Tensorflow完全连接在单元测试用例中不起作用？

Question

我必须错过一些基本的东西。这是我的测试用例 -

    def test_silly(self):
        with self.test_session() as sess:
            init_op = tf.global_variables_initializer()
            out = tf.contrib.layers.fully_connected(
                inputs=tf.zeros([self.batch_size, 10]),
                num_outputs=20,
                activation_fn=None,
                scope="hmmm")
            sess.run(init_op)
            print(out.eval())

我不明白为什么我一直得到如下错误。

FailedPreconditionError (see above for traceback): Attempting to use uninitialized value hmmm/weights
 [[Node: hmmm/weights/read = Identity[T=DT_FLOAT, _class=["loc:@hmmm/weights"], _device="/job:localhost/replica:0/task:0/cpu:0"](hmmm/weights)]]

Answer 1

hmmm/weights未初始化。

def test_silly(self):
    with self.test_session() as sess:
        out = tf.contrib.layers.fully_connected(
            inputs=tf.zeros([self.batch_size, 10]),
            num_outputs=20,
            activation_fn=None,
            scope="hmmm")
        init_op = tf.global_variables_initializer()
        sess.run(init_op)
        print(out.eval())

交换init_op的位置会使其有效。有关更多见解，请参阅What does tf.global_variables_initializer() do under the hood?。