def _buildTree(ra, dec, leafsize=100, scale=None):
"""
Build KD tree on simDataRA/Dec and set radius (via setRad) for matching.
Parameters
----------
ra, dec : float (or arrays)
RA and Dec values (in radians).
leafsize : int (100)
The number of Ra/Dec pointings in each leaf node.
scale : float (None)
If set, the values are scaled up, rounded, and converted to integers. Useful for
forcing a set precision and preventing machine precision differences
"""
if np.any(np.abs(ra) > np.pi * 2.0) or np.any(np.abs(dec) > np.pi * 2.0):
raise ValueError('Expecting RA and Dec values to be in radians.')
x, y, z = _xyz_from_ra_dec(ra, dec)
if scale is not None:
x = np.round(x * scale).astype(int)
y = np.round(y * scale).astype(int)
z = np.round(z * scale).astype(int)
data = list(zip(x, y, z))
if np.size(data) > 0:
try:
tree = kdTree(data,
leafsize=leafsize,
balanced_tree=False,
compact_nodes=False)
except TypeError:
tree = kdTree(data, leafsize=leafsize)
else:
raise ValueError('ra and dec should have length greater than 0.')
return tree
def _update_kdtree(self):
self.empty = False
#all_particle_coords = self.data
# self.swarm.shadow_particles_fetch()
# dims = self.swarm.particleCoordinates.data.shape[1]
# pc = np.append(self.swarm.particleCoordinates.data,
# self.swarm.particleCoordinates.data_shadow)
# all_particle_coords = pc.reshape(-1,dims)
all_particle_coords = self.data
if len(all_particle_coords) < 3:
self.empty = True
#self.kdtree = lambda x: float('inf')
#trying this instead,
self.kdtree = kdTree(np.empty((2, 0)))
else:
self.kdtree = kdTree(all_particle_coords)
return
def _update_kdtree(self):
self.empty = False
all_particle_coords = self.particleCoordinates
if len(all_particle_coords) < 3:
self.empty = True
self.kdtree = kdTree(np.empty((2,0)))
else:
self.kdtree = kdTree(all_particle_coords)
return
def nn_evaluation(_fromCoords, _toCoords, n=1, weighted=False):
"""
This function provides nearest neighbour information for uw swarms,
given the "_toCoords", which could be the .data handle (coordinates) of a mesh or a different swarm,
this function returns the indices of the n nearest neighbours in "_fromCoords" (will usually be swarm.particleCoordinates.data )
it also returns the inverse-distance weights if weighted=True.
The function works in parallel, if the example below is followed
Usage
------------
#get the n indexes, weights and distances
ix, weights, d = nn_evaluation(swarm.particleCoordinates.data, toSwarm.particleCoordinates.data, n=n, weighted=False)
#apply to the 'toSwarm' variable in a parallel-safe way
if len(weights): #parallel safety
toSwarmVar.data[:,0] = np.average(fromSwarmVar.evaluate(fromSwarm)[:,0][ix], weights=weights, axis=len((weights.shape)) - 1)
"""
if len(_toCoords) > 0: #this is required for safety in parallel
tree = kdTree(_fromCoords)
d, ix = tree.query(_toCoords, n)
if n == 1:
weights = np.ones(_toCoords.shape[0])
elif not weighted:
weights = np.ones((_toCoords.shape[0], n))*(1./n)
else:
weights = (1./d[:])/(1./d[:]).sum(axis=1)[:,None]
return ix, weights, d
else:
return np.empty(0, dtype="int"), np.empty(0, dtype="int"), np.empty(0, dtype="int")
def _buildTree(ra, dec, leafsize=100):
"""
Build KD tree on simDataRA/Dec and set radius (via setRad) for matching.
Parameters
----------
ra, dec = RA and Dec values (in radians).
leafsize = the number of Ra/Dec pointings in each leaf node.
