Some Questions about re-implement VectorNet

[Hugh-Cai]

Hi,

Are there some details (or tricks) when construct the model vectornet? I tried it according to the paper description, but the loss converges very slowly whatever learning rate. The loss term using mse_loss on (x_offset, y_offset) is

Epoch: 0 | loss: 0.2395   ADE: 33.5584  FDE: 98.2625
Epoch: 1 | loss: 0.2312   ADE: 34.1547  FDE: 99.9695
Epoch: 2 | loss: 0.2270   ADE: 32.7290  FDE: 96.1623
Epoch: 3 | loss: 0.2246   ADE: 31.6781  FDE: 93.0649
Epoch: 4 | loss: 0.2222   ADE: 30.9941  FDE: 91.0348
...

(decay only very little per epoch, and hyper-params is the same as the reference)

I‘m appreciate if you give me an answer! Thanks

[Henry1iu]

Hi,

Could you please share more information, such as your trajectory representation and your loss function?

Best Regards, Jianbang

[Hugh-Cai]

Hi,

I used a standard MLP as decoder to predict the offsets (both x and y) in future 3 seconds, and used mse_loss as prediction loss and smooth_l1_loss as graph completion loss.

I implemented this only using torch without torch_geometric，so if did any questions appear in the encoder? Can I email the details to you if it's convenient？

Thanks for answering me!

Best Regards

[Henry1iu]

Hi,

The result (ADE and FDE) doesn't make sense if your model predicts the offsets. I suggest you double-check your offset ground truth and the post-prediction processing (the offset accumulation).

You can contact me via my email: henryliu@link.cuhk.edu.hk. I'll try my best to help you identify the problem : )

Best Regards, Jianbang

[Hugh-Cai]

Thanks again!

And I described my understanding and details of implementation in the email. Please correct if I have any bias.

Best Regards

[JingyiShen99]

Hi, I have the same problem now, and I also implemented this only using torch without torch_geometric. Have you solved this problem? Could you give me some suggestions or ideas?

Best Regards.