Same cell for word and sentence level about hierarchical-attention-networks HOT 3 OPEN

AFAIR Cell is/was a template/factory, not a layer or parameter container

from hierarchical-attention-networks.

I think I am getting you. BNLSTMCell is just a declaration of a class. Still when you say call BNLSTMCell.call(), it will still call same set of parameters which are already defined in the graph at word level as they fall in same namescope of defined parameters. So treated as one unique set of parameters, not two (word and sentence).

I am seeing a big performance difference by making this change. ( i am trying in a little more complex problem of multi-class multi-label )

cell_word = BNLSTMCell(40, is_training) # h-h batchnorm LSTMCell
cell = GRUCell(30)
cell_word = MultiRNNCell([cell]*5)

cell_sent = BNLSTMCell(40, is_training) # h-h batchnorm LSTMCell
cell = GRUCell(30)
cell_sent = MultiRNNCell([cell]*5)

Similarly if you expect a different cell for forward and backward then you should define two more cells.

please correct me if I am wrong.

Other small side thing, according to my understanding, it's a general practice to not to dropout at eval time but it's happening as defined in the code here.

from hierarchical-attention-networks.

You might be right, my memory of TF's conventions is quite vague at this point.

from hierarchical-attention-networks.