def cluster_experiment():
p = [0.33,0.33,0.34]
max_k = 5
n_factors = 1
res = {'cv':np.zeros(max_k-1),'ca':np.zeros(max_k-1),'cstd':np.zeros(max_k-1),'aic':np.zeros(max_k-1),'bic':np.zeros(max_k-1)}
result_list = []
count = 100
distance_dist = np.zeros((1, max_k-1),dtype=int)
for X, Y in sample_generator(2, p, 600):
print count
result = {'id':count}
# recorde the input data
# np.savetxt("intput_data_%d.csv" % count, X, delimiter=',')
# fit mfa model with validation
mfa_cluster = MFACluster(X,max_k,n_factors,Y,iters=10)
mfa_cluster.fit()
res['cv'][mfa_cluster.best_k("voting")-2] += 1
res['ca'][mfa_cluster.best_k("averaging")-2] += 1
res['cstd'][mfa_cluster.best_k("std")-2] += 1
result['cv'] = mfa_cluster.best_k("voting")
result['ca'] = mfa_cluster.best_k("averaging")
result['cstd'] = mfa_cluster.best_k("std")
distance_dist = np.vstack((distance_dist, mfa_cluster.result_matrix))
# fit mfa model to whole dataset with aic and bic
min_aic = min_bic = sys.maxint
k_aic = k_bic = 0
for k in xrange(2, max_k+1):
m = MixtureFA(k, n_factors)
m.fit(X)
result['lik_%d' % k] = m.ll[-1]
result['aic_%d' % k] = m.aic()
result['bic_%d' % k] = m.bic()
if m.aic() < min_aic:
k_aic = k
min_aic = m.aic()
if m.bic() < min_bic:
k_bic = k
min_bic = m.bic()
res['aic'][k_aic-2] += 1
res['bic'][k_bic-2] += 1
result['aic'] = k_aic
result['bic'] = k_bic
# collect result
# save iter number, likelihood_k, aic_k, bic_k, cv, ca, cstd into csv file
result_list.append(result)
count += 1
save_dict_to_csv(result_list)
np.savetxt(os.path.join('.','result',"distance_dist.csv"), distance_dist, delimiter=',')
print res
def mfa_experiment(n_mixtures, output=False):
p = [0.1, 0.3, 0.6]
mus = []
pis = []
pis_diff = []
phis = []
lls = []
lambdas = []
count = 0
for X, ratio in sample_generator(1, p, 100):
count += 1
mfa = MixtureFA(n_mixtures, 1)
mfa.fit(X)
# np.savetxt("data.csv", X, delimiter=',')
mus.append(mfa.mu)
pis.append(mfa.pi)
phis.append(mfa.phi)
pis_diff.append(np.abs(mfa.pi - np.array(ratio)))
lambdas.append(np.array([l.dot(l.T) for l in np.vsplit(mfa.lambdas,n_mixtures)]))
# pdb.set_trace()
lls.append(mfa.ll[-1])
if count % 10 == 0:
print("finished %d" % count)
print mfa.mu
print mfa.pi
if output:
with open(os.path.join('.','result_seed0.txt'), 'w') as outfile:
outfile.write("Mean of MU\n")
outfile.write(np.array2string(np.mean(np.array(mus), axis=0), separator=',') + '\n')
# outfile.write("std of MU\n")
# outfile.write(np.array2string(np.std(np.array(mus), axis=0), separator=',') + '\n')
outfile.write("Mean of lambda.T.dot(lambda)\n")
outfile.write(np.array2string(np.mean(np.array(lambdas), axis=0), separator=',') + '\n')
# outfile.write("std of lambda.T.dot(lambda)\n")
# outfile.write(np.array2string(np.std(np.array(lambdas), axis=0), separator=',') + '\n')
outfile.write("Mean of Phi\n")
outfile.write(np.array2string(np.mean(np.array(phis), axis=0), separator=',') + '\n')
# outfile.write("std of Phi\n")
# outfile.write(np.array2string(np.std(np.array(phis), axis=0), separator=',') + '\n')
outfile.write("Mean of Pi\n")
outfile.write(np.array2string(np.mean(np.array(pis), axis=0), separator=',') + '\n')
def fit(self):
iteration = 0
permu_list = range(self.n_obs)
for iteration in xrange(self.n_iters):
m1_data, m2_data, validation, m1_label, m2_label, val_label = self._split_data(permu_list)
# paralle model fit?
for k in xrange(2, self.max_k + 1):
print ("%d's iteration with k=%d" % (iteration, k))
m1 = MixtureFA(k, self.n_factors)
m1.fit(m1_data)
m2 = MixtureFA(k, self.n_factors)
m2.fit(m2_data)
self.result_matrix[iteration, k - 2] = self._mismatch(m1, m2, validation)
