Dear author,
Thank you for your incomparable work. I want to ask you a question. How is is_revisit_D-3_T-30.json constructed?
Thank you very much for your reply.

Hi, thanks for your great job!

I am gonna to test your work in our project. We have done some investigation before. In some lidar-based place recognition method, lidar with 360 degree data with at least 36 lines is recommended. So how about the adaptability to different lidar sensor of your method?

Thank your for your attention and keep waiting for kind response!

Can you provide an example how to train on a custom dataset?

Dear Authors,

At the outset, I would like to thank you for your great work and for making this code public.
But when I was trying to execute training for KITI dataset with the default training parameters, I am getting the following errors:
[mpiexec@node0006] match_arg (utils/args/args.c:163): unrecognized argument allow-run-as-root
[mpiexec@node0006] HYDU_parse_array (utils/args/args.c:178): argument matching returned error
[mpiexec@node0006] parse_args (ui/mpich/utils.c:1642): error parsing input array
[mpiexec@node0006] HYD_uii_mpx_get_parameters (ui/mpich/utils.c:1694): unable to parse user arguments
[mpiexec@node0006] main (ui/mpich/mpiexec.c:148): error parsing parameter

Any feedback/ insights on the same are appreciated.

Thanks for your time and your work.

I want to compare results with PointNetVlad. How to run it on pretrained weights?

Hello, I want to first of all thank you for your great work!

I found a problem when I trained the network with batch_size=2:

Exception has occurred: AttributeError       (note: full exception trace is shown but execution is paused at: <module>)
'list' object has no attribute 'to'
  File "/home/stud/maiy/LoGG3D-Net/training/train.py", line 100, in main
    batch_st = batch[0].to('cuda:%d' % dist.local_rank())
  File "/home/stud/maiy/LoGG3D-Net/training/train.py", line 174, in <module> (Current frame)
    main()

After I looked into the code, I think the problem lies in the collation function of dataloader which is incompatible with indexing in train.py:

Could you please check it? Any feedback is appreciated and please let me know if I misunderstood anything. Thanks again for your work!

how to Visualize the t-SNE embeddings of the local features？ Have any code provide？Thanks！

I have got some CAD model point clouds. But when I applied the model to them, in some cases it succeeded but in some other cases it would report a channel value error.

Here is the code segment:

save_path = os.path.join(os.path.dirname(file), 'checkpoints')
save_path = str(save_path) + cfg.checkpoint_name
checkpoint = torch.load(save_path)
model = get_pipeline(cfg.eval_pipeline)
model.load_state_dict(checkpoint['model_state_dict'])

base_file = data_dir + '/' + sorted(os.listdir(data_dir))[1]
base_pcd = np.load(base_file)
base_pcd = np.hstack((base_pcd, np.zeros(base_pcd.shape[0]).reshape(-1, 1))).astype(np.float32) # My pcds do not have intensity.
current_input = make_sparse_tensor(base_pcd, cfg.voxel_size).cuda()
output_desc, output_feats = model(current_input)

And here is the error report:

