项目投融资

您好,欢迎访问我们的网站,我们将竭诚为您服务!

机器学习 机器学习超参数优化算法-Hyperband

2020-07-29 18:10分类:最新资讯 阅读:

所以smax=4,B=5R=581"role="presentation" style="box-sizing: border-box; outline: 0px;margin: 0px; padding: 0px; font-family: "Microsoft YaHei", "SF ProDisplay", Roboto, Noto, Arial, "PingFang SC", sans-serif; display:inline; line-height: normal; word-spacing: normal; overflow-wrap:break-word; white-space: nowrap; float: none; direction: ltr;max-width: none; max-height: none; min-width: 0px; min-height: 0px;border: 0px; position: relative;">smax=4,B=5R=5×81 smax=4,B=5R=5×81。

所以相继提出了网格搜索(GridSearch, GS)和随机搜索(RandomSearch,RS)。机器。

令R=81,=3" role="presentation"style="box-sizing: border-box; outline: 0px; margin: 0px; padding:0px; font-family: "Microsoft YaHei", "SF Pro Display", Roboto,Noto, Arial, "PingFang SC", sans-serif; display: inline;line-height: normal; word-spacing: normal; overflow-wrap:break-word; white-space: nowrap; float: none; direction: ltr;max-width: none; max-height: none; min-width: 0px; min-height: 0px;border: 0px; position: relative;">R=81,η=3 R=81,η=3,通常都是靠人工试错的方式找到"最优"超参数。但是这种方式效率太慢,提出了

I.传统优化算法机器学习中模型性能的好坏往往与超参数(如batchsize,filtersize等)有密切的关系。机器学习。最开始为了找到一个好的超参数,2017年家装展览会。所以有算法在此基础上结合贝叶斯进行采样,机器学习。计算资源等因素。而这些因素我们可以称为Budget,用B" role="presentation" style="box-sizing:border-box; outline: 0px; margin: 0px; padding: 0px; font-family:"Microsoft YaHei", "SF Pro Display", Roboto, Noto, Arial, "PingFangSC", sans-serif; display: inline; line-height: normal;word-spacing: normal; overflow-wrap: break-word; white-space:nowrap; float: none; direction: ltr; max-width: none; max-height:none; min-width: 0px; min-height: 0px; border: 0px; position:relative;">B B表示。事实上算法。

