Hello everyone, I have done a set of experiments on KITTI using OpenPCDet framework. There is some improvement in the evaluation metrics on the validation set, but the metrics of cyclists and pedestrians on the test set have greatly decreased. I guess that might be overfitting, so I want to ask how...

I want to utilize nuscenes 1/7 dataset. Is there any arg in openpcdet repository that supports nuscenes split dataset?

I ran kitti_dataSet.py to debug, and found an error： from . import kitti_utils ImportError: attempted relative import with no known parent package。Why can't I find this bag? How should I modify or ignore this error? In fact, the code can be trained and reasoned normally

Supported in the latest verison with `--infer_time` keyword Originally posted by @jihanyang in #997 (comment) When I used this args, some digits come out. How can I interpret this digits??

Hi, when training on the Waymo Open dataset (pointpillar_1x as a simple test), OOM in RAM occurs. I checked the memory increase when launching the training script. The memory usage keeps increasing until the maximum value (The RAM used is 256G), then the launching fails. Is it normal? The Waymo data...

The evaluation results as follows: PV-RCNN ++ (trained with 20% waymo training data): 2022-09-25 13:47:13,688 INFO *************** Performance of EPOCH 30 ***************** 2022-09-25 13:47:13,688 INFO Generate label finished(sec_per_example: 0.0324 second). 2022-09-25 13:47:13,689 INFO rec...

Hello, where can I download the premodal "deeplabv3_resnet50_coco-cd0a2569.pth", mentioned in voxel_rcnn_nultimodel yaml

I run the training on AWS EC2 instance, g4dnx12large with Tesla T4 gpu (4x16) gpu memory. batch_size=8 is ok. But the training was still kind of slow, the multi-gpu communication is slower and does not use NVLink. It's expensive though, $4/hr. Better to use P3 multi-GPU instances, it uses NVLink f...

"<THC/THC.h> no such file or directory" occured when I try to build OpenPCDet with torch1.12. I solve this problem as following: remove all <THC/THC.h> in ops dir. replace all xxx.data<type>() in ops dir with xxx.data_ptr<type>(). replace all x.type().is_cuda() in ops dir wi...

hi, The training works ok for batch_size=2 in 2 gpu machine. However, the training fails with OOM error for batch_size=4 or more. What should I change, the log details are provided below: /////////////////////////////////////////////////////////////////////////////////////////// batch_size=4 ///////...