Theano implementation for PTB

[shimisalant]

Hi,

I've added 2 files: theano_rhn.py (where the model is defined), and theano_rhn_ptb.py (main program, which follows the code in rhn_train.py, e.g. it also uses Sacred). This implementation reaches to same SOTA results.

One thing to consider is how the existing data/reader.py was updated:

• It's dependence on Tensorflow was removed (so it could be used also by the Theano code). The single consequence here is that instead of doing tf.gfile.GFile(filename, "r"), we're doing open(filename, "r").
• A few functions from Tensorflow's code for reading PTB were copied.

Hopefully, this is okay, let me know otherwise.

Thanks for the great paper and code!

[flukeskywalker]

Fantastic, we were hoping for a Theano implementation from the community! I'll happily merge this. Two questions:

• Looks like the exact result you have reproduced is recurrence depth 10 with WT. Is this correct?
• enwik8 experiment incoming? :)

[shimisalant]

• Yes, that's correct.
• Sure :)

[flukeskywalker]

Great! :D