# -*- coding: utf-8 -*-
import numpy as np
from jax import vmap, jit, numpy as jnp
from functools import partial
from brainpy import math as bm
from brainpy.tools import to_size, size2num
from .base import _IntraLayerInitializer
__all__ = [
'GaussianDecay',
'DOGDecay',
]
@jit
@partial(vmap, in_axes=(0, None, None))
def gaussian_decay_dist_cal1(i_value, post_values, sigma):
dists = jnp.abs(i_value - post_values)
exp_dists = jnp.exp(-(jnp.sqrt(jnp.sum(dists ** 2, axis=0)) / sigma) ** 2 / 2)
return bm.asarray(exp_dists)
@jit
@partial(vmap, in_axes=(0, None, None, None))
def gaussian_decay_dist_cal2(i_value, post_values, value_sizes, sigma):
dists = jnp.abs(i_value - post_values)
dists = jnp.where(dists > (value_sizes / 2), value_sizes - dists, dists)
exp_dists = jnp.exp(-(jnp.sqrt(jnp.sum(dists ** 2, axis=0)) / sigma) ** 2 / 2)
return bm.asarray(exp_dists)
[docs]
class GaussianDecay(_IntraLayerInitializer):
r"""Builds a Gaussian connectivity pattern within a population of neurons,
where the weights decay with gaussian function.
Specifically, for any pair of neurons :math:`(i, j)`, the weight is computed as
.. math::
w(i, j) = w_{max} \cdot \exp(-\frac{\sum_{k=1}^n |v_k^i - v_k^j|^2 }{2\sigma^2})
where :math:`v_k^i` is the $i$-th neuron's encoded value at dimension $k$.
Parameters
----------
sigma : float
Width of the Gaussian function.
max_w : float
The weight amplitude of the Gaussian function.
min_w : float, None
The minimum weight value below which synapses are not created (default: :math:`0.005 * max\_w`).
include_self : bool
Whether create the conn at the same position.
encoding_values : optional, list, tuple, int, float
The value ranges to encode for neurons at each axis.
- If `values` is not provided, the neuron only encodes each positional
information, i.e., :math:`(i, j, k, ...)`, where :math:`i, j, k` is
the index in the high-dimensional space.
- If `values` is a single tuple/list of int/float, neurons at each dimension
will encode the same range of values. For example, `values=(0, np.pi)`,
neurons at each dimension will encode a continuous value space `[0, np.pi]`.
- If `values` is a tuple/list of list/tuple, it means the value space will be
different for each dimension. For example, `values=((-np.pi, np.pi), (10, 20), (0, 2 * np.pi))`.
periodic_boundary : bool
Whether the neuron encode the value space with the periodic boundary.
normalize : bool
Whether normalize the connection probability.
"""
def __init__(self, sigma, max_w, min_w=None, encoding_values=None,
periodic_boundary=False, include_self=True, normalize=False):
super(GaussianDecay, self).__init__()
self.sigma = sigma
self.max_w = max_w
self.min_w = max_w * 0.005 if min_w is None else min_w
self.encoding_values = encoding_values
self.periodic_boundary = periodic_boundary
self.include_self = include_self
self.normalize = normalize
def __call__(self, shape, dtype=None):
"""Build the weights.
Parameters
----------
shape : tuple of int, list of int, int
The network shape. Note, this is not the weight shape.
"""
shape = to_size(shape)
net_size = size2num(shape)
# value ranges to encode
if self.encoding_values is None:
value_ranges = tuple([(0, s) for s in shape])
elif isinstance(self.encoding_values, (tuple, list)):
if len(self.encoding_values) == 0:
raise ValueError
elif isinstance(self.encoding_values[0], (int, float)):
assert len(self.encoding_values) == 2
assert self.encoding_values[0] < self.encoding_values[1]
value_ranges = tuple([self.encoding_values for _ in shape])
elif isinstance(self.encoding_values[0], (tuple, list)):
if len(self.encoding_values) != len(shape):
raise ValueError(f'The network size has {len(shape)} dimensions, while '
f'the encoded values provided only has {len(self.encoding_values)}-D. '
f'Error in {str(self)}.')
