def test_network_remove():
x = Counter()
y = Counter()
net = Network(x, y)
net.remove(y)
net.run(1*ms)
assert_equal(x.count, 10)
assert_equal(y.count, 0)
# the relevance of this test is when we use weakref.proxy objects in
# Network.objects, we should be able to add and remove these from
# the Network just as much as the original objects
for obj in copy.copy(net.objects):
net.remove(obj)
net.run(1*ms)
assert_equal(x.count, 10)
assert_equal(y.count, 0)
def test_network_remove():
x = Counter()
y = Counter()
net = Network(x, y)
net.remove(y)
net.run(1*ms)
assert_equal(x.count, 10)
assert_equal(y.count, 0)
# the relevance of this test is when we use weakref.proxy objects in
# Network.objects, we should be able to add and remove these from
# the Network just as much as the original objects
# TODO: Does this test make sense now that Network does not store weak
# references by default?
for obj in copy.copy(net.objects):
net.remove(obj)
net.run(1*ms)
assert_equal(x.count, 10)
assert_equal(y.count, 0)
def test_incorrect_network_use():
'''Test some wrong uses of `Network` and `MagicNetwork`'''
assert_raises(TypeError, lambda: Network(name='mynet',
anotherkwd='does not exist'))
assert_raises(TypeError, lambda: Network('not a BrianObject'))
net = Network()
assert_raises(TypeError, lambda: net.add('not a BrianObject'))
assert_raises(ValueError, lambda: MagicNetwork())
G = NeuronGroup(10, 'v:1')
net.add(G)
assert_raises(TypeError, lambda: net.remove(object()))
assert_raises(MagicError, lambda: magic_network.add(G))
assert_raises(MagicError, lambda: magic_network.remove(G))
def run_task_hierarchical(task_info, taskdir, tempdir):
# imports
from brian2 import defaultclock, set_device, seed, TimedArray, Network, profiling_summary
from brian2.monitors import SpikeMonitor, PopulationRateMonitor, StateMonitor
from brian2.synapses import Synapses
from brian2.core.magic import start_scope
from brian2.units import second, ms, amp
from integration_circuit import mk_intcircuit
from sensory_circuit import mk_sencircuit, mk_sencircuit_2c, mk_sencircuit_2cplastic
from burstQuant import spks2neurometric
from scipy import interpolate
# if you want to put something in the taskdir, you must create it first
os.mkdir(taskdir)
print(taskdir)
# parallel code and flag to start
set_device('cpp_standalone', directory=tempdir)
#prefs.devices.cpp_standalone.openmp_threads = max_tasks
start_scope()
# simulation parameters
seedcon = task_info['simulation']['seedcon']
runtime = task_info['simulation']['runtime']
runtime_ = runtime / second
settletime = task_info['simulation']['settletime']
settletime_ = settletime / second
stimon = task_info['simulation']['stimon']
stimoff = task_info['simulation']['stimoff']
stimoff_ = stimoff / second
stimdur = stimoff - stimon
smoothwin = task_info['simulation']['smoothwin']
nummethod = task_info['simulation']['nummethod']
# -------------------------------------
# Construct hierarchical network
# -------------------------------------
# set connection seed
seed(
seedcon
) # set specific seed to test the same network, this way we also have the same synapses!
# decision circuit
Dgroups, Dsynapses, Dsubgroups = mk_intcircuit(task_info)
decE = Dgroups['DE']
decI = Dgroups['DI']
decE1 = Dsubgroups['DE1']
decE2 = Dsubgroups['DE2']
# sensory circuit, ff and fb connections
eps = 0.2 # connection probability
d = 1 * ms # transmission delays of E synapses
if task_info['simulation']['2cmodel']:
if task_info['simulation']['plasticdend']:
