def __init__(self, *argL, **argD ):
self.gp = GP(*argL, **argD)
def __init__(self, *argL, **argD):
self.gp = GP(*argL, **argD)
class MyGP:
def __init__(self, *argL, **argD ):
self.gp = GP(*argL, **argD)
def set_hypers(self, state):
print 'setting chooser state'
self.gp.__dict__.update(state)
print self
self.hypers_initialized = True
def __str__(self):
infoD = {}
for field in ['mean', 'amp2', 'noise', 'ls']:
infoD[ field ] = getattr(self.gp, field)
return '\n'.join([ "%s : %s"%(key, str(val)) for key, val in infoD.items() ])
def fit(self, x, y):
# y, self.mean, self.std = normalize(y)
m,d = x.shape
assert y.shape == (m,), "m = %d, y.shape = %s"%(m, str(y.shape))
self.x = x
self.y = y
self.gp.real_init(d,y)
if not self.hypers_initialized:
self.gp.optimize_hypers(x,y)
# The primary covariances for prediction.
K = self.gp.cov(x)
assert K.shape == (m,m)
# Compute the required Cholesky.
obsv_cov = K + self.gp.noise*np.eye(x.shape[0])
obsv_chol = spla.cholesky( obsv_cov, lower=True )
# Solve the linear systems.
self.alpha = spla.cho_solve((obsv_chol, True), y - self.gp.mean)
self.obsv_chol = obsv_chol
return self
def predict(self,x):
cand_cross = self.gp.cov(self.x, x) # O(mdn)
assert cand_cross.shape == (self.x.shape[0], x.shape[0])
y_mean = np.dot(cand_cross.T, self.alpha) + self.gp.mean # O(mn)
return y_mean
def predict_proba(self,x):
cand_cross = self.gp.cov(self.x, x) # O(mdn)
beta = spla.solve_triangular(self.obsv_chol, cand_cross, lower=True) # O(mmn)
assert cand_cross.shape == (self.x.shape[0], x.shape[0])
assert beta.shape == (self.x.shape[0],x.shape[0])
# Predict the marginal means and variances at candidates.
y_mean = np.dot(cand_cross.T, self.alpha) + self.gp.mean # O(mn)
y_var = self.gp.amp2*(1+1e-6) - np.sum(beta**2, axis=0) # O(mn)
return y_mean, y_var
def ei(self,x):
best = np.min(self.x)
y_mean, y_var = self.predict_proba(x)
y_std = np.sqrt(y_var)
u = (best - y_mean) / y_std
ncdf = sps.norm.cdf(u)
npdf = sps.norm.pdf(u)
ei = y_std*( u*ncdf + npdf)
return ei, y_mean, y_var
class MyGP:
def __init__(self, *argL, **argD):
self.gp = GP(*argL, **argD)
def set_hypers(self, state):
print 'setting chooser state'
self.gp.__dict__.update(state)
print self
self.hypers_initialized = True
def __str__(self):
infoD = {}
for field in ['mean', 'amp2', 'noise', 'ls']:
infoD[field] = getattr(self.gp, field)
return '\n'.join(
["%s : %s" % (key, str(val)) for key, val in infoD.items()])
def fit(self, x, y):
# y, self.mean, self.std = normalize(y)
m, d = x.shape
assert y.shape == (m, ), "m = %d, y.shape = %s" % (m, str(y.shape))
self.x = x
self.y = y
self.gp.real_init(d, y)
if not self.hypers_initialized:
self.gp.optimize_hypers(x, y)
# The primary covariances for prediction.
K = self.gp.cov(x)
assert K.shape == (m, m)
# Compute the required Cholesky.
obsv_cov = K + self.gp.noise * np.eye(x.shape[0])
obsv_chol = spla.cholesky(obsv_cov, lower=True)
# Solve the linear systems.
self.alpha = spla.cho_solve((obsv_chol, True), y - self.gp.mean)
self.obsv_chol = obsv_chol
return self
def predict(self, x):
cand_cross = self.gp.cov(self.x, x) # O(mdn)
assert cand_cross.shape == (self.x.shape[0], x.shape[0])
y_mean = np.dot(cand_cross.T, self.alpha) + self.gp.mean # O(mn)
return y_mean
def predict_proba(self, x):
cand_cross = self.gp.cov(self.x, x) # O(mdn)
beta = spla.solve_triangular(self.obsv_chol, cand_cross,
lower=True) # O(mmn)
assert cand_cross.shape == (self.x.shape[0], x.shape[0])
assert beta.shape == (self.x.shape[0], x.shape[0])
# Predict the marginal means and variances at candidates.
y_mean = np.dot(cand_cross.T, self.alpha) + self.gp.mean # O(mn)
y_var = self.gp.amp2 * (1 + 1e-6) - np.sum(beta**2, axis=0) # O(mn)
return y_mean, y_var
def ei(self, x):
best = np.min(self.x)
y_mean, y_var = self.predict_proba(x)
y_std = np.sqrt(y_var)
u = (best - y_mean) / y_std
ncdf = sps.norm.cdf(u)
npdf = sps.norm.pdf(u)
ei = y_std * (u * ncdf + npdf)
return ei, y_mean, y_var
