def gpmaprecc(optstate, persist, **para):
if para['onlyafter'] >= optstate.n or not optstate.n % para['everyn'] == 0:
return argminrecc(optstate, persist, **para)
#return [sp.NaN for i in para['lb']],{'didnotrun':True}
logger.info('gpmap reccomender')
d = len(para['lb'])
lb = para['lb']
ub = para['ub']
maxf = para['maxf']
x = sp.vstack(optstate.x)
y = sp.vstack(optstate.y)
s = sp.vstack([e['s'] + 10**optstate.condition for e in optstate.ev])
dx = [e['d'] for e in optstate.ev]
MAP = GPdc.searchMAPhyp(x, y, s, dx, para['mprior'], para['sprior'],
para['kindex'])
logger.info('MAPHYP {}'.format(MAP))
G = GPdc.GPcore(x, y, s, dx, GPdc.kernel(para['kindex'], d, MAP))
def wrap(x):
xq = copy.copy(x)
xq.resize([1, d])
a = G.infer_m_post(xq, [[sp.NaN]])
return a[0, 0]
xmin, ymin, ierror = gpbo.core.optutils.twopartopt(wrap, para['lb'],
para['ub'],
para['dpara'],
para['lpara'])
return [i for i in xmin], persist, {'MAPHYP': MAP, 'ymin': ymin}
def traincfn1dll(x, c):
#cost modeled in a latent space c=exp(cl)
n = x.size
cl = sp.log(c)
MAP = GPdc.searchMAPhyp(x, cl, sp.array([1e-3] * n), [[sp.NaN]] * n,
sp.array([1., 0., -1.]), sp.array([2., 2., 2.]),
GPdc.MAT52CS)
print('MAPhyp in costfn {}'.format(MAP))
g = GPdc.GPcore(x, cl, sp.array([1e-3] * n), [[sp.NaN]] * n,
GPdc.kernel(GPdc.MAT52CS, 1, MAP))
if gpbo.core.debugoutput['cost1d']:
print('plotting cost1d...')
import time
from matplotlib import pyplot as plt
f, a = plt.subplots(2)
low = min(0, min(x))
high = max(1, max(x))
xaxis = sp.linspace(low, high, 100)
y, cy = g.infer_diag_post(xaxis, [[sp.NaN]] * 100)
s = 2. * sp.sqrt(cy)
u = sp.empty(100)
l = sp.empty(100)
for i in xrange(100):
s = sp.sqrt(cy[0, i])
u[i] = y[0, i] + 2. * s
l[i] = y[0, i] - 2. * s
a[0].plot(xaxis, y[0, :], 'b')
a[0].fill_between(xaxis,
l,
u,
facecolor='lightblue',
edgecolor='lightblue',
alpha=0.5)
for i in xrange(n):
a[0].plot(x[i], cl[i], 'r.')
a[0].set_ylabel('latent')
a[1].plot(xaxis, sp.exp(y[0, :]), 'b')
a[1].fill_between(xaxis,
sp.exp(l),
sp.exp(u),
facecolor='lightblue',
edgecolor='lightblue',
alpha=0.5)
for i in xrange(n):
a[1].plot(x[i], c[i], 'r.')
a[1].set_ylabel('out')
f.savefig(
os.path.join(
gpbo.core.debugoutput['path'],
'cost1d' + time.strftime('%d_%m_%y_%H:%M:%S') + '.png'))
f.clf()
plt.close(f)
del (f)
return logcfnobj(g)
def gpmap2upperrecc(optstate, persist, **para):
if para['onlyafter'] >= optstate.n:
print('{} <= {} : switch to argmin'.format(optstate.n,
para['onlyafter']))
return argminrecc(optstate, persist, **para)
#return [sp.NaN for i in para['lb']],{'didnotrun':True}
logger.info('gpmapucb2 reccomender')
d = len(para['lb'])
x = sp.vstack(optstate.x)
y = sp.vstack(optstate.y)
s = sp.vstack([e['s'] + 10**optstate.condition for e in optstate.ev])
dx = [e['d'] for e in optstate.ev]
MAP = GPdc.searchMAPhyp(x, y, s, dx, para['mprior'], para['sprior'],
para['kindex'])
logger.info('MAPHYP {}'.format(MAP))
G = GPdc.GPcore(x, y, s, dx, GPdc.kernel(para['kindex'], d, MAP))
count = 0
def wrap(xq):
xq.resize([1, d])
a, v = G.infer_diag_post(xq, [[sp.NaN]])
return a[0, 0] + 2. * sp.sqrt(v[0, 0])
#print('nevals={}\n\n'.format(count))
#[xmin,ymin,ierror] = direct(directwrap,para['lb'],para['ub'],user_data=[], algmethod=1, maxf=para['maxf'], logfilename='/dev/null')
xmin, ymin, ierror = gpbo.core.optutils.twopartopt(wrap, para['lb'],
para['ub'],
para['dpara'],
para['lpara'])
logger.info('DIRECT found post. min {} at {} {}'.format(
ymin, xmin, ierror))
a, v = G.infer_diag_post(sp.array([[xmin]]), [[sp.NaN]])
print('mean {} std{} '.format(a[0, 0], sp.sqrt(v[0, 0])))
x, p, di = argminrecc(optstate, persist, **para)
a, v = G.infer_diag_post(sp.array([[x]]), [[sp.NaN]])
print('at argmin mean {} std {} wasactually {}'.format(
a[0, 0], sp.sqrt(v[0, 0]), di['yinc']))
import sys
sys.stdout.flush()
return [i for i in xmin], persist, {'MAPHYP': MAP, 'ymin': ymin}
def gpmapasrecc(optstate, persist, **para):
if para['onlyafter'] >= optstate.n or not optstate.n % para['everyn'] == 0:
return argminrecc(optstate, persist, **para)
#return [sp.NaN for i in para['lb']],{'didnotrun':True}
logger.info('gpmapas reccomender')
d = len(para['lb'])
x = sp.hstack(
[sp.vstack(optstate.x),
sp.vstack([e['xa'] for e in optstate.ev])])
y = sp.vstack(optstate.y)
s = sp.vstack([e['s'] + 10**optstate.condition for e in optstate.ev])
dx = [e['d'] for e in optstate.ev]
MAP = GPdc.searchMAPhyp(x, y, s, dx, para['mprior'], para['sprior'],
para['kindex'])
logger.info('MAPHYP {}'.format(MAP))
G = GPdc.GPcore(x, y, s, dx, GPdc.kernel(para['kindex'], d + 1, MAP))
def wrap(xq):
xq.resize([1, d])
xe = sp.hstack([xq, sp.array([[0.]])])
