I set the syn_bn as True when doing multi GPU training, but the 3D AP is lower.

TRAILab / CaDDN

Categorical Depth Distribution Network for Monocular 3D Object Detection (CVPR 2021 Oral)

Apache License 2.0

359 stars 62 forks source link

I set the syn_bn as True when doing multi GPU training, but the 3D AP is lower. #72

Closed rockywind closed 2 years ago

rockywind commented 2 years ago

My environment is: 8*A6000 BATCH_SIZE_PER_GPU:4 batch_size:32 pytorch1.7.1 cuda11.1

codyreading commented 2 years ago

I didn't do experiments with this large of a batch size, so unsure how performance is affected at these large batch sizes. I would recommend trying multiple runs at different batch sizes and different settings of sync_bn to see the performance if you are interested at these large batch sizes.