[{"data":1,"prerenderedAt":229},["ShallowReactive",2],{"content-query-bIJxjrRvLA":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"cover":11,"type":12,"category":13,"body":14,"_type":223,"_id":224,"_source":225,"_file":226,"_stem":227,"_extension":228},"/technology-blogs/zh/1916","zh",false,"","【MindSpore易点通】应用实践系列-ResNet50图像分类精讲（连载一）","随着网络深度的增加，模型精度并不总是提升，并且这个问题显然不是由过拟合造成的，因为网络加深后不仅测试误差变高了，它的训练误差竟然也变高了。","2022-10-14","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2022/10/24/bccebdeb08154f0e8c34675e72984b33.png","technology-blogs","开发者分享",{"type":15,"children":16,"toc":214},"root",[17,25,42,55,60,65,70,79,90,95,100,105,113,124,129,134,142,147,155,160,165,170,178,183,191,196,201,209],{"type":18,"tag":19,"props":20,"children":22},"element","h1",{"id":21},"mindspore易点通应用实践系列-resnet50图像分类精讲连载一",[23],{"type":24,"value":8},"text",{"type":18,"tag":26,"props":27,"children":28},"p",{},[29,31,40],{"type":24,"value":30},"何凯明等人在2015年提出的ResNet，在ImageNet比赛classification任务上获得第一名，获评CVPR2016最佳论文(",{"type":18,"tag":32,"props":33,"children":37},"a",{"href":34,"rel":35},"https://arxiv.org/pdf/1512.03385.pdf",[36],"nofollow",[38],{"type":24,"value":39},"ResNet论文",{"type":24,"value":41},")。因为它“简单与实用”并存，之后许多目标检测、图像分类任务都是建立在ResNet的基础上完成的，成为计算机视觉领域重要的基石结构。",{"type":18,"tag":43,"props":44,"children":46},"h4",{"id":45},"一resnet要解决什么问题",[47,49],{"type":24,"value":48},"一、",{"type":18,"tag":50,"props":51,"children":52},"strong",{},[53],{"type":24,"value":54},"ResNet要解决什么问题？",{"type":18,"tag":26,"props":56,"children":57},{},[58],{"type":24,"value":59},"自从深度神经网络在ImageNet大放异彩之后，后来问世的深度神经网络就朝着网络层数越来越深的方向发展。直觉上我们不难得出结论：增加网络深度后，网络可以进行更加复杂的特征提取，因此更深的模型可以取得更好的结果。",{"type":18,"tag":26,"props":61,"children":62},{},[63],{"type":24,"value":64},"但事实并非如此，人们发现随着网络深度的增加，模型精度并不总是提升，并且这个问题显然不是由过拟合造成的，因为网络加深后不仅测试误差变高了，它的训练误差竟然也变高了。作者提出，这可能是因为更深的网络会伴随梯度消失/爆炸问题，从而阻碍网络的收敛。作者将这种加深网络深度但网络性能却下降的现象称为退化问题。",{"type":18,"tag":26,"props":66,"children":67},{},[68],{"type":24,"value":69},"下图是在CIFAR-10数据集上使用56层网络与20层网络训练误差和测试误差图，由图中数据可以看出，56层网络比20层网络训练误差和测试误差更大，随着网络的加深，其误差并没有如预想的一样减小。",{"type":18,"tag":26,"props":71,"children":72},{},[73],{"type":18,"tag":74,"props":75,"children":78},"img",{"alt":76,"src":77},"resnet_1.png","https://fileserver.developer.huaweicloud.com/FileServer/getFile/cmtybbs/001/057/1f0/09dd3443210010571f0ac01982dd2c01.20220925091912.55138679367394433882135701561813:50531023083928:2400:CBA530AF4F8DEBAE20475974ADDB938A9F91F0FC8EC6271786EFDEE68147E6CA.png",[],{"type":18,"tag":43,"props":80,"children":82},{"id":81},"二resnet怎么解决网络退化问题",[83,85],{"type":24,"value":84},"二、",{"type":18,"tag":50,"props":86,"children":87},{},[88],{"type":24,"value":89},"ResNet怎么解决网络退化问题",{"type":18,"tag":26,"props":91,"children":92},{},[93],{"type":24,"value":94},"随着网络层数的增加，梯度爆炸和梯度消失问题严重制约了神经网络的性能，研究人员通过提出包括Batch normalization在内的方法，已经一定程度上缓解了这个问题，但依然不足以满足需求。",{"type":18,"tag":26,"props":96,"children":97},{},[98],{"type":24,"value":99},"作者想到了构建恒等映射（Identity mapping）来解决这个问题，问题解决的标志是：增加网络层数，但训练误差不增加。