Different way to generate cubic lattices?

[jgostick]

I just stumbled across the sklearn.neighbors.knearest function. If this function is fast then it could be used to replace the current cubic/fcc/bcc generators, as follows:

import numpy as np
from sklearn.neighbors import kneighbors_graph

Nx = 3
Ny = 5
Nz = 7
i = np.arange(Nx*Ny*Nz)

# Fortran ordering
# X = i % Nx
# Y = (i // Nx) % Ny
# Z = (i // (Nx * Ny))  % Nz

# C/numpy ordering
X = (i // (Ny * Nz))  % Nx
Y = (i // Nx) % Ny
Z = i % Nz
xyz = np.vstack([X, Y, Z]).T

kn = kneighbors_graph(xyz,n_neighbors=6).tocoo()
conns = np.sort(np.vstack((kn.row, kn.col)).T, axis=1)
kn.row, kn.col = conns.T
kn = (kn.tocsr()).tocoo()

[jgostick]

I have just done a quick time comparison, and the above approach is about 200x SLOWER than our current way! The kneighbors_graph function can be parallelized though, so if I use all my cores then it's only 20x slower. Ouch. Not sure where the bottleneck is.

The above is nice and concise, but the performance is just too poor for use in our cubic generator...however, it's still worth adding as a new type of random network generator. Instead of finding xyz on a lattice, we can generate them randomly and find the n nearest neighbors. If n were some high-ish number, the neighbors could then be trimmed to be within some threshold distance.