Skip to content

Commit

Permalink
残差网络
Browse files Browse the repository at this point in the history
  • Loading branch information
shataowei committed Dec 2, 2019
1 parent 6414b2f commit 69885b1
Show file tree
Hide file tree
Showing 3 changed files with 82 additions and 26 deletions.
81 changes: 55 additions & 26 deletions .idea/workspace.xml

Some generated files are not rendered by default. Learn more about how customized files appear on GitHub.

5 changes: 5 additions & 0 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -288,6 +288,11 @@
- embedding
- softmax
- 梯度消失/爆炸
- 残差网络
- [介绍残差网络](深度学习/残差网络.md#L164)
- [残差网络为什么能解决梯度消失的问题](深度学习/残差网络.md#L164)
- [残差网络残差作用](深度学习/残差网络.md#L164)
- [你平时有用过么?或者你在哪些地方遇到了](深度学习/残差网络.md#L164)
- Attention
- [Attention对比RNN和CNN,分别有哪点你觉得的优势](深度学习/Attention.md#L164)
- [写出Attention的公式](深度学习/Attention.md#L164)
Expand Down
22 changes: 22 additions & 0 deletions 深度学习/残差网络.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,22 @@
# 介绍残差网络
- 常见结构,CV里面用的比较多
- y=F(x)+x
- y=F(x)+indentity `*` x

# 残差网络为什么能解决梯度消失的问题
- ![](https://tva1.sinaimg.cn/large/006tNbRwly1g9is4x344xj304j01jglf.jpg)
- ![](https://tva1.sinaimg.cn/large/006tNbRwly1g9is8cigikj305s0180sl.jpg)
- 虽然是对![](https://tva1.sinaimg.cn/large/006tNbRwly1g9is8rtnxuj300l00k3y9.jpg)求偏导数,但是存在一项只和![](https://tva1.sinaimg.cn/large/006tNbRwly1g9is984bi0j300l00k3y9.jpg)相关的项,之间避免了何中间权重矩阵变换导致梯度消失的问题

# 残差网络残差作用
- 防止梯度消失
- 恒等映射使得网络突破层数限制,避免网络退化
- 对输出的变化更敏感
- X=5;F(X)=5.1;F(X)=H(X)+X=>H(X)=0.1
- X=5;F(X)=5.2;F(X)=H(X)+X=>H(X)=0.2
- H(X)变换了100%,去掉相同的主体部分,从而突出微小的变化

# 你平时有用过么?或者你在哪些地方遇到了
- 我在做xdeepfm的输出层的时候做到了,因为当时做CIN的时候,我设置了layers为5层,担心层数过深造成网络退化,在output的时候加了残差网络
- Bert和Transform中attention部分残差网络用的比较频繁

0 comments on commit 69885b1

Please sign in to comment.