Traceback (most recent call last):
File "/media/uds/extra_space/LoGG3D-Net/Calc_Logg3D_Descs.py", line 50, in
output_desc, output_feats = model(current_input)
File "/media/uds/extra_space/LoGG3D-Net/venv/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1501, in _call_impl
return forward_call(*args, **kwargs)
File "/media/uds/extra_space/LoGG3D-Net/models/pipelines/LOGG3D.py", line 25, in forward
x = self.spvcnn(x)
File "/media/uds/extra_space/LoGG3D-Net/venv/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1501, in _call_impl
return forward_call(*args, **kwargs)
File "/media/uds/extra_space/LoGG3D-Net/models/backbones/spvnas/core/models/semantic_kitti/spvcnn.py", line 197, in forward
x4 = self.stage4(x3)
File "/media/uds/extra_space/LoGG3D-Net/venv/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1501, in _call_impl
return forward_call(*args, **kwargs)
File "/media/uds/extra_space/LoGG3D-Net/venv/lib/python3.10/site-packages/torch/nn/modules/container.py", line 217, in forward
input = module(input)
File "/media/uds/extra_space/LoGG3D-Net/venv/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1501, in _call_impl
return forward_call(*args, **kwargs)
File "/media/uds/extra_space/LoGG3D-Net/models/backbones/spvnas/core/models/semantic_kitti/spvcnn.py", line 27, in forward
out = self.net(x)
File "/media/uds/extra_space/LoGG3D-Net/venv/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1501, in _call_impl
return forward_call(*args, **kwargs)
File "/media/uds/extra_space/LoGG3D-Net/venv/lib/python3.10/site-packages/torch/nn/modules/container.py", line 217, in forward
input = module(input)
File "/media/uds/extra_space/LoGG3D-Net/venv/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1501, in _call_impl
return forward_call(*args, **kwargs)
File "/media/uds/extra_space/LoGG3D-Net/venv/lib/python3.10/site-packages/torchsparse/nn/modules/norm.py", line 13, in forward
return fapply(input, super().forward)
File "/media/uds/extra_space/LoGG3D-Net/venv/lib/python3.10/site-packages/torchsparse/nn/utils/apply.py", line 12, in fapply
feats = fn(input.feats, *args, **kwargs)
File "/media/uds/extra_space/LoGG3D-Net/venv/lib/python3.10/site-packages/torch/nn/modules/batchnorm.py", line 171, in forward
return F.batch_norm(
File "/media/uds/extra_space/LoGG3D-Net/venv/lib/python3.10/site-packages/torch/nn/functional.py", line 2448, in batch_norm
_verify_batch_size(input.size())
File "/media/uds/extra_space/LoGG3D-Net/venv/lib/python3.10/site-packages/torch/nn/functional.py", line 2416, in _verify_batch_size
raise ValueError("Expected more than 1 value per channel when training, got input size {}".format(size))
ValueError: Expected more than 1 value per channel when training, got input size torch.Size([1, 81])

Hi,

I'm currently trying to train PointNetVLAD on different dataset (KittiSparseTupleDataset and KittiTupleDataset).
I have no problems for LoGG3D-net training but I have those two errors for PointNetVLAD:

Command number 1 :

python training/train.py --train_pipeline 'PointNetVLAD' --train_num_workers 8 --point_loss_weight 0 --dataset "KittiSparseTupleDataset"
Error :
Traceback (most recent call last):
File "/gpfs7kw/linkhome/rech/genvsy01/ufm44cu/code/LoGG3D-Net/training/train.py", line 161, in
main()
File "/gpfs7kw/linkhome/rech/genvsy01/ufm44cu/code/LoGG3D-Net/training/train.py", line 98, in main
output = model(batch_st)
File "/gpfslocalsup/pub/anaconda-py3/2021.05/envs/pytorch-gpu-1.10.1+py3.9.7/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
return forward_call(*input, **kwargs)
File "/gpfslocalsup/pub/anaconda-py3/2021.05/envs/pytorch-gpu-1.10.1+py3.9.7/lib/python3.9/site-packages/torch/nn/parallel/distributed.py", line 886, in forward
output = self.module(*inputs[0], **kwargs[0])
File "/gpfslocalsup/pub/anaconda-py3/2021.05/envs/pytorch-gpu-1.10.1+py3.9.7/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
return forward_call(*input, **kwargs)
File "/gpfs7kw/linkhome/rech/genvsy01/ufm44cu/code/LoGG3D-Net/training/../models/pipelines/PointNetVLAD.py", line 19, in forward
x = self.point_net(x)
File "/gpfslocalsup/pub/anaconda-py3/2021.05/envs/pytorch-gpu-1.10.1+py3.9.7/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
return forward_call(*input, **kwargs)
File "/gpfs7kw/linkhome/rech/genvsy01/ufm44cu/code/LoGG3D-Net/training/../models/backbones/PointNet.py", line 86, in forward
batchsize = x.size()[0]
AttributeError: 'SparseTensor' object has no attribute 'size'

