深度学习名词2：batch & epoches &dacay &iteration

参考google group：https://groups.google.com/forum/#!searchin/darknet/subdivisions|sort:relevance/darknet/WgFJDkVc7z4/V56hc5w_AwAJ
https://groups.google.com/forum/#!searchin/darknet/subdivisions|sort:relevance/darknet/qZp5P0I4iGA/OexXrkK7DwAJ

batch sizes

https://www.zhihu.com/question/32673260

batch: how many images load at once;

epoches

一个epoch是指把所有训练数据完整的过一遍；

一次epoch=所有训练数据forward+backward后更新参数的过程。
一次iteration=[batch size]个训练数据forward+backward后更新参数过程

dacay

iterations（迭代）

每一次迭代都是一次权重更新，每一次权重更新需要batch size个数据进行Forward运算得到损失函数，再BP算法更新参数

 one epoch = numbers of iterations = N = 训练样本的数量/batch size

subdivisions:

how many times to break image batch into smaller pieces;
batch / subdivisions:how many images are loaded and feed to network.