flare.rnn

build-seq

(build-seq cell inputs)(build-seq cell inputs bidrectional?)

return sequence of add-input outputs for a given RNNCell and sequence of inputs. Can optionally make sequence building bidrectional using bidirectional? optional third argument.

Returned sequence will drop the init-hidden element which doesn’t correspond to an input.

lstm-cell

(lstm-cell model input-dim hidden-dim)

Standard LSTM cell https://en.wikipedia.org/wiki/Long_short-term_memory without any peepholes or other adaptations.

The hidden state is a piar (output, cell-state)

RNNCell

protocol

members

add-input

(add-input this input last-hidden)

take last hidden and reutrn new hidden

cell-model

(cell-model this)

return model underlying cell

init-hidden

(init-hidden this)

return initial (output, cell-state) pair

input-dim

(input-dim this)

dimension of inputs

output-dim

(output-dim this)

output dimension of cell