Sampling decoding about nmt-keras

[lvapeab]

The models defined here are meant to be used with beam search, as they are autoreggressive (see Sec. 2 of this paper).

Of course, the keras-wrapper library supports regular sampling (sample from a distribution probability, given by a model), as done for example by a classifier.

Non-autoreggressive NMT is not implemented here. If you want to use it, you should modify the models and implement something similar to the paper above.

If you want to stick to autoreggressive NMT, simply generating the most-likely token (or a token sampled from the model's distribution), you can set the beam size to 1. Moreover, if that's you use case, you can create some sort of predictGreedySearchNet function, that doesn't require all the beam search logic. PRs are welcome :)