def test_lbp():
exms = 30
train = 10
x, y, z = get_1d_toy_data(exms, plot=False)
x -= np.mean(x, axis=0)
x = np.hstack([x, np.ones((exms, 1))])
print x.shape
# ...and corresponding transition matrix
A = np.zeros((exms, exms), dtype=np.int32)
for j in range(1, 2):
for i in range(j, exms):
A[i - j, i] = 1
A[i, i - j] = 1
A = co.sparse(co.matrix(A, tc='i'))
inds = np.random.permutation(exms)
uinds = inds[train:]
linds = inds[:train]
cluster = [np.arange(exms)]
qp = TCRFR_QP(x.T.copy(),
y[linds].copy(),
linds,
states=2,
A=A,
reg_gamma=10.,
reg_theta=0.9,
trans_sym=[1])
lbpa = TCRFR_lbpa(x.T.copy(),
y[linds].copy(),
linds,
states=2,
A=A,
reg_gamma=10.,
reg_theta=0.9,
trans_sym=[1])
lbpa_iset = TCRFR_lbpa_iset(cluster,
x.T.copy(),
y[linds].copy(),
linds,
states=2,
A=A,
reg_gamma=10.,
reg_theta=0.9,
trans_sym=[1])
qp.fit()
lbpa.fit()
lbpa_iset.fit()
# lbpa_iset.map_inference(lbpa.u, lbpa.unpack_v(lbpa.v))
lid = lbpa.map_indep(qp.u, lbpa.unpack_v(qp.v))
print 'STATES ------------------------'
foo = np.zeros(lbpa.samples, dtype=np.int8)
foo[lbpa.label_inds] = 1
print 'Train = ', foo
print 'True = ', z
print 'QP = ', qp.latent
print 'LBPA = ', lbpa.latent
print 'LBPAc = ', lbpa_iset.latent
print 'Indep = ', lid
def method_tcrfr_qp(vecX, vecy, train, test, states=2, params=[0.9, 0.00001, 0.5, 10], true_latent=None, plot=False):
A = co.spmatrix(0, [], [], (vecX.shape[0], vecX.shape[0]), tc='d')
for k in range(1, params[3]):
for i in range(vecX.shape[0]-k):
A[i, i+k] = 1
A[i+k, i] = 1
tcrfr = TCRFR_QP(data=vecX.T, labels=vecy[train], label_inds=train, states=states, A=A,
reg_theta=params[0], reg_lambda=params[1], reg_gamma=params[2]*float(len(train)+len(test)),
trans_regs=[[1., 1.]], trans_sym=[1])
tcrfr.fit(max_iter=20, use_grads=False, auto_adjust=False)
y_preds, lats = tcrfr.predict()
print lats
if plot:
plt.figure(1)
plt.subplot(1, 2, 1)
plt.plot(vecX[:, 0], vecy, '.g', alpha=0.1, markersize=10.0)
plt.plot(vecX[test, 0], vecy[test], 'or', alpha=0.6, markersize=10.0)
plt.plot(vecX[test, 0], y_preds[test], 'oc', alpha=0.6, markersize=6.0)
plt.plot(vecX[test, 0], lats[test], 'ob', alpha=0.6, markersize=6.0)
plt.subplot(1, 2, 2)
plt.plot(vecX[train, 0], vecy[train], 'or', alpha=0.6, markersize=10.0)
plt.plot(vecX[train, 0], lats[train], 'ob', alpha=0.6, markersize=6.0)
plt.plot(vecX[train, 0], y_preds[train], 'xg', alpha=0.8, markersize=10.0)
print('Test performance: ')
print evaluate(vecy[test], y_preds[test], true_latent[test], lats[test])
print('Training performance: ')
print evaluate(vecy[train], y_preds[train], true_latent[train], lats[train])
plt.show()
