We develop a novel ensembling technique by creating libraries of models using snapshot ensembles. We conduct k runs and take n snapshots during each run. We try to select optimal combinations of models from this library. However, it fails to outperform a naive deep ensemble baseline.