(BENCHMARK_BINARY))
 $(BENCHMARK_BINARY) : $(BENCHMARK_LIB)
     @mkdir -p $(dir $@)
     $(CXX) $(CXXFLAGS) $(INCLUDES) \
     -o $(BENCHMARK_BINARY) \
     $(LIBFLAGS) $(BENCHMARK_LIB) $(LDFLAGS) $(LIBS)

 benchmark: $(BENCHMARK_BINARY)

 # Gets rid of all generated files.
 clean:
     rm -rf $(MAKEFILE_DIR)/gen

 # Gets rid of target files only, leaving the host alone. Also leaves the lib
 # directory untouched deliberately, so we can persist multiple architectures
 # across builds for iOS and Android.
 cleantarget:
     rm -rf $(OBJDIR)
     rm -rf $(BINDIR)

 $(DEPDIR)/%.d: ;
 .PRECIOUS: $(DEPDIR)/%.d

 -include $(patsubst %,$(DEPDIR)/%.d,$(basename $(TF_CC_SRCS)))

执行： ./tensorflow/contrib/lite/tools/make/build_imx6_lib.sh

编译过程中可能会出现一些问题，依次解决即可！

编译完成后，在./tensorflow/contrib/lite/tools/make/gen/rpi_arm7l/bin  目录下生成可执行文件label_img

附上自己碰到的一些问题：

1、a/tensorflow/contrib/lite/tools/benchmark/benchmark_params.h

 diff --git a/tensorflow/contrib/lite/profiling/profile_summarizer.cc b/tensorflow/contrib/lite/profiling/profile_summarizer.cc
 old mode 100644
 new mode 100755
 index c37a096..590cd21
 --- a/tensorflow/contrib/lite/profiling/profile_summarizer.cc
 +++ b/tensorflow/contrib/lite/profiling/profile_summarizer.cc
 @@ -83,7 +83,8 @@ OperatorDetails GetOperatorDetails(const tflite::Interpreter& interpreter,
    OperatorDetails details;
    details.name = op_name;
    if (profiling_string) {
 -    details.name += ":" + string(profiling_string);
 +    //wly
 +    details.name += ":" + std::string(profiling_string);
    }
    details.inputs = GetTensorNames(interpreter, inputs);
    details.outputs = GetTensorNames(interpreter, outputs);

2、./tensorflow/contrib/lite/profiling/profile_summarizer.cc

 diff --git a/tensorflow/contrib/lite/tools/benchmark/benchmark_params.h b/tensorflow/contrib/lite/tools/benchmark/benchmark_params.h
 old mode 100644
 new mode 100755
 index 33448dd..e7f63ff
 --- a/tensorflow/contrib/lite/tools/benchmark/benchmark_params.h
 +++ b/tensorflow/contrib/lite/tools/benchmark/benchmark_params.h
 @@ -47,7 +47,8 @@ class BenchmarkParam {
    virtual ~BenchmarkParam() {}
    BenchmarkParam(ParamType type) : type_(type) {}

 - private:
 + //private:
 + public:        //wly
    static void AssertHasSameType(ParamType a, ParamType b);
    template <typename T>
    static ParamType GetValueType();

3、./tensorflow/contrib/lite/kernels/internal/reference/reference_ops.h

 diff --git a/tensorflow/contrib/lite/kernels/internal/reference/reference_ops.h b/tensorflow/contrib/lite/kernels/internal/reference/reference_ops.h
 old mode 100644
 new mode 100755
 index 2d40f17..43c54dc
 --- a/tensorflow/contrib/lite/kernels/internal/reference/reference_ops.h
 +++ b/tensorflow/contrib/lite/kernels/internal/reference/reference_ops.h
 @@ -24,8 +24,9 @@ limitations under the License.
  #include <type_traits>

 - #include "third_party/eigen3/Eigen/Core"
 +//#include "third_party/eigen3/Eigen/Core""
  #include "tensorflow/contrib/lite/kernels/internal/common.h"
  #include "tensorflow/contrib/lite/kernels/internal/quantization_util.h"
  #include "tensorflow/contrib/lite/kernels/internal/round.h"

4、需要下载包：protobuf

5、将c++11  改为c++0x

12、在PC上测试label_image

./label_image -v 1 -m ./mobilenet_v1_1.0_224.tflite -i ./grace_hopper.jpg -l ./imagenet_slim_labels.txt

报错信息： bash ./label_image: cannot execute binary file:Exec format error

原因有两个：

一是GCC编译时多加了一个-C，生成了二进制文件；

解决方法:找到GCC编译处，去除-C选项。

二就是编译环境不同（平台芯片不一致）导致

解决方法：需要在对应平台编译。

在tensorflow根目录执行：bazel build tensorflow/examples/label_image:label_image

如果是第一次编译，时间较久；

编译完成后，生成可执行文件：bazel_bin/tensorflow/examples/label_image/label_image

本地测试该文件：

拷贝label_image和libtensorflow_framework.so到tensorflow/examples/label_image下

(第一次测试时，未拷贝libtensorflow_framework.so，直接提示：error while loading shared libraryies:libtensorflow_framework.so:cannot open shared object file:No such file or directory)

再次运行./tensorflow/examples/label_image/label_image

显示结果：military uniform(653):0.834306

mortarboard(668):0.0218695

academic gown(401):0.0103581

pickelhaube(716):0.00800814

bulletproof vest(466):0.00535084

说明测试OK！

以下操作在ARM板子上：

1、拷贝生成的label_image到板子上（bazel_bin/tensorflow/contrib/lite/examples/label_image/）(1.7M)

拷贝生成的label_image到板子上（bazel_bin/tensorflow/examples/label_image）(52.2M)

2、拷贝图片./tensorflow/examples/label_image/data/grace_hopper.jpg到板子上

3、下载模型mobilenet_quant_v1_224.tflite

（如果下载其它模型，可参考该文件：/tensorflow/contrib/lite/g3doc/models.md）

4、下载模型所需文件：

curl -L "https://storage.googleapis.com/download.tensorflow.org/models/inception_v3_2016_08_28_frozen.pb.tar.gz" | tar -C tensorflow/examples/label_image/data -xz

5、拷贝libtensorflow_framework.so到板子上，文件位于目录bazel_bin/tensorflow/下（14M）

4、运行label_image

确保如下所需文件都已完成：

（拷贝文件1：label_image）

（拷贝文件2：grace_hopper.jpg）

（拷贝文件3：mobilenet_quant_v1_224.tflite）

（拷贝文件4：imagenet_slim_labels.txt）

（拷贝文件5：libtensorflow_framework.so）

ARM板上常用操作：

mount /dev/sba1 /mnt/usb

ls /mnt/usb

cp /mnt/usb/   .

执行脚本：

./label_image

如果出现-sh: ./label_image: not found，可能是编译器不一致导致。

尝试方法1：重定向:ln -s ld-linux.so.3 ld-linux-armhf.so.3

新报错：./label_image:/lib/libm.so.6: version 'GLIBC_2.27' not found (required by ./label_image)

出现以下信息，说明已经OK，恭喜你!