Qingyun
Qingyun
> I no longer focus on this problem. But ddim results usually are a little worser than ddpm ones. okkkk The results are indeed different. It seems that my implementation...
@Janspiry Does this repo consider reviewing a PR about supporting ddim for sr3? I can create a pr, but I require some supports about knowledge, experience and code reviewing.
@wangzhen699 You can refer to [denoising-diffusion-pytorch](https://github.com/lucidrains/denoising-diffusion-pytorch). The code has been used in a colleague's private project.
没碰见过bug
> > 没碰见过bug > > 你好，我想问下，在改为ddim后，模型是不是还要重新训练，即使是只改了推断过程，好像也不能直接使用之前sr3的权重了 不是吧，我理解ddim是用来eval的吧
> > > > 我理解的ddim是改变了DDPM中的inference，对p(x{t-1} | x{t},x{0})进行了改变，然后获取正向过程的一段子序列作为实际的采样次数来缩短inference的时间。所以我一直在对model中的diffusion文件进行改动。 嗯嗯，差不多
@wangzhen699 没出现，我实现的很简单就work了，我是参考 pytorch denoising diffusion 的仓库实现的
OK, thanks for your reply! I have understood.
@Czm369 This work is directed by @jshilong.
## Align Inference Accuracy configs/maskformer/maskformer_r50_ms-16xb1-75e_coco.py develop/new_maskformer_r50_mstrain_16x1_75e_coco_20220221_141956-bc2699cb.pth (need to be converted) ### MaskFormer ``` +--------+--------+--------+--------+------------+ | | PQ | SQ | RQ | categories | +--------+--------+--------+--------+------------+ | All | 46.854 |...