Permalink
Branch: master
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
609 lines (494 sloc) 25.1 KB
#from brian import Equations, NeuronGroup, Clock, CoincidenceCounter, Network, zeros, array, \
# ones, kron, ms, second, concatenate, hstack, sort, nonzero, diff, TimedArray, \
# reshape, sum, log
from brian import *
from brian.tools.statistics import firing_rate, get_gamma_factor
try:
from playdoh import *
except Exception, e:
print e
raise ImportError("Playdoh must be installed (https://code.google.com/p/playdoh/)")
try:
import pycuda
from brian.library.modelfitting.gpu_modelfitting import GPUModelFitting
can_use_gpu = True
except ImportError:
can_use_gpu = False
from brian.experimental.codegen.integration_schemes import *
import sys, cPickle
__all__ = ['modelfitting', 'print_table', 'get_spikes', 'predict', 'PSO', 'GA','CMAES',
'MAXCPU', 'MAXGPU',
'debug_level', 'info_level', 'warning_level', 'open_server']
class ModelFitting(Fitness):
def initialize(self, **kwds):
self.use_gpu = self.unit_type=='GPU'
# Gets the key,value pairs in shared_data
for key, val in self.shared_data.iteritems():
setattr(self, key, val)
# Gets the key,value pairs in **kwds
for key, val in kwds.iteritems():
setattr(self, key, val)
# log_info(self.model)
self.model = cPickle.loads(self.model)
# log_info(self.model)
# if model is a string
if type(self.model) is str:
self.model = Equations(self.model)
self.total_steps = int(self.duration / self.dt)
self.neurons = self.nodesize
self.groups = self.groups
# Time slicing
self.input = self.input[0:self.slices * (len(self.input) / self.slices)] # makes sure that len(input) is a multiple of slices
self.duration = len(self.input) * self.dt # duration of the input
self.sliced_steps = len(self.input) / self.slices # timesteps per slice
self.overlap_steps = int(self.overlap / self.dt) # timesteps during the overlap
self.total_steps = self.sliced_steps + self.overlap_steps # total number of timesteps
self.sliced_duration = self.overlap + self.duration / self.slices # duration of the vectorized simulation
self.N = self.neurons * self.slices # TOTAL number of neurons in this worker
self.input = hstack((zeros(self.overlap_steps), self.input)) # add zeros at the beginning because there is no overlap from the previous slice
# Prepares data (generates I_offset, spiketimes, spiketimes_offset)
self.prepare_data()
# Add 'refractory' parameter on the CPU on the CPU only
if not self.use_gpu:
if self.max_refractory is not None:
refractory = 'refractory'
self.model.add_param('refractory', second)
else:
refractory = self.refractory
else:
if self.max_refractory is not None:
refractory = 0*ms
else:
refractory = self.refractory
# Must recompile the Equations : the functions are not transfered after pickling/unpickling
self.model.compile_functions()
self.group = NeuronGroup(self.N,
model=self.model,
reset=self.reset,
threshold=self.threshold,
refractory=refractory,
max_refractory = self.max_refractory,
method = self.method,
clock=Clock(dt=self.dt))
if self.initial_values is not None:
for param, value in self.initial_values.iteritems():
self.group.state(param)[:] = value
# Injects current in consecutive subgroups, where I_offset have the same value
# on successive intervals
k = -1
for i in hstack((nonzero(diff(self.I_offset))[0], len(self.I_offset) - 1)):
I_offset_subgroup_value = self.I_offset[i]
I_offset_subgroup_length = i - k
sliced_subgroup = self.group.subgroup(I_offset_subgroup_length)
input_sliced_values = self.input[I_offset_subgroup_value:I_offset_subgroup_value + self.total_steps]
sliced_subgroup.set_var_by_array(self.input_var, TimedArray(input_sliced_values, clock=self.group.clock))
k = i
if self.use_gpu:
# Select integration scheme according to method
if self.method == 'Euler': scheme = euler_scheme
elif self.method == 'RK': scheme = rk2_scheme
elif self.method == 'exponential_Euler': scheme = exp_euler_scheme
else: raise Exception("The numerical integration method is not valid")
self.mf = GPUModelFitting(self.group, self.model, self.input, self.I_offset,
self.spiketimes, self.spiketimes_offset, zeros(self.neurons), 0*ms, self.delta,
precision=self.precision, scheme=scheme)
else:
self.cc = CoincidenceCounter(self.group, self.spiketimes, self.spiketimes_offset,
onset=self.onset, delta=self.delta)
def prepare_data(self):
"""
Generates I_offset, spiketimes, spiketimes_offset from data,
and also target_length and target_rates.
