# Copyright (c) 2017 The University of Manchester
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import numpy
from pyNN.standardmodels.synapses import StaticSynapse
from spinn_utilities.overrides import overrides
from spynnaker.pyNN.exceptions import SynapticConfigurationException
from spynnaker.pyNN.utilities.utility_calls import create_mars_kiss_seeds
from .abstract_synapse_dynamics_structural import (
AbstractSynapseDynamicsStructural)
from .synapse_dynamics_stdp import SynapseDynamicsSTDP
from .synapse_dynamics_structural_common import (
DEFAULT_F_REW, DEFAULT_INITIAL_WEIGHT, DEFAULT_INITIAL_DELAY,
DEFAULT_S_MAX, SynapseDynamicsStructuralCommon)
from .synapse_dynamics_neuromodulation import SynapseDynamicsNeuromodulation
from spynnaker.pyNN.utilities.constants import SPIKE_PARTITION_ID
class SynapseDynamicsStructuralSTDP(
SynapseDynamicsSTDP, SynapseDynamicsStructuralCommon):
"""
Class that enables synaptic rewiring in the presence of STDP.
It acts as a wrapper around SynapseDynamicsSTDP, meaning rewiring can
operate in parallel with STDP synapses.
Written by Petrut Bogdan.
"""
__slots__ = [
# Frequency of rewiring (Hz)
"__f_rew",
# Initial weight assigned to a newly formed connection
"__initial_weight",
# Delay assigned to a newly formed connection
"__initial_delay",
# Maximum fan-in per target layer neuron
"__s_max",
# The seed
"__seed",
# Holds initial connectivity as defined via connector
"__connections",
# The actual type of weights: static through the simulation or those
# that can be change through STDP
"__weight_dynamics",
# Shared RNG seed to be written on all cores
"__seeds",
# The RNG used with the seed that is passed in
"__rng",
# The partner selection rule
"__partner_selection",
# The formation rule
"__formation",
# The elimination rule
"__elimination"
]
def __init__(
self, partner_selection, formation, elimination,
timing_dependence=None, weight_dependence=None,
voltage_dependence=None, dendritic_delay_fraction=1.0,
f_rew=DEFAULT_F_REW, initial_weight=DEFAULT_INITIAL_WEIGHT,
initial_delay=DEFAULT_INITIAL_DELAY, s_max=DEFAULT_S_MAX,
with_replacement=True, seed=None,
weight=StaticSynapse.default_parameters['weight'], delay=None,
backprop_delay=True):
"""
:param AbstractPartnerSelection partner_selection:
The partner selection rule
:param AbstractFormation formation: The formation rule
:param AbstractElimination elimination: The elimination rule
:param AbstractTimingDependence timing_dependence:
The STDP timing dependence rule
:param AbstractWeightDependence weight_dependence:
The STDP weight dependence rule
:param None voltage_dependence:
The STDP voltage dependence (unsupported)
:param float dendritic_delay_fraction:
The STDP dendritic delay fraction
:param float f_rew:
How many rewiring attempts will be done per second.
:param float initial_weight:
Weight assigned to a newly formed connection
:param initial_delay:
Delay assigned to a newly formed connection; a single value means
a fixed delay value, or a tuple of two values means the delay will
be chosen at random from a uniform distribution between the given
values
:type initial_delay: float or tuple(float, float)
:param int s_max: Maximum fan-in per target layer neuron
:param bool with_replacement:
If set to True (default), a new synapse can be formed in a
location where a connection already exists; if False, then it must
form where no connection already exists
:param seed: seed for the random number generators
:type seed: int or None
:param float weight: The weight of connections formed by the connector
:param delay: The delay of connections formed by the connector
Use ``None`` to get the simulator default minimum delay.
:type delay: float or None
:param bool backprop_delay: Whether back-propagated delays are used
"""
super().__init__(
timing_dependence, weight_dependence, voltage_dependence,
dendritic_delay_fraction, weight, delay, pad_to_length=s_max,
backprop_delay=backprop_delay)
self.__partner_selection = partner_selection
self.__formation = formation
self.__elimination = elimination
self.__f_rew = float(f_rew)
self.__initial_weight = initial_weight
self.__initial_delay = initial_delay
self.__s_max = s_max
self.__with_replacement = with_replacement
self.__seed = seed
self.__connections = dict()
self.__rng = numpy.random.RandomState(seed)
self.__seeds = dict()
[docs]
@overrides(SynapseDynamicsSTDP.merge)
def merge(self, synapse_dynamics):
# If dynamics is Neuromodulation, merge with other neuromodulation,
# and then return ourselves, as neuromodulation can't be used by
# itself
if isinstance(synapse_dynamics, SynapseDynamicsNeuromodulation):
super().merge_neuromodulation(synapse_dynamics)
return self
# If other is structural, check structural matches
if isinstance(synapse_dynamics, AbstractSynapseDynamicsStructural):
if not SynapseDynamicsStructuralCommon.is_same_as(
self, synapse_dynamics):
raise SynapticConfigurationException(
"Synapse dynamics must match exactly when using multiple"
" edges to the same population")
# If other is STDP, check STDP matches
if isinstance(synapse_dynamics, SynapseDynamicsSTDP):
if not SynapseDynamicsSTDP.is_same_as(self, synapse_dynamics):
raise SynapticConfigurationException(
"Synapse dynamics must match exactly when using multiple"
" edges to the same population")
