def run(lowThreshold):
    runtime = 1000.0
    pynn.setup()

    # set STDP params for low threshold  -> fails when vcthigh-vctlow < 0.04
    if lowThreshold:
        pynn.hardware.hwa.setSTDPParams(0.0, default.tpcsec,
                                        default.tpcorperiod, 1.0, 1.0, 1.0,
                                        0.98, 2.5)
    else:
        pynn.hardware.hwa.setSTDPParams(0.0, default.tpcsec,
                                        default.tpcorperiod, 1.0, 1.0, 1.0,
                                        0.85, 2.5)

    neuron = pynn.Population(1, pynn.IF_facets_hardware1)
    spikeArray = pynn.Population(1, pynn.SpikeSourceArray)

    stdp_model = pynn.STDPMechanism(
        timing_dependence=pynn.SpikePairRule(),
        weight_dependence=pynn.AdditiveWeightDependence())

    prj = pynn.Projection(
        spikeArray,
        neuron,
        method=pynn.AllToAllConnector(weights=pynn.minExcWeight() * 0),
        target='excitatory',
        synapse_dynamics=pynn.SynapseDynamics(slow=stdp_model))

    pynn.run(runtime)
    pynn.end()
def run(withSTDP):
    runtime = 1000.0

    pynn.setup()
    stim = pynn.Population(1, pynn.SpikeSourcePoisson, {
                           'start': 0, 'duration': runtime, 'rate': 100.0})
    neuron = pynn.Population(1, pynn.IF_facets_hardware1)

    if withSTDP:
        stdp_model = pynn.STDPMechanism(timing_dependence=pynn.SpikePairRule(),
                                        weight_dependence=pynn.AdditiveWeightDependence())
        pynn.Projection(stim, neuron,
                        method=pynn.AllToAllConnector(
                            weights=pynn.maxExcWeight()),
                        target='excitatory',
                        synapse_dynamics=pynn.SynapseDynamics(slow=stdp_model))
    else:
        pynn.Projection(stim, neuron,
                        method=pynn.AllToAllConnector(
                            weights=pynn.maxExcWeight()),
                        target='excitatory')

    pynn.run(runtime)
    pynn.end()
Esempio n. 3
0
    def runTest(self):
        import numpy as np

        column = 4
        row = 4
        n = 20  # number of spike pairs
        deltaTList = [-1.0, 1.0]  # ms
        deltaTLimit = 0.3  # allowed deviation
        delay = 2.9  # ms (between stimulus and post)
        # at beginning in ms (should be larger than max deltaT)
        experimentOffset = 100.0
        deltaTPairs = 100.0  # time between pre-post-pairs in ms

        noStimulators = 3
        weightStimulator = 15  # weight for stimulator neurons
        weightMeasure = 0  # weight for measured neuron
        procCorrOffset = 100.0  # time after experiment until correlations are processed in ms

        for deltaT in deltaTList:
            stimulus = np.arange(experimentOffset,
                                 (n - 0.5) * deltaTPairs + experimentOffset,
                                 deltaTPairs)
            self.assertTrue(len(stimulus) == n)
            stimulusMeasure = stimulus + delay - deltaT

            import pyNN.hardware.spikey as pynn
            import hwconfig_default_s1v2 as default

            pynn.setup()

            if column > 0:
                pynn.Population(column, pynn.IF_facets_hardware1)
            # stimulated neuron
            neuron = pynn.Population(1, pynn.IF_facets_hardware1)

            spikeSourceStim = None
            spikeSourceMeasure = None
            # stimulators above measured synapse
            if row < noStimulators:
                if row > 0:
                    dummy = pynn.Population(row, pynn.SpikeSourceArray)
                spikeSourceMeasure = pynn.Population(
                    1, pynn.SpikeSourceArray, {'spike_times': stimulusMeasure})

            spikeSourceStim = pynn.Population(noStimulators,
                                              pynn.SpikeSourceArray,
                                              {'spike_times': stimulus})

            # stimulators below measured synapse
            if row >= noStimulators:
                if row > noStimulators:
                    dummy = pynn.Population(row - noStimulators,
                                            pynn.SpikeSourceArray)
                spikeSourceMeasure = pynn.Population(
                    1, pynn.SpikeSourceArray, {'spike_times': stimulusMeasure})

            # connect and record
            stdp_model = pynn.STDPMechanism(
                timing_dependence=pynn.SpikePairRule(),
                weight_dependence=pynn.AdditiveWeightDependence())
            pynn.Projection(
                spikeSourceStim,
                neuron,
                method=pynn.AllToAllConnector(weights=pynn.minExcWeight() *
                                              weightStimulator),
                target='excitatory')
            prj = pynn.Projection(
                spikeSourceMeasure,
                neuron,
                method=pynn.AllToAllConnector(weights=pynn.minExcWeight() *
                                              weightMeasure),
                target='excitatory',
                synapse_dynamics=pynn.SynapseDynamics(slow=stdp_model))
            neuron.record()

