def spike_call(self, spiking_neurons):
dt = self.dt
t = self.step() * dt
spiking_neurons = array(spiking_neurons)
if len(spiking_neurons) > 0:
if t >= self.onset:
self.spike_count[spiking_neurons] += 1
T_spiking = array(rint((t + self.delays[spiking_neurons]) / dt),
dtype=int)
remaining_neurons = spiking_neurons
remaining_T_spiking = T_spiking
while True:
remaining_indices, = (
remaining_T_spiking >
self.next_spike_time[remaining_neurons]).nonzero()
if len(remaining_indices):
indices = remaining_neurons[remaining_indices]
self.spiketime_index[indices] += 1
self.last_spike_time[indices] = self.next_spike_time[
indices]
self.next_spike_time[indices] = array(rint(
self.spikes_inline[self.spiketime_index[indices] + 1] /
dt),
dtype=int)
if self.algorithm == 'exclusive':
self.last_spike_allowed[
indices] = self.next_spike_allowed[indices]
self.next_spike_allowed[indices] = True
remaining_neurons = remaining_neurons[remaining_indices]
remaining_T_spiking = remaining_T_spiking[
remaining_indices]
else:
break
# Updates coincidences count
near_last_spike = self.last_spike_time[
spiking_neurons] + self.delta >= T_spiking
near_next_spike = self.next_spike_time[
spiking_neurons] - self.delta <= T_spiking
last_spike_allowed = self.last_spike_allowed[spiking_neurons]
next_spike_allowed = self.next_spike_allowed[spiking_neurons]
I = (near_last_spike & last_spike_allowed) | (near_next_spike
& next_spike_allowed)
if t >= self.onset:
self.coincidences[spiking_neurons[I]] += 1
if self.algorithm == 'exclusive':
near_both_allowed = (near_last_spike & last_spike_allowed) & (
near_next_spike & next_spike_allowed)
self.last_spike_allowed[
spiking_neurons] = last_spike_allowed & -near_last_spike
self.next_spike_allowed[spiking_neurons] = (
next_spike_allowed & -near_next_spike) | near_both_allowed
def __init__(self,
group,
traces=None,
spikes=None,
targets_count=1,
duration=None,
onset=0 * ms,
spikes_inline=None,
spikes_offset=None,
traces_inline=None,
traces_offset=None,
delays=None,
when='end',
**params):
NetworkOperation.__init__(self, None, clock=group.clock, when=when)
self.group = group
# needed by SpikeMonitor
self.source = group
self.source.set_max_delay(0)
self.delay = 0
self.N = len(group) # number of neurons
self.K = targets_count # number of targets
self.dt = self.clock.dt
if traces is not None:
self.traces = array(traces) # KxT array
if self.traces.ndim == 1:
self.traces = self.traces.reshape((1, -1))
assert targets_count == self.traces.shape[0]
self.spikes = spikes
# get the duration from the traces if duration is not specified in the constructor
if duration is None:
duration = self.traces.shape[1] # total number of steps
self.duration = duration
self.total_steps = int(duration / self.dt)
if delays is None: delays = zeros(self.n)
self.delays = delays
self.onset = int(onset / self.dt)
self.intdelays = array(self.delays / self.clock.dt, dtype=int)
self.mindelay = min(delays)
self.maxdelay = max(delays)
# the following data is sliced
self.spikes_inline = spikes_inline
self.spikes_offset = spikes_offset
self.traces_inline = traces_inline
self.traces_offset = traces_offset
if self.spikes is not None:
# target spike count and rates
self.target_spikes_count = self.get_spikes_count(self.spikes)
self.target_spikes_rate = self.get_spikes_rate(self.spikes)
self.initialize(**params)
def reinit_vars(self,
criterion,
I,
I_offset,
spiketimes=None,
spiketimes_offset=None,
traces=None,
traces_offset=None,
spikedelays=None,
refractory=None):
self.criterion = criterion
self.kernel_module = SourceModule(self.kernel_src)
self.kernel_func = self.kernel_module.get_function('runsim')
blocksize = BLOCKSIZE
# try:
# blocksize = self.kernel_func.get_attribute(pycuda.driver.function_attribute.MAX_THREADS_PER_BLOCK)
# except: # above won't work unless CUDA>=2.2
# pass
self.block = (blocksize, 1, 1)
self.grid = (int(ceil(float(self.N) / blocksize)), 1)
self.kernel_func_kwds = {'block': self.block, 'grid': self.grid}
mydtype = self.mydtype
N = self.N
eqs = self.eqs
statevars_arr = gpuarray.to_gpu(
array(self.G._S.flatten(), dtype=mydtype))
self.I = gpuarray.to_gpu(array(I, dtype=mydtype))
self.statevars_arr = statevars_arr
