youngskkim / CRN

[ICCV'23] Official implementation of CRN: Camera Radar Net for Accurate, Robust, Efficient 3D Perception
MIT License
95 stars 14 forks source link

can't reproduce the r50 #19

Open loserhou opened 1 month ago

loserhou commented 1 month ago

I use 4 a800 and bs=8 mAP: 0.4681 mATE: 0.5202 mASE: 0.2827 mAOE: 0.5189 mAVE: 0.2834 mAAE: 0.1873 NDS: 0.5548 Eval time: 111.6s but your map is 47.3 and nds is 56.2,can you give me some advice?

adeelajmal2468 commented 1 month ago

@loserhou image results on CRN with r50

loserhou commented 1 month ago

@loserhou image These are my results on CRN with r50 how to set the learning rate?

loserhou commented 1 month ago

@loserhou image These are my results on CRN with r50 how to set the learning rate?

the original lr/8?

loserhou commented 1 month ago

@loserhou image These are my results on CRN with r50

is this the weights given by the author or did yiu train them yourself?

adeelajmal2468 commented 1 month ago

The checkpoints were given

loserhou commented 1 month ago

The checkpoints were given

have you trained by yourself?

adeelajmal2468 commented 1 month ago

Yes I did try it but it was taking too long for 1 epoch. So for now I just tested it

loserhou commented 1 month ago

Yes I did try it but it was taking too long for 1 epoch. So for now I just tested it

thx

youngskkim commented 3 weeks ago

Different seeds or GPU architectures may result in performance differences of up to a few percent.