def optimize_GLM( gtol=1e-6 , maxiter=500):
objective = objV1_data( R , obj_V1, v2=np.zeros(Ncones) , u=np.eye(Ncones) )
optimizer = optimize.optimizer( objective )
init_params = {'V1': np.vstack( R['statistics']['stimulus']['STA'] ) }
params = optimizer(init_params=init_params,maxiter=maxiter,gtol=gtol)
opt_V1 = objective.unflat(params)
return opt_V1['V1']
def MaxLike(init_params,data):
'''Maximum Likelihood optimization of parameters theta and M'''
print
print
print 'Starting Max-Likelihood optimization in theta and M, compiling...'
# Compile symbolic objectives into numerical objective.
objective = kolia_theano.Objective(init_params=init_params[0],differentiate=['f'],
callback=callback,
f=quadratic_Poisson, barrier=eig_barrier)
# Set fixed arguments STA and STC in optimizer object.
optimizers = [ optimize.optimizer( objective.where(**dat) ) for dat in data ]
# Maximize likelihood separately for each RGC; run the optimization twice just to be sure.
params = init_params
for i in range(2):
params = [opt(init_params=par) for opt,par in zip(optimizers,params)]
# Unpack flat parameters into meaningful dictionary.
params = [objective.unflat(par) for par in params]
# print norm of inferred and true theta
optU = [param['theta'] for param in params]
print
print 'true ||theta||^2 : ', np.sum(U*U,axis=1)
print 'optimal ||theta||^2 : ', [np.sum(optu*optu) for optu in optU]
print
print
return params
def optimize_u( v1, init_u, v2 , T, gtol=1e-7 , maxiter=500):
objective = kb.Sum_objective( [objU_data( ustats[rgc_type] , v1 , v2 , T , i )
for i in range(NRGC)] )
optimizer = optimize.optimizer( objective )
# debug_here()
params = optimizer(init_params={'u': init_u },maxiter=maxiter,gtol=gtol)
opt_u = single_objective.unflat(params)
return opt_u['u']
def optimize_UV1( objective, v1,u,v2 ):
data = {'STAs':np.vstack(R['statistics']['stimulus']['STA']) ,
'STCs':np.vstack([stc[np.newaxis,:] for stc in R['statistics']['stimulus']['STC']]),
'V2':v2 , 'N':NRGC , 'N_spikes':R['N_spikes'] ,
'Ncones':Ncones , 'centers':centers , 'indices':indices , 'lam':lam }
optimizer = optimize.optimizer( objective.where(**data).with_callback(callback) )
params = optimizer(init_params={'V1': v1 , 'U': u },maxiter=10000,gtol=1.1e-7)
opt_U = objective.unflat(params)
return opt_U
def optimize_U( objective, v1,init_U, v2 ):
data = {'STAs':np.vstack(R['statistics']['stimulus']['STA']) ,
'STCs':np.vstack([stc[np.newaxis,:] for stc in R['statistics']['stimulus']['STC']]),
'V2':v2 , 'V1': v1 , 'N':NRGC , 'N_spikes':R['N_spikes'] }
optimizer = optimize.optimizer( objective.where(**data).with_callback(callback) )
params = optimizer(init_params={'U': init_U },maxiter=2000,gtol=1.1e-6)
opt_U = objective.unflat(params)
return opt_U['U']
def optimize_V1( u,v2, T, init=V1 , gtol=1e-4 , maxiter=100):
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 , 'T':T, 'N':NRGC , 'N_spikes':R['N_spikes'] }
objective = obj_V1u.where(**data).with_callback(callback)
optimizer = optimize.optimizer( objective )
