Lowest validation loss from a previous session is not being carried over when continuing training in a new session

When the training is being interrupted and being resumed at a later point, the lowest validation loss from the previous session will not be carried over to the new session. This leads potentially to the replacement of the model with the lowest validation loss from the previous session with a model that has the lowest validation loss in the new training session. But this model from the new session does not necessarily have to have the lowest validation loss compared to the models from the previous session.

Temporary fix is to always backup the lowest validation loss model from an interrupted session before resuming the training, and compare the lowest validation errors manually after the full training has been completed to choose the model with the lowest validation loss.