Overview of our main CNN architecture. The input has four different filter images for each lens system and each image a size of 64 × 64 pixels. The network contains two convolutional layers (conv) each followed by a max-pooling layer (max pool) with kernel size f and stride S values indicated in the figure. This is then followed, after flattening the data cube, by three fully connected (FC) layers to finally obtain the five output values of the SIE η, containing the lens center x and y, the complex ellipticity e_x and e_y, and the Einstein radius θ_E.