Closed gngdb closed 9 years ago
Could not find a bug in the csv writing process. Trained a model from scratch: alexnet_based_objective
. Watches objective to save best model. Got better score doing this than before, moved 26 places up leaderboard. Suspect problems were due to pickle saving problems but not totally sure at this point.
Turns out only models applying normalisation were having problems because normalisation was not being applied to test set on load.
Recent submissions were terrible, but logs indicated they would be OK. Either they're broken or the classes are getting garbled. Needs fixed.