Different RNN architecture for each layer of Encoder


Is it possible to use a bi-directional RNN in the first layer, and uni-directional RNN in the remaining layers of the encoder as done by Wu et al. (2016): ?


No, this is not currently implemented but it could be quite easily added in the future.

Available architectures are described here.

I just added the Google’s NMT encoder by the way. See:


You also noticed the new -encoder_type option that replaces the boolean flags -brnn, -dbrnn and -pdbrnn.