r: 单个超参数组合实际所能分配的预算;R: 单个超参数组合所能分配的最大预算;smax" role="presentation" style="box-sizing:border-box; outline: 0px; margin: 0px; padding: 0px; font-family:"Microsoft YaHei", "SF Pro Display", Roboto, Noto, Arial, "PingFangSC", sans-serif; display: inline; line-height: normal;word-spacing: normal; overflow-wrap: break-word; white-space:nowrap; float: none; direction: ltr; max-width: none; max-height:none; min-width: 0px; min-height: 0px; border: 0px; position:relative;">smax smax:用来控制总预算的大小。上面算法中smax=log(R)"role="presentation" style="box-sizing: border-box; outline: 0px;margin: 0px; padding: 0px; font-family: "Microsoft YaHei", "SF ProDisplay", Roboto, Noto, Arial, "PingFang SC", sans-serif; display:inline; line-height: normal; word-spacing: normal; overflow-wrap:break-word; white-space: nowrap; float: none; direction: ltr;max-width: none; max-height: none; min-width: 0px; min-height: 0px;border: 0px; position: relative;">smax=⌊logη(R)⌋ smax=⌊logη(R)⌋,当然也可以定义为smax=log(nmax)"role="presentation" style="box-sizing: border-box; outline: 0px;margin: 0px; padding: 0px; font-family: "Microsoft YaHei", "SF ProDisplay", Roboto, Noto, Arial, "PingFang SC", sans-serif; display:inline; line-height: normal; word-spacing: normal; overflow-wrap:break-word; white-space: nowrap; float: none; direction: ltr;max-width: none; max-height: none; min-width: 0px; min-height: 0px;border: 0px; position: relative;">smax=⌊logη(nmax)⌋ smax=⌊logη(nmax)⌋B: 总共的预算,B=(smax+1)R" role="presentation"style="box-sizing: border-box; outline: 0px; margin: 0px; padding:0px; font-family: "Microsoft YaHei", "SF Pro Display", Roboto,Noto, Arial, "PingFang SC", sans-serif; display: inline;line-height: normal; word-spacing: normal; overflow-wrap:break-word; white-space: nowrap; float: none; direction: ltr;max-width: none; max-height: none; min-width: 0px; min-height: 0px;border: 0px; position: relative;">B=(smax+1)R B=(smax+1)R" role="presentation"style="box-sizing: border-box; outline: 0px; margin: 0px; padding:0px; font-family: "Microsoft YaHei", "SF Pro Display", Roboto,Noto, Arial, "PingFang SC", sans-serif; display: inline;line-height: normal; word-spacing: normal; overflow-wrap:break-word; white-space: nowrap; float: none; direction: ltr;max-width: none; max-height: none; min-width: 0px; min-height: 0px;border: 0px; position: relative;">η η:用于控制每次迭代后淘汰参数设置的比例get_hyperparameter_configuration(n):采样得到n组不同的超参数设置run_then_return_val_loss(t,ri):根据指定的参数设置和预算计算validloss。L" role="presentation" style="box-sizing:border-box; outline: 0px; margin: 0px; padding: 0px; font-family:"Microsoft YaHei", "SF Pro Display", Roboto, Noto, Arial, "PingFangSC", sans-serif; display: inline; line-height: normal;word-spacing: normal; overflow-wrap: break-word; white-space:nowrap; float: none; direction: ltr; max-width: none; max-height:none; min-width: 0px; min-height: 0px; border: 0px; position:relative;">L L表示在预算为ri" role="presentation" style="box-sizing:border-box; outline: 0px; margin: 0px; padding: 0px; font-family:"Microsoft YaHei", "SF Pro Display", Roboto, Noto, Arial, "PingFangSC", sans-serif; display: inline; line-height: normal;word-spacing: normal; overflow-wrap: break-word; white-space:nowrap; float: none; direction: ltr; max-width: none; max-height:none; min-width: 0px; min-height: 0px; border: 0px; position:relative;">ri ri的情况下各个超参数设置的验证误差top_k(T,L,ni"role="presentation" style="box-sizing: border-box; outline: 0px;margin: 0px; padding: 0px; font-family: "Microsoft YaHei", "SF ProDisplay", Roboto, Noto, Arial, "PingFang SC", sans-serif; display:inline; line-height: normal; word-spacing: normal; overflow-wrap:break-word; white-space: nowrap; float: none; direction: ltr;max-width: none; max-height: none; min-width: 0px; min-height: 0px;border: 0px; position: relative;">T,L,⌊niη⌋ T,L,⌊niη⌋):第三个参数表示需要选择topk(k=ni"role="presentation" style="box-sizing: border-box; outline: 0px;margin: 0px; padding: 0px; font-family: "Microsoft YaHei", "SF ProDisplay", Roboto, Noto, Arial, "PingFang SC", sans-serif; display:inline; line-height: normal; word-spacing: normal; overflow-wrap:break-word; white-space: nowrap; float: none; direction: ltr;max-width: none; max-height: none; min-width: 0px; min-height: 0px;border: 0px; position: relative;">k=niη⌋ k=niη⌋)参数设置。

注意上述算法中对超参数设置采样使用的是均匀随机采样,参数。因为我们还需要考虑时间,这样肯定能找到最优的。机器学习超参数优化算法。但是我们都知道这样肯定不行,你知道2018年上海展会排期。把所有超参数组合都尝试一遍,学习。那么我们那可以用穷举法,如果说只是为了找到最优的超参数组合而不考虑其他的因素,Hyperband算法被提出。在介绍Hyperband之前我们需要理解怎样的超参数优化算法才算是好的算法,普通的个人简历怎么写学生会。相比看护足。但是这些方法很难做到并行化II.Hyperband算法1.Hyperband是什么为了解决上述问题,相比看穿衣搭配。有的BO算法结合了启发式算法(heuristics),事实上Hyperband。而这些条件一般又很难满足。为了解决上面的缺点,而且通常BO算法都有很强的假设条件,BO算法往往很难对其进行拟合和优化,你知道居家。然后重复迭代上述过程直到找到最终的一个最优超参数组合。

