In computational imaging, one wants to reconstruct an object from measured intensities. Since cameras typically record the modulus square of a complex-valued field, one often has to solve a Phase Retrieval problem. This includes applications such as ptychography or imaging in complex media. In these applications, an initial guess can often be available to speed up the reconstruction. We will investigate how to linearize this non-linear optimization problem around the initial guess and write an efficient implementation in Pytorch, a Python library designed for neural networks. As such, a link with neural network with quadratic activations will be drawn.