mirror of
https://github.com/SheffieldML/GPy.git
synced 2026-06-29 15:59:41 +02:00
Merge branch 'devel' of github.com:SheffieldML/GPy into devel
This commit is contained in:
James Hensman
commit
6f06646437
10 changed files with 528 additions and 49 deletions
|
|
@ -161,10 +161,13 @@ def bgplvm_oil(optimize=True, verbose=1, plot=True, N=200, Q=7, num_inducing=40,
|
|||
import GPy
|
||||
from matplotlib import pyplot as plt
|
||||
import numpy as np
|
||||
import pods
|
||||
|
||||
_np.random.seed(0)
|
||||
data = pods.datasets.oil()
|
||||
try:
|
||||
import pods
|
||||
data = pods.datasets.oil()
|
||||
except ImportError:
|
||||
data = GPy.util.datasets.oil()
|
||||
|
||||
|
||||
kernel = GPy.kern.RBF(Q, 1., 1./_np.random.uniform(0,1,(Q,)), ARD=True)# + GPy.kern.Bias(Q, _np.exp(-2))
|
||||
Y = data['X'][:N]
|
||||
|
|
|
|||
|
|
@ -34,7 +34,9 @@ class VarDTC(LatentFunctionInference):
|
|||
self.get_YYTfactor.limit = limit
|
||||
|
||||
def _get_trYYT(self, Y):
|
||||
return np.sum(np.square(Y))
|
||||
return np.einsum("ij,ij->", Y, Y)
|
||||
# faster than, but same as:
|
||||
# return np.sum(np.square(Y))
|
||||
|
||||
def __getstate__(self):
|
||||
# has to be overridden, as Cacher objects cannot be pickled.
|
||||
|
|
@ -103,7 +105,7 @@ class VarDTC(LatentFunctionInference):
|
|||
psi0 = kern.Kdiag(X)
|
||||
psi1 = kern.K(X, Z)
|
||||
if het_noise:
|
||||
tmp = psi1 * (np.sqrt(beta.reshape(num_data, 1)))
|
||||
tmp = psi1 * (np.sqrt(beta))
|
||||
else:
|
||||
tmp = psi1 * (np.sqrt(beta))
|
||||
tmp, _ = dtrtrs(Lm, tmp.T, lower=1)
|
||||
|
|
@ -137,14 +139,14 @@ class VarDTC(LatentFunctionInference):
|
|||
|
||||
# log marginal likelihood
|
||||
log_marginal = _compute_log_marginal_likelihood(likelihood, num_data, output_dim, beta, het_noise,
|
||||
psi0, A, LB, trYYT, data_fit, VVT_factor)
|
||||
psi0, A, LB, trYYT, data_fit, Y)
|
||||
|
||||
#noise derivatives
|
||||
dL_dR = _compute_dL_dR(likelihood,
|
||||
het_noise, uncertain_inputs, LB,
|
||||
_LBi_Lmi_psi1Vf, DBi_plus_BiPBi, Lm, A,
|
||||
psi0, psi1, beta,
|
||||
data_fit, num_data, output_dim, trYYT, Y)
|
||||
data_fit, num_data, output_dim, trYYT, Y, VVT_factor)
|
||||
|
||||
dL_dthetaL = likelihood.exact_inference_gradients(dL_dR,Y_metadata)
|
||||
|
||||
|
|
@ -184,14 +186,14 @@ class VarDTC(LatentFunctionInference):
|
|||
return post, log_marginal, grad_dict
|
||||
|
||||
def _compute_dL_dpsi(num_inducing, num_data, output_dim, beta, Lm, VVT_factor, Cpsi1Vf, DBi_plus_BiPBi, psi1, het_noise, uncertain_inputs):
|
||||
dL_dpsi0 = -0.5 * output_dim * (beta[:,None] * np.ones([num_data, 1])).flatten()
|
||||
dL_dpsi0 = -0.5 * output_dim * (beta* np.ones([num_data, 1])).flatten()
|
||||
dL_dpsi1 = np.dot(VVT_factor, Cpsi1Vf.T)
|
||||
dL_dpsi2_beta = 0.5 * backsub_both_sides(Lm, output_dim * np.eye(num_inducing) - DBi_plus_BiPBi)
|
||||
if het_noise:
|
||||
if uncertain_inputs:
|
||||
dL_dpsi2 = beta[:, None, None] * dL_dpsi2_beta[None, :, :]
|
||||
dL_dpsi2 = beta[:, None] * dL_dpsi2_beta[None, :, :]
|
||||
else:
|
||||
dL_dpsi1 += 2.*np.dot(dL_dpsi2_beta, (psi1 * beta.reshape(num_data, 1)).T).T
|
||||
dL_dpsi1 += 2.*np.dot(dL_dpsi2_beta, (psi1 * beta).T).T
|
||||
dL_dpsi2 = None
|
||||
else:
|
||||
dL_dpsi2 = beta * dL_dpsi2_beta
|
||||
|
|
@ -202,7 +204,7 @@ def _compute_dL_dpsi(num_inducing, num_data, output_dim, beta, Lm, VVT_factor, C
|
|||
return dL_dpsi0, dL_dpsi1, dL_dpsi2
|
||||
|
||||
|
||||
def _compute_dL_dR(likelihood, het_noise, uncertain_inputs, LB, _LBi_Lmi_psi1Vf, DBi_plus_BiPBi, Lm, A, psi0, psi1, beta, data_fit, num_data, output_dim, trYYT, Y):
|
||||
def _compute_dL_dR(likelihood, het_noise, uncertain_inputs, LB, _LBi_Lmi_psi1Vf, DBi_plus_BiPBi, Lm, A, psi0, psi1, beta, data_fit, num_data, output_dim, trYYT, Y, VVT_factr=None):
|
||||
# the partial derivative vector for the likelihood
|
||||
if likelihood.size == 0:
|
||||
# save computation here.
|
||||
|
|
@ -217,8 +219,7 @@ def _compute_dL_dR(likelihood, het_noise, uncertain_inputs, LB, _LBi_Lmi_psi1Vf,
|
|||
|
||||
Lmi_psi1, nil = dtrtrs(Lm, psi1.T, lower=1, trans=0)
|
||||
_LBi_Lmi_psi1, _ = dtrtrs(LB, Lmi_psi1, lower=1, trans=0)
|
||||
|
||||
dL_dR = -0.5 * beta + 0.5 * (beta*Y)**2
|
||||
dL_dR = -0.5 * beta + 0.5 * VVT_factr**2
|
||||
dL_dR += 0.5 * output_dim * (psi0 - np.sum(Lmi_psi1**2,0))[:,None] * beta**2
|
||||
|
||||
dL_dR += 0.5*np.sum(mdot(LBi.T,LBi,Lmi_psi1)*Lmi_psi1,0)[:,None]*beta**2
|
||||
|
|
@ -232,10 +233,10 @@ def _compute_dL_dR(likelihood, het_noise, uncertain_inputs, LB, _LBi_Lmi_psi1Vf,
|
|||
dL_dR += beta * (0.5 * np.sum(A * DBi_plus_BiPBi) - data_fit)
|
||||
return dL_dR
|
||||
|
||||
def _compute_log_marginal_likelihood(likelihood, num_data, output_dim, beta, het_noise, psi0, A, LB, trYYT, data_fit,Y):
|
||||
def _compute_log_marginal_likelihood(likelihood, num_data, output_dim, beta, het_noise, psi0, A, LB, trYYT, data_fit, Y):
|
||||
#compute log marginal likelihood
|
||||
if het_noise:
|
||||
lik_1 = -0.5 * num_data * output_dim * np.log(2. * np.pi) + 0.5 * np.sum(np.log(beta)) - 0.5 * np.sum(beta * np.square(Y).sum(axis=-1))
|
||||
lik_1 = -0.5 * num_data * output_dim * np.log(2. * np.pi) + 0.5 * output_dim * np.sum(np.log(beta)) - 0.5 * np.sum(beta.ravel() * np.square(Y).sum(axis=-1))
|
||||
lik_2 = -0.5 * output_dim * (np.sum(beta.flatten() * psi0) - np.trace(A))
|
||||
else:
|
||||
lik_1 = -0.5 * num_data * output_dim * (np.log(2. * np.pi) - np.log(beta)) - 0.5 * beta * trYYT
|
||||
|
|
|
|||
|
|
@ -23,7 +23,7 @@ class MixedNoise(Likelihood):
|
|||
variance = np.zeros(ind.size)
|
||||
for lik, j in zip(self.likelihoods_list, range(len(self.likelihoods_list))):
|
||||
variance[ind==j] = lik.variance
|
||||
return variance
|
||||
return variance[:,None]
|
||||
|
||||
def betaY(self,Y,Y_metadata):
|
||||
#TODO not here.