self.I_offset = gpuarray.to_gpu(array(I_offset, dtype=int32))
# SPIKES
if self.criterion.type == 'spikes':
self.initialize_spikes(spiketimes, spiketimes_offset)
# TRACES
if self.criterion.type == 'traces':
self.initialize_traces(traces, traces_offset)
self.criterion.initialize_cuda_variables()
self.initialize_delays(spikedelays)
self.initialize_refractory(refractory)
if self.statemonitor_var is not None:
self.initialize_statemonitor()
if self.spikemonitor is True:
self.initialize_spikemonitor()
self.initialize_kernel_arguments()
def transform_traces(self, traces):
# from standard structure to inline structure
K, T = traces.shape
traces_inline = traces.flatten()
traces_offset = array(kron(arange(K), T * ones(self.subpopsize)),
dtype=int)
return traces_inline, traces_offset
def initialize(self, delta, coincidence_count_algorithm):
self.algorithm = coincidence_count_algorithm
self.delta = int(rint(delta / self.dt))
self.spike_count = zeros(self.N, dtype='int')
self.coincidences = zeros(self.N, dtype='int')
self.spiketime_index = self.spikes_offset
self.last_spike_time = array(rint(self.spikes_inline[self.spiketime_index] / self.dt), dtype=int)
self.next_spike_time = array(rint(self.spikes_inline[self.spiketime_index + 1] / self.dt), dtype=int)
# First target spikes (needed for the computation of
# the target train firing rates)
# self.first_target_spike = zeros(self.N)
self.last_spike_allowed = ones(self.N, dtype='bool')
self.next_spike_allowed = ones(self.N, dtype='bool')
def initialize_refractory(self, refractory):
if isinstance(refractory, float):
refractory = refractory * ones(self.N)
self.refractory_arr = gpuarray.to_gpu(
array(rint(refractory / self.dt), dtype=int32))
self.next_allowed_spiketime_arr = gpuarray.to_gpu(
-ones(self.N, dtype=int32))
def spike_call(self, spiking_neurons):
dt = self.dt
t = self.step()*dt
spiking_neurons = array(spiking_neurons)
if len(spiking_neurons)>0:
if t >= self.onset:
self.spike_count[spiking_neurons] += 1
T_spiking = array(rint((t + self.delays[spiking_neurons]) / dt), dtype=int)
remaining_neurons = spiking_neurons
remaining_T_spiking = T_spiking
while True:
remaining_indices, = (remaining_T_spiking > self.next_spike_time[remaining_neurons]).nonzero()
if len(remaining_indices):
indices = remaining_neurons[remaining_indices]
self.spiketime_index[indices] += 1
self.last_spike_time[indices] = self.next_spike_time[indices]
self.next_spike_time[indices] = array(rint(self.spikes_inline[self.spiketime_index[indices] + 1] / dt), dtype=int)
if self.algorithm == 'exclusive':
self.last_spike_allowed[indices] = self.next_spike_allowed[indices]
self.next_spike_allowed[indices] = True
remaining_neurons = remaining_neurons[remaining_indices]
remaining_T_spiking = remaining_T_spiking[remaining_indices]
else:
break
# Updates coincidences count
near_last_spike = self.last_spike_time[spiking_neurons] + self.delta >= T_spiking
near_next_spike = self.next_spike_time[spiking_neurons] - self.delta <= T_spiking
last_spike_allowed = self.last_spike_allowed[spiking_neurons]
next_spike_allowed = self.next_spike_allowed[spiking_neurons]
I = (near_last_spike & last_spike_allowed) | (near_next_spike & next_spike_allowed)
if t >= self.onset:
self.coincidences[spiking_neurons[I]] += 1
if self.algorithm == 'exclusive':
near_both_allowed = (near_last_spike & last_spike_allowed) & (near_next_spike & next_spike_allowed)
self.last_spike_allowed[spiking_neurons] = last_spike_allowed & -near_last_spike
self.next_spike_allowed[spiking_neurons] = (next_spike_allowed & -near_next_spike) | near_both_allowed
def reinit_vars(self, criterion,
I, I_offset,
spiketimes = None, spiketimes_offset = None,
traces = None, traces_offset = None,
spikedelays = None, refractory = None):
self.criterion = criterion
self.kernel_module = SourceModule(self.kernel_src)
self.kernel_func = self.kernel_module.get_function('runsim')
blocksize = BLOCKSIZE
# try:
# blocksize = self.kernel_func.get_attribute(pycuda.driver.function_attribute.MAX_THREADS_PER_BLOCK)
# except: # above won't work unless CUDA>=2.2
# pass
self.block = (blocksize, 1, 1)
self.grid = (int(ceil(float(self.N) / blocksize)), 1)
