-
Notifications
You must be signed in to change notification settings - Fork 0
Home
Jonathan Vance edited this page Apr 22, 2021
·
1 revision
Welcome to the karn-p3 wiki!
We achieved our highest IoU score of 7.7 using U-Net. Counterintuitively, we found that our accuracy decreased with more training, even as losses decreased. We began with high momentums in our SGD optimization function, around .99, and high loss rates of .1, but we gradually lowered these significantly, settling on .69 with a loss of .00001. We had similar experiences when we swapped out SGD for Adam, where we got the best results with lr = .00001 and betas about .6 & .799.
We tried to use Team Linden's implementations of OpticalFlow and Tiramisu, but we were unsuccessful at getting the former running, and we were unsuccessful seeing the latter through complete training (ran out of memory).