# encoding: utf-8
"""
Tools for performing spatial/topographical calculations.
Classes:
Space - representation of a Cartesian space for use in calculating
distances
Line - represents a structure with neurons distributed evenly on a
straight line.
Grid2D - represents a structure with neurons distributed on a 2D grid.
Grid3D - represents a structure with neurons distributed on a 3D grid.
RandomStructure - represents a structure with neurons distributed randomly
within a given volume.
Cuboid - representation of a cuboidal volume, for use with RandomStructure.
Sphere - representation of a spherical volume, for use with RandomStructure.
:copyright: Copyright 2006-2023 by the PyNN team, see AUTHORS.
:license: CeCILL, see LICENSE for details.
"""
# There must be some Python package out there that provides most of this stuff.
# Distance computations are provided by scipy.spatial, but scipy is a fairly heavy dependency.
from functools import reduce
import math
from operator import and_
import logging
import numpy as np
from .random import NumpyRNG
from . import descriptions
logger = logging.getLogger("PyNN")
def distance(src, tgt, mask=None, scale_factor=1.0, offset=0.0,
periodic_boundaries=None): # may need to add an offset parameter
"""
Return the Euclidian distance between two cells.
`mask` allows only certain dimensions to be considered, e.g.::
* to ignore the z-dimension, use `mask=array([0,1])`
* to ignore y, `mask=array([0,2])`
* to just consider z-distance, `mask=array([2])`
`scale_factor` allows for different units in the pre- and post- position
(the post-synaptic position is multipied by this quantity).
"""
d = src.position - scale_factor * (tgt.position + offset)
if periodic_boundaries is not None:
d = np.minimum(abs(d), periodic_boundaries - abs(d))
if mask is not None:
d = d[mask]
return np.sqrt(np.dot(d, d))
[docs]class Space(object):
"""
Class representing a space within distances can be calculated. The space
is Cartesian, may be 1-, 2- or 3-dimensional, and may have periodic
boundaries in any of the dimensions.
Arguments:
axes:
if not supplied, then the 3D distance is calculated. If supplied,
axes should be a string containing the axes to be used, e.g. 'x', or
'yz'. axes='xyz' is the same as axes=None.
scale_factor:
it may be that the pre and post populations use
different units for position, e.g. degrees and µm. In this case,
`scale_factor` can be specified, which is applied to the positions
in the post-synaptic population.
offset:
if the origins of the coordinate systems of the pre- and post-
synaptic populations are different, `offset` can be used to adjust
for this difference. The offset is applied before any scaling.
periodic_boundaries:
either `None`, or a tuple giving the boundaries for each dimension,
e.g. `((x_min, x_max), None, (z_min, z_max))`.
"""
AXES = {'x': [0], 'y': [1], 'z': [2],
'xy': [0, 1], 'yz': [1, 2], 'xz': [0, 2], 'xyz': range(3), None: range(3)}
def __init__(self, axes=None, scale_factor=1.0, offset=0.0,
periodic_boundaries=None):
"""
"""
self.periodic_boundaries = periodic_boundaries
self.axes = np.array(Space.AXES[axes])
self.scale_factor = scale_factor
self.offset = offset
[docs] def distances(self, A, B, expand=False):
"""
Calculate the distance matrix between two sets of coordinates, given
the topology of the current space.
From http://projects.scipy.org/pipermail/numpy-discussion/2007-April/027203.html
"""
assert A.ndim <= 2
assert B.ndim <= 2
assert A.shape[-1] == 3
if len(A.shape) == 1:
A = A.reshape(1, 3)
if len(B.shape) == 1:
B = B.reshape(1, 3)
B = self.scale_factor * (B + self.offset)
d = np.zeros((len(self.axes), A.shape[0], B.shape[0]), dtype=A.dtype)
for i, axis in enumerate(self.axes):
diff2 = A[:, None, axis] - B[:, axis]
if self.periodic_boundaries is not None:
boundaries = self.periodic_boundaries[axis]
if boundaries is not None:
range = boundaries[1] - boundaries[0]
ad2 = abs(diff2)
diff2 = np.minimum(ad2, range - ad2)
diff2 **= 2
d[i] = diff2
if not expand:
d = np.sum(d, 0)
np.sqrt(d, d)
return d.flatten()
[docs] def distance_generator(self, f, g):
def distance_map(i, j):
shape = []
if isinstance(i, np.ndarray) and i.ndim == 2:
i = i[:, 0]
shape.append(i.size)
if isinstance(j, np.ndarray) and j.ndim == 2:
j = j[0, :]
shape.append(j.size)
d = self.distances(f(i), g(j))
if shape:
return d.reshape(shape)
else:
return d
return distance_map
[docs]class BaseStructure(object):
def __repr__(self):
return "%s(%s)" % (self.__class__.__name__,
", ".join("%s=%r" % item for item in self.get_parameters().items()))
def __eq__(self, other):
return reduce(and_, (getattr(self, attr) == getattr(other, attr)
for attr in self.parameter_names))
[docs] def get_parameters(self):
"""Return a dict containing the parameters of the :class:`Structure`."""
