U_Week_2_Day_8

수업 정리 　

강의 목록

[DL Basic] Convolution은 무엇인가?

• Convolution
  
  
• RGB Image Convolution
  
  
• Stack of Convolutions
  
• Convolutional Neural Networks
  • 피처를 추출하는 convolution layer, pooling layer, 마지막에 분류 또는 회귀 등의 역할을 위한 fully connected layer로 구성됨
  • 학습 파라미터의 수를 줄이기 위해 fully connected layer를 최소화하는 추세
    
• Stride
  • convolution filter(kernal)의 이동 간격
• Padding
  • covolution operation 후 버려지는 boundary 정보를 채워 input과 output의 shape을 동일하게 함.(일반적으로 zero-padding 진행)
• Convolution Arithmetic
  
  • kernal width * kernal height * number of input channel * number of output channel
    
    • gpu 2장에 나눠 넣어 *2 발생
    • $11\times11\times3\times48*2\approx35k$
    • $5\times5\times48\times128*2\approx307k$
    • $3\times3\times128*2\times192*2\approx884k$
    • $3\times3\times192\times192*2\approx663k$
    • $3\times3\times192\times128*2\approx442k$
    • $13*13*128*2\times2048*2\approx177M$
    • $2048*2\times2048*2\approx16M$
    • $2048*2\times1000\approx4M$
• 1x1 Convolution
  
  • 차원(채널) 축소
  • convolution layer를 깊게 쌓으면서 parameter 수를 줄일 수 있음

[DL Basic] Modern CNN - 1x1 convolution의 중요성

• AlexNet
  
  • Rectified Linear Unit(ReLU) activation
  • GPU implementation(2GPUs)
  • Local response normalization, Overlapping pooling
  • Data augmentation
  • Dropout
• VGGNet
  
  • Increasing depth with $3\times3$ convolution filters(with stride 1)
  • 1x1 convolution for fully connected layers
  • Dropout(p=0.5)
  • VGG16, VGG19
  • why $3\times3$ convolution?
    
• GoogLeNet
  
  • Inception blocks
    
    • can be seen as channel-wise dimension reduction
    • 
• ResNet
  • Add an identity map(skip connection)
    
  • result
    
  • 일반적으로 simple shortcut 사용
    
  • Bottleneck architecture
    
• DenseNet
  • Using concatenation instead of addition
    
    
  • Dense Block
    • each layer concatenates feature maps of all preceding layers
    • the number of channels increases geometrically
  • Transition Block
    • BatchNorm -> 1x1 Conv -> 2x2 AvgPooling
    • Dimension reduction
      

[DL Basic] Computer Vision Applications

• Semantic Segmentation
  
  • Full Convolutional Network
    • 
    • 
    • 
    • 
    • 
  • Deconvolution(conv transpose)
    
    
• Detection
  • R-CNN
    
    • 1. takes an input image
    • 2. extracts around 2,000 region proposals(using Selective search)
    • 3. compute features for each proposal(using AlexNet)
    • 4. classifies with linear SVMs
         
  • SPPnet
    
    • In R-CNN, the number of crop/warp is usually over 2,000 meaning that CNN must run more than 2,000 times
    • However, in SPPNet, CNN runs once
  • Fast R-CNN
    
    • 1. Takes an input and a set of bounding boxes
    • 2. Generated convolutional feature map
    • 3. For each region, get a fixed length feature from ROI pooling
    • 4. Two outputs: class and bouding box regressor
  • Faster R-CNN
    • Region Proposal Network + Fast R-CNN
    • 
    • 
    • 
    • YOLO
      
    • YOLO(v1) is an extremely fast object detection algorithm
      • baseline : 45fps / smaller version: 155fps
    • It simultaneously predicts multiple bounding boxes and class probabilities
      • No explicit bounding box sampling(compared with Faster R-CNN)
    • 
    • 1. Given an image, YOLO divides it into SxS grid
      • if the center of an object falls into the grid cell, that grid cell is responsible for detection
    • 2-1. Each cell predicts B bounding boxes(B=5)
      • each bounding box predicsts 1)box refinement(x/y/w/h), 2) confidence(of objectness)
    • 2-2. Each cell predicts C class probabilities
    • 3. 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
• Bounding box와 Class를 동시에 찾는 방향으로 논문이 발전하고 있다.
  
  

과제

• Convolution
  - 강의 보면서 완료

피어세션 정리

• 어제 optimize 강의에서 cost와 variance의 trade-off 관계 그리고 여기서 noise가 noise robustness 의 noise와 동일한건지
• 이고잉님 github 특강 관련 질문
• git 기초학습(https://learngitbranching.js.org)
• git 커밋규약(https://medium.com/hdackorea/commit-history%EB%A5%BC-%ED%9A%A8%EA%B3%BC%EC%A0%81%EC%9C%BC%EB%A1%9C-%EA%B4%80%EB%A6%AC%ED%95%98%EA%B8%B0-%EC%9C%84%ED%95%9C-%EA%B7%9C%EC%95%BD-conventional-commits-67b2114ac8e4)
• git 브랜치관리(https://techblog.woowahan.com/2553/)
• git 커밋팁(https://meetup.toast.com/posts/106)

느낀점

vscode을 통해 git을 사용하는 방법에 대해 특강을 진행해주신 이고잉님과 자리 마련해주신 운영진님들한테 너무 감사합니다ㅠㅠㅠ 내일도 열심히 들을게요!
여태까지 자연어만 공부했었고 처음으로 비젼 관련 강의를 들어보았는데, 아직은 낯설고 어렵기만 하다(다들 존경스럽다,,, 몇번 더 들어봐야겠다,,,) 아직 갈 길이 멀다 힘내자!!!!!