# plasticity rule in dendrites --> FB synapses will be removed from the network!
Sgroups, Ssynapses, Ssubgroups = mk_sencircuit_2cplastic(task_info)
else:
# 2c model (Naud)
Sgroups, Ssynapses, Ssubgroups = mk_sencircuit_2c(task_info)
senE = Sgroups['soma']
dend = Sgroups['dend']
senI = Sgroups['SI']
senE1 = Ssubgroups['soma1']
senE2 = Ssubgroups['soma2']
dend1 = Ssubgroups['dend1']
dend2 = Ssubgroups['dend2']
# FB
wDS = 0.003 # synaptic weight of FB synapses, 0.0668 nS when scaled by gleakE of sencircuit_2c
synDE1SE1 = Synapses(decE1,
dend1,
model='w : 1',
method=nummethod,
on_pre='x_ea += w',
delay=d)
synDE2SE2 = Synapses(decE2,
dend2,
model='w : 1',
method=nummethod,
on_pre='x_ea += w',
delay=d)
else:
# normal sensory circuit (Wimmer)
Sgroups, Ssynapses, Ssubgroups = mk_sencircuit(task_info)
senE = Sgroups['SE']
senI = Sgroups['SI']
senE1 = Ssubgroups['SE1']
senE2 = Ssubgroups['SE2']
# FB
wDS = 0.004 # synaptic weight of FB synapses, 0.0668 nS when scaled by gleakE of sencircuit
synDE1SE1 = Synapses(decE1,
senE1,
model='w : 1',
method=nummethod,
on_pre='x_ea += w',
delay=d)
synDE2SE2 = Synapses(decE2,
senE2,
model='w : 1',
method=nummethod,
on_pre='x_ea += w',
delay=d)
# feedforward synapses from sensory to integration
wSD = 0.0036 # synaptic weight of FF synapses, 0.09 nS when scaled by gleakE of intcircuit
synSE1DE1 = Synapses(senE1,
decE1,
model='w : 1',
method=nummethod,
on_pre='g_ea += w',
delay=d)
synSE1DE1.connect(p='eps')
synSE1DE1.w = 'wSD'
synSE2DE2 = Synapses(senE2,
decE2,
model='w : 1',
method=nummethod,
on_pre='g_ea += w',
delay=d)
synSE2DE2.connect(p='eps')
synSE2DE2.w = 'wSD'
# feedback synapses from integration to sensory
b_fb = task_info['bfb'] # feedback strength, between 0 and 6
wDS *= b_fb # synaptic weight of FB synapses, 0.0668 nS when scaled by gleakE of sencircuit
synDE1SE1.connect(p='eps')
synDE1SE1.w = 'wDS'
synDE2SE2.connect(p='eps')
synDE2SE2.w = 'wDS'
# -------------------------------------
# Create stimuli
# -------------------------------------
if task_info['stimulus']['replicate']:
# replicated stimuli across iters()
np.random.seed(task_info['seed']) # numpy seed for OU process
else:
# every trials has different stimuli
np.random.seed()
