def build_lems_for_model(src):
model = lems.Model()
model.add(lems.Dimension('time', t=1))
# model.add(lems.Dimension('au'))
# primary element of the model is a mass model component
mass = lems.ComponentType(src.name, extends="baseCellMembPot")
model.add(mass)
######### Adding v is required to ease mapping to NEURON...
mass.dynamics.add(lems.StateVariable(name="v", dimension="voltage", exposure="v"))
mass.add(lems.Attachments(name="synapses",type_="basePointCurrentDL"))
for input in src.input:
mass.dynamics.add(lems.DerivedVariable(name=input,
dimension='none',
exposure=input,
select='synapses[*]/I',
reduce='add'))
mass.add(lems.Exposure(input, 'none'))
for key, val in src.const.items():
mass.add(lems.Parameter(key, 'none')) # TODO units
mass.add(lems.Constant(name="MSEC", dimension="time", value="1ms"))
mass.add(lems.Constant(name="PI", dimension="none", value="3.14159265359"))
states = []
der_vars = []
# for key in src.param:
# mass.add(lems.Parameter(key, 'au')) # TODO units
for key, val in src.auxex:
val = val.replace('**', '^')
mass.dynamics.add(lems.DerivedVariable(key, value=val))
for key in src.obsrv:
name_dv = key.replace('(','_').replace(')','').replace(' - ','_min_')
mass.dynamics.add(lems.DerivedVariable(name_dv, value=key, exposure=name_dv))
mass.add(lems.Exposure(name_dv, 'none'))
for src_svar in src.state_space:
name = src_svar.name
ddt = src_svar.drift.replace('**', '^')
mass.dynamics.add(lems.StateVariable(name, 'none', name))
mass.dynamics.add(lems.TimeDerivative(name, '(%s)/MSEC'%ddt))
mass.add(lems.Exposure(name, 'none'))
''' On condition is not need on the model but NeuroML requires its definition -->
<OnCondition test="r .lt. 0">
<EventOut port="spike"/>
</OnCondition>'''
oc = lems.OnCondition(test='v .gt. 0')
oc.actions.append(lems.EventOut(port='spike'))
mass.dynamics.add(oc)
return model
model.add(lems.Dimension('voltage', m=1, l=3, t=-3, i=-1))
model.add(lems.Dimension('time', t=1))
model.add(lems.Dimension('capacitance', m=-1, l=-2, t=4, i=2))
model.add(lems.Unit('milliVolt', 'mV', 'voltage', -3))
model.add(lems.Unit('milliSecond', 'ms', 'time', -3))
model.add(lems.Unit('microFarad', 'uF', 'capacitance', -12))
iaf1 = lems.ComponentType('iaf1')
model.add(iaf1)
iaf1.add(lems.Parameter('threshold', 'voltage'))
iaf1.add(lems.Parameter('reset', 'voltage'))
iaf1.add(lems.Parameter('refractoryPeriod', 'time'))
iaf1.add(lems.Parameter('capacitance', 'capacitance'))
iaf1.add(lems.Exposure('vexp', 'voltage'))
dp = lems.DerivedParameter('range', 'threshold - reset', 'voltage')
iaf1.add(dp)
iaf1.dynamics.add(lems.StateVariable('v','voltage', 'vexp'))
iaf1.dynamics.add(lems.DerivedVariable('v2',dimension='voltage', value='v*2'))
cdv = lems.ConditionalDerivedVariable('v_abs','voltage')
cdv.add(lems.Case('v .geq. 0','v'))
cdv.add(lems.Case('v .lt. 0','-1*v'))
iaf1.dynamics.add(cdv)
model.add(lems.Component('celltype_a', iaf1.name))
model.add(lems.Component('celltype_b', iaf1.name, threshold="20mV"))
fn = '/tmp/model.xml'
def add_neurongroup(self, neurongrp, index_neurongrp, initializers):
"""
Add NeuronGroup to self._model
If number of elements is 1 it adds component of that type, if
it's bigger, the network is created by calling:
`make_multiinstantiate`.