"""
if np.any(np.abs(ra) > np.pi * 2.0) or np.any(np.abs(dec) > np.pi * 2.0):
raise ValueError('Expecting RA and Dec values to be in radians.')
x, y, z = _xyz_from_ra_dec(ra, dec)
data = list(zip(x, y, z))
if np.size(data) > 0:
try:
tree = kdTree(data, leafsize=leafsize, balanced_tree=False, compact_nodes=False)
except TypeError:
tree = kdTree(data, leafsize=leafsize)
else:
raise ValueError('ra and dec should have length greater than 0.')
return tree
def pointToRegionsDistances(p0, points, regions, kd=3):
region_distances = []
for rid in range(len(regions)):
region = regions[rid]
data = np.insert([points[i] for i in region], 0, p0,
axis=0) #prepending the search point to the region
kdtree = kdTree(data)
distances, i = kdtree.query(p0, k=2)
region_distances.append(
distances[1]) # distances[0][0] will be to the point itself
return region_distances
def removeOutliersIndices(points, indices, mode='sor', max=2, k=8):
if mode == 'median':
z_data = points[:, 2]
mask = abs(z_data - np.median(z_data)) < max
return points[mask]
elif mode == 'sor':
kdtree = kdTree(points)
distances, i = kdtree.query(kdtree.data, k=k, n_jobs=-1)
z_distances = stats.zscore(np.mean(distances, axis=1))
sor_filter = abs(z_distances) < max
print("SOR removing: ",
len([bol for bol in sor_filter if bol == False]))
return indices[sor_filter]
else:
print("Undefined outlierremoval mode")
def nn_evaluation(fromSwarm, toSwarm, n=1, weighted=False):
"""
This function provides nearest neighbour information for uw swarms,
given the "toSwarm", this function returns the indices of the n nearest neighbours in "fromSwarm"
it also returns the inverse-distance if weighted=True.
The function works in parallel.
The arrays come out a bit differently when used in nearest neighbour form
(n = 1), or IDW: (n > 1). The examples belowe show how to fill out a swarm variable in each case.
Usage n == 1:
------------
ix, weights = nn_evaluation(swarm, data, n=1, weighted=False)
toSwarmVar.data[:][:,0] = np.average(fromSwarmVar[ix][:,0], weights=weights)
Usage n > 1:
------------
ix, weights = nn_evaluation(swarm, data, n=2, weighted=False)
toSwarmVar.data[:][:,0] = np.average(fromSwarmVar[ix][:,:,0], weights=weights, axis=1)
"""
if len(toSwarm) > 0: #this is required for safety in parallel
#this should avoid building the tree again when this function is called multiple times.
try:
tree = fromSwarm.tree
#print(1)
except:
#print(2)
fromSwarm.tree = kdTree(fromSwarm.particleCoordinates.data)
tree = fromSwarm.tree
d, ix = tree.query(toSwarm, n)
if n == 1:
weights = np.ones(toSwarm.shape[0])
elif not weighted:
weights = np.ones((toSwarm.shape[0], n)) * (1. / n)
else:
weights = (1. / d[:]) / (1. / d[:]).sum(axis=1)[:, None]
return ix, weights
else:
return [], []
def _update_kdtree(self):
self.empty = False
self.swarm.shadow_particles_fetch()
dims = self.swarm.particleCoordinates.data.shape[1]
pc = np.append(self.swarm.particleCoordinates.data,
self.swarm.particleCoordinates.data_shadow)
all_particle_coords = pc.reshape(-1, dims)
if len(all_particle_coords) < 4:
self.empty = True
self.kdtree = lambda x: float('inf')
else:
self.kdtree = kdTree(all_particle_coords)
return
def _update_kdtree(self):
self.empty = False
self.swarm.shadow_particles_fetch()
dims = self.swarm.particleCoordinates.data.shape[1]
pc = np.append(self.swarm.particleCoordinates.data,
self.swarm.particleCoordinates.data_shadow)
all_particle_coords = pc.reshape(-1,dims)
