Shape error when making out-of-sample predictions

Questions version agnostic
,
14

Thanks. I am trying to use the scan function as follows:

            cores = np.empty(self.D, dtype=object)
            for d in range(self.D):
                cores[d] = (sqrt_prior_covariance / np.sqrt(np.sqrt(R[d]*R[d+1]))) @ pm.Normal(f'factor_{d}',mu=0, sigma = 1, shape = (self.M,self.R[d]*self.R[d+1]))

            # Model output
            def tt_loop(D, X, M, cores, R):
                temp = pm.math.matmul(feature_TT(X[0], M), cores[0]).reshape((R[0], R[1]))
                for d in range(1, D, 1):
                    temp = temp @ pm.math.matmul(feature_TT(X[d], M), cores[d]).reshape((R[d], R[d + 1]))
                return temp.flatten()

            result, _ = pytensor.scan(fn=tt_loop,
                        sequences=[X],
                        outputs_info=None,
                        non_sequences=[self.D, self.M, cores, self.R])

but I am running into some problems due to how cores is defined, namely as a numpy object array (in fact I get TypeError: Unsupported dtype for TensorType: object). What is the best practice in this case?