๋ชฉ๋กSwin Transformer (1)

SJ_Koding

ConvNeXt (A ConvNet for the 2020s, facebook) ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ

- ๋ถ€์ œ: ConvNeXt ์ดํ•ดํ•˜๊ธฐ ์ตœ์ข… - Image Classification model์„ ํ•™์Šตํ•  ์ƒํ™ฉ์ด ์ƒ๊ฒจ์„œ ConvNeXt๋ฅผ ์‚ฌ์šฉํ•ด๋ดค๋Š”๋ฐ ์„ฑ๋Šฅ์ด ์ •๋ง ์ข‹๊ฒŒ ๋‚˜์™”๋‹ค. ๋Œ€์ฒด ์ด ๋ชจ๋ธ์ด ๋ญ๊ธธ๋ž˜? ํ•˜๊ณ  ๋ดค๋”๋‹ˆ Facebook์ด ๋ฐœํ‘œํ•œ ๋…ผ๋ฌธ์ด์–ด์„œ ํฅ๋ฏธ๊ฐ€ ์ƒ๊ฒจ ์ด ๋…ผ๋ฌธ์„ ๊ณต๋ถ€ํ•˜๊ฒŒ๋˜์—ˆ๋‹ค. ๋…ผ๋ฌธ์„ ์ฝ์œผ๋ฉด์„œ ์ •๋ง CNN์— ๋Œ€ํ•ด ๊ณต๋ถ€ํ•˜๊ธฐ ์ข‹์€ ๋…ผ๋ฌธ์ด๋ผ๋Š” ์ƒ๊ฐ์ด ๋“ค์—ˆ๋‹ค. ๋˜, ๋…ผ๋ฌธ๊ตฌ์„ฑ์ด ์ •๋ง ์žฌ๋ฐŒ๋‹ค. ์—ฐ๊ตฌ์— ๋”ฐ๋ฅธ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ์ˆœ์ฐจ์ ์œผ๋กœ ๋ฐฐ์น˜ํ•˜์—ฌ ๋งˆ์น˜ ํ•จ๊ป˜ ์—ฐ๊ตฌ๋ฅผ ํ•œ ๊ฒƒ๊ณผ ๊ฐ™์€ ๋Š๋‚Œ์„ ์ค€๋‹ค. ์–ด๋–ค ๋‚ด์šฉ์ด๊ธธ๋ž˜? extreme ์š”์•ฝ: CNN๊ณ„์—ด ๋ชจ๋ธ(ResNet50์‚ฌ์šฉ)์„ ์ตœ์‹ ์— ์•Œ๋ ค์ง„ ๊ธฐ๋ฒ•์„ ์ ์šฉํ•˜๋ฉด์„œ ๊ทนํ•œ์œผ๋กœ ๋Œ์–ด์˜ฌ๋ ค๋ณด์ž~! ๋Œ€์‹ , pure CNN์œผ๋กœ. ์„ฑ๊ณต~! Abstract 2020๋…„๋Œ€, Vision Task์—์„œ..

Deep Learning 2024. 4. 4. 13:46