RNN size with BLSTM/PDBRNN

Hi,

When RNN size is set to X and BLSTM type cell is used (pyramidal encoder), does it mean we have X units in each direction or total X units?

Also, could you please include the feature to specify different number of units for encoder and different for decoder? Recent models show lesser number of units in decoder helps.

Thank you!

Hello,

When -rnn_size is set to x it means the RNN layer output has a dimension of x. So the number of units in each direction depends on the -brnn_merge option:

  • -brnn_merge sum (default): x units in each direction
  • -brnn_merge concat: x/2 units in each direction

Maybe setting a different number of layers would also for you. See the options -enc_layers, -dec_layers, and the bridge feature.