Generative Modeling by Estimating Gradients of the Data Distribution (Noise Conditional Score Network)

Diffusion Model의 시초인 Diffusion Probabilistic Models부터 Score-based Generative Model(NCSN), Denoising Diffusion Probabilistic Models(DDPM) 그리고 Denoising Diffusion Implicit Models(DDIM)까지 정리하는 시리즈의 세 번째 글에서는 Score-based Generative Model(NCSN)에 관해 리뷰해 볼 것이다.

 

이 논문을 이해하는 데 도움을 주는 전반적인 배경 지식과 내용은 아래 저자의 웹사이트에 잘 소개되어 있다.

https://yang-song.net/blog/2021/score/

Generative Modeling by Estimating Gradients of the Data Distribution | Yang Song

 

 

 

Noise Conditional Score Network

 

(1) Key Point

1) Yang Song et al. “Generative Modeling by Estimating Gradients of the Data Distribution.” arXiv:1907.05600 (2019)

 

 

(2) Score Function

1) Yang Song et al. “Generative Modeling by Estimating Gradients of the Data Distribution.” arXiv:1907.05600 (2019)

 

 

(3) Training and Inference

1) Yang Song et al. “Generative Modeling by Estimating Gradients of the Data Distribution.” arXiv:1907.05600 (2019)

 

 

 

(4) Code from scratch

 

https://github.com/Glanceyes/ML-Paper-Review/blob/main/ComputerVision/Diffusion/NCSN/NCSN.ipynb

GitHub - Glanceyes/ML-Paper-Review: Implementation of ML&DL models in machine learning that I have studied and written source co

 

 

 

(5) Summary

 

 

 

 

 

 

 

출처
1. Yang Song et al. “Generative Modeling by Estimating Gradients of the Data Distribution.” arXiv:1907.05600 (2019)
2. https://yang-song.net/blog/2021/score/