Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Residual Attention Networks for Image Classification #35

Open
chullhwan-song opened this issue Aug 3, 2018 · 1 comment
Open

Residual Attention Networks for Image Classification #35

chullhwan-song opened this issue Aug 3, 2018 · 1 comment

Comments

@chullhwan-song
Copy link
Owner

chullhwan-song commented Aug 3, 2018

https://arxiv.org/abs/1704.06904
http://cs231n.stanford.edu/reports/2017/pdfs/939.pdf

@chullhwan-song
Copy link
Owner Author

chullhwan-song commented Aug 3, 2018

What?

  • CNN 계열 attention 초기 유명 논문 > spatial attention 부분에서
  • image classification task
  • 모듈처럼 기존 cnn net에 끼워넣음. > resnet

Residual Attention Network

  • Residual Unit, ResNeXt, Inception 에 쉽게 모듈하여 결합할수 있다.
    • 그럼 내가 필요한 CNN 모듈에서는 ㅠ
    • 이렇게 주장하지만, 실제로는 그렇지 않다라고 생각됨.(해보니 어려움)
  • 크게 두가지 부분 > 이두부분이 joint하여 구성 -> highway network의 방식 모사.
    image
    • mask branch(image)= soft mask branch
      • bottom-up and Top-Down 구조
      • residual unit과 max pooling의 조합
        image
      • 일종의 de-convolution를 이용한 segmentation과 구조와 유사
        • down sampling vs up sampling 구조
        • up sampling 할때, Linear interpolation
          • Linear interpolation의 수는 max pooling동일하다.
        • 이후, 1x1 conv 를 두번한다. 왜 하지?? > 차원을 줄이려는듯~
        • 최종적으로, sigmoid layer -normalize [0, 1]
    • trunk branch(image)
      • bottom-up 과 top-down (M(x))와 T(x)는 skip-connection

image
* i ranges over all spatial positions
* c is channel index
* 수식 (1)의 back-propa 가능
image
* image 는 branch parameter, image 는 trunk parameter

  • Attention Residual Learning = resnet과의 결합.
    image
    • F는 original feature
      image
    • 이 두구조를 가지는 속성은 robust to noisy labels.
      image
  • Spatial Attention and Channel Attention
    • soft mask output 전에 activation function 안에서 normalization step 형태로써 변환를 통해 mask branch와 결합가능 > 그니까 여기서는 이 activation function를 애기하려는듯~
      • ?? > 이부분이 저런
    • three types of activation functions > constrains to attention can still be added to mask branch by changing normalization step in activation function before soft mask output.
      image
    • Mixed attention f1 : Mixed attention f1 without additional restriction use simple sigmoid for each channel and spatial position
    • Channel attention f2 performs L2 normalization within all channels for each spatial position to remove spatial information.
    • Spatial attention f3 performs normalization within feature map from each channel
    • i ranges over all spatial positions
    • c ranges over all channels.
    • mean_c and std_c denotes the mean value and standard deviation of feature map from c-th channel
    • x_i denotes the feature vector at the ith spatial position.
  • 전체 network 구조
    image

실험

image
image

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

No branches or pull requests

1 participant