Based on this line of code:
https://github.com/ufal/neuralmonkey/blob/master/neuralmonkey/decoders/output_projection.py#L125

Current implementation isn't flexible enough; if we train a "submodel" (e.g. decoder without attention - not containing any ctx_tensors) we cannot use the trained variables to initialize model with attention defined because the size of the dense layer matrix input become

https://github.com/JayParks/tf-seq2seq/blob/master/seq2seq_model.py#L368
It gives that the dimension 0 of inputs and attention do not match (as we are tile_batching it to batch_size * beam_width). Didn't you get any error while running with beam_search?

Thanks for this wonderful library, but it would be much more intuitive for users to get started by providing some simple but clearly training process on self-defined dataset.

Make clear in documentation and small.yaml that encoder (and decoder?) embedding section takes a separate dropout argument and defaults to encoder dropout argument if missing.

Will send a pull request myself eventually and just document all issues I find along the way of setting joey up for me.

This could be visualized in another table, where e.g. the confidence of the system across different documents could be compared and contrasted.