Continuous-variable systems can be characterized through phase-space representations, e.g. the Husimi Q representation. Current experiments make such measurements easier, though the data is typically sparse and noisy. The readout can be optimized using machine learning methods, with invertible neural networks being especially suitable for reconstructing the Husimi Q function. Ultimately, quantum resources such as non-Gaussianity and noni-classicality can be quantified from the reconstructed state.