A weird problem about GPU memory usage

WHU-USI3DV / CoFiI2P

[IEEE RA-L 2024] CoFiI2P: Coarse-to-Fine Correspondences-Based Image-to-Point Cloud Registration

https://whu-usi3dv.github.io/CoFiI2P/

56 stars 3 forks source link

A weird problem about GPU memory usage #13

Closed YhQIAO closed 2 days ago

YhQIAO commented 1 week ago

if i change batch_size in eval_all.py as 2, There will be the following error Increasing the batch_sizes will increase the memory usage, but 1200GB is very abnormal. I found that the batch_size in the training stage is also 1 (I haven't run the train code yet). Will this problem occur if the batch size is set to be greater than 1 on your computer?

martin-liao commented 1 week ago

Due to the GPU memory limitation, we trained the model with batch_size=1 in our experiments. I will work on solving the issues caused by increasing the batch size.

martin-liao commented 1 week ago

As we follow the KPConv implementation of Geotransformer here, CoFiI2P only supports batch_size=1 now. We will try our best to support batch_size >1shortly.