Giters

wqi / WIMP

[arXiv] What-If Motion Prediction for Autonomous Driving ❓🚗💨

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Reproducing the Map-Free and only Social-Context Results form the Ablation Study

SchDevel opened this issue · comments

commented

Hey there,

I want to reproduce the results of your ablation study, where you only used Social-Context with EWTA-Loss.

image

However, I habe problems training the model only with social context.
What are the correct flags I need to set for preprocessing (run_preprocess.py) and for training (main.py)?

Looking forward hearing from you soon!

Best regards

SchDevel

commented

Hi SchDevel, apologies for the delayed response - my CMU inbox is no longer active so I missed your initial email.

Unfortunately, there isn't a "simple" way to run this ablation with only social context using the current codebase. You'd likely have to remove the waypoint prediction step within WIMP_encoder.py and make the corresponding changes in WIMP_decoder.py.

I would've liked for this to be easier to reproduce and I'll try to make this something that's runnable with a single flag in the future, but I haven't been actively worked on this codebase in quite some time. Thanks for understanding!

commented

Hey,

thank you for your response.
I want to compare an approach I am currently working on with your WIMP approach (because WIMP is very well-thought and has publicly available code).
One part of this is a comparison of the amount of learnable model parameters on your map-free experiment.

I tried to count these parameters in two different ways:

  • Reimplementation of your model (without the polyline attention) with the implementation details given in the paper: ~23.000.000 learnable parameters
  • Following your suggestion, I removed the whole centerline processing and the waypoint prediction part of your publicly available code: ~25.000.000 learnable parameters.

Therefore, in my comparison, I would conservatively describe that the results of your map-free evaluation were achieved with >20.000.000 learnable parameters.
Would you agree with this statement?

As already written, I will of course cite your paper accordingly.

Looking forward hearing from you soon!

Best regards

SchDevel