trailab / pdv Goto Github PK

when I train PDV on waymo dataset, I meet this error.
The yaml file is "cfgs/waymo_models/pdv.yaml". I have cost much time on this.

error details:
Traceback (most recent call last):
File "train.py", line 232, in
main()
File "train.py", line 176, in main
train_model(
File "/home/rcvlab/lly/PDV/tools/train_utils/train_utils.py", line 86, in train_model
accumulated_iter = train_one_epoch(
File "/home/rcvlab/lly/PDV/tools/train_utils/train_utils.py", line 38, in train_one_epoch
loss, tb_dict, disp_dict = model_func(model, batch)
File "/home/rcvlab/lly/PDV/pcdet/models/init.py", line 42, in model_func
ret_dict, tb_dict, disp_dict = model(batch_dict)
File "/home/rcvlab/anaconda3/envs/lly_torch17/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1110, in _call_impl
return forward_call(*input, **kwargs)
File "/home/rcvlab/lly/PDV/pcdet/models/detectors/pdv.py", line 11, in forward
batch_dict = cur_module(batch_dict)
File "/home/rcvlab/anaconda3/envs/lly_torch17/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1110, in _call_impl
return forward_call(*input, **kwargs)
File "/home/rcvlab/lly/PDV/pcdet/models/roi_heads/pdv_head.py", line 239, in forward
targets_dict = self.assign_targets(batch_dict)
File "/home/rcvlab/lly/PDV/pcdet/models/roi_heads/roi_head_template.py", line 110, in assign_targets
targets_dict = self.proposal_target_layer.forward(batch_dict)
File "/home/rcvlab/lly/PDV/pcdet/models/roi_heads/target_assigner/proposal_target_layer.py", line 32, in forward
batch_rois, batch_gt_of_rois, batch_roi_ious, batch_roi_scores, batch_roi_labels = self.sample_rois_for_rcnn(
File "/home/rcvlab/lly/PDV/pcdet/models/roi_heads/target_assigner/proposal_target_layer.py", line 113, in sample_rois_for_rcnn
batch_gt_of_rois[index] = cur_gt[gt_assignment[sampled_inds]]
RuntimeError: The expanded size of the tensor (8) must match the existing size (10) at non-singleton dimension 1. Target sizes: [128, 8]. Tensor sizes: [128, 10]

For my custom dataset, the pointcloud range and voxel size are defined by what? How can i calculate these two arguments?

Hi, nice work, and thank you for sharing the code and a trained checkpoint.

Despite your kind walkthrough, I'm having difficulties running your code with the given model-147M weights.
I had some version issues regarding spconv, pytorch, and pcdet.
Then I encountered some parameter load failure problems. (I was able to train the model from scratch.)

Can you inform me full evaluation results I could achieve when I run your model?
Sharing your evaluation log (as below) would be more than sufficient.

INFO  Generate label finished(sec_per_example: 0.0109 second).
INFO  recall_roi_0.3: xxxxxxx
INFO  recall_rcnn_0.3: xxxxxxx
INFO  recall_roi_0.5: xxxxxxx
INFO  recall_rcnn_0.5: xxxxxxx
INFO  recall_roi_0.7: xxxxxxx
INFO  recall_rcnn_0.7: xxxxxxx
INFO  Average predicted number of objects(3769 samples): xxxxxxx
INFO  Car [email protected], 0.70, 0.70:
bbox AP:xxxxxxx, xxxxxxx, xxxxxxx
bev  AP:xxxxxxx, xxxxxxx, xxxxxxx
3d   AP:xxxxxxx, xxxxxxx, xxxxxxx
aos  AP:xxxxxxx, xxxxxxx, xxxxxxx
Car [email protected], 0.70, 0.70:
bbox AP:xxxxxxx, xxxxxxx, xxxxxxx
bev  AP:xxxxxxx, xxxxxxx, xxxxxxx
3d   AP:xxxxxxx, 85.05, xxxxxxx
aos  AP:xxxxxxx, xxxxxxx, xxxxxxx
.
.
.

Thank you!

Hi, I tried to tun your source code，but the following error occurred：“AttributeError: 'Tensor' object has no attribute 'isnan”
The wrong sentence is “points_out_of_range = ((xyz_local_grid < 0) | (xyz_local_grid >= grid_size) | (xyz_local_grid.isnan())) .any(-1).flatten()”in the folder “pcdet/utils/density_utils.py".

What is the cause and how to improve it?

Hello, authors.
The ground truth of the kitti test set is not available, how do you calculate the accuracy of the network on the test set?

Hi,
Nice work!
May I ask where the results of PVRCNN and PVRCNN++ come from? The numbers look a little bit strange to me since I can not find them in the original papers of PVRCNN(++).