if len(all_particle_coords) < 3:
self.empty = True
self.kdtree = lambda x: float('inf')
else:
self.kdtree = kdTree(all_particle_coords)
return
def _update_kdtree_v2(self):
self.empty = False
self.swarm.shadow_particles_fetch()
if self.swarm.particleLocalCount == 0:
self.empty = True
self.kdtree = lambda x: float('inf')
return
if self.swarm.particleCoordinates.data_shadow.shape[0] == 0:
all_particle_coords = self.swarm.particleCoordinates.data
else:
all_particle_coords = np.concatenate((self.swarm.particleCoordinates.data,
self.swarm.particleCoordinates.data_shadow))
if len(all_particle_coords) < 3:
self.empty = True
self.kdtree = lambda x: float('inf')
else:
self.kdtree = kdTree(all_particle_coords)
return
def _update_surface_normals(self):
"""
Rebuilds the normals for the string of points
"""
# This is the case if there are too few points to
# compute normals so there can be values to remove
all_particle_coords = self.data
#can be important for parallel
#self.swarm.shadow_particles_fetch()
if self.empty:
self.director.data[...] = 0.0
else:
#before looping through particles, we want to set up some things
#local coords
particle_coords = self.swarm.particleCoordinates.data
if isinstance(self.insidePt, interface2D):
#local insidePoints
inside_particle_coords = self.insidePt.swarm.particleCoordinates.data
if inside_particle_coords.shape[0] > 1:
localkdtree = kdTree(inside_particle_coords)
#get the full set of insidepoint NNs
#self.insidePt._update_kdtree() #make sure kdtree is synced
r, nindex = localkdtree.query(particle_coords, k=1)
insidepoints = inside_particle_coords[nindex]
#these will hold the normal vector compenents
Nx = np.empty(self.swarm.particleLocalCount)
Ny = np.empty(self.swarm.particleLocalCount)
for i, xy in enumerate(particle_coords):
r, neighbours = self.kdtree.query(particle_coords[i], k=3)
# neighbour points are neighbours[1] and neighbours[2]
XY1 = all_particle_coords[neighbours[1]]
XY2 = all_particle_coords[neighbours[2]]
#XY1 = self.kdtree.data[neighbours[1]]
#XY2 = self.kdtree.data[neighbours[2]]
dXY = XY2 - XY1
Nx[i] = dXY[1]
Ny[i] = -dXY[0]
#if the inside point is another interface2D object
#use the nearest point to define the orientation
if isinstance(self.insidePt, interface2D):
if inside_particle_coords.shape[0] > 1:
#if True is False:
#only proceed if there are particles in the insidePt
#if not self.insidePt.empty:
#r, nindex = self.insidePt.kdtree.query(particle_coords[i], k=1)
ip = insidepoints[i]
#uw.barrier()
sign = np.sign((ip[0] - xy[0]) * Nx[i] +
(ip[1] - xy[1]) * Ny[i])
#if not self.insidePt.empty:
Nx[i] *= sign
Ny[i] *= sign
#else:
#elif (self.insidePt):
#elif hands(self.insidePt) == 2:
#elif isinstance(self.insidePt, tuple):
elif hasattr(self.insidePt, '__len__'):
#print('shouldt be here test')
sign = np.sign((self.insidePt[0] - xy[0]) * Nx[i] +
(self.insidePt[1] - xy[1]) * Ny[i])
Nx[i] *= sign
Ny[i] *= sign
for i in range(0, self.swarm.particleLocalCount):
scale = 1.0 / np.sqrt(Nx[i]**2 + Ny[i]**2)
Nx[i] *= scale
Ny[i] *= scale
self.director.data[:, 0] = Nx[:]
self.director.data[:, 1] = Ny[:]
return
import numpy as np
from scipy.spatial import cKDTree as kdTree
np.random.seed(42)
# Make a square of random points
npts = int(1e7)
x = np.random.random(npts) - 0.5
y = np.random.random(npts) - 0.5
data = list(zip(x, y))
tree = kdTree(data)
count = np.size(tree.query_ball_point([0, 0], 0.2))
print('count = %i' % count)