<?xml version="1.0" encoding="utf-8" ?><rss version="2.0"><channel><title><![CDATA[LuffysMan的博客]]></title><description><![CDATA[]]></description><link>https://blog.youkuaiyun.com/LuffysMan</link><language>zh-cn</language><generator>https://blog.youkuaiyun.com/</generator><copyright><![CDATA[Copyright &copy; LuffysMan]]></copyright><item><title><![CDATA[快速排序(荷兰国旗法)]]></title><link>https://blog.youkuaiyun.com/LuffysMan/article/details/117080044</link><guid>https://blog.youkuaiyun.com/LuffysMan/article/details/117080044</guid><author>LuffysMan</author><pubDate>Thu, 20 May 2021 16:47:38 +0800</pubDate><description><![CDATA[快排(QuickSort)
快排是经典的二分法思想的排序方法, 它是不稳定的(即a0, a1的值均为1, 但是排序后可能a1在a0前面, 不保证相对顺序不变).
代码实现(python, 荷兰国旗法)
"""
输入: 数组A, 长度为L. 
输出: 排序后的数组A.
基本思想: 
	①select pivot: 从输入数组中随机取一个数字作为基准(pivot);
	②partition: 将所有&lt;pivot的数字放到pivot左边, 记为子数组A_left, 所有&gt;pivot的数字放到pivot]]></description><category></category></item><item><title><![CDATA[并查集(DisjointSet)]]></title><link>https://blog.youkuaiyun.com/LuffysMan/article/details/113687704</link><guid>https://blog.youkuaiyun.com/LuffysMan/article/details/113687704</guid><author>LuffysMan</author><pubDate>Thu, 20 May 2021 15:54:14 +0800</pubDate><description><![CDATA[1. 并查集
 并查集可谓非常有用了, 比如用于判断两个节点是否连通, 计算图的连通分量个数, 甚至有一些巧妙的操作"逆向"的解决问题, 比如leetcode803:打砖块
 以下是代码实现
class UnionFind:
    def __init__(self, n):
        """
        :n: 并查集节点总数
        """
        self.__parent = [i for i in range(n)]
        self.__size = [1]*]]></description><category></category></item><item><title><![CDATA[pytorch中torch.Module().eval()有何用处?是否设置之后可以保证模型参数不被更新?]]></title><link>https://blog.youkuaiyun.com/LuffysMan/article/details/116773937</link><guid>https://blog.youkuaiyun.com/LuffysMan/article/details/116773937</guid><author>LuffysMan</author><pubDate>Fri, 14 May 2021 01:44:54 +0800</pubDate><description><![CDATA[关于pytorch Module类eval方法只是一个标志位实测效果结论
 最近看一篇论文的源码, 作者为了将某几层bn层冻结, 居然对每个bn层调用了 bn层的eval()方法. 我有所怀疑是否有效, 查看了pytorch对应的源码, 并做了测试, 确定这样做没什么卵用.
只是一个标志位
model.eval()只是设置一个标志位, 同理model.train()也是一样. 查看Module类源码, eval()方法只做了一件事情, 调用self.train(False).
    def eval(se]]></description><category></category></item><item><title><![CDATA[Pytorch autograd.backward理解]]></title><link>https://blog.youkuaiyun.com/LuffysMan/article/details/112597086</link><guid>https://blog.youkuaiyun.com/LuffysMan/article/details/112597086</guid><author>LuffysMan</author><pubDate>Thu, 14 Jan 2021 01:09:47 +0800</pubDate><description><![CDATA[参考了解释Pytorch autograd,backward详解
通常在训练的时候, 最后的loss是一个标量, 无脑使用loss.backward()进行反向传播计算梯度即可. 但是碰到有些代码中出现了多个loss, 比如这样的代码:
torch.autograd.backward([loss0, loss1], [torch.ones_like(loss0), torch.ones_like(loss1)])

