Schematic representation of the architecture employed in this paper. Sampled amplitude and phase distributions serve as input. After a preconvolution followed by a PRelu activation (He et al. 2015b), the main part of the architecture is composed of eight residual blocks (He et al. 2015a; Gross & Wilber 2016). One block consists of two convolution batch norm pairs with a PRelu activation in between. The input data is bypassed and added element-wise to the output for every residual block. The last residual block is followed by a convolution batch norm pair. Another skip connection is placed around the main part of the architecture. After a final convolution the reconstructed amplitude and phase distributions are the output of the architecture. Diagram reproduced from (Ledig et al. 2016).