kyungheee / 2024-Samsung-AI-Challenge-Black-box-Optimization

2024 Samsung AI Challenge : Black-box Optimization
0 stars 0 forks source link

understanding diffusion model #30

Open rhycha opened 1 month ago

rhycha commented 1 month ago

image https://www.youtube.com/watch?v=HoKDTa5jHvg https://udlbook.github.io/udlbook/

rhycha commented 1 month ago

image

rhycha commented 1 month ago

image

rhycha commented 1 month ago

image

rhycha commented 1 month ago

동영상 23분 까지의 결론 log likelihood에서 온몸비틀기 하고 나면 노이즈 difference 제곱 오차로 변함.

image

rhycha commented 1 month ago

어차피 1d라서, 레이어마다 신경망 만들어도 될 것 같음. image

rhycha commented 1 month ago

only neural net(not diffusion model) causal architecture in neural net that allow you to train on all of these generation steps while only evaluating causal convolutional neural nets, and causal transformers.


other x variable can be conditional variable classifier free guidance make diffusion model work better

구현불가능 - 우린 프롬프트 없음 image

rhycha commented 1 month ago

이거 참조하시고, https://github.com/udlbook/udlbook/blob/main/Notebooks/Chap18/18_2_1D_Diffusion_Model.ipynb 원래 논문에서 superconductor 구현 어떻게 했는지 살펴봐야 함. (현재 작성하는 코드 굳이 작성 X)