Config mrt_ml_mix seems to be unused

[luou-wen]

Hello, I noticed that the config mrt_ml_mix doesn’t seem to be referenced anywhere outside of the config code. I was wondering if this was by design, or if there is a mrt/mile mixed loss that is yet to be implemented.

I would be interested in using / helping develop the feature if possible.

[rsennrich]

thanks for pointing this out! Yes, this parameter was used in the Theano branch, but hasn't been implemented in the Tensorflow branch. A pull request adding it would be very welcome!

You should be able to implement this in mrt_cost() in mrt_utils.py , making use of the translation probabilities that are stored in cost. There's already code to include the reference in the list of samples (config.mrt_reference), so make use of that, identify the reference, and add its cost to MRTloss (see the Theano implementation for reference:

nematus/nematus/nmt.py

Line 643 in 5727727

def mrt_cost(cost, y_mask, options):

).