CUDA out of memory though there supposed to be enough VRAM about eccv2022-rife HOT 7 CLOSED

BTW, one piece of video in 20FPS triggerred this problem but another one in 25FPS didn't. They are both 720p and encoded by x265, file size about 20MB.

from eccv2022-rife.

Hi, I think we should fall back to the non-parallel version, because single-card parallel is slower on most devices. And the parallel version significantly increases the memory overhead. If you can upload the video to google drive for me to download, I can check the error.

from eccv2022-rife.

In other words, we think that continuing to maintain a single-card parallel version is likely to introduce new problems, so please re-integrate inference_video.py.

from eccv2022-rife.

Got it. You can close this issue if won't fix. Thanks.

from eccv2022-rife.

Got it. You can close this issue if won't fix. Thanks.

From your log, CUDA OOM is seems not a bug. I think inference_video will not cause it.

from eccv2022-rife.

Hi, I think we should fall back to the non-parallel version, because single-card parallel is slower on most devices. And the parallel version significantly increases the memory overhead. If you can upload the video to google drive for me to download, I can check the error.

Hmmm, on my RTX 2070S, the parallel code is noticably faster.
The non-parallel code only reaches about 75-80% GPU utilization.

from eccv2022-rife.

Hmmm, on my RTX 2070S, the parallel code is noticably faster.
The non-parallel code only reaches about 75-80% GPU utilization.

OK. I will add a flag to support that.

from eccv2022-rife.