DerrickXuNu / OpenCOOD

[ICRA 2022] An opensource framework for cooperative detection. Official implementation for OPV2V.
https://mobility-lab.seas.ucla.edu/opv2v/
Other
663 stars 99 forks source link

Training v2vnet got the wrong result #93

Closed rujiawang329 closed 1 year ago

rujiawang329 commented 1 year ago

I have trained v2vnet for 60 epochs. However, when I test the last epoch results, I get that all IOU is 0.

The output of the result is in the below picture. 862291d09de027ddeadb5accbc1de81

Here is some background information about my training.

First, about the environment, spconv is spconv-cu113 2.3.6, pytorch is 1.11.0+cu113. The more detailed environment information is in the bellow picture. 3c9ec5fc004e6769827d22ffd6d85d3

Next, about the training and inference code. I use python opencood/tools/train.py --hypes_yaml opencood/hypes_yaml/point_pillar_v2vnet.yaml for train. I use python opencood/tools/inference.py --model_dir opencood/logs/point_pillar_v2vnet_2023_06_05_06_40_58/ --fusion_method intermediate for test.

Last, the training loss can be seen in the below picture. 7975395262f8b2c8e7816e2a949236e

Thanks a lot for helping me to solve this problem.

DerrickXuNu commented 1 year ago

V2VNet is hard to train, sometime you need some luck to have good weight initialization you train to coverage. I suggest directly using the checkpoint that I provide as a start point.

rujiawang329 commented 1 year ago

OK, thanks!