def setupLayerRN(params, neuronModel, cell_params, injectionPopulations, popPoissionNoiseSource, populationsRN):
#create a single RN population divided into virtual clusters one per VR
#this will be fed by the noise population and modulated by the relevant ratecoded neuron
#to create a rate coded population
numVR = params['NUM_VR']
rnClusterSize = int(params['CLUSTER_SIZE']) #* params['NETWORK_SCALE']
rnPopSize = rnClusterSize * numVR
popName = 'popRN'
popRN = spynnaker.Population(rnPopSize, neuronModel, cell_params, label=popName)
populationsRN.append(popRN)
#connect one random poisson neuron to each RN neuron
weight = params['WEIGHT_POISSON_TO_CLUSTER_RN']
delay = params['DELAY_POISSON_TO_CLUSTER_RN']
connections = utils.fromList_OneRandomSrcForEachTarget(popPoissionNoiseSource._size,popRN._size,weight,delay)
projPoissonToClusterRN = spynnaker.Projection(popPoissionNoiseSource, popRN, spynnaker.FromListConnector(connections), target='excitatory')
vr = 0
for injectionPopn in injectionPopulations:
connections = list()
for fromNeuronIdx in range(injectionPopn._size):
#connect the correct VR ratecode neuron in popRateCodeSpikes to corresponding subsection (cluster) of the RN population
weight = params['WEIGHT_RATECODE_TO_CLUSTER_RN']
firstIndex = vr * rnClusterSize
lastIndex = firstIndex + rnClusterSize - 1
connections += utils.fromList_SpecificNeuronToRange(fromNeuronIdx,firstIndex,lastIndex,weight,params['MIN_DELAY_RATECODE_TO_CLUSTER_RN'],params['MAX_DELAY_RATECODE_TO_CLUSTER_RN'])
vr = vr + 1
#after the last neuron in the current injection pop, create a projection to the RN
projRateToClusterRN = spynnaker.Projection(injectionPopn, popRN, spynnaker.FromListConnector(connections), target='excitatory')
print 'Added projection to RN of ', len(connections), " connections from injection pop ", injectionPopn.label, "(size ", injectionPopn._size,")"
def createIntraPopulationWTA(popn, numClusters, weight, delay, connectivity,
createSingleProjection):
allConnections = []
clusterSize = popn._size / numClusters
for i in range(numClusters):
for j in range(numClusters):
if i != j:
srcFirstIdx = i * clusterSize
srcLastIdx = srcFirstIdx + clusterSize - 1
targFirstIdx = j * clusterSize
targLastIdx = targFirstIdx + clusterSize - 1
connections = fromList_fromRangeToRange(
srcFirstIdx, srcLastIdx, targFirstIdx, targLastIdx, weight,
delay, delay, connectivity)
if (createSingleProjection):
allConnections += connections
else:
projInhibitory = spynnaker.Projection(
popn,
popn,
spynnaker.FromListConnector(connections),
target='inhibitory')
#print ' WTA connections', projInhibitory.getWeights()
if (createSingleProjection):
projInhibitory = spynnaker.Projection(
popn,
popn,
spynnaker.FromListConnector(allConnections),
target='inhibitory')
def setupLayerAN(params, settings, neuronModel, cell_params, popClassActivation, popPoissionNoiseSource, populationsPN, populationsAN,learning,projectionsPNAN):
#create an Association Neuron AN cluster population per class
#this will be fed by:
#1) PN clusters via plastic synapses
#2) Class activation to innervate the correct AN cluster for a given input
#3) laterally inhibit between AN clusters
numClasses = params['NUM_CLASSES']
anClusterSize = int(params['CLUSTER_SIZE']) #* params['NETWORK_SCALE']
for an in range(numClasses):
popName = 'popClusterAN_' + str(an) ;
popClusterAN = spynnaker.Population(anClusterSize, neuronModel, cell_params, label=popName)
populationsAN.append(popClusterAN)
#connect neurons in every PN popn to x% (e.g 50%) neurons in this AN cluster
for pn in range(len(populationsPN)):
if learning:
projLabel = 'Proj_PN' + str(pn) + '_AN' + str(an)
projClusterPNToClusterAN = connectClusterPNtoAN(params,populationsPN[pn],popClusterAN,float(settings['OBSERVATION_EXPOSURE_TIME_MS']),projLabel)
projectionsPNAN.append(projClusterPNToClusterAN) #keep handle to use later for saving off weights at end of learning
else:
#Without plasticity, create PNAN FromList connectors using weights saved during learning stage
connections = utils.loadListFromFile(getWeightsFilename(settings,'PNAN',pn,an))
#print 'Loaded weightsList[',pn,',',an,']',connections
tupleList = utils.createListOfTuples(connections) #new version only accepts list of tuples not list of lists
#print 'tupleList[',pn,',',an,']',tupleList
conn = spynnaker.FromListConnector(tupleList)
projClusterPNToClusterAN = spynnaker.Projection(populationsPN[pn], popClusterAN,conn, target='excitatory')
if learning:
#use the class activity input neurons to create correlated activity during learining in the corresponding class cluster
weight = params['WEIGHT_CLASS_EXCITATION_TO_CLUSTER_AN']
connections = utils.fromList_SpecificNeuronToAll(an,anClusterSize,weight,params['MIN_DELAY_CLASS_ACTIVITY_TO_CLUSTER_AN'],params['MAX_DELAY_CLASS_ACTIVITY_TO_CLUSTER_AN'])
projClassActivityToClusterAN = spynnaker.Projection(popClassActivation, popClusterAN, spynnaker.FromListConnector(connections), target='excitatory')
else: #testing
#send spikes on these outputs back to correct host port , these will be used to determine winner etc
anHostReceivePort = int(settings['AN_HOST_RECEIVE_PORT'])
ExternalDevices.activate_live_output_for(popClusterAN,port=anHostReceivePort)
#connect each AN cluster to inhibit every other AN cluster
utils.createInterPopulationWTA(populationsAN,params['WEIGHT_WTA_AN_AN'],params['DELAY_WTA_AN_AN'],float(params['CONNECTIVITY_WTA_AN_AN']))
#inhibit other non-corresponding class clusters
if learning:
weight = params['WEIGHT_CLASS_INHIBITION_TO_CLUSTER_AN']
for activeCls in range(numClasses):
connections = utils.fromList_SpecificNeuronToAll(activeCls,anClusterSize,weight,params['MIN_DELAY_CLASS_ACTIVITY_TO_CLUSTER_AN'],params['MAX_DELAY_CLASS_ACTIVITY_TO_CLUSTER_AN'])
for an in range(numClasses):
if an != activeCls:
projClassActivityToClusterAN = spynnaker.Projection(popClassActivation, populationsAN[an], spynnaker.FromListConnector(connections), target='inhibitory')
def connectClusterPNtoAN(params,popClusterPN,popClusterAN, observationExposureTimeMs, projLabel=''):
#Using custom Hebbian-style plasticity, connect neurons in specfied PN cluster to x% neurons in specified AN cluster
startWeightPNAN = float(params['STARTING_WEIGHT_PN_AN'])
delayPNAN = int(params['DELAY_PN_AN'])
connectivity = float(params['CONNECTIVITY_PN_AN'])
#STDP curve parameters
tau = float(params['STDP_TAU_PN_AN'])
wMin = float(params['STDP_WMIN_PN_AN'])
wMax = float(params['STDP_WMAX_PN_AN'])
gainScaling = float(params['STDP_SCALING_PN_AN'])
'''
#this setting was tuned for a 120ms learning window
#rescale it according to actual window used. ie for longer window, slow down learning rate
tweak = 120.0/float(observationExposureTimeMs)
gainScaling = gainScaling * tweak
print "Weight gain scaled by factor of ", tweak
'''
timingDependence = spynnaker.SpikePairRule(tau_plus=tau, tau_minus=tau, nearest=True)
weightDependence = spynnaker.AdditiveWeightDependence(w_min=wMin, w_max=wMax, A_plus=gainScaling, A_minus=-gainScaling)
stdp_model = spynnaker.STDPMechanism(timing_dependence = timingDependence, weight_dependence = weightDependence)
probConnector = spynnaker.FixedProbabilityConnector(connectivity, weights=startWeightPNAN, delays=delayPNAN, allow_self_connections=True)
projClusterPNToClusterAN = spynnaker.Projection(popClusterPN, popClusterAN,probConnector,synapse_dynamics = spynnaker.SynapseDynamics(slow = stdp_model), target='excitatory', label=projLabel)
return projClusterPNToClusterAN
def nealprojection(self, pre_neurons, post_neurons, connector_list,
inh_exc):
#projection method for SpiNNaker
conn_list = spinn.FromListConnector(connector_list)
spinn.Projection(pre_neurons,
post_neurons,
conn_list,
receptor_type=inh_exc)
def test(spikeTimes, trained_weights,label):
#spikeTimes = extractSpikes(sample)
runTime = int(max(max(spikeTimes)))+100
##########################################
sim.setup(timestep=1)
pre_pop = sim.Population(input_size, sim.SpikeSourceArray, {'spike_times': spikeTimes}, label="pre_pop")
post_pop = sim.Population(output_size, sim.IF_curr_exp , cell_params_lif, label="post_pop")
if len(trained_weights) > input_size:
weigths = [[0 for j in range(output_size)] for i in range(input_size)] #np array? size 1024x25
k=0
for i in range(input_size):
for j in range(output_size):
weigths[i][j] = trained_weights[k]
k += 1
else:
weigths = trained_weights
connections = []
#k = 0
for n_pre in range(input_size): # len(untrained_weights) = input_size
for n_post in range(output_size): # len(untrained_weight[0]) = output_size; 0 or any n_pre
#connections.append((n_pre, n_post, weigths[n_pre][n_post]*(wMax), __delay__))
connections.append((n_pre, n_post, weigths[n_pre][n_post]*(wMax)/max(trained_weights), __delay__)) #
#k += 1
prepost_proj = sim.Projection(pre_pop, post_pop, sim.FromListConnector(connections), synapse_type=sim.StaticSynapse(), receptor_type='excitatory') # no more learning !!
