LinRNN.py

import numpy as np

class LinRNN:
    def __init__(self, alpha, A, Omega):
        self.alpha = alpha
        self.A = A
        self.Omega = Omega
        self.num_states = alpha.shape[0]
        self.input_dim = A.shape[1]
        self.output_dim = Omega.shape[1] if Omega.ndim > 1 else 1


    def update_dynamics(self, prev, obs):
        next = np.tensordot(prev, self.A, [prev.ndim - 1, 0])
        obs = obs.reshape(len(obs), -1)
        next = np.tensordot(next, obs, [0, 0])
        next = next.reshape(-1, )
        return next

    def term_dynamics(self, prev):
        term = np.tensordot(prev, self.Omega, [prev.ndim - 1, 0])
        return term

    def predict(self, obs_sequences):
        current_state = self.alpha
        for o in obs_sequences:

            current_state = self.update_dynamics(current_state, o)
        term = self.term_dynamics(current_state).ravel()
        pred = term if self.output_dim > 1 else term[0]
        return np.asarray(pred)

    def build_true_Hankel_tensor(self,l):
        H = self.alpha
        for i in range(l):
            H = np.tensordot(H,self.A,[H.ndim-1,0])
        H = np.tensordot(H,self.Omega,[H.ndim-1,0])
        return H