๋ณธ๋ฌธ ๋ฐ”๋กœ๊ฐ€๊ธฐ

728x90

VGG16

4/21 ๋ชฉ ๋ชฉ์š”์ผ! ์˜ค๋Š˜์€ Fine Tuning์„ ๋ฐฐ์šฐ๊ณ  CNN์„ ๋งˆ๋ฌด๋ฆฌ ์ง“๋Š”๋‹ค. DNN - ๊ฐ€์ง„ ์ด๋ฏธ์ง€๋“ค์˜ ํ”ฝ์…€์„ ํ•™์Šต CNN - ๊ฐ€์ง„ ์ด๋ฏธ์ง€๋“ค์˜ ํŠน์„ฑ์„ ์ถ”์ถœํ•˜์—ฌ ํ•™์Šต ์ถฉ๋ถ„ํ•œ ์–‘์˜ ๋ฐ์ดํ„ฐ๊ฐ€ ์—†๋Š” ๊ฒฝ์šฐ ์ด๋ฏธ์ง€ ์ฆ์‹(Augmentation) ๊ธฐ๋ฒ•์„ ์‚ฌ์šฉ ํ•™์Šต ์‹œ๊ฐ„์„ ์ค„์ด๊ณ  ๋” ์ข‹์€ filter๋ฅผ ์ด์šฉํ•˜๊ธฐ ์œ„ํ•ด ์ „์ดํ•™์Šต(Transfer Learning)์„ ์‚ฌ์šฉ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ์…‹ ํ•™์Šต ์‹œ Feature Extraction(CNN) ์ดํ›„ Classification(๋ถ„๋ฅ˜๊ธฐ)๋กœ DNN์ด ์ ํ•ฉํ•˜์—ฌ ์‚ฌ์šฉํ•˜๊ณ  ์žˆ์ง€๋งŒ, ๋‹ค๋ฅธ ๋ถ„๋ฅ˜๊ธฐ๋“ค์ด ๋งŽ์ด ์žˆ์Œ(SVM, Decision Tree, KNN, Naive Bayes, ์•™์ƒ๋ธ” ๊ธฐ๋ฒ•) * ์ „์ดํ•™์Šต์˜ Pretrained Network ์ค‘ EfficientNet๊ณผ ResNet ์„ฑ๋Šฅ์ด ์ข‹์Œ 1. .. ๋”๋ณด๊ธฐ
4/20 ์ˆ˜ ์ˆ˜์š”์ผ! ์˜ค๋Š˜์€ ์ด๋ฏธ์ง€ ์ฆ์‹(Image Augmentation)๊ณผ ์ „์ดํ•™์Šต(Transfer Learning)์— ๋Œ€ํ•ด ๋ฐฐ์šด๋‹ค. AWS ์„œ๋ฒ„ ์ผœ๊ณ  PuTTY๋กœ ๊ฐ€์ƒํ™˜๊ฒฝ ์—ด์ž~ (AWS GPU ์‚ฌ์šฉํ•˜๋ ค๋ฉด, ์ฟ ๋‹ค ์ ์šฉํ•˜๊ณ  ์ฝ”๋“œ ์“ธ ๋•Œ ์„ค์ • ์ถ”๊ฐ€ํ•ด์•ผ ํ•จ) Over-Fitting ์ค„์ด๋Š” ๋ฐฉ๋ฒ•์€ ๋งŽ์€ ์–‘์˜ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉ, feature์˜ ๊ฐœ์ˆ˜๋ฅผ ์ค„์ž„, ๊ทœ์ œ(L1, L2) ์‚ฌ์šฉ, Dropout ์‚ฌ์šฉ! conda activate machine_TF2_18 jupyter notebook --ip=0.0.0.0 --no-browser --port=8918 1. 4000๊ฐœ์˜ ์ ์€ ๋ฐ์ดํ„ฐ๋กœ ๋ชจ๋ธ๋ง # ์ผ๋ถ€ ์ด๋ฏธ์ง€ ๋ถ„๋ฆฌ(์ด 4000๊ฐœ) import os, shutil original_dataset_dir = './data/cat.. ๋”๋ณด๊ธฐ

728x90