了解 tf.name_scope

Question

I am trying to understand tf.name_scope. The documentation mentions the following:

"This context manager pushes a name scope, which will make the name of all operations added within it have a prefix.

For example, to define a new Python op called my_op:

def my_op(a, b, c, name=None):
  with tf.name_scope("MyOp") as scope:
    a = tf.convert_to_tensor(a, name="a")
    b = tf.convert_to_tensor(b, name="b")
    c = tf.convert_to_tensor(c, name="c")
    # Define some computation that uses `a`, `b`, and `c`.
    return foo_op(..., name=scope)

When executed, the Tensors a, b, c, will have names MyOp/a, MyOp/b, and MyOp/c."

My understanding is that the with block does not introduce a new local scope in Python. Under normal situation, the tensor variable a will also refer to the local parameter a of function my_op. How is the name prefixing with "MyOp/" implemented using Python context? In the source code link for tf.name_scope ( https://github.com/tensorflow/tensorflow/blob/v2.2.0/tensorflow/python/framework/ops.py#L6423-L6442 ) there is an invocation of

ctx = context.context()

but I could not find the semantics of context.context(). Most context manager discussion talk about enter and exit , but no mention of variable renaming with some prefix. Is this some introspective mechanism in Python that allows the manipulation of Python variable scopes? Many thanks for any insights.

Answer 1

Name scopes are entirely for naming convenience. It has nothing to do with python scope. Whenever you perform an operation in tensorflow it will get a name according to some state maintained by tensorflow, but if you want those operations to be named clearly for your use case then name scope is used. Any operation within the name scope will have the provided value appended in it's name, there is no use more than that. Regarding the implementation, I highly doubt that those scope implementations will be exposed in python. All tensor creation operations are executed by C++ backend of tensorflow, so scopes should be mostly handled by them