๐Ÿ“ AI & Bigdata

๐Ÿ“ AI & Bigdata/AI & ML & DL

[ML] GBM์ด๋ž€

๋ถ€์ŠคํŒ… ์•Œ๊ณ ๋ฆฌ์ฆ˜: ์—ฌ๋Ÿฌ ๊ฐœ์˜ ๋จธ์‹ ๋Ÿฌ๋‹์„ ์ˆœ์ฐจ์ ์œผ๋กœ ํ•™์Šตํ•˜๋ฉด์„œ ๊ฐ step์—์„œ ์ž˜๋ชป ์˜ˆ์ธก๋œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ๊ฐ€์ค‘์น˜๋ฅผ ๋ถ€์—ฌํ•ด ์˜ค๋ฅ˜๋ฅผ ๊ฐœ์„ ํ•ด ๋‚˜๊ฐ€๋Š” ํ•™์Šต ๋ฐฉ์‹ AdaBoost(Adaptive boosting) Gradient boost AdaBoost๊ฐ€ ํ•™์Šต ๋ฐฉ๋ฒ• GBM์—์„œ ํŒŒ์ƒ๋œ ๋ชจ๋ธ XGBoost LightGBM CatBoost GBM(Gradient Boost Model) -. ์ด๋•Œ ์—ฌ๋Ÿฌ ๊ฐœ์˜ ๋จธ์‹ ๋Ÿฌ๋‹์€ ์ˆœ์ฐจ์ ์œผ๋กœ ๊ตฌ์„ฑ์ด ๋˜๋ฉฐ, ๋จธ์‹ ๋Ÿฌ๋‹์ด ์˜ˆ์ธกํ•œ ๊ฐ’๊ณผ ์‹ค์ œ ๋ฐ์ดํ„ฐ ๊ฐ’ ์‚ฌ์ด์˜ ์ฐจ์ด๋ฅผ "์ž”์ฐจ"๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค. -. ์ฒซ๋ฒˆ์งธ ์ƒ์„ฑ๋œ ๋จธ์‹ ๋Ÿฌ๋‹์ด ๋ฐœ์ƒ์‹œํ‚จ ์ž”์ฐจ๋ฅผ ๋‘๋ฒˆ์งธ ์ƒ์„ฑ๋œ ๋จธ์‹ ๋Ÿฌ๋‹์ด ํ•™์Šตํ•˜๊ฒŒ ๋˜๋ฉฐ, N๋ฒˆ์งธ๊นŒ์ง€ ๋ฐ˜๋ณต์ ์œผ๋กœ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•์œผ๋กœ ์ƒ์„ฑ๋œ ๋ชจ๋ธ์€ ์ž”์ฐจ๋ฅผ ์ตœ์†Œํ™”์‹œํ‚ค๋ฉฐ, GBM์ด๋ผ ๋ถ€๋ฆ…๋‹ˆ๋‹ค. โ€ป ์˜ค์ฐจ์™€ ..

