view article
Figure 2
Application of the BraggNN deep neural network to an input patch yields a peak center position (y, z). All convolutions are 2D of size 3 × 3, with a rectifier activation function. Each fully connected layer, except for the output layer, also has a rectifier activation function. The non-local attention block will be discussed in Appendix A[link].

IUCrJ
Volume 9| Part 1| January 2022| Pages 104-113
ISSN: 2052-2525
https://doi.org/10.1107/S2052252521011258