self.kernel_func_kwds = {'block':self.block, 'grid':self.grid}
mydtype = self.mydtype
N = self.N
eqs = self.eqs
statevars_arr = gpuarray.to_gpu(array(self.G._S.flatten(), dtype=mydtype))
self.I = gpuarray.to_gpu(array(I, dtype=mydtype))
self.statevars_arr = statevars_arr
self.I_offset = gpuarray.to_gpu(array(I_offset, dtype=int32))
# SPIKES
if self.criterion.type == 'spikes':
self.initialize_spikes(spiketimes, spiketimes_offset)
# TRACES
if self.criterion.type == 'traces':
self.initialize_traces(traces, traces_offset)
self.criterion.initialize_cuda_variables()
self.initialize_delays(spikedelays)
self.initialize_refractory(refractory)
if self.statemonitor_var is not None:
self.initialize_statemonitor()
if self.spikemonitor is True:
self.initialize_spikemonitor()
self.initialize_kernel_arguments()
def initialize(self, delta, coincidence_count_algorithm):
self.algorithm = coincidence_count_algorithm
self.delta = int(rint(delta / self.dt))
self.spike_count = zeros(self.N, dtype='int')
self.coincidences = zeros(self.N, dtype='int')
self.spiketime_index = self.spikes_offset
self.last_spike_time = array(rint(
self.spikes_inline[self.spiketime_index] / self.dt),
dtype=int)
self.next_spike_time = array(rint(
self.spikes_inline[self.spiketime_index + 1] / self.dt),
dtype=int)
# First target spikes (needed for the computation of
# the target train firing rates)
# self.first_target_spike = zeros(self.N)
self.last_spike_allowed = ones(self.N, dtype='bool')
self.next_spike_allowed = ones(self.N, dtype='bool')
def __init__(self, group, traces=None, spikes=None, targets_count=1, duration=None, onset=0*ms,
spikes_inline=None, spikes_offset=None,
traces_inline=None, traces_offset=None,
delays=None, when='end', **params):
NetworkOperation.__init__(self, None, clock=group.clock, when=when)
self.group = group
# needed by SpikeMonitor
self.source = group
self.source.set_max_delay(0)
self.delay = 0
self.N = len(group) # number of neurons
self.K = targets_count # number of targets
self.dt = self.clock.dt
if traces is not None:
self.traces = array(traces) # KxT array
if self.traces.ndim == 1:
self.traces = self.traces.reshape((1,-1))
assert targets_count==self.traces.shape[0]
self.spikes = spikes
# get the duration from the traces if duration is not specified in the constructor
if duration is None: duration = self.traces.shape[1] # total number of steps
self.duration = duration
self.total_steps = int(duration/self.dt)
if delays is None: delays = zeros(self.n)
self.delays = delays
self.onset = int(onset/self.dt)
self.intdelays = array(self.delays/self.clock.dt, dtype=int)
self.mindelay = min(delays)
self.maxdelay = max(delays)
# the following data is sliced
self.spikes_inline = spikes_inline
self.spikes_offset = spikes_offset
self.traces_inline = traces_inline
self.traces_offset = traces_offset
if self.spikes is not None:
# target spike count and rates
self.target_spikes_count = self.get_spikes_count(self.spikes)
self.target_spikes_rate = self.get_spikes_rate(self.spikes)
self.initialize(**params)
def transform_spikes(self, spikes):
# from standard structure to inline structure
i, t = zip(*spikes)
i = array(i)
t = array(t)
alls = []
n = 0
pointers = []
model_target = []
for j in xrange(self.groups):
s = sort(t[i == j])
s = hstack((-1 * second, s, self.duration + 1 * second))
model_target.extend([j] * self.subpopsize)
alls.append(s)
pointers.append(n)
n += len(s)
pointers = array(pointers, dtype=int)
model_target = array(hstack(model_target), dtype=int)
spikes_inline = hstack(alls)
spikes_offset = pointers[model_target]
return spikes_inline, spikes_offset
def transform_spikes(self, spikes):
# from standard structure to inline structure
i, t = zip(*spikes)
i = array(i)
t = array(t)
alls = []
n = 0
pointers = []
model_target = []
for j in xrange(self.groups):
s = sort(t[i == j])
s = hstack((-1 * second, s, self.duration + 1 * second))
model_target.extend([j] * self.subpopsize)
alls.append(s)
pointers.append(n)
n += len(s)
pointers = array(pointers, dtype=int)
model_target = array(hstack(model_target), dtype=int)
spikes_inline = hstack(alls)
spikes_offset = pointers[model_target]
return spikes_inline, spikes_offset