#inhib_proj = sim.Projection(post_pop, post_pop, sim.AllToAllConnector(), synapse_type=sim.StaticSynapse(weight=inhibWeight, delay=__delay__), receptor_type='inhibitory')
# no more lateral inhib
post_pop.record(['v', 'spikes'])
sim.run(runTime)
neo = post_pop.get_data(['v', 'spikes'])
spikes = neo.segments[0].spiketrains
v = neo.segments[0].filter(name='v')[0]
f1=pplt.Figure(
# plot voltage
pplt.Panel(v, ylabel="Membrane potential (mV)", xticks=True, yticks=True, xlim=(0, runTime+100)),
# raster plot
pplt.Panel(spikes, xlabel="Time (ms)", xticks=True, yticks=True, markersize=2, xlim=(0, runTime+100)),
title='Test with label ' + str(label),
annotations='Test with label ' + str(label)
)
f1.save('plot/'+str(trylabel)+str(label)+'_test.png')
f1.fig.texts=[]
print("Weights:{}".format(prepost_proj.get('weight', 'list')))
weight_list = [prepost_proj.get('weight', 'list'), prepost_proj.get('weight', format='list', with_address=False)]
#predict_label=
sim.end()
return spikes
def load(path, filename, input_layer):
layers = load_assembly(path, filename)
projections = []
for i in range(len(layers) - 1):
filepath = os.path.join(path, layers[i + 1].label)
assert os.path.isfile(filepath), \
"Connections were not found at specified location."
with warnings.catch_warnings():
warnings.simplefilter('ignore')
warnings.warn('deprecated', UserWarning)
if i == 0:
projections.append(
sim.Projection(input_layer, layers[i + 1],
sim.FromFileConnector(filepath)))
#projections.append(sim.Projection(input_layer, layers[i+1], sim.OneToOneConnector(), synapse_type=sim.StaticSynapse(weight=5, delay=1)))
else:
projections.append(
sim.Projection(layers[i], layers[i + 1],
sim.FromFileConnector(filepath)))
return layers, projections
def createInterPopulationWTA(populations, weight, delay, connectivity):
probConnector = spynnaker.FixedProbabilityConnector(
connectivity,
weights=weight,
delays=delay,
allow_self_connections=True)
for i in range(len(populations)):
for j in range(len(populations)):
if i != j:
projInhibitory = spynnaker.Projection(populations[i],
populations[j],
probConnector,
target='inhibitory')
def setupLayerAN(params, settings, neuronModel, cell_params, popClassActivation, popPoissionNoiseSource, populationsPN, populationsAN,learning,projectionsPNAN):
#create an Association Neuron AN cluster population per class
#this will be fed by:
#1) PN clusters via plastic synapses
#2) Class activation to innervate the correct AN cluster for a given input
#3) laterally inhibit between AN clusters
numClasses = params['NUM_CLASSES']
anClusterSize = params['CLUSTER_SIZE'] * params['NETWORK_SCALE']
for an in range(numClasses):
popName = 'popClusterAN_' + str(an) ;
popClusterAN = spynnaker.Population(anClusterSize, neuronModel, cell_params, label=popName)
populationsAN.append(popClusterAN)
#connect neurons in every PN popn to x% (e.g 50%) neurons in this AN cluster
for pn in range(len(populationsPN)):
if learning:
projLabel = 'Proj_PN' + str(pn) + '_AN' + str(an)
projClusterPNToClusterAN = connectClusterPNtoAN(params,populationsPN[pn],popClusterAN,projLabel)
projectionsPNAN.append(projClusterPNToClusterAN) #keep handle to use later for saving off weights at end of learning
else:
#Without plasticity, create PNAN FromList connectors using weights saved during learning stage
connections = utils.loadListFromFile(getWeightsFilename(settings,'PNAN',pn,an))
#print 'Loaded weightsList[',pn,',',an,']',connections
projClusterPNToClusterAN = spynnaker.Projection(populationsPN[pn], popClusterAN,spynnaker.FromListConnector(connections), target='excitatory')
if learning:
#use the class activity input neurons to create correlated activity during learining in the corresponding class cluster
weight = params['WEIGHT_CLASS_ACTIVITY_TO_CLUSTER_AN']
connections = utils.fromList_SpecificNeuronToAll(an,anClusterSize,weight,params['MIN_DELAY_CLASS_ACTIVITY_TO_CLUSTER_AN'],params['MAX_DELAY_CLASS_ACTIVITY_TO_CLUSTER_AN'])
projClassActivityToClusterAN = spynnaker.Projection(popClassActivation, popClusterAN, spynnaker.FromListConnector(connections), target='excitatory')
#connect each AN cluster to inhibit every other AN cluster
utils.createInterPopulationWTA(populationsAN,params['WEIGHT_WTA_AN_AN'],params['DELAY_WTA_AN_AN'],float(params['CONNECTIVITY_WTA_AN_AN']))
def setupProjection_PN_KC(pn_population,kc_population):
connectionList = list() # Connection list between PN and KC
for each_kc_cell in xrange(NUM_KC_CELLS):
count = 6
selectedCells = random.sample(xrange(NUM_PN_CELLS),count)
for each_pn_cell in selectedCells:
single_coonection = (each_pn_cell,each_kc_cell)
connectionList.append(single_coonection)
pnkcProjection = spynnaker.Projection(pn_population,
kc_population,
spynnaker.FromListConnector(connectionList),
spynnaker.StaticSynapse(weight=WEIGHT_PN_KC, delay=DELAY_PN_KC))
return pnkcProjection
def connectClusterPNtoAN(params,popClusterPN,popClusterAN, projLabel=''):
#Using custom Hebbian-style plasticity, connect neurons in specfied PN cluster to x% neurons in specified AN cluster
startWeightPNAN = float(params['STARTING_WEIGHT_PN_AN'])
delayPNAN = int(params['DELAY_PN_AN'])
connectivity = float(params['CONNECTIVITY_PN_AN'])
#STDP curve parameters
tau = float(params['STDP_TAU_PN_AN'])
wMin = float(params['STDP_WMIN_PN_AN'])
wMax = float(params['STDP_WMAX_PN_AN'])
gainScaling = float(params['STDP_SCALING_PN_AN'])
timingDependence = spynnaker.SpikePairRule(tau_plus=tau, tau_minus=tau, nearest=True)
weightDependence = spynnaker.AdditiveWeightDependence(w_min=wMin, w_max=wMax, A_plus=gainScaling, A_minus=-gainScaling)
stdp_model = spynnaker.STDPMechanism(timing_dependence = timingDependence, weight_dependence = weightDependence)
probConnector = spynnaker.FixedProbabilityConnector(connectivity, weights=startWeightPNAN, delays=delayPNAN, allow_self_connections=True)
projClusterPNToClusterAN = spynnaker.Projection(popClusterPN, popClusterAN,probConnector,synapse_dynamics = spynnaker.SynapseDynamics(slow = stdp_model), target='excitatory', label=projLabel)
return projClusterPNToClusterAN
def createInterPopulationWTA(populations, weight, delay, connectivity):
probConnector = spynnaker.FixedProbabilityConnector(
connectivity,
weights=weight,
delays=delay,
allow_self_connections=True)
numPopulations = len(populations)
if numPopulations > 1:
popCreatedCount = 0
totalProjections = numPopulations * (numPopulations - 1)
for i in range(numPopulations):
for j in range(numPopulations):
if i != j:
projInhibitory = spynnaker.Projection(populations[i],
populations[j],
probConnector,
target='inhibitory')
popCreatedCount = popCreatedCount + 1
print 'Created ', popCreatedCount, ' of ', totalProjections, ' (', 100 * float(
popCreatedCount) / float(totalProjections), '%)'
#print ' WTA connections', projInhibitory.getWeights()
else:
print 'No inter-projections created , only one population provided'
def setupProjection_PN_KC(pn_population, kc_population):
WEIGHT_PN_KC = 5
DELAY_PN_KC = 1.0
NUM_KC_CELLS = 2000
NUM_PN_CELLS = 784
connectionList = list() # Build up a connection list between PN and KC
for each_kc_cell in xrange(NUM_KC_CELLS):
count = random.randint(
5, 7) # How many pn_cells will connect to this kc_cell
selectedCells = random.sample(
xrange(NUM_PN_CELLS), count) # Index of randomly selected PN cells
for each_pn_cell in selectedCells:
single_coonection = (each_pn_cell, each_kc_cell)
connectionList.append(single_coonection)
pnkcProjection = spynnaker.Projection(
pn_population, kc_population,
spynnaker.FromListConnector(connectionList),
spynnaker.StaticSynapse(weight=WEIGHT_PN_KC, delay=DELAY_PN_KC))
return pnkcProjection
weight_to_spike = 2.0
delay = 17
loopConnections = list()
for i in range(0, nNeurons):
singleConnection = (i, ((i + 1) % nNeurons), weight_to_spike, delay)
loopConnections.append(singleConnection)
injectionConnection = [(0, 0, weight_to_spike, 1)]
spikeArray = {'spike_times': [[0, 1050]]}
populations.append(p.Population(nNeurons, p.IF_curr_exp, cell_params_lif,
label='pop_1'))
populations.append(p.Population(1, p.SpikeSourceArray, spikeArray,
label='inputSpikes_1'))
projections.append(p.Projection(populations[0], populations[0],
p.FromListConnector(loopConnections)))
projections.append(p.Projection(populations[1], populations[0],
p.FromListConnector(injectionConnection)))
populations[0].record_v()
populations[0].record_gsyn()
populations[0].record()
p.run(runtime)
v = None
gsyn = None
spikes = None
v = populations[0].get_v(compatible_output=True)
gsyn = populations[0].get_gsyn(compatible_output=True)
testpop = p.Population(200, p.IF_curr_exp, cell_params_lif, label='ifcurr')
testpop.record()
inppop = p.Population(100,
p.SpikeSourcePoisson, {
'rate': 120,
'duration': 8000
},
label="poisson_PLOT")
inppop.stream()
inp2pop = p.Population(100, p.SpikeSourcePoisson, {
'rate': 50,
'duration': 8000
})
proj = p.Projection(inppop, myopop,
p.OneToOneConnector(weights=1.00, delays=1.0))
proj2 = p.Projection(inp2pop, myopop2,
p.OneToOneConnector(weights=1.0, delays=1.0))
#projout=p.Projection(myopop,testpop,p.OneToOneConnector(weights=0.5,delays=1.0))
#emptypop = p.Population(200, p.IF_curr_exp, cell_params_lif, label='dummy')
#dummyproj = p.Projection(emptypop, testpop, p.OneToOneConnector(weights=1.0,delays=1.0))
#inppop.record()
#myopop.record()
poispops = []
for k, (mini, maxi) in [(0xFEFFFE21, (1220., 2880.))]: # angle measurement
# (0xFEFFFE03,(-20.,1000.)), # displacement
# (0xFEFFFE07,(-20.,1000.)), # should not go negative, but can overflow
database_notify_port_num=19996)
sim.external_devices.activate_live_output_for(input1,
database_notify_host="localhost",
database_notify_port_num=19998)
timing_rule = sim.SpikePairRule(tau_plus=20.0,
tau_minus=20.0,
A_plus=0.5,
A_minus=0.5)
weight_rule = sim.AdditiveWeightDependence(w_max=25.0, w_min=0.0)
stdp_model = sim.STDPMechanism(timing_dependence=timing_rule,
weight_dependence=weight_rule,
weight=2.0,
delay=1)
stdp_projection = sim.Projection(pre_pop,
