"warms up" the final layer for two epochs before doing the hyperparameter search. this should keep the pre-trained weights from over-adjusting based because of the randomly initialized final layer. this seems to help finetuning performance on several molnet tasks when pretraining using MTR especially
"warms up" the final layer for two epochs before doing the hyperparameter search. this should keep the pre-trained weights from over-adjusting based because of the randomly initialized final layer. this seems to help finetuning performance on several molnet tasks when pretraining using MTR especially