open-mmlab
KeyError: "NuScenesDataset: 'infos'"
Hi @Wuziyi616 && @sunjiahao1999 && @lijain
I encountered a problem during execution, but I don't know what to do. Thank you for your help
Now the problem is that it does not understand the word infos in the last line of the above message, why? What should I do to solve this problem?
This looks like a version issue, make sure the mmdet3d version of bevfusion is the same as the corresponding annotation file. By the way, we now highly recommend using mmdetection3d version 1.1, which fully supports the Openmmlab 2.0 ecosystem and is much simpler to use.
BEVFusion training will support in next version (to be released next week). Welcome to follow.
Hi @sunjiahao1999 I made some changes to the code to work with mini data. The following photo is my changes:
I tried to change its version to mini, and after that, we don't have infos in the data keys, it only had data_list and metainfo, which I changed in the circled part so that it can understand it, in addition to that, gcc I upgraded it as if above 5 was intended, the numba library was also necessary, I installed it, then I typed the following code for evaluation: (from this link https://github.com/mit-han-lab/bevfusion#evaluation)
torchpack dist-run -np 8 python tools/test.py configs/nuscenes/det/transfusion/secfpn/camera+lidar/swint_v0p075/convfuser.yaml pretrained/bevfusion-det.pth --eval bbox
After downloading a cache for Torch still remains in this section:
What needs to be done?
I got the pkls from this link that I asked one of your friends and I am working with them
https://github.com/open-mmlab/mmdetection3d/issues/2454#issuecomment-1517743436
Hi @sunjiahao1999 Based on the question I asked, the problem was solved: https://github.com/mit-han-lab/bevfusion/issues/409
But I ran into another problem:
Now I checked the code, there is a problem in this part:
As you can see, we don't have lidar_path or location or sweeps at all.
what should I do ?
BEVFusion training will support in next version (to be released next week). Welcome to follow.
Hello again, can you send me the link if it is published?
