CUDA out of memory

[ListIndexOutOfRange]

CUDA out of memory

[ma-xu]

@ListIndexOutOfRange Thanks for your interest.

This is not expected behavior, and PointMLP does need large memory.
With the default setting, the peak memory will be ~19GB for pointMLP and ~4.5GB for pointMLP-elite. See the screenshots.

Note that the tensor shape in each stage would be [batch_size, number_of_selected_points, number_of_neighbors, dimension] (e.g., [32, 64, 24, 512] in last stage of PointMLP, see Fig.6). Considering the FC operation, this is not small. The residual connection also increases the memory cost.

[ma-xu]

I will close this issue since no further discussions. Feel free to reopen it if necessary.

[mapi829c]

Hi!, I see your CUDA Version is 11.4 when used the command: nvidia-smi

May I ask, which version appears when using the following.

1. Command: nvcc --version
2. Inside python command: torch.version.cuda

my question is because I get the following every time I try to install 'pip3 install pointnet2_ops_lib/':
'The detected CUDA version (11.4) mismatches the version that was used to compile PyTorch (11.3). Please make sure to use the same CUDA versions.'

Thank you!