report.md

typora-copy-images-to
./acc.png

Abstract

Using VGG16 and ResNet18 to make image prediction.

Process

Raw Data Process

Using Data augmentation to make regularitzation

train dataset

• randomly crops the 32x32
• randomly horizontonal flips

####test dataset

• resize to 32x32
• crops the center with the size 32x32

Set Hyperparamater

• batch_size: 256
• epochs: 300
• loss function: cross entropy
• optimizer: SGD
  • weight_decay 0.0001
  • momentum: 0.9
  • initial learning rate: 0.1
    • learning rate shrink 0.1 when epoch reach 90th, 175th and 225 respectively

###Network Preparation

####VGG16

There are 16 layers, for each layers has [64, 64, M, 128, 128, M, 256, 256, M, 512, 512, 512, M, 512, 512, 512, M] channels with 3x3 filters, where 'M' is the maxpool.

For the classifier, there are 3 fully connected layer [4096, 4096, 10].

Using ReLU and dropout after every FC

ResNet18

There are 18 layers in ResNet18. Specificially, there are 8 blocks which have [64, 64, 128, 128, 256, 256, 512, 512] channels, and each block has two layers with szie 1x1 and 3x3 respectively.

And there are only one fuuly connected layer at the end with the size 10

##Result

VGG16

ResNet18