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
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)
rng = p.NumpyRNG()
grcpc_weights_distribution = p.RandomDistribution('uniform', [0.05, 0.5], rng)
pro_grcpcsynapsis_connector = p.FixedProbabilityConnector(
0.8, weights=grcpc_weights_distribution)
pro_grcpcsynapsis_left = p.Projection(prepop,
postpop,
pro_grcpcsynapsis_connector,
target="excitatory",
synapse_dynamics=syndyn_grcpcsynapsis,
label="grcpcsynapsis")
proj = pro_grcpcsynapsis_left
proj.projection_edge.postvertex.custom_max_atoms_per_core = max(
1, 4000 / proj.projection_edge.prevertex.atoms)
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()
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
w_max=max_weight,
A_plus=a_plus,
A_minus=a_minus))
rand_weights = sim.RandomDistribution(boundaries=(min_weight, max_weight))
prob_conn_10pc = sim.FixedProbabilityConnector(conn_prob,
allow_self_connections=False,
weights=rand_weights)
m2s_proj = sim.Projection(
master_pop,
slave_pop,
prob_conn_10pc,
target="excitatory",
synapse_dynamics=sim.SynapseDynamics(slow=stdp_model))
s2i_proj = sim.Projection(
slave_pop,
inhibit_pop,
prob_conn_10pc,
target="excitatory",
synapse_dynamics=sim.SynapseDynamics(slow=stdp_model))
i2s_proj = sim.Projection(
inhibit_pop,
slave_pop,
prob_conn_10pc,
target="inhibitory",
synapse_dynamics=sim.SynapseDynamics(slow=stdp_model))
#max_delay = 1
#
ssa2_times = {'spike_times': [[8, 18, 28, 38, 48, 58, 68, 78, 88, 98]]}
lif1 = p.Population(1, p.IF_curr_exp, cell_params_lif, label='lif1')
lif2 = p.Population(1, p.IF_curr_exp, cell_params_lif, label='lif2')
ssa1 = p.Population(1, p.SpikeSourceArray, ssa1_times, label='ssa1')
ssa2 = p.Population(1, p.SpikeSourceArray, ssa2_times, label='ssa2')
t_rule = p.SpikePairRule(tau_plus=5, tau_minus=5)
w_rule = p.AdditiveWeightDependence(w_min=0.0,
w_max=weight_to_spike,
A_plus=weight_to_spike,
A_minus=weight_to_spike)
stdp_model = p.STDPMechanism(timing_dependence=t_rule,
weight_dependence=w_rule)
s_d = p.SynapseDynamics(slow=stdp_model)
input_proj = p.Projection(lif1,
lif2,
p.AllToAllConnector(weights=(weight_to_spike / 2.0),
delays=1),
synapse_dynamics=s_d,
target="excitatory")
start_proj = p.Projection(ssa1,
lif1,
p.AllToAllConnector(weights=weight_to_spike,
delays=1),
target="excitatory")
teaching_proj = p.Projection(ssa2,
lif2,
p.AllToAllConnector(weights=weight_to_spike,
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