为什么是恒等映射呢？20层的网络是56层网络的一个子集，56层网络的解空间包含着20层网络的解空间。如果我们将56层网络的最后36层全部短接，这些层进来是什么出来也是什么（也就是做一个恒等映射），那这个56层网络不就等效于20层网络了吗，至少效果不会相比原先的20层网络差吧。同样是56层网络，不引入恒等映射为什么就不行呢？因为梯度消失现象使得网络难以训练，虽然网络的深度加深了，但是实际上无法有效训练网络，训练不充分的网络不但无法提升性能，甚至降低了性能。",{"type":18,"tag":26,"props":101,"children":102},{},[103],{"type":24,"value":104},"那怎么构建恒等映射呢？简单地说，原先的网络输入x，希望输出H(x)。现在我们改一改，我们令H(x)=F(x)+x，那么我们的网络就只需要学习并输出一个残差F(x)=H(x)-x。作者提出，学习残差F(x)=H(x)-x会比直接学习原始特征H(x)简单的多。",{"type":18,"tag":26,"props":106,"children":107},{},[108],{"type":18,"tag":74,"props":109,"children":112},{"alt":110,"src":111},"2.png","https://fileserver.developer.huaweicloud.com/FileServer/getFile/cmtybbs/001/057/1f0/09dd3443210010571f0ac01982dd2c01.20220925091926.13101320590915761164140054351787:50531023083928:2400:2C58CD8548514102B59192F1ABF42C6EF7B5BE50F8F84FE9125501FC2B16323C.png",[],{"type":18,"tag":43,"props":114,"children":116},{"id":115},"三resnet网络结构与代码实现",[117,119],{"type":24,"value":118},"三、",{"type":18,"tag":50,"props":120,"children":121},{},[122],{"type":24,"value":123},"ResNet网络结构与代码实现",{"type":18,"tag":26,"props":125,"children":126},{},[127],{"type":24,"value":128},"ResNet主要有五种主要形式：Res18，Res34，Res50，Res101，Res152",{"type":18,"tag":26,"props":130,"children":131},{},[132],{"type":24,"value":133},"每个网络都包括三个主要部分：输入部分、输出部分和中间卷积部分（中间卷积部分包括如图所示的Stage1到Stage4共计四个stage）。尽管ResNet的变种形式丰富，但都遵循上述的结构特点，网络之间的不同主要在于中间卷积部分的block参数和个数存在差异。",{"type":18,"tag":26,"props":135,"children":136},{},[137],{"type":18,"tag":74,"props":138,"children":141},{"alt":139,"src":140},"4.png","https://fileserver.developer.huaweicloud.com/FileServer/getFile/cmtybbs/001/057/1f0/09dd3443210010571f0ac01982dd2c01.20220925091940.49717348586702939368517950771819:50531023083928:2400:68678CE7B2A0E30D1CD37C8E843F0DA43FFD07E5DA45A6C74A13BA6D9A6BEDB0.png",[],{"type":18,"tag":26,"props":143,"children":144},{},[145],{"type":24,"value":146},"最终网络结构如下：",{"type":18,"tag":26,"props":148,"children":149},{},[150],{"type":18,"tag":74,"props":151,"children":154},{"alt":152,"src":153},"1.png","https://fileserver.developer.huaweicloud.com/FileServer/getFile/cmtybbs/001/057/1f0/09dd3443210010571f0ac01982dd2c01.20220925091951.85205486564998282600414905389066:50531023083928:2400:AE3E639ECA83810932D09E8E666CDFC8B92C771F06B2264EBDE2F2DF16FE104C.png",[],{"type":18,"tag":26,"props":156,"children":157},{},[158],{"type":24,"value":159},"从上图可以看出，残差网络结构主要由两种，一种是Building