<?xml version="1.0" encoding="utf-8" ?><rss version="2.0"><channel><title><![CDATA[chikeshi的博客]]></title><description><![CDATA[]]></description><link>https://blog.youkuaiyun.com/chikeshi</link><language>zh-cn</language><generator>https://blog.youkuaiyun.com/</generator><copyright><![CDATA[Copyright &copy; chikeshi]]></copyright><item><title><![CDATA[CS234 RL Winter]]></title><link>https://blog.youkuaiyun.com/chikeshi/article/details/118437572</link><guid>https://blog.youkuaiyun.com/chikeshi/article/details/118437572</guid><author>chikeshi</author><pubDate>Sat, 17 Jul 2021 10:01:51 +0800</pubDate><description><![CDATA[文章目录IntroMarkov (Decision / reward) process

lecture 1

Intro

rrr Reward function, can be a function of $r(s) $or r(s,a)r(s,a)r(s,a)
sss, state
aaa, action
hhh: history
γ\gammaγ: discount factor

mathematically convenient. γ=0\gamma = 0γ=0: only care abou]]></description><category></category></item><item><title><![CDATA[cs120]]></title><link>https://blog.youkuaiyun.com/chikeshi/article/details/115535821</link><guid>https://blog.youkuaiyun.com/chikeshi/article/details/115535821</guid><author>chikeshi</author><pubDate>Mon, 21 Jun 2021 06:39:00 +0800</pubDate><description><![CDATA[CSE 120前言Lecture 3Reading [chapter 6](https://pages.cs.wisc.edu/~remzi/OSTEP/cpu-mechanisms.pdf)Two modes: User mode and kernel modetrap tabletwo phases in the limited direct execution (LDE) protocol.切换进程concurrencyLecture 3Handling FaulsSystem callsLectur]]></description><category></category></item><item><title><![CDATA[cse123 computer networks]]></title><link>https://blog.youkuaiyun.com/chikeshi/article/details/115500476</link><guid>https://blog.youkuaiyun.com/chikeshi/article/details/115500476</guid><author>chikeshi</author><pubDate>Wed, 28 Apr 2021 15:44:27 +0800</pubDate><description><![CDATA[文章目录Lecture 5
Lecture 5

stop and wait- 一个一个发

收到之后/time out发下一个
sender要加上sequence number
receiver的ack也要加上sequence
propagation dellay是吧package发过去再收回来。最快是光速
Go back N：receiver只ack最高的连续的收到的package
如果中间的丢失，那么sender要重发后面的
sliding window
window length = bandwid]]></description><category></category></item><item><title><![CDATA[Bert源码解读]]></title><link>https://blog.youkuaiyun.com/chikeshi/article/details/115500543</link><guid>https://blog.youkuaiyun.com/chikeshi/article/details/115500543</guid><author>chikeshi</author><pubDate>Sat, 10 Apr 2021 00:36:22 +0800</pubDate><description><![CDATA[谈一谈自己的理解。
Transformer
讲的最好的一篇（我认为） 用pytorch一步步实现

他的multihead应该是错的，不应该吧embediing分开，而应该有多个Q，K，V

错误解决方案请看这篇，或者中文版，并且讲述了bert,吧所有head的output拼在一起，并训练WoW_oWo​使其回归到原来的大小
attention的输出是每一个词的embedding，somehow有其他词的attention。每个词是v1*attention_v1 + v2*attention_v2 …

第]]></description><category></category></item><item><title><![CDATA[hands on machine learning]]></title><link>https://blog.youkuaiyun.com/chikeshi/article/details/110746005</link><guid>https://blog.youkuaiyun.com/chikeshi/article/details/110746005</guid><author>chikeshi</author><pubDate>Mon, 08 Mar 2021 06:51:36 +0800</pubDate><description><![CDATA[test]]></description><category></category></item><item><title><![CDATA[CS244n NLP with Deep Learning | Winter 2019]]></title><link>https://blog.youkuaiyun.com/chikeshi/article/details/114244347</link><guid>https://blog.youkuaiyun.com/chikeshi/article/details/114244347</guid><author>chikeshi</author><pubDate>Fri, 05 Mar 2021 19:58:18 +0800</pubDate><description><![CDATA[CS244n  NLP with Deep Learning | Winter 2019 Lecture 1word2vecLecture 2
Lecture 1
word2vec

每一个单词可以变成一个vector, 相似的word vector也相似
word = [0.230.52−0.41−0.310.270.48]\begin{bmatrix} 0.23 \\ 0.52 \\ -0.41 \\ -0.31 \\ 0.27 \\ 0.48 \end{bmatrix}⎣⎢⎢⎢⎢⎢⎢⎡​0.230.5]]></description><category></category></item></channel></rss>