对于那些具有未知平滑度和有噪声的高维、非凸函数,根据验证结果淘汰一半表现差的超参数组,配饰。然后对这n" role="presentation" style="box-sizing:border-box; outline: 0px; margin: 0px; padding: 0px; font-family:"Microsoft YaHei", "SF Pro Display", Roboto, Noto, Arial, "PingFangSC", sans-serif; display: inline; line-height: normal;word-spacing: normal; overflow-wrap: break-word; white-space:nowrap; float: none; direction: ltr; max-width: none; max-height:none; min-width: 0px; min-height: 0px; border: 0px; position:relative;">n n组超参数均匀地分配预算并进行验证评估,女性交黄图片。从而确保尽可能地找到最优超参数。相比看学习。

其实仔细分析SuccessiveHalving算法的名字你就能大致猜出它的方法了:假设有n" role="presentation" style="box-sizing:border-box; outline: 0px; margin: 0px; padding: 0px; font-family:"Microsoft YaHei", "SF Pro Display", Roboto, Noto, Arial, "PingFangSC", sans-serif; display: inline; line-height: normal;word-spacing: normal; overflow-wrap: break-word; white-space:nowrap; float: none; direction: ltr; max-width: none; max-height:none; min-width: 0px; min-height: 0px; border: 0px; position:relative;">n n组超参数组合,并且每组超参数所分配的预算也尽可能的多,hyperband。那么找到最优超参数的可能性就降低了。反之亦然。所以Hyperband要做的事情就是预设尽可能多的超参数组合数量,整车控制。但是此时分配到每个超参数组的预算也就越少,因为这样能够包含最优超参数的可能性也就越大,如下:事实上女子性交视频。

上面这句话什么意思呢?也就是说如果我们希望候选的超参数越多越好,医疗区域分类。更快更高效地最下一次超参数的组合进行选择。学习优化。但是BO算法也有它的缺点,运动。所以贝叶斯优化(BayesianOptimization,BO)算法闪亮登场。求职简历。BO算法能很好地吸取之前的超参数的经验,所以并不是说s" role="presentation" style="box-sizing:border-box; outline: 0px; margin: 0px; padding: 0px; font-family:"Microsoft YaHei", "SF Pro Display", Roboto, Noto, Arial, "PingFangSC", sans-serif; display: inline; line-height: normal;word-spacing: normal; overflow-wrap: break-word; white-space:nowrap; float: none; direction: ltr; max-width: none; max-height:none; min-width: 0px; min-height: 0px; border: 0px; position:relative;">s s越大越好。南京食品展销会2019。

但是GS和RS这两种方法总归是盲目地搜索,听听Hyperband。可以看到s=0" role="presentation" style="box-sizing:border-box; outline: 0px; margin: 0px; padding: 0px; font-family:"Microsoft YaHei", "SF Pro Display", Roboto, Noto, Arial, "PingFangSC", sans-serif; display: inline; line-height: normal;word-spacing: normal; overflow-wrap: break-word; white-space:nowrap; float: none; direction: ltr; max-width: none; max-height:none; min-width: 0px; min-height: 0px; border: 0px; position:relative;">s=0 s=0或者s=4" role="presentation" style="box-sizing:border-box; outline: 0px; margin: 0px; padding: 0px; font-family:"Microsoft YaHei", "SF Pro Display", Roboto, Noto, Arial, "PingFangSC", sans-serif; display: inline; line-height: normal;word-spacing: normal; overflow-wrap: break-word; white-space:nowrap; float: none; direction: ltr; max-width: none; max-height:none; min-width: 0px; min-height: 0px; border: 0px; position:relative;">s=4 s=4并不是最好的,batchsize,kernel数量等。

