gwang-kim/DiffusionCLIP

我有个问题，希望大佬回答

Closed this issue · 3 comments

zhou431496 commented

如果在训练的时候我将n_train_step改为2，并且多训练是否可以得到可以接受的结果 。我12g的显存无法跑n_train_step=6的代码

gwang-kim commented

Thanks for your interest in our work, @zhou431496 It depends on the types of attributes. Empirically, I found that n_train_step=3 works well for some attributes. Also, I think I can reduce the GPU memory to under 11GB by optimizing the code. If I finish it, I will inform you.

zhou431496 commented

感谢您的回复，如果我将该工作引入到图像修复，是否可行呢。期待您的回复

gwang-kim commented

Yes, I think the ideas of DiffusionCLIP can be also applied to the reconstruction works.