post_pop,
sim.OneToOneConnector(),
synapse_type=stdp_model)
input_projection1 = sim.Projection(input1,
pre_pop,
sim.OneToOneConnector(),
synapse_type=sim.StaticSynapse(weight=5,
delay=1))
pre_pop.record(["spikes", "v"])
post_pop.record(["spikes", "v"])
simtime = 100
k = PyKeyboard()
def receive_spikes(label, time, neuron_ids):
teachpop = p.Population(nn_teach, p.SpikeSourcePoisson, {
'rate': 100,
'duration': duration
})
#teachpop.record()
postpop = p.Population(nn_post, p.IF_cond_exp, cell_params)
#postpop.record()
connteach = p.OneToOneConnector(
weights=0.0,
delays=1.0) #FromListConnector([(i,i,0.0,1.0) for i in range(nn)])
#randconn = p.FixedProbabilityConnector(0.5,weights=0.0,delays=1.0)
#noisesyn = p.Projection(noisepop,postpop,connteach,target='inhibitory')
teachsyn = p.Projection(teachpop, postpop, connteach, target='inhibitory')
# plasticity
wdep_grcpcsynapsis = p.AdditiveWeightDependence(w_min=0.0,
w_max=0.5,
A_plus=0.0015,
A_minus=0.0018)
tdep_grcpcsynapsis = p.SpikePairRuleSinAdd(tau_minus=50.,
tau_plus=50.,
delay=100.0,
nearest=False) # delay 70-100
stdp_grcpcsynapsis = p.STDPMechanism(timing_dependence=tdep_grcpcsynapsis,
weight_dependence=wdep_grcpcsynapsis,
voltage_dependence=None)
syndyn_grcpcsynapsis = p.SynapseDynamics(slow=stdp_grcpcsynapsis)
source = sim.Population(n_atoms, model, cell_params, label='source_layer')
target = sim.Population(n_atoms, model, cell_params, label='target_layer')
target.set_constraint(PlacerChipAndCoreConstraint(0, 1))
# Define learning
# Plastic Connections between pre_pop and post_pop
stdp_model = sim.STDPMechanism(
timing_dependence=sim.SpikePairRule(tau_plus=20.,
tau_minus=20.0,
nearest=True),
weight_dependence=sim.AdditiveWeightDependence(w_min=0,
w_max=0.9,
A_plus=0.02,
A_minus=0.02))
structure_model_w_stdp = sim.StructuralMechanism(stdp_model=stdp_model)
# Define connections
plastic_projection = sim.Projection(
source,
source,
sim.FixedNumberPreConnector(32),
synapse_dynamics=sim.SynapseDynamics(slow=structure_model_w_stdp),
label="plastic_projection")
# Add a sprinkle of Poisson noise
# Add some spatial pattern to be repeated
# Start Simulation
# Recover provenance
# End simulation
try:
import pyNN.spiNNaker as p
except Exception as e:
import spynnaker8 as p
# set up the tools
p.setup(timestep=1.0, min_delay=1.0, max_delay=32.0)
# set up the virtual chip coordinates for the motor
connected_chip_coords = {'x': 0, 'y': 0}
link = 4
populations = list()
projections = list()
input_population = p.Population(6, p.SpikeSourcePoisson(rate=10))
control_population = p.Population(6, p.IF_curr_exp())
motor_device = p.Population(
6, p.external_devices.MunichMotorDevice(spinnaker_link_id=0))
p.Projection(input_population,
control_population,
p.OneToOneConnector(),
synapse_type=p.StaticSynapse(weight=5.0))
p.external_devices.activate_live_output_to(control_population, motor_device)
p.run(1000)
p.end()
'v_reset': -70.0,
'v_rest': -65.0,
'v_thresh': -55.0
}
neurons = sim.Population(100, sim.IF_cond_exp(**cell_params))
inputs = sim.Population(100, sim.SpikeSourcePoisson(rate=0.0))
# set input firing rates as a linear function of cell index
input_firing_rates = np.linspace(0.0, 1000.0, num=inputs.size)
inputs.set(rate=input_firing_rates)
# create one-to-one connections
wiring = sim.OneToOneConnector()
static_synapse = sim.StaticSynapse(weight=0.1, delay=2.0)
connections = sim.Projection(inputs, neurons, wiring, static_synapse)
# configure recording
neurons.record('spikes')
# run simulation
sim_duration = 10.0 # seconds
sim.run(sim_duration * 1000.0)
# retrieve recorded data
spike_counts = neurons.get_spike_counts()
print(spike_counts)
output_firing_rates = np.array(
[value for (key, value) in sorted(spike_counts.items())]) / sim_duration
sim.end()
# declare python code when received spikes for a timer tick
def receive_spikes(label, time, neuron_ids):
for neuron_id in neuron_ids:
print("Received spike at time {} from {}-{}"
"".format(time, label, neuron_id))
p.setup(timestep=1.0)
p1 = p.Population(1, p.IF_curr_exp(), label="pop_1")
input_injector = p.Population(1, p.external_devices.SpikeInjector(),
label=INJECTOR_LABEL)
# set up python live spike connection
live_spikes_connection = p.external_devices.SpynnakerLiveSpikesConnection(
receive_labels=[RECEIVER_LABEL])
# register python receiver with live spike connection
live_spikes_connection.add_receive_callback(RECEIVER_LABEL, receive_spikes)
input_proj = p.Projection(input, p1, p.OneToOneConnector(),
p.StaticSynapse(weight=5, delay=3))
p1.record(["spikes", "v"])
p.run(50)
neo = p1.get_data(["spikes", "v"])
spikes = neo.segments[0].spiketrains
print(spikes)
v = neo.segments[0].filter(name='v')[0]
print(v)
post_pop = sim.Population(1, model, cell_params)
# Stimulating populations
pre_times = [i for i in range(pre_phase, sim_time, time_between_pairs)]
post_times = [i for i in range(post_phase, sim_time, time_between_pairs)]
pre_stim = sim.Population(
1, sim.SpikeSourceArray(spike_times=[pre_times]))
post_stim = sim.Population(
1, sim.SpikeSourceArray(spike_times=[post_times]))
weight = 0.035
# Connections between spike sources and neuron populations
ee_connector = sim.OneToOneConnector()
sim.Projection(
pre_stim, pre_pop, ee_connector, receptor_type='excitatory',
synapse_type=sim.StaticSynapse(weight=weight))
sim.Projection(
post_stim, post_pop, ee_connector, receptor_type='excitatory',
synapse_type=sim.StaticSynapse(weight=weight))
# Plastic Connection between pre_pop and post_pop
stdp_model = sim.STDPMechanism(
timing_dependence=sim.SpikePairRule(
tau_plus=16.7, tau_minus=33.7, A_plus=0.005, A_minus=0.005),
weight_dependence=sim.AdditiveWeightDependence(
w_min=0.0, w_max=0.0175), weight=start_w)
projections.append(sim.Projection(
pre_pop, post_pop, sim.OneToOneConnector(),
synapse_type=stdp_model))
def run_test(w_list, cell_para, spike_source_data):
#Du.set_trace()
pop_list = []
p.setup(timestep=1.0, min_delay=1.0, max_delay=3.0)
#input poisson layer
input_size = w_list[0].shape[0]
#print w_list[0].shape[0]
#print w_list[1].shape[0]
list = []
for j in range(input_size):
list.append(spike_source_data[j])
pop_in = p.Population(input_size, p.SpikeSourceArray,
{'spike_times': list})
pop_list.append(pop_in)
#for j in range(input_size):
#pop_in[j].spike_times = spike_source_data[j]
#pop_in = p.Population(input_size, p.SpikeSourceArray, {'spike_times' : []})
#for j in range(input_size):
#pop_in[j].spike_times = spike_source_data[j]
#pop_list.append(pop_in)
#count =0
#print w_list[0].shape[0]
for w in w_list:
input_size = w.shape[0]
#count = count+1
#print count
output_size = w.shape[1]
#pos_w = np.copy(w)
#pos_w[pos_w < 0] = 0
#neg_w = np.copy(w)
#neg_w[neg_w > 0] = 0
conn_list_exci = []
conn_list_inhi = []
#k_size=in_size-out_size+1
for x_ind in range(input_size):
for y_ind in range(output_size):
weights = w[x_ind][y_ind]
#for i in range(w.shape[1]):
if weights > 0:
conn_list_exci.append((x_ind, y_ind, weights, 1.))
elif weights < 0:
conn_list_inhi.append((x_ind, y_ind, weights, 1.))
#print output_size
pop_out = p.Population(output_size, p.IF_curr_exp, cell_para)
if len(conn_list_exci) > 0:
p.Projection(pop_in,
pop_out,
p.FromListConnector(conn_list_exci),
target='excitatory')
if len(conn_list_inhi) > 0:
p.Projection(pop_in,
pop_out,
p.FromListConnector(conn_list_inhi),
target='inhibitory')
#p.Projection(pop_in, pop_out, p.AllToAllConnector(weights = pos_w), target='excitatory')
#p.Projection(pop_in, pop_out, p.AllToAllConnector(weights = neg_w), target='inhibitory')
pop_list.append(pop_out)
pop_in = pop_out
pop_out.record()
run_time = np.ceil(np.max(spike_source_data)[0] / 1000.) * 1000
#print run_time
p.run(run_time)
spikes = pop_out.getSpikes(compatible_output=True)
return spikes
def train_snn(### Settings
data_dir = "data/X_train_zied.npy",
cls_dir = "data/y_train_zied.npy",
data = "load", # pass data as argument
cls = "load", # pass labels as argument
save = True, # True to save all parameters of the network
randomness = True,
reverse_src_del = False,
use_old_weights = False,
rand_data = False,
### Parameters
n_training = 2, # How many times the samples will be iterated
ts = 1., # Timestep of Spinnaker (ms)
trial_num = 10, # How many samples (trials) from data used
# Network
n_feature = 80, # Number of features (= 4 features * 20 neurons)
# Weights
wei_src_enc = .2, # From Source Array at input to Encoding Layer(Exc)
wei_enc_filt = .6, # From Encoding Layer to Filtering Layer Exc neurons (Exc)
wei_filt_inh = 0.03, # From Filtering Layer Inh neurons to Exc neurons (Inh)
wei_init_stdp = .0, # From Filtering Layer Exc neurons to Output Layer (Exc)
wei_cls_exc = 0.9, # From Output Layer Exc neurons to Inh neurons (Exc)
wei_cls_inh = 50,#0.1,#,10 # From Output Layer Inh neurons to Exc neurons (Inh)
wei_source_outp = 10., # From Source Array at output to Output Layer Exc neurons (Exc)
wei_noise_poi = 0.02,
# Delays
del_init_stdp = 1.,
del_source_outp = 1.,
del_noise_poi = 1.,
# Connection Probabilities
prob_filt_inh = .4, # Prob of connectivity inhibitory connections at FilT_Layer
prob_stdp = 1., # Prob of STDP connections
prob_output_inh = .7, # Prob of inhibitory connections at Output Layer
prob_noise_poi_conn = 0.02,
## STDP Parameters
tau_pl = 5.,
stdp_w_max = 0.4, # default 0.4
stdp_w_min = 0.0, # default 0.0
stdp_A_pl = 2,#0.02,# 0.01, # default 0.01 (below 0.01 weights don't change)
# => minus in order to get symmetric curve
# Data Extraction
scale_data = 2.): # Scale features into [0-scale_data] range
# BUG fix:
# n_feature is somehow a tuple
# try:
# trial_num = trial_num[0]
# except Exception as e:
# print("\n\n\n EXCEPTION TRIGGERED !!!! \n\n\n")
# pass
############################################################################
## Function Definitions
############################################################################
def gaussian(x, mu, sig):
return np.float16(np.exp(-np.power(x - mu, 2.) /
(2 * np.power(sig, 2.))))
