why the eval result mAP that i run out is always about 0.005?

[IndigoChildren]

[DerrickXuNu]

Does this just happen to Where2comm or any other model?

[IndigoChildren]

any other model, actually i think maybe there is some wrong in the eval_utils.py, because it puts all the prediction results together, but the confidence score descend operation is used in every single batch.

[DerrickXuNu]

No, I don't think eval_util has an issue, as it runs well on several of my machines, and there are already many research groups using this without any issue. What is your spconv version?

[IndigoChildren]

spconv-cu113, 2.3.3

[DerrickXuNu]

try pytorch 1.12.0 + spconv 2.2.6

[DerrickXuNu]

also, can you show your inference command?

[socialism-redstar]

python opencood/tools/inference.py --model_dir opencood/opencood/checkpoints/point_pillar_where2comm_v2xset --fusion_method 'intermediate'

[IndigoChildren]

I try pytorch 1.12.0 + spconv 2.2.6 again, but get a same result. And i try to inference a single frame, superiseingly get a correct result about 0.88.

[IndigoChildren]

[DerrickXuNu]

@Little-Podi have you met this before? Could you have a try to see whether you will get such result with current opencood version? Many thanks!

[Little-Podi]

I never met this issue before. I can test the up-to-date version within this week. I'm taking a vacation in mainland right now ^_^. @IndigoChildren Have you tried the previous version of OpenCOOD? I have verified the codebase just 3 or 4 commits back. I'm using pytorch 1.13 + cuda 11.6 + spconv 2.2.6.

[IndigoChildren]

Totally, I only use one version. Thanks a lot for you spending time to deal with my problem, have a nice trip in mainland. By the way, welcome to ShenZhen! It's really hard to get a concert ticket of Jay Zhou.

[Little-Podi]

Haha, I know! I have never successfully got one even there are some many concerts hold in Hong Kong. I just tested the evaluation code today and found no problem in the current version. Did you customize sth?

[IndigoChildren]

I try to train a fcooper model myself, and i get a same bad validation result, i dont know what happened to my code.

[IndigoChildren]

I will try the current version and thanks for your help.

[IndigoChildren]

@Little-Podi 我昨天试了一下最新版本的代码，然后试着train了一个point_pillar_fcooper.yaml这个文件，各种配置都没有更改，训完以后用logs里面的文件在test集上测试了一下精度，发现精度几乎为0。请问能不能麻烦您也试着这样做一下，为什么我一直出bug，真的绷不住了。

[IndigoChildren]

而且在训这个fcooper模型的时候，好几次validation_loss直接跑飞了

[IndigoChildren]

[sidiangongyuan]

而且在训这个fcooper模型的时候，好几次validation_loss直接跑飞了

解决了这个问题吗？ 我现在也有这个问题。 在CoCa3D 上 的效果非常差