๐Ÿ“ AI & Bigdata/AI & ML & DL

[AI] XAI, eXplanable AI์ด๋ž€

์„ค๋ช… ๊ฐ€๋Šฅํ•œ ์ธ๊ณต์ง€๋Šฅ: XAI ๋จธ์‹ ๋Ÿฌ๋‹: ์ปดํ“จํ„ฐ ๋น„์ „, ํšŒ๊ท€ ๋ถ„์„, ์‹œ๊ณ„์—ด ์˜ˆ์ธก, ๋ถ„๋ฅ˜, ์Œ์„ฑ ์ธ์‹, ๋ฌธ์ž ์ธ์‹ ์— ์‚ฌ์šฉ ๋˜์—ˆ๋‹ค. ํ•˜์ง€๋งŒ ๋จธ์‹ ๋Ÿฌ๋‹, ๋ช‡๋ช‡ ๋”ฅ๋Ÿฌ๋‹ ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ์ž‘๋™ ๋ฐฉ์‹์€ black box์™€๋„ ๊ฐ™๋‹ค. ์ด๋Š” ๋ชจ๋ธ์˜ ์ž…๋ ฅ๊ฐ’, ์ถœ๋ ฅ๊ฐ’ ์ด์™ธ์˜ ์ž‘๋™ ์›๋ฆฌ๋ฅผ ์•Œ๊ธฐ ํž˜๋“ค๋‹ค๋Š” ๊ฒƒ์ด๋‹ค. ๋”ฅ๋Ÿฌ๋‹์€ ํ•™์Šต ๋ฐฉ์‹์€ ๊ฐ„๋‹จํ•œ ๋ฏธ๋ถ„ ๊ณ„์‚ฐ๋“ค๋กœ ์ด๋ฃจ์–ด์ ธ ์žˆ์ง€๋งŒ, ์ •ํ™•๋„ ๊ฐœ์„ ์„ ์œ„ํ•ด ๋ฌด์ˆ˜ํžˆ ๋งŽ์€ layer๋ฅผ ์Œ“์•„ ๊ฒฐ๊ณผ์ ์œผ๋กœ ๋ณต์žกํ•œ ๋ชจ๋ธ์ด ๋งŒ๋“ค์–ด ์ง„๋‹ค. ML ๋ชจ๋ธ black box: XGBoost, Random Forest XGBoost๋Š” Decision Tree์˜ Overfitting ๋‹จ์ ์„ ๊ทน๋ณตํ•˜๊ธฐ ์œ„ํ•ด Single Decision Tree๋ฅผ ์Œ“์€ ๊ฒƒ์ด๋‹ค. white box: Decision Tree XAI ๋ชจ๋ธ์˜ ์ •ํ™•๋„ ..

๐Ÿ“ AI & Bigdata/AI & ML & DL

[DL] TensorFlow- ์‹ฌ์ธต์‹ ๊ฒฝ๋ง ๋ชจ๋ธ ์ฝ”๋“œ ๊ตฌํ˜„

import: ํ•„์š”ํ•œ ๋ชจ๋“ˆ import ์ „์ฒ˜๋ฆฌ: ํ•™์Šต์— ํ•„์š”ํ•œ ๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ๋ง(model): ๋ชจ๋ธ์„ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค. ์ปดํŒŒ์ผ(compile): ๋ชจ๋ธ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ํ•™์Šต (fit): ๋ชจ๋ธ์„ ํ•™์Šต์‹œํ‚ต๋‹ˆ๋‹ค. ์ด๋ฏธ์ง€: ๋ฐ์ดํ„ฐ ๋กœ๋“œ 1) load_data()๋ฅผ ํ†ตํ•ด train, validation ๋‚˜๋ˆˆ๋‹ค. (x_train, y_train), (x_valid, y_valid) = fashion_mnist.load_data() ์ „์ฒ˜๋ฆฌ 1) ์ด๋ฏธ์ง€ ์ •๊ทœํ™” (Normalization)- ๋ชจ๋“  ํ”ฝ์…€์€ 0~255(8bit) x_train = x_train / 255.0 x_valid = x_valid / 255.0 3) Flatten- 2D> 1D x = Flatten(input_shape=(28, 28)) ..

๐Ÿ“ AI & Bigdata/AI & ML & DL

[DL] Faster R-CNN ๋„คํŠธ์›Œํฌ ์„ธ๋ถ€ ๊ตฌ์„ฑ

R-CNN(Regions with CNN features)์ด ์–ด๋–ป๊ฒŒ ์ž‘๋™ํ•˜๋Š”์ง€ ์ž์„ธํžˆ ์„ค๋ช…ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค. R-CNN (Girshick et al. 2013)* Fast R-CNN (Girshick 2015)* Faster R-CNN (Ren et al. 2015)* Section 1: Image Pre-Processing R-CNN ์ฃผ์š” ๋ฌธ์ œ ROI(Region Proposal Interest) ์‹๋ณ„: ์ด๋ฏธ์ง€์—์„œ ๊ฐœ์ฒด ROI(Region Proposal Interest) : class ํ™•๋ฅ  ๋ถ„ํฌ ๊ณ„์‚ฐ(ROI์— calss์˜ ๊ฐœ์ฒด๊ฐ€ ํฌํ•จ๋  ํ™•๋ฅ ), ๋ถ„๋ฅ˜ ๊ฒฐ๊ณผ ํ™•๋ฅ ์ด ๊ฐ€์žฅ ๋†’์€ class ์„ ํƒ Section 2: Network Organization Head Region Proposal Network (RPN) ..

SOIT
'๐Ÿ“ AI & Bigdata' ์นดํ…Œ๊ณ ๋ฆฌ์˜ ๊ธ€ ๋ชฉ๋ก (4 Page)