return 'TCRFR-QP', y_preds[test], lats[test]
def method_tcrfr_qp(vecX,
vecy,
train,
test,
states=2,
params=[0.9, 0.00001, 0.5, 10],
true_latent=None,
plot=False):
A = co.spmatrix(0, [], [], (vecX.shape[0], vecX.shape[0]), tc='d')
for k in range(1, params[3]):
for i in range(vecX.shape[0] - k):
A[i, i + k] = 1
A[i + k, i] = 1
tcrfr = TCRFR_QP(data=vecX.T,
labels=vecy[train],
label_inds=train,
states=states,
A=A,
reg_theta=params[0],
reg_lambda=params[1],
reg_gamma=params[2] * float(len(train) + len(test)),
trans_regs=[[1., 1.]],
trans_sym=[1])
tcrfr.fit(max_iter=20, use_grads=False, auto_adjust=False)
y_preds, lats = tcrfr.predict()
print lats
if plot:
plt.figure(1)
plt.subplot(1, 2, 1)
plt.plot(vecX[:, 0], vecy, '.g', alpha=0.1, markersize=10.0)
plt.plot(vecX[test, 0], vecy[test], 'or', alpha=0.6, markersize=10.0)
plt.plot(vecX[test, 0], y_preds[test], 'oc', alpha=0.6, markersize=6.0)
plt.plot(vecX[test, 0], lats[test], 'ob', alpha=0.6, markersize=6.0)
plt.subplot(1, 2, 2)
plt.plot(vecX[train, 0], vecy[train], 'or', alpha=0.6, markersize=10.0)
plt.plot(vecX[train, 0], lats[train], 'ob', alpha=0.6, markersize=6.0)
plt.plot(vecX[train, 0],
y_preds[train],
'xg',
alpha=0.8,
markersize=10.0)
print('Test performance: ')
print evaluate(vecy[test], y_preds[test], true_latent[test],
lats[test])
print('Training performance: ')
print evaluate(vecy[train], y_preds[train], true_latent[train],
lats[train])
plt.show()
return 'TCRFR-QP', y_preds[test], lats[test]
def test_cluster_fixed_latent_states():
print "Test Cluster setting."
# setup cluster 0 = labeled, 1 = unlabeled (2 latent states)
cluster = [np.arange(0, 32)]
map_types = [TCRFR_lbpa_iset.MAP_ISET_FULL]
labeled_inds = np.random.permutation(32)
labeled_inds = labeled_inds[:8]
qp = TCRFR_QP(x.T, y[labeled_inds], labeled_inds, states=2, A=A, \
reg_theta=0.8, reg_gamma=1., trans_sym=[1])
qp.fit(use_grads=False)
lbpa_iset = TCRFR_lbpa_iset(cluster, map_types, x.T, y[labeled_inds], labeled_inds, states=2, A=A, \
reg_theta=0.8, reg_gamma=1., trans_sym=[1], verbosity_level=3)
fixed = -np.ones(y.size, dtype=np.int)
fixed[4:10] = z[4:10]
lbpa_iset.set_fixed_latent_states(fixed)
lbpa_iset.fit(use_grads=False)
print '------------'
inds = np.zeros(y.size, dtype=np.int8)
inds[labeled_inds] = 1
print inds
print '------------'
print fixed
print '------------'
print 'Truth: ', z
print 'Lbpa : ', lbpa_iset.latent
print 'Qp : ', qp.latent
print '------------'
print lbpa_iset.log_partition(qp.unpack_v(qp.v))
print qp.log_partition_pl(qp.unpack_v(qp.v))
print '------------'
def test_cluster_setting():
print "Test Cluster setting."