def transform_trials(self, spikes):
# from standard structure to inline structure
i, t = zip(*spikes)
i = array(i)
t = array(t)
alls = []
n = 0
pointers = []
model_target = []
pointers.append(0)
for j in xrange(self.ntrials):
s = sort(t[i == j])
s = hstack((-1 * second, s, self.duration + 1 * second))
alls.append(s)
pointers.append(pointers[j]+len(s))
pointers.pop()
spikes_inline = hstack(alls)
# print pointers
# print nonzero(spikes_inline==-1)
# show()
return spikes_inline, pointers
def initialize_spikes(self, spiketimes, spiketimes_indices):
self.spiketimes = gpuarray.to_gpu(array(rint(spiketimes / self.dt), dtype=int32))
self.spiketime_indices = gpuarray.to_gpu(array(spiketimes_indices, dtype=int32))
def transform_traces(self, traces):
# from standard structure to inline structure
K, T = traces.shape
traces_inline = traces.flatten()
traces_offset = array(kron(arange(K), T*ones(self.subpopsize)), dtype=int)
return traces_inline, traces_offset
def simulate(model,
reset=None,
threshold=None,
input=None,
input_var='I',
dt=defaultclock.dt,
refractory=0 * ms,
max_refractory=None,
data=None,
groups=1,
slices=1,
overlap=0 * second,
neurons=1,
initial_values=None,
use_gpu=False,
stepsize=100 * ms,
precision='double',
criterion=None,
record=None,
method='Euler',
**params):
unit_type = 'CPU'
if use_gpu: unit_type = 'GPU'
for param in params.keys():
if (param not in model._diffeq_names) and (param != 'delays'):
raise Exception(
"Parameter %s must be defined as a parameter in the model" %
param)
if criterion is None:
criterion = GammaFactor()
data = array(data)
if criterion.type == 'spikes':
# Make sure that 'data' is a N*2-array
if data.ndim == 1:
data = concatenate((zeros((len(data), 1)), data.reshape((-1, 1))),
axis=1)
spikes = data
traces = None
groups = int(array(data)[:, 0].max() + 1) # number of target trains
elif criterion.type == 'trace':
if data.ndim == 1:
data = data.reshape((1, -1))
spikes = None
traces = data
groups = data.shape[0]
elif criterion.type == 'both':
# TODO
log_warn("Not implemented yet")
pass
inputs = input
if inputs.ndim == 1:
inputs = inputs.reshape((1, -1))
# dt must be set
if dt is None:
raise Exception('dt (sampling frequency of the input) must be set')
# default overlap when no time slicing
if slices == 1:
overlap = 0 * ms
# check numerical integration method
if use_gpu and method not in ['Euler', 'RK', 'exponential_Euler']:
raise Exception(
"The method can only be 'Euler', 'RK', or 'exponential_Euler' when using the GPU"
)
if method not in [
'Euler', 'RK', 'exponential_Euler', 'linear', 'nonlinear'
]:
raise Exception(
"The method can only be 'Euler', 'RK', 'exponential_Euler', 'linear', or 'nonlinear'"
)
# determines whether optimization over refractoriness or not
if type(refractory) is tuple or type(refractory) is list:
params['refractory'] = refractory
max_refractory = refractory[-1]
else:
max_refractory = None
# Initialize GPU
if use_gpu:
set_gpu_device(0)
simulator = Simulator(model,
reset,
threshold,
inputs,
input_var=input_var,
dt=dt,
refractory=refractory,
max_refractory=max_refractory,
spikes=spikes,
traces=traces,
groups=groups,
slices=slices,
overlap=overlap,
onset=overlap,
neurons=neurons,
initial_values=initial_values,
unit_type=unit_type,
stepsize=stepsize,
precision=precision,
criterion=criterion,
statemonitor_var=record,
method=method)
criterion_values = simulator.run(**params)
if record is not None:
record_values = simulator.get_statemonitor_values()
return criterion_values, record_values
else:
return criterion_values
def simulate( model, reset = None, threshold = None,
input = None, input_var = 'I', dt = defaultclock.dt,
refractory = 0*ms, max_refractory = None,
data = None,
groups = 1,
slices = 1,
overlap = 0*second,
onset = None,
neurons = 1,
initial_values = None,
use_gpu = False,
stepsize = 128*ms,
precision = 'double',
criterion = None,
ntrials=1,
record = None,
spikemonitor = False,
nbr_spikes = 200,
method = 'Euler',
stand_alone=False,
# neuron_group = none,
**params):
unit_type = 'CPU'
if use_gpu: unit_type = 'GPU'
for param in params.keys():
if (param not in model._diffeq_names) and (param != 'delays') and (param != 'tau_metric'):
