import numpy as np
#import matplotlib.pylab as plt
import json
import os
#####################################
def save_obj(obj, name ):
with open('dat/'+ name + '.txt', 'w') as f:
json.dump(obj, f)
def load_obj(name ):
with open('dat/' + name + '.txt', 'r') as f:
return json.load(f)
def load_params(name ):
script_dir = os.path.abspath(os.path.join(os.path.dirname(os.path.abspath(__file__)), os.pardir))
with open(os.path.join(script_dir,'params/' + name + '.txt'), 'r') as f:
return json.load(f)
np.random.seed()
print('===================================')
print('START Evolution sim')
print('===================================')
#####################################
###### Parameters
#####################################
store_name = 'dat_totalSim_evo_s6c'
store_osi = 'dat_totalSim_OSI_s6c'
m_par = load_params('model_params')
timeStep = m_par['timeStep']
stimLen = m_par['stimLen']
deltaStim = m_par['deltaStim']
NrE_L3 = m_par['NrE_L3']
NrE_L4 = m_par['NrE_L4']
NrI_L3 = m_par['NrI_L3']
Nr4to3 = m_par['Nr4to3']
tau_e = m_par['tau_e']
tau_i = m_par['tau_i']
beta_v = m_par['beta_v']
beta_inh = m_par['beta_inh']
theta_v = m_par['theta_v']
vis_amp = m_par['vis_amp']
depr_value = m_par['depr_value']
nr_Ori = m_par['nr_Ori']
NrE_L4_tot =nr_Ori*NrE_L4
bck_current = m_par['bck_current']
layerfactor = m_par['layerfactor']
## Plasticity
p_par = load_params('plast_params')
wmax_taro = p_par['wmax_taro']
wmin_taro = p_par['wmin_taro']
Theta_eeMax_H = p_par['Theta_eeMax_H']-1
Theta_eeMax_L = p_par['Theta_eeMax_L']-1
l_rate_ee = p_par['l_rate_ee']
wmin_ei = p_par['wmin_ei']
deltath = p_par['deltath']
theta_ei_H = p_par['theta_ei_H']
lrate_ei = p_par['lrate_ei']
wmax_ie = p_par['wmax_ie']
wmin_ie = p_par['wmin_ie']
lrate_ie_Max = p_par['lrate_ie_Max']
ie_target = p_par['ie_target']
tau_avg = p_par['tau_avg']
I3avg_init = p_par['I3_avg_ini']
wee_ini = 0#p_par['wee_ini']
wee_L4to3_ini = p_par['wee_L4to3_ini']
wie_ini = p_par['wie_ini']
w_ie_4to3_ini = p_par['w_ie_4to3_ini']
wei_ini = p_par['wei_ini']
tmax1 = 50000
tmax2 = 0
tmax = tmax1 + tmax2
deprtime = tmax1/100 + 10
storeDT = timeStep
stimWindow = int(deltaStim/timeStep)
nr_per_Ori = int(NrE_L3/nr_Ori)
CL_mod = 1
IL_mod = 1
I3_avg = I3avg_init*np.ones(NrI_L3)
activity_shift = 1 + .1*np.random.randn(NrE_L3) #.15*np.random.randn(NrE_L3) #different firing rates
corr_act = np.outer(activity_shift,np.array(np.ones(NrE_L4_tot)))
backge = .1#1#0
backgi = .1#1
unos = (np.diag(np.ones(nr_Ori)) + backge*(np.ones((nr_Ori,nr_Ori))-np.diag(np.ones(nr_Ori))))
unos[0,-1] = backge
unos[-1,0] = backge
unos=np.expand_dims(unos,axis=1)
spec_Ori = np.repeat(np.repeat(unos,nr_per_Ori,axis=0),NrE_L4,axis=1)
spec_Ori = spec_Ori + np.random.rand(nr_Ori*nr_per_Ori,NrE_L4,nr_Ori)*.5
if NrI_L3>1:
unos = backgi*(np.ones((NrI_L3,nr_Ori)))
for ii in np.arange(NrI_L3):
unos[ii,int(np.floor(np.random.rand()*nr_Ori))] = 1
# print(unos)
unos=np.expand_dims(unos,axis=1)
spec_Ori_inh = np.repeat(unos,NrE_L4,axis=1)