for v in self.encoding_values:
assert isinstance(v[0], (int, float))
assert len(v) == 2
value_ranges = tuple(self.encoding_values)
else:
raise ValueError(f'Unsupported encoding values: {self.encoding_values}')
else:
raise ValueError(f'Unsupported encoding values: {self.encoding_values}')
# values
values = [np.linspace(vs[0], vs[1], n + 1)[:n] for vs, n in zip(value_ranges, shape)]
post_values = np.stack([v.flatten() for v in np.meshgrid(*values)])
value_sizes = np.array([v[1] - v[0] for v in value_ranges])
if value_sizes.ndim < post_values.ndim:
value_sizes = np.expand_dims(value_sizes, axis=tuple([i + 1 for i in range(post_values.ndim - 1)]))
# connectivity matrix
i_value_list = np.zeros(shape=(net_size, len(shape), 1))
for i in range(net_size):
list_index = i
# values for node i
i_coordinate = tuple()
for s in shape[:-1]:
i, pos = divmod(i, s)
i_coordinate += (pos,)
i_coordinate += (i,)
i_value = np.array([values[i][c] for i, c in enumerate(i_coordinate)])
if i_value.ndim < post_values.ndim:
i_value = np.expand_dims(i_value, axis=tuple([i + 1 for i in range(post_values.ndim - 1)]))
i_value_list[list_index] = i_value
if self.periodic_boundary:
conn_mat = gaussian_decay_dist_cal2(i_value_list, post_values, value_sizes, self.sigma)
else:
conn_mat = gaussian_decay_dist_cal1(i_value_list, post_values, self.sigma)
if self.normalize:
conn_mat /= conn_mat.max()
if not self.include_self:
bm.fill_diagonal(conn_mat, 0.)
# connectivity weights
conn_mat *= self.max_w
conn_mat = bm.where(conn_mat < self.min_w, 0., conn_mat)
return bm.asarray(conn_mat, dtype=dtype)
def __repr__(self):
name = self.__class__.__name__
bank = ' ' * len(name)
return (f'{name}(sigma={self.sigma}, max_w={self.max_w}, min_w={self.min_w}, \n'
f'{bank}periodic_boundary={self.periodic_boundary}, '
f'include_self={self.include_self}, '
f'normalize={self.normalize})')
@jit
@partial(vmap, in_axes=(0, None, None, None, None, None, None, None))
def _dog_decay_pd(voxel_ids,
values, post_values, value_sizes,
max_w_p, sigma_p,
max_w_n, sigma_n):
i_value = []
for i in range(len(voxel_ids)):
p_id = voxel_ids[i] # position id
i_value.append(values[i][p_id])
i_value = bm.array(i_value)
if i_value.ndim < post_values.ndim:
i_value = bm.expand_dims(i_value, axis=tuple([i + 1 for i in range(post_values.ndim - 1)]))
# distances
dists = bm.abs(i_value - post_values)
dists = bm.where(dists > value_sizes / 2, value_sizes - dists, dists)
dists_exp_p = max_w_p * bm.exp(-(bm.linalg.norm(dists, axis=0) / sigma_p) ** 2 / 2)
dists_exp_n = max_w_n * bm.exp(-(bm.linalg.norm(dists, axis=0) / sigma_n) ** 2 / 2)
return dists_exp_p - dists_exp_n
@jit
@partial(vmap, in_axes=(0, None, None, None, None, None, None))
def _dog_decay(voxel_ids,
values, post_values,
max_w_p, sigma_p,
max_w_n, sigma_n):
i_value = []
for i in range(len(voxel_ids)):
p_id = voxel_ids[i] # position id
i_value.append(values[i][p_id])
i_value = bm.array(i_value)
if i_value.ndim < post_values.ndim:
i_value = bm.expand_dims(i_value, axis=tuple([i + 1 for i in range(post_values.ndim - 1)]))
# distances
dists = bm.abs(i_value - post_values)
dists_exp_p = max_w_p * bm.exp(-(bm.linalg.norm(dists, axis=0) / sigma_p) ** 2 / 2)
dists_exp_n = max_w_n * bm.exp(-(bm.linalg.norm(dists, axis=0) / sigma_n) ** 2 / 2)
return dists_exp_p - dists_exp_n
[docs]
class DOGDecay(_IntraLayerInitializer):
r"""Builds a Difference-Of-Gaussian (dog) connectivity pattern within a population of neurons.
Mathematically, for the given pair of neurons :math:`(i, j)`, the weight between them is computed as
.. math::
w(i, j) = w_{max}^+ \cdot \exp(-\frac{\sum_{k=1}^n |v_k^i - v_k^j|^2}{2\sigma_+^2}) -
w_{max}^- \cdot \exp(-\frac{\sum_{k=1}^n |v_k^i - v_k^j|^2}{2\sigma_-^2})
where weights smaller than :math:`0.005 * max(w_{max}, w_{min})` are not created and
self-connections are avoided by default (parameter allow_self_connections).
