RuntimeError : Cuda out of memory

[seoultechJS]

Ours : GTX TitanX Ubuntu 20.04 RuntimeError: CUDA out of memory. Tried to allocate 62.00 MiB (GPU 0; 11.92 GiB total capacity; 8.81 GiB already allocated; 50.31 MiB free; 9.09 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

We got this Error while running Replica or TumRGBD. In the Nice-SLAM model, this error only appeared in the dataset we constructed ourselves, and not in TumRGBD or Replica. How can we fix it?

[shaoxiang777]

Ours : GTX TitanX Ubuntu 20.04 RuntimeError: CUDA out of memory. Tried to allocate 62.00 MiB (GPU 0; 11.92 GiB total capacity; 8.81 GiB already allocated; 50.31 MiB free; 9.09 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

We got this Error while running Replica or TumRGBD. In the Nice-SLAM model, this error only appeared in the dataset we constructed ourselves, and not in TumRGBD or Replica. How can we fix it?

I think the possible reason is they use different sampling strategy. In NICE-SLAM they select total 1000 rays for all keyframes in the window(mapping_window_size: 5). But in E-slam, they author select 4000 rays for every keyrame in a window(mapping_window_size: 20), so 4000*20 rays in total. So I suggest you can reduce the number or selected rays or change a stronger GPU.

[MohammadJohari]

Hello,

We actually select 4000 rays for all keyframes. That said, the best approach to deal with the memory limitation is indeed reducing the number of rays in mapping and tracking.

[seoultechJS]

Thanks for fast apply! I'll try it what you said