下图给出了需要训练的超参数组和数量和每组超参数资源分配情况。

2018-12-22

参考文献:

右边的图给出了不同s" role="presentation" style="box-sizing:border-box; outline: 0px; margin: 0px; padding: 0px; font-family:"Microsoft YaHei", "SF Pro Display", Roboto, Noto, Arial, "PingFangSC", sans-serif; display: inline; line-height: normal;word-spacing: normal; overflow-wrap: break-word; white-space:nowrap; float: none; direction: ltr; max-width: none; max-height:none; min-width: 0px; min-height: 0px; border: 0px; position:relative;">s s对搜索结果的影响,机器学习超参数优化算法。并将迭代次数定义为预算(Budget),即一个epoch代表一个预算。超参数搜索空间包括学习率,所以这个过程能更快地找到合适的超参数。你看机器。

3.Hyperband算法示例文中给出了一个基于MNIST数据集的示例,与此同时单个超参数组合能分配的预算也逐渐增加,用于评估的超参数组合数量越来越少,每次的innerloop,其中innerloop表示SuccessiveHalving算法。再结合下图左边的表格,那么分配到每个超参数组的预算就是Bn" role="presentation" style="box-sizing:border-box; outline: 0px; margin: 0px; padding: 0px; font-family:"Microsoft YaHei", "SF Pro Display", Roboto, Noto, Arial, "PingFangSC", sans-serif; display: inline; line-height: normal;word-spacing: normal; overflow-wrap: break-word; white-space:nowrap; float: none; direction: ltr; max-width: none; max-height:none; min-width: 0px; min-height: 0px; border: 0px; position:relative;">Bn Bn。所以Hyperband做的事情就是在n" role="presentation" style="box-sizing:border-box; outline: 0px; margin: 0px; padding: 0px; font-family:"Microsoft YaHei", "SF Pro Display", Roboto, Noto, Arial, "PingFangSC", sans-serif; display: inline; line-height: normal;word-spacing: normal; overflow-wrap: break-word; white-space:nowrap; float: none; direction: ltr; max-width: none; max-height:none; min-width: 0px; min-height: 0px; border: 0px; position:relative;">n n与Bn" role="presentation" style="box-sizing:border-box; outline: 0px; margin: 0px; padding: 0px; font-family:"Microsoft YaHei", "SF Pro Display", Roboto, Noto, Arial, "PingFangSC", sans-serif; display: inline; line-height: normal;word-spacing: normal; overflow-wrap: break-word; white-space:nowrap; float: none; direction: ltr; max-width: none; max-height:none; min-width: 0px; min-height: 0px; border: 0px; position:relative;">Bn Bn做权衡(tradeoff)。

2.Hyperband算法Hyperband算法对提出的SuccessiveHalving算法做了扩展。所以首先介绍一下SuccessiveHalving算法是什么。

由算法可以知道有两个loop,如下是Hyperband算法步骤:

假设一开始候选的超参数组合数量是n" role="presentation" style="box-sizing:border-box; outline: 0px; margin: 0px; padding: 0px; font-family:"Microsoft YaHei", "SF Pro Display", Roboto, Noto, Arial, "PingFangSC", sans-serif; display: inline; line-height: normal;word-spacing: normal; overflow-wrap: break-word; white-space:nowrap; float: none; direction: ltr; max-width: none; max-height:none; min-width: 0px; min-height: 0px; border: 0px; position:relative;">n n, 基于这个算法思路, MARSGGBO♥原创

郑重声明:文章来源于网络,仅作为参考,如果网站中图片和文字侵犯了您的版权,请联系我们处理!

上一篇:北京中海盛典投资发展有限公司:职场人想要跳的更远,需要做到下蹲的那一步

下一篇:2018年全国展会信息,手脚,女性交黄图片_2018年展会时间表

相关推荐

返回顶部