Parameters
----------
sigmas : tuple
Widths of the positive and negative Gaussian functions.
max_ws : tuple
The weight amplitudes of the positive and negative Gaussian functions.
min_w : float, None
The minimum weight value below which synapses are not created (default: :math:`0.005 * min(max\_ws)`).
include_self : bool
Whether create the connections at the same position (self-connections).
normalize : bool
Whether normalize the connection probability .
encoding_values : optional, list, tuple, int, float
The value ranges to encode for neurons at each axis.
- If `values` is not provided, the neuron only encodes each positional
information, i.e., :math:`(i, j, k, ...)`, where :math:`i, j, k` is
the index in the high-dimensional space.
- If `values` is a single tuple/list of int/float, neurons at each dimension
will encode the same range of values. For example, `values=(0, np.pi)`,
neurons at each dimension will encode a continuous value space `[0, np.pi]`.
- If `values` is a tuple/list of list/tuple, it means the value space will be
different for each dimension. For example, `values=((-np.pi, np.pi), (10, 20), (0, 2 * np.pi))`.
periodic_boundary : bool
Whether the neuron encode the value space with the periodic boundary.
"""
def __init__(self, sigmas, max_ws, min_w=None, encoding_values=None,
periodic_boundary=False, normalize=True, include_self=True):
super(DOGDecay, self).__init__()
self.sigma_p, self.sigma_n = sigmas
self.max_w_p, self.max_w_n = max_ws
self.min_w = 0.005 * min(self.max_w_p, self.max_w_n) if min_w is None else min_w
self.normalize = normalize
self.include_self = include_self
self.encoding_values = encoding_values
self.periodic_boundary = periodic_boundary
def __call__(self, shape, dtype=None):
"""Build the weights.
Parameters
----------
shape : tuple of int, list of int, int
The network shape. Note, this is not the weight shape.
"""
shape = to_size(shape)
# value ranges to encode
if self.encoding_values is None:
value_ranges = tuple([(0, s) for s in shape])
elif isinstance(self.encoding_values, (tuple, list)):
if len(self.encoding_values) == 0:
raise ValueError
elif isinstance(self.encoding_values[0], (int, float)):
assert len(self.encoding_values) == 2
assert self.encoding_values[0] < self.encoding_values[1]
value_ranges = tuple([self.encoding_values for _ in shape])
elif isinstance(self.encoding_values[0], (tuple, list)):
if len(self.encoding_values) != len(shape):
raise ValueError(f'The network size has {len(shape)} dimensions, while '
f'the encoded values provided only has {len(self.encoding_values)}-D. '
f'Error in {str(self)}.')
for v in self.encoding_values:
assert isinstance(v[0], (int, float))
assert len(v) == 2
value_ranges = tuple(self.encoding_values)
else:
raise ValueError(f'Unsupported encoding values: {self.encoding_values}')
else:
raise ValueError(f'Unsupported encoding values: {self.encoding_values}')
# values
values = [np.linspace(vs[0], vs[1], n + 1)[:n] for vs, n in zip(value_ranges, shape)]
post_values = np.stack([v.flatten() for v in np.meshgrid(*values)])
value_sizes = np.array([v[1] - v[0] for v in value_ranges])
if value_sizes.ndim < post_values.ndim:
value_sizes = np.expand_dims(value_sizes, axis=tuple([i + 1 for i in range(post_values.ndim - 1)]))
voxel_ids = np.meshgrid(*[np.arange(s) for s in shape])
if np.ndim(voxel_ids[0]) > 1:
voxel_ids = tuple(np.moveaxis(m, 0, 1).flatten() for m in voxel_ids)
# connectivity matrix
if self.periodic_boundary:
conn_weights = _dog_decay_pd(voxel_ids, values, post_values, value_sizes,
self.max_w_p, self.sigma_p,
self.max_w_n, self.sigma_n)
else:
conn_weights = _dog_decay(voxel_ids, values, post_values,
self.max_w_p, self.sigma_p,
self.max_w_n, self.sigma_n)
if not self.include_self:
conn_weights = bm.asarray(conn_weights)
bm.fill_diagonal(conn_weights, 0.)
# connectivity weights
conn_weights = bm.where(np.abs(conn_weights) < self.min_w, 0., conn_weights)
return bm.asarray(conn_weights, dtype=dtype)
def __repr__(self):
name = self.__class__.__name__
bank = ' ' * len(name)
return (f'{name}(sigmas={(self.sigma_p, self.sigma_n)}, '
f'max_ws={(self.max_w_p, self.max_w_n)}, min_w={self.min_w}, \n'
f'{bank}periodic_boundary={self.periodic_boundary}, '
f'include_self={self.include_self}, '
f'normalize={self.normalize})')