# If everything matches, return ourselves as supreme!
return self
[docs]
def set_projection_parameter(self, param, value):
"""
:param str param:
:param value:
"""
for item in (self.partner_selection, self.__formation,
self.__elimination):
if hasattr(item, param):
setattr(item, param, value)
break
else:
raise ValueError(f"Unknown parameter {param}")
[docs]
@overrides(SynapseDynamicsSTDP.is_same_as)
def is_same_as(self, synapse_dynamics):
if (isinstance(synapse_dynamics, SynapseDynamicsSTDP) and
not super().is_same_as(synapse_dynamics)):
return False
return SynapseDynamicsStructuralCommon.is_same_as(
self, synapse_dynamics)
[docs]
@overrides(SynapseDynamicsSTDP.get_vertex_executable_suffix)
def get_vertex_executable_suffix(self):
return (super().get_vertex_executable_suffix() +
SynapseDynamicsStructuralCommon.get_vertex_executable_suffix(
self))
[docs]
@overrides(AbstractSynapseDynamicsStructural.set_connections)
def set_connections(
self, connections, post_vertex_slice, app_edge, synapse_info):
if not isinstance(synapse_info.synapse_dynamics,
AbstractSynapseDynamicsStructural):
return
collector = self.__connections.setdefault(
(app_edge.post_vertex, post_vertex_slice.lo_atom), [])
collector.append((connections, app_edge, synapse_info))
[docs]
@overrides(SynapseDynamicsSTDP.get_parameter_names)
def get_parameter_names(self):
names = super().get_parameter_names()
names.extend(SynapseDynamicsStructuralCommon.get_parameter_names(self))
return names
@property
@overrides(AbstractSynapseDynamicsStructural.f_rew)
def f_rew(self):
return self.__f_rew
@property
@overrides(AbstractSynapseDynamicsStructural.s_max)
def s_max(self):
return self.__s_max
@property
@overrides(AbstractSynapseDynamicsStructural.with_replacement)
def with_replacement(self):
return self.__with_replacement
@property
@overrides(AbstractSynapseDynamicsStructural.seed)
def seed(self):
return self.__seed
@property
@overrides(AbstractSynapseDynamicsStructural.initial_weight)
def initial_weight(self):
return self.__initial_weight
@property
@overrides(AbstractSynapseDynamicsStructural.initial_delay)
def initial_delay(self):
return self.__initial_delay
@property
@overrides(AbstractSynapseDynamicsStructural.partner_selection)
def partner_selection(self):
return self.__partner_selection
@property
@overrides(AbstractSynapseDynamicsStructural.formation)
def formation(self):
return self.__formation
@property
@overrides(AbstractSynapseDynamicsStructural.elimination)
def elimination(self):
return self.__elimination
@property
@overrides(SynapseDynamicsStructuralCommon.connections)
def connections(self):
return self.__connections
[docs]
@overrides(SynapseDynamicsSTDP.get_weight_mean)
def get_weight_mean(self, connector, synapse_info):
return self.get_weight_maximum(connector, synapse_info)
[docs]
@overrides(SynapseDynamicsSTDP.get_weight_maximum)
def get_weight_maximum(self, connector, synapse_info):
w_max = super().get_weight_maximum(connector, synapse_info)
return max(w_max, self.__initial_weight)
[docs]
@overrides(SynapseDynamicsSTDP.get_delay_maximum)
def get_delay_maximum(self, connector, synapse_info):
d_m = super().get_delay_maximum(connector, synapse_info)
return max(d_m, self.__initial_delay)
[docs]
@overrides(SynapseDynamicsSTDP.get_delay_minimum)
def get_delay_minimum(self, connector, synapse_info):
d_m = super().get_delay_minimum(connector, synapse_info)
return min(d_m, self.__initial_delay)
[docs]
@overrides(SynapseDynamicsSTDP.get_delay_variance)
def get_delay_variance(self, connector, delays, synapse_info):
return 0.0
[docs]
@overrides(SynapseDynamicsStructuralCommon.get_seeds)
def get_seeds(self, app_vertex=None):
if app_vertex:
if app_vertex not in self.__seeds.keys():
self.__seeds[app_vertex] = (
create_mars_kiss_seeds(self.__rng))
return self.__seeds[app_vertex]
else:
return create_mars_kiss_seeds(self.__rng)
[docs]
@overrides(SynapseDynamicsSTDP.generate_on_machine)
def generate_on_machine(self):
# Never generate structural connections on the machine
return False
[docs]
@overrides(SynapseDynamicsSTDP.get_connected_vertices)
def get_connected_vertices(self, s_info, source_vertex, target_vertex):
# Things change, so assume all connected
return [(m_vertex, [source_vertex])
for m_vertex in target_vertex.splitter.get_in_coming_vertices(
SPIKE_PARTITION_ID)]
@property
@overrides(SynapseDynamicsSTDP.is_combined_core_capable)
def is_combined_core_capable(self):
return False