else:
spec_Ori_inh = np.ones((NrI_L3,NrE_L4,nr_Ori))
#*** Other
NrIp = 2 # total nr of eyes
#####################################
###### Variables
#####################################
#################################################################################
# LAYER 3
#################################################################################
Ev_L3 = np.zeros(NrE_L3) # vector of excitatory activities
Iv_L3 = np.zeros(NrI_L3) # vector of inhibitory activities
w_ie_4to3 = w_ie_4to3_ini*spec_Ori_inh#*np.ones((NrI_L3,NrE_L4,nr_Ori)) # #ffw weight matrix
w_ie_4to3[w_ie_4to3<wmin_ie] = wmin_ie
w_ie_4to3 =np.reshape(w_ie_4to3,(NrI_L3,nr_Ori*NrE_L4),order='F')
w_ei_L3 = wei_ini*np.ones((NrE_L3,NrI_L3)) # ffw weight matrix
w_ie_L3 = wie_ini*np.ones((NrI_L3,NrE_L3))
w_ii_L3 = np.zeros((NrI_L3,NrI_L3))
L3_connections = np.zeros((NrE_L3,NrE_L4_tot))
subfam = int(NrE_L4/5)
for ii in range(0,int(.1*NrE_L3)):
arr = []
distr = [5,5,5,10,25]
distr = np.array(distr)*Nr4to3/50
sp = len(distr)
for _ in range(nr_Ori):
for jj in range(sp):
a = (np.random.permutation(np.arange(subfam))<distr[jj])*1
arr = arr+a.tolist()
L3_connections[ii,:] = np.array(arr)
for ii in range(int(.1*NrE_L3),int(.2*NrE_L3)):
arr = []
distr = [5,10,10,10,15]
distr = np.array(distr)*Nr4to3/50
sp = len(distr)
for _ in range(nr_Ori):
for jj in range(sp):
a = (np.random.permutation(np.arange(subfam))<distr[jj])*1
arr = arr+a.tolist()
L3_connections[ii,:] = np.array(arr)
for ii in range(int(.2*NrE_L3),int(.4*NrE_L3)):
arr = []
distr = [10,10,10,10,10]
distr = np.array(distr)*Nr4to3/50
sp = len(distr)
for _ in range(nr_Ori):
for jj in range(sp):
a = (np.random.permutation(np.arange(subfam))<distr[jj])*1
arr = arr+a.tolist()
L3_connections[ii,:] = np.array(arr)
for ii in range(int(.4*NrE_L3),int(.6*NrE_L3)):
arr = []
distr = [10,15,10,10,5]
distr = np.array(distr)*Nr4to3/50
sp = len(distr)
for _ in range(nr_Ori):
for jj in range(sp):
a = (np.random.permutation(np.arange(subfam))<distr[jj])*1
arr = arr+a.tolist()
L3_connections[ii,:] = np.array(arr)
for ii in range(int(.6*NrE_L3),int(.8*NrE_L3)):
arr = []
distr = [10,15,15,5,5]
distr = np.array(distr)*Nr4to3/50
sp = len(distr)
for _ in range(nr_Ori):
for jj in range(sp):
a = (np.random.permutation(np.arange(subfam))<distr[jj])*1
arr = arr+a.tolist()
L3_connections[ii,:] = np.array(arr)
for ii in range(int(.8*NrE_L3),int(1*NrE_L3)):
arr = []
distr = [25,10,5,5,5]
distr = np.array(distr)*Nr4to3/50
sp = len(distr)
for _ in range(nr_Ori):
for jj in range(sp):
a = (np.random.permutation(np.arange(subfam))<distr[jj])*1
arr = arr+a.tolist()
L3_connections[ii,:] = np.array(arr)
np.random.shuffle(L3_connections)
L3_connections_inh = np.zeros((NrI_L3,NrE_L4_tot))# np.append(np.ones((NrI_L3,Nr4to3)),np.zeros((NrI_L3,NrE_L4-Nr4to3)),axis=1)
for ii in range(NrI_L3):
L3_connections_inh[ii,:] = np.tile(np.append(np.append(np.ones((1,int(.6*Nr4to3))),np.zeros((1,NrE_L4-Nr4to3))
,axis=1),np.ones((1,int(.4*Nr4to3))),axis=1),(1,nr_Ori))