def calc_pop_code(feature, rng1, rng2, num):
interval=np.float(rng2-rng1)/num
means=np.arange(rng1+interval, rng2+interval, interval)
pop_code=[gaussian(feature,mu,0.025) for mu in means]
return pop_code
def PoissonTimes2(t_str=0., t_end=100., rate=10., seed=1.):
times = [t_str]
rng = np.random.RandomState(seed=seed)
cont = True
while cont == True:
t_next = np.floor(times[-1] + 1000. * next_spike_times(rng,rate))
if t_next < t_end - 30:
times.append(t_next[0])
else:
cont=False
return times[1:]
def PoissonTimes(t_str=0., t_end=100., rate=10., seed=1., max_rate=0):
if rate>0:
interval = (t_end - t_str + 0.) / rate
# Add additional reverse_src_del
if reverse_src_del == True:
times = np.arange(t_str + 30, t_end - 40, interval)
# add reverse proportional delay
rev_del = np.ceil(max_rate / rate)
if rev_del != np.inf:
times += rev_del
else:
times = np.arange(t_str + 30, t_end - 40, interval)
return list(times)
else:
return []
def next_spike_times(rng,rate):
return -np.log(1.0-rng.rand(1)) / rate
def ismember(a, b):
b=[b]
bind = {}
for i, elt in enumerate(b):
if elt not in bind:
bind[elt] = i
aa=[bind.get(itm, -1) for itm in a]
return sum(np.array(aa)+1.)
def get_data(trial_num, test_num=10):
# trial_num: number of training samples
# test_num: number of test samples
pass
def rand_sample_of_train_set(n):
# n: number of features
# Return: np.array containing n samples of the training set
X = np.load(data_dir)
y = np.load(cls_dir)
idx = np.random.randint(len(X), size=n)
return X[idx], y[idx]
############################################################################
## Parameters
############################################################################
# Load training data
# only load n_rand_data features of training set
if rand_data == True:
data, cls = rand_sample_of_train_set(trial_num)
# load all features of training set
else: # load data if its not in passed as fuct argument
if data == "load" and cls == "load":
data = np.load(data_dir)
cls = np.load(cls_dir)
# Simulation Parameters
trial_num = len(cls) # How many samples (trials) from data will be presented
#n_training = 1 # How many times the samples will be iterated
n_trials = n_training * trial_num # Total trials
time_int_trials = 200. # (ms) Time to present each trial data
SIM_TIME = n_trials * time_int_trials # Total simulation time (ms)
#ts = 1. # Timestep of Spinnaker (ms)
min_del = ts
max_del = 144 * ts
p.setup(timestep=ts, min_delay=min_del, max_delay=max_del)
## Neuron Numbers
#n_feature = 80 # Number of features (= 4 features * 20 neurons)
# => 20 neuros: resolution of encoding
n_pop = data.shape[1] #4 # Number of neurons in one population (X dim)
n_cl = 2 # Number of classes at the output
## Connection Parameters
# Weights
# wei_src_enc = .2 # From Source Array at input to Encoding Layer(Exc)
# wei_enc_filt = .6 # From Encoding Layer to Filtering Layer Exc neurons (Exc)
# wei_filt_inh = 0.03 # From Filtering Layer Inh neurons to Exc neurons (Inh)
# wei_init_stdp = .0 # From Filtering Layer Exc neurons to Output Layer (Exc)
# wei_cls_exc = 0.9 # From Output Layer Exc neurons to Inh neurons (Exc)
# wei_cls_inh = 10 # 0.1 # From Output Layer Inh neurons to Exc neurons (Inh)
# wei_source_outp = 10. # From Source Array at output to Output Layer Exc neurons (Exc)
# wei_noise_poi = 0.02
# Delays
if randomness == True: # if True: calculate "del_src_enc" (randomly) new
# if False: load previously saved "del_src_enc"
if reverse_src_del == True:
# calc delays erversly proportional to feature value
del_src_enc = np.zeros(n_feature*n_pop)
else:
del_src_enc = [int(np.random.randint(n_pop)+1)
for _ in range(n_feature*n_pop)]
np.save("output_files/del_src_enc.npy", del_src_enc)
else:
#del_src_enc = np.load("output_files/del_src_enc.npy")
del_src_enc = np.ones(n_feature*n_pop).astype(int) #[1 for _ in range(n_feature*n_pop)]
del_enc_filt = ts
del_filt_inh = ts
# del_init_stdp = 1.
del_cls_exc = ts
del_cls_inh = ts
# del_source_outp = 1.
# del_noise_poi = 1.
# Firing Rates
noise_poi_rate = 10.
max_fr_input = 100. # maximum firing rate at the input layer
max_fr_rate_output = 20. # Maximum firing rate at output (supervisory signal)
## Connection Probabilities
# prob_filt_inh = .4 # Prob of connectivity inhibitory connections at FilT_Layer
# prob_stdp = 1. # Prob of STDP connections
# prob_output_inh = .7 # Prob of inhibitory connections at Output Layer
# prob_noise_poi_conn = 0.02
## STDP Parameters
# tau_pl = 0.3 # (0.2 - 0.3 works)
tau_min = tau_pl # default tau_pl
# stdp_w_max = 0.4 # default 0.4
# stdp_w_min = 0.0 # default 0.0
# stdp_A_pl = 0.01 # default 0.01 (below 0.01 weights don't change)
stdp_A_min = -stdp_A_pl # default - stdp_A_pl
# => minus in order to get symmetric curve
## Neuron Parameters
cell_params_lif = {'cm': 0.25,# 0.25,
'i_offset': 0.0,
'tau_m': 20.,
'tau_refrac': 2.0,
'tau_syn_E': 5.0,
'tau_syn_I': 5.0,
'v_reset': -70.0,
'v_rest': -65.0,
'v_thresh': -50#-50
}
############################################################################
## Data Extraction
############################################################################
## Extract Feature Data
# scale_data = 2. # Scale features into [0-scale_data] range
r,c = np.shape(data)
data_rates = np.reshape(data, (1, r*c))[0]
# Threshold (to keep spikes in range)
thr_data_plus = 30
thr_data_minus = -10
#dd = [d if d<thr_data_plus else thr_data_plus for d in data_rates]
#dd = [d if d>thr_data_minus else thr_data_minus for d in dd]
# Shift and normalize data
#dd2 = np.array(dd) - min(dd)
dd2 = np.array(data_rates) - min(data_rates)
dd2 = dd2 / max(dd2) * 2
new_data_rates = []
for r in dd2:
new_data_rates += calc_pop_code(r, 0., scale_data, n_feature /
(n_pop + 0.0))
data_rates = list(max_fr_input*np.array(new_data_rates))
## Extract Class Data
# load class vector
#cls = np.load(path_y)
cls = np.reshape(cls, (len(cls),1)) # create col vector
r_cl, c_cl = np.shape(cls)
#cls = list(np.reshape(cls, (1, r_cl * c_cl))[0] - 1)
cls = list(np.reshape(cls, (1, r_cl * c_cl))[0])
## The class and rate information to be used during the simulation
outputs = n_training * cls[0:trial_num] # positiv, ints
poi_rate = n_training * data_rates[0:trial_num * n_feature]
## Save parameters to be used in test
parameter_dict = {"n_feature":n_feature, "n_pop":n_pop,"n_cl":n_cl,
"wei_src_enc":wei_src_enc, "wei_enc_filt":wei_enc_filt,
"wei_filt_inh":wei_filt_inh, "wei_cls_exc":wei_cls_exc,
"wei_cls_inh":wei_cls_inh, "del_enc_filt":del_enc_filt,
"del_init_stdp":del_init_stdp, "del_cls_exc":del_cls_exc,
"del_cls_inh":del_cls_inh, "trial_num":trial_num,
"time_int_trials":time_int_trials, "scale_data":scale_data,
"ts":ts,"max_fr_input":max_fr_input,
"max_fr_rate_output":max_fr_rate_output,
"noise_poi_rate":noise_poi_rate, "max_fr_input":max_fr_input,
"max_fr_rate_output":max_fr_rate_output, "prob_filt_inh":prob_filt_inh,
"prob_stdp":prob_stdp, "prob_output_inh":prob_output_inh,
"prob_noise_poi_conn":prob_noise_poi_conn, "tau_pl":tau_pl,
"stdp_w_max":stdp_w_max, "stdp_w_min":stdp_w_min, "stdp_A_pl":stdp_A_pl,
"wei_noise_poi":wei_noise_poi, "del_noise_poi":del_noise_poi,
"thr_data_plus":thr_data_plus, "thr_data_minus":thr_data_minus
}
if save == True:
np.save("output_files/parameters1",parameter_dict)
np.save("output_files/parameters2",del_src_enc)
############################################################################
## Create populations for different layers
############################################################################
poi_layer = []
enc_layer = []
filt_layer_exc = []
out_layer_exc = []
out_layer_inh = []
out_spike_source = []
# Calculate spike times at the input using the rate information coming from features
spike_times = [[] for i in range(n_feature)]
for i in range(n_trials):
t_st = i * time_int_trials
t_end = t_st + time_int_trials
ind = i * n_feature
for j in range(n_feature):
times = PoissonTimes(t_st, t_end, poi_rate[ind+j],
np.random.randint(100), max_rate=max(poi_rate))
for t in times:
spike_times[j].append(t)
if randomness == True: # if True: calculate "spike_times" (randomly) new
# uf False: load previously saved "spike_times"
np.save('output_files/spike_times_train.npy', spike_times)
else:
spike_times = np.load('output_files/spike_times_train.npy')
# Calculate spike times at the output (as supervisory signal)
out_spike_times=[[] for i in range(n_cl)]
for i in range(n_trials):
t_st = i * time_int_trials
t_end = t_st + time_int_trials
ind = outputs[i]
times = PoissonTimes(t_st, t_end, max_fr_rate_output,
np.random.randint(100))
for t in times:
out_spike_times[int(ind)].append(t)
if randomness == True: # if True: calculate "out_spike_times" (randomly) new
# uf False: load previously saved "out_spike_times"
np.save('output_files/out_spike_times.npy', out_spike_times)
else:
out_spike_times = np.load('output_files/out_spike_times.npy')
# Spike source of input layer
spike_source=p.Population(n_feature,
p.SpikeSourceArray,
{'spike_times':spike_times},
label='spike_source')
# Spike source of output layer (Supervisory signal)
for i in range(n_cl):
out_spike_source.append(p.Population(1, p.SpikeSourceArray,
{'spike_times':[out_spike_times[i]]}, label='out_spike_source'))
# Encoding layer and Filtering Layer definitions
enc_layer = p.Population(n_feature * n_pop,
p.IF_curr_exp,
cell_params_lif,
label='enc_layer')
filt_layer = p.Population(n_feature * n_pop,
p.IF_curr_exp,
cell_params_lif,
label='filt_layer')
# Excitatory and Inhibitory population definitions at the output
for i in range(n_cl):
out_layer_exc.append(p.Population(n_pop,
p.IF_curr_exp,
cell_params_lif,
label='out_layer_exc{}'.format(i)))
out_layer_inh.append(p.Population(n_pop,
p.IF_curr_exp,
cell_params_lif,
label='out_layer_inh{}'.format(i)))
out_layer_exc[i].record()
# Noisy poisson population at the input
poisson_input = p.Population(n_pop * 2,
p.SpikeSourcePoisson,
{"rate":noise_poi_rate})
# Record Spikes
enc_layer.record()
filt_layer.record()
#enc_layer.initialize('v',p.RandomDistribution('uniform',[-51.,-69.]))