# setup cluster 0 = labeled, 1 = unlabeled (2 latent states)
cluster = [np.arange(0, 10), np.arange(10, 16)]
A[9, 10] = 0
A[10, 9] = 0
print z
assert any(z[:6] == 0) and any(z[6:16] == 1)
labeled_inds = np.arange(0, 10)
qp = TCRFR_QP(x.T, y[labeled_inds], labeled_inds, states=2, A=A, \
reg_theta=0.6, reg_gamma=1., trans_sym=[1])
qp.fit(use_grads=False)
map_types = [TCRFR_lbpa_iset.MAP_ISET_FULL, TCRFR_lbpa_iset.MAP_ISET_FULL]
lbpa_iset = TCRFR_lbpa_iset(cluster, map_types, x.T, y[labeled_inds], labeled_inds, states=2, A=A, \
reg_theta=0.6, reg_gamma=1., trans_sym=[1])
lbpa_iset.fit(use_grads=False)
print '------------'
inds = np.zeros(y.size, dtype=np.int8)
inds[labeled_inds] = 1
print inds
print '------------'
print lbpa_iset.latent
print qp.latent
print z
print '------------'
print lbpa_iset.log_partition_pl(qp.unpack_v(qp.v))
print qp.log_partition_pl(qp.unpack_v(qp.v))
print '------------'
def test_cluster_large_setting():
print "Test Cluster setting."
# setup cluster 0 = labeled, 1 = unlabeled (2 latent states)
cluster = [np.arange(0, 2501), np.arange(2501, 5000)]
A[2500, 2501] = 0
A[2501, 2500] = 0
labeled_inds = np.random.permutation(y.size)
labeled_inds = labeled_inds[:1500]
qp = TCRFR_QP(x.T, y[labeled_inds], labeled_inds, states=2, A=A, \
reg_theta=0.9, reg_gamma=1., trans_sym=[1])
#qp.fit(use_grads=False)
map_types = [TCRFR_lbpa_iset.MAP_ISET_FULL, TCRFR_lbpa_iset.MAP_ISET_FULL]
lbpa_iset = TCRFR_lbpa_iset(cluster, map_types, x.T, y[labeled_inds], labeled_inds, states=2, A=A, \
reg_theta=0.9, reg_gamma=1., trans_sym=[1])
lbpa_iset.fit(use_grads=False)
print '------------'
print np.sum(np.abs(lbpa_iset.latent - z))
print np.sum(np.abs((1 - lbpa_iset.latent) - z))
print '------------'
print np.sum(np.abs(qp.latent - z))
print np.sum(np.abs((1 - qp.latent) - z))
print '------------'
def test_cluster_fixed_latent_states():
print "Test Cluster setting."
# setup cluster 0 = labeled, 1 = unlabeled (2 latent states)
cluster = [np.arange(0, 32)]
map_types = [TCRFR_lbpa_iset.MAP_ISET_FULL]
labeled_inds = np.random.permutation(32)
labeled_inds = labeled_inds[:8]
qp = TCRFR_QP(x.T, y[labeled_inds], labeled_inds, states=2, A=A, \
reg_theta=0.8, reg_gamma=1., trans_sym=[1])
qp.fit(use_grads=False)
lbpa_iset = TCRFR_lbpa_iset(cluster, map_types, x.T, y[labeled_inds], labeled_inds, states=2, A=A, \
reg_theta=0.8, reg_gamma=1., trans_sym=[1], verbosity_level=3)
fixed = -np.ones(y.size, dtype=np.int)
fixed[4:10] = z[4:10]
lbpa_iset.set_fixed_latent_states(fixed)
lbpa_iset.fit(use_grads=False)
print '------------'
inds = np.zeros(y.size, dtype=np.int8)
inds[labeled_inds] = 1
print inds
print '------------'
print fixed
print '------------'
print 'Truth: ', z
print 'Lbpa : ', lbpa_iset.latent
print 'Qp : ', qp.latent
print '------------'
print lbpa_iset.log_partition(qp.unpack_v(qp.v))
print qp.log_partition_pl(qp.unpack_v(qp.v))
print '------------'
def test_cluster_setting():
print "Test Cluster setting."