raise Exception("Parameter %s must be defined as a parameter in the model" % param)
if criterion is None:
criterion = GammaFactor()
data = array(data)
if criterion.type == 'spikes':
# Make sure that 'data' is a N*2-array
if data.ndim == 1:
data = concatenate((zeros((len(data), 1)), data.reshape((-1, 1))), axis=1)
spikes = data
traces = None
groups = int(array(data)[:, 0].max() + 1) # number of target trains
elif criterion.type == 'trace':
if data.ndim == 1:
data = data.reshape((1,-1))
spikes = None
traces = data
groups = data.shape[0]
elif criterion.type == 'both':
# TODO
log_warn("Not implemented yet")
pass
inputs = input
if inputs.ndim==1:
inputs = inputs.reshape((1,-1))
# dt must be set
if dt is None:
raise Exception('dt (sampling frequency of the input) must be set')
# default overlap when no time slicing
if slices == 1:
overlap = 0*ms
if onset is None:
onset = overlap
# check numerical integration method
if use_gpu and method not in ['Euler', 'RK', 'exponential_Euler']:
raise Exception("The method can only be 'Euler', 'RK', or 'exponential_Euler' when using the GPU")
if method not in ['Euler', 'RK', 'exponential_Euler', 'linear', 'nonlinear']:
raise Exception("The method can only be 'Euler', 'RK', 'exponential_Euler', 'linear', or 'nonlinear'")
# determines whether optimization over refractoriness or not
if type(refractory) is tuple or type(refractory) is list:
params['refractory'] = refractory
max_refractory = refractory[-1]
else:
max_refractory = None
# Initialize GPU
if use_gpu:
set_gpu_device(0)
# if neuron_group is not None:
# self.neuron_group = neuron_group
# self.given_neuron_group = True
# else:
# self.given_neuron_group = False
simulator = Simulator(model, reset, threshold,
inputs,
input_var = input_var,
dt = dt,
refractory = refractory,
max_refractory = max_refractory,
spikes = spikes,
traces = traces,
groups = groups,
slices = slices,
overlap = overlap,
onset = onset,
neurons = neurons,
initial_values = initial_values,
unit_type = unit_type,
stepsize = stepsize,
precision = precision,
ntrials=ntrials,
criterion = criterion,
statemonitor_var = record,
spikemonitor = spikemonitor,
nbr_spikes = nbr_spikes,
method = method,
# stand_alone=stand_alone,
# self.neuron_group,
# self.given_neuron_group
)
criterion_values = simulator.run(**params)
if record is not None and spikemonitor is False:
record_values = simulator.get_statemonitor_values()
return criterion_values, record_values
elif record is not None and spikemonitor is True:
record_values = simulator.get_statemonitor_values()
spike_times = simulator.get_spikemonitor_values()
return criterion_values, record_values,spike_times
elif record is None and spikemonitor is True:
spike_times = simulator.get_spikemonitor_values()
return criterion_values,spike_times
else:
return criterion_values
For more details on the gamma factor, see
`Jolivet et al. 2008, "A benchmark test for a quantitative assessment of simple neuron models", J. Neurosci. Methods <http://www.ncbi.nlm.nih.gov/pubmed/18160135>`__ (available in PDF
`here <http://icwww.epfl.ch/~gerstner/PUBLICATIONS/Jolivet08.pdf>`__).
"""
for param in params.keys():
if (param not in model._diffeq_names) and (param != 'delays'):
raise Exception(
"Parameter %s must be defined as a parameter in the model" %
param)
if criterion is None:
criterion = GammaFactor()
data = array(data)
if criterion.type == 'spikes':
# Make sure that 'data' is a N*2-array
if data.ndim == 1:
data = concatenate((zeros((len(data), 1)), data.reshape((-1, 1))),
axis=1)
spikes = data
traces = None
groups = int(array(data)[:, 0].max() + 1) # number of target trains
elif criterion.type == 'trace':
if data.ndim == 1:
data = data.reshape((1, -1))
spikes = None
traces = data
groups = data.shape[0]
elif criterion.type == 'both':
``or[i][key]``
where ``i`` is a group index, is the same as ``or[i].best_pos[key]``.
For more details on the gamma factor, see
`Jolivet et al. 2008, "A benchmark test for a quantitative assessment of simple neuron models", J. Neurosci. Methods <http://www.ncbi.nlm.nih.gov/pubmed/18160135>`__ (available in PDF
`here <http://icwww.epfl.ch/~gerstner/PUBLICATIONS/Jolivet08.pdf>`__).