#filt_layer.initialize('v',p.RandomDistribution('uniform',[-51.,-69.]))
############################################################################
## Projections
############################################################################
## Connection List from Spike Source Array to Encoding Layer
conn_inp_enc=[]
for i in range(n_feature):
ind=i*n_pop
for j in range(n_pop):
conn_inp_enc.append([i,ind+j,wei_src_enc,del_src_enc[ind+j]])
if save == True:
np.save("output_files/conn_inp_enc",conn_inp_enc)
## Connection List for Filtering Layer Inhibitory
if randomness == True: # if True: calculate conn_filt_inh (randomly) new
# uf False: load previously saved conn_filt_inh
conn_filt_inh=[]
for i in range(n_feature):
rng1=i*n_pop
rng2=rng1+n_pop
inp=range(rng1,rng2)
outp=range(0,rng1)+range(rng2,n_feature*n_pop)
for ii in inp:
for jj in outp:
if prob_filt_inh>np.random.rand():
conn_filt_inh.append([ii,jj,wei_filt_inh,del_filt_inh])
if save == True:
np.save('output_files/conn_filt_inh.npy', conn_filt_inh)
else:
conn_filt_inh = np.load('output_files/conn_filt_inh.npy')
## STDP Connection List
if randomness == True: # if True: calculate conn_stdp_list (randomly) new
# uf False: load previously saved conn_stdp_list
conn_stdp_list=[[] for i in range(n_cl)]
for i in range(n_cl): # For each population at output layer
if use_old_weights == True:
cl_weights = np.load(
"output_files/stdp_weights{}.npy".format(i))
w = 0
for ii in range(n_pop * n_feature): # For each neuron in filtering layer
for jj in range(n_pop): # For each neuron in each population of output layer
if prob_stdp > np.random.rand(): # If the prob of connection is satiesfied
# Make the connection
if use_old_weights == True:
conn_stdp_list[i].append([ii,
jj,
cl_weights[w],
del_init_stdp])
w += 1
else:
conn_stdp_list[i].append([ii,
jj,
wei_init_stdp,
del_init_stdp])
if use_old_weights == False or save == True:
np.save('output_files/conn_stdp_list.npy', conn_stdp_list)
else:
conn_stdp_list = np.load('output_files/conn_stdp_list.npy')
## Output Layer Inhibitory Connection List
if randomness == True: # if True: calculate conn_stdp_list (randomly) new
# uf False: load previously saved conn_stdp_list
conn_output_inh = [[] for i in range(n_cl) for j in range(n_cl) if i!=j]
c = 0
for i in range(n_cl):
for j in range(n_cl):
if i != j:
for ii in range(n_pop):
for jj in range(n_pop):
if prob_output_inh > np.random.rand():
conn_output_inh[c].append([ii,
jj,
wei_cls_inh,
del_cls_inh])
c += 1
if save == True:
np.save("output_files/conn_output_inh.npy",conn_output_inh)
else:
conn_output_inh = np.load("output_files/conn_output_inh.npy")
## Spike Source to Encoding Layer
p.Projection(spike_source, enc_layer,
p.FromListConnector(conn_inp_enc))
## Encoding Layer to Filtering Layer
p.Projection(enc_layer, filt_layer,
p.OneToOneConnector(weights=wei_enc_filt,
delays=del_enc_filt))
## Filtering Layer Inhibitory
p.Projection(filt_layer, filt_layer,
p.FromListConnector(conn_filt_inh),
target="inhibitory")
## STDP Connection between Filtering Layer and Output Layer
timing_rule = p.SpikePairRule(tau_plus=tau_pl,
tau_minus=tau_min)
weight_rule = p.AdditiveWeightDependence(w_max=stdp_w_max,
w_min=stdp_w_min,
A_plus=stdp_A_pl,
A_minus=stdp_A_min)
stdp_model = p.STDPMechanism(timing_dependence=timing_rule,
weight_dependence=weight_rule)
# STDP connection
stdp_proj = []
for j in range(n_cl):
stdp_proj.append(
p.Projection(filt_layer,out_layer_exc[j],
p.FromListConnector(conn_stdp_list[j]),
synapse_dynamics = p.SynapseDynamics(slow=stdp_model)))
## Connection between Output Layer neurons
c = 0
for i in range(n_cl):
p.Projection(out_layer_exc[i], out_layer_inh[i],
p.OneToOneConnector(weights=wei_cls_exc,
delays=del_cls_exc))
iter_array=[j for j in range(n_cl) if j!=i]
for j in iter_array:
p.Projection(out_layer_exc[i], out_layer_exc[j],
p.FromListConnector(conn_output_inh[c]),
target="inhibitory")
c += 1
## Spike Source Array to Output
for i in range(n_cl):
p.Projection(out_spike_source[i],
out_layer_exc[i],
p.AllToAllConnector(weights=wei_source_outp,
delays=del_source_outp))
iter_array = [j for j in range(n_cl) if j != i]
for j in iter_array:
p.Projection(out_spike_source[i],
out_layer_exc[j],
p.AllToAllConnector(weights=wei_source_outp,
delays=del_source_outp),
target="inhibitory")
#for i in range(n_cl):
# p.Projection(out_spike_source[i], out_layer_exc[i], p.AllToAllConnector\
# (weights=wei_source_outp, delays=del_source_outp))
# p.Projection(out_spike_source[i], out_layer_exc[1-i], p.AllToAllConnector\
# (weights=wei_source_outp, delays=del_source_outp),target="inhibitory")
## Noisy poisson connection to encoding layer
if randomness == True: # if True: connect noise to network
# if False: don't use noise in network
p.Projection(poisson_input, enc_layer,
p.FixedProbabilityConnector(p_connect=prob_noise_poi_conn,
weights=wei_noise_poi,
delays=del_noise_poi))
############################################################################
## Simulation
############################################################################
p.run(SIM_TIME)
Enc_Spikes = enc_layer.getSpikes()
Filt_Exc_Spikes = filt_layer.getSpikes()
Out_Spikes = [[] for i in range(n_cl)]
for i in range(n_cl):
Out_Spikes[i] = out_layer_exc[i].getSpikes()
wei = []
for i in range(n_cl):
ww = stdp_proj[i].getWeights()
if save == True:
np.save("output_files/stdp_weights{}".format(i), ww)
wei.append(ww)
p.end()
############################################################################
## Plot
############################################################################
## Plot 1: Encoding Layer Raster Plot
if 0:
pylab.figure()
pylab.xlabel('Time (ms)')
pylab.ylabel('Neuron ID')
pylab.title('Encoding Layer Raster Plot')
pylab.hold(True)
pylab.plot([i[1] for i in Enc_Spikes], [i[0] for i in Enc_Spikes], ".b")
pylab.hold(False)
#pylab.axis([-10,c*SIM_TIME+100,-1,numInp+numOut+numInp+3])
pylab.show()
## Plot 2-1: Filtering Layer Raster Plot
if 0:
pylab.figure()
pylab.xlabel('Time (ms)')
pylab.ylabel('Neuron ID')
pylab.title('Filtering Layer Raster Plot')
pylab.plot([i[1] for i in Filt_Exc_Spikes],
[i[0] for i in Filt_Exc_Spikes], ".b")
#pylab.axis([-10,c*SIM_TIME+100,-1,numInp+numOut+numInp+3])
pylab.show()
## Plot 2-2: Filtering Layer Layer Raster Plot
if 0:
pylab.figure()
pylab.xlabel('Time (ms)')
pylab.ylabel('Neuron ID')
pylab.title('Filtering Layer Layer Raster Plot')
pylab.hold(True)
pylab.plot([i[1] for i in Filt_Exc_Spikes],
[i[0] for i in Filt_Exc_Spikes], ".b")
time_ind=[i*time_int_trials for i in range(len(outputs))]
for i in range(len(time_ind)):
pylab.plot([time_ind[i],time_ind[i]],[0,2000],"r")
pylab.hold(False)
#pylab.axis([-10,c*SIM_TIME+100,-1,numInp+numOut+numInp+3])
pylab.show()
## Plot 3-1: Output Layer Raster Plot
if 0:
pylab.figure()
pylab.xlabel('Time (ms)')
pylab.ylabel('Neuron')
pylab.title('Output Layer Raster Plot')
pylab.hold(True)
c=0
for array in Out_Spikes:
pylab.plot([i[1] for i in array], [i[0]+c for i in array], ".b")
c+=0.2
pylab.hold(False)
pylab.axis([-10,SIM_TIME+100,-1,n_pop+3])
pylab.show()
## Plot 4: STDP WEIGHTS
if 1:
pylab.figure()
pylab.xlabel('Weight ID')
pylab.ylabel('Weight Value')
pylab.title('STDP weights at the end')
#pylab.title('STDP weights at the end' + ' (trail_num=' + str(trial_num) + ')')
pylab.hold(True)
for i in range(n_cl):
pylab.plot(wei[i])
pylab.hold(False)
pylab.axis([-10,n_pop*n_feature*n_pop*0.5+10,-stdp_w_max,2*stdp_w_max])
str_legend=["To Cl {}".format(i+1) for i in range(n_cl)]
pylab.legend(str_legend)
#pylab.show()
fname = 'plots/weights_1.png'
while True:
if os.path.isfile(fname): # if file already exists
new_num = int(fname.split('.')[0].split('_')[1]) + 1
fname = fname.split('_')[0] + '_' +str(new_num) + '.png'
# if len(fname) == 19:
# new_num = int(fname.split('.')[0][-1]) + 1
# fname = fname.split('.')[0][:-1] + str(new_num) + '.png'
# elif len(fname) == 20:
# new_num = int(fname.split('.')[0][-2:]) + 1
# fname = fname.split('.')[0][:-2] + str(new_num) + '.png'
# else:
# new_num = int(fname.split('.')[0][-3:]) + 1
# fname = fname.split('.')[0][:-3] + str(new_num) + '.png'
else:
pylab.savefig(fname)
break
#pylab.figure()
#pylab.xlabel('Weight ID')
#pylab.ylabel('Weight Value')
#pylab.title('STDP weights at the end')
#pylab.hold(True)
#pylab.plot(wei[0], "b")
#pylab.plot(wei[1], "g")
#pylab.hold(False)
#pylab.axis([-10, n_pop * n_feature * n_pop * 0.5 + 10,
# -stdp_w_max, 2 * stdp_w_max])
#pylab.legend(['To Cl 1','To Cl 2'])
#pylab.show()
## Plot 5: Spike Source Spiking Times
if 0:
pylab.figure()
pylab.hold(True)
pylab.plot(out_spike_times[0],
[1 for i in range(len(out_spike_times[0]))],"x")
pylab.plot(out_spike_times[1],
[1.05 for i in range(len(out_spike_times[1]))],"x")
pylab.hold(False)
pylab.title("Spike Source Spiking Times")
pylab.axis([-100,SIM_TIME+100,-2,3])
pylab.show()
## Calculate spiking activity of each neuron to each class inputs
sum_filt=[[0 for i in range(n_feature*n_pop)] for j in range(n_cl)]
sum_filt=np.array(sum_filt)
for i in range(n_trials):
t_st = i * time_int_trials