# setup cluster 0 = labeled, 1 = unlabeled (2 latent states)
cluster = [np.arange(0, 10), np.arange(10, 16)]
A[9, 10] = 0
A[10, 9] = 0
print z
assert any(z[:6]==0) and any(z[6:16]==1)
labeled_inds = np.arange(0, 10)
qp = TCRFR_QP(x.T, y[labeled_inds], labeled_inds, states=2, A=A, \
reg_theta=0.6, reg_gamma=1., trans_sym=[1])
qp.fit(use_grads=False)
map_types = [TCRFR_lbpa_iset.MAP_ISET_FULL, TCRFR_lbpa_iset.MAP_ISET_FULL]
lbpa_iset = TCRFR_lbpa_iset(cluster, map_types, x.T, y[labeled_inds], labeled_inds, states=2, A=A, \
reg_theta=0.6, reg_gamma=1., trans_sym=[1])
lbpa_iset.fit(use_grads=False)
print '------------'
inds = np.zeros(y.size, dtype=np.int8)
inds[labeled_inds] = 1
print inds
print '------------'
print lbpa_iset.latent
print qp.latent
print z
print '------------'
print lbpa_iset.log_partition_pl(qp.unpack_v(qp.v))
print qp.log_partition_pl(qp.unpack_v(qp.v))
print '------------'
def test_transition_conversion():
print "Test transition conversion."
num_exms = 100
num_edges = 200
# create a (valid) random transition matrix
edges = np.zeros((num_edges, 3), dtype=np.int64)
neighbors = np.zeros(num_edges, dtype=np.int32)
A = co.spmatrix(0, [], [], (num_exms, num_exms), tc='d')
cnt_edges = 0
while cnt_edges < num_edges:
e1 = np.random.randint(0, num_exms)
e2 = np.random.randint(0, num_exms)
if e1 != e2 and A[e1, e2] == 0:
val = np.random.randint(1, 3)
A[e1, e2] = val
A[e2, e1] = val
edges[cnt_edges, :] = (e1, e2, val)
neighbors[e1] += 1
neighbors[e2] += 1
cnt_edges += 1
qp = TCRFR_QP(x.T, y[labeled_inds], labeled_inds, states=2, A=A, \
reg_theta=0.9, reg_gamma=1., trans_sym=[1])
assert qp.N.shape[1] == np.max(neighbors)
assert qp.E.shape[0] == num_edges # correct number of edges detected?
# check if the correct edges are in the edge-array
for (e1, e2, t) in edges:
cnt_1 = 0
cnt_2 = 0
for e in range(qp.E.shape[0]):
# check edges in both direction (only one should be in there)
if qp.E[e, 0] == e1 and qp.E[e, 1] == e2 and qp.E[e, 2] == t:
cnt_1 += 1
if qp.E[e, 0] == e2 and qp.E[e, 1] == e1 and qp.E[e, 2] == t:
cnt_2 += 1
assert cnt_1 == 1 or cnt_2 == 1
assert cnt_1 + cnt_2 == 1
for i in range(num_exms):
assert neighbors[i] == np.sum(qp.N_weights[i, :])
for (e1, e2, t) in edges:
if e1 == i or e2 == i:
if e2 == i:
e2 = e1
e1 = i
assert e2 in qp.N[i, :neighbors[i]]
ind = np.where(qp.N[i, :neighbors[i]] == e2)[0][0]
# print qp.N_inv[i, ind]
assert qp.N[e2, qp.N_inv[i, ind]] == i
def test_lbp():
exms = 30
train = 10
x, y, z = get_1d_toy_data(exms, plot=False)
x -= np.mean(x, axis=0)
x = np.hstack([x, np.ones((exms, 1))])
print x.shape
# ...and corresponding transition matrix
A = np.zeros((exms, exms), dtype=np.int32)
for j in range(1, 2):
for i in range(j, exms):
A[i-j, i] = 1
A[i, i-j] = 1
A = co.sparse(co.matrix(A, tc='i'))
