c++ - Tensorflow 操作文档

Question

在 Tensorflow 中 source ，我明白了，

REGISTER_OP("BroadcastGradientArgs")
    .Input("s0: T")
    .Input("s1: T")
    .Output("r0: T")
    .Output("r1: T")
    .Attr("T: {int32, int64} = DT_INT32")
    .SetShapeFn([](InferenceContext* c) {
      ... uninteresting details ...
    })
    .Doc(R"doc(                                                                                                              
Return the reduction indices for computing gradients of s0 op s1 with broadcast.                                             

This is typically used by gradient computations for a broadcasting operation.                                                
)doc");

在 Python 中，我可以执行以下操作，

>>> from tensorflow.python.ops import gen_array_ops
>>> gen_array_ops._InitOpDefLibrary()._ops['BroadcastGradientArgs'].op_def
name: "BroadcastGradientArgs"
input_arg {
  name: "s0"
  type_attr: "T"
}
... more stuff ...
attr {
  name: "T"
  type: "type"
  ... uninteresting details ...
}

请注意，我在 Python 中获取了 TF 操作的 Protobuf 定义(为简洁起见，我删除了一些定义)。我想获取我在 C++ 代码中看到的定义的文档部分。我如何获得它？

Answer 1

那是痛苦的。你需要给 TF 和 Protobuf 打补丁

https://github.com/tensorflow/tensorflow/issues/8207 https://github.com/google/protobuf/issues/2798

然后你需要另外注释掉这一行， https://github.com/tensorflow/tensorflow/blob/a3e636c0f561e2ac6d9f8a0044fbe09acb003803/tensorflow/python/framework/python_op_gen.cc#L708

重建和运行给了，

bazel build --config=opt //tensorflow/tools/pip_package:build_pip_package && \
bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg && \
sudo pip install --upgrade /tmp/tensorflow_pkg/tensorflow-1.0.0*.whl

并执行我的测试，

$ python -c "from tensorflow.python.ops import gen_array_ops; print gen_array_ops._InitOpDefLibrary()._ops['BroadcastGradientArgs'].op_def"
name: "BroadcastGradientArgs"
input_arg {
  name: "s0"
  type_attr: "T"
}
input_arg {
  name: "s1"
  type_attr: "T"
}
output_arg {
  name: "r0"
  type_attr: "T"
}
output_arg {
  name: "r1"
  type_attr: "T"
}
attr {
  name: "T"
  type: "type"
  default_value {
    type: DT_INT32
  }
  allowed_values {
    list {
      type: DT_INT32
      type: DT_INT64
    }
  }
}
summary: "Return the reduction indices for computing gradients of s0 op s1 with broadcast."
description: "This is typically used by gradient computations for a broadcasting operation."

这用了太多正则表达式......