## Backpropagation learning on shuffled data
# to be run on a server or cluster
# Run as: python test_mf_grc_backprop_shuffle.py r_ix
# Where r_ix is index of vector of radii, 1 corresponding to 0um and 7 to 30um
import numpy as np
import pickle as pkl
import scipy.io as io
from datetime import datetime
import sys
# Partial shuffling algorithm
# See MATLAB code for more detailed comments
def part_shuffle(x,cov_x,cov_desired):
N,T = x.shape
#
switch_cells = np.zeros((N))
for i in range(N):
if len(np.unique(x[i,:]))>1:
switch_cells[i] = 1
whichswitch = np.where(switch_cells)[0]
#
cov_new = cov_x.copy()
#
if cov_desired < cov_x: # Decrease correlations
while cov_new > cov_desired:
num_to_switch_simul = int(np.max([1,np.min([N,np.round((cov_new-cov_desired)/cov_new*N/0.5)])]))
for i in range(num_to_switch_simul):
randcell = np.random.choice(whichswitch)
T1 = np.random.choice(T)
T2 = np.random.choice(np.where(x[randcell,:] != x[randcell,T1])[0])
x_T1 = x[randcell,T1]
x_T2 = x[randcell,T2]
x[randcell,T1] = x_T2
x[randcell,T2] = x_T1
cov_new = get_cov(x)
elif cov_desired > cov_x: # Increase correlations
while cov_new < cov_desired:
Ts = np.random.choice(T,2) # 1st value is low, 2nd value is high
for i in range(len(whichswitch)):
x_T1 = x[whichswitch[i],Ts[0]]
x_T2 = x[whichswitch[i],Ts[1]]
x_mean = x[whichswitch[i],:].mean()
if x_T1 > x_mean and x_T2 < x_mean:
x[whichswitch[i],Ts[0]] = x_T2
x[whichswitch[i],Ts[1]] = x_T1
cov_new = get_cov(x)
#
var_new = get_var(x)
return x, cov_new, var_new
def get_cov(x):
N = x.shape[0]
L,V = np.linalg.eig(np.cov(x)); L = np.real(np.sqrt(L+0J))
cov_x = (np.max(L)/np.sum(L) - 1./N)/(1.-1./N)
return cov_x
def get_var(x):
L,V = np.linalg.eig(np.cov(x)); L = np.real(np.sqrt(L+0J))
var_x = np.sum(L**2)
return var_x
def get_samples(r,f_mf_ix,N_syn,num_patterns,NADT):
file = open('../network_structures/GCLconnectivity_'+str(N_syn)+'.pkl')
p = pkl.load(file); conn_mat = p['conn_mat']; glom_pos = p['glom_pos']
N_mf, N_grc = conn_mat.shape
theta = 3. + NADT*f_mf[f_mf_ix]
if r == 0:
x_mf = np.zeros((N_mf,num_patterns))
for i in range(num_patterns):
mf_on = np.random.choice(N_mf,int(round(f_mf[f_mf_ix]*N_mf)),replace=False)
x_mf[mf_on,i] = 1.
