Fourier Ptychography is an emerging microscope technique which is increasingly used for biological samples of high importance. The technique allows for stunning resolution without having to cope with a restricted field of view. This result is achieved by combining phase retrieval with an aperture synthesis procedure, indeed casting the technique as one of the most advanced computational imaging investigation methods. Due to the nature of the inverse problem involved (phase retrieval), currently only iterative algorithms can be deployed for reconstruction, thus delaying the image visualization from the acquisition process. In this paper, we propose a deep learning method to seed the iterative reconstruction and obtain a higher quality result in a shorter time. A parameter agnostic CNN is trained to produce an initial estimate for the iterative process. The final reconstructions exhibit a reduced number of artefacts, even for a limited illumination Numerical Aperture. Our method is decisive to relax the design of the illumination array.