Transition

ns_sdn.typing.Transition = Transition

Alias of Literal['gru', 'timeaware'].

Latent transition \(F_\psi\).

"gru" is a gated recurrent cell on the input row \(x_t = [y; x]\), the configuration of the IEEE-ORSS experiments,

\[r_t = \sigma(W_r x_t + U_r h_{t-1} + b_r), \quad z_t = \sigma(W_z x_t + U_z h_{t-1} + b_z), \quad \tilde h_t = \tanh\bigl(W_h x_t + U_h (r_t \odot h_{t-1}) + b_h\bigr),\]

\[h_t = (1 - z_t) \odot \tilde h_t + z_t \odot h_{t-1},\]

in which the step size enters only the phase recursion. "timeaware" is the residual Euler update

\[h_t = h_{t-1} + \Delta t_t \, g_\psi\bigl([h_{t-1}; x_t]\bigr),\]

with \(g_\psi\) a one-hidden-layer tanh network, the discrete-time analogue of a neural ODE, in which irregular spacing enters the dynamics explicitly. The GRU runs the teacher-forced pass as one recurrent call; the time-aware cell loops over time in Python.