Question

我看看yotube上的移动和嵌入式TensorFlow（TensorFlow Dev Summit 2017）视频，在这里video link。

在视频中，我学习了一些在Android上减少张量流量以便文件大小的功能。

我在这里

"""Prints a header file to be used with SELECTIVE_REGISTRATION.

Example usage:
print_selective_registration_header \
--graphs=path/to/graph.pb > ops_to_register.h

Then when compiling tensorflow, include ops_to_register.h in the     include
 search path and pass -DSELECTIVE_REGISTRATION  - see
 core/framework/selective_registration.h for more details.
 """

* .pb文件是我自己，然后我在这里得到ops_to_register.h文件

  #ifndef OPS_TO_REGISTER
  #define OPS_TO_REGISTER
  constexpr inline bool ShouldRegisterOp(const char op[]) {
      return false
     || (strcmp(op, "Add") == 0)
     || (strcmp(op, "Const") == 0)
     || (strcmp(op, "Conv2D") == 0)
     || (strcmp(op, "Exp") == 0)
     || (strcmp(op, "Identity") == 0)
     || (strcmp(op, "Max") == 0)
     || (strcmp(op, "MaxPool") == 0)
     || (strcmp(op, "NoOp") == 0)
     || (strcmp(op, "Placeholder") == 0)
     || (strcmp(op, "RealDiv") == 0)
     || (strcmp(op, "Relu") == 0)
     || (strcmp(op, "Reshape") == 0)
     || (strcmp(op, "Sub") == 0)
     || (strcmp(op, "Sum") == 0)
     || (strcmp(op, "_Recv") == 0)
     || (strcmp(op, "_Send") == 0)
     ;
    }
  #define SHOULD_REGISTER_OP(op) ShouldRegisterOp(op)

  const char kNecessaryOpKernelClasses[] = ","
"BinaryOp< CPUDevice, functor::add<float>>,"
"ConstantOp,"
"Conv2DOp<CPUDevice, float>,"
 "UnaryOp< CPUDevice, functor::exp<float>>,"
  "IdentityOp,"
 "ReductionOp<CPUDevice, float, Eigen::internal::MaxReducer<float>>,"
"MaxPoolingOp<CPUDevice, float>,"
 "NoOp,"
  "PlaceholderOp,"
 "BinaryOp< CPUDevice, functor::div<float>>,"
  "ReluOp<CPUDevice, float>,"
"ReshapeOp,"
"BinaryOp< CPUDevice, functor::sub<float>>,"
 "ReductionOp<CPUDevice, float, Eigen::internal::SumReducer<float>>,"
 "RecvOp,"
"SendOp,"
 ;
#define SHOULD_REGISTER_OP_KERNEL(clz)               (strstr(kNecessaryOpKernelClasses, "," clz ",") != nullptr)

 #define SHOULD_REGISTER_OP_GRADIENT false
 #endif

我将ops_to_register.h放在tensorflow / tensorflow / core / framework目录中，并在selective_registration.h中定义SELECTIVE_REGISTRATION。

然后我跑

bazel build -c opt //tensorflow/contrib/android:libtensorflow_inference.so    --crosstool_top=//external:android/crosstool    --host_crosstool_top=@bazel_tools//tools/cpp:toolchain    --cpu=armeabi-v7a --verbose_failures

在android项目中，我使用libtensorflow_inference.so运行我的.pb模型，但获取失败的信息：

native: tensorflow_inference_jni.cc:145 Could not create TensorFlow graph: Invalid argument: No OpKernel was registered to support Op 'Add' with these attrs.  Registered devices: [CPU], Registered kernels:
                                                        <no registered kernels>

                                                         [[Node: add_1 = Add[T=DT_FLOAT](Conv2D, Reshape)]]

Answer 1

这个错误因为tesorflow分支中的一些问题有一些错误，解决问题很容易。

BinaryOp< CPUDevice, functor::div<float>> change to BinaryOp<CPUDevice, functor::div<float>>, the change is no space。

没有注册OpKernel来支持Op＆＃39; Add＆＃39; Android上使用这些功能

1 个答案: