search

kbrodt / waymo-motion-prediction-2021

Waymo: Motion Prediction 2021
186 stars 45 forks source link

Loss #15

Closed zsgj-Xxx closed 2 years ago

zsgj-Xxx commented 2 years ago

Hi! I tried your method and I observed that in training, l2loss and log_softmax have so large difference. so my network does not learn multimodal tracks, only one best track is fitted. Do you have any solution?

kbrodt commented 2 years ago

Hi! We don't train with L_2 loss. We train with negative log-likelihood loss of Gaussian mixtures to capture the multimodality.

zsgj-Xxx commented 2 years ago

Oh. Sorry for my expression errors. What I use is the Loss in the code. However, my training results are inconsistent with the results in the paper. Among my 6 trajs, there is only one close to gt, others pred are near the original point ![Uploading mmexport1661165075704.png…]()

kbrodt commented 2 years ago

Do you observe this for all examples?

zsgj-Xxx commented 2 years ago

Yes, I observed nearly 2,000 samples in val, all got similar results, and they all got the best score in the same pred id.

kbrodt commented 2 years ago

It seems your model is overfitted. We didn't observe that behaviour. Do you train on full training dataset?

zsgj-Xxx commented 2 years ago

no yet, I just used 1/10 training data to verify the effect

zsgj-Xxx commented 2 years ago

Oh I found the problem. I put the gt at the vehicle point instead of the origin in the preprocessing stage. Thank you for your reply.

Best,