search

NotACracker / COTR

[CVPR24] COTR: Compact Occupancy TRansformer for Vision-based 3D Occupancy Prediction
Apache License 2.0
47 stars 4 forks source link

GPU memory for training #6

Closed lvchuandong closed 2 months ago

lvchuandong commented 2 months ago

Thank you for your great work. I want to know why a GFLOPs of 740.89 is not very high, but when training, when the batch size is 1, the GPU used is already 18.2G. I am using 4090, and if the batch size is 2, there will be an out of memory. flops.log

NotACracker commented 2 months ago

Thank you for your great work. I want to know why a GFLOPs of 740.89 is not very high, but when training, when the batch size is 1, the GPU used is already 18.2G. I am using 4090, and if the batch size is 2, there will be an out of memory. flops.log

The increase in memory comes mainly from IVT and Group transformer decoder, you can reduce the memory by modifying the _dim_ value in config, but there may be a slightly drop in accuracy

lvchuandong commented 2 months ago

Thank you very much