Pytorch's implementation of add_hparams isn't really quite correctly. In particular, it creates a new events file for every write, which is not quite what we want. We should write our own implementation that writes everything to the same events file. Related to #157 .