search

Megvii-BaseDetection / BEVDepth

Official code for BEVDepth.
MIT License
710 stars 98 forks source link

BEVDepth perform very bad when training with batch size 1! #164

Closed cskkxjk closed 1 year ago

cskkxjk commented 1 year ago

Hi! I'm currently training the bevdepth models with different batch size. I used 1,2,4,8 GPUs with batch size=1, 2, 4. All the models trained with batch size=1 seems to perform bad on nuScenes val set (0\~3mAP). All the models trained with batch size != 1 perform reasonable (20\~30mAP).

can you give me some advices about how to change the config to make it work when batch size=1?

hutchinsonian commented 1 year ago

When batch size=1, what is your GPU memory usage？

cskkxjk commented 1 year ago

When batch size=1, what is your GPU memory usage？

the GPU memory is about 6G when batch size is 1, the experiment config is bev_depth_lss_r50_256x704_128x128_24e.py modified from bev_depth_lss_r50_256x704_128x128_24e_ema.py