t_end = t_st + time_int_trials
cl = outputs[i]
for n,t in Filt_Exc_Spikes:
if t >= t_st and t < t_end:
sum_filt[int(cl),int(n)] = sum_filt[int(cl), int(n)] + 1
a4=sum_filt[0]
b4=sum_filt[1]
thr=20
diff_vec=np.abs(a4 - b4)
diff_thr=[i if i>thr else 0. for i in diff_vec]
diff_ind=[i for i in range(len(diff_thr)) if diff_thr[i]!=0]
if save == True:
np.save("output_files/diff_ind_filt",diff_ind)
diff2 = a4 - b4
diff_thr2=[i if i > thr or i <- thr else 0. for i in diff2]
diff_ind2=[i for i in range(len(diff_thr2)) if diff_thr2[i] != 0]
if save == True:
np.save("output_files/diff_ind_filt2",diff_ind2)
np.save("output_files/diff_thr2",diff_thr2)
## Plot 6: Total Spiking Activity of Neurons at Decomposition Layer for Each Class
if 0:
a4=sum_filt[0]
b4=sum_filt[1]
pylab.figure()
pylab.hold(True)
pylab.plot(a4,"b")
pylab.plot(b4,"r")
pylab.xlabel('Neuron ID')
pylab.ylabel('Total Firing Rates Through Trials')
pylab.title("Total Spiking Activity of Neurons at Decomposition Layer ",
"for Each Class")
pylab.hold(False)
pylab.legend(["Activity to AN1","Activity to AN2"])
pylab.show()
import pyNN.spiNNaker as sim
sim.setup()
p1 = sim.Population(3, sim.SpikeSourceArray, {"spike_times": [1.0, 2.0, 3.0]})
p2 = sim.Population(3, sim.SpikeSourceArray, {"spike_times": [[10.0], [20.0], [30.0]]})
p3 = sim.Population(4, sim.IF_cond_exp, {})
sim.Projection(p2, p3, sim.FromListConnector([
(0, 0, 0.1, 1.0), (1, 1, 0.1, 1.0), (2, 2, 0.1, 1.0)]))
#sim.Projection(p1, p3, sim.FromListConnector([(0, 3, 0.1, 1.0)])) # works if this line is added
sim.run(100.0)
def train(spikeTimes,untrained_weights=None):
organisedStim = {}
labelSpikes = []
#spikeTimes = generate_data()
#for j in range(5):
# labelSpikes
#labelSpikes[label] = [(input_len-1)*v_co+1,(input_len-1)*v_co*2+1,(input_len-1)*v_co*3+1,]
if untrained_weights == None:
untrained_weights = RandomDistribution('uniform', low=wMin, high=wMaxInit).next(input_size*output_size)
#untrained_weights = RandomDistribution('normal_clipped', mu=0.1, sigma=0.05, low=wMin, high=wMaxInit).next(input_size*output_size)
untrained_weights = np.around(untrained_weights, 3)
#saveWeights(untrained_weights, 'untrained_weightssupmodel1traj')
print ("init!")
print "length untrained_weights :", len(untrained_weights)
if len(untrained_weights)>input_size:
training_weights = [[0 for j in range(output_size)] for i in range(input_size)] #np array? size 1024x25
k=0
for i in range(input_size):
for j in range(output_size):
training_weights[i][j] = untrained_weights[k]
k += 1
else:
training_weights = untrained_weights
connections = []
for n_pre in range(input_size): # len(untrained_weights) = input_size
for n_post in range(output_size): # len(untrained_weight[0]) = output_size; 0 or any n_pre
connections.append((n_pre, n_post, training_weights[n_pre][n_post], __delay__)) #index
runTime = int(max(max(spikeTimes))/3)+100
#####################
sim.setup(timestep=1)
#def populations
layer1=sim.Population(input_size,sim.SpikeSourceArray, {'spike_times': spikeTimes},label='inputspikes')
layer2=sim.Population(output_size,sim.IF_curr_exp,cellparams=cell_params_lif,label='outputspikes')
#supsignal=sim.Population(output_size,sim.SpikeSourceArray, {'spike_times': labelSpikes},label='supersignal')
#def learning rule
stdp = sim.STDPMechanism(
weight=untrained_weights,
#weight=0.02, # this is the initial value of the weight
#delay="0.2 + 0.01*d",
timing_dependence=sim.SpikePairRule(tau_plus=tauPlus, tau_minus=tauMinus,A_plus=aPlus, A_minus=aMinus),
#weight_dependence=sim.MultiplicativeWeightDependence(w_min=wMin, w_max=wMax),
weight_dependence=sim.AdditiveWeightDependence(w_min=wMin, w_max=wMax),
#weight_dependence=sim.AdditiveWeightDependence(w_min=0, w_max=0.4),
dendritic_delay_fraction=1.0)
#def projections
#stdp_proj = sim.Projection(layer1, layer2, sim.FromListConnector(connections), synapse_type=stdp)
stdp_proj = sim.Projection(layer1, layer2, sim.AllToAllConnector(), synapse_type=stdp)
inhibitory_connections = sim.Projection(layer2, layer2, sim.AllToAllConnector(allow_self_connections=False),
synapse_type=sim.StaticSynapse(weight=inhibWeight, delay=__delay__),
receptor_type='inhibitory')
#stim_proj = sim.Projection(supsignal, layer2, sim.OneToOneConnector(),
# synapse_type=sim.StaticSynapse(weight=stimWeight, delay=__delay__))
layer1.record(['spikes'])
layer2.record(['v','spikes'])
#supsignal.record(['spikes'])
sim.run(runTime)
print("Weights:{}".format(stdp_proj.get('weight', 'list')))
weight_list = [stdp_proj.get('weight', 'list'), stdp_proj.get('weight', format='list', with_address=False)]
neo = layer2.get_data(["spikes", "v"])
spikes = neo.segments[0].spiketrains
v = neo.segments[0].filter(name='v')[0]
#neostim = supsignal.get_data(["spikes"])
#spikestim = neostim.segments[0].spiketrains
neoinput= layer1.get_data(["spikes"])
spikesinput = neoinput.segments[0].spiketrains
plt.close('all')
pplt.Figure(
pplt.Panel(spikesinput,xticks=True, yticks=True, markersize=2, xlim=(0,runTime),xlabel='(a) Spikes of Input Layer'),
#pplt.Panel(spikestim, xticks=True, yticks=True, markersize=2, xlim=(0,runTime),xlabel='(c) Spikes of Supervised Layer'),
pplt.Panel(spikes, xticks=True, xlabel="(b) Spikes of Output Layer", yticks=True, markersize=2, xlim=(0,runTime)),
pplt.Panel(v, ylabel="Membrane potential (mV)", xticks=True, yticks=True, xlim=(0,runTime),xlabel='(c) Membrane Potential of Output Layer\nTime (ms)'),
title="Two Training",
annotations="Twoway Training"
).save('SNN_DVS_un/plot_for_twoway/'+str(trylabel)+'_training.png')
#plt.hist(weight_list[1], bins=100)
plt.close('all')
plt.hist([weight_list[1][0:input_size], weight_list[1][input_size:input_size*2], weight_list[1][input_size*2:]], bins=20, label=['neuron 0', 'neuron 1', 'neuron 2'], range=(0, wMax))
plt.title('weight distribution')
plt.xlabel('Weight value')
plt.ylabel('Weight count')
#plt.show()
#plt.show()
sim.end()
return weight_list[1]
def estimate_kb(cell_params_lif):
cell_para = copy.deepcopy(cell_params_lif)
random.seed(0)
p.setup(timestep=1.0, min_delay=1.0, max_delay=16.0)
run_s = 10.
runtime = 1000. * run_s
max_rate = 1000.
ee_connector = p.OneToOneConnector(weights=1.0, delays=2.0)
pop_list = []
pop_output = []
pop_source = []
x = np.arange(0., 1.01, 0.1)
count = 0
trail = 10
for i in x:
for j in range(trail): #trails for average
pop_output.append(p.Population(1, p.IF_curr_exp, cell_para))
poisson_spikes = mu.poisson_generator(i * max_rate, 0, runtime)
pop_source.append(
p.Population(1, p.SpikeSourceArray,
{'spike_times': poisson_spikes}))
p.Projection(pop_source[count],
pop_output[count],
ee_connector,
target='excitatory')
pop_output[count].record()
count += 1
count = 0
for i in x:
cell_para['i_offset'] = i
pop_list.append(p.Population(1, p.IF_curr_exp, cell_para))
pop_list[count].record()
count += 1
pop_list[count - 1].record_v()
p.run(runtime)
rate_I = np.zeros(count)
rate_P = np.zeros(count)
rate_P_max = np.zeros(count)
rate_P_min = np.ones(count) * 1000.
for i in range(count):
spikes = pop_list[i].getSpikes(compatible_output=True)
rate_I[i] = len(spikes) / run_s
for j in range(trail):
spikes = pop_output[i * trail +
j].getSpikes(compatible_output=True)
spike_num = len(spikes) / run_s
rate_P[i] += spike_num
if spike_num > rate_P_max[i]:
rate_P_max[i] = spike_num
if spike_num < rate_P_min[i]:
rate_P_min[i] = spike_num
rate_P[i] /= trail
'''
#plot_spikes(spikes, 'Current = 10. mA')
plt.plot(x, rate_I, label='current',)
plt.plot(x, rate_P, label='Poisson input')
plt.fill_between(x, rate_P_min, rate_P_max, facecolor = 'green', alpha=0.3)
'''
x0 = np.where(rate_P > 1.)[0][0]
x1 = 4
k = (rate_P[x1] - rate_P[x0]) / (x[x1] - x[x0])
'''
plt.plot(x, k*(x-x[x0])+rate_P[x0], label='linear')
plt.legend(loc='upper left', shadow=True)
plt.grid('on')
plt.show()
'''
p.end()
return k, x[x0], rate_P[x0]
'tau_m': 20, 'tau_refrac': 2, 'tau_syn_E': 50, 'tau_syn_I': 5, 'v_reset': -70, 'v_rest': -65, 'v_thresh': -55 } sim.set_number_of_neurons_per_core(sim.IF_curr_exp, 100) G1_1= sim.Population(1,sim.IF_curr_exp(**cell_params_lif), label="G1_1") G2_2= sim.Population(10,sim.IF_curr_exp(**cell_params_lif), label="G2_2") GEN1_3= sim.Population(1,simSpikeSourceArray(spike_times=[0,8], label="GEN1_3") input_G1_1GEN1_3=sim.Projection(G1_1,GEN1_3, sim.OneToOneConnector(), synapse_type.StaticSynapse(weight=1.25, delay=0)) input_G1_1G2_2=sim.Projection(G1_1,G2_2, sim.OneToOneConnector(), synapse_type.StaticSynapse(weight=0.8, delay=0)) input_G2_2G1_1=sim.Projection(G2_2,G1_1, sim.OneToOneConnector(), synapse_type.StaticSynapse(weight=0.1, delay=0)) G1_1.record(["spikes","v","gsyn_exc"]) simtime =50 sim.run(simtime) neo = G1_1.get_data(variables=["spikes","v","gsync_exc"]) spikes = neo.segments[0].spiketrains print spikes v = neo.segments[0].filter(name='v')[0] print v gsync = neo.segments[0].filter(name='gsync_exc')[0] print gsync
label="expoisson")
# +-------------------------------------------------------------------+
# | Creation of connections |
# +-------------------------------------------------------------------+
# Connection parameters
JEE = 3.
