Skip to content
New issue

Have a question about this project? # for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “#”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? # to your account

[20211219] Weekly AI ArXiv 만담 - 34회차 #34

Closed
jungwoo-ha opened this issue Dec 19, 2021 · 4 comments
Closed

[20211219] Weekly AI ArXiv 만담 - 34회차 #34

jungwoo-ha opened this issue Dec 19, 2021 · 4 comments

Comments

@jungwoo-ha
Copy link
Owner

  • News
  • Arxiv
@kimyoungdo0122
Copy link

kimyoungdo0122 commented Dec 19, 2021

@veritas9872
Copy link

veritas9872 commented Dec 19, 2021

Self-attention Does Not Need O(n2) Memory

image

Arxiv: https://arxiv.org/abs/2112.05682v2

Unofficial GitHub (by me): https://github.com/veritas9872/Memory-Efficient-Self-Attention
Screenshot (5)

구글 브레인에서 Transformer 구조의 Self-Attention을 memory efficient하게 연산하는 새로운 방법을 제안했습니다.

널리 알려진 block matrix multiply를 softmax와 backpropagation에 적용되도록 수정을 하여 매우 단순한 방법으로 memory를 너무 많이 차치하는 문제점을 O(\sqrt(n))로 바꿀 수 있는 방법을 제시했습니다.

초거대 AI의 접근성 문제를 해결하는데 많은 도움을 줄 것 같습니다.

@veritas9872
Copy link

veritas9872 commented Dec 19, 2021

한국어로 된 대규모 음성 데이터를 소개해드립니다.

단, 데이터 전처리가 어려우니 유의하시길 바랍니다.

http://aihub-competition.or.kr/hangeul

https://aihub.or.kr/aihub-data/natural-language/about

@veritas9872
Copy link

확인해보니 Efficient attention PyTorch를 다른 분께서 12월 19일 이미 구현하셨습니다.

제가 JAX 문법을 잘못 이해해서 계속 오류가 난 것 같습니다. ㅎㅎㅎ

https://github.com/AminRezaei0x443/memory-efficient-attention

# for free to join this conversation on GitHub. Already have an account? # to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

3 participants