ImageNet xiazai

Download Data

Training set (138G)

The validation set (6.3G-50000 sheets)

train_label.txt

validation_label.txt

ps with the Thunder pretty fast, three days to get

Data decompression

tar xvf ILSVRC2012_img_train.tar -C ./train

tar xvf ILSVRC2012_img_val.tar -C ./val

For train data sets, unzip 1000 tar file, you need to unpack again, SFX script follows unzip.sh

dir=/data/srd/data/Image/ImageNet/train

for x in ls $dir/*tar
do
filename=basename $x .tar
mkdir $to you/$filename
tar -xvf $x -C $to you/$filename
done

tar rm *

• 1
• 2
• 3
• 4
• 5
• 6
• 7
• 8
• 9
• 10

Using data sets

Each file in the download folder is a good training set class picture, name tags corresponding folder in the download file meta.mat good label, which is a matlab file, scipy.io.loadmat can read the contents of the documents, verify 5000 is set at pictures, each picture in the corresponding label in ILSVRC2012_validation_ground_truth.txt.
Enhanced Data: randomly when taking pictures, and the picture zoom short side is 256, then randomized 224x224 cropped images, then the average value of each channel by subtracting the corresponding channel, randomly flipping around

Neural network model selection

Because DenseNet achieved over again, this time to play a game ResNeXt and Inception-ResNet-v2:

ResNeXt: feeling a little look online code implementation issue, split channel feeling and does not match the original meaning, but I trained a bit cifar-100 results and conclusions of the paper are not the same, so according to their own understanding to engage in an in imagenet on training results and compare the original fit

• 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: shining paper line and three block Reduction There are two kinds of stem these modules

      </div>
      <link href="https://csdnimg.cn/release/phoenix/mdeditor/markdown_views-258a4616f7.css" rel="stylesheet">
              </div>

Download Data

Training set (138G)

The validation set (6.3G-50000 sheets)

train_label.txt

validation_label.txt

ps with the Thunder pretty fast, three days to get

Data decompression

tar xvf ILSVRC2012_img_train.tar -C ./train

tar xvf ILSVRC2012_img_val.tar -C ./val

For train data sets, unzip 1000 tar file, you need to unpack again, SFX script follows unzip.sh

dir=/data/srd/data/Image/ImageNet/train