# Connection type between noise poisson generator and excitatory populations
ee_connector = sim.OneToOneConnector()
# Noise projections
sim.Projection(INoisePre,
pre_pop,
ee_connector,
receptor_type='excitatory',
synapse_type=sim.StaticSynapse(weight=JEE * 0.05))
sim.Projection(INoisePost,
post_pop,
ee_connector,
receptor_type='excitatory',
synapse_type=sim.StaticSynapse(weight=JEE * 0.05))
# Additional Inputs projections
for i in range(len(IAddPre)):
sim.Projection(IAddPre[i],
pre_pop,
ee_connector,
receptor_type='excitatory',
synapse_type=sim.StaticSynapse(weight=JEE * 0.05))
def test_snn(randomness = False,
data_dir = "data/X_test_zied.npy",
cls_dir = "data/y_test_zied.npy",
data = "load", # pass data as argument
cls = "load"): # pass labels as argument
###############################################################################
## Function Definitions
###############################################################################
def gaussian(x, mu, sig):
return np.float16(np.exp(-np.power(x - mu, 2.) / (2 * np.power(sig, 2.))))
def calc_pop_code(feature, rng1, rng2, num):
interval = np.float(rng2 - rng1) / num
means = np.arange(rng1 + interval,rng2 + interval, interval)
pop_code = [gaussian(feature, mu, 0.025) for mu in means]
return pop_code
def PoissonTimes2(t_str=0., t_end=100., rate=10., seed=1.):
times = [t_str]
rng = np.random.RandomState(seed=seed)
cont = True
while cont == True:
t_next = np.floor(times[-1] + 1000. * next_spike_times(rng, rate))
if t_next < t_end - 30:
times.append(t_next[0])
else:
cont = False
return times[1:]
def PoissonTimes(t_str=0., t_end=100., rate=10., seed=1.):
if rate > 0:
interval = (t_end - t_str+0.) / rate
times = np.arange(t_str + 30, t_end - 40, interval)
return list(times)
else:
return []
def next_spike_times(rng,rate):
return -np.log(1.0 - rng.rand(1)) / rate
def ismember(a, b):
b = [b]
bind = {}
for i, elt in enumerate(b):
if elt not in bind:
bind[elt] = i
aa=[bind.get(itm, -1) for itm in a]
return sum(np.array(aa) + 1.)
###############################################################################
## Parameters
###############################################################################
# Load Parameter
parameters = np.load("output_files/parameters1.npy")
parameters = parameters.item()
# Load test data
if data == "load" and cls == "load":
data = np.load(data_dir)
cls = np.load(cls_dir)
# Simulation Parameters
trial_num = parameters["trial_num"] # How many samples (trials) from data will be presented
n_trials = len(cls)#10#20 #int(trial_num) # Total trials
time_int_trials = parameters["time_int_trials"] # (ms) Time to present each trial data
SIM_TIME = n_trials * time_int_trials # Total simulation time (ms)
ts = parameters["ts"] # Timestep of Spinnaker (ms)
min_del = ts
max_del = 144 * ts
p.setup(timestep=ts, min_delay=min_del, max_delay=max_del)
## Neuron Numbers
n_feature = parameters["n_feature"] # Number of features
n_pop = parameters["n_pop"] # Number of neurons in one population
n_cl = parameters["n_cl"] # Number of classes at the output
## Connection Parameters
# Weights
wei_src_enc = parameters["wei_src_enc"] # From Source Array at input to Encoding Layer(Exc)
wei_enc_filt = parameters["wei_enc_filt"] # From Encoding Layer to Filtering Layer Exc neurons (Exc)
wei_filt_inh = parameters["wei_filt_inh"] # From Filtering Layer Inh neurons to Exc neurons (Inh)
wei_cls_exc = parameters["wei_cls_exc"] # From Output Layer Exc neurons to Inh neurons (Exc)
wei_cls_inh = parameters["wei_cls_inh"] # From Output Layer Inh neurons to Exc neurons (Inh)
wei_noise_poi = parameters["wei_noise_poi"]
# Delays
del_src_enc = np.load("output_files/parameters2.npy")
del_enc_filt = parameters["del_enc_filt"]
del_init_stdp = parameters["del_init_stdp"]
del_cls_exc = parameters["del_cls_exc"]
del_cls_inh = parameters["del_cls_inh"]
del_noise_poi = parameters["del_noise_poi"]
# Firing Rates
noise_poi_rate = parameters["noise_poi_rate"]
max_fr_input = parameters["max_fr_input"] # maximum firing rate at the input layer
max_fr_rate_output = parameters["max_fr_rate_output"] # Maximum firing rate at output (supervisory signal)
## Connection Probabilities
prob_filt_inh = parameters["prob_filt_inh"] # Prob of connectivity inhi-connections at Filtering Layer
prob_stdp = parameters["prob_stdp"] # Probability of STDP connections
prob_output_inh = parameters["prob_output_inh"] # Prob of inhi-connections at Output Layer
prob_noise_poi_conn = parameters["prob_noise_poi_conn"]
## STDP Parameters
tau_pl = parameters["tau_pl"] #5
tau_min = tau_pl
stdp_w_max = parameters["stdp_w_max"]
stdp_w_min = parameters["stdp_w_min"]
stdp_A_pl = parameters["stdp_A_pl"]
stdp_A_min = -stdp_A_pl # minus in order to get symmetric curve
## Neuron Parameters
cell_params_lif = {'cm': 1.,
'i_offset': 0.0,
'tau_m': 20.,
'tau_refrac': 2.0,
'tau_syn_E': 5.0,
'tau_syn_I': 5.0,
'v_reset': -70.0,
'v_rest': -65.0,
'v_thresh': -65.0
}
###############################################################################
## Data Extraction
###############################################################################
## Extract Feature Data
scale_data = parameters["scale_data"] # Scale features into [0-scale_data] range
#data = np.load("features_without_artifact.npy")
#data = np.load('X_test.npy')
r, c = np.shape(data)
# Threshold (to keep spikes amplitudes in range)
thr_data_plus = parameters["thr_data_plus"]
thr_data_minus = parameters["thr_data_minus"]
data_rates = np.reshape(data, (1, r * c))[0]
# Shift an normalize the data
#dd = [d if d<thr_data_plus else thr_data_plus for d in data_rates]
#dd = [d if d>thr_data_minus else thr_data_minus for d in dd]
#dd2 = np.array(dd) - min(dd)
#dd2 = dd2 / max(dd2) * 2
dd2 = np.array(data_rates) - min(data_rates)
dd2 = dd2 / max(dd2) * 2
new_data_rates = []
for r in dd2:
new_data_rates += calc_pop_code(r, 0., scale_data,
n_feature / (n_pop + 0.0))
data_rates = list(max_fr_input * np.array(new_data_rates))
## Extract Class Data
#cls = np.load("classes_without_artifact.npy")
#cls = np.load("y_test.npy")
cls = cls.reshape(len(cls), 1)
r_cl, c_cl = np.shape(cls)
cls = list(np.reshape(cls, (1, r_cl * c_cl))[0])
outputs = cls[:n_trials]
poi_rate = data_rates[:n_feature * n_trials]
t1 = 0#70
t2 = int(t1 + n_trials)
outputs = cls[t1:t2]
poi_rate = data_rates[t1 * n_feature:n_feature * t2]
###############################################################################
## Create populations for different layers
###############################################################################
poi_layer = []
enc_layer = []
filt_layer_exc = []
out_layer_exc = []
out_layer_inh = []
# Calculate poisson spike times for features
spike_times = [[] for i in range(n_feature)]
for i in range(n_trials):
t_st = i * time_int_trials
t_end = t_st + time_int_trials
ind = i * n_feature
for j in range(n_feature):
times = PoissonTimes(t_st, t_end, poi_rate[ind+j],
np.random.randint(100))
for t in times:
spike_times[j].append(t)
if randomness == True: # if True: calculate "spike_times" (randomly) new
# if False: load previously saved "spike_times"
np.save('output_files/spike_times_test.npy', spike_times)
else:
spike_times = np.load('output_files/spike_times_test.npy')
# Spike source of input layer
spike_source = p.Population(n_feature,
p.SpikeSourceArray,
{'spike_times':spike_times},
label='spike_source')
enc_layer = p.Population(n_feature * n_pop,
p.IF_curr_exp,
cell_params_lif,
label='enc_layer')
filt_layer = p.Population(n_feature * n_pop,
p.IF_curr_exp,
cell_params_lif,
label='filt_layer')
#filt_layer_inh=p.Population(n_feature*n_pop, p.IF_curr_exp, cell_params_lif, label='filt_layer_inh')
for i in range(n_cl):
out_layer_exc.append(p.Population(n_pop,
p.IF_curr_exp,
cell_params_lif,
label='out_layer_exc{}'.format(i)))
out_layer_inh.append(p.Population(n_pop,
p.IF_curr_exp,
cell_params_lif,
label='out_layer_inh{}'.format(i)))
out_layer_exc[i].record()
poisson_input = p.Population(n_pop * 2,
p.SpikeSourcePoisson,
{"rate":noise_poi_rate})
enc_layer.record()
filt_layer.record()
###############################################################################
## Projections
###############################################################################
## Connection List from Spike Source Array to Encoding Layer
conn_inp_enc = np.load("output_files/conn_inp_enc.npy")
#Connection List for Filtering Layer Inhibitory
conn_filt_inh = np.load("output_files/conn_filt_inh.npy")
## STDP Connection List
conn_stdp_list = np.load("output_files/conn_stdp_list.npy")
diff_ind = np.load("output_files/diff_ind_filt.npy")
diff_ind2 = np.load("output_files/diff_ind_filt2.npy")
diff_thr2 = np.load("output_files/diff_thr2.npy")
c1 = 0
for cls_list in conn_stdp_list:
c2 = 0
cls_wei = np.load("output_files/stdp_weights{}.npy".format(c1))
mx = max(cls_wei)
for conn in cls_list:
# if ismember(diff_ind,conn[0]):
if (ismember(diff_ind2,conn[0]) and
np.sign(c1-0.5) * np.sign(diff_thr2[int(conn[0])]) == -1.):
# conn[2]=0.08*cls_wei[c2]/mx
conn[2] = 0.08#*diff_thr2[conn[0]]/36.
