During the training (in the trainer.Trainer.train() method), if the loss is nan (such as when using very few datapoints, like in #528), checkpoint_model is never created, thus the crashes and gives the following error:

UnboundLocalError: local variable 'checkpoint_model' referenced before assignment

TODO:

• Handle such cases
• Add a test
• Release a new fixed version (patch)
• Check that docker installation works with the new version, by running the tutorials (we expect that using cutoff 8 and only 15 data points the loss is not nan)