BEVFusion: Wrong pytorch model/keys missing or misaligned while running 'demo'

[pravn]

Prerequisite

• [X] I have searched Issues and Discussions but cannot get the expected help.
• [X] I have read the FAQ documentation but cannot get the expected help.
• [X] The bug has not been fixed in the latest version (dev-1.x) or latest version (dev-1.0).

Task

I'm using the official example scripts/configs for the officially supported tasks/models/datasets.

Branch

main branch https://github.com/open-mmlab/mmdetection3d

Environment

I am unable to run the demo example in the README. It appears that the model has incompatible/missing keys:

The following error is obtained: `Loads checkpoint by local backend from path: bevfusion_converted.pth
The model and loaded state dict do not match exactly

Reproduces the problem - code sample

python demo/multi_modality_demo.py demo/data/nuscenes/n015-2018-07-24-11-22-45+0800__LIDAR_TOP__1532402927647951.pcd.bin demo/data/nuscenes/ demo/data/nuscenes/n015-2018-07-24-11-22-45+0800.pkl projects/BEVFusion/configs/bevfusion_lidar-cam_voxel0075_second_secfpn_8xb4-cyclic-20e_nus-3d.py ${CHECKPOINT_FILE} --cam-type all --score-thr 0.2 --show

Reproduces the problem - command or script

python demo/multi_modality_demo.py demo/data/nuscenes/n015-2018-07-24-11-22-45+0800__LIDAR_TOP__1532402927647951.pcd.bin demo/data/nuscenes/ demo/data/nuscenes/n015-2018-07-24-11-22-45+0800.pkl projects/BEVFusion/configs/bevfusion_lidar-cam_voxel0075_second_secfpn_8xb4-cyclic-20e_nus-3d.py ${CHECKPOINT_FILE} --cam-type all --score-thr 0.2 --show

Reproduces the problem - error message

The following error is obtained:

The model and loaded state dict do not match exactly
                                                          
unexpected key in source state_dict: vtransform.dx, vtransform.bx, vtransform.nx, vtransform.frustum, vtransform.dtransform.0.
weight, vtransform.dtransform.0.bias, vtransform.dtransform.1.weight, vtransform.dtransform.1.bias, vtransform.dtransform.1.ru
nning_mean, vtransform.dtransform.1.running_var, vtransform.dtransform.1.num_batches_tracked, vtransform.dtransform.3.weight,
vtransform.dtransform.3.bias, vtransform.dtransform.4.weight, vtransform.dtransform.4.bias, vtransform.dtransform.4.running_mean, vtransform.dtransform.4.running_var, vtransform.dtransform.4.num_batches_tracked, vtransform.dtransform.6.weight, vtransfo
rm.dtransform.6.bias, vtransform.dtransform.7.weight, vtransform.dtransform.7.bias, vtransform.dtransform.7.running_mean, vtra
nsform.dtransform.7.running_var, vtransform.dtransform.7.num_batches_tracked, vtransform.depthnet.0.weight, vtransform.depthne
t.0.bias, vtransform.depthnet.1.weight, vtransform.depthnet.1.bias, vtransform.depthnet.1.running_mean, vtransform.depthnet.1.
running_var, vtransform.depthnet.1.num_batches_tracked, vtransform.depthnet.3.weight, vtransform.depthnet.3.bias, vtransform.d
epthnet.4.weight, vtransform.depthnet.4.bias, vtransform.depthnet.4.running_mean, vtransform.depthnet.4.running_var, vtransfor
m.depthnet.4.num_batches_tracked, vtransform.depthnet.6.weight, vtransform.depthnet.6.bias, vtransform.downsample.0.weight, vt
ransform.downsample.1.weight, vtransform.downsample.1.bias, vtransform.downsample.1.running_mean, vtransform.downsample.1.runn
ing_var, vtransform.downsample.1.num_batches_tracked, vtransform.downsample.3.weight, vtransform.downsample.4.weight, vtransfo
rm.downsample.4.bias, vtransform.downsample.4.running_mean, vtransform.downsample.4.running_var, vtransform.downsample.4.num_b
atches_tracked, vtransform.downsample.6.weight, vtransform.downsample.7.weight, vtransform.downsample.7.bias, vtransform.downs
ample.7.running_mean, vtransform.downsample.7.running_var, vtransform.downsample.7.num_batches_tracked

Additional information

No response

[sunjiahao1999]

You should use new checkpoints in BEVFusion README.md

[hepingpeace]

You should use new checkpoints in BEVFusion README.md

Thank you new checkpoints, but when I use this checkpoint and use demo I find bboxes did not match to objects. it is demo find annotation file is different from that after tools/create_data.py nuscenes programs file?? and the bevfusion_converted.pth also has same problem.

[pravn]

I got the keys to align by hacking it. Load checkpoints and replace vtransform with view_transform to make it work. However, @hepingpeace notes, the boxes produced are incorrect.

Using the new checkpoint given above by @sunjiahao1999 seems to produce the correct result.

[jeromefan]

Hi, I encountered the same issue, when I use the pretrained model (name is bevfusion_converted.pth) from the doc of mmdet3d, the result is the same as @pravn 's first pic:

After check this issue, I tried the pretrained model (bevfusion_lidar-cam_voxel0075_second_secfpn_8xb4-cyclic-20e_nus-3d-5239b1af.pth) from @sunjiahao1999 ,but the performance is worse.

The pred_score_thr is set to 0.0;

The env I tried includes: python=3.9, pytorch=2.0.1, mmcv=2.1.0, mmengine=0.9.0, mmdet=3.2.0, mmdet3d=1.3.0, and python=3.9, pytorch=2.1.1, mmcv=2.1.0, mmengine=0.9.0, mmdet=3.2.0, mmdet3d=1.2.0, the mmdet3d v1.3.0 is downloaded from github repository Tags: v1.3.0, and the mmdet3d v1.2.0 is downloaded from github repository Branches: dev-1.x;

system: Ubuntu 20.04, gpu: 4090, cuda version: 11.8.

[HydrogenWasser]

hi, i've tried both bevfusion_lidar-cam_voxel0075_second_secfpn_8xb4-cyclic-20e_nus-3d-5239b1af.pth and bevfusion_converted.pth, but both returned "The model and loaded state dict do not match exactly" as posted above.

I am using the latest version of mm3d, I wonder which version of mm3d are you using when you can succesfully run the demo?

[zxzheng826]

I found out what go wrong. It seems it is crucial to install spconv 2.0 to run BEVFusion project correctly. It is mentioned in installing mmdet3d guideline Note 2. After I install cumm-cuxx and spconv-cuxx with pip, the problem fade instantly. So checking spconv 2.0 flag in environment may help you to diagnose.

[HydrogenWasser]

many thx. i'll go check and give it a try.

[jeromefan]

I found out what go wrong. It seems it is crucial to install spconv 2.0 to run BEVFusion project correctly. It is mentioned in installing mmdet3d guideline Note 2. After I install cumm-cuxx and spconv-cuxx with pip, the problem fade instantly. So checking spconv 2.0 flag in environment may help you to diagnose.

@zxzheng826 very useful with the checkpoint provided by @sunjiahao1999, can't thx more!!!! 牛逼!!!