# Note that in standalone we need to specify np seed because it's not taken care with Brian's seed() function!
if task_info['simulation']['2cmodel']:
I0 = task_info['stimulus']['I0s']
last_muOUd = np.loadtxt("last_muOUd.csv") # save the mean
else:
I0 = task_info['stimulus'][
'I0'] # mean input current for zero-coherence stim
c = task_info['c'] # stim coherence (between 0 and 1)
mu1 = task_info['stimulus'][
'mu1'] # av. additional input current to senE1 at highest coherence (c=1)
mu2 = task_info['stimulus'][
'mu2'] # av. additional input current to senE2 at highest coherence (c=1)
sigma = task_info['stimulus'][
'sigma'] # amplitude of temporal modulations of stim
sigmastim = 0.212 * sigma # std of modulation of stim inputs
sigmaind = 0.212 * sigma # std of modulations in individual inputs
taustim = task_info['stimulus'][
'taustim'] # correlation time constant of Ornstein-Uhlenbeck process
# generate stim from OU process
N_stim = int(senE1.__len__())
z1, z2, zk1, zk2 = generate_stim(N_stim, stimdur, taustim)
# stim2exc
i1 = I0 * (1 + c * mu1 + sigmastim * z1 + sigmaind * zk1)
i2 = I0 * (1 + c * mu2 + sigmastim * z2 + sigmaind * zk2)
stim_dt = 1 * ms
i1t = np.concatenate((np.zeros((int(stimon / ms), N_stim)), i1.T,
np.zeros((int(
(runtime - stimoff) / stim_dt), N_stim))),
axis=0)
i2t = np.concatenate((np.zeros((int(stimon / ms), N_stim)), i2.T,
np.zeros((int(
(runtime - stimoff) / stim_dt), N_stim))),
axis=0)
Irec = TimedArray(np.concatenate((i1t, i2t), axis=1) * amp, dt=stim_dt)
# -------------------------------------
# Simulation
# -------------------------------------
# set initial conditions (different for evert trial)
seed()
decE.g_ea = '0.2 * rand()'
decI.g_ea = '0.2 * rand()'
decE.V = '-52*mV + 2*mV * rand()'
decI.V = '-52*mV + 2*mV * rand()' # random initialization near 0, prevent an early decision!
senE.g_ea = '0.05 * (1 + 0.2*rand())'
senI.g_ea = '0.05 * (1 + 0.2*rand())'
senE.V = '-52*mV + 2*mV*rand()' # random initialization near Vt, avoid initial bump!
senI.V = '-52*mV + 2*mV*rand()'
if task_info['simulation']['2cmodel']:
dend.g_ea = '0.05 * (1 + 0.2*rand())'
dend.V_d = '-72*mV + 2*mV*rand()'
dend.muOUd = np.tile(last_muOUd, 2) * amp
# create monitors
rateDE1 = PopulationRateMonitor(decE1)
rateDE2 = PopulationRateMonitor(decE2)
rateSE1 = PopulationRateMonitor(senE1)
rateSE2 = PopulationRateMonitor(senE2)
subSE = int(senE1.__len__())
spksSE = SpikeMonitor(senE[subSE - 100:subSE +
100]) # last 100 of SE1 and first 100 of SE2
# construct network
net = Network(Dgroups.values(),
Dsynapses.values(),
Sgroups.values(),
Ssynapses.values(),
synSE1DE1,
synSE2DE2,
synDE1SE1,
synDE2SE2,
rateDE1,
rateDE2,
rateSE1,
rateSE2,
spksSE,
name='hierarchicalnet')
# create more monitors for plot
if task_info['simulation']['pltfig1']:
# inh
rateDI = PopulationRateMonitor(decI)
rateSI = PopulationRateMonitor(senI)
# spk monitors
subDE = int(decE1.__len__() * 2)
spksDE = SpikeMonitor(decE[:subDE])
spksSE = SpikeMonitor(senE)
# state mons no more, just the arrays
stim1 = i1t.T
stim2 = i2t.T
stimtime = np.linspace(0, runtime_, stim1.shape[1])
# construct network
net = Network(Dgroups.values(),
Dsynapses.values(),
Sgroups.values(),
Ssynapses.values(),
synSE1DE1,
synSE2DE2,
synDE1SE1,
synDE2SE2,
spksDE,
rateDE1,
rateDE2,
rateDI,
spksSE,
rateSE1,
rateSE2,
rateSI,
name='hierarchicalnet')
if task_info['simulation']['plasticdend']:
# create state monitor to follow muOUd and add it to the networks
dend_mon = StateMonitor(dend1,
variables=['muOUd', 'Ibg', 'g_ea', 'B'],
record=True,
dt=1 * ms)
net.add(dend_mon)
# remove FB synapses!
net.remove([synDE1SE1, synDE2SE2, Dsynapses.values()])
print(
" FB synapses and synapses of decision circuit are ignored in this simulation!"
