Phase-retrieval techniques aim to recover the original signal from just the modulus of its Fourier transform, which is usually much easier to measure than its phase, but the standard iterative techniques tend to fail if only part of the modulus information is available. We show that a neural network can be trained to perform phase retrieval using only incomplete information, and we discuss advantages and limitations of this approach.