apunkt/GPy
1
0
Fork 0
mirror of https://github.com/SheffieldML/GPy.git synced 2026-06-08 15:05:15 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merged conflict of model.py

Browse source
This commit is contained in:
Neil Lawrence 2013-09-14 20:05:42 +01:00
commit dffa9541c2
42 changed files with 2297 additions and 453 deletions
Download patch file Download diff file Expand all files Collapse all files

1
GPy/core/fitc.py
View file

@ -140,7 +140,6 @@ class FITC(SparseGP):
dA_dnoise = 0.5 * self.input_dim * (dbstar_dnoise/self.beta_star).sum() - 0.5 * self.input_dim * np.sum(self.likelihood.Y**2 * dbstar_dnoise)
dC_dnoise = -0.5 * np.sum(mdot(self.LBi.T,self.LBi,Lmi_psi1) * Lmi_psi1 * dbstar_dnoise.T)
dC_dnoise = -0.5 * np.sum(mdot(self.LBi.T,self.LBi,Lmi_psi1) * Lmi_psi1 * dbstar_dnoise.T)
dD_dnoise_1 = mdot(self.V_star*LBiLmipsi1.T,LBiLmipsi1*dbstar_dnoise.T*self.likelihood.Y.T)
alpha = mdot(LBiLmipsi1,self.V_star)

80
GPy/core/gp.py
View file

@ -19,9 +19,6 @@ class GP(GPBase):
:param normalize_X: whether to normalize the input data before computing (predictions will be in original scales)
:type normalize_X: False|True
:rtype: model object
:param epsilon_ep: convergence criterion for the Expectation Propagation algorithm, defaults to 0.1
:param powerep: power-EP parameters [$\eta$,$\delta$], defaults to [1.,1.]
:type powerep: list
.. Note:: Multiple independent outputs are allowed using columns of Y
@ -105,7 +102,12 @@ class GP(GPBase):
Note, we use the chain rule: dL_dtheta = dL_dK * d_K_dtheta
"""
return np.hstack((self.kern.dK_dtheta(dL_dK=self.dL_dK, X=self.X), self.likelihood._gradients(partial=np.diag(self.dL_dK))))
#return np.hstack((self.kern.dK_dtheta(dL_dK=self.dL_dK, X=self.X), self.likelihood._gradients(partial=np.diag(self.dL_dK))))
if not isinstance(self.likelihood,EP):
tmp = np.hstack((self.kern.dK_dtheta(dL_dK=self.dL_dK, X=self.X), self.likelihood._gradients(partial=np.diag(self.dL_dK))))
else:
tmp = np.hstack((self.kern.dK_dtheta(dL_dK=self.dL_dK, X=self.X), self.likelihood._gradients(partial=np.diag(self.dL_dK))))
return tmp
def _raw_predict(self, _Xnew, which_parts='all', full_cov=False, stop=False):
"""
@ -136,7 +138,7 @@ class GP(GPBase):
:type Xnew: np.ndarray, Nnew x self.input_dim
:param which_parts: specifies which outputs kernel(s) to use in prediction
:type which_parts: ('all', list of bools)
:param full_cov: whether to return the folll covariance matrix, or just the diagonal
:param full_cov: whether to return the full covariance matrix, or just the diagonal
:type full_cov: bool
:rtype: posterior mean, a Numpy array, Nnew x self.input_dim
:rtype: posterior variance, a Numpy array, Nnew x 1 if full_cov=False, Nnew x Nnew otherwise
@ -153,5 +155,71 @@ class GP(GPBase):
# now push through likelihood
mean, var, _025pm, _975pm = self.likelihood.predictive_values(mu, var, full_cov, **likelihood_args)
return mean, var, _025pm, _975pm
def predict_single_output(self, Xnew, output=0, which_parts='all', full_cov=False):
"""
For a specific output, predict the function at the new point(s) Xnew.
Arguments
---------
:param Xnew: The points at which to make a prediction
:type Xnew: np.ndarray, Nnew x self.input_dim
:param output: output to predict
:type output: integer in {0,..., num_outputs-1}
:param which_parts: specifies which outputs kernel(s) to use in prediction
:type which_parts: ('all', list of bools)
:param full_cov: whether to return the full covariance matrix, or just the diagonal
:type full_cov: bool
:rtype: posterior mean, a Numpy array, Nnew x self.input_dim
:rtype: posterior variance, a Numpy array, Nnew x 1 if full_cov=False, Nnew x Nnew otherwise
:rtype: lower and upper boundaries of the 95% confidence intervals, Numpy arrays, Nnew x self.input_dim
.. Note:: For multiple output models only
"""
assert hasattr(self,'multioutput')
index = np.ones_like(Xnew)*output
Xnew = np.hstack((Xnew,index))
# normalize X values
Xnew = (Xnew.copy() - self._Xoffset) / self._Xscale
mu, var = self._raw_predict(Xnew, full_cov=full_cov, which_parts=which_parts)
# now push through likelihood
mean, var, _025pm, _975pm = self.likelihood.predictive_values(mu, var, full_cov, noise_model = output)
return mean, var, _025pm, _975pm
def _raw_predict_single_output(self, _Xnew, output=0, which_parts='all', full_cov=False,stop=False):
"""
Internal helper function for making predictions for a specific output,
does not account for normalization or likelihood
---------
:param Xnew: The points at which to make a prediction
:type Xnew: np.ndarray, Nnew x self.input_dim
:param output: output to predict
:type output: integer in {0,..., num_outputs-1}
:param which_parts: specifies which outputs kernel(s) to use in prediction
:type which_parts: ('all', list of bools)
:param full_cov: whether to return the full covariance matrix, or just the diagonal
.. Note:: For multiple output models only
"""
assert hasattr(self,'multioutput')
# creates an index column and appends it to _Xnew
index = np.ones_like(_Xnew)*output
_Xnew = np.hstack((_Xnew,index))
Kx = self.kern.K(_Xnew,self.X,which_parts=which_parts).T
KiKx, _ = dpotrs(self.L, np.asfortranarray(Kx), lower=1)
mu = np.dot(KiKx.T, self.likelihood.Y)
if full_cov:
Kxx = self.kern.K(_Xnew, which_parts=which_parts)
var = Kxx - np.dot(KiKx.T, Kx)
else:
Kxx = self.kern.Kdiag(_Xnew, which_parts=which_parts)
var = Kxx - np.sum(np.multiply(KiKx, Kx), 0)
var = var[:, None]
if stop:
debug_this # @UndefinedVariable
return mu, var

194
GPy/core/gp_base.py
View file

@ -57,34 +57,30 @@ class GPBase(Model):
self.X = state.pop()
Model.setstate(self, state)
def plot_f(self, samples=0, plot_limits=None, which_data='all', which_parts='all', resolution=