The neurons are first grouped by time slice : there are group_size*group_count
per group/time slice
Within each time slice, the neurons are grouped by target train : there are
group_size neurons per group/target train
"""
# Generates I_offset
self.I_offset = zeros(self.N, dtype=int)
for slice in range(self.slices):
self.I_offset[self.neurons * slice:self.neurons * (slice + 1)] = self.sliced_steps * slice
# Generates spiketimes, spiketimes_offset, target_length, target_rates
i, t = zip(*self.data)
i = array(i)
t = array(t)
alls = []
n = 0
pointers = []
dt = self.dt
target_length = []
target_rates = []
model_target = []
group_index = 0
neurons_in_group = self.subpopsize
for j in xrange(self.groups):
# neurons_in_group = self.groups[j] # number of neurons in the current group and current worker
s = sort(t[i == j])
target_length.extend([len(s)] * neurons_in_group)
target_rates.extend([firing_rate(s)] * neurons_in_group)
for k in xrange(self.slices):
# first sliced group : 0...0, second_train...second_train, ...
# second sliced group : first_train_second_slice...first_train_second_slice, second_train_second_slice...
spikeindices = (s >= k * self.sliced_steps * dt) & (s < (k + 1) * self.sliced_steps * dt) # spikes targeted by sliced neuron number k, for target j
targeted_spikes = s[spikeindices] - k * self.sliced_steps * dt + self.overlap_steps * dt # targeted spikes in the "local clock" for sliced neuron k
targeted_spikes = hstack((-1 * second, targeted_spikes, self.sliced_duration + 1 * second))
model_target.extend([k + group_index * self.slices] * neurons_in_group)
alls.append(targeted_spikes)
pointers.append(n)
n += len(targeted_spikes)
group_index += 1
pointers = array(pointers, dtype=int)
model_target = array(hstack(model_target), dtype=int)
self.spiketimes = hstack(alls)
self.spiketimes_offset = pointers[model_target] # [pointers[i] for i in model_target]
# Duplicates each target_length value 'group_size' times so that target_length[i]
# is the length of the train targeted by neuron i
self.target_length = array(target_length, dtype=int)
self.target_rates = array(target_rates)
def evaluate(self, **param_values):
"""
Use fitparams['delays'] to take delays into account
Use fitparams['refractory'] to take refractory into account
"""
delays = param_values.pop('delays', zeros(self.neurons))
refractory = param_values.pop('refractory', zeros(self.neurons))
# kron spike delays
delays = kron(delays, ones(self.slices))
refractory = kron(refractory, ones(self.slices))
# Sets the parameter values in the NeuronGroup object
self.group.reinit()
for param, value in param_values.iteritems():
self.group.state(param)[:] = kron(value, ones(self.slices)) # kron param_values if slicing
# Reinitializes the model variables
if self.initial_values is not None:
for param, value in self.initial_values.iteritems():
self.group.state(param)[:] = value
if self.use_gpu:
# Reinitializes the simulation object
self.mf.reinit_vars(self.input, self.I_offset, self.spiketimes, self.spiketimes_offset, delays, refractory)
# LAUNCHES the simulation on the GPU
self.mf.launch(self.duration, self.stepsize)
coincidence_count = self.mf.coincidence_count
spike_count = self.mf.spike_count
else:
# set the refractory period
if self.max_refractory is not None:
self.group.refractory = refractory
self.cc = CoincidenceCounter(self.group, self.spiketimes, self.spiketimes_offset,
onset=self.onset, delta=self.delta)
# Sets the spike delay values
self.cc.spikedelays = delays
# Reinitializes the simulation objects
self.group.clock.reinit()
# self.cc.reinit()
net = Network(self.group, self.cc)
# LAUNCHES the simulation on the CPU
net.run(self.duration)
coincidence_count = self.cc.coincidences
spike_count = self.cc.model_length
coincidence_count = sum(reshape(coincidence_count, (self.slices, -1)), axis=0)
spike_count = sum(reshape(spike_count, (self.slices, -1)), axis=0)
gamma = get_gamma_factor(coincidence_count, spike_count, self.target_length, self.target_rates, self.delta)
return gamma
def modelfitting(model=None,
reset=None,
threshold=None,
refractory=0*ms,
data=None,
input_var='I',
input=None,
dt=None,
popsize=1000,
maxiter=10,
delta=4*ms,
slices=1,
overlap=0*second,
initial_values=None,
stepsize=100 * ms,
unit_type=None,
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,
async = 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``
The optimization algorithm. It can be :class:`PSO`, :class:`GA` or :class:`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>`__).
