24.07.30 CNN 3

 2D Dealing with shapes 

• C의 값은 Input과 Kernels에서 동일해야함 (즉, kernal이 input과 같은 차원의 행렬로 이뤄져있다)
• 코드에 익숙해지기 전까지 항상 shape의 개수를 따져보는 연습을 하자
• Numbers of parameters: (w x h x C +1) x D
• 1은 bias 

Multi-channel 2D Convolution

• 1D signal is convered into a 1D signal
• 2D signal into a 2D
• Input signal에서 neighboring parts인 것은 output signal에서도 이웃임  

Structure of Convolution Layer

 

• Input 
• Conv. blocks 
  • Convolution + activation (relu)
  • Convolution + activation (relu)
  • .. 반복 
  • Maxpooling
• Output
  • Fully connected layers
  • Softmax 

Summary

• Learn features in input image through convolution
  (input image에서 convolution을 통해 features를 배우고)
• Introduce non-linearity though activation function
  (activation fuctiond로 non-liearity를 주고)
• Reduce dimensionality and preserve spatial invariance with pooling
  (pooling을 통해 공간상의 invariance를 줄임)
• Convolution and pooling layers output high-level features of input
  (Convolution 과 pooling layers 을 통해 high level features를 출력) 
• Fully connected layer uses these features for classifying input image
  (fully connected layer는 이 features를 사용해서input image를 분류)
• Express output as probability of image belonging to a particular class  
  (outputs를 특정 class에 속할 확률로 표시)

CNN 구조

 

CNN with Tensorflow 

• MNIST example

 

#input layer

input_h = 28    # Input height

input_w = 28   # Input width

input_ch = 1    # Input channel (gray scale)

# (None, 28, 28, 1)

 

# first convolution layer

k1_h = 3

k1_w = 3

k1_ch = 32

p1_h = 2

p1_2 = 2

# (None, 14, 14, 32)

 

# second convolution layer

k2_h = 3

k2_w = 3

k2_ch = 64

p2_h = 2

p2_w = 2

# (None, 7, 7, 64)

 

## fully connected 

# flatten the features -> (None, 7*7*64)

conv_result_size = int((28/2*2)) * (28/(2*2)) *k2_ch)

n_hidden = 100

n_output = 10