CNN ์˜ˆ์‹œ

CNN

  • 32x32x3 ์ด๋ฏธ์ง€์— 0~9 ๊นŒ์ง€์˜ 10๊ฐ€์ง€ ์ˆซ์ž ์ค‘ ๋ฌด์—‡์˜ ์ด๋ฏธ์ง€์ธ์ง€ ๋งž์ถ”๋Š” ์‹ ๊ฒฝ๋ง์„ ๊ตฌ์„ฑํ•ด๋ณด์ž
  • input : 32ร—32ร—332\times{}32\times{}3
  • Layer 1
    1. conv1 : f=5,s=1โ†’convolutionโ†’28ร—28ร—6f=5, s=1\rightarrow{}\text{convolution}\rightarrow{}28\times{}28\times{}6
    2. pool1 : f=2,s=2โ†’maxย poolingโ†’14ร—14ร—6f=2, s=2\rightarrow\text{max pooling}\rightarrow{}14\times14\times6
  • Layer 2
    1. conv2 : f=5,s=1โ†’convolutionโ†’10ร—10ร—16f=5, s=1\rightarrow{}\text{convolution}\rightarrow{}10\times{}10\times{}16
    2. pool2 : f=2,s=2โ†’maxย poolingโ†’5ร—5ร—16f=2, s=2\rightarrow\text{max pooling}\rightarrow{}5\times5\times16
  • Layer 3 (FC3)
    • ๋ ˆ์ด์–ด 2์˜ ๊ฒฐ๊ณผ ํ–‰๋ ฌ์„ (400,1)(400, 1) ์˜ ๋ฒกํ„ฐ๋กœ ๋งŒ๋“ค๊ณ  ๋ ˆ์ด์–ด 3์˜ ์ธํ’‹ ๋ฐ์ดํ„ฐ๋กœ ์‚ฌ์šฉํ•œ๋‹ค.
    • W[3]โ†’(120,400)W^{[3]}\rightarrow{}(120, 400)
    • b[3]โ†’(120,1)b^{[3]}\rightarrow{}(120, 1)
  • Layer 4 (FC4)
    • W[4]โ†’(120,84)W^{[4]}\rightarrow{}(120, 84)
    • b[4]โ†’(84,1)b^{[4]}\rightarrow{}(84, 1)
  • Layer N
    • Softmax (10 outputs)
LayerActivation ShapeActivation SizeParameters
Input32x32x33,0720
Conv128x28x64,704(5x5x3)x6 + 6 = 456
Pool114x14x61,1760
Conv210x10x161,600(5x5x6)x16 + 16 = 2,416
Pool25x5x164000
FC3120x1120120x400 + 120 = 48,120
FC484x184120x84 + 84 = 10,164
Output10x11084x10 + 10 = 850

Total Parameters: 62,006