|
||||
|
|
|
|||
|
|
@ -75,7 +75,7 @@ def dparam_checkgrad(func, dfunc, params, params_names, args, constraints=None,
|
|||
if verbose:
|
||||
print grad
|
||||
grad.checkgrad(verbose=1)
|
||||
if not grad.checkgrad():
|
||||
if not grad.checkgrad(verbose=True):
|
||||
gradchecking = False
|
||||
|
||||
return gradchecking
|
||||
|
|
@ -364,9 +364,8 @@ class TestNoiseModels(object):
|
|||
dlogpdf_df = functools.partial(model.dlogpdf_df, y=Y)
|
||||
grad = GradientChecker(logpdf, dlogpdf_df, f.copy(), 'g')
|
||||
grad.randomize()
|
||||
grad.checkgrad(verbose=1)
|
||||
print model
|
||||
assert grad.checkgrad()
|
||||
assert grad.checkgrad(verbose=1)
|
||||
|
||||
@with_setup(setUp, tearDown)
|
||||
def t_d2logpdf_df2(self, model, Y, f):
|
||||
|
|
@ -375,9 +374,8 @@ class TestNoiseModels(object):
|
|||
d2logpdf_df2 = functools.partial(model.d2logpdf_df2, y=Y)
|
||||
grad = GradientChecker(dlogpdf_df, d2logpdf_df2, f.copy(), 'g')
|
||||
grad.randomize()
|
||||
grad.checkgrad(verbose=1)
|
||||
print model
|
||||
assert grad.checkgrad()
|
||||
assert grad.checkgrad(verbose=1)
|
||||
|
||||
@with_setup(setUp, tearDown)
|
||||
def t_d3logpdf_df3(self, model, Y, f):
|
||||
|
|
@ -386,9 +384,8 @@ class TestNoiseModels(object):
|
|||
d3logpdf_df3 = functools.partial(model.d3logpdf_df3, y=Y)
|
||||
grad = GradientChecker(d2logpdf_df2, d3logpdf_df3, f.copy(), 'g')
|
||||
grad.randomize()
|
||||
grad.checkgrad(verbose=1)
|
||||
print model
|
||||
assert grad.checkgrad()
|
||||
assert grad.checkgrad(verbose=1)
|
||||
|
||||
##############
|
||||
# df_dparams #
|
||||
|
|
@ -439,8 +436,8 @@ class TestNoiseModels(object):
|
|||
|
||||
grad.randomize()
|
||||
print grad
|
||||
grad.checkgrad(verbose=1)
|
||||
assert grad.checkgrad()
|
||||
print model
|
||||
assert grad.checkgrad(verbose=1)
|
||||
|
||||
@with_setup(setUp, tearDown)
|
||||
def t_d2logpdf_dlink2(self, model, Y, f, link_f_constraints):
|
||||
|
|
@ -454,9 +451,9 @@ class TestNoiseModels(object):
|
|||
constraint('g', grad)
|
||||
|
||||
grad.randomize()
|
||||
grad.checkgrad(verbose=1)
|
||||
print grad
|
||||
assert grad.checkgrad()
|
||||
print model
|
||||
assert grad.checkgrad(verbose=1)
|
||||
|
||||
@with_setup(setUp, tearDown)
|
||||
def t_d3logpdf_dlink3(self, model, Y, f, link_f_constraints):
|
||||
|
|
@ -470,9 +467,9 @@ class TestNoiseModels(object):
|
|||
constraint('g', grad)
|
||||
|
||||
grad.randomize()
|
||||
grad.checkgrad(verbose=1)
|
||||
print grad
|
||||
assert grad.checkgrad()
|
||||
print model
|
||||
assert grad.checkgrad(verbose=1)
|
||||
|
||||
#################
|
||||
# dlink_dparams #
|
||||
|
|
@ -535,13 +532,12 @@ class TestNoiseModels(object):
|
|||
|
||||
#m.optimize(max_iters=8)
|
||||
print m
|
||||
m.checkgrad(verbose=1, step=step)
|
||||
#if not m.checkgrad(step=step):
|
||||
#m.checkgrad(verbose=1, step=step)
|
||||
#NOTE this test appears to be stochastic for some likelihoods (student t?)
|
||||
# appears to all be working in test mode right now...
|
||||
#if isinstance(model, GPy.likelihoods.StudentT):
|
||||
assert m.checkgrad(step=step)
|
||||
assert m.checkgrad(verbose=1, step=step)
|
||||
|
||||
###########
|
||||
# EP test #
|
||||
|
|
@ -563,9 +559,8 @@ class TestNoiseModels(object):
|
|||
constraints[param_num](name, m)
|
||||
|
||||
m.randomize()
|
||||
m.checkgrad(verbose=1, step=step)
|
||||
print m
|
||||
assert m.checkgrad(step=step)
|
||||
assert m.checkgrad(verbose=1, step=step)
|
||||
|
||||
|
||||
class LaplaceTests(unittest.TestCase):
|
||||
|
|
@ -616,8 +611,8 @@ class LaplaceTests(unittest.TestCase):
|
|||
d2logpdf_df2 = functools.partial(self.gauss.d2logpdf_df2, y=self.Y)
|
||||
grad = GradientChecker(dlogpdf_df, d2logpdf_df2, self.f.copy(), 'g')
|
||||
grad.randomize()
|
||||
grad.checkgrad(verbose=1)
|
||||
self.assertTrue(grad.checkgrad())
|
||||
|
||||
self.assertTrue(grad.checkgrad(verbose=1))
|
||||
|
||||
def test_laplace_log_likelihood(self):
|
||||
debug = False
|
||||
|
|
@ -705,8 +700,8 @@ class LaplaceTests(unittest.TestCase):
|
|||
#Check they are checkgradding
|
||||
#m1.checkgrad(verbose=1)
|
||||
#m2.checkgrad(verbose=1)
|
||||
self.assertTrue(m1.checkgrad())
|
||||
self.assertTrue(m2.checkgrad())
|
||||
self.assertTrue(m1.checkgrad(verbose=True))
|
||||
self.assertTrue(m2.checkgrad(verbose=True))
|
||||
|
||||
if __name__ == "__main__":
|
||||
print "Running unit tests"
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue