Deep Learning/Network

EfficientNet - Network ๊ตฌ์กฐ ๋ฐ ๊ตฌํ˜„

KimTory 2023. 2. 23. 00:30

๐Ÿ’กNetwork (์•„ํ‚คํ…์ฒ˜ ๊ตฌ์กฐ)

EfficientNet Network๋Š” ๋ฉ”์ธ์œผ๋กœ mobile inverted bottleneck convolution(MBConv)
block์„ ์‚ฌ์šฉํ•œ๋‹ค. MBConv block Depthwise separable conv์™€
Squeeze-and-excitation(se) ๊ฐœ๋…์„ ์ ์šฉํ•œ ๋ฐฉ์‹์ด๋‹ค.

 

๐Ÿ“ https://github.com/qubvel/efficientnet

 

GitHub - qubvel/efficientnet: Implementation of EfficientNet model. Keras and TensorFlow Keras.

Implementation of EfficientNet model. Keras and TensorFlow Keras. - GitHub - qubvel/efficientnet: Implementation of EfficientNet model. Keras and TensorFlow Keras.

github.com

๐Ÿ“ https://keras.io/examples/vision/image_classification_efficientnet_fine_tuning/

 

Keras documentation: Image classification via fine-tuning with EfficientNet

Image classification via fine-tuning with EfficientNet Author: Yixing Fu Date created: 2020/06/30 Last modified: 2020/07/16 Description: Use EfficientNet with weights pre-trained on imagenet for Stanford Dogs classification. View in Colab • GitHub source

keras.io


 

๐Ÿ‘‰ ํŠน์ง•

EfficientNet ๋ชจ๋ธ์€ Network Depth, Filter ๊ฐœ์ˆ˜(width), Image Resolution ํฌ๊ธฐ๋ฅผ
์ตœ์ ์œผ๋กœ ์กฐํ•ฉํ•˜์—ฌ ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ๊ทน๋Œ€ํ™” ํ–ˆ๋‹ค.

 ๐Ÿ‘‰ Depth, Width, Resolution์˜ ๋ชจ๋“  ์ฐจ์›์„ ๊ท ์ผํ•˜๊ฒŒ scalingํ•˜๋Š”
์ƒˆ๋กœ์šด ๋ฐฉ๋ฒ•์„ ์ œ์•ˆ

2019๋…„ ๊ธฐ์ค€ (FLOFS๋Š” ์ดˆ๋‹น ๋ถ€๋™์†Œ์ˆ˜์  ์—ฐ์‚ฐ ํšŸ์ˆ˜)

โœ๏ธ ๊ฐœ๋ณ„ Scaling ์š”์†Œ์— ๋”ฐ๋ฅธ ์„ฑ๋Šฅ ํ–ฅ์ƒ ํ…Œ์ŠคํŠธ

  • Depth, ๋„คํŠธ์›Œํฌ ๊นŠ์ด๋ฅผ ํ™•์žฅํ•˜๋Š” ๊ฒƒ์€ CNN์—์„œ ์ผ๋ฐ˜์ ์ธ ๋ฐฉ์‹์ด๋ฉฐ, ๊นŠ์ด๊ฐ€ ๊นŠ์„์ˆ˜๋ก ์ผ๋ฐ˜ํ™”์ ์ธ feature map์„ ์ถ”์ถœ ํ•  ์ˆ˜ ์žˆ๋‹ค. ํ•˜์ง€๋งŒ ๊ธฐ์šธ๊ธฐ ์†์‹ค์— ๋Œ€ํ•œ ๋ฌธ์ œ๊ฐ€ ์กด์žฌ
  • Width, scalingํ•˜๋Š” ๋ฐฉ๋ฒ•์€ ์ฃผ๋กœ ์†Œํ˜• ๋ชจ๋ธ์— ์‚ฌ์šฉ๋˜๋ฉฐ, width๊ฐ€ ํด์ˆ˜๋ก network๊ฐ€ ํ›จ์”ฌ ๋„“์–ด์ง€๋ฉด acc๊ฐ€ ๋น ๋ฅด๊ฒŒ saturate ๋˜๋Š” ๊ฒƒ์„ ํ™•์ธ ํ•  ์ˆ˜ ์žˆ๋‹ค.
  • Resolution, CNN์—์„œ ๊ณ ํ•ด์ƒ๋„ input image๋ฅผ ์‚ฌ์šฉ ํ• ์ˆ˜๋ก ์ •ํ™•๋„๊ฐ€ ํ–ฅ์ƒ ๋˜์ง€๋งŒ, ๋งค์šฐ ๋†’์€ ํ•ด์ƒ๋„๋Š” acc gain์ด ๋–จ์–ด์ง€๋Š” ๊ฒƒ์œผ๋กœ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๊ฐ€ ๋‚˜์™€ ์žˆ๋‹ค.


 ๐Ÿ‘‰ filter ์ˆ˜, network depth๋ฅผ ์ผ์ • ์ˆ˜์ค€ ์ด์ƒ ๋Š˜๋ ค๋„ ์„ฑ๋Šฅ ํ–ฅ์ƒ์ด ๋ฏธ๋น„ํ•˜๋‹ค. Resolution ๊ฒฝ์šฐ ์„ฑ๋Šฅ ํ–ฅ์ƒ์ด ์ง€์†๋จ
       imagenet dataset ๊ธฐ์ค€ 80% ์ •ํ™•๋„์—์„œ ๊ฐœ๋ณ„ scaling ์š”์†Œ๋ฅผ ์ฆ๊ฐ€ ์‹œํ‚ค๋”๋ผ๋„ ์„ฑ๋Šฅ ํ–ฅ์ƒ์ด ์–ด๋ ค์›€.

inflearn ๊ฐ•์˜ ์ž๋ฃŒ

โœ๏ธ Compound Scaling

  • Image ํ•ด์ƒ๋„๊ฐ€ ๋†’์„ ๊ฒฝ์šฐ, ๋” ํฐ Receptive Field๊ฐ€ ๋” ๋งŽ์€ Pixel์„ ํฌํ•จํ•˜๋Š” ๋น„์Šทํ•œ Feature๋ฅผ ์ž˜ Capture ํ•จ.
  • ๋” ๋งŽ์€ Pixel ์ˆ˜๋ฅผ ๊ฐ€์ง€๋ฉด ๋†’์€ ์ด๋ฏธ์ง€ ํ•ด์ƒ๋„์˜ ๋งŽ์€ ํ”ฝ์…€๋“ค์— ๋Œ€ํ•ด์„œ ์„ธ๋ฐ€ํ•œ ํŒจํ„ด์„ ์ž˜ Capture ํ•  ์ˆ˜ ์žˆ์Œ.
  • depth + resolution์„ ๊ฐ๊ฐ 1.0์œผ๋กœ ๊ณ ์ •, width๋งŒ ์ฆ๊ฐ€ํ•˜๊ฒŒ ๋˜๋ฉด acc ์„ฑ๋Šฅ 80% ์ˆ˜๋ ด
    (๊ฐ๊ฐ 2.0์œผ๋กœ ๊ณ ์ •ํ•˜๊ณ , width๋งŒ ๋ณ€ํ™” ์‹œ, ๋น„์Šทํ•œ FLOPS์ƒ์—์„œ ๋” ๋‚˜์€ ์„ฑ๋Šฅ ํ–ฅ์ƒ์ด ๋œ๋‹ค.)

๐Ÿ‘‰ Receptive Field(์ˆ˜์šฉ์˜์—ญ, ์ˆ˜์šฉ์žฅ)์€ output layer์˜ ๋‰ด๋Ÿฐ ํ•˜๋‚˜์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” input ๋‰ด๋Ÿฐ๋“ค์˜ ๊ณต๊ฐ„ ํฌ๊ธฐ


// ์ฐธ๊ณ  ์ž๋ฃŒ

๐Ÿ“ https://lynnshin.tistory.com/13

 

EfficientNet ๋ชจ๋ธ ๊ตฌ์กฐ

EfficientNet ์ •๋ฆฌ ๊ธ€ : [AI Research Paper Review/More] - EfficientNet ์ •๋ฆฌ EfficientNet ์ •๋ฆฌ ์ด์ „๊ธ€ : [AI/Self-Study] - EfficientNet ๋ชจ๋ธ ๊ตฌ์กฐ EfficientNet ๋ชจ๋ธ ๊ตฌ์กฐ EfficientNet - B0 baseline ๋„คํŠธ์›Œํฌ ๊ตฌ์กฐ EfficientNet B0 ์ „์ฒด

lynnshin.tistory.com

 

'Deep Learning > Network' ์นดํ…Œ๊ณ ๋ฆฌ์˜ ๋‹ค๋ฅธ ๊ธ€

ResNet - Network ๊ตฌ์กฐ ๋ฐ ๊ตฌํ˜„  (0) 2023.02.23
VGGNet16 - Network ๊ตฌ์กฐ ๋ฐ ๊ตฌํ˜„  (0) 2023.02.23