Source

• (22) The Key Equation Behind Probability - YouTube a good youtube video of cross-entropy
• Huggingface perplexity code example
• Code: wikitext2 perplexity_llm_claude.py

Source

• (22) The Key Equation Behind Probability - YouTube a good youtube video of cross-entropy
• Huggingface perplexity code example
• Code: wikitext2 perplexity_llm_claude.py

Source

Flow Sampling (from $\mathbf{x}_0$ and $u_t$ to get $\mathbf{x}_t$)

Flow Sampling (from $\mathbf{x}_0$ and $u_t$ to get $\mathbf{x}_t$)

Flow $\phi_t$ = 連續的 (自己到自己) 座標變換

Flow $\phi_t$ 定義: $\mathbf{x}_t \triangleq \phi_t\left(\mathbf{x}_0\right)$, 顯然 $t=0 \to \mathbf{x}_0 = \phi_0\left(\mathbf{x}_0\right)$ $\phi_t(\mathbf{x})$ 必須可微分且可逆。

![[Pasted image 20250612202606.png]] 是是否直接用 flow + diffusion mixed approach 解決這個問題？

https://mlg.eng.cam.ac.uk/blog/2024/01/20/flow-matching.html

幾個常用公式：

Takeaways