Layort
Layort
> 你好,不是通过sample,而是直接用t step预测出来的噪声epsilon_t预估z'_0,然后VAE decode得到图像x'_0。由于此时x'_0是预估的,根据t的变化会存在模糊,所以用一个自适应权重函数来调节loss,具体参见[OpenReview](https://openreview.net/forum?id=ezBH9WE9s2¬eId=FNIu1OQmHo)关于Q4的回复。 那请问您用了多久来训练呢?
> @Layort 在8XA100上总计312h,具体参见[OpenReview](https://openreview.net/forum?id=ezBH9WE9s2¬eId=FNIu1OQmHo)关于Q5的回复。 好的,感谢
I have met the same problem. Did anybody solve it?
> Follow the new [#621](https://github.com/NVIDIA/MinkowskiEngine/issues/621) and I successfully installed it. You need to install the newest PyTorch through I tried but failed. Is it suitable for CUDA 12.8?
me too. I crushed for several times in one hour.