As discussed in the char-nn repo document, the training_loss mark in the checkpoint filename is very useful for sampling.
would be great to bring this back for quicker exploration together with a -output_dir option to store checkpoint files trained from different parameters