search

InhwanBae / GPGraph

Official Code for "Learning Pedestrian Group Representations for Multi-modal Trajectory Prediction (ECCV 2022)"
https://ihbae.com/publication/gpgraph/
MIT License
59 stars 7 forks source link

loss problem #2

Closed ximi1211 closed 1 year ago

ximi1211 commented 1 year ago

Hello, may I ask why my loss is always 0 since the first epoch? image image

InhwanBae commented 1 year ago

Hi @ximi1211!

Thank you for letting me know about this issue. In the last few hours, I've analyzed the cause of this problem. Finally, I figured out that one line of code that I rewrote to clean up before public release caused a problem. The equation is the same, but I still don't know why nan appears during back-propagation (I replaced the loss with 0 if its value is nan). Please replace the 59th line in model_groupwrapper.py with the PyTorch built-in function as follows:

# sig = 1 / (1 + ((dist_mat - self.th) / tau).exp())
sig = (-(dist_mat - self.th) / tau).sigmoid()
InhwanBae commented 1 year ago

I'm closing this issue for now. Please feel free to open another issue for other questions.