w_ee_4to3 = wee_L4to3_ini*spec_Ori#np.ones((NrE_L3,NrE_L4))
w_ee_4to3[w_ee_4to3<wmin_taro] = wmin_taro#np.ones((NrE_L3,NrE_L4))
w_ee_4to3 = np.reshape(w_ee_4to3,(NrE_L3,nr_Ori*NrE_L4),order='F')
w_taroee_4to3 = w_ee_4to3
w_ee_L3 = np.zeros((NrE_L3,NrE_L3))
#################################################################################
#################################################################################
# LAYER 4
#################################################################################
Ev_L4 = np.zeros(NrE_L4_tot) # vector of excitatory activities
L4_connections = []
meancontra = 30
stdvcontra = 35
for _ in range(nr_Ori):
q = meancontra +np.random.randn(NrE_L4)*stdvcontra
q[q<0] = 100*np.random.rand(len(q[q<0])) #0
q[q>100] = 100
q = (np.sort(q)).tolist()
L4_connections.append(q)
q = np.reshape(L4_connections,[NrE_L4_tot])
L4_connections = .01*np.transpose(np.array([q,100-q]))
#################################################################################
storeV = np.zeros((15,int(tmax/storeDT))) # storage of data
weightV = np.zeros((15,int(tmax/storeDT))) # storage of data
weightV2 = np.zeros((15,int(tmax/storeDT))) # storage of data
weightV3 = np.zeros((15,int(tmax/storeDT))) # storage of data
#####################################
###### functions
#####################################
def f_IEvol_WilCow2(y,ipt,exc,iiw,ipw,ip_corr,iew,tauy,ori,bckg):
tot_inputs = np.dot(ipw*ip_corr,ipt)
y = y + (timeStep/tauy)*(-y + gainF_relu( tot_inputs+np.dot(iew,exc)-np.dot(iiw,y)+bckg,beta_inh ) )
return y
def f_EEvol_WilCow2(y,ipt,inh,eew,ipw,ip_corr,eiw,tauy,ori,bckg):
tot_inputs = np.dot(ipw*ip_corr,ipt)
y = y + (timeStep/tauy)*(-y + gainF_relu( tot_inputs+np.dot(eew,y)
-np.dot(eiw,inh)+bckg,beta_v ) )
return y
def gainF_relu(x,slopef):
out = slopef*(x-theta_v)
out = out*(x>theta_v)
return out
def f_e_to_i_plasticity(x,y,yavg,weight):
theta_bcm = yavg**2/ie_target
aux = np.outer(y-theta_bcm,2*(x>0))*(np.outer(y-theta_bcm,2)<0)+np.outer(y-theta_bcm,x*(x>3.2))*(np.outer(y-theta_bcm,2)>0)
dw = np.multiply(np.outer(y,np.ones(np.shape(x))),aux)
weight = weight + timeStep*lrate_ie_Max*dw
weight[weight>wmax_ie] = wmax_ie
weight[weight<wmin_ie] = wmin_ie
return weight
def f_i_to_e_plasticity(x,y,weight,activityshift):
pre_o = np.ones(np.shape(x))
dw = np.outer(y>theta_ei_H*activityshift-deltath,pre_o)*np.outer(y-theta_ei_H*activityshift,x)
weight = weight + timeStep*lrate_ei*dw
weight[weight<wmin_ei] = wmin_ei
return weight
def f_ee_plasticity(pre=0,post=0,whebb=0,recurrent='no'):
dw = ((np.outer(post,pre)>Theta_eeMax_L)*(np.outer(post,pre)-Theta_eeMax_H)
*(np.outer(post,pre)-Theta_eeMax_L))
dw = np.tanh(dw/0.0001)
whebb = whebb + timeStep*l_rate_ee*dw
whebb[whebb>wmax_taro] = wmax_taro
whebb[whebb<wmin_taro] = wmin_taro
if recurrent == 'yes':
if np.size(whebb,axis=0) == np.size(whebb,axis=1):
whebb = whebb - np.diag(np.diag(whebb))
else:
print('Problem with recurrent connections: no square matrix!')
