Closed rginjapan closed 1 year ago
@rginjapan - Thanks for raising this issue. What exactly do you mean by "debug the code during training"? The code you shared is only for inference during ObjectNav testing.
Thanks for reply, I see, so how to balance these two pfs durning training?
Thanks for quick response, I see, could you plz explain why not keep the consistency design of balancing two pfs between training and testing?
If you recall, the key motivation of our work is to perform supervised training of our potential functions without any environment interactions, i.e., we are not performing ObjectNav during training (no balancing). The balancing is needed only while performing ObjectNav during inference. The balancing is quite simplistic since the coefficient is fixed throughout.
Thanks! So during training, there is 1object_pf + 1area_pf, but in inference is 0.5object_pf+0,5area_pf? What is the difference?
Do you still have more questions related to this issue, @rginjapan?
Closing this issue due to inactivity.
https://github.com/srama2512/PONI/blob/30682c2bdcd820eec8f72043b2579eb045d547bf/semexp/model_pf.py#L169C17-L169C55
I am interested in the combination of area_pf and object_pf, I think it is on the line of above link, but when I debug the code durning training, it seems that this comination(the code) is not used?