GPU memory leakage issue

[ikodoh]

Thank you for sharing a nice work!

I'm currently using monst3r by repetitively reconstructing 3D point clouds for a series of videos.
However, I observed that there is a GPU memory leakage, where the GPU memory gradually increases and finally the CUDA OOM error occurs after handling several dozens of images.
I've tried to delete the 'scene' variable memory by inserting del scene after global_alignment, but it doesn't address the issue.
Can you help me to address this GPU memory leakage issue?

Thank you in advance.

[Junyi42]

Hi @ikodoh,

Thanks for the feedback. Could you please elaborate more on the issue? For example, is it possible for you to process the longest video of your series? Also, if you are encountering the issue when using demo.py, you could probably try to use or refer to pose_eval.py, which may help avoid this problem.

Thanks.

[ikodoh]

Actually, I made a custom function as an attached image, and repetitively called the inference_most3r function for a series of videos.
Then, this job causes the CUDA OOM error after processing several dozens of images.
I think this code seems pretty similar to pose_eval.py as you mentioned.
How can I fix it?

Thank you for your help.