def sample_gp_with_slice( gp, params ):
logposterior_generating_function_wrt_params = progapy.gp.generate_gp_logposterior_using_params
p = gp.get_params()
thetas = np.zeros( (params["N"],len(gp.get_params())))
if params.has_key("ids"):
ids = params["ids"]
else:
ids = range(len(gp.get_params()))
cur_p = gp.get_params()
L,R,stepsizes = gp.get_range_of_params()
for n in range(params["N"]):
for i in ids:
X, logprob = slice_sample( logposterior_generating_function_wrt_params(gp, idx=i), \
i, \
cur_p[i], \
L[i], \
R[i], \
stepsizes[i], \
params["nbrSteps"], \
params["MODE"] )
cur_p[i] = X[-1]
thetas[n,:] = cur_p
if params.has_key("set_to_last_sample"):
if params["set_to_last_sample"]:
gp.set_params(thetas[-1])
else:
gp.set_params(p)
else:
gp.set_params(p)
return thetas
def drawhyp_plk(X, Y, S, D, ki, hm, hs, n, burn=80, subsam=5):
def f(loghyp):
ub = hm + 1.8 * hs
lb = hm - 1.8 * hs
if all(loghyp < ub) and all(loghyp > lb):
r = GPdc.GP_LKonly(X, Y, S, D, GPdc.kernel(ki, X.shape[1], [10 ** i for i in loghyp])).plk(hm, hs)
if sp.isnan(r):
class MJMError(Exception):
pass
print "nan from GPLKonly with input"
print [X, Y, S, D, ki, hm, hs, n, burn, subsam]
raise MJMError("nan from GPLKonly with input")
else:
r = -1e99
# print [loghyp, r]
return r
X = slice.slice_sample(f, hm, n, 0.05 * hs, burn=burn, subsam=subsam)
return 10 ** X
def drawhyp_plk(X, Y, S, D, ki, hm, hs, n, burn=80, subsam=5):
def f(loghyp):
ub = hm + 1.8 * hs
lb = hm - 1.8 * hs
if all(loghyp < ub) and all(loghyp > lb):
r = GPdc.GP_LKonly(
X, Y, S, D, GPdc.kernel(ki, X.shape[1],
[10**i for i in loghyp])).plk(hm, hs)
if sp.isnan(r):
class MJMError(Exception):
pass
print 'nan from GPLKonly with input'
print[X, Y, S, D, ki, hm, hs, n, burn, subsam]
raise MJMError('nan from GPLKonly with input')
else:
r = -1e99
#print [loghyp, r]
return r
X = slice.slice_sample(f, hm, n, 0.05 * hs, burn=burn, subsam=subsam)
return 10**X
def draw_support(g, lb, ub, n, method, para=1.0):
# para is the std confidence bound
if type(g) is int:
d = g
else:
d = g.D
# print 'Draw support input GP with d={} lb {} ub {}'.format(d,lb,ub)
if method == SUPPORT_UNIFORM:
print "Drawing support using uniform:"
X = sp.random.uniform(size=[n, d])
for i in xrange(d):
X[:, i] *= ub[i] - lb[i]
X[:, i] += lb[i]
elif method == SUPPORT_LAPAPR:
print "Drawing support using lapapr:"
# start with 4 times as many points as needed
# print 'a'
para = 5 * int(para)
over = 4
Xsto = sp.random.uniform(size=[over * para, d])
for i in xrange(d):
Xsto[:, i] *= ub[i] - lb[i]
Xsto[:, i] += lb[i]
# eval mean at the points
# print 'b'
fs = sp.empty(para * over)
for i in xrange(para * over):
fs[i] = g.infer_m_post(Xsto[i, :], [[sp.NaN]])[0, 0]
# print 'mean at random draws {}'.format(fs)
Xst = sp.empty([2 * para, d])
# keep the lowest
# print 'c'
for i in xrange(para):
j = fs.argmin()
Xst[i, :] = Xsto[j, :]
fs[j] = 1e99
# minimize the posterior mean from each start
# print 'd'
def f(x):
y = g.infer_m_post(sp.array(x), [[sp.NaN]])[0, 0]
bound = 0
r = max(abs(x))
if r > 1:
# print 'offedge {}'.format(x)
bound = (1e3 * (r - 1)) ** 6
return y + bound
for i in xrange(para):
res = spomin(f, Xst[i, :], method="Nelder-Mead", options={"xtol": 0.0001, "maxfev": 2000})
if not res.success:
class MJMError(Exception):
pass
print res
if not res.status == 2:
raise MJMError("failed in opt in support lapapr")
else:
print "warn, lapapr opt did not fully vconverge "
Xst[i + int(para), :] = res.x
# find endpoints that are unique
# print 'e'
Xst
unq = [Xst[0 + para, :]]
for i in xrange(para):
tmp = []
for xm in unq:
tmp.append(abs((xm - Xst[i + para, :])).max())
if min(tmp) > 0.0002:
unq.append(Xst[i + para, :])
# get the alligned gaussian approx of pmin
# print 'f'
# print unq
cls = []
for xm in unq:
ls = []
for i in xrange(d):
vg = g.infer_diag_post(xm, [[i]])[1][0, 0]
gg = g.infer_diag_post(xm, [[i, i]])[0][0, 0]
ls.append(sp.sqrt(vg) / gg)
cls.append(ls)
# print 'g'
X = mnv.rvs(size=n, mean=[0.0] * d)
if d == 1:
X.resize([n, 1])
neach = int(n / len(unq))
for i in xrange(len(unq)):
for j in xrange(d):
X[i * neach : (i + 1) * neach, j] *= min(2.0, cls[i][j])
X[i * neach : (i + 1) * neach, j] += unq[i][j]
sp.clip(X, -1, 1, out=X)
if True:
# print 'inits'
# print Xst
# print 'cls'
# print cls
np = para
# print 'para{}'.format(para)
# print Xst.shape
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]]), [[sp.NaN]])
z_[i, j] = m_[0, 0]
s_[i, j] = sp.sqrt(v_[0, 0])
fig, ax = plt.subplots(nrows=2, ncols=1, figsize=(10, 20))
CS = ax[0].contour(x_, y_, z_, 20)
ax[0].clabel(CS, inline=1, fontsize=10)
CS = ax[1].contour(x_, y_, s_, 20)
ax[0].axis([-1.0, 1.0, -1.0, 1.0])
ax[1].clabel(CS, inline=1, fontsize=10)
for i in xrange(np):
ax[0].plot([Xst[i, 0], Xst[i + np, 0]], [Xst[i, 1], Xst[i + np, 1]], "b.-")
for j in xrange(len(unq)):
x = unq[j]
ax[0].plot(x[0], x[1], "ro")
xp = [x[0] + cls[j][0], x[0], x[0] - cls[j][0], x[0], x[0] + cls[j][0]]
yp = [x[1], x[1] + cls[j][1], x[1], x[1] - cls[j][1], x[1]]
ax[0].plot(xp, yp, "r-")
fig.savefig(
os.path.join(
os.path.expanduser("~"),
"Dropbox/debugoutput",
"drawlapapr" + time.strftime("%d_%m_%y_%H:%M:%S") + ".pdf",
)
)
elif method == SUPPORT_SLICELCB:
def f(x):
if all(x > lb) and all(x < ub):
try:
return -g.infer_LCB_post(sp.array(x), [[sp.NaN]], para)[0, 0]
except:
g.infer_LCB_post(sp.array(x), [[sp.NaN]], para)[0, 0]
g.printc()
raise
else:
return -1e99
print "Drawing support using slice sample over LCB:"
X = slice.slice_sample(f, 0.5 * (ub + lb), n, 0.1 * (ub - lb))
elif method == SUPPORT_SLICEEI:
def f(x):
if all(x > lb) and all(x < ub):
try:
ei = g.infer_EI(sp.array(x), [[sp.NaN]])[0, 0]
# print ei
return sp.log(ei)
except:
# ei=g.infer_EI(sp.array(x),[[sp.NaN]])[0,0]
g.printc()
raise
else:
return -1e99
print "Drawing support using slice sample over EI:"
X = slice.slice_sample(f, 0.5 * (ub + lb), n, 0.1 * (ub - lb))
elif method == SUPPORT_SLICEPM:
def f(x):
if all(x > lb) and all(x < ub):
[m, v] = g.infer_diag_post(sp.vstack([sp.array(x)] * d), [[i] for i in xrange(d)])
p = 0.0
for i in xrange(d):
p += -0.5 * (m[0, i] ** 2) / v[0, i]
ym = g.infer_m_post(sp.array(x), [[sp.NaN]])[0, 0]
if not sp.isfinite(p):
print [m, V, p]
# raise ValueError
return -10 * ym + 0.01 * p
else:
return -1e99
if False:
A = sp.empty([100, 100])
sup = sp.linspace(-0.999, 0.999, 100)
for i in xrange(100):
for j in xrange(100):
print sp.array([sup[i], sup[j]])
A[99 - j, i] = f([sup[i], sup[j]])
print A[99 - j, i]
print A
plt.figure()
plt.imshow(A)
plt.figure()
print "Drawing support using slice sample over PM:"
X = slice.slice_sample(f, 0.5 * (ub + lb), n, 0.1 * (ub - lb))
else:
raise RuntimeError("draw_support method invalid")
