Inception block used in our full network architecture, Figure B.1. The block is composed of convolutional layers (purple) and a max pooling layer (turquoise). The inception block uses a mix of 1 × 1, 3 × 3, and 5 × 5 convolution, all with same padding and a stride of one to maintain the input dimensions H × W. We show the number of channels as a fraction of the output number of channels as this is variable, as shown in Figure B.1.