            #######
            # RUN #
            #######
            # correlation flags:
            # 0: no weight change
            # 1: causal weight change
            # 2: anti-causal weight change
            pynn.hardware.hwa.setLUT(
                [1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
                [2, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0])

            lastInputSpike = np.max(np.concatenate(
                (stimulus, stimulusMeasure)))
            runtime = lastInputSpike + procCorrOffset
            pynn.hardware.hwa.autoSTDPFrequency = runtime
            print 'runtime: ' + str(runtime) + '; last input spike: ' + str(
                lastInputSpike) + '; STDP readout: ' + str(runtime)
            pynn.run(runtime)

            # get flag and spikes
            corrFlag = (
                np.array(prj.getWeightsHW(readHW=True, format='list')) /
                pynn.minExcWeight())[0]
            spikes = neuron.getSpikes()[:, 1]
            print 'stimulus:', stimulus
            print 'measure:', stimulusMeasure
            print 'post:', spikes
            self.assertTrue(
                len(stimulusMeasure) == len(spikes), 'No proper spiking!')
            print 'correlation flag: ' + str(corrFlag)
            print 'deltaT (is / should / limit):', np.mean(
                spikes - stimulusMeasure), '/', deltaT, '/', deltaTLimit
            self.assertTrue(
                abs(np.mean(spikes - stimulusMeasure) - deltaT) <= deltaTLimit,
                'No precise spiking!')
            if deltaT > 0:  # causal
                self.assertTrue(corrFlag == 1, 'Wrong correlation flag!')
            else:  # anti-causal
                self.assertTrue(corrFlag == 2, 'Wrong correlation flag!')

            pynn.end()
Esempio n. 4
0
runtime = 1000.0

pynn.setup(mappingOffset=column)

neuron = pynn.Population(1, pynn.IF_facets_hardware1)
dummy = pynn.Population(row, pynn.SpikeSourceArray, stimParams)
stimulus = pynn.Population(1, pynn.SpikeSourceArray, stimParams)

# enable and configure STP
stp_model = pynn.TsodyksMarkramMechanism(**stpParams)
pynn.Projection(
    stimulus,
    neuron,
    method=pynn.AllToAllConnector(weights=weight * pynn.minExcWeight()),
    target='excitatory',
    synapse_dynamics=pynn.SynapseDynamics(
        fast=stp_model)  #hier auskommentieren
)

pynn.record_v(neuron[0], '')

pynn.run(runtime)

membrane = np.array(zip(pynn.timeMembraneOutput, pynn.membraneOutput))

pynn.end()

# plot
import matplotlib.pyplot as plt
plt.plot(membrane[:, 0], membrane[:, 1])
plt.xlabel('time (ms)')
plt.ylabel('membrane potential (mV)')
Esempio n. 5
0
stdp_model = pynn.STDPMechanism(
    timing_dependence=pynn.SpikePairRule(),
    weight_dependence=pynn.AdditiveWeightDependence())
# connect stimulus
pynn.Projection(spikeSourceStim,
                neuron,
                method=pynn.AllToAllConnector(weights=pynn.minExcWeight() *
                                              weightStim),
                target='excitatory')
# create plastic synapse
prj = pynn.Projection(
    spikeSourcePlastic,
    neuron,
    method=pynn.AllToAllConnector(weights=pynn.minExcWeight() * weightPlastic),
    target='excitatory',
    synapse_dynamics=pynn.SynapseDynamics(slow=stdp_model))

neuron.record()

## custom correlation flags:
## 0: no weight change
## 1: one (or multiple) weight changes triggered by pre-post spike pairs
## 2: one (or multiple) weight changes triggered by post-pre spike pairs
#pynn.hardware.hwa.setLUT([1,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
#[2,0,2,0,0,0,0,0,0,0,0,0,0,0,0,0])

lastInputSpike = np.max(np.concatenate((stimulus, stimulusPlastic)))
runtime = lastInputSpike + stimulusOffset
# configure frequency of STDP controller manually, otherwise maximum frequency is used
pynn.hardware.hwa.autoSTDPFrequency = runtime
pynn.run(runtime)
Esempio n. 6
0
stimParams = {'spike_times': np.concatenate((np.arange(100.0, 401.0, 50.0), [700.0]))}
stpParams = {'U': 0.4, 'tau_rec': 100.0}
runtime = 1000.0

pynn.setup(mappingOffset=column)

neuron = pynn.Population(1, pynn.IF_facets_hardware1)
dummy = pynn.Population(row, pynn.SpikeSourceArray, stimParams)
stimulus = pynn.Population(1, pynn.SpikeSourceArray, stimParams)

# enable and configure STP
stp_model = pynn.TsodyksMarkramMechanism(**stpParams)
pynn.Projection(stimulus, neuron,
                method=pynn.AllToAllConnector(weights=weight * pynn.minExcWeight()),
                target='excitatory',
                synapse_dynamics=pynn.SynapseDynamics(fast=stp_model))

pynn.record_v(neuron[0], '')

pynn.run(runtime)

membrane = np.array(list(zip(pynn.timeMembraneOutput, pynn.membraneOutput)))

pynn.end()

# plot
import matplotlib.pyplot as plt
plt.plot(membrane[:,0], membrane[:,1])
plt.xlabel('time (ms)')
plt.ylabel('membrane potential (mV)')
plt.savefig('stp.png')