python - tensorflow 初始化变量错误

Question

众所周知，在 tensorflow 中初始化变量有多种方法。我尝试了一些与图形定义相结合的东西。请参阅下面的代码。

def Graph1a():
    g1 = tf.Graph()
    with g1.as_default() as g:
        matrix1 = tf.constant([[3., 3.]])
        matrix2 = tf.constant([[2.],[2.]])
        product = tf.matmul( matrix1, matrix2, name = "product")

    sess = tf.Session( graph = g )
    sess.run(tf.global_variables_initializer())
    return product

def Graph1b():
    g1 = tf.Graph()
    with g1.as_default() as g:
        matrix1 = tf.constant([[3., 3.]])
        matrix2 = tf.constant([[2.],[2.]])
        product = tf.matmul( matrix1, matrix2, name = "product")

    sess = tf.Session( graph = g )
    sess.run(tf.initialize_all_variables())
    return product

def Graph1c():
    g1 = tf.Graph()
    with g1.as_default() as g:
        matrix1 = tf.constant([[3., 3.]])
        matrix2 = tf.constant([[2.],[2.]])
        product = tf.matmul( matrix1, matrix2, name = "product")

    with tf.Session( graph = g ) as sess:
        tf.global_variables_initializer().run()
        return product

为什么 Graph1a() 和 Graph1b() 不会返回乘积，而 Graph1c() 会返回乘积？我认为这些陈述是等效的。

Answer 1

问题是 global_variables_initializer 需要与 session 相同的图关联。在 Graph1c 中，发生这种情况是因为 global_variables_initializer 位于 session 的 with 语句的范围内。要让 Graph1a 工作，需要像这样重写

def Graph1a():
    g1 = tf.Graph()
    with g1.as_default() as g:
        matrix1 = tf.constant([[3., 3.]])
        matrix2 = tf.constant([[2.],[2.]])
        product = tf.matmul( matrix1, matrix2, name = "product")
        init_op = tf.global_variables_initializer()

    sess = tf.Session( graph = g )
    sess.run(init_op)
    return product