We study the training dynamic of sparse neural networks through the lens of Graphon where masks of pruning methods converge to graphons as networks' width approach infinite