uber-research / LaneGCN

[ECCV2020 Oral] Learning Lane Graph Representations for Motion Forecasting
https://arxiv.org/abs/2007.13732
Other
507 stars 131 forks source link

Learning Rate Drop #22

Closed ziqipang closed 1 year ago

ziqipang commented 3 years ago

Hi! Thank you for your cool project!

I have a question which might be stupid. I notice that in the repository no warm-up and dropping learning rate on every iteration are involved. Since these techniques are normal for other deep learning applications, I am wondering if the current style in LaneGCN works better or the structure of LaneGCN can already be good without such tricks.

Thanks!