编译tensorflow-lite-with-select-tf-ops遇到的坑

前言

最近在将tf训练的模型迁移到Android端，使用的是tensorflow-lite，由于模型用到了一些tflite还没有支持的op，所以需要让tflite支持tf的op，官方没有直接给出aar，而是让自己用bazel去编译一个，实在是有点坑啊，官方编译网址：

https://www.tensorflow.org/lite/using_select_tf_ops

大致方法就是：

【1】 安装bazel，【 https://docs.bazel.build/versions/master/install.html 】

【2】 下载tensorflow源码，【 https://github.com/tensorflow/tensorflow 】

【3】 进入源码所在目录，修改WORKSPACE文件，增加sdk路径：

android_sdk_repository (
    name = "androidsdk",
    api_level = 26,
    build_tools_version = "26.0.2",
    path = "/Users/vell/Library/Android/sdk/",
)

android_ndk_repository(
    name = "androidndk",
    path = "/Users/vell/Library/Android/sdk/ndk-bundle",
    api_level = 19,
)

注意修改自己的sdk路径和版本

【4】 进入源码所在的目录执行如下命令：

bazel build --cxxopt='--std=c++11' -c opt             \
  --config=android_arm --config=monolithic          \
  //tensorflow/lite/java:tensorflow-lite-with-select-tf-ops

【5】 如果你运气足够好的话，你将在如下目录找到编译好的aar：

bazel-genfiles/tensorflow/lite/java/tensorflow-lite-with-select-tf-ops.aar

【6】 大功告成，但是，我就是那个运气极其差的，总是遇到些奇怪的问题，我这算是有两个吧

“undeclared inclusion(s)” error

ERROR: /data/vellhe/tensorflow-master/tensorflow/core/common_runtime/eager/BUILD:40:1: undeclared inclusion(s) in rule '//tensorflow/core/common_runtime/eager:context':
this rule is missing dependency declarations for the following files included by 'tensorflow/core/common_runtime/eager/context.cc':
  'tensorflow/core/distributed_runtime/collective_param_resolver_distributed.h'
  'tensorflow/core/distributed_runtime/device_resolver_distributed.h'
  'tensorflow/core/distributed_runtime/rpc_collective_executor_mgr.h'
Target //tensorflow/lite/java:tensorflow-lite-with-select-tf-ops failed to build
Use --verbose_failures to see the command lines of failed build steps.
INFO: Elapsed time: 85.812s, Critical Path: 33.15s
INFO: 666 processes: 655 local, 11 worker.
FAILED: Build did NOT complete successfully

这个是我遇到最坑的，由于对bazel工具不熟悉，各种google，花了整整一个上午，才搞出点名堂了，如果编译遇到坑的话，还是建议先去熟悉下bazel

其实stackoverflow上已经有大神给到了解决方法：

How to resolve bazel “undeclared inclusion(s)” error?

只不过，如果没有bazel背景是很难真正看懂怎么操作

问题原因

错误信息里可以看出，是.h文件没有声明， tensorflow/core/common_runtime/eager/context.cc 里缺三个.h文件的声明，缺的.h文件都在 tensorflow/core/distributed_runtime 目录下， tensorflow/core/distributed_runtime 里有一个BUILD文件，这个目录就相当于也是一个模块，所以需要做的就是将 tensorflow/core/distributed_runtime 里的.h文件告诉 tensorflow/core/common_runtime/eager/context.cc

解决方法：

【1】 编辑 tensorflow/core/distributed_runtime/BUILD 文件，新增如下：

cc_library(
    name = "headers",
    hdrs = glob(["*.h"])
)

【2】 编辑 tensorflow/core/common_runtime/eager/BUILD 文件，找到 //tensorflow/core/common_runtime/eager:context 的定义，在deps里增加一条， "//tensorflow/core/distributed_runtime:headers", ，修改后如下：

【3】 大功告成

no member named 'round' in namespace 'std';

ERROR: /Users/vell/workspace/github/tensorflow-master/tensorflow/lite/kernels/internal/BUILD:418:1: C++ compilation of rule '//tensorflow/lite/kernels/internal:neon_tensor_utils' failed (Exit 1)
In file included from tensorflow/lite/kernels/internal/reference/portable_tensor_utils.cc:19:
./tensorflow/lite/c/builtin_op_data.h:154:9: warning: empty struct has size 0 in C, size 1 in C++ [-Wextern-c-compat]
typedef struct {
        ^
./tensorflow/lite/c/builtin_op_data.h:157:9: warning: empty struct has size 0 in C, size 1 in C++ [-Wextern-c-compat]
typedef struct {
        ^
./tensorflow/lite/c/builtin_op_data.h:232:9: warning: empty struct has size 0 in C, size 1 in C++ [-Wextern-c-compat]
typedef struct {
        ^
./tensorflow/lite/c/builtin_op_data.h:235:9: warning: empty struct has size 0 in C, size 1 in C++ [-Wextern-c-compat]
typedef struct {
        ^
./tensorflow/lite/c/builtin_op_data.h:274:9: warning: empty struct has size 0 in C, size 1 in C++ [-Wextern-c-compat]
typedef struct {
        ^
In file included from tensorflow/lite/kernels/internal/reference/portable_tensor_utils.cc:21:
./tensorflow/lite/kernels/internal/round.h:28:10: error: no member named 'round' in namespace 'std'; did you mean simply 'round'?
  return std::round(x);
         ^~~~~~~~~~
         round
external/androidndk/ndk/sysroot/usr/include/math.h:255:8: note: 'round' declared here
double round(double __x);
       ^
5 warnings and 1 error generated.
Target //tensorflow/lite/java:tensorflow-lite-with-select-tf-ops failed to build
Use --verbose_failures to see the command lines of failed build steps.
INFO: Elapsed time: 6.732s, Critical Path: 3.36s
INFO: 0 processes.
FAILED: Build did NOT complete successfully

问题原因

在 tensorflow/lite/kernels/internal/round.h:28:10 处，调用了 std::round(x); ，而在std的namespace里没有round这个函数，所以报错了

解决方法

将std去掉，如下：