Unexpectedly bad results for billiards.py variant
Opened this issue · 5 comments
In billiards.py, I've done some changes to the starting condition of the cue ball and the target position. These changes are available here.
What I observed was that in almost every iteration, the result got worse than before not better.
i.e. Loss seems to be increasing:
The ending stable state is not that surprising (no gradient) but the initial trend is concerning me.
The billiards example is not implemented with the time-of-impact fix (see our paper) so I'm not surprised that it's making inverse progress. I'll fix this in a few days.
Hi @DomNomNom ,
I added the TOI fix to billiards.py. Gradient quality is significantly improved. Could you have a try? (PS: please also upgrade taichi to 0.4.6+). Thanks.
Hi @yuanming-hu
I've updated taichi and merged your changes, however I'm still seeing similar symptoms:
Here's what I'm running: https://github.com/DomNomNom/difftaichi/blob/5326ebfc1c8bd5bfe56b3ba5b68ad891b158aa20/examples/billiards.py
Note: after updating, pip3 list | grep taichi
gave me taichi-nightly-cuda-10-1 0.5.0
.
Any thoughts on this?