Obtain best model right after training
Following #138 (closed) , several checkpoints of the model are created during the training procdeure. However, in order to save memory on disk, we would only like to retain the best and the last model checkpoint, while the others should be deleted.
Instead of performing this step in the postprocessing, we decided to shift the task to the training-substep of the workflow. Thus, after training, a new model graph (in test-mode) will be created and the quick postprocessing will be executed for each model checkpoint to detect the best performing model. This procedure will be based on the validation dataset since optimization on the test-dataset is actually forbidden. After checking the model checkpoints, all checkpoints, but the last and the best ones will be removed from disk.