We propose a methodology for reconstruction of compressive spectral imaging from an uncalibrated optical system, where the propagation model is learned and included as regularizer to improve the reconstruction quality.