We were using the wrong learning rate for fully-supervised because we were setting the consistency model to mean_teacher with max_cons_multiplier=0. This ended up causing a hyperparameter override to set the wrong learning rate. This PR adds a consistency model none which explicitly avoids using any consistency model and also allows us to have a separate hyperpamater override dict for fully-supervised learning (hopefully preventing this kind of issue in the future).
We were using the wrong learning rate for fully-supervised because we were setting the consistency model to
mean_teacherwithmax_cons_multiplier=0. This ended up causing a hyperparameter override to set the wrong learning rate. This PR adds a consistency modelnonewhich explicitly avoids using any consistency model and also allows us to have a separate hyperpamater override dict for fully-supervised learning (hopefully preventing this kind of issue in the future).Fixes #19