当loss是一个标量值的时候
此时反向传播求梯度比较好理解, 假设
&gt;&gt;&gt;]]></description><category></category></item><item><title><![CDATA[论文阅读: Interpretable and Generalizable Person Re-Identification with Query-Adaptive Convolution]]></title><link>https://blog.youkuaiyun.com/LuffysMan/article/details/112492317</link><guid>https://blog.youkuaiyun.com/LuffysMan/article/details/112492317</guid><author>LuffysMan</author><pubDate>Tue, 12 Jan 2021 02:01:32 +0800</pubDate><description><![CDATA[论文阅读: Interpretable and Generalizable Person Re-Identification with Query-Adaptive Convolution and Temporal Lifting
ECCV2020的一篇文章, 关于可泛化行人重识别. 提出了"查询自适应卷积"和"时序提举". 其中"查询自适应卷积"方法比较新颖, 放弃了传统的对高层特征进行分类, 计算损失的方法; 转而关注中间的特征图对一个class_memory的匹配, 一次来提升模型对未知场景的泛化能力]]></description><category></category></item><item><title><![CDATA[猴子分桃问题]]></title><link>https://blog.youkuaiyun.com/LuffysMan/article/details/111939982</link><guid>https://blog.youkuaiyun.com/LuffysMan/article/details/111939982</guid><author>LuffysMan</author><pubDate>Wed, 30 Dec 2020 00:48:22 +0800</pubDate><description><![CDATA[猴子分桃问题问题描述:方案1方案2
笔试的时候碰到一道题, 死活不得其解, 后来看了知乎的一个解答才茅塞顿开;
问题描述:
有未知数量香蕉一堆, 有m只猴子, 第一只猴子来了之后, 把香蕉分为5堆, 刚好多1个, 于是它吃掉1个, 拿走其中一堆;   第二只猴子来了之后, 同样把香蕉分5堆, 刚好多一个, 于是它吃掉1个, 拿走其中一堆;  剩余猴子依次到来, 做同样的操作;  问这堆香蕉至少有多少个?
方案1
暴力解: 设置香蕉初始值为1, 看是否满足同样的m次操作; 由于m可能非常大, 通常会超时;
方]]></description><category></category></item><item><title><![CDATA[论文阅读-Visualizing and Understanding Convolutional Networks]]></title><link>https://blog.youkuaiyun.com/LuffysMan/article/details/94711920</link><guid>https://blog.youkuaiyun.com/LuffysMan/article/details/94711920</guid><author>LuffysMan</author><pubDate>Sat, 06 Jul 2019 16:44:13 +0800</pubDate><description><![CDATA[作者: Matthew D.Zeiler
来源: ECCV 2014
评价: ZF-Net, 反卷积, 卷积层可视化
论文链接: PDF
1 Problem
There is no clear understanding of why large convolution network models work so well.
2 The proposed method

A novel way ...]]></description><category></category></item><item><title><![CDATA[多线程使用keras训练模型错误-"is not an element of this graph"]]></title><link>https://blog.youkuaiyun.com/LuffysMan/article/details/94591023</link><guid>https://blog.youkuaiyun.com/LuffysMan/article/details/94591023</guid><author>LuffysMan</author><pubDate>Wed, 03 Jul 2019 22:26:52 +0800</pubDate><description><![CDATA[错误场景
为了每隔固定时间训练一次模型, luffy在线程函数中设置timer再次调用线程函数, 简化版代码如下
def _thread_func(interval=10):
	model = train_model()				
	timer = threading.Timer(interval, _thread_func, args=(interval))		#设置定时器间隔interval后再...]]></description><category></category></item><item><title><![CDATA[关于keras中Embedding层的思考]]></title><link>https://blog.youkuaiyun.com/LuffysMan/article/details/94555029</link><guid>https://blog.youkuaiyun.com/LuffysMan/article/details/94555029</guid><author>LuffysMan</author><pubDate>Wed, 03 Jul 2019 15:01:38 +0800</pubDate><description><![CDATA[ luffy在使用keras中Embedding层的时候, 碰到了这样的问题:

如果我的vocabulary_size是10, 但是我训练集中的数据的词汇编号并不能完全覆盖词汇表, 比方说训练集里面只使用到了6-10的词;
vocabulary: {1,2,3,4,5,6,7,8,9,10}
train_pairs: [[6, 1], [7, 1], [8,1], [9,1], [10, 1],...]]></description><category></category></item><item><title><![CDATA[论文阅读-Rich feature hierarchies for accurate object detection and semantic segmentation]]></title><link>https://blog.youkuaiyun.com/LuffysMan/article/details/94356401</link><guid>https://blog.youkuaiyun.com/LuffysMan/article/details/94356401</guid><author>LuffysMan</author><pubDate>Mon, 01 Jul 2019 09:46:08 +0800</pubDate><description><![CDATA[作者: Ross Gir
来源: CVPR 2014
评价: RCNN
论文链接: PDF
1 Problem

labeled data is scarce.

