请问训练的时候，怎么训练？

[henbucuoshanghai]

训练的时候，三阶段怎么划分的？依旧三阶段？0-0.3第一个？0.3-0.6第二个？不同阶段的dit输出后上采样进下一个stage？
感谢

[feifeiobama]

• The $K$ time windows are simply divided with uniformly spaced endpoints, namely $e_k=1-\frac{k}{K}$. According to Equation (26) in the paper, these time windows are $[s_k,e_k]=\big[\frac{K-k-1}{K+k+1}, 1-\frac{k}{K}\big]$. For $K=3$, these time windows are $\big[0,\frac{1}{3}\big],\big[\frac{1}{5},\frac{2}{3}\big],\big[\frac{1}{2},1\big]$.
• During training, the loss is directly computed by Equations (9-11) in the paper. During inference, we upsample and renoise the output from previous stage, according to Equation (15) in the paper.