return X
#!/usr/bin/env python2
#encoding: UTF-8
# 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.
from matplotlib import pyplot as plt
import slice
import scipy as sp
def llk(x):
return sp.log( sp.exp(-(0.5*(x-1)**2)/0.5**2)+sp.exp(-(0.5*(x+2)**2)/0.25**2))
S = slice.slice_sample(llk,sp.array([0.]),5000,sp.array([0.05]))
#print S
plt.hist(sp.array(S).flatten(),bins=30)
def llk2(x):
return sp.log( sp.exp(-0.5*((x[0]+1.5)**2+(x[1]-1.5)**2)/0.5**2)+sp.exp(-0.5*(((x[0]-1.5)**2)/0.75**2+((x[1]+1.5)**2)/0.25**2)))
S = slice.slice_sample(llk2,sp.array([0.,0.]),10000,sp.array([0.15,0.15]))
A = sp.zeros([20,20])
for i in xrange(S.shape[0]):
a1 = int((S[i,0]+3)*20/6)
a2 = int((S[i,1]+3)*20/6)
try:
A[a1,a2]+=1
except:
pass
#plt.figure()
#for i in xrange(A.shape[0]):
# plt.plot(S[i,0],S[i,1],'r.')
def draw_support(g, lb, ub, n, method, para=1.):
#para is the std confidence bound
if (type(g) is int):
d = g
else:
d = g.D
#print 'Draw support input GP with d={} lb {} ub {}'.format(d,lb,ub)
if method == SUPPORT_UNIFORM:
print "Drawing support using uniform:"
X = sp.random.uniform(size=[n, d])
for i in xrange(d):
X[:, i] *= ub[i] - lb[i]
X[:, i] += lb[i]
elif method == SUPPORT_LAPAPR:
print "Drawing support using lapapr:"
#start with 4 times as many points as needed
#print 'a'
para = 5 * int(para)
over = 4
Xsto = sp.random.uniform(size=[over * para, d])
for i in xrange(d):
Xsto[:, i] *= ub[i] - lb[i]
Xsto[:, i] += lb[i]
#eval mean at the points
#print 'b'
fs = sp.empty(para * over)
for i in xrange(para * over):
fs[i] = g.infer_m_post(Xsto[i, :], [[sp.NaN]])[0, 0]
#print 'mean at random draws {}'.format(fs)
Xst = sp.empty([2 * para, d])
#keep the lowest
#print 'c'
for i in xrange(para):
j = fs.argmin()
Xst[i, :] = Xsto[j, :]
fs[j] = 1e99
#minimize the posterior mean from each start
#print 'd'
def f(x):
y = g.infer_m_post(sp.array(x), [[sp.NaN]])[0, 0]
bound = 0
r = max(abs(x))
if r > 1:
#print 'offedge {}'.format(x)
bound = (1e3 * (r - 1))**6
return y + bound
for i in xrange(para):
res = spomin(f,
Xst[i, :],
method='Nelder-Mead',
options={
'xtol': 0.0001,
'maxfev': 2000
})
if not res.success:
class MJMError(Exception):
pass
print res
if not res.status == 2:
raise MJMError('failed in opt in support lapapr')
else:
print "warn, lapapr opt did not fully vconverge "
Xst[i + int(para), :] = res.x
#find endpoints that are unique
#print 'e'
Xst
unq = [Xst[0 + para, :]]
for i in xrange(para):
tmp = []
for xm in unq:
tmp.append(abs((xm - Xst[i + para, :])).max())
if min(tmp) > 0.0002:
unq.append(Xst[i + para, :])
#get the alligned gaussian approx of pmin
#print 'f'
#print unq
cls = []
for xm in unq:
ls = []
for i in xrange(d):
vg = g.infer_diag_post(xm, [[i]])[1][0, 0]
gg = g.infer_diag_post(xm, [[i, i]])[0][0, 0]
ls.append(sp.sqrt(vg) / gg)
cls.append(ls)
#print 'g'
X = mnv.rvs(size=n, mean=[0.] * d)
if d == 1:
X.resize([n, 1])
neach = int(n / len(unq))
for i in xrange(len(unq)):
for j in xrange(d):
X[i * neach:(i + 1) * neach, j] *= min(2., cls[i][j])
X[i * neach:(i + 1) * neach, j] += unq[i][j]
sp.clip(X, -1, 1, out=X)
if True:
#print 'inits'
#print Xst
#print 'cls'
#print cls
np = para
#print 'para{}'.format(para)
#print Xst.shape
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]]),
[[sp.NaN]])
z_[i, j] = m_[0, 0]
s_[i, j] = sp.sqrt(v_[0, 0])
fig, ax = plt.subplots(nrows=2, ncols=1, figsize=(10, 20))
CS = ax[0].contour(x_, y_, z_, 20)
ax[0].clabel(CS, inline=1, fontsize=10)
CS = ax[1].contour(x_, y_, s_, 20)
ax[0].axis([-1., 1., -1., 1.])
