3.3. Calculator

class fitsnap3lib.calculators.calculator.Calculator(name, pt, config)

Class for allocating, calculating, and collating descriptors.

allocate_per_config(data: list)

Allocate shared arrays for total number of atoms. This is only needed when doing big A matrix fits (need number of atoms) or nonlinear fits.

Parameters

data – List of data dictionaries.

collect_distributed_lists(allgather: bool = False)

Gathers all the distributed lists on each proc to the root proc. For each distributed list (fitsnap dicts) this will create a concatenated list on the root proc. We use this function in fitsnap.py after processing configs.

Parameters

allgather – Whether to gather lists on all nodes or just the head node.

create_a()

Allocate shared arrays for calculator.

create_dicts(nconfigs)

Create dictionaries for certain distributed lists. Each list should be of size nconfigs of a single proc.

Parameters

nconfigs – int number of configs on this proc.

3.3.1. LammpsBase

class fitsnap3lib.calculators.lammps_base.LammpsBase(name, pt, config)
create_a()

Allocate shared arrays for calculator.

preprocess_allocate(nconfigs: int)

Allocate arrays to be used by this proc. These arrays have size nconfigs.

Parameters

nconfigs – number of configs on this proc

process_configs(data, i)

Calculate descriptors for a given configuration. Action of this function is altered by certain attributes.

Parameters

transpose_trick – Don’t touch shared arrays in _collect_lammps() if true. Instead store smaller matrices self.aw, self.bw.

process_single(data, i=0)

Calculate descriptors on a single configuration without touching the shraed arrays.

Parameters

  • data – dictionary of structural and fitting info for a configuration in fitsnap data dictionary format.

  • i – integer index which is optional, mainly for debugging purposes.

Returns

  • A matrix of descriptors depending on settings declared in CALCULATOR. If bikflag is 0 (default) then A has 1 and 0s in the first column since it is ready to fit with linear solvers; the descriptors are also divided by no. atoms in this case. If bikflag is 1, then A is simply an unaltered per-atom descriptor matrix.

  • b vector of truths

  • w vector of weights