Skip to content
Seongsu Park edited this page Jul 22, 2023 · 36 revisions

Diffusers ๋ฒˆ์—ญ

๋‹จ์–ด ์ •๋ฆฌ

์˜์–ด ๊ทธ๋Œ€๋กœ ์‚ฌ์šฉํ•  ๋‹จ์–ด๋“ค

  • (ํŠน์ • ์šฉ์–ด๋ฅผ ๊ฐ€๋ฆฌํ‚ค๋Š”) ๋Œ€๋ฌธ์ž๋กœ ์‹œ์ž‘ํ•˜๋Š” ๊ธ€์ž๋“ค
  • Hugging Face
  • Diffusers
  • text-to-image
  • sliced attention
  • memory-efficient attention
  • tracing
  • forward pass
  • just-in-time
  • Channels Last (๋ฉ”๋ชจ๋ฆฌ ํ˜•์‹)
  • diffusion
  • unconditional, conditional
  • toolbox
  • shape
  • optimizer
  • textual-inversion
  • LoRA
  • PyTorch
  • ToMe, Token Merging
  • Python
  • snapshot

ํ•œ๊ธ€๋กœ ๋ฒˆ์—ญํ•  ๋‹จ์–ด๋“ค

  • push: ํ‘ธ์‹œ
  • step: ์Šคํ…
  • convolution: ์ปจ๋ณผ๋ฃจ์…˜
  • section: ์„น์…˜
  • fine tuning: ํŒŒ์ธํŠœ๋‹
  • licence: ๋ผ์ด์„ ์Šค
  • load: ๋ถˆ๋Ÿฌ์˜ค๊ธฐ
  • inference: ์ถ”๋ก 
  • inference time: ์ถ”๋ก ์‹œ๊ฐ„
  • custom: ์ปค์Šคํ…€
  • code snippet: ์˜ˆ์‹œ ์ฝ”๋“œ...
  • submodule: ์„œ๋ธŒ๋ชจ๋“ˆ
  • branch: ๋ธŒ๋žœ์น˜
  • directory: ๋””๋ ‰ํ† ๋ฆฌ
  • checkpoint: ์ฒดํฌํฌ์ธํŠธ
  • resume: ์žฌ๊ฐœ
  • gradient: ๊ทธ๋ž˜๋””์–ธํŠธ
  • model: ๋ชจ๋ธ
  • system: ์‹œ์Šคํ…œ
  • lession: ๋‹จ์›
  • reproducibility: ์žฌํ˜„์„ฑ
  • prompt: ํ”„๋กฌํ”„ํŠธ
  • pre-train: ์‚ฌ์ „ํ•™์Šต
  • instance: ์ธ์Šคํ„ด์Šค
  • inpainting: ์ธํŽ˜์ธํŒ…
  • denoise: ๋””๋…ธ์ด์ฆˆ
  • repository: ๋ฆฌํฌ์ง€ํ† ๋ฆฌ

ํ•œ๊ธ€๊ณผ ์˜์–ด ๋ณ‘๊ธฐํ•  ๋‹จ์–ด๋“ค

  • ์ฐจ์›์ˆœ์„œ(dimensions ordering)
  • ๋‹ค์ค‘ ์ด๋ฏธ์ง€ ๋ฐฐ์น˜(multi-image batches)
  • ํ˜ผํ•ฉ ์ •๋ฐ€๋„(mixed precision) // ํ˜ผํ•ฉ ์ •๋ฐ€๋„๋ผ๋Š” ํ‘œํ˜„์ด ์žˆ๋Š” ์ค„ ๋ชฐ๋ž๋„ค์š”(ํ™๊ทœ)
  • ๋ฐ˜์ •๋ฐ€๋„(half precision)
  • attention: ์–ดํ…์…˜
  • prior-preserving(์‚ฌ์ „ ๋ณด์กด)
  • batch(๋ฐฐ์น˜)
  • ์ข…์†์„ฑ(dependency)

๋…ผ์˜๊ฐ€ ํ•„์š”ํ•œ ๋‹จ์–ด๋“ค