To Learn More-Optimization for Deep Learning

何为优化?

找个loss surface的最小值

To Learn More-Optimization for Deep Learning

On-line一次可以拿到一组训练资料,Off-line一次拿到所有的训练资料

To Learn More-Optimization for Deep Learning

To Learn More-Optimization for Deep Learning

Gradient的方向就是L增加的方法,所以我们要往反方向走,就是L减少的方向走,目标找到一个To Learn More-Optimization for Deep Learning可以有最小的L

复习一下SGD

To Learn More-Optimization for Deep Learning

加入了动量之后

To Learn More-Optimization for Deep Learning

To Learn More-Optimization for Deep Learning

Adagard

To Learn More-Optimization for Deep Learning

RMSProp

相比于Adagrad,如果Adagrad刚开始g很大,导致learning rate 一直很小,就会很容易卡住;RMSProp就很好的解决了这个问题,这个Optimizer不会在走没几步以后就因为前几步

gradient太大,所以停下来

To Learn More-Optimization for Deep Learning

Adam

To Learn More-Optimization for Deep Learning

对比一下Adam和SGDM

To Learn More-Optimization for Deep Learning

 

 

 

上一篇:Deep Glow for mac(AE高级辉光特效插件)支持ae2021


下一篇:element ui el-table中默认的排序三角形改变样式