Using large dataset such as SensatUrban

[aymanmuk]

Hi, thank you for sharing your work! I have been trying to implement your code on large datasets such as SensatUrban and Dales. However, I encountered a couple of challenges:

In scene_seg.py, while computing nearest neighbors per tile:

_, idxs = tiled_knn(q_pts, s_pts, k=1, tile_size=20.5, margin=2 * dl) # 3.5
Dividing the area by 3.5 meters takes a significant amount of time. Could you recommend any adjustments or optimizations for this parameter to improve performance on large datasets?

Also, in Kpconv_blocks.py, I ran into the following error:
output_feats = torch.sum(neighbor_feats * neighbors_weights, dim=1) # -> (M, G, 0//G)
RuntimeError: The size of tensor a (59584) must match the size of tensor b (931) at non-singleton dimension 0

Could you provide guidance on resolving this issue?

[HuguesTHOMAS]

Hi @aymanmuk,

Thx for your interest!

For the tiled knn function, you can increase the tile size as long as there is no OOM error, it should be faster. For larger scenes, with smaller density, larger tiels makes more sense. The parameters in the code were chosen for indoor scenes.

For the block, you can use is the error happening as well if you use the summation mode?

Best,
Hugues

[aymanmuk]

Hi @HuguesTHOMAS,
Thank you for the quick responce.
For the second point, I think it there is a mistake in the code. It should be:
neighbor_feats = neighbor_feats.view(-1, H, self.groups, 1) # (M, H, C, 1)
instead of:
neighbor_feats = neighbors_weights.view(-1, H, self.groups, 1) # (M, H, C, 1)

For the first point, I will try tile_size=50.
According to previous Dales implemetation, I changed the following:
cfg.model.kp_radius = 20
cfg.data.init_sub_size = 0.025
cfg.model.in_sub_size = 0.025
cfg.model.input_channels = 1

Please advise if these values need further adjustments. Thanks

[HuguesTHOMAS]

Indeed you are right for the second point.

For you parameters, this is not right. You should keep cfg.model.kp_radius to a relatively small value, between 2 and 4, and use cfg.train.in_radius = 20

[aymanmuk]

Thanks @HuguesTHOMAS.
Using cfg.train.in_radius = 20 has significantly increased memory requirements. To address this, I am using:
cfg.train.batch_limit = 1. Hopefully that will not cause any training instability.

[narges-tk]

Thanks @HuguesTHOMAS. Using cfg.train.in_radius = 20 has significantly increased memory requirements. To address this, I am using: cfg.train.batch_limit = 1. Hopefully that will not cause any training instability.

Hi! Could you please let me know whether you are using the Standalone or Pointcept option? Did you encounter the "no attribute 'batch_grid_partitioning'" or a similar error, and if so, how did you resolve it? Thanks in advance for your help!

[aymanmuk]

@narges-tk I am using the Standalone version and I didnt get any error duing training the S3DIS datast.

[narges-tk]

@narges-tk I am using the Standalone version and I didnt get any error duing training the S3DIS datast.

Thanks a lot for your repl! What is this line in your cpp_funcs.py?
import cpp_wrappers.cpp_subsampling.cpp_subsampling as cpp_subsampling
Did you make any change/compilation in cpp_wrappers>cpp_subsampling folder?

[aymanmuk]

@narges-tk , Install both numpy version 1.22.1 and setuptools 59.5.0 and try to build the cpp folders again.

[aymanmuk]

@HuguesTHOMAS

I finally managed to get good results. I used init_sub_size 0.25, not 0.025, and now I can fit everthing on the gpu.

To summerize, I changed the following:

cfg.data.init_sub_size = 0.25
cfg.model.in_sub_size = 0.25
cfg.train.in_radius = 20
cfg.model.input_channels = 1

tile_size=25

[narges-tk]

@narges-tk , Install both numpy version 1.22.1 and setuptools 59.5.0 and try to build the cpp folders again.

Many thanks! It's now working nicely for me!

[narges-tk]

@HuguesTHOMAS

I finally managed to get good results. I used init_sub_size 0.25, not 0.025, and now I can fit everthing on the gpu.

To summerize, I changed the following:

cfg.data.init_sub_size = 0.25 cfg.model.in_sub_size = 0.25 cfg.train.in_radius = 20 cfg.model.input_channels = 1

tile_size=25

Hi! I wonder if you did not face any inconsistency during the prediction using your trained model. In my case, I see that while the accuracies shown while training are fine, they are by far bad in prediction. I should note that I have considered test data as validation.
Thanks in advance for your reply!