为什么 tf.get_variable('test') 返回名称为 test_1 的变量？

Question

我用tf.Variable创建了一个张量流变量。 我想知道为什么如果我用相同的名称调用tf.get_variable，就不会引发异常，并且会创建一个名称递增的新变量？

import tensorflow as tf
class QuestionTest(tf.test.TestCase):
def test_version(self):
self.assertEqual(tf.__version__, '1.10.1')
def test_variable(self):
a = tf.Variable(0., trainable=False, name='test')
self.assertEqual(a.name, "test:0")
b = tf.get_variable('test', shape=(), trainable=False)
self.assertEqual(b.name, "test_1:0")
self.assertNotEqual(a, b, msg='`a` is not `b`')
with self.assertRaises(ValueError) as ecm:
tf.get_variable('test', shape=(), trainable=False)
exception = ecm.exception
self.assertStartsWith(str(exception), "Variable test already exists, disallowed.")

Answer 1

这是因为tf.Variable是一个低级方法，它将创建的变量存储在GLOBALS(或LOCALS(集合中，而tf.get_variable则通过将它们存储在变量存储中来记录它创建的变量。

当你第一次调用tf.Variable时，创建的变量不会添加到变量存储中，让人认为没有创建名为"test"的变量。

因此，当您稍后调用tf.get_variable("test")时，它将查看变量存储区，看到其中没有名称为"test"的变量。
因此，它将调用tf.Variable，这将创建一个名称递增的变量"test_1"存储在键"test"下的变量存储中。

import tensorflow as tf
class AnswerTest(tf.test.TestCase):
def test_version(self):
self.assertEqual(tf.__version__, '1.10.1')    
def test_variable_answer(self):
"""Using the default variable scope"""
# Let first check the __variable_store and the GLOBALS collections.
self.assertListEqual(tf.get_collection(("__variable_store",)), [], 
"No variable store.")
self.assertListEqual(tf.global_variables(), [],
"No global variables")
a = tf.Variable(0., trainable=False, name='test')
self.assertEqual(a.name, "test:0")
self.assertListEqual(tf.get_collection(("__variable_store",)), [],
"No variable store.")
self.assertListEqual(tf.global_variables(), [a],
"but `a` is in global variables.")
b = tf.get_variable('test', shape=(), trainable=False)
self.assertNotEqual(a, b, msg='`a` is not `b`')
self.assertEqual(b.name, "test_1:0", msg="`b`'s name is not 'test'.")
self.assertTrue(len(tf.get_collection(("__variable_store",))) > 0,
"There is now a variable store.")
var_store = tf.get_collection(("__variable_store",))[0]
self.assertDictEqual(var_store._vars, {"test": b},
"and variable `b` is in it.")
self.assertListEqual(tf.global_variables(), [a, b],
"while `a` and `b` are in global variables.")
with self.assertRaises(ValueError) as exception_context_manager:
tf.get_variable('test', shape=(), trainable=False)
exception = exception_context_manager.exception
self.assertStartsWith(str(exception),
"Variable test already exists, disallowed.")

使用显式变量作用域时也是如此。

def test_variable_answer_with_variable_scope(self):
"""Using now a variable scope"""
self.assertListEqual(tf.get_collection(("__variable_store",)), [], 
"No variable store.")
with tf.variable_scope("my_scope") as scope:
self.assertTrue(len(tf.get_collection(("__variable_store",))) > 0, 
"There is now a variable store.")
var_store = tf.get_collection(("__variable_store",))[0]
self.assertDictEqual(var_store._vars, {},
"but with variable in it.")
a = tf.Variable(0., trainable=False, name='test')
self.assertEqual(a.name, "my_scope/test:0")
var_store = tf.get_collection(("__variable_store",))[0]
self.assertDictEqual(var_store._vars, {},
"Still no variable in the store.")

b = tf.get_variable('test', shape=(), trainable=False)
self.assertEqual(b.name, "my_scope/test_1:0")
var_store = tf.get_collection(("__variable_store",))[0]
self.assertDictEqual(
var_store._vars, {"my_scope/test": b},
"`b` is in the store, but notice the difference between its name and its key in the store.")
with self.assertRaises(ValueError) as exception_context_manager:
tf.get_variable('test', shape=(), trainable=False)
exception = exception_context_manager.exception
self.assertStartsWith(str(exception),
"Variable my_scope/test already exists, disallowed.")