return whebb
#% Calculate OSI
def f_OSI(rates):
nr_Ori = np.size(rates,axis=1)
rt_var = np.sort(rates,axis=1)
epsilon = 1e-10
tot_rt = np.divide(1,np.sum(rates,axis=1)+epsilon)
ori = np.arange(0,2*np.pi,2*np.pi/nr_Ori)
R = np.dot((np.tile(np.expand_dims(tot_rt,axis=1),(1,nr_Ori))*rt_var),np.exp(1j*ori))
return np.abs(R)
def f_calc_odi(w_ee_L3b,w_ee_L4to3b,w_ei_L3b,w_ie_L4to3b,w_ie_L3_effb,w_ii_L3b,bckg):
tttime= deltaStim#nr_Ori*deltaStim
respv = np.zeros((NrE_L3,2)) # storage of data
respv2 = np.zeros((NrE_L4_tot,2)) # storage of data
storV = np.zeros((NrE_L3,int(tttime/timeStep),nr_Ori)) # storage of data
storVI = np.zeros((NrI_L3,int(tttime/timeStep),nr_Ori)) # storage of data
storV4 = np.zeros((NrE_L4_tot,int(tttime/timeStep),nr_Ori)) # storage of data
for ippp in range(2):
if ippp == 1:
eye = 'ipsi_closed'
else:
eye = 'contra_closed'
if eye == 'ipsi_closed':
a = 0
b = 1
else:
a = 1
b = 0
for ori in np.arange(nr_Ori):
EvL4 = np.zeros(NrE_L4_tot) # vector of excitatory activities
EvL3 = np.zeros(NrE_L3) # vector of excitatory activities
IvL3 = np.zeros(NrI_L3) # vector of inhibitory activities
for tt in np.arange(0,deltaStim,timeStep):
if tt%stimLen == 0:
Iptv = np.zeros(NrE_L4_tot)
if tt%deltaStim == 0:
Ori = ori
aux_IL = np.zeros((nr_Ori,NrE_L4))
aux_IL[Ori,:] = vis_amp
aux_IL = np.reshape(aux_IL,(NrE_L4_tot))
aux_CL = np.zeros((nr_Ori,NrE_L4))
aux_CL[Ori,:] = vis_amp
aux_CL = np.reshape(aux_CL,(NrE_L4_tot))
Visual_Ipt_gauss_ipsi = L4_connections[:,0]*aux_IL
Visual_Ipt_gauss_contra = L4_connections[:,1]*aux_CL
Visual_Ipt_gauss = a*Visual_Ipt_gauss_ipsi + b*Visual_Ipt_gauss_contra
Iptv = Visual_Ipt_gauss + Iptv
Iptv = 1*Iptv*(Iptv>0)
EvL4 = beta_v*(Iptv)
IvL3 = f_IEvol_WilCow2(IvL3,layerfactor*EvL4,EvL3,w_ii_L3b,w_ie_L4to3b,L3_connections_inh,
w_ie_L3_effb,tau_i,ori,0)
EvL3 = f_EEvol_WilCow2(EvL3,layerfactor*EvL4,IvL3,w_ee_L3b,
w_ee_L4to3b,L3_connections*corr_act,w_ei_L3b,tau_e,ori,0)
storV[:,int(tt/timeStep),ori] = EvL3
storVI[:,int(tt/timeStep),ori] = IvL3
storV4[:,int(tt/timeStep),ori] = EvL4
maxTime = np.amax(storV,axis=1)
maxOri = np.argmax(maxTime,axis=1)
maxTime4 = np.amax(storV4,axis=1)
maxOri4 = np.argmax(maxTime4,axis=1)
for xx in range(np.size(storV,axis=0)):
respv[xx,ippp]= maxTime[xx,maxOri[xx]] - storV[xx,-1,maxOri[xx]]
for xx in range(np.size(storV4,axis=0)):
respv2[xx,ippp]= maxTime4[xx,maxOri4[xx]] - storV4[xx,-1,maxOri4[xx]]
return respv,respv2
def f_calc_osi(w_ee_L3b,w_ee_L4to3b,w_ei_L3b,w_ie_L4to3b,w_ie_L3_effb,w_ii_L3b,nrOri,bckgr):
tttime= deltaStim
storV = np.zeros((NrE_L3,int(tttime/timeStep))) # storage of data
storVI = np.zeros((NrI_L3,int(tttime/timeStep))) # storage of data