ax[1].clabel(CS, inline=1, fontsize=10)
for i in xrange(np):
ax[0].plot([Xst[i, 0], Xst[i + np, 0]],
[Xst[i, 1], Xst[i + np, 1]], 'b.-')
for j in xrange(len(unq)):
x = unq[j]
ax[0].plot(x[0], x[1], 'ro')
xp = [
x[0] + cls[j][0], x[0], x[0] - cls[j][0], x[0],
x[0] + cls[j][0]
]
yp = [x[1], x[1] + cls[j][1], x[1], x[1] - cls[j][1], x[1]]
ax[0].plot(xp, yp, 'r-')
fig.savefig(
os.path.join(
os.path.expanduser('~'), 'Dropbox/debugoutput',
'drawlapapr' + time.strftime('%d_%m_%y_%H:%M:%S') +
'.pdf'))
elif method == SUPPORT_SLICELCB:
def f(x):
if all(x > lb) and all(x < ub):
try:
return -g.infer_LCB_post(sp.array(x), [[sp.NaN]], para)[0,
0]
except:
g.infer_LCB_post(sp.array(x), [[sp.NaN]], para)[0, 0]
g.printc()
raise
else:
return -1e99
print "Drawing support using slice sample over LCB:"
X = slice.slice_sample(f, 0.5 * (ub + lb), n, 0.1 * (ub - lb))
elif method == SUPPORT_SLICEEI:
def f(x):
if all(x > lb) and all(x < ub):
try:
ei = g.infer_EI(sp.array(x), [[sp.NaN]])[0, 0]
#print ei
return sp.log(ei)
except:
#ei=g.infer_EI(sp.array(x),[[sp.NaN]])[0,0]
g.printc()
raise
else:
return -1e99
print "Drawing support using slice sample over EI:"
X = slice.slice_sample(f, 0.5 * (ub + lb), n, 0.1 * (ub - lb))
elif method == SUPPORT_SLICEPM:
def f(x):
if all(x > lb) and all(x < ub):
[m, v] = g.infer_diag_post(sp.vstack([sp.array(x)] * d),
[[i] for i in xrange(d)])
p = 0.
for i in xrange(d):
p += -0.5 * (m[0, i]**2) / v[0, i]
ym = g.infer_m_post(sp.array(x), [[sp.NaN]])[0, 0]
if not sp.isfinite(p):
print[m, V, p]
#raise ValueError
return -10 * ym + 0.01 * p
else:
return -1e99
if False:
A = sp.empty([100, 100])
sup = sp.linspace(-0.999, 0.999, 100)
for i in xrange(100):
for j in xrange(100):
print sp.array([sup[i], sup[j]])
A[99 - j, i] = f([sup[i], sup[j]])
print A[99 - j, i]
print A
plt.figure()
plt.imshow(A)
plt.figure()
print "Drawing support using slice sample over PM:"
X = slice.slice_sample(f, 0.5 * (ub + lb), n, 0.1 * (ub - lb))
else:
raise RuntimeError("draw_support method invalid")
return X
# 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.
from matplotlib import pyplot as plt
import slice
import scipy as sp
def llk(x):
return sp.log(
sp.exp(-(0.5 * (x - 1)**2) / 0.5**2) + sp.exp(-(0.5 *
(x + 2)**2) / 0.25**2))
S = slice.slice_sample(llk, sp.array([0.]), 5000, sp.array([0.05]))
#print S
plt.hist(sp.array(S).flatten(), bins=30)
def llk2(x):
return sp.log(
sp.exp(-0.5 * ((x[0] + 1.5)**2 + (x[1] - 1.5)**2) / 0.5**2) +
sp.exp(-0.5 * (((x[0] - 1.5)**2) / 0.75**2 +
((x[1] + 1.5)**2) / 0.25**2)))
S = slice.slice_sample(llk2, sp.array([0., 0.]), 10000, sp.array([0.15, 0.15]))
A = sp.zeros([20, 20])
for i in xrange(S.shape[0]):
a1 = int((S[i, 0] + 3) * 20 / 6)