"""
for param in params.keys():
if (param not in model._diffeq_names) and (param != 'delays'):
raise Exception("Parameter %s must be defined as a parameter in the model" % param)
if criterion is None:
criterion = GammaFactor()
data = array(data)
if criterion.type == 'spikes':
# Make sure that 'data' is a N*2-array
if data.ndim == 1:
data = concatenate((zeros((len(data), 1)), data.reshape((-1, 1))), axis=1)
spikes = data
traces = None
groups = int(array(data)[:, 0].max() + 1) # number of target trains
elif criterion.type == 'trace':
if data.ndim == 1:
data = data.reshape((1,-1))
spikes = None
traces = data
groups = data.shape[0]
elif criterion.type == 'both':
# TODO
def initialize_refractory(self, refractory):
if isinstance(refractory, float):
refractory = refractory*ones(self.N)
self.refractory_arr = gpuarray.to_gpu(array(rint(refractory / self.dt), dtype=int32))
self.next_allowed_spiketime_arr = gpuarray.to_gpu(-ones(self.N, dtype=int32))
def modelfitting(model=None,
reset=None,
threshold=None,
refractory=0*ms,
data=None,
input_var='I',
input=None,
dt=None,
popsize=1000,
maxiter=10,
slices=1,
overlap=0*second,
initial_values=None,
stepsize=100 * ms,
unit_type='CPU',
total_units=None,
cpu=None,
gpu=None,
precision='double', # set to 'float' or 'double' to specify single or double precision on the GPU
machines=[],
allocation=None,
returninfo=False,
scaling=None,
algorithm=CMAES,
criterion=None,
optparams={},
method='Euler',
**params):
"""
Model fitting function.
Fits a spiking neuron model to electrophysiological data (injected current and spikes).
See also the section :ref:`model-fitting-library` in the user manual.
**Arguments**
``model``
An :class:`~brian.Equations` object containing the equations defining the model.
``reset``
A reset value for the membrane potential, or a string containing the reset
equations.
``threshold``
A threshold value for the membrane potential, or a string containing the threshold
equations.
``refractory``
The refractory period in second. If it's a single value, the same refractory will be
used in all the simulations. If it's a list or a tuple, the fitting will also
optimize the refractory period (see ``**params`` below).
Warning: when using a refractory period, you can't use a custom reset, only a fixed one.
``data``
A list of spike times, or a list of several spike trains as a list of pairs (index, spike time)
if the fit must be performed in parallel over several target spike trains. In this case,
the modelfitting function returns as many parameters sets as target spike trains.
``input_var='I'``
The variable name used in the equations for the input current.
``input``
A vector of values containing the time-varying signal the neuron responds to (generally
an injected current).
``dt``
The time step of the input (the inverse of the sampling frequency).
``**params``
The list of parameters to fit the model with. Each parameter must be set as follows:
``param_name=[bound_min, min, max, bound_max]``
where ``bound_min`` and ``bound_max`` are the boundaries, and ``min`` and ``max``
specify the interval from which the parameter values are uniformly sampled at
the beginning of the optimization algorithm.
If not using boundaries, set ``param_name=[min, max]``.
Also, you can add a fit parameter which is a spike delay for all spikes :
add the special parameter ``delays`` in ``**params``, for example
``modelfitting(..., delays=[-10*ms, 10*ms])``.
You can also add fit the refractory period by specifying
``modelfitting(..., refractory=[-10*ms, 10*ms])``.
``popsize``
Size of the population (number of particles) per target train used by the optimization algorithm.
``maxiter``
Number of iterations in the optimization algorithm.
``optparams``
Optimization algorithm parameters. It is a dictionary: keys are parameter names,
values are parameter values or lists of parameters (one value per group).
This argument is specific to the optimization
algorithm used. See :class:`PSO`, :class:`GA`, :class:`CMAES`.
``delta=4*ms``
The precision factor delta (a scalar value in second).
``slices=1``
The number of time slices to use.
``overlap=0*ms``
When using several time slices, the overlap between consecutive slices, in seconds.
``initial_values``
A dictionary containing the initial values for the state variables.
``cpu``
The number of CPUs to use in parallel. It is set to the number of CPUs in the machine by default.
``gpu``
The number of GPUs to use in parallel. It is set to the number of GPUs in the machine by default.
``precision``
GPU only: a string set to either ``float`` or ``double`` to specify whether to use
single or double precision on the GPU. If it is not specified, it will
use the best precision available.
``returninfo=False``
Boolean indicating whether the modelfitting function should return technical information
about the optimization.
``scaling=None``
Specify the scaling used for the parameters during the optimization.
It can be ``None`` or ``'mapminmax'``. It is ``None``
by default (no scaling), and ``mapminmax`` by default for the CMAES algorithm.
``algorithm=CMAES``
``optparams={}``
Optimization parameters. See
``method='Euler'``
Integration scheme used on the CPU and GPU: ``'Euler'`` (default), ``RK``,
or ``exponential_Euler``.