init_params = {'V1': init }
params = optimizer(init_params=init_params,maxiter=maxiter,gtol=gtol)
opt_V1 = objective.unflat(params)
return opt_V1['V1']
def optimize_L2(X, Y, mu, **optimize_params):
def L2_f(x, *_):
return getattr(objective, f)(x) + 0.5 * sum((x - Y) * (x - Y)) / mu
def L2_df(x, *_):
return getattr(objective, df)(x) + (x - Y) / mu
objective.L2_f = L2_f
objective.L2_df = L2_df
opter = optimizer(
objective, init_params=X, f="L2_f", df="L2_df", maxiter=100, gtol=gtol, disp=0, **optimize_params
)
return opter()
def main():
"""
Main function
"""
# Download data for NLTK if not already done
#nltk.download('all')
# Read
imdb = Indexer()
imdb_file = 'data/data.json'
logging.basicConfig(format='%(levelname)s: %(message)s', level=logging.INFO)
logging.info('Reading file %s' % imdb_file)
imdb.read_file(imdb_file)
logging.info('File %s read' % imdb_file)
(vocab_size, user_list, movie_list, \
rating_matrix, review_matrix, review_map) = imdb.get_mappings()
# Get number of users and movies
Users = len(user_list)
Movies = len(movie_list)
logging.info('No. of users U = %d' % Users)
logging.info('No. of movies M = %d' % Movies)
# Run Gibbs EM
for it in xrange(1,MAX_ITER+1):
logging.info('Running iteration %d of Gibbs EM' % it)
logging.info('Running E-Step - Gibbs Sampling')
gibbs_sampler = GibbsSampler(5,A,2)
gibbs_sampler.run(rating_matrix)
logging.info('Running M-Step - Gradient Descent')
for i in xrange(1,MAX_OPT_ITER+1):
optimizer()
# Output Predicted Ratings
for u in range(U):
for m in range(M):
pred_rate = predicted_rating(u, m)
print "Predicted Rating of user " + str(u) + " and movie " + str(m) + ": " + str(pred_rate)
def optimize_V1( objective, u,v2, lambdas=None , init=None , gtol=1e-4 , maxiter=100):
if lambdas is None:
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 }
optimizer = optimize.optimizer( objective.where(**data).with_callback(callback) )
if init is None:
init_params = {'V1': 0.001 + 0.01 * Rand.rand(NRGC,u.shape[0]) }
else:
init_params = {'V1': init }
params = optimizer(init_params=init_params,maxiter=maxiter,gtol=gtol)
opt_V1 = objective.unflat(params)
return opt_V1['V1']
def __init__(self,data,Nsub,prior=lambda U:0):
self.DATA = data
self.N = len(data[-2][0])
self.Nsub = Nsub
self.NRGC = len(data[-2])
self.prior = prior
self.mindet= -2.
U = Th.dmatrix() # SYMBOLIC variables #
V1 = Th.dvector() #
V2 = Th.dvector() #
STA = Th.dvector() #
STC = Th.dmatrix() #
theta= Th.dot( U.T , V1 ) #
M = Th.dot( V1 * U.T , (V2 * U.T).T ) #
detM = Th.dscalar()
invM = Th.dmatrix()
invMtheta = Th.as_tensor_variable(Th.dot(invM,theta),ndim=2)
prior = self.prior(U) #
post = ( Th.log(detM) \
# - 0.01 / (Th.log(detM)-self.mindet) \
- 1./(Th.log(detM)+6)**2 \
- Th.sum(invMtheta*theta) \
+ 2. * Th.sum( theta * STA ) \
+ Th.sum( M * (STC + Th.outer(STA,STA)) )) / 2. \
#+ prior \
#+ Th.sum(0.001*Th.log(V1)) #+ Th.sum(0.001*Th.log(U))
dpost_dM = ( invM + invMtheta * invMtheta.T \
# + 0.01 * invM / ((Th.log(detM)-self.mindet)**2) \
+ 2. * invM / (Th.log(detM)+6)**3 \
) / 2.