a = G.infer_m_post(xe, [[sp.NaN]])
return a[0, 0]
#[xmin,ymin,ierror] = direct(directwrap,para['lb'],para['ub'],user_data=[], algmethod=1, maxf=para['maxf'], logfilename='/dev/null')
xmin, ymin, ierror = gpbo.core.optutils.twopartopt(wrap, para['lb'],
para['ub'],
para['dpara'],
para['lpara'])
logger.info('reccsearchresult: {}'.format([xmin, ymin, ierror]))
from gpbo.core import debugoutput
if debugoutput['datavis']:
A2 = MAP[0]
l = MAP[3]
from matplotlib import pyplot as plt
fig, ax = plt.subplots(nrows=3, ncols=1, figsize=(10, 30))
n = 200
x_ = sp.linspace(-1, 1, n)
y_ = sp.linspace(-1, 1, n)
z_ = sp.empty([n, n])
s_ = sp.empty([n, n])
for i in xrange(n):
for j in xrange(n):
m_, v_ = G.infer_diag_post(sp.array([y_[j], x_[i], 0.]),
[[sp.NaN]])
z_[i, j] = m_[0, 0]
s_[i, j] = sp.sqrt(v_[0, 0])
CS = ax[1].contour(x_, y_, z_, 20)
ax[1].clabel(CS, inline=1, fontsize=10)
CS = ax[2].contour(x_, y_, s_, 20)
ax[2].clabel(CS, inline=1, fontsize=10)
for i in xrange(x.shape[0]):
ax[0].plot(x[i, 0], x[i, 1], 'b.')
circle = plt.Circle([x[i, 0], x[i, 1]],
radius=0.5 * x[i, 2] * l,
edgecolor="none",
color='lightblue',
alpha=0.8 - 0.6 * x[i, 2])
ax[0].add_patch(circle)
ax[0].axis([-1., 1., -1., 1.])
fig.savefig(
os.path.join(
os.path.expanduser('~'), 'Dropbox/debugoutput',
'datavis' + time.strftime('%d_%m_%y_%H:%M:%S') + '.png'))
fig.clf()
plt.close(fig)
del (fig)
return [i for i in xmin], persist, {'MAPHYP': MAP, 'ymin': ymin}
def traincfnfull(x, c):
#cost modeled in a latent space c=exp(cl)
n, d = x.shape
c_ = sp.log(c)
off = sp.mean(c_)
cl = c_ - off
MAP = GPdc.searchMAPhyp(x, cl, sp.array([1e-6] * n), [[sp.NaN]] * n,
sp.array([1.] + [-0.] * d),
sp.array([2.] * (d + 1)), GPdc.MAT52)
print('MAPhyp in costfn {}'.format(MAP))
g = GPdc.GPcore(x, cl, sp.array([1e-3] * n), [[sp.NaN]] * n,
GPdc.kernel(GPdc.MAT52, 1, MAP))
if gpbo.core.debugoutput['cost1d']:
print('plotting cost...')
import time
from matplotlib import pyplot as plt
f, a = plt.subplots(d, 2)
n = 60
x = sp.linspace(-1, 1, n)
xa = sp.linspace(0, 1, n)
z = sp.empty([n, n])
vz = sp.empty([n, n])
for D in xrange(d - 1):
for i in xrange(n):
for j in xrange(n):
q = sp.zeros([1, d])
q[0, 0] = xa[j]
q[0, D + 1] = x[i]
m, v = g.infer_diag(q, [[sp.NaN]])
z[i, j] = m[0, 0]
vz[i, j] = v[0, 0]
try:
CS = a[D, 0].contour(xa, x, z, 30)
a[D, 0].clabel(CS, inline=1, fontsize=8)
except ValueError:
pass
try:
CS = a[D, 1].contour(xa, x, vz, 30)
a[D, 1].clabel(CS, inline=1, fontsize=8)
except ValueError:
pass
for i in xrange(n):
for j in xrange(n):
q = sp.zeros([1, d])
q[0, 0] = 0.
q[0, 1] = x[j]
q[0, 2] = x[i]
m, v = g.infer_diag(q, [[sp.NaN]])
z[i, j] = m[0, 0]
vz[i, j] = v[0, 0]
try:
CS = a[d - 1, 0].contour(x, x, z, 30)
a[d - 1, 0].clabel(CS, inline=1, fontsize=8)
except ValueError:
pass
try:
CS = a[d - 1, 1].contour(x, x, vz, 30)
a[d - 1, 1].clabel(CS, inline=1, fontsize=8)
except ValueError:
pass
f.savefig(
os.path.join(
debugoutput['path'],
'cost1d' + time.strftime('%d_%m_%y_%H:%M:%S') + '.png'))
f.clf()
plt.close(f)
del (f)
return logcfnobjfull(g, offset=off)