python - 为什么 tf.get_variable ('test' ) 返回名为 test_1 的变量？

Question

我使用 tf.Variable 创建了一个 tensorflow 变量。我想知道为什么如果我使用相同的名称调用 tf.get_variable 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 是一种低级方法，它将创建的变量存储在 GLOBALS(或 LOCALS)集合中，而 tf.get_variable 则记录它所拥有的变量通过将它们存储在变量存储中来创建。

当您第一次调用tf.Variable时，创建的变量不会添加到变量存储中，让我们认为尚未创建名为“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，发现名称为 "test" 的变量不存在它。
因此，它将调用 tf.Variable，这将创建一个具有递增名称 "test_1" 的变量，该变量存储在变量存储中的键 "test".

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.")