def dpost(dX):
return Th.grad( cost = post , wrt = dX ,
consider_constant=[invM,detM,STC,STA] ) \
- Th.grad( cost = Th.sum( dpost_dM * M ) , wrt = dX ,
consider_constant=[dpost_dM,STA,STC,invM,invMtheta])
self.M = function( [ U,V2,V1] , M ) #
self.posterior = function( [ U,V2,V1,invM,detM,STA,STC], post ) #
self.dpost_dU = function( [ U,V2,V1,invM,detM,STA,STC], dpost(U) ) #
self.dpost_dV1 = function( [ U,V2,V1,invM,detM,STA,STC], dpost(V1)) #
# self.dpost_dV2 = function( [ U,V2,V1,invM,detM,STA,STC], dpost(V2)) #
self.dpost_dM = function( [ U,V2,V1,invM,detM,STA,STC], dpost_dM) #
self.optimize = optimizer( self )
def _test_LNP( rgc_type='off parasol' ):
vardict = LNP( **thetaM( **linear_reparameterization()))
init_LNP = LNP_model( init_sv1(rgc_type) )
indices = extract( linear_stats( rgc_type, (5,0) ), ['sparse_index', 'subunit_index'] )
indices['N_subunits'] = len(cones)
unknown = extract(init_LNP,['sv1'])
train_LNP = global_objective( unknown, extract(init_LNP,['u','V2']),
vardict, run=linear_stats( rgc_type, (5,0) ),
indices=indices)
train_LNP.with_callback(callback)
train_LNP.description = 'LNP'
sv1 = optimize.optimizer( train_LNP )( init_params=unknown, maxiter=5000, gtol=1e-7 )
model = LNP_model( train_LNP.unflat( sv1 )['sv1'] )
model['LL'] = global_objective( unknown, extract(init_LNP,['u','V2']),
vardict, run=linear_stats( rgc_type, (-5,0)),
indices=indices).LL(sv1)
save(model,'LNP_'+rgc_type)
return model
def optimize_u( v1, init_u, v2 , T, gtol=1e-7 , maxiter=500):
optimizer = optimize.optimizer( objU_data( R , v1 , v2 , T ) )
# debug_here()
params = optimizer(init_params={'u': init_u },maxiter=maxiter,gtol=gtol)
opt_u = obj_u.unflat(params)
return opt_u['u']
def optimize():
content = request.json
data = content['data']
solution = optimizer(data)
return jsonify({"solution": solution})
# aqfn,aqpara = acquisitions.EIMAP
aqpara["lb"] = [-1.0, -1.0]
aqpara["ub"] = [1.0, 1.0]
aqpara["ev"]["s"] = 1e-12
cfn = objectives.cfaexp(1.0, 0.2)
ojf, xmin, ymin = objectives.genbiasedmat52ojf(len(aqpara["lb"]), aqpara["lb"], aqpara["ub"], 1.0)
ojfn = objectives.costfnwrap(ojf, cfn)
ojfchar = {"dx": len(aqpara["lb"]), "dev": len(aqpara["ev"])}
aqpara["cfn"] = cfn
aqpara["xau"] = 1.0
aqpara["xal"] = 0.0
# cfn = objectives.cf42
# ojfn,xmin,ymin = objectives.genmat52ojf(len(aqpara['lb']),aqpara['lb'],aqpara['ub'])
# ojfchar = {'dx':len(aqpara['lb']),'dev':len(aqpara['ev'])}
stoppara = {"nmax": 20}
stopfn = optimize.nstopfn
reccfn, reccpara = reccomenders.gpasmap
reccpara["lb"] = aqpara["lb"]
reccpara["ub"] = aqpara["ub"]
O = optimize.optimizer(path, aqpara, aqfn, stoppara, stopfn, reccpara, reccfn, ojfn, ojfchar, checkrecc=True)
O.run()
def optimize_V1( objective, true_run, u , v2, gtol=1e-4 , maxiter=500):
optimizer = optimize.optimizer( objective )
init_params = {'V1': 0.001 + 0.01 * Rand.rand(*true_run['model']['V'].shape) }
params = optimizer(init_params=init_params,maxiter=maxiter,gtol=gtol)
return objective.unflat(params)
def main():
"""
Main function
"""
