Skip to content
New issue

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

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

Already on GitHub? Sign in to your account

[20230205] Weekly AI ArXiv 만담 시즌2 - 4회차 #70

Open
jungwoo-ha opened this issue Feb 4, 2023 · 4 comments
Open

[20230205] Weekly AI ArXiv 만담 시즌2 - 4회차 #70

jungwoo-ha opened this issue Feb 4, 2023 · 4 comments

Comments

@jungwoo-ha
Copy link
Owner

jungwoo-ha commented Feb 4, 2023

News

ArXiv

  • Dreamix: Video Diffusion Models are General Video Editors

    • Inferece time에 이미지, 비디오 와 text prompt를 입력받아 비디오를 변환하는 Video diffusion model (VDM)연구 (from Google Research)
    • VDM을 video로만 finetune하면 모션 변화에 취약하다고함. 그래서 Temporal attention FT도 하지만 비디오의 Unordered frame들로 Temporal Attention/Conv 를 Frozen하고 Temporal attention masking도 함께 joint 학습
    • 기본적으로 물체 변환도 잘되고 prompt semantic에 맞게 잘 변환되는 것으로 보임
    • 변환된 비디오에서 디테일이 상당히 사라지는 (배경도 물체도) 부분은 개선해가야할 점으로 보임
    • 데모페이지: https://dreamix-video-editing.github.io/
      image
      image
  • Accelerating Large Language Model Decoding with Speculative Sampling --> 소개는 이동현님이 @dhlee347

    • 초거대 언어모델에서 디코딩 속도 올리기 위해 제안하는 새로운 sampling 기법 (from DeepMind)
    • 순서는 아래와 같습니다.
      • Generating a short draft of length 𝐾. This can be attained with either a parallel model (Stern
        et al., 2018) or by calling a faster, auto-regressive model 𝐾 times. We shall refer to this model
        as the draft model, and focus on the case where it is auto-regressive.
      • Scoring the draft using the larger, more powerful model from we wish to sample from. We shall
        refer to this model as the target model.
      • Using a modified rejection sampling scheme, accept a subset of the 𝐾 draft tokens from left to
        right, recovering the distribution of the target model in the process.
    • 그런데 사실 Google의 Fast Inference from Transformers via Speculative Decoding 랑 거의 같음. LAMDA vs. Chinchilla
      image
      image
  • 신기한 Unconditional Infinite Outdoor Scene Generation 연구

@ghlee3401
Copy link

ghlee3401 commented Feb 4, 2023

ArXiv (Audio and Speech Processing)

Text-To-Audio Papers

@scene-the-ella
Copy link

scene-the-ella commented Feb 5, 2023

News

@dhlee347
Copy link

dhlee347 commented Feb 5, 2023

Accelerating Large Language Model Decoding with Speculative Sampling

  • LLM의 Sampling latency를 줄이려는 노력 (2~2.5x decoding speedup)
  • 훨씬 가벼운 draft model이 길이 K의 짧은 draft 를 생성하게 한 후, 이것의 logit(혹은 prob)을 무거운 target model도 (병렬적으로) 계산하게 함.
  • 그 비율을 이용해서 modified rejection sampling을 진행하여 최종적으로 target model의 확률분포에서 샘플링한 효과를 내게 함.
  • 쉽고 간단한 토큰은 가벼운 draft model에서, 근데 그것이 너무 Target Model의 분포에서 벗어나려하면 무거운 target model에서 샘플링.
    image

image

  • Result
    image

  • Acceptance Rate
    image

Open Assistant

Flan Collection

image

image

image

image

image

@scene-the-ella
Copy link

Arxiv Prompt 4th.pdf

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

No branches or pull requests

4 participants