# conn[2]=2.*cls_wei[c2]
c2 += 1
c1 += 1
conn_stdp_list = list(conn_stdp_list)
## Output Layer Inhibitory Connection List
conn_output_inh = np.load("output_files/conn_output_inh.npy")
## Spike Source to Encoding Layer
p.Projection(spike_source,enc_layer,p.FromListConnector(conn_inp_enc))
## Encoding Layer to Filtering Layer
p.Projection(enc_layer, filt_layer,
p.OneToOneConnector(weights=wei_enc_filt,
delays=del_enc_filt))
## Filtering Layer Inhibitory
p.Projection(filt_layer,filt_layer,
p.FromListConnector(conn_filt_inh),
target="inhibitory")
## STDP Connection between Filtering Layer and Output Layer
stdp_proj = []
for j in range(n_cl):
stdp_proj.append(p.Projection(filt_layer, out_layer_exc[j],
p.FromListConnector(conn_stdp_list[j])))
## Connection between Output Layer neurons
c = 0
for i in range(n_cl):
p.Projection(out_layer_exc[i], out_layer_inh[i],
p.OneToOneConnector(weights=wei_cls_exc,
delays=del_cls_exc))
iter_array = [j for j in range(n_cl) if j != i]
for j in iter_array:
p.Projection(out_layer_inh[i], out_layer_exc[j],
p.FromListConnector(conn_output_inh[c]),
target="inhibitory")
c+=1
## Noisy poisson connection to encoding layer
if randomness == True: # if True: connect noise to network
# if False: don't use noise in network
p.Projection(poisson_input,
enc_layer,
p.FixedProbabilityConnector(p_connect=prob_noise_poi_conn,
weights=wei_noise_poi,
delays = del_noise_poi))
###############################################################################
## Simulation
###############################################################################
p.run(SIM_TIME)
Enc_Spikes = enc_layer.getSpikes()
Filt_Exc_Spikes = filt_layer.getSpikes()
#Filt_Inh_Spikes = filt_layer_inh.getSpikes()
Out_Spikes = [[] for i in range(n_cl)]
for i in range(n_cl):
Out_Spikes[i] = out_layer_exc[i].getSpikes()
p.end()
###############################################################################
## Plot
###############################################################################
## Plot 1
if 0:
pylab.figure()
pylab.xlabel('Time (ms)')
pylab.ylabel('Neuron ID')
pylab.title('Encoding Layer Raster Plot')
pylab.hold(True)
pylab.plot([i[1] for i in Enc_Spikes], [i[0] for i in Enc_Spikes], ".b")
pylab.hold(False)
#pylab.axis([-10,c*SIM_TIME+100,-1,numInp+numOut+numInp+3])
pylab.show()
## Plot 2-1
if 0:
pylab.figure()
pylab.xlabel('Time (ms)')
pylab.ylabel('Neuron ID')
pylab.title('Filtering Layer Raster Plot')
pylab.plot([i[1] for i in Filt_Exc_Spikes], [i[0] for i in Filt_Exc_Spikes], ".b")
#pylab.axis([-10,c*SIM_TIME+100,-1,numInp+numOut+numInp+3])
pylab.show()
## Plot 2-2
pylab.figure()
pylab.xlabel('Time (ms)')
pylab.ylabel('Neuron ID')
pylab.title('Filtering Layer Raster Plot')
pylab.hold(True)
pylab.plot([i[1] for i in Filt_Exc_Spikes], [i[0] for i in Filt_Exc_Spikes], ".b")
time_ind=[i*time_int_trials for i in range(len(outputs))]
for i in range(len(time_ind)):
pylab.plot([time_ind[i],time_ind[i]],[0,2000],"r")
pylab.hold(False)
#pylab.axis([-10,c*SIM_TIME+100,-1,numInp+numOut+numInp+3])
pylab.show()
## Plot 3-1
if 0:
pylab.figure()
pylab.xlabel('Time (ms)')
pylab.ylabel('Neuron ID')
pylab.title('Association Layer Raster Plot\nTest for Trial Numbers {}-{}'.format(t1,t2))
pylab.hold(True)
c=0
for array in Out_Spikes:
pylab.plot([i[1] for i in array], [i[0]+c for i in array], ".b")
c+=0.2
time_cls=[j*time_int_trials+i for j in range(len(outputs)) for i in range(int(time_int_trials))]
cls_lb=[outputs[j]+0.4 for j in range(len(outputs)) for i in range(int(time_int_trials))]
time_ind=[i*time_int_trials for i in range(len(outputs))]
for i in range(len(time_ind)):
pylab.plot([time_ind[i],time_ind[i]],[0,10],"r")
#pylab.plot(time_cls,cls_lb,".")
pylab.hold(False)
pylab.axis([-10,SIM_TIME+100,-1,n_pop+2])
pylab.show()
## Plot 3-2
pylab.figure()
pylab.xlabel('Time (ms)')
pylab.ylabel('Neuron ID')
pylab.title(('Association Layer Raster Plot\n',
'Test for Samples {}-{}').format(t1,t2))
pylab.hold(True)
pylab.plot([i[1] for i in Out_Spikes[0]],
[i[0] for i in Out_Spikes[0]],
".b")
pylab.plot([i[1] for i in Out_Spikes[1]],
[i[0] + 0.2 for i in Out_Spikes[1]],
".r")
time_ind = [i * time_int_trials for i in range(len(outputs))]
for i in range(len(time_ind)):
pylab.plot([time_ind[i], time_ind[i]], [0,n_pop], "k")
#pylab.plot(time_cls,cls_lb,".")
pylab.hold(False)
pylab.axis([-10, SIM_TIME+100, -1, n_pop + 2])
pylab.legend(["AN1","AN2" ])
pylab.show()
sum_output = [[] for i in range(n_cl)]
for i in range(n_trials):
t_st = i * time_int_trials
t_end = t_st + time_int_trials
for j in range(n_cl):
sum_output[j].append(np.sum(
[1 for n, t in Out_Spikes[j] if t >= t_st and t < t_end])
)
## Plot 4
if 0:
# pylab.figure()
# pylab.hold(True)
# pylab.plot(sum_output[0], "b.")
# pylab.plot(sum_output[1], "r.")
# out_cl0 = [i for i in range(len(outputs)) if outputs[i] == 0]
# out_cl1 = [i for i in range(len(outputs)) if outputs[i] == 1]
# pylab.plot(out_cl0,[-2 for i in range(len(out_cl0))], "xb")
# pylab.plot(out_cl1,[-2 for i in range(len(out_cl1))], "xr")
# pylab.hold(False)
# pylab.title("Total spikes at each AN population for each trial")
# pylab.xlabel("Trials")
# pylab.ylabel("Firing Rate")
# pylab.legend(["Cl0","Cl1","Winning Cl 0", "Winning Cl 1"])
# pylab.axis([-2, n_trials + 2, -4, max(max(sum_output)) + 30])
# pylab.show()
pylab.figure()
pylab.hold(True)
pylab.plot(sum_output[0], "b^")
pylab.plot(sum_output[1], "r^")
#pylab.plot(sum_output[0],"b")
#pylab.plot(sum_output[1],"r")
ppp0 = np.array(sum_output[0])
ppp1 = np.array(sum_output[1])
out_cl0 = [i for i in range(len(outputs)) if outputs[i] == 0]
out_cl1 = [i for i in range(len(outputs)) if outputs[i] == 1]
pylab.plot(out_cl0, ppp0[out_cl0], "bs")
pylab.plot(out_cl1, ppp1[out_cl1], "rs")
pylab.hold(False)
pylab.title("Total spikes at each AN population for each trial")
pylab.xlabel("Trials")
pylab.ylabel("Spike Count for Each Trial")
pylab.legend(["Cls 0", "Cls 1", "Actual Winner Cls 0",
"Actual Winner Cls 1"])
pylab.axis([-2, n_trials + 2, -4, max(max(sum_output)) + 30])
pylab.show()
## Check Classification rate
s = np.array(sum_output)
cl = np.floor((np.sign(s[1] - s[0]) + 1) / 2)
r_cl = np.array(outputs)
wrong = np.sum(np.abs(cl - r_cl))
rate = (n_trials - wrong) / n_trials
print("success rate: {}%".format(abs(rate)*100.))
print("cl:\n", cl)
print("r_cl:\n", r_cl)
## Plot 5
if 0:
pylab.figure()
cf = 0.1
pylab.hold(True)
cls_wei0 = np.load("output_files/stdp_weights{}.npy".format(0))
mx = max(cls_wei0)
cls_wei0 = cf * cls_wei0 / mx
cls_wei1 = np.load("output_files/stdp_weights{}.npy".format(1))
mx = max(cls_wei1)
cls_wei1 = cf * cls_wei1/ mx
l = min(len(cls_wei0), len(cls_wei1))
new_array0 = [cls_wei0[i] for i in range(l) if cls_wei0[i] > cls_wei1[i]]
x0 = [i for i in range(l) if cls_wei0[i] > cls_wei1[i]]
new_array1 = [cls_wei1[i] for i in range(l) if cls_wei1[i] > cls_wei0[i]]
x1 = [i for i in range(l) if cls_wei1[i] > cls_wei0[i]]
pylab.plot(x0, new_array0, "gx")
pylab.plot(x1, new_array1, "bx")
#for i in range(2):
# cls_wei=np.load("stdp_weights{}.npy".format(i))
# mx=max(cls_wei)
# cls_wei=0.05*cls_wei/mx
# pylab.plot(cls_wei,"x")
pylab.axis([-10, 2000, -0.1, 0.15])
pylab.hold(False)
pylab.show()
## Plot 7
if 0:
sum_filt = [[0 for i in range(n_feature * n_pop)] for j in range(n_cl)]
sum_filt = np.array(sum_filt)
for i in range(n_trials):
t_st = i * time_int_trials
t_end = t_st + time_int_trials
cl = outputs[i]
for n, t in Filt_Exc_Spikes:
if t >= t_st and t < t_end:
sum_filt[int(cl),int(n)] = sum_filt[(cl),int(n)] + 1
a4=sum_filt[0]
b4=sum_filt[1]
pylab.figure()
pylab.hold(True)
pylab.plot(a4,"b.")
pylab.plot(b4,"r.")
pylab.xlabel('Neuron ID')
pylab.ylabel('Total Firing Rates Through Trials')
pylab.title("Total Spiking Activity of Neurons at Decomposition Layer for Each Class")
pylab.hold(False)
pylab.legend(["Activity to AN1","Activity to AN2"])
pylab.show()
return rate