# Download data for NLTK if not already done
# nltk.download('all')
# Read
np.random.seed(5)
baseline = False ## Make this true if you want to run the baseline, which is a simple latent factor model
path_to_save_results = './test/'
imdb = Indexer()
imdb_file = 'data/clothing_data_small.json' ## path to data file
logging.basicConfig(format='%(levelname)s: %(message)s',
level=logging.INFO)
logging.info('Reading file %s' % imdb_file)
imdb.read_file(imdb_file)
logging.info('File %s read' % imdb_file)
(
vocab_size,
user_list, # remove
movie_list,
review_matrix,
review_map,
user_dict,
movie_dict,
rating_list,
t_mean,
movie_reviews,
word_dictionary,
U,
M,
R,
test_indices) = imdb.get_mappings(path_to_save_results)
mul_factor = 0.1
## Initialize
alpha_vu = np.random.normal(0, sigma_u, (U, K)) * mul_factor
alpha_bu = np.random.normal(0, sigma_u, (U, 1)) * mul_factor
alpha_tu = np.random.normal(0, sigma_u, (U, A)) * mul_factor
# User
v_u = np.random.normal(0, sigma_u,
(U, K)) * mul_factor # Latent factor vector
b_u = np.random.normal(0, sigma_bu,
(U, 1)) * mul_factor # Common bias vector
theta_u = np.random.normal(0, sigma_ua,
(U, A)) * mul_factor # Aspect specific vector
# Movie
v_m = np.random.normal(0, sigma_m,
(M, K)) * mul_factor # Latent factor vector
b_m = np.random.normal(0, sigma_bm,
(M, 1)) * mul_factor # Common bias vector
theta_m = np.random.normal(0, sigma_ma,
(M, A)) * mul_factor # Aspect specific vector
# Common bias
b_o = np.random.normal(0, sigma_b0) * mul_factor
# Scaling Matrix
M_a = np.random.normal(0, sigma_Ma, (A, K)) * mul_factor
params = numpy.concatenate(
(alpha_vu.flatten('F'), v_u.flatten('F'), alpha_bu.flatten('F'),
b_u.flatten('F'), alpha_tu.flatten('F'), theta_u.flatten('F'),
v_m.flatten('F'), b_m.flatten('F'), theta_m.flatten('F'),
M_a.flatten('F'), np.array([b_o]).flatten('F')))
save_test_rmse = []
# Get number of users and movies
Users = len(user_list)
Movies = len(movie_list)
logging.info('No. of users U = %d' % Users)
logging.info('No. of movies M = %d' % Movies)
# change gibbs sampler initialization
gibbs_sampler = GibbsSampler(vocab_size, review_matrix, rating_list,
movie_dict, user_dict, movie_reviews,
word_dictionary, U, M, R, test_indices)
# Run Gibbs EM
for it in range(1, MAX_ITER + 1):
print('Running iteration %d of Gibbs EM' % it)
print('Running E-Step - Gibbs Sampling')
if baseline != True:
Nums, Numas, Numa = gibbs_sampler.run(vocab_size, review_matrix,
rating_list, user_dict,
movie_dict, movie_reviews,
word_dictionary, t_mean,
params, test_indices,
path_to_save_results)
else:
Nums = np.zeros((R, 2))
Numas = np.zeros((R, A, 2))
Numa = np.zeros((R, A))
print('Running M-Step - Gradient Descent')
for i in range(1, MAX_OPT_ITER + 1):
params, save_test_rmse = optimizer(Nums, Numas, Numa, rating_list,
t_mean, params, U, M, R,
test_indices, save_test_rmse)
np.save(path_to_save_results + 'params.npy', params)
np.save(
path_to_save_results +
'performance_notime_medium_noreg_seed5.npy', save_test_rmse)
def testPlanesEjecucion(self):
optima = optimize.optimizer({"@nombre": "S30", "@ciudad": "Sevilla"}, self.tabla)
