Tensorflow中实现BN为什么需要加入这个额外依赖？见CS231N作业源码

batch normalization in tensorflow requires this extra dependency

为什么加上这两句？

extra_update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(extra_update_ops):
train_step = optimizer.minimize(mean_loss)