hustvl / MapTR

[ICLR'23 Spotlight & IJCV'24] MapTR: Structured Modeling and Learning for Online Vectorized HD Map Construction
MIT License
1.08k stars 163 forks source link

Using vovnet99 as backbone #122

Open ChenccLOL opened 12 months ago

ChenccLOL commented 12 months ago

Thanks for your great work!

When I try to train maptrv2 with vovnet99 as backbone,the grad_norm is always nan. Reducing the learning rate and applying gradient clipping are both ineffective. Can you provide the config file for training using vovnet99 as the backbone? If possible, I would greatly appreciate it.

brianlan commented 10 months ago

I met exactly the same problem.

brianlan commented 10 months ago

Thanks for your great work!

When I try to train maptrv2 with vovnet99 as backbone,the grad_norm is always nan. Reducing the learning rate and applying gradient clipping are both ineffective. Can you provide the config file for training using vovnet99 as the backbone? If possible, I would greatly appreciate it.

Do you finally got the solution to this problem?

ChenccLOL commented 8 months ago

Thanks for your great work! When I try to train maptrv2 with vovnet99 as backbone,the grad_norm is always nan. Reducing the learning rate and applying gradient clipping are both ineffective. Can you provide the config file for training using vovnet99 as the backbone? If possible, I would greatly appreciate it.

Do you finally got the solution to this problem?

Not yet, how about you?

Wolfybox commented 6 months ago

@ChenccLOL @brianlan Did you try using pretrained weights from https://github.com/TRI-ML/dd3d ?

GeMap provides config file for using V2-99:https://github.com/cnzzx/GeMap

brianlan commented 6 months ago

@ChenccLOL @brianlan Did you try using pretrained weights from https://github.com/TRI-ML/dd3d ?

GeMap provides config file for using V2-99:https://github.com/cnzzx/GeMap

Oh, that's good to know. Thanks for the information, buddy.

Wolfybox commented 6 months ago

@ChenccLOL @brianlan Did you try using pretrained weights from https://github.com/TRI-ML/dd3d ? GeMap provides config file for using V2-99:https://github.com/cnzzx/GeMap

Oh, that's good to know. Thanks for the information, buddy.

I managed to load pretrain weights correctly with config file from GeMap and pretrained weights from dd3d, so i guess you guys could also give a try. Good luck.