def FALMopt( objective , u, v2 , mus=np.array([30.,3.,0.3])): #mus=np.array([0.1,1.,10.])): #
lambdas = np.zeros(v2.size)
data = {'STAs':np.vstack(R['statistics']['stimulus']['STA']) ,
'STCs':np.vstack([stc[np.newaxis,:] for stc in R['statistics']['stimulus']['STC']]),
'V2':v2 , 'U': u , 'N':NRGC , 'N_spikes':R['N_spikes'] ,
'Ncones':Ncones , 'centers':centers , 'indices':indices , 'lam':lam , 'lambdas':lambdas }
obj_V1 = objective.where(**data).with_callback(callback)
optimize_L2 = FALMS.add_L2_term( obj_V1 , gtol=1e-3 )
init = (0.001 + 0.01 * Rand.rand(NRGC,u.shape[0])).flatten()
r = {}
def falms_callback(falmer): print ' L0: ', falmer[1][falmer[1]<>0].shape[0]
for mu in mus:
print
print
print 'mu: ', mu
reg_L1 = FALMS.L1_L2( mu )
params = FALMS.falms(init,optimize_L2,reg_L1,mu=0.01,ftol=2e-7,callback=falms_callback)
r[mu] = params
print [ (mu,par['M'][par['M']<>0].shape[0]) for mu,par in sorted(r.items())]
return r
data = [{ 'STA':STA[i] , \
'STC':STC[i] } for i in range(NRGC)]
init_params = [{'theta':data[i]['STA'] * 0.1 , \
'M':data[i]['STC'] * 0.1} for i in range(NRGC)]
# quadratic Poisson LL in theta,M:
objective = kolia_theano.Objective(init_params=init_params[0],differentiate=['f'],
f=quadratic_Poisson, barrier=eig_barrier)
# fix particular data:
obj = objective.where(**data[0])
# LL + quadratic optimizer, for FALMS
optimize_L2 = FALMS.add_L2_term( obj )
# callback after each FALMS iteration:
def falms_callback(falmer): print ' L0: ', falmer[1][falmer[1]<>0].shape[0]
r = {'L1':[{} for _ in init_params] , 'true':true}
def save():
savefile = open('../../../Desktop/results.pyckle','w')
cPickle.dump(r,savefile)
savefile.close()
r['ML'] = MaxLike(init_params,data)
r['analytical'] = analytical_ML(data)
objective = kolia_theano.sum_objective(terms,callback=callback_one)
objective = FALMS.set_argument( objective , data )
#optimize = optimize.optimizer( objective , callback=callback_one )
mu = 1
rho = 1
true = [{ 'theta' : np.dot( U.T , V1[i,:] ) , 'M' : 0.1*np.dot( U.T * V1[i,:] , U ) } for i in range(iii)]
def nuclearMatrix(params):
return np.reshape(params,(Ncones,-1))
optimize_L2 = FALMS.add_L2_term( objective )
falmer = FALMS.initialize( objective.flatten(true) )
ranks = {}
span = {}
result = {}
for i,rho in enumerate( 0.1 * np.array([2.9,2.9,2.9])):
#for rho in 0.001 * np.array([2.,2.5,2.8,3.,3.5]):
#for rho in 0.005 * np.array([1.,2.,4.,8.,16.]):
nuclear_L2 = FALMS.nuclear_L2( rho , get_matrix=nuclearMatrix )
falmer = FALMS.initialize( objective.flatten(init_params) + 0.03*R.randn(756) )
print
print
print 'RHO : ' , rho
for j in range(100):
