search

Neuroglycerin / neukrill-net-tools

Tools coded as part of the NDSB competition.
MIT License
0 stars 0 forks source link

Mixing in prior #109

Closed gngdb closed 9 years ago

gngdb commented 9 years ago

Before submission, could slightly improve generalization performance by mixing in a weighted average of the prior to regularise. Should be easy to do this to submission csv files, but should be able to specify in run settings so it plugs into the full pipeline; just runs at test time.

May have to store class priors from training set in settings.json. They are calculated in the notebook plotting the data from the start of the competition.

scottclowe commented 9 years ago

I have already specified class_priors as a property in the utils.Settings class, so you should use that.

gngdb commented 9 years ago

Tried this and it did not improve the score. See leaderboard results.