As discussed in issue #141 training takes too long and spends most of the compute time generating the pairlist. To move this from a per-epoch cost to a one-time cost this PR moves the calculation of the pairlist to the dataset and saves the pairlist.
The pairlist is then passed, as an optional parameter, as input to the network. The Neighborlist class takes it as optional input and calculates distances to further filter it to the local neighborhood for each atom within a cutoff.
Description
As discussed in issue #141 training takes too long and spends most of the compute time generating the pairlist. To move this from a per-epoch cost to a one-time cost this PR moves the calculation of the pairlist to the dataset and saves the pairlist. The pairlist is then passed, as an optional parameter, as input to the network. The Neighborlist class takes it as optional input and calculates distances to further filter it to the local neighborhood for each atom within a cutoff.
Todos.
Status