P = {}
for name in self.parameter_names:
P[name] = getattr(self, name)
return P
[docs] def describe(self, template='structure_default.txt', engine='default'):
"""
Returns a human-readable description of the network structure.
The output may be customized by specifying a different template
togther with an associated template engine (see ``pyNN.descriptions``).
If template is None, then a dictionary containing the template context
will be returned.
"""
context = {'name': self.__class__.__name__,
'parameters': self.get_parameters()}
return descriptions.render(engine, template, context)
[docs] def generate_positions(self, n):
"""
Calculate and return the positions of `n` neurons positioned according
to this structure.
"""
raise NotImplementedError
[docs]class Line(BaseStructure):
"""
Represents a structure with neurons distributed evenly on a straight line.
Arguments:
`dx`:
distance between points in the line.
`y`, `z`,:
y- and z-coordinates of all points in the line.
`x0`:
x-coordinate of the first point in the line.
"""
parameter_names = ("dx", "x0", "y", "z")
def __init__(self, dx=1.0, x0=0.0, y=0.0, z=0.0):
self.dx = dx
self.x0 = x0
self.y = y
self.z = z
[docs] def generate_positions(self, n):
x = self.dx * np.arange(n, dtype=float) + self.x0
y = np.zeros(n) + self.y
z = np.zeros(n) + self.z
return np.array((x, y, z))
generate_positions.__doc__ = BaseStructure.generate_positions.__doc__
[docs]class Grid2D(BaseStructure):
"""
Represents a structure with neurons distributed on a 2D grid.
Arguments:
`dx`, `dy`:
distances between points in the x, y directions.
`x0`, `y0`:
coordinates of the starting corner of the grid.
`z`:
the z-coordinate of all points in the grid.
`aspect_ratio`:
ratio of the number of grid points per side (not the ratio of the
side lengths, unless ``dx == dy``)
`fill_order`:
may be 'sequential' or 'random'
"""
parameter_names = ("aspect_ratio", "dx", "dy", "x0", "y0", "z", "fill_order")
def __init__(self, aspect_ratio=1.0, dx=1.0, dy=1.0, x0=0.0, y0=0.0, z=0,
fill_order="sequential", rng=None):
self.aspect_ratio = aspect_ratio
assert fill_order in ('sequential', 'random')
self.fill_order = fill_order
self.rng = rng
self.dx = dx
self.dy = dy
self.x0 = x0
self.y0 = y0
self.z = z
[docs] def calculate_size(self, n):
"""docstring goes here"""
nx = math.sqrt(n * self.aspect_ratio)
if not math.isclose(n % nx, 0, rel_tol=1e-12, abs_tol=1e-12):
raise Exception(f"Invalid size: n={n}, nx={nx}")
nx = int(round(nx))
ny = n // nx
return nx, ny
[docs] def generate_positions(self, n):
nx, ny = self.calculate_size(n)
x, y, z = np.indices((nx, ny, 1), dtype=float)
x = self.x0 + self.dx * x.flatten()
y = self.y0 + self.dy * y.flatten()
z = self.z + z.flatten()
# use column_stack, if we decide to switch from (3,n) to (n,3)
positions = np.array((x, y, z))
if self.fill_order == 'sequential':
return positions
else: # random
if self.rng is None:
self.rng = NumpyRNG()
return self.rng.permutation(positions.T).T
generate_positions.__doc__ = BaseStructure.generate_positions.__doc__
[docs]class Grid3D(BaseStructure):
"""
Represents a structure with neurons distributed on a 3D grid.
