network中的weight initialization

因为自建的网络很深导致网络不能很好的训练，于是寻找解决方法，
（１）好的初始参数
tanh的初始化用Xavier initialization
ReLU初始化用He initialization
（２）batch normalization

博文还没有完善，等我有空
https://www.leiphone.com/news/201703/3qMp45aQtbxTdzmK.html
http://www.imooc.com/article/details/id/20307
http://m.blog.csdn.net/hongbin_xu/article/details/78708388
http://blog.csdn.net/u011324454/article/details/78648096