See also :ref:`numerical-integration`.
``machines=[]``
A list of machine names to use in parallel. See :ref:`modelfitting-clusters`.
**Return values**
Return an :class:`OptimizationResult` object with the following attributes:
``best_pos``
Minimizing position found by the algorithm. For array-like fitness functions,
it is a single vector if there is one group, or a list of vectors.
For keyword-like fitness functions, it is a dictionary
where keys are parameter names and values are numeric values. If there are several groups,
it is a list of dictionaries.
``best_fit``
The value of the fitness function for the best positions. It is a single value if
there is one group, or it is a list if there are several groups.
``info``
A dictionary containing various information about the optimization.
Also, the following syntax is possible with an ``OptimizationResult`` instance ``or``.
The ``key`` is either an optimizing parameter name for keyword-like fitness functions,
or a dimension index for array-like fitness functions.
``or[key]``
it is the best ``key`` parameter found (single value), or the list
of the best parameters ``key`` found for all groups.
``or[i]``
where ``i`` is a group index. This object has attributes ``best_pos``, ``best_fit``,
``info`` but only for group ``i``.
``or[i][key]``
where ``i`` is a group index, is the same as ``or[i].best_pos[key]``.
For more details on the gamma factor, see
`Jolivet et al. 2008, "A benchmark test for a quantitative assessment of simple neuron models", J. Neurosci. Methods <http://www.ncbi.nlm.nih.gov/pubmed/18160135>`__ (available in PDF
`here <http://icwww.epfl.ch/~gerstner/PUBLICATIONS/Jolivet08.pdf>`__).
"""
if criterion is None:
criterion = GammaFactor()
data = array(data)
if criterion.type == 'spikes':
# Make sure that 'data' is a N*2-array
if data.ndim == 1:
data = concatenate((zeros((len(data), 1)), data.reshape((-1, 1))), axis=1)
spikes = data
traces = None
elif criterion.type == 'trace':
if data.ndim == 1:
data = data.reshape((1,-1))
spikes = None
traces = data
elif criterion.type == 'both':
# TODO
log_warn("Not implemented yet")
pass
inputs = input
if inputs.ndim==1:
inputs = inputs.reshape((1,-1))
# dt must be set
if dt is None:
raise Exception('dt (sampling frequency of the input) must be set')
# default overlap when no time slicing
if slices == 1:
overlap = 0*ms
# default allocation
if cpu is None and gpu is None:
if CANUSEGPU: gpu = 1
else: cpu = MAXCPU-1
# check numerical integration method
if (gpu>0 or unit_type == 'GPU') and method not in ['Euler', 'RK', 'exponential_Euler']:
raise Exception("The method can only be 'Euler', 'RK', or 'exponential_Euler' when using the GPU")
if method not in ['Euler', 'RK', 'exponential_Euler', 'linear', 'nonlinear']:
raise Exception("The method can only be 'Euler', 'RK', 'exponential_Euler', 'linear', or 'nonlinear'")
if (algorithm == CMAES) & (scaling is None):
scaling = 'mapminmax'
# determines whether optimization over refractoriness or not
if type(refractory) is tuple or type(refractory) is list:
params['refractory'] = refractory
max_refractory = refractory[-1]
# refractory = 'refractory'
else:
max_refractory = None
# common values
# group_size = particles # Number of particles per target train
groups = int(array(data)[:, 0].max() + 1) # number of target trains
# N = group_size * group_count # number of neurons
duration = len(input) * dt # duration of the input
# keyword arguments for Modelfitting initialize
kwds = dict( model=cPickle.dumps(model),
threshold=threshold,
reset=reset,
refractory=refractory,
max_refractory=max_refractory,
input_var=input_var, dt=dt,
duration=duration,
criterion=criterion,
slices=slices,
overlap=overlap,
returninfo=returninfo,
precision=precision,
stepsize=stepsize,
method=method,
onset=overlap)
shared_data = dict(inputs=inputs,
traces=traces,
spikes=spikes,
initial_values=initial_values)
r = maximize( ModelFitting,
shared_data=shared_data,
kwds = kwds,
groups=groups,
popsize=popsize,
maxiter=maxiter,
optparams=optparams,
unit_type = unit_type,
machines=machines,
allocation=allocation,
cpu=cpu,
gpu=gpu,
returninfo=returninfo,
codedependencies=[],
algorithm=algorithm,
scaling=scaling,
**params)
# r is (results, fitinfo) or (results)
return r
def initialize_delays(self, spikedelays):
self.spikedelay_arr = gpuarray.to_gpu(
array(rint(spikedelays / self.dt), dtype=int32))
def initialize_spikes(self, spiketimes, spiketimes_indices):
self.spiketimes = gpuarray.to_gpu(
array(rint(spiketimes / self.dt), dtype=int32))
self.spiketime_indices = gpuarray.to_gpu(
array(spiketimes_indices, dtype=int32))
For more details on the gamma factor, see
`Jolivet et al. 2008, "A benchmark test for a quantitative assessment of simple neuron models", J. Neurosci. Methods <http://www.ncbi.nlm.nih.gov/pubmed/18160135>`__ (available in PDF
`here <http://icwww.epfl.ch/~gerstner/PUBLICATIONS/Jolivet08.pdf>`__).