3 The proposed method
A single and scalable detection algorithm - RCNN.

features:

can apply high capacity convolutio...]]></description><category></category></item><item><title><![CDATA[论文阅读-Deep Residual Learning for Image Recognition]]></title><link>https://blog.youkuaiyun.com/LuffysMan/article/details/93973700</link><guid>https://blog.youkuaiyun.com/LuffysMan/article/details/93973700</guid><author>LuffysMan</author><pubDate>Fri, 28 Jun 2019 15:36:52 +0800</pubDate><description><![CDATA[作者: Kaiming He et al.
来源: CVPR 2015
评价: ResNet,Very very deep networks, CVPR best paper
论文链接: https://arxiv.org/pdf/1512.03385.pdf
1 Problem

the problem and solution:

problem: the degradation proble...]]></description><category></category></item><item><title><![CDATA[论文阅读-Going Deeper with Convolutions]]></title><link>https://blog.youkuaiyun.com/LuffysMan/article/details/93166387</link><guid>https://blog.youkuaiyun.com/LuffysMan/article/details/93166387</guid><author>LuffysMan</author><pubDate>Sun, 23 Jun 2019 21:35:07 +0800</pubDate><description><![CDATA[作者: Christian Szegedy(贾扬清大神是三作哦)
日期: 2015
类型: conference article
来源: CVPR
评价:
论文链接:
1 Purpose
The efficiency fo the algorightms matters, especially on mobile and embedded devices. The author’s aim is ...]]></description><category></category></item><item><title><![CDATA[论文阅读-VERY DEEP CONVOLUTIONAL NETWORKS FOR LARGE-SCALE IMAGE RECOGNITION]]></title><link>https://blog.youkuaiyun.com/LuffysMan/article/details/92806661</link><guid>https://blog.youkuaiyun.com/LuffysMan/article/details/92806661</guid><author>LuffysMan</author><pubDate>Thu, 20 Jun 2019 10:41:04 +0800</pubDate><description><![CDATA[作者: Karen Simonyan et al.
日期: 2015
类型: conference article
来源: ICLR
评价:
论文链接:
1 Purpose

Investigating the effect of the convolutional network depth on its accuracy in the large-scale image recognition...]]></description><category></category></item><item><title><![CDATA[论文阅读-ImageNet Classification with Deep Convolutional Neural Networks]]></title><link>https://blog.youkuaiyun.com/LuffysMan/article/details/92698135</link><guid>https://blog.youkuaiyun.com/LuffysMan/article/details/92698135</guid><author>LuffysMan</author><pubDate>Mon, 17 Jun 2019 21:37:01 +0800</pubDate><description><![CDATA[作者: Alex Krizhevsky
日期: 2012.1.1
类型: article
来源: NIPS
评价:AlexNet, deeplearning breakthrough
论文链接:
1 Purpose

