search

USTC3DV / FlashAvatar-code

[CVPR 2024] The official repo for FlashAvatar
MIT License
143 stars 18 forks source link

CUDA out of memory. Tried to allocate 21355.80 GiB. #18

Closed liaopuyun closed 4 months ago

liaopuyun commented 4 months ago

Great work! Can't wait to run it on my own data.

I run the script test.py but find it require overmuch CUDA memory when render the gaussian. I try to render with the first 100 points, or decrease the fov by 100 times, yet it makes not much difference. How can I figure out what is going on?

xiangjun-xj commented 4 months ago

Someone else has successfully run the code, so I believe the script test.py itself is right. Maybe you should check places you changed. The main memory cost is from images loaded in the test dataset, you can load paths instead. But I think this is not the point.

liaopuyun commented 4 months ago

Thank you for your kindly answer. I carefully checked the code and make sure there is no change. So I viewed the cuda code in diff-gaussian-rasterization. In the code rasterizer_impl.cu, I found no mater how many gaussian I input, the num_rendered in line 282 is always 393217, and the binning.sorting_size in line 192 is always about 20000G. What is that supposed to mean?

xiangjun-xj commented 4 months ago

Sorry to say that I have no idea. What about creating the environment again.