Parameters
----------
neurongrp : dict
Standard dictionary representation of NeuronGroup object
index_neurongrp : int
Index of neurongroup in the network
initializers : list
List of initializers defined in the network
"""
# get name of the neurongrp
component_name = neurongrp['name']
self._model_namespace["neuronname"] = component_name
# add BASE_CELL component
self._component_type = lems.ComponentType(component_name,
extends=BASE_CELL)
# get identifiers attached to neurongrp and create special_properties dict
identifiers = []
special_properties = {}
if 'identifiers' in neurongrp:
identifiers = neurongrp['identifiers']
for initializer in initializers:
if 'identifiers' in initializer:
identifiers.update(initializer['identifiers'])
for identifier in identifiers.keys():
special_properties.update({identifier: None})
# add the identifers as properties of the component
for param in self._determine_properties(identifiers):
self._component_type.add(param)
# common things for every neuron definition
# TODO: Is this same for custom events too?
self._component_type.add(lems.EventPort(name='spike', direction='out'))
# dynamics of the network
dynamics = lems.Dynamics()
# get neurongrp equations
equations = neurongrp['equations']
# loop over the variables
initializer_vars = []
for initializer in initializers:
if initializer['source'] == neurongrp['name']:
initializer_vars.append(initializer['variable'])
for var in equations.keys():
# determine the dimension
dimension = _determine_dimension(equations[var]['unit'])
# add to all_params_unit
self._all_params_unit[var] = dimension
# identify diff eqns to add Exposure
if equations[var]['type'] == 'differential equation':
state_var = lems.StateVariable(var,
dimension=dimension,
exposure=var)
self._component_type.add(
lems.Exposure(var, dimension=dimension))
dynamics.add_state_variable(state_var)
else:
if var in initializer_vars and 'i' in initializer['value']:
self._component_type.add(lems.Property(var, dimension))
special_properties[var] = initializer['value']
continue
state_var = lems.StateVariable(var, dimension=dimension)
dynamics.add_state_variable(state_var)
# what happens at initialization
onstart = lems.OnStart()
# loop over the initializers
for var in equations.keys():
if var in (NOT_REFRACTORY, LAST_SPIKE):
continue
# check the initializer is connected to this neurongrp
if var not in initializer_vars:
continue
if var in special_properties:
continue
for initializer in initializers:
if initializer['variable'] == var and initializer[
'source'] == neurongrp['name']:
init_value = initializer['value']
if type(init_value) != str:
value = brian_unit_to_lems(init_value)
else:
value = renderer.render_expr(str(init_value))
# add to onstart
onstart.add(lems.StateAssignment(var, value))
dynamics.add(onstart)
# check whether refractoriness is defined
if ('events' in neurongrp and 'spike' in neurongrp['events']
and 'refractory' in neurongrp['events']['spike']):
# if refractoriness, we create separate regimes for integrating
integr_regime = lems.Regime(INTEGRATING, dynamics,
True) # True -> initial regime
# check spike event
# NOTE: isn't refractory only for spike events?
for spike_flag, on_cond in self._event_builder(
neurongrp['events']):
if spike_flag:
# if spike occured we make transition to refractory regime
on_cond.add_action(lems.Transition(REFRACTORY))
integr_regime.add_event_handler(on_cond)
# add refractory regime
refrac_regime = lems.Regime(REFRACTORY, dynamics)
# make lastspike variable and initialize it
refrac_regime.add_state_variable(
lems.StateVariable(LAST_SPIKE, dimension='time'))
oe = lems.OnEntry()
oe.add(lems.StateAssignment(LAST_SPIKE, 't'))
refrac_regime.add(oe)
# after refractory time we make transition to integrating regime
refractory_code = neurongrp['events']['spike']['refractory']
if not _equation_separator(str(refractory_code)):
# if there is no specific variable given, we assume
# that this is time condition
ref_oc = lems.OnCondition('t .gt. ( {0} + {1} )'.format(
LAST_SPIKE, brian_unit_to_lems(refractory_code)))
else:
ref_oc = lems.OnCondition(
renderer.render_expr(refractory_code))
ref_trans = lems.Transition(INTEGRATING)
ref_oc.add_action(ref_trans)
refrac_regime.add_event_handler(ref_oc)
# identify variables with differential equation
for var in neurongrp['equations']:
if neurongrp['equations'][var][
'type'] == 'differential equation':
diff_var = neurongrp['equations'][var]
# get the expression
td = lems.TimeDerivative(
var, renderer.render_expr(diff_var['expr']))