# To change this license header, choose License Headers in Project Properties.
# To change this template file, choose Tools | Templates
# and open the template in the editor.
import optimize
import acquisitions
import reccomenders
import objectives
import scipy as sp
import os
import sys
import logging
logging.basicConfig(level=logging.DEBUG)
sys.path.append('configs')
#import randomsh as optconfig
#import gridsh as optconfig
#import EIMLsh as optconfig
import PESbssh as optconfig
#import EIpumppid as optconfig
#import PESpump as optconfig
#import PBpump as optconfig
O = optimize.optimizer(optconfig.path,optconfig.aqpara,optconfig.aqfn,optconfig.stoppara,optconfig.stopfn,optconfig.reccpara,optconfig.reccfn,optconfig.ojf,optconfig.ojfchar,checkrecc=True)
O.run()
def f(params,args):
return np.sum([term.f(params[i*N*(N+1):(i+1)*N*(N+1)],args[i]) for i in range(iii)],0)
def barrier(params,args):
return np.sum([term.barrier(params[i*N*(N+1):(i+1)*N*(N+1)],args[i]) for i in range(iii)],0)
def df(params,args):
return np.concatenate([term.df(params[i*N*(N+1):(i+1)*N*(N+1)],args[i]) for i in range(iii)])
objective.f = f
objective.df = df
objective.barrier = barrier
def callback_one(ip,d): pass
optimize = optimize.optimizer( objective , callback=callback_one )
#true = [{ 'theta' : np.dot( U.T , V1[i,:] ) , 'M' : 0.1*np.dot( U.T * V1[i,:] , U ) } for i in range(iii)]
#
#for t in true:
# w,v = eig( np.eye(t['M'].shape[0]) - t['M'] )
# print 'eig true M' , w.real
#
#trupar = true
#for i in range(5):
# trupar = optimize(init_params=trupar,args=data)
# callback_one(trupar,data)
#trupar = objective.inflate(trupar)
params = np.concatenate( [term.flatten(ip) for ip in init_params] )
aqpara['ub'], 1.)
ojfn = objectives.costfnwrap(ojf, cfn)
ojfchar = {'dx': len(aqpara['lb']), 'dev': len(aqpara['ev'])}
aqpara['cfn'] = cfn
aqpara['xau'] = 1.
aqpara['xal'] = 0.
#cfn = objectives.cf42
#ojfn,xmin,ymin = objectives.genmat52ojf(len(aqpara['lb']),aqpara['lb'],aqpara['ub'])
#ojfchar = {'dx':len(aqpara['lb']),'dev':len(aqpara['ev'])}
stoppara = {'nmax': 20}
stopfn = optimize.nstopfn
reccfn, reccpara = reccomenders.gpasmap
reccpara['lb'] = aqpara['lb']
reccpara['ub'] = aqpara['ub']
O = optimize.optimizer(path,
aqpara,
aqfn,
stoppara,
stopfn,
reccpara,
reccfn,
ojfn,
ojfchar,
checkrecc=True)
O.run()
def callback( objective , params ):
print 'Objective: ' , objective.f(params) , ' barrier: ', objective.barrier(params)
true = {'U' : U , 'V1': V1 }
data = {'STAs':np.vstack(STA) , 'STCs':np.vstack([stc[np.newaxis,:] for stc in STC]),
'V2':V2*np.ones(Nsub) , 'N':NRGC }
targets = { 'f':quadratic_Poisson, 'barrier':eig_barrier }
targets = kolia_theano.reparameterize(targets,UVs(NRGC))
objective = kolia_theano.Objective( init_params=true, differentiate=['f'],
callback=callback, **targets )
optimizer = optimize.optimizer( objective.where(**data) )
trupar = true
for i in range(2):
trupar = optimizer(init_params=trupar)
trupar = objective.unflat(trupar)
init_params = {'U' : 0.0001+0.05*R.random(size=U.shape ) ,
'V1': 0.0001+0.05*R.random(size=V1.shape) }
params = init_params
for i in range(10):
params = optimizer(init_params=params)
params = objective.unflat(params)
