Interesting.

So, to decode you would start with the unoptimized model and decode the first batch. You then train using that first batch to get a better model and use your improved model to decode the next batch.

Okay, that makes sense; thanks for the response!