[Week2] DL Basic - CNN [Day3]

*Convolution

• Contiunuous convolution
  

  

  
  
  
• Discrete convolution
  

  

  
  
  
• 2D image convolution
  

  

  
• 2D convolution in action
  

  

• RGB Image Convolution
  

  

 

 

 

*Convolutional Neural Networks

• Convolution and pooling layers : feature extraction
• Fully connected layer : decision making (e.g., classification) 
  • 최소화 시키는 추세 (parameter dependency : 파라미터가 많으면 학습이 잘 안되고, Generalization performance 감소)

 

 

 

*Stride

 

 

 

 

 

*Padding

 

 

 

*Convolution Arithmetic

• Padding (1),  Stride (1),  3 x 3 Kernel
  

  

• What is the number of parameters of this model?
  • The answer is 3 x 3 x 128 x 64 = 73,728
• Excercise
  

  

  number of parameters : 2048 * 2 x 1000 = 약 4M

  • parameter수를 줄이기 위해서는 dense layer를 줄여야 함

 

 

 

 

* why 1x1 Convolution?

• Dimension reduction
• To reduce the number of parameters while increasing the depth
• e.g., bottleneck architecture

 

 

 

 

 

*Modern CNN

 

*AlexNet

• Key ideas
  • ReLu activation
    1. Preserves properties of linear models
    2. Easy to optimize with gradient descent
    3. Good generalization
    4. Overcome the vanishing gradient problem
  • GPU implementation (2 GPUs)
  • Local response normalization, Overlapping pooling
  • Data augmentation
  • Dropout

 

 

 

 

*VGGNet

• Increasing depth with 3 x 3 convolution filters (with stride 1)
  • Receptive field : 커널이 커짐에 따라 고려되는 인풋의 크기
  • why 3 x 3 convolution?
    

    

    
    
    
• 1 x 1 convolution for fully connected layers
• Dropout (p=0.5)
• VGG16, VGG19

 

 

 

*GoogLeNet

• network-in-network (NiN) with inception blocks
• Inception blocks
  

  

  
  
  • 1 x 1 Conv를 적절히 섞어주며 dimension reducntion을 통해 parameter를 줄임
    

    

    1 x 1 convolution enables about 30% reduce of the number of parameters

    
    
    

 

 

*ResNet

• Deeper neural networks are hard to train
  • overfitting은 아니지만 layer가 깊어져도 학습률이 좋지 않음
    

    

• Add an identity map (skip connection)
  

  

  skip connection

  
  
• Bottleneck architecture
  • dimension reduction
    

    

 

*DenseNet

• DenseNet uses concatenation instead of addition
  

  

  concatenation

  
  
  
  • Dense Block
    • concatenates the feature maps of all preceding layers
    • The number of channels increases geometrically
  • Transition Block
    • BatchNorm -> 1x1 Conv -> 2x2 AvgPooling
    • Dimension reduction
      

      

 

 

 

 

*Computer Vision Applications

 

*Semantic Segmentation

 

 

• *Fully Convolutional Network
  • Convolutionalization
    

    

    
    
    • Left : 4 x 4 x 16 x 10 = 2,560
    • Right : 4 x 4 x 16 x 10 = 2,560
  • Transforming fully connected layers into convolution layers enables a classfication net to output a heat map
    

    

  • Deconvolution (conv transpose)
    • Convolutionalization을 수행하면 파라미터 수는 같지만 spatial dimension이 줄게 됨
    • 따라서 spatial dimension을 키우기 위한 Deconvolution을 함
      

      

  • Result
    

    

 

 

*Detection

• R-CNN

 

• SPPNet
  • R-CNN은 이미지 내의 bbox를 CNN을 다 통과 시켜야함
  • SPPNet은 CNN을 한번만 수행 
    

    

    
    
    
• Fast R-CNN
  • Selective search를 통한 bbox 제안
  • CNN 한번 수행 (SPPNet과 같음)
  • For each region, get a fixed length feature from ROI pooling
  • Two outputs : class and bounding-box regressor 
    

    

    
    
    
• Fatser R-CNN
  • 기존 Selective search단계를 Region Proposal Network로 바꿈
    

    

  • Region Proposal Network
    

    

    
    

    

    
    
    
  • 9 : Thress different region sizes (128, 256, 512) with three different ratios (1:1, 1:2, 2:1)
  • 4 : four bounding box regression parameters
  • 2 : box classification (wheter to use it or not)

 

 

• YOLO
  • No explicit bounding box sampling (compared with Fatser R-CNN) -> speed up
    

    

  • Given an imgae, YOLO divides it into SxS grid
    

    

  • Each cell predicts B bounding boxes (B=5)
    • box refinement (x / y / w / h)
    • confidence (of objectness)
  • Each cell predicts C class probabilities
  • In total, it becomes a tensor with SxSx(B*5+C) size
    • SxS : Number of cells of the grid
    • B*5 : B bounding boxes with offsets(x,y,w,h) and confidence
    • C : Number of classes
  • Result