bazel编译android用的TF library，添加rfft op的编译

模型中使用了函数

tensorflow.contrib.signal.stft

在安卓平台中推断出现错误如下：

E/AndroidRuntime: FATAL EXCEPTION: main
    Process: com.example.hhc.myapplication, PID: 10317
    java.lang.IllegalArgumentException: No OpKernel was registered to support Op 'RFFT' used by {{node stft/rfft}}with these attrs: []
    Registered devices: [CPU]
    Registered kernels:
      <no registered kernels>
    
         [[{{node stft/rfft}}]]
        at org.tensorflow.Session.run(Native Method)
        at org.tensorflow.Session.access$100(Session.java:48)
        at org.tensorflow.Session$Runner.runHelper(Session.java:314)
        at org.tensorflow.Session$Runner.run(Session.java:264)
        at org.tensorflow.contrib.android.TensorFlowInferenceInterface.run(TensorFlowInferenceInterface.java:228)
        at org.tensorflow.contrib.android.TensorFlowInferenceInterface.run(TensorFlowInferenceInterface.java:197)
        at org.tensorflow.contrib.android.TensorFlowInferenceInterface.run(TensorFlowInferenceInterface.java:187)
        at com.example.hhc.myapplication.MainActivity.recognizeRecord(MainActivity.java:193)
        at com.example.hhc.myapplication.MainActivity$2.onClick(MainActivity.java:104)
        at android.view.View.performClick(View.java:4848)
        at android.view.View$PerformClick.run(View.java:20262)
        at android.os.Handler.handleCallback(Handler.java:815)
        at android.os.Handler.dispatchMessage(Handler.java:104)
        at android.os.Looper.loop(Looper.java:194)
        at android.app.ActivityThread.main(ActivityThread.java:5631)
        at java.lang.reflect.Method.invoke(Native Method)
        at java.lang.reflect.Method.invoke(Method.java:372)
        at com.android.internal.os.ZygoteInit$MethodAndArgsCaller.run(ZygoteInit.java:959)
        at com.android.internal.os.ZygoteInit.main(ZygoteInit.java:754)

原因：默认编译的android用TF library，只实现了部分op，当使用的op不在默认范围内时，需要自己添加对应的op。

图片来源：https://github.com/tensorflow/tensorflow/issues/11804

 

解决方案：经过网上查找，对应我使用的函数，需要添加的是“fft_op.cc”，其路径为tf的源码目录中的"/tensorflow/tensorflow/core/kernels/fft_op.cc"。添加这个文件的编译，需要修改“/tensorflow/tensorflow/core/framework/register_types.h”文件，在其中添加"fft_op"，如下。然后，按照原来的命令编译即可。

filegroup(
    name = "android_extended_ops_group2",
    srcs = [
        ...
        "maxpooling_op.cc",
        "mfcc.cc",
        "mfcc_dct.cc",
        "mfcc_mel_filterbank.cc",
        "mfcc_op.cc",
        "fft_ops.cc", //添加ffc_ops
        "mirror_pad_op.cc",
        "mirror_pad_op_cpu_impl_1.cc",
        "mirror_pad_op_cpu_impl_2.cc",
        "mirror_pad_op_cpu_impl_3.cc",
        "mirror_pad_op_cpu_impl_4.cc",
        "mirror_pad_op_cpu_impl_5.cc",
        ...
        ":android_extended_ops_headers",
    ],
)

参考链接：https://github.com/tensorflow/models/issues/1740

 

另附

android 用TF library 编译教程连接：https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/android/README.md