"""
for param in params.keys():
if (param not in model._diffeq_names) and (param != 'delays') and (
param != 'tau_metric'):
raise Exception(
"Parameter %s must be defined as a parameter in the model" %
param)
if criterion is None:
criterion = GammaFactor()
data = array(data)
if criterion.type == 'spikes':
# Make sure that 'data' is a N*2-array
if data.ndim == 1:
data = concatenate((zeros((len(data), 1)), data.reshape((-1, 1))),
axis=1)
spikes = data
traces = None
if ntrials > 1:
groups = 1
else:
groups = int(array(data)[:, 0].max() +
1) # number of target trains
elif criterion.type == 'trace':
if data.ndim == 1:
def initialize_traces(self, traces, traces_offset):
self.traces = gpuarray.to_gpu(array(traces, dtype=double))
self.traces_offset = gpuarray.to_gpu(array(traces_offset, dtype=int32))
def initialize_traces(self, traces, traces_offset):
self.traces = gpuarray.to_gpu(array(traces, dtype=double))
self.traces_offset = gpuarray.to_gpu(array(traces_offset, dtype=int32))
def initialize_delays(self, spikedelays):
self.spikedelay_arr = gpuarray.to_gpu(array(rint(spikedelays / self.dt), dtype=int32))
def simulate(
model,
reset=None,
threshold=None,
input=None,
input_var="I",
dt=defaultclock.dt,
refractory=0 * ms,
max_refractory=None,
data=None,
groups=1,
slices=1,
overlap=0 * second,
neurons=1,
initial_values=None,
use_gpu=False,
stepsize=100 * ms,
precision="double",
criterion=None,
record=None,
method="Euler",
**params
):
unit_type = "CPU"
if use_gpu:
unit_type = "GPU"
for param in params.keys():
if (param not in model._diffeq_names) and (param != "delays"):
raise Exception("Parameter %s must be defined as a parameter in the model" % param)
if criterion is None:
criterion = GammaFactor()
data = array(data)
if criterion.type == "spikes":
# Make sure that 'data' is a N*2-array
if data.ndim == 1:
data = concatenate((zeros((len(data), 1)), data.reshape((-1, 1))), axis=1)
spikes = data
traces = None
groups = int(array(data)[:, 0].max() + 1) # number of target trains
elif criterion.type == "trace":
if data.ndim == 1:
data = data.reshape((1, -1))
spikes = None
traces = data
groups = data.shape[0]
elif criterion.type == "both":
# TODO
log_warn("Not implemented yet")
pass
inputs = input
if inputs.ndim == 1:
inputs = inputs.reshape((1, -1))
# dt must be set
if dt is None:
raise Exception("dt (sampling frequency of the input) must be set")
# default overlap when no time slicing
if slices == 1:
overlap = 0 * ms
# check numerical integration method
if use_gpu and method not in ["Euler", "RK", "exponential_Euler"]:
raise Exception("The method can only be 'Euler', 'RK', or 'exponential_Euler' when using the GPU")
if method not in ["Euler", "RK", "exponential_Euler", "linear", "nonlinear"]:
raise Exception("The method can only be 'Euler', 'RK', 'exponential_Euler', 'linear', or 'nonlinear'")
# determines whether optimization over refractoriness or not
if type(refractory) is tuple or type(refractory) is list:
params["refractory"] = refractory
max_refractory = refractory[-1]
else:
max_refractory = None
# Initialize GPU
if use_gpu:
set_gpu_device(0)
simulator = Simulator(
model,
reset,
threshold,
inputs,
input_var=input_var,
dt=dt,
refractory=refractory,
max_refractory=max_refractory,
spikes=spikes,
traces=traces,
groups=groups,
slices=slices,
overlap=overlap,
onset=overlap,
neurons=neurons,
initial_values=initial_values,
unit_type=unit_type,
stepsize=stepsize,
precision=precision,
criterion=criterion,
statemonitor_var=record,
method=method,
)
criterion_values = simulator.run(**params)
if record is not None:
record_values = simulator.get_statemonitor_values()
return criterion_values, record_values
else:
return criterion_values