"""
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)
# Make sure that 'data' is a N*2-array
data = array(data)
if data.ndim == 1:
data = concatenate((zeros((len(data), 1)), data.reshape((-1, 1))), axis=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 and unit_type is None:
if CANUSEGPU: unit_type = 'GPU'
else: unit_type = 'CPU'
# 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,delta=delta,
slices=slices,
overlap=overlap,
returninfo=returninfo,
precision=precision,
stepsize=stepsize,
method=method,
onset=0 * ms)
shared_data = dict(input=input,
data=data,
initial_values=initial_values)
if async:
r = maximize_async( ModelFitting,
shared_data=shared_data,
kwds = kwds,
groups=groups,
popsize=popsize,
maxiter=maxiter,
optparams=optparams,
unit_type = unit_type,
total_units = total_units,
machines=machines,
allocation=allocation,
cpu=cpu,
gpu=gpu,
returninfo=returninfo,
codedependencies=[],
algorithm=algorithm,
scaling=scaling,
**params)
else:
r = maximize( ModelFitting,
shared_data=shared_data,
kwds = kwds,
groups=groups,
popsize=popsize,
maxiter=maxiter,
optparams=optparams,
unit_type = unit_type,
total_units = total_units,
machines=machines,
allocation=allocation,
cpu=cpu,
gpu=gpu,
returninfo=returninfo,
codedependencies=[],
algorithm=algorithm,
scaling=scaling,
**params)
return r
def get_spikes(model=None, reset=None, threshold=None,
input=None, input_var='I', dt=None,
initial_values=None,
**params):
"""
Retrieves the spike times corresponding to the best parameters found by
the modelfitting function.
**Arguments**
``model``, ``reset``, ``threshold``, ``input``, ``input_var``, ``dt``, ``initial_values``
Same parameters as for the ``modelfitting`` function.
``**params``
The best parameters returned by the ``modelfitting`` function.
**Returns**
``spiketimes``
The spike times of the model with the given input and parameters.
"""
duration = len(input) * dt
ngroups = len(params[params.keys()[0]])
group = NeuronGroup(N=ngroups, model=model, reset=reset, threshold=threshold,
clock=Clock(dt=dt))
group.set_var_by_array(input_var, TimedArray(input, clock=group.clock))
for param, values in params.iteritems():
if (param == 'delays') | (param == 'fitness'):
continue
group.state(param)[:] = values
if initial_values is not None:
for param, value in initial_values.iteritems():
group.state(param)[:] = value
M = SpikeMonitor(group)
net = Network(group, M)
net.run(duration)
reinit_default_clock()
return M.spikes
def predict(model=None, reset=None, threshold=None,
data=None, delta=4 * ms,
input=None, input_var='I', dt=None,
**params):
"""
Predicts the gamma factor of a fitted model with respect to the data with
a different input current.
**Arguments**
``model``, ``reset``, ``threshold``, ``input_var``, ``dt``
Same parameters as for the ``modelfitting`` function.
``input``
The input current, that can be different from the current used for the fitting
procedure.
``data``
The experimental spike times to compute the gamma factor against. They have
been obtained with the current ``input``.
``**params``
The best parameters returned by the ``modelfitting`` function.
**Returns**
``gamma``
The gamma factor of the model spike trains against the data.
If there were several groups in the fitting procedure, it is a vector
containing the gamma factor for each group.
"""
spikes = get_spikes(model=model, reset=reset, threshold=threshold,
input=input, input_var=input_var, dt=dt,
**params)
ngroups = len(params[params.keys()[0]])
gamma = zeros(ngroups)
for i in xrange(ngroups):
spk = [t for j, t in spikes if j == i]
gamma[i] = gamma_factor(spk, data, delta, normalize=True, dt=dt)
if len(gamma) == 1:
return gamma[0]
else:
return gamma