ImageNet训练完整流程 - 好文

<>下载数据

训练集(138G)
<http://www.image-net.org/challenges/LSVRC/2012/nnoupb/ILSVRC2012_img_train.tar>

验证集(6.3G-50000张)
<http://www.image-net.org/challenges/LSVRC/2012/nnoupb/ILSVRC2012_img_val.tar>

train_label.txt
<https://d11.baidupcs.com/file/de4dd213b25d1f86f66d81c48de41ea2?bkt=p3-000084b9708ef38ccbbda14ca1c08258c4e6&xcode=ce089cdfd351835eb41b685df317c6f569ccde56bc8cb64393d160629a08e20933e935411b0944a7a5c4786ed8c965620b2977702d3e6764&fid=4164754243-250528-605837973406253&time=1533027144&sign=FDTAXGERLQBHSK-DCb740ccc5511e5e8fedcff06b081203-qjT6Km7qOAjAMics6S9y60LP3VA%3D&to=d11&size=43832878&sta_dx=43832878&sta_cs=224&sta_ft=txt&sta_ct=5&sta_mt=5&fm2=MH%2CYangquan%2CAnywhere%2C%2Cbeijing%2Cany&resv0=cdnback&resv1=0&vuk=282335&iv=-2&newver=1&newfm=1&secfm=1&flow_ver=3&pkey=000084b9708ef38ccbbda14ca1c08258c4e6&sl=82640974&expires=8h&rt=sh&r=527814912&mlogid=298305419616337108&vbdid=3286963511&fin=train_label.txt&fn=train_label.txt&rtype=1&dp-logid=298305419616337108&dp-callid=0.1.1&hps=1&tsl=50&csl=78&csign=0vnYzTYv2VV%2Ff%2FRkrbacf8q2JPs%3D&so=0&ut=8&uter=4&serv=0&uc=1125290866&ic=4162093333&ti=76168191086d6f29c2f8d31801ca0ebd048b4e699a8b2eee&by=themis>

validation_label.txt
<https://d11.baidupcs.com/file/b72a388e3fda14a6f9f50b2d4a8b00fe?bkt=p3-1400b72a388e3fda14a6f9f50b2d4a8b00fe1b1bf78a0000001917d4&xcode=ce089cdfd351835ef94fde216b2bba0edc3c9774f14f8cb7756236dfbd40e18f23107062d7c87f046a06497ca5895130316128a2cdfcce4d&fid=4164754243-250528-6345911560019&time=1533027134&sign=FDTAXGERLQBHSK-DCb740ccc5511e5e8fedcff06b081203-rI4WwktTnyDvUI2sE8c45agVKKk%3D&to=d11&size=1644500&sta_dx=1644500&sta_cs=436&sta_ft=txt&sta_ct=5&sta_mt=5&fm2=MH%2CYangquan%2CAnywhere%2C%2Cbeijing%2Cany&resv0=cdnback&resv1=0&vuk=282335&iv=-2&newver=1&newfm=1&secfm=1&flow_ver=3&pkey=1400b72a388e3fda14a6f9f50b2d4a8b00fe1b1bf78a0000001917d4&sl=82640974&expires=8h&rt=sh&r=650735058&mlogid=298302698122890737&vbdid=3286963511&fin=validation_label.txt&fn=validation_label.txt&rtype=1&dp-logid=298302698122890737&dp-callid=0.1.1&hps=1&tsl=50&csl=78&csign=0vnYzTYv2VV%2Ff%2FRkrbacf8q2JPs%3D&so=0&ut=8&uter=4&serv=0&uc=1125290866&ic=4162093333&ti=e3357e20d22cf84873280fc2893a49764acd85a43c2f0539&by=themis>

p.s. 用迅雷下还挺快的，3天搞定

<>数据解压

tar xvf ILSVRC2012_img_train.tar -C ./train

tar xvf ILSVRC2012_img_val.tar -C ./val

对于train数据集，解压后是1000个tar文件，需要再次解压，解压脚本unzip.sh如下
dir=/data/srd/data/Image/ImageNet/train for x in `ls $dir/*tar` do
filename=`basename $x .tar` mkdir $dir/$filename tar -xvf $x -C $dir/$filename
done rm *.tar
<>使用数据集

下载好的训练集下的每个文件夹是一类图片，文件夹名对应的标签在下载好标签文件meta.mat中，这是一个matlab文件，scipy.io.loadmat可以读取文件内容，验证集下是5000张图片，每张图片对应的标签在ILSVRC2012_validation_ground_truth.txt中。
数据增强：取图片时随机取，然后将图片放缩为短边为256，然后再随机裁剪224x224的图片，再把每个通道减去相应通道的平均值，随机左右翻转

<>神经网络模型选择

因为DenseNet实现过了，这次来玩一玩ResNeXt和Inception-ResNet-v2：

ResNeXt：感觉看网上代码实现都有点问题，split通道感觉都和原文的意思不符，而且我训练了一下cifar-100结果和论文中的结论也不一样，所以就按自己的理解搞了一个，在imagenet上训练结果和原文比较吻合

* blocks of ResNeXt:
256d(in)-(256,1x1,128)-(3x3,32x4d)-(128,1x1,256)-256d(out)
* Downsampling is done by stride-2 convolutions in the 3×3 layer of the first
block in each stage.（shortcut用stride-2的2x2的平均池化）
* The identity shortcuts can be directly used when the input and output are
of the same dimensions. When the dimensions increase, we consider two options:
(A) The shortcut still performs identity mapping, with extra zero entries
padded for increasing dimensions. This option introduces no extra parameter;
(B) The projection shortcut is used to match dimensions (done by 1×1
convolutions). For both options, when the shortcuts go across feature maps of
two sizes, they are performed with a stride of 2.（我采用了直接补0通道的方式）
result: 50-layer

* top 5 acc: 0.92708
* top 1 acc: 0.7562
Inception-ResNet-v2：照着论文撸，三种block两种Reduction还有stem这几个模块

热门工具换一换