The competition: To classify the 1.2million high-resolution images in the ImageNet LSVRC-20...]]></description><category></category></item><item><title><![CDATA[论文阅读-A Fast Learning Algorithm for Deep Belief Nets]]></title><link>https://blog.youkuaiyun.com/LuffysMan/article/details/91895270</link><guid>https://blog.youkuaiyun.com/LuffysMan/article/details/91895270</guid><author>LuffysMan</author><pubDate>Fri, 14 Jun 2019 08:45:11 +0800</pubDate><description><![CDATA[作者: G.E.Hinton et. al.
日期: 2006
类型: article
来源: Neural Computation
评价: Deep Learning eve(深度学习前夕)
论文链接: http://www.cs.toronto.edu/~hinton/absps/ncfast.pdf
文章比较"硬核", 各种算法原理解释, 数学公式和术语. 而且作者真的是很喜欢用一些生物上的...]]></description><category></category></item><item><title><![CDATA[论文阅读-Reducing the Dimensionality of Data with Neural Networks]]></title><link>https://blog.youkuaiyun.com/LuffysMan/article/details/91441508</link><guid>https://blog.youkuaiyun.com/LuffysMan/article/details/91441508</guid><author>LuffysMan</author><pubDate>Tue, 11 Jun 2019 16:39:20 +0800</pubDate><description><![CDATA[作者: G.E.Hinton, R.R.Salakhutinov
日期: 2006.7
类型: reports
来源: Science(Journal)
Problems
Biggest problem
Reduce the dimensionality of data.
other problem
Gradient descent can be used for fine-tuing ‘auto...]]></description><category></category></item><item><title><![CDATA[论文阅读-Deep learning]]></title><link>https://blog.youkuaiyun.com/LuffysMan/article/details/91385104</link><guid>https://blog.youkuaiyun.com/LuffysMan/article/details/91385104</guid><author>LuffysMan</author><pubDate>Mon, 10 Jun 2019 21:42:18 +0800</pubDate><description><![CDATA[作者: Yann lecun, Yoshua Bengio, Geoffery Hinton
日期: 2015.5
类型: Review
来源: Nature(Journal)
Summary
Abstract
Deep learning allows models to learn representation of data automatically. It ha brought break...]]></description><category></category></item><item><title><![CDATA[如何读一篇学术论文]]></title><link>https://blog.youkuaiyun.com/LuffysMan/article/details/91325997</link><guid>https://blog.youkuaiyun.com/LuffysMan/article/details/91325997</guid><author>LuffysMan</author><pubDate>Sun, 09 Jun 2019 14:14:13 +0800</pubDate><description><![CDATA[ 这个秋天我在给研究生一年级的学术教授机构和组织经济学. 这门课的阅读任务相当繁重, 比起他们所习惯的本科课程和第一年的微观经济学, 计量经济学等. 我给他们讲, 你们不能仅仅成为有热情的读者, 而且要高效, 能够做最少的工作从一篇学术论文中提取出最有用的信息. 换句话说, 他们需要学习浏览的艺术.
 过去给学生讲这些的时候, 他们反馈说不知道如何去浏览. 所以几年之后汇总了一个小手册, 如何读一...]]></description><category></category></item><item><title><![CDATA[MarkText--一款还不错的markdown编辑器]]></title><link>https://blog.youkuaiyun.com/LuffysMan/article/details/90216484</link><guid>https://blog.youkuaiyun.com/LuffysMan/article/details/90216484</guid><author>LuffysMan</author><pubDate>Tue, 14 May 2019 22:17:28 +0800</pubDate><description><![CDATA[ 尝试了几款markdown编辑器, 先是markdownpad2, 写到后面老是出现无法预览的情况, 检查语法又看不出毛病. 又试了小书匠, 好像不支持表格, 或者收费版可以. 后来发现了MarkText, 据说类似typora, 试了下, 表格没问题, 语法基本都规范化了, 不用自己敲.
 但是遇到一个问题, 插入图片的时候, 拖拽的化, 会将本地图片上传到服务器, 然后生成一段链接. 几次之...]]></description><category></category></item><item><title><![CDATA[深度学习与应用--coolsite]]></title><link>https://blog.youkuaiyun.com/LuffysMan/article/details/90140829</link><guid>https://blog.youkuaiyun.com/LuffysMan/article/details/90140829</guid><author>LuffysMan</author><pubDate>Sun, 12 May 2019 16:09:40 +0800</pubDate><description><![CDATA[ZF-Net 可视化
conv_arithmetic: 一个技术关于卷积网络的技术报告, 对卷积, 反卷积, 空洞卷积有生动的动图演示, 有代码和论文.

ConvNetJS CIFAR-10 demo: 基于CIFAR-10来实现的卷积特征图的可视化.
Understanding Neural Networks Through Deep Visualization: 通过加入图像先验知识来可视化...]]></description><category></category></item></channel></rss>