Block，适用于较浅的ResNet网络，如ResNet18和ResNet34；另一种是Bottleneck，适用于层数较深的ResNet网络，如ResNet50、ResNet101和ResNet152。",{"type":18,"tag":26,"props":161,"children":162},{},[163],{"type":24,"value":164},"Building Block，也成为基本模块，输入特征图的尺寸和输出特征层的尺寸相同，两个特征图可以直接叠加；",{"type":18,"tag":26,"props":166,"children":167},{},[168],{"type":24,"value":169},"Bottleneck，下采样模块，即主干部分对输入特征图使用 stride=2 的下采样卷积，使得输入特征图的尺寸变成原来的一半，那么残差边部分也需要对输入特征图进行下采样操作，使得输入特征图经过残差边处理后的 shape 能和主干部分处理后的特征图 shape 相同，从而能够将残差边输出和主干输出直接叠加。",{"type":18,"tag":26,"props":171,"children":172},{},[173],{"type":18,"tag":74,"props":174,"children":177},{"alt":175,"src":176},"cke_1870.png","https://fileserver.developer.huaweicloud.com/FileServer/getFile/cmtybbs/001/057/1f0/09dd3443210010571f0ac01982dd2c01.20220925091513.30128454908272255395505007456481:50531023083928:2400:563BA4EF49121CF087FFFB015E7659B420C6FEC51F380AD0B33F08D1B63CD0EE.png",[],{"type":18,"tag":26,"props":179,"children":180},{},[181],{"type":24,"value":182},"以下图基本残差块为例，先对输入图像使用 1*1 卷积下降通道数，在低维空间下使用 3*3 卷积提取特征，然后再使用 1*1 卷积上升通道数，残差连接输入和输出，将叠加后的结果进过 relu 激活函数。",{"type":18,"tag":26,"props":184,"children":185},{},[186],{"type":18,"tag":74,"props":187,"children":190},{"alt":188,"src":189},"cke_3405.png","https://fileserver.developer.huaweicloud.com/FileServer/getFile/cmtybbs/001/057/1f0/09dd3443210010571f0ac01982dd2c01.20220925091639.86718360097921387458372571330174:50531023083928:2400:5B7179F3EB594C9CC2CAC888447FE937330F7FC59F14C5AD72C9DD98B3D70B91.png",[],{"type":18,"tag":26,"props":192,"children":193},{},[194],{"type":24,"value":195},"我们已经成功构建完单个残差单元的类，而残差结构就是由多个残差单元堆叠而来的，ResnNet50 中有 4 组残差结构，每个残差结构分别堆叠了 3，4，6，3 个残差单元，如下图所示。",{"type":18,"tag":26,"props":197,"children":198},{},[199],{"type":24,"value":200},"第一个残差结构中的第一个残差单元只需要调整输入特征图的通道数，不需要改变特征图的尺寸；而其他三个的残差结构的第一个残差单元不仅需要对输入特征图调整通道数，还要对输入特征图进行下采样操作。",{"type":18,"tag":26,"props":202,"children":203},{},[204],{"type":18,"tag":74,"props":205,"children":208},{"alt":206,"src":207},"ccc851d13a274730831db21cbcc0d332.png","https://fileserver.developer.huaweicloud.com/FileServer/getFile/cmtybbs/001/057/1f0/09dd3443210010571f0ac01982dd2c01.20220925092318.88819016630758705328082285001788:50531023083928:2400:C6686479C71A54A498A8173F261AF3EB0AB8FB1D309187606E08FB6630C8A087.png",[],{"type":18,"tag":26,"props":210,"children":211},{},[212],{"type":24,"value":213},"最后经过全局平均池化以及全连接后，网络构建完成。",{"title":7,"searchDepth":215,"depth":215,"links":216},4,[217,219,221],{"id":45,"depth":215,"text":218},"一、ResNet要解决什么问题？",{"id":81,"depth":215,"text":220},"二、ResNet怎么解决网络退化问题",{"id":115,"depth":215,"text":222},"三、ResNet网络结构与代码实现","markdown","content:technology-blogs:zh:1916.md","content","technology-blogs/zh/1916.md","technology-blogs/zh/1916","md",1776506116978]