inds = np.random.permutation(exms)
uinds = inds[train:]
linds = inds[:train]
cluster = [np.arange(exms)]
qp = TCRFR_QP(x.T.copy(), y[linds].copy(), linds, states=2, A=A, reg_gamma=10., reg_theta=0.9, trans_sym=[1])
lbpa = TCRFR_lbpa(x.T.copy(), y[linds].copy(), linds, states=2, A=A, reg_gamma=10., reg_theta=0.9, trans_sym=[1])
lbpa_iset = TCRFR_lbpa_iset(cluster, x.T.copy(), y[linds].copy(), linds, states=2, A=A, reg_gamma=10., reg_theta=0.9, trans_sym=[1])
qp.fit()
lbpa.fit()
lbpa_iset.fit()
# lbpa_iset.map_inference(lbpa.u, lbpa.unpack_v(lbpa.v))
lid = lbpa.map_indep(qp.u, lbpa.unpack_v(qp.v))
print 'STATES ------------------------'
foo = np.zeros(lbpa.samples, dtype=np.int8)
foo[lbpa.label_inds] = 1
print 'Train = ', foo
print 'True = ', z
print 'QP = ', qp.latent
print 'LBPA = ', lbpa.latent
print 'LBPAc = ', lbpa_iset.latent
print 'Indep = ', lid
def get_test_data(exms, train):
# generate toy niidbox-data
x, y, z = get_1d_toy_data(exms, plot=False)
y -= np.mean(y, axis=0)
x -= np.mean(x, axis=0)
x = np.hstack([x, np.ones((exms, 1))])
print x.shape
# ...and corresponding transition matrix
A = np.zeros((exms, exms), dtype=np.int32)
for j in range(1, 2):
for i in range(j, exms):
A[i - j, i] = 1
A[i, i - j] = 1
A = co.sparse(co.matrix(A))
inds = np.random.permutation(exms)
linds = inds[:train]
lbpa = TCRFR_lbpa(x.T.copy(),
y[linds].copy(),
linds,
states=2,
A=A,
reg_gamma=10.,
reg_theta=0.8,
trans_sym=[1])
qp = TCRFR_QP(x.T.copy(),
y[linds].copy(),
linds,
states=2,
A=A,
reg_gamma=10.,
reg_theta=0.8,
trans_sym=[1])
bf = TCRFR_BF(x.T.copy(),
y[linds].copy(),
linds,
states=2,
A=A,
reg_gamma=10.,
reg_theta=0.8,
trans_sym=[1])
return lbpa, qp, bf, x, y, z
def test_smiley():
# x = np.loadtxt('../../Projects/si.txt')
# y = np.loadtxt('../../Projects/phi.txt')
# z = np.loadtxt('../../Projects/facies.txt')
# height, width = x.shape
# exms = x.size
#
# x = x.reshape((x.size, 1), order='C')
# y = y.reshape(y.size, order='C')
# z = z.reshape(z.size, order='C')
#
# x[np.where(z==0)] -= 0.9
#
# y -= np.mean(y, axis=0)
# x -= np.mean(x, axis=0)
# y /= np.max(np.abs(y))
# y *= 1.0
# x /= np.max(np.abs(x))
#
# x = np.hstack([x, np.ones((exms, 1))])
# print x.shape, y.shape, z.shape
data = np.load('niidbox-data/data_smiley.npz')
x = data['x']
y = data['y']
z = data['latent']
width = data['width']
height = data['height']
exms = x.shape[0]
linds = np.random.permutation(exms)[:np.int(0.1 * exms)]
A = co.spmatrix(0, [], [], (exms, exms), tc='d')
for k in range(1, 3):
for i in range(height):
for j in range(width):
idx = i * width + j
idx1 = (i + k) * width + j
idx2 = i * width + j + k
if k == 1 or (k > 1 and idx in linds) or (
k > 1 and idx1 in linds) or (k > 1 and idx2 in linds):
if i < height - k:
A[idx, idx1] = 1
A[idx1, idx] = 1
if j < width - k:
A[idx, idx2] = 1
A[idx2, idx] = 1
# qp = TCRFR_QP(x.T.copy(), y[linds].copy(), linds, states=2, A=A,
# reg_gamma=10000., reg_theta=0.95, trans_sym=[1], trans_regs=[[20., 4.]])
yyy = 1.0 * np.ones(x.shape[0])
print x.shape
qp = TCRFR_QP(x[:, 0].reshape((x.shape[0], 1)).T.copy(),
yyy.copy(),
np.arange(x.shape[0]),
states=2,
A=A,
reg_lambda=2. * (x.shape[0] * 0.99),
reg_gamma=10000.,
reg_theta=0.85,
trans_sym=[1],
trans_regs=[[20., 4.]])
qp.set_log_partition(qp.LOGZ_PL_SUM)
# u = np.random.randn(qp.get_num_feats()*qp.S)
# v = np.random.randn(qp.get_num_compressed_dims())
#np.savez('../../Projects/hotstart.npz', start=(u, v, linds))
(u, v, linds) = np.load('../../Projects/hotstart.npz')['start']
# qp.fit(use_grads=False, hotstart=(u, v), auto_adjust=True)
qp.fit(use_grads=False, hotstart=None, auto_adjust=False)
lbpa = TCRFR_lbpa(x.T.copy(),
y[linds].copy(),
linds,
states=2,
A=A,
reg_gamma=10000.,
reg_theta=0.49995,
trans_sym=[1],
trans_regs=[[20., 4.]])
lbpa.verbosity_level = 3
lbpa.set_log_partition(lbpa.LOGZ_PL_SUM)
lbpa.fit(use_grads=False, hotstart=(u, v), auto_adjust=False)
#lbpa.fit(use_grads=False, hotstart=None, auto_adjust=True)
#lbpa.map_inference(qp.u, lbpa.unpack_v(qp.v))
# initialize all non-fixed latent variables with random states
import sklearn.cluster as cl
kmeans = cl.KMeans(n_clusters=2,
init='random',
n_init=4,
max_iter=100,
tol=0.0001)
kmeans.fit(x)
import matplotlib.pyplot as plt
plt.figure(1)
plt.subplot(1, 7, 1)
plt.imshow(y.reshape((height, width), order='C'))
plt.subplot(1, 7, 2)
plt.imshow(x[:, 0].reshape((height, width), order='C'))
plt.subplot(1, 7, 3)
plt.imshow(z.reshape((height, width), order='C'))
plt.subplot(1, 7, 4)
plt.imshow(lbpa.latent.reshape((height, width), order='C'))
plt.title('LBPA')
plt.subplot(1, 7, 5)
plt.imshow(kmeans.labels_.reshape((height, width), order='C'))
plt.title('KMeans')
plt.subplot(1, 7, 6)
plt.imshow(qp.latent.reshape((height, width), order='C'))
plt.title('QP')
plt.subplot(1, 7, 7)
res, _ = lbpa.predict()
plt.title('LBPA prediction')
print 'Labels:', linds.size
print 'RESULT:', np.sum((res - y) * (res - y)) / y.size
plt.imshow(res.reshape((height, width), order='C'))
plt.show()
def test_smiley():
# x = np.loadtxt('../../Projects/si.txt')
# y = np.loadtxt('../../Projects/phi.txt')
# z = np.loadtxt('../../Projects/facies.txt')
# height, width = x.shape
# exms = x.size
#
# x = x.reshape((x.size, 1), order='C')
# y = y.reshape(y.size, order='C')
# z = z.reshape(z.size, order='C')
#
# x[np.where(z==0)] -= 0.9
#
# y -= np.mean(y, axis=0)
# x -= np.mean(x, axis=0)
# y /= np.max(np.abs(y))
# y *= 1.0
# x /= np.max(np.abs(x))
#
# x = np.hstack([x, np.ones((exms, 1))])
# print x.shape, y.shape, z.shape
data = np.load('niidbox-data/data_smiley.npz')
x = data['x']
y = data['y']
z = data['latent']
width = data['width']
height = data['height']
exms = x.shape[0]
linds = np.random.permutation(exms)[:np.int(0.1*exms)]
A = co.spmatrix(0, [], [], (exms, exms), tc='d')
for k in range(1, 3):
for i in range(height):
for j in range(width):
idx = i*width + j
idx1 = (i+k)*width + j
idx2 = i*width + j + k
if k == 1 or (k>1 and idx in linds) or (k>1 and idx1 in linds) or (k>1 and idx2 in linds):
if i < height-k:
A[idx, idx1] = 1
A[idx1, idx] = 1
if j < width-k:
A[idx, idx2] = 1
A[idx2, idx] = 1
# qp = TCRFR_QP(x.T.copy(), y[linds].copy(), linds, states=2, A=A,
# reg_gamma=10000., reg_theta=0.95, trans_sym=[1], trans_regs=[[20., 4.]])
yyy = 1.0*np.ones(x.shape[0])
print x.shape
qp = TCRFR_QP(x[:, 0].reshape((x.shape[0], 1)).T.copy(), yyy.copy(), np.arange(x.shape[0]), states=2, A=A,
reg_lambda=2.*(x.shape[0]*0.99), reg_gamma=10000., reg_theta=0.85, trans_sym=[1], trans_regs=[[20., 4.]])
qp.set_log_partition(qp.LOGZ_PL_SUM)
# u = np.random.randn(qp.get_num_feats()*qp.S)
# v = np.random.randn(qp.get_num_compressed_dims())
#np.savez('../../Projects/hotstart.npz', start=(u, v, linds))
(u, v, linds) = np.load('../../Projects/hotstart.npz')['start']
# qp.fit(use_grads=False, hotstart=(u, v), auto_adjust=True)
qp.fit(use_grads=False, hotstart=None, auto_adjust=False)
lbpa = TCRFR_lbpa(x.T.copy(), y[linds].copy(), linds, states=2, A=A,
reg_gamma=10000., reg_theta=0.49995, trans_sym=[1], trans_regs=[[20., 4.]])
lbpa.verbosity_level = 3
lbpa.set_log_partition(lbpa.LOGZ_PL_SUM)
lbpa.fit(use_grads=False, hotstart=(u, v), auto_adjust=False)
#lbpa.fit(use_grads=False, hotstart=None, auto_adjust=True)
#lbpa.map_inference(qp.u, lbpa.unpack_v(qp.v))
# initialize all non-fixed latent variables with random states
import sklearn.cluster as cl
kmeans = cl.KMeans(n_clusters=2, init='random', n_init=4, max_iter=100, tol=0.0001)
kmeans.fit(x)
import matplotlib.pyplot as plt
plt.figure(1)
plt.subplot(1, 7, 1)
plt.imshow(y.reshape((height, width), order='C'))
plt.subplot(1, 7, 2)
plt.imshow(x[:, 0].reshape((height, width), order='C'))
plt.subplot(1, 7, 3)
plt.imshow(z.reshape((height, width), order='C'))
plt.subplot(1, 7, 4)
plt.imshow(lbpa.latent.reshape((height, width), order='C'))
plt.title('LBPA')
plt.subplot(1, 7, 5)
plt.imshow(kmeans.labels_.reshape((height, width), order='C'))
plt.title('KMeans')
plt.subplot(1, 7, 6)
plt.imshow(qp.latent.reshape((height, width), order='C'))
plt.title('QP')
plt.subplot(1, 7, 7)
res, _ = lbpa.predict()
plt.title('LBPA prediction')
print 'Labels:', linds.size
print 'RESULT:', np.sum((res - y)*(res - y))/y.size
plt.imshow(res.reshape((height, width), order='C'))
plt.show()