tf.summary.experimental.summary_scope

View source on GitHub

Experimental context manager for use when defining a custom summary op.

@tf_contextlib.contextmanager
tf.summary.experimental.summary_scope(
    name, default_name='summary', values=None
)

This behaves similarly to tf.name_scope, except that it returns a generated summary tag in addition to the scope name. The tag is structurally similar to the scope name - derived from the user-provided name, prefixed with enclosing name scopes if any - but we relax the constraint that it be uniquified, as well as the character set limitation (so the user-provided name can contain characters not legal for scope names; in the scope name these are removed).

This makes the summary tag more predictable and consistent for the user.

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

def my_op(name, my_value, step):
  with tf.summary.summary_scope(name, "MyOp", [my_value]) as (tag, scope):
    my_value = tf.convert_to_tensor(my_value)
    return tf.summary.write(tag, my_value, step=step)

Yields:

A tuple (tag, scope) as described above.