2020/05/14
Three key points of Neural Network:DataSet,Network,Loss.
The component that works as engine is the training method! In other word, the method while using to minimize the loss function.
最近做实验,训练小数据集,发现,根本不需要调learning rate 和 decay,用keras包自己带的“Adam” and "SGD"就够用了,而且BatchNormalization都可以在小数据集上不使用。
BN的优势在大数据集上,可以有效避免训练过程中loss的不收敛。
----to be continued!