search

Zongwei97 / SCTNet

[ICCV 23] Official implementation of Alignment-free HDR Deghosting with Sematnics Consistent Transformer (Code & Dataset)
24 stars 3 forks source link

CUDA out of memory #2

Open asher-bit opened 4 months ago

asher-bit commented 4 months ago

你好， 首先，感谢您精彩的工作 我在尝试复现模型的时候，出现了显存报错。debug时发现在G-SAB模块中，tensor的尺寸过大（150万），导致显存溢出。但是我并未修改您的模型。请问是什么问题呢？ 期待您的回复！

asher-bit commented 4 months ago

image image

josair21 commented 2 months ago

150M is not too big in computational terms, what GPU are u using?, training with a batch size of 22 (defult is 22) needs a los of memory, maybe u will need 2 x 40GB of GPU memory for training it.