Wu Yifan

> > > i met the same problem when train bevfusion , could you please teel me how to solve it? > > > The para sweeps_num of loadpointsfrommultisweeps function...

Parallel training may affect performance. Try adding `CUDA_VISIBLE-DEVICES=0, 1, 2, 3` before the training code. Another reason may be that The model does not converge. Try adding the epoch. PS:...

Modify the https://github.com/mit-han-lab/bevfusion/blob/main/configs/default.yaml#L17 to continuing training with the previous model.

I modified the [https://github.com/mit-han-lab/bevfusion/blob/main/configs/nuscenes/det/transfusion/secfpn/camera%2Blidar/default.yaml](url). ``` lidar: voxelize_reduce: false voxelize: max_num_points: 20 point_cloud_range: ${point_cloud_range} voxel_size: ${voxel_size} max_voxels: ${max_voxels} backbone: type: PointPillarsEncoder pts_voxel_encoder: type: PillarFeatureNet in_channels: 5 feat_channels: [64, 64] with_distance: false...

I have the same problem, too. And when I increase the `max_epoch`, the corresponding `lr` needs to be smaller. I have to adjust the `lr` again and again.

As for me, 4090. It needs about 20000 MB to train. sample_per_gpu=1

HI, @alexzhuuuu @AlexIlis I face the problem, too. I'm running a lidar-only model and the backbone is pointpillars. Here is the running code: `torchpack dist-run -np 1 python tools/train.py configs/nuscenes/det/transfusion/secfpn/lidar/pointpillars.yaml`...

Just comment the code `from flash_attn.flash_attention import FlashMHA` and the code `from .radar_encoder import *`. Because they are used for radar. Hope this can help you.

> I have the same problem, could you tell me how did you solve it ? thanks I I haven't found a solution yet.

No, I met this error when I used a 3090 GPU rented on AutoDL.[https://www.autodl.com/login?url=/market/list](url) Now I'm using the GPU of my own. And this error didn't occur. Maybe this error...