Training descriptions serialization

[shuttle1987]

Related to #153

As it stands the training descriptions are made in an ad-hoc format:

persephone/persephone/model.py

Lines 202 to 215 in aa69575

	#Get information about training for the names of output files.
	frame = inspect.currentframe()
	# pylint: disable=deprecated-method
	# It was a mistake to deprecate this in Python 3.5
	args, _, _, values = inspect.getargvalues(frame)
	with open(os.path.join(self.exp_dir, "train_description.txt"), "w") as desc_f:
	for arg in args:
	if type(values[arg]) in [str, int, float] or isinstance(
	values[arg], type(None)):
	print("%s=%s" % (arg, values[arg]), file=desc_f)
	else:
	print("%s=%s" % (arg, values[arg].__dict__), file=desc_f)
	print("num_train=%s" % (self.corpus_reader.num_train), file=desc_f)
	print("batch_size=%s" % (self.corpus_reader.batch_size), file=desc_f)

There are a few concerning things here, especially the use of getting info directly out of the frame. I think this was a really smart way of dealing with this initially but I'd like to move over to use a more specific method from Tensorflow for extracting relevant variables, especially if more than one model is available at a time.

I propose that we make this a more easily parseable format when we write this to a file.

I'm thinking something like YAML or JSON works well here. Is there a preference for serialization format?

JSON comes in the standard library but has a bit less flexibility. That said flexibility isn't necessarily a benefit here.

[oadams]

I don't have a format preference. But whatever we do here it's important that the output file gives a complete picture of the experiment so it can be reproduced. This should include, but may not be limited to: the model information and the corpus information (as well as perhaps a hash of the data too).