having .s() return nothing on an RNN's initial state is very confusing (at least to me), as there must be some kind of value from which the RNN starts and not revealing it just hides what's going on (making things like automaton-extraction very difficult).
to my understanding, RNNStates with empty .s() are effectively treated as having zero-vectors in .add_input() computations (the matrix multiplication with the missing c/h vectors is simply skipped). It would be nice if .s(), .h(), .output(), etc reflected this.
having .s() return nothing on an RNN's initial state is very confusing (at least to me), as there must be some kind of value from which the RNN starts and not revealing it just hides what's going on (making things like automaton-extraction very difficult).
to my understanding, RNNStates with empty .s() are effectively treated as having zero-vectors in .add_input() computations (the matrix multiplication with the missing c/h vectors is simply skipped). It would be nice if .s(), .h(), .output(), etc reflected this.