在写系统时候遇到一个问题：模仿的版本在测试阶段，把所有权重乘以一个保留概率（做法1）。而GJH他看的Keras里面测试阶段什么都不做（做法2），所以比较好奇，怎么做法不同？
但是随手搜了几个中文博客，都是按做法1来的，训练阶段按保留概率随机生成一个保留矩阵(元素要么0-丢弃，要么1-保留)，然后测试阶段所有权重乘以保留概率。
在回头看TensorFlow的Dropout函数https://www.tensorflow.org/api_docs/python/tf/nn/dropout，怎么涉及一个scale操作，懵了：
With probability keep_prob, outputs the input element scaled up by 1 / keep_prob, otherwise outputs 0. The scaling is so that the expected sum is unchanged.
Google了下，找到类似的疑惑提问：
Why input is scaled in tf.nn.dropout in tensorflow?
http://techqa.info/programming/question/34597316/why-input-is-scaled-in-tf.nn.dropout-in-tensorflow
顺藤摸瓜去扫了一眼Dropout原文：http://www.cs.toronto.edu/~rsalakhu/papers/srivastava14a.pdf，才发现是Dropout的两种策略：
In this paper, we described dropout as a method where we retain units with probability pat training time and scale down the weights by multiplying them by a factor of p at test time.Another way to achieve the same effect is to scale up the retained activations by multiplying by 1/p at training time and not modifying the weights at test time. These methods are equivalent with appropriate scaling of the learning rate and weight initializations at each layer.
简单讲就是：
训练有scale，测试不用管。
训练没scale，测试要乘p。
不过，读了下原文摘要，比那些博客写的好多了... 还是要多看论文啊.....