I notice that for each dataset, info files (.pkl) are generated. Do you have any rules about generating these info files? Because I want to train this model on my own dataset. Thanks a lot.

Dear author: could you tell me how to reproduce your cyclist and pedstrian results in your paper? I wonder you only train 1 class seperately, or you train 3 class together in a single training process?

Hi author!
I have a problem of the PDV's results on the KITTI val. I see the results of 3d detection performance for three classes(car, pedestrian, cyclist) of moderate difficulty on KITTI val are 85.05, 57.41, 75.95 in Github. But I find the results on paper are 85.29, 60.80, 74.23.

They are different! Could you tell me the differences between these two results?
Thanks!

when I do python -m pcdet.datasets.kitti.kitti_dataset create_kitti_infos tools/cfgs/dataset_configs/kitti_dataset.yaml
the error like this
/mnt/Anaconda3/envs/PDV/lib/python3.8/runpy.py:127: RuntimeWarning: 'pcdet.datasets.kitti.kitti_dataset' found in sys.modules after import of package 'pcdet.datasets.kitti', but prior to execution of 'pcdet.datasets.kitti.kitti_dataset'; this may result in unpredictable behaviour
warn(RuntimeWarning(msg))
Traceback (most recent call last):
File "/mnt/wg/PDV/pcdet/datasets/processor/data_processor.py", line 50, in transform_points_to_voxels
from spconv.utils import VoxelGeneratorV2 as VoxelGenerator
ImportError: cannot import name 'VoxelGeneratorV2' from 'spconv.utils' (/mnt/Anaconda3/envs/PDV/lib/python3.8/site-packages/spconv/utils/init.py)

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
File "/mnt/Anaconda3/envs/PDV/lib/python3.8/runpy.py", line 194, in _run_module_as_main
return _run_code(code, main_globals, None,
File "/mnt/Anaconda3/envs/PDV/lib/python3.8/runpy.py", line 87, in _run_code
exec(code, run_globals)
File "/mnt/wg/PDV/pcdet/datasets/kitti/kitti_dataset.py", line 501, in
create_kitti_infos(
File "/mnt/wg/PDV/pcdet/datasets/kitti/kitti_dataset.py", line 454, in create_kitti_infos
dataset = KittiDataset(dataset_cfg=dataset_cfg, class_names=class_names, root_path=data_path, training=False)
File "/mnt/wg/PDV/pcdet/datasets/kitti/kitti_dataset.py", line 23, in init
super().init(
File "/mnt/wg/PDV/pcdet/datasets/dataset.py", line 33, in init
self.data_processor = DataProcessor(
File "/mnt/wg/PDV/pcdet/datasets/processor/data_processor.py", line 17, in init
cur_processor = getattr(self, cur_cfg.NAME)(config=cur_cfg)
File "/mnt/wg/PDV/pcdet/datasets/processor/data_processor.py", line 52, in transform_points_to_voxels
from spconv.utils import VoxelGenerator
ImportError: cannot import name 'VoxelGenerator' from 'spconv.utils' (/mnt/Anaconda3/envs/PDV/lib/python3.8/site-packages/spconv/utils/init.py)

Hello，authors. This is the first time I've trained a pdv on a waymo dataset, but the network loss changes are strange, and the loss of rcnn_rpn_cls is increasing. Is this normal? If not, what might be the cause of this? Could you please give me your training log of tensorboard for reference?

I find the following code in this repo:

return self.kde_func.log_prob(input).sum(-1).exp()

However, I think it should be performed exp before the sum operation.
Is there any problem about my understanding?

env as follow
CUDA 11.3
Pytorch 1.7.1
the error like
OSError: /mnt/Anaconda3/envs/PDVT/lib/python3.8/site-packages/spconv/libspconv.so: undefined symbol: _ZNK3c104Type14isSubtypeOfExtESt10shared_ptrIS0_EPSo
Does this mean that my spconv is not compatible with pytorch or cuda versions and that I have to install cuda10.2 ？
thanks

Hello, I use all KITTI data for training. How can I find the model with the best performance from the trained models.

I found it in mean_vfe.py, I wonder what it does and how can I use him

May I ask how to calculate the fps of the model

Hello, i've met a problem in running demo.py
2022-04-04 14:55:26.887 ( 15.100s) [ F0A55E80]vtkOpenGLRenderWindow.c:493 ERR| vtkEGLRenderWindow (0x561980223bb0): GLEW could not be initialized: Missing GL version
run-demo.sh: line 19: 7579 Segmentation fault (core dumped)
I've tried several approaches but all of them failed. Do you have any idea about this bug?

Thanks for your work, I would like to know what is the role of MAX_NUM_BOXES, is it because the RoIs may have overlap, causing a point to fall into different RoI?

Hi, thanks for your wonderful job.Can you push a training log to verify our training results?
Thank you again!