# check flags for UNLESS_REFRACTORY TODO: is this available in 'flags' key?
if 'flags' in diff_var:
# if unless refratory we add only do integration regime
if UNLESS_REFRACTORY in diff_var['flags']:
integr_regime.add_time_derivative(td)
continue
# add time derivative to both regimes
integr_regime.add_time_derivative(td)
refrac_regime.add_time_derivative(td)
# add the regimes to dynamics
dynamics.add_regime(integr_regime)
dynamics.add_regime(refrac_regime)
else:
# adding events directly to dynamics
for spike_flag, on_cond in self._event_builder(
neurongrp['events']):
dynamics.add_event_handler(on_cond)
# get variables with diff eqns
for var in neurongrp['equations']:
if neurongrp['equations'][var][
'type'] == 'differential equation':
diff_var = neurongrp['equations'][var]
td = lems.TimeDerivative(
var, renderer.render_expr(diff_var['expr']))
# add to dynamics
dynamics.add_time_derivative(td)
# add dynamics to _component_type
self._component_type.dynamics = dynamics
# add _component_type to _model
self._model.add_component_type(self._component_type)
# get identifiers
paramdict = dict()
for ident_name, ident_value in neurongrp['identifiers'].items():
paramdict[ident_name] = self._unit_lems_validator(ident_value)
# if more than one neuron use multiinstantiate
if neurongrp['N'] == 1:
self._model.add(
lems.Component("n{}".format(index_neurongrp), component_name,
**paramdict))
else:
self.make_multiinstantiate(special_properties, component_name,
paramdict, neurongrp['N'])
description="Prandtl Number"))
lorenz.add(
lems.Parameter(name="beta",
dimension="none",
description="Also named b elsewhere"))
lorenz.add(
lems.Parameter(
name="rho",
dimension="none",
description="Related to the Rayleigh number, also named r elsewhere"))
lorenz.add(lems.Parameter(name="x0", dimension="none"))
lorenz.add(lems.Parameter(name="y0", dimension="none"))
lorenz.add(lems.Parameter(name="z0", dimension="none"))
lorenz.add(lems.Exposure(name="x", dimension="none"))
lorenz.add(lems.Exposure(name="y", dimension="none"))
lorenz.add(lems.Exposure(name="z", dimension="none"))
lorenz.add(lems.Constant(name="sec", value="1s", dimension="time"))
lorenz.dynamics.add(
lems.StateVariable(name="x", dimension="none", exposure="x"))
lorenz.dynamics.add(
lems.StateVariable(name="y", dimension="none", exposure="y"))
lorenz.dynamics.add(
lems.StateVariable(name="z", dimension="none", exposure="z"))
lorenz.dynamics.add(
lems.TimeDerivative(variable="x", value="( sigma * (y - x)) / sec"))
lorenz.dynamics.add(
def add_neurongroup(self, obj, idx_of_ng, namespace, initializers):
"""
Adds NeuronGroup object *obj* to self._model.
If number of elements is 1 it adds component of that type, if
it's bigger, the network is created by calling:
`make_multiinstantiate`.
Parameters
----------
obj : brian2.NeuronGroup
NeuronGroup object to parse
idx_of_ng : int
index of neurongroup
namespace : dict
dictionary with all neccassary variables definition
initializers : dict
initial values for all model variables
"""
if hasattr(obj, "namespace") and not obj.namespace:
obj.namespace = namespace
self._nr_of_neurons = obj.N # maybe not the most robust solution
ct_name = "neuron{}".format(idx_of_ng+1)
self._model_namespace["neuronname"] = ct_name
self._component_type = lems.ComponentType(ct_name, extends=BASE_CELL)
# adding parameters
special_properties = {}
for key in obj.namespace.keys():
special_properties[key] = None
for param in self._determine_properties(obj.namespace):
self._component_type.add(param)
# common things for every neuron definition
self._component_type.add(lems.EventPort(name='spike', direction='out'))
# dynamics of the network
dynamics = lems.Dynamics()
ng_equations = obj.user_equations
for var in ng_equations:
if ng_equations[var].type == DIFFERENTIAL_EQUATION:
dim_ = _determine_dimension(ng_equations[var].unit)
sv_ = lems.StateVariable(var, dimension=dim_, exposure=var)
self._all_params_unit[var] = dim_
dynamics.add_state_variable(sv_)
self._component_type.add(lems.Exposure(var, dimension=dim_))
elif ng_equations[var].type in (PARAMETER, SUBEXPRESSION):
if var == NOT_REFRACTORY:
continue
dim_ = _determine_dimension(ng_equations[var].unit)
self._all_params_unit[var] = dim_
# all initializers contatining iterator need to be assigned
# as a property
# i is default iterator in Brian2
if var in initializers and "i" in get_identifiers(str(initializers[var])):