Arguments:
`dx`, `dy`, `dz`:
distances between points in the x, y, z directions.
`x0`, `y0`. `z0`:
coordinates of the starting corner of the grid.
`aspect_ratioXY`, `aspect_ratioXZ`:
ratios of the number of grid points per side (not the ratio of the
side lengths, unless ``dx == dy == dz``)
`fill_order`:
may be 'sequential' or 'random'.
If `fill_order` is 'sequential', the z-index will be filled first, then y
then x, i.e. the first cell will be at (0,0,0) (given default values for
the other arguments), the second at (0,0,1), etc.
"""
parameter_names = ("aspect_ratios", "dx", "dy", "dz", "x0", "y0", "z0", "fill_order")
def __init__(self, aspect_ratioXY=1.0, aspect_ratioXZ=1.0, dx=1.0, dy=1.0,
dz=1.0, x0=0.0, y0=0.0, z0=0, fill_order="sequential", rng=None):
self.aspect_ratios = (aspect_ratioXY, aspect_ratioXZ)
assert fill_order in ('sequential', 'random')
self.fill_order = fill_order
self.rng = rng
self.dx = dx
self.dy = dy
self.dz = dz
self.x0 = x0
self.y0 = y0
self.z0 = z0
[docs] def calculate_size(self, n):
"""docstring goes here"""
a, b = self.aspect_ratios
nx = int(round(math.pow(n * a * b, 1 / 3.0)))
ny = int(round(nx / a))
nz = int(round(nx / b))
assert nx * ny * nz == n, str((nx, ny, nz, nx * ny * nz, n, a, b))
return nx, ny, nz
[docs] def generate_positions(self, n):
nx, ny, nz = self.calculate_size(n)
x, y, z = np.indices((nx, ny, nz), dtype=float)
x = self.x0 + self.dx * x.flatten()
y = self.y0 + self.dy * y.flatten()
z = self.z0 + self.dz * z.flatten()
positions = np.array((x, y, z))
if self.fill_order == 'sequential':
return positions
else:
if self.rng is None:
self.rng = NumpyRNG()
return self.rng.permutation(positions.T).T
generate_positions.__doc__ = BaseStructure.generate_positions.__doc__
class Shape(object):
pass
[docs]class Cuboid(Shape):
"""
Represents a cuboidal volume within which neurons may be distributed.
Arguments:
height:
extent in y direction
width:
extent in x direction
depth:
extent in z direction
"""
def __init__(self, width, height, depth):
self.height = height
self.width = width
self.depth = depth
def __repr__(self):
return "Cuboid(width=%r, height=%r, depth=%r)" % (self.width, self.height, self.depth)
[docs] def sample(self, n, rng):
"""Return `n` points distributed randomly with uniform density within the cuboid."""
return 0.5 * rng.uniform(-1, 1, size=(n, 3)) * (self.width, self.height, self.depth)
[docs]class Sphere(Shape):
"""
Represents a spherical volume within which neurons may be distributed.
"""
def __init__(self, radius):
Shape.__init__(self)
self.radius = radius
def __repr__(self):
return "Sphere(radius=%r)" % self.radius
[docs] def sample(self, n, rng):
"""Return `n` points distributed randomly with uniform density within the sphere."""
# this implementation is wasteful, as it throws away a lot of numbers,
# but simple. More efficient implementations welcome.
positions = np.empty((n, 3))
i = 0
while i < n:
candidate = rng.uniform(-1, 1, size=(1, 3))
if (candidate**2).sum() < 1:
positions[i] = candidate
i += 1
return self.radius * positions
[docs]class RandomStructure(BaseStructure):
"""
Represents a structure with neurons distributed randomly within a given
volume.
Arguments:
`boundary` - a subclass of :class:`Shape`.
`origin` - the coordinates (x,y,z) of the centre of the volume.
"""
parameter_names = ('boundary', 'origin', 'rng')
def __init__(self, boundary, origin=(0.0, 0.0, 0.0), rng=None):
assert isinstance(boundary, Shape)
assert len(origin) == 3
self.boundary = boundary
self.origin = origin
self.rng = rng or NumpyRNG()
[docs] def generate_positions(self, n):
return (np.array(self.origin) + self.boundary.sample(n, self.rng)).T
generate_positions.__doc__ = BaseStructure.generate_positions.__doc__
# what about rotations?
