PAB超参数如何确定

[nebuladream]

对于PAB的threshold、gap应该如何确定合适的超参数，需要对比不同step的att数值变化么，有什么经验？以及如果使用full attention是否仍然适用？

[nebuladream]

另外这些模型的threshold都是几百，timestep只有训练的时候才会生效吧

[oahzxl]

1. the timestep in both training and inference are from 1000 to 0. so the default value is workable for inference.
2. to determine the hyper param, we will quantilize the difference of attention outputs for adjacent diffusion timesteps. Based on the difference, we can then finetune the threshold and gap by visualizing the results.
3. suitable for any attention