)
# run hierarchical net
net.run(runtime, report='stdout', profile=True)
print(profiling_summary(net=net, show=10))
# nice plots on cluster
if task_info['simulation']['pltfig1']:
plot_fig1b([
rateDE1, rateDE2, rateDI, spksDE, rateSE1, rateSE2, rateSI, spksSE,
stim1, stim2, stimtime
], smoothwin, taskdir)
# -------------------------------------
# Burst quantification
# -------------------------------------
events = np.zeros(1)
bursts = np.zeros(1)
singles = np.zeros(1)
spikes = np.zeros(1)
last_muOUd = np.zeros(1)
# neurometric params
dt = spksSE.clock.dt
validburst = task_info['sen']['2c']['validburst']
smoothwin_ = smoothwin / second
if task_info['simulation']['burstanalysis']:
if task_info['simulation']['2cmodel']:
last_muOUd = np.array(dend_mon.muOUd[:, -int(1e3):].mean(axis=1))
if task_info['simulation']['plasticdend']:
# calculate neurometric info per population
events, bursts, singles, spikes, isis = spks2neurometric(
spksSE,
runtime,
settletime,
validburst,
smoothwin=smoothwin_,
raster=False)
# plot & save weigths after convergence
eta0 = task_info['sen']['2c']['eta0']
tauB = task_info['sen']['2c']['tauB']
targetB = task_info['targetB']
B0 = tauB * targetB
tau_update = task_info['sen']['2c']['tau_update']
eta = eta0 * tau_update / tauB
plot_weights(dend_mon, events, bursts, spikes,
[targetB, B0, eta, tauB, tau_update, smoothwin_],
taskdir)
plot_rasters(spksSE, bursts, targetB, isis, runtime_, taskdir)
else:
# calculate neurometric per neuron
events, bursts, singles, spikes, isis = spks2neurometric(
spksSE,
runtime,
settletime,
validburst,
smoothwin=smoothwin_,
raster=True)
plot_neurometric(events, bursts, spikes, stim1, stim2, stimtime,
(settletime_, runtime_), taskdir, smoothwin_)
plot_isis(isis, bursts, events, (settletime_, runtime_), taskdir)
# -------------------------------------
# Choice selection
# -------------------------------------
# population rates and downsample
originaltime = rateDE1.t / second
interptime = np.linspace(0, originaltime[-1],
originaltime[-1] * 100) # every 10 ms
fDE1 = interpolate.interp1d(
originaltime, rateDE1.smooth_rate(window='flat', width=smoothwin))
fDE2 = interpolate.interp1d(
originaltime, rateDE2.smooth_rate(window='flat', width=smoothwin))
fSE1 = interpolate.interp1d(
originaltime, rateSE1.smooth_rate(window='flat', width=smoothwin))
fSE2 = interpolate.interp1d(
originaltime, rateSE2.smooth_rate(window='flat', width=smoothwin))
rateDE = np.array([f(interptime) for f in [fDE1, fDE2]])
rateSE = np.array([f(interptime) for f in [fSE1, fSE2]])
# select the last half second of the stimulus
newdt = runtime_ / rateDE.shape[1]
settletimeidx = int(settletime_ / newdt)
dec_ival = np.array([(stimoff_ - 0.5) / newdt, stimoff_ / newdt],
dtype=int)
who_wins = rateDE[:, dec_ival[0]:dec_ival[1]].mean(axis=1)
# divide trls into preferred and non-preferred
pref_msk = np.argmax(who_wins)
poprates_dec = np.array([rateDE[pref_msk],
rateDE[~pref_msk]]) # 0: pref, 1: npref
poprates_sen = np.array([rateSE[pref_msk], rateSE[~pref_msk]])
results = {
'raw_data': {
'poprates_dec': poprates_dec[:, settletimeidx:],
'poprates_sen': poprates_sen[:, settletimeidx:],
'pref_msk': np.array([pref_msk]),
'last_muOUd': last_muOUd
},
'sim_state': np.zeros(1),
'computed': {
'events': events,
'bursts': bursts,
'singles': singles,
'spikes': spikes,
'isis': np.array(isis)
}
}
return results
