search

dvlab-research / VoxelNeXt

VoxelNeXt: Fully Sparse VoxelNet for 3D Object Detection and Tracking (CVPR 2023)
https://arxiv.org/abs/2303.11301
Apache License 2.0
735 stars 64 forks source link

About the FLOPs. #21

Closed AlmoonYsl closed 1 year ago

AlmoonYsl commented 1 year ago

Hi, thanks for your excellent work. As shown in Tab. 1, the FLOPs is calculated for sparse cnn and head separately. I noticed that the FLOPs of sparse cnn and head (VoxelNeXt) is 33.6G and 5.1G. So, I want to know whether the sparse conv in 3D backbone and head (VoxelNeXt) is considered in calculation. AFAK, the mmcv and fvcore do not support the FLOPs calculation for sparse conv. Besides, the head of CenterPoint is consists of 2D conv, which is supported by mmcv and fvcore.

yukang2017 commented 1 year ago

Hi,

spconv is included in FLOPs calculation. We use the FLOPs calculation tool in https://github.com/CVMI-Lab/SparseKD

AlmoonYsl commented 1 year ago

Hi,

spconv is included in FLOPs calculation. We use the FLOPs calculation tool in https://github.com/CVMI-Lab/SparseKD

Got it. Thanks for your reply!

yBeOne commented 1 year ago

Where is the FLOPs calculation tool in SparseKD?

yukang2017 commented 1 year ago

Please refer to this.

https://github.com/CVMI-Lab/SparseKD/blob/8572093edc77d4217f7596d9e1de68c6dc77e420/pcdet/utils/common_utils.py#L344