Command number 2 :

python training/train.py --train_pipeline 'PointNetVLAD' --train_num_workers 8 --point_loss_weight 0 --dataset "KittiTupleDataset"
Error :
Traceback (most recent call last):
File "/gpfs7kw/linkhome/rech/genvsy01/ufm44cu/code/LoGG3D-Net/training/train.py", line 161, in
main()
File "/gpfs7kw/linkhome/rech/genvsy01/ufm44cu/code/LoGG3D-Net/training/train.py", line 93, in main
for i, batch in enumerate(train_loader, 0):
File "/gpfslocalsup/pub/anaconda-py3/2021.05/envs/pytorch-gpu-1.10.1+py3.9.7/lib/python3.9/site-packages/torch/utils/data/dataloader.py", line 521, in next
data = self._next_data()
File "/gpfslocalsup/pub/anaconda-py3/2021.05/envs/pytorch-gpu-1.10.1+py3.9.7/lib/python3.9/site-packages/torch/utils/data/dataloader.py", line 1203, in _next_data
return self._process_data(data)
File "/gpfslocalsup/pub/anaconda-py3/2021.05/envs/pytorch-gpu-1.10.1+py3.9.7/lib/python3.9/site-packages/torch/utils/data/dataloader.py", line 1229, in _process_data
data.reraise()
File "/gpfslocalsup/pub/anaconda-py3/2021.05/envs/pytorch-gpu-1.10.1+py3.9.7/lib/python3.9/site-packages/torch/_utils.py", line 434, in reraise
raise exception
AttributeError: Caught AttributeError in DataLoader worker process 0.
Original Traceback (most recent call last):
File "/gpfslocalsup/pub/anaconda-py3/2021.05/envs/pytorch-gpu-1.10.1+py3.9.7/lib/python3.9/site-packages/torch/utils/data/_utils/worker.py", line 287, in _worker_loop
data = fetcher.fetch(index)
File "/gpfslocalsup/pub/anaconda-py3/2021.05/envs/pytorch-gpu-1.10.1+py3.9.7/lib/python3.9/site-packages/torch/utils/data/_utils/fetch.py", line 52, in fetch
return self.collate_fn(data)
File "/gpfs7kw/linkhome/rech/genvsy01/ufm44cu/code/LoGG3D-Net/training/../utils/data_loaders/pointcloud_dataset.py", line 126, in call
return self.collation_fn(list_data)
File "/gpfs7kw/linkhome/rech/genvsy01/ufm44cu/code/LoGG3D-Net/training/../utils/data_loaders/pointcloud_dataset.py", line 143, in collate_tuple
outputs.append(sparse_collate(contrastive_tuple))
File "/gpfslocalsup/pub/anaconda-py3/2021.05/envs/pytorch-gpu-1.10.1+py3.9.7/lib/python3.9/site-packages/torchsparse/utils/collate.py", line 13, in sparse_collate
stride = inputs[0].stride
AttributeError: 'numpy.ndarray' object has no attribute 'stride'

Problems

=> Even when I change the dataset to be "non-sparse", the code call the function sparse_collate
def sparse_collate(inputs: List[SparseTensor]) -> SparseTensor:
in Torchsparse.

collate_tuple (in pointcloud_dataset.py) call sparse_collate, which need a SparseTensor object as an input,
Shouldn't I call the sparse_collate_fn function instead?
def sparse_collate_fn(inputs: List[Any]) -> Any:

And, kitti_dataset.py don't return a pytorch tensor but a numpy.ndarray, is it normal?

Originally posted by @Chahine-Nicolas in #2 (comment)