[Enhancement] faster but nondeterministic version of hard voxelization

[zhanggefan]

Small enhancement -- reimplement hard voxelization and make it faster at the cost of deterministic property

It is now an optional feature and is disabled by default. Set deterministic=False inside Voxelization config to enable it. It only works for hard voxelization and has no effect when using dynamic voxelization.

Performance comparison:

conf	training iter time tested on	iter time before	iter time after
pointpillars-kitti-3class	2080Ti bs=6	0.340	0.303
pointpillars-kitti-car	2080Ti bs=6	0.281	0.246
pointpillars-waymo-3class	3090 bs=2	0.412	0.296
pointpillars-waymo-car	3090 bs=2	0.365	0.237
centerpoint-pillar0.2-nuscenes	3090 bs=4	0.858	0.441
centerpoint-pillar0.2-nuscenes	3090 bs=8	1.681	0.832
centerpoint-voxel0.075-nuscenes	3090 bs=4	1.638	1.212

Score comparison with MMDet3D Original (May Not Precise!!!!! I also made a lot of code revisions besides voxelization and I messed them up)

conf	training on	score of this version	score by MMDet3D
pointpillars-kitti-3class	3090*4 bs=12	mAP=61.3359	mAP=59.5
pointpillars-kitti-car	3090*4 bs=12	mAP=76.7144	mAP=77.1
pointpillars-waymo-3class	3090*4 bs=8 fp16	L1-mAP=70.04 L1-mAPH=63.94 L2-mAP=63.62 L2-mAPH=58.08	L1-mAP=68.843 L1-mAPH=63.318 L2-mAP=62.619 L2-mAPH=57.573
centerpoint-pillar0.2-nuscenes	3090*4 bs=8	mAP=49.43 NDS=60.14	mAP=49.07 NDS=59.66

[zhanggefan]

Seems something went wrong with your CICD pipeline when building MMCV exts

[codecov]

Codecov Report

Merging #904 (72c0541) into master (2efdb6b) will increase coverage by 0.00%.
The diff coverage is 50.00%.

@@           Coverage Diff           @@
##           master     #904   +/-   ##
=======================================
  Coverage   49.33%   49.33%           
=======================================
  Files         210      210           
  Lines       16021    16023    +2     
  Branches     2556     2556           
=======================================
+ Hits         7904     7905    +1     
- Misses       7614     7615    +1     
  Partials      503      503           

Flag	Coverage Δ
unittests	49.33% <50.00%> (+<0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
mmdet3d/ops/voxel/voxelize.py	80.76% <50.00%> (-1.24%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 2efdb6b...72c0541. Read the comment docs.

[ZwwWayne]

Hi @zhanggefan ,
Thanks for the efforts. Faster implementation of voxelization is definitely beneficial to the model training and testing with MMDetection3D. For completeness, shall we also leave a comparison of accuracy in this PR? For example, it is better to leave a table to compare the AP of a similar model as a reference, and we may also leave more detailed comments to indicate why it is faster but non-deterministic in the code and documentation.

[zhanggefan]

Hi, @ZwwWayne
This PR is related to issue #894. Some details of the implementation can be found there.
The score comparison table is added in the PR description now. It may not be precise!!! Because I made many code optimizations(to improve the training speed, but I cannot guarantee that they have no impact on scores) and the voxelization is just one of them.

[ZwwWayne]

Hi, @ZwwWayne
This PR is related to issue #894. Some details of the implementation can be found there.
The score comparison table is added in the PR description now. It may not be precise!!! Because I made many code optimizations(to improve the training speed, but I cannot guarantee that they have no impact on scores) and the voxelization is just one of them.

I see. Can we refer to this PR and issue #894 in the docstring of the module? So that future developers could know the trade-off and discussion. Furthermore, we also welcome you to create PRs for optimization to improve the training speed.

[zhanggefan]

Docstring is revised. Besides, a bug similar to #768 in the previous commit is fixed.