self._component_type.add(lems.Property(var, dim_))
special_properties[var] = initializers[var]
continue
sv_ = lems.StateVariable(var, dimension=dim_)
dynamics.add_state_variable(sv_)
# what happens at initialization
onstart = lems.OnStart()
for var in obj.equations.names:
if var in (NOT_REFRACTORY, LAST_SPIKE):
continue
if var not in initializers:
continue
if var in special_properties:
continue
init_value = initializers[var]
if type(init_value) != str:
init_value = brian_unit_to_lems(init_value)
else:
init_value = renderer.render_expr(str(init_value))
onstart.add(lems.StateAssignment(var, init_value))
dynamics.add(onstart)
if obj._refractory:
# if refractoriness, we create separate regimes
# - for integrating
integr_regime = lems.Regime(INTEGRATING, dynamics, True) # True -> initial regime
for spike_flag, oc in self._event_builder(obj.events, obj.event_codes):
if spike_flag:
# if spike occured we make transition to refractory regime
oc.add_action(lems.Transition(REFRACTORY))
integr_regime.add_event_handler(oc)
# - for refractory
refrac_regime = lems.Regime(REFRACTORY, dynamics)
# we make lastspike variable and initialize it
refrac_regime.add_state_variable(lems.StateVariable(LAST_SPIKE, dimension='time'))
oe = lems.OnEntry()
oe.add(lems.StateAssignment(LAST_SPIKE, 't'))
refrac_regime.add(oe)
# after time spiecified in _refractory we make transition
# to integrating regime
if not _equation_separator(str(obj._refractory)):
# if there is no specific variable given, we assume
# that this is time condition
ref_oc = lems.OnCondition('t .gt. ( {0} + {1} )'.format(LAST_SPIKE, brian_unit_to_lems(obj._refractory)))
else:
ref_oc = lems.OnCondition(renderer.render_expr(obj._refractory))
ref_trans = lems.Transition(INTEGRATING)
ref_oc.add_action(ref_trans)
refrac_regime.add_event_handler(ref_oc)
for var in obj.user_equations.diff_eq_names:
td = lems.TimeDerivative(var, renderer.render_expr(str(ng_equations[var].expr)))
# if unless refratory we add only do integration regime
if UNLESS_REFRACTORY in ng_equations[var].flags:
integr_regime.add_time_derivative(td)
continue
integr_regime.add_time_derivative(td)
refrac_regime.add_time_derivative(td)
dynamics.add_regime(integr_regime)
dynamics.add_regime(refrac_regime)
else:
# here we add events directly to dynamics
for spike_flag, oc in self._event_builder(obj.events, obj.event_codes):
dynamics.add_event_handler(oc)
for var in obj.user_equations.diff_eq_names:
td = lems.TimeDerivative(var, renderer.render_expr(str(ng_equations[var].expr)))
dynamics.add_time_derivative(td)
self._component_type.dynamics = dynamics
# making componenttype is done so we add it to the model
self._model.add_component_type(self._component_type)
obj.namespace.pop("init", None) # kick out init
# adding component to the model
paramdict = dict()
for param in obj.namespace:
paramdict[param] = self._unit_lems_validator(obj.namespace[param])
if obj.N == 1:
self._model.add(lems.Component("n{}".format(idx_of_ng), ct_name, **paramdict))
else:
self.make_multiinstantiate(special_properties, ct_name, paramdict)
def mdf_to_neuroml(graph, save_to=None, format=None, run_duration_sec=2):
print("Converting graph: %s to NeuroML" % (graph.id))
net = neuromllite.Network(id=graph.id)
net.notes = "NeuroMLlite export of {} graph: {}".format(
format if format else "MDF",
graph.id,
)
model = lems.Model()
lems_definitions = "%s_lems_definitions.xml" % graph.id
for node in graph.nodes:
print(" Node: %s" % node.id)
node_comp_type = "%s__definition" % node.id
node_comp = "%s__instance" % node.id
# Create the ComponentType which defines behaviour of the general class
ct = lems.ComponentType(node_comp_type, extends="baseCellMembPotDL")
ct.add(lems.Attachments("only_input_port", "basePointCurrentDL"))
ct.dynamics.add(
lems.DerivedVariable(name="V",
dimension="none",
value="0",
exposure="V"))
model.add(ct)
# Define the Component - an instance of the ComponentType
comp = lems.Component(node_comp, node_comp_type)
model.add(comp)
cell = neuromllite.Cell(id=node_comp,
lems_source_file=lems_definitions)
net.cells.append(cell)
pop = neuromllite.Population(
id=node.id,
size=1,
component=cell.id,
properties={
"color": "0.2 0.2 0.2",
"radius": 3
},
)
net.populations.append(pop)
if len(node.input_ports) > 1:
raise Exception(
"Currently only max 1 input port supported in NeuroML...")
