--- tags: Diffusion model ---   # DDPM(Denoising Diffusion Probabilistic Models) 定義一個噪聲過程,並嘗試學習一個神經網路來恢復噪聲過程 ## original loss function $$L_{vlb} := L_0 + L_1 + \dots + L_{T-1}+L_T \\ L_0 := -\log p_\theta(x_0|x_1) \\ L_{t-1} := D_{KL}(q(x_{t-1}|x_t,x_0) || p_\theta(x_{t-1}|x_t)) \\ L_T := D_{KL}(q(x_T|x_0) || p(x_T))\\ $$ ## simple loss function $$L_{simple} = E_{t,x_0,\epsilon}[||\epsilon - \epsilon_\theta(x_t,t)||^2]$$ 由神經網路添加的噪點應該要匹配到前向加噪的噪點  $$p_\theta(x_{t-1}|x_t) := \mathscr{N}(x_{t-1};\color{blue}{\mu_\theta}(x_t,t),\color{blue}{\Sigma_\theta}(x_t,t))$$ 猜測:想要找出 $x_{t-1} \rightarrow x_{t}$ 時的高斯雜訊的高斯分佈,並將 $x_{t-1} = x_{t} + \text{雜訊}$,以此來求出$x_{t-1}$ (U-net) 在論文中,$\Sigma_\theta$ 是被固定住的,真正要求的只有 $\mu_\theta$。 * $\Sigma_\theta(x_t,t) = \beta_t$(待確認) $$p_\theta(x_{0:T}):=p(X_T)\prod\limits_{t=1}^T p_\theta(x_{t-1}|x_t)$$ $$p(x_T)=\mathscr{N}(x_T;0,I) \leftarrow 高斯分佈$$
×
Sign in
Email
Password
Forgot password
or
By clicking below, you agree to our
terms of service
.
Sign in via Facebook
Sign in via Twitter
Sign in via GitHub
Sign in via Dropbox
Sign in with Wallet
Wallet (
)
Connect another wallet
New to HackMD?
Sign up