Why is the GPU occupancy rate during training very low, only 10%

WHU-USI3DV / SGHR

[CVPR 2023] Robust Multiview Point Cloud Registration with Reliable Pose Graph Initialization and History Reweighting

193 stars 16 forks source link

Why is the GPU occupancy rate during training very low, only 10% #5

Closed horizonfly99 closed 7 months ago

horizonfly99 commented 1 year ago

Thank you for your work，Why is the GPU occupancy rate during training very low, only 10%？

Wansit99 commented 1 year ago

I have the same question

HpWang-whu commented 1 year ago

Hi @horizonfly99 and @Wansit99, Sorry for the late reply! The local features (YOHO) in SGHR are pre-extracted and saved. Thus we only train a quite tiny network: only one NetVLAD layer and thus requires low GPU memory.

Yours,

horizonfly99 commented 1 year ago

Why is the validation value after running Train.py only about 0.75?