Log hp_metric only if improved

[lukasfolle]

Hi 👋

I like to use Tensorboard to log my runs and compare different hyperparameter configurations. Sorting by hp_metric help identifying good runs, but since the last hp_metric might not be the best of a given run I would like to only log it if the metric acutally improved.

I couldn't find out how to access past stored metrics using self.logger.experiment but this should be possible right?

Then I could do:

if past_hp_metric < current_hp_metric:  # max
    self.log("hp_metric", current_hp_metric)