storV4 = np.zeros((NrE_L4_tot,int(tttime/timeStep))) # storage of data
store_Osi_E = np.zeros((NrE_L3,nrOri))
store_Osi_I = np.zeros((NrI_L3,nrOri))
for ori in np.arange(nrOri):
EvL4 = np.zeros(NrE_L4_tot) # vector of excitatory activities
EvL3 = np.zeros(NrE_L3) # vector of excitatory activities
IvL3 = np.zeros(NrI_L3) # vector of inhibitory activities
for tt in np.arange(0,deltaStim,timeStep):
if tt%stimLen == 0:
Iptv = np.zeros(NrE_L4_tot)
if tt%deltaStim == 0:
Ori = ori
aux_IL = np.zeros((nr_Ori,NrE_L4))
aux_IL[Ori,:] = vis_amp#+bckgr
aux_IL = np.reshape(aux_IL,(NrE_L4_tot))
aux_CL = np.zeros((nr_Ori,NrE_L4))
aux_CL[Ori,:] = vis_amp#+bckgr
aux_CL = np.reshape(aux_CL,(NrE_L4_tot))
Visual_Ipt_gauss_ipsi = L4_connections[:,0]*aux_IL
Visual_Ipt_gauss_contra = L4_connections[:,1]*aux_CL
Visual_Ipt_gauss = Visual_Ipt_gauss_ipsi + Visual_Ipt_gauss_contra
Iptv = Visual_Ipt_gauss + Iptv
Iptv = 1*Iptv*(Iptv>0)
EvL4 = beta_v*(Iptv)
IvL3 = f_IEvol_WilCow2(IvL3,layerfactor*EvL4,EvL3,w_ii_L3b,w_ie_L4to3b,L3_connections_inh,
w_ie_L3_effb,tau_i,ori,0)
EvL3 = f_EEvol_WilCow2(EvL3,layerfactor*EvL4,IvL3,w_ee_L3b,
w_ee_L4to3b,L3_connections*corr_act,w_ei_L3b,tau_e,ori,0)
storV[:,int(tt/timeStep)] = EvL3
storVI[:,int(tt/timeStep)] = IvL3
storV4[:,int(tt/timeStep)] = EvL4
store_Osi_E[:,ori] = np.amax(storV,axis=1)
store_Osi_I[:,ori] = np.amax(storVI,axis=1)
return store_Osi_E,store_Osi_I
######################################
#----- Run -----#
######################################
Ori = int(np.floor(np.random.rand(1)*nr_Ori)[0])
w_ie_L3_eff = w_ie_L3
strV3 = np.zeros((NrE_L3,stimWindow)) # storage of data
strV4 = np.zeros((NrE_L4_tot,stimWindow)) # storage of data
strI3 = np.zeros((NrI_L3,stimWindow))
OSI_evolution_E = np.zeros(20)
OSI_evolution_I = np.zeros(20)
osi_index = 0
bckg = bck_current
rvv_d0, rvv2 = f_calc_odi(w_ee_L3,w_ee_4to3,w_ei_L3,w_ie_4to3,w_ie_L3_eff,w_ii_L3,bck_current)
for tt in np.arange(0,int(tmax),timeStep):
if tt%stimLen == 0:
Iptv = np.zeros(NrE_L4_tot)
if tt%deltaStim == 0:
Ori = int(np.floor(np.random.rand(1)*nr_Ori)[0])
Ori_vect = np.zeros((nr_Ori,NrE_L4))
Ori_vect[Ori,:] = 1
Ori_vect = np.reshape(Ori_vect,(NrE_L4_tot))
Visual_Ipt_gauss_ipsi = L4_connections[:,0]*Ori_vect*IL_mod*vis_amp
Visual_Ipt_gauss_contra = L4_connections[:,1]*Ori_vect*CL_mod*vis_amp
Visual_Ipt_gauss = Visual_Ipt_gauss_ipsi + Visual_Ipt_gauss_contra
Iptv = Visual_Ipt_gauss + Iptv +bckg*Ori_vect
Iptv = 1*Iptv*(Iptv>0)
strV3[:,int((tt%deltaStim)/timeStep)] = Ev_L3
strV4[:,int((tt%deltaStim)/timeStep)] = Ev_L4
strI3[:,int((tt%deltaStim)/timeStep)] = Iv_L3
I3_avg = I3_avg + (timeStep/tau_avg)*(Iv_L3-I3_avg)
w_ie_L3_eff = w_ie_L3
#####################################################################################################
# LAYER 4
#####################################################################################################
Ev_L4 = beta_v*(Iptv)
#####################################################################################################
# LAYER 3
#####################################################################################################
Iv_L3 = f_IEvol_WilCow2(Iv_L3,layerfactor*Ev_L4,Ev_L3,w_ii_L3,w_ie_4to3,
L3_connections_inh,w_ie_L3_eff,tau_i,Ori,0)
Ev_L3 = f_EEvol_WilCow2(Ev_L3,layerfactor*Ev_L4,Iv_L3,w_ee_L3,
w_ee_4to3,L3_connections*corr_act,w_ei_L3,tau_e,Ori,0)
w_ie_4to3 = f_e_to_i_plasticity(Ev_L4,Iv_L3,I3_avg,w_ie_4to3)
w_ie_L3 = f_e_to_i_plasticity(Ev_L3,Iv_L3,I3_avg,w_ie_L3)
# w_ee_L3 = f_ee_plasticity(pre=Ev_L3,post=Ev_L3,whebb=w_ee_L3,recurrent='no')
w_ee_4to3 = f_ee_plasticity(pre=Ev_L4,post=Ev_L3,whebb=w_ee_4to3,recurrent='no')
w_ei_L3 = f_i_to_e_plasticity(Iv_L3,Ev_L3,w_ei_L3,activity_shift)
#####################################################################################################
#####################################################################################################
##################################
# Display the progress
###################################
if np.abs(tt-tmax*0.05*np.round(20*tt/tmax)) < 0.5*timeStep:
print('> '+str(int(np.round(100*tt/tmax)))+'%')
# store data
nrpoints = .05
if np.abs(tt-timeStep-tmax*nrpoints*np.round((1/nrpoints)*tt/tmax)) < 0.5*timeStep:
osi_E,osi_I = f_calc_osi(w_ee_L3,w_ee_4to3,w_ei_L3,w_ie_4to3,w_ie_L3_eff,w_ii_L3,nr_Ori,bck_current)
osi_E = f_OSI(osi_E)
osi_I = f_OSI(osi_I)
OSI_evolution_E[osi_index] = np.mean(osi_E)
OSI_evolution_I[osi_index] = np.mean(osi_I)
osi_index = osi_index + 1
rvv_d3, rvv2 = f_calc_odi(w_ee_L3,w_ee_4to3,w_ei_L3,w_ie_4to3,w_ie_L3_eff,w_ii_L3,bck_current)
odi_v = {
'OSI_evolution_E' : OSI_evolution_E.tolist(), #
'OSI_evolution_I' : OSI_evolution_I.tolist() #
}
save_obj(odi_v, store_osi )
w_v = {
'w_ee_L3' : w_ee_L3.tolist(), #
'w_ei_L3' : w_ei_L3.tolist(), #
'w_ie_L3' : w_ie_L3.tolist(), #
'w_ii_L3' : w_ii_L3.tolist(), #
'w_ee_4to3' : w_ee_4to3.tolist(), #
'w_ie_4to3' : w_ie_4to3.tolist(), #
'activity_shift' : activity_shift.tolist(), #
'L3_connections' : L3_connections.tolist(), #
'L3_connections_inh' : L3_connections_inh.tolist(), #
'spec_Ori' : spec_Ori.tolist(), #
'spec_Ori_inh' : spec_Ori_inh.tolist(), #
'I3_avg' : I3_avg.tolist(), #
'L4_connections' : L4_connections.tolist(), #
}
save_obj(w_v, store_name )