gpu memory not realesed

[royzhang12]

Hi @wenbowen123,
thank you for your nice work. I've been trying to run Bundletrack on a collections of video clips.
To automatically swith between each clips, I simply declare Bundler by
*Bundler bundler = nullptr; before the main loop of video
and construct it again after swith to to a new video clips by
delete bundler;
bundler = new Bundler(yml,&data_loader); . The pose estimation result seem well. However, I find that the GPU memory is not realesed after swith to a new video clips and the usage keeps increaseting. Untill 10G GPU memory was used, the programe will be automatically killed by no reason.

Do you have any idea of how to resolve this issue?

[wenbowen123]

Hi, there might be a memory leak. I'll try to find a time to investigate further after a deadline recently.
If you are running for some offline data, perhaps you can run each video sequence by bash script? Each time modify the data_dir in the config file to run.

[royzhang12]

Hi, there might be a memory leak. I'll try to find a time to investigate further after a deadline recently. If you are running for some offline data, perhaps you can run each video sequence by bash script? Each time modify the data_dir in the config file to run.

Hi, @wenbowen123. Thank you for the swift reply. I am doing test on both online and offline data. The suggestion is a great solution for offline data, thanks. However, for oneline test, it does not work.
I tried to restart the process autoamtically online, which consumes much time and leads to really slow inference speed. Hope you can help to find the reason cause the memory leak after your deadline. Many thanks.

[ChrisSun99]

I also have this problem right now. I tried the workaround to split my dataset into smaller segments, but it would be great if there's a solid solution.

[wenbowen123]

FYI, we recently released a follow up version BundleSDF, which should not have such issue.