for ip in node.input_ports:
ct.add(lems.Exposure(ip.id, "none"))
ct.dynamics.add(
lems.DerivedVariable(
name=ip.id,
dimension="none",
select="only_input_port[*]/I",
reduce="add",
exposure=ip.id,
))
on_start = None
for p in node.parameters:
print("Converting %s" % p)
if p.value is not None:
try:
v_num = float(p.value)
ct.add(lems.Parameter(p.id, "none"))
comp.parameters[p.id] = v_num
print(comp.parameters[p.id])
except Exception as e:
ct.add(lems.Exposure(p.id, "none"))
dv = lems.DerivedVariable(
name=p.id,
dimension="none",
value="%s" % (p.value),
exposure=p.id,
)
ct.dynamics.add(dv)
elif p.function is not None:
ct.add(lems.Exposure(p.id, "none"))
func_info = mdf_functions[p.function]
expr = func_info["expression_string"]
expr2 = substitute_args(expr, p.args)
for arg in p.args:
expr += ";{}={}".format(arg, p.args[arg])
dv = lems.DerivedVariable(name=p.id,
dimension="none",
value="%s" % (expr2),
exposure=p.id)
ct.dynamics.add(dv)
else:
ct.add(lems.Exposure(p.id, "none"))
ct.dynamics.add(
lems.StateVariable(name=p.id,
dimension="none",
exposure=p.id))
if p.default_initial_value:
if on_start is None:
on_start = lems.OnStart()
ct.dynamics.add(on_start)
sa = lems.StateAssignment(
variable=p.id,
value=str(evaluate_expr(p.default_initial_value)))
on_start.actions.append(sa)
if p.time_derivative:
td = lems.TimeDerivative(variable=p.id,
value=p.time_derivative)
ct.dynamics.add(td)
if len(node.output_ports) > 1:
raise Exception(
"Currently only max 1 output port supported in NeuroML...")
for op in node.output_ports:
ct.add(lems.Exposure(op.id, "none"))
ct.dynamics.add(
lems.DerivedVariable(name=op.id,
dimension="none",
value=op.value,
exposure=op.id))
only_output_port = "only_output_port"
ct.add(lems.Exposure(only_output_port, "none"))
ct.dynamics.add(
lems.DerivedVariable(
name=only_output_port,
dimension="none",
value=op.id,
exposure=only_output_port,
))
if len(graph.edges) > 0:
model.add(
lems.Include(
os.path.join(os.path.dirname(__file__),
"syn_definitions.xml")))
rsDL = neuromllite.Synapse(id="rsDL",
lems_source_file=lems_definitions)
net.synapses.append(rsDL)
# syn_id = 'silentSyn'
# silentSynDL = neuromllite.Synapse(id=syn_id, lems_source_file=lems_definitions)
for edge in graph.edges:
print(f" Edge: {edge.id} connects {edge.sender} to {edge.receiver}")
ssyn_id = "silentSyn_proj_%s" % edge.id
ssyn_id = "silentSyn_proj_%s" % edge.id
# ssyn_id = 'silentSynX'
silentDLin = neuromllite.Synapse(id=ssyn_id,
lems_source_file=lems_definitions)
model.add(lems.Component(ssyn_id, "silentRateSynapseDL"))
net.synapses.append(silentDLin)
net.projections.append(
neuromllite.Projection(
id="proj_%s" % edge.id,
presynaptic=edge.sender,
postsynaptic=edge.receiver,
synapse=rsDL.id,
pre_synapse=silentDLin.id,
type="continuousProjection",
weight=1,
random_connectivity=neuromllite.RandomConnectivity(
probability=1),
))
# Much more todo...
model.export_to_file(lems_definitions)
print("Nml net: %s" % net)
if save_to:
new_file = net.to_json_file(save_to)
print("Saved NML to: %s" % save_to)
################################################################################
### Build Simulation object & save as JSON
simtime = 1000 * run_duration_sec
dt = 0.1
sim = neuromllite.Simulation(
id="Sim%s" % net.id,
network=new_file,
duration=simtime,
dt=dt,
seed=123,
recordVariables={"OUTPUT": {
"all": "*"
}},
)
recordVariables = {}
for node in graph.nodes:
for ip in node.input_ports:
if not ip.id in recordVariables:
recordVariables[ip.id] = {}
recordVariables[ip.id][node.id] = 0
for p in node.parameters:
if p.is_stateful():
if not p.id in recordVariables:
recordVariables[p.id] = {}
recordVariables[p.id][node.id] = 0
for op in node.output_ports:
if not op.id in recordVariables:
recordVariables[op.id] = {}
recordVariables[op.id][node.id] = 0
sim.recordVariables = recordVariables
if save_to:
sf = sim.to_json_file()
print("Saved Simulation to: %s" % sf)
return net, sim