else:
p=io.loadmat('../input_statistics/mf_patterns_r'+str(r)+'.mat')
g = p['gs'][f_mf_ix]; R = p['Rs'][:,:,f_mf_ix]
t = np.dot(R.transpose(), np.random.randn(N_mf,num_patterns))
x_mf = 1.*(t>-g*np.ones((N_mf,num_patterns)))
#
inp = 4./N_syn*np.dot(conn_mat.transpose(),x_mf)
x_grc = np.maximum(inp-theta,0)
#
return x_mf, x_grc
def backprop_step_nohid(W_out,gamma,training_pattern,target_pattern):
# Dynamics of units
s = lambda x: 1./(1.+np.exp(-x)); ds = lambda s: s*(1.-s) # sigmoidal
#####################################
# First step: feedforward propagation
o_in = training_pattern
o_out = s(np.dot(np.append(o_in,1),W_out))
# Second step: output layer backpropagation
D = np.diag(ds(o_out))
err = o_out - target_pattern
err_d = np.prod((target_pattern==target_pattern.max()) == (o_out==o_out.max()))
delta_out = np.dot(D,err)
dW_out = - gamma * np.outer(np.append(o_in,1),delta_out)
# Third step: update weights
W_out = W_out + dW_out;
#
return err, err_d, W_out
def backprop_nohid(training_set,target,n_epochs,gamma):
#
n_in = training_set.shape[0]; n_out = target.shape[0]
W_out = np.random.uniform(-1.,1.,size=(n_in+1,n_out))*1./(n_in+1)
# Shuffle order of training set
temp = range(training_set.shape[1])
np.random.shuffle(temp)
training_set = training_set[:,temp]
target = target[:,temp]
#
errors_rms = np.zeros((n_epochs),float)
errors_discrim = np.zeros((n_epochs),float)
for ep in range(n_epochs):
errors_temp = np.zeros((target.shape[1]),float)
errors_d_temp = np.zeros((target.shape[1]),float)
for k in range(target.shape[1]):
# Backpropagation step
err, err_d, W_out = backprop_step_nohid(W_out,gamma,training_set[:,k],target[:,k])
# Record errors
errors_temp[k] = np.sqrt(np.mean(err**2)) # RMS error
errors_d_temp[k] = err_d # Discrimination error
# Record average error for the epoch
errors_rms[ep] = errors_temp.mean()
errors_discrim[ep] = errors_d_temp.mean()
# Reshuffle order of training data
temp = range(training_set.shape[1])
np.random.shuffle(temp)
training_set = training_set[:,temp]
target = target[:,temp]
#
return errors_rms, errors_discrim, W_out
# Input parameters
r_ix = int(sys.argv[1])-1
r = ((r_ix)*5)
# Network parameters, vary for different network instances
N_grc = 487; N_mf = 187
N_syn_range = range(1,21)
f_mf = np.linspace(.05,.95,19)
# Backprop parameters
gamma = 0.01
N_epochs = 5000
C = 10
N_out = C
num_patterns = 64*C
filename = 'results_bp_shuff/grc_toy_r'+str(r)+'_shuff'
err_sh = np.zeros((len(N_syn_range),len(f_mf),N_epochs),float)
err_rms_sh = np.zeros((len(N_syn_range),len(f_mf),N_epochs),float)
cov_sh = np.zeros((len(N_syn_range),len(f_mf)),float)
var_sh = np.zeros((len(N_syn_range),len(f_mf)),float)
# array to save whether shuffling to remove or add correlations
which_sh = np.zeros((len(N_syn_range),len(f_mf)),float)
for k1 in range(len(N_syn_range)):
N_syn = N_syn_range[k1]
print N_syn
for k2 in range(len(f_mf)):
# Get activity samples x_mf and x_grc
samples_mf, samples_grc = get_samples(r,k2,N_syn,num_patterns,0)
# MF total variance and population correlation
var_mf = get_var(samples_mf)
cov_mf = get_cov(samples_mf)
#
if samples_grc.max() == 0:
err_rms_sh[k1,k2,:] = np.nan; err_sh[k1,k2,:] = np.nan
else:
# GC total variance and population correlation
var_grc = get_var(samples_grc)
cov_grc = get_cov(samples_grc)
# Determine which partial shuffling algorithm to use
if cov_grc > cov_mf:
which_sh[k1,k2] = 1
elif cov_mf > cov_grc:
which_sh[k1,k2] = -1
# Shuffle
samples_sh, cov_sh[k1,k2], var_sh[k1,k2] = part_shuffle(samples_grc,cov_grc,cov_mf)
# Get random pattern classifications
target = np.zeros((C,num_patterns))
for k in range(num_patterns):
target[np.random.choice(C),k] = 1
# Single layer backpropagation
err_rms_sh[k1,k2,:], err_sh[k1,k2,:], W_sh = backprop_nohid(samples_sh,target,N_epochs,gamma)
#
# Save results
p = {'err_rms_sh':err_rms_sh, 'err_sh':err_sh, 'var_sh':var_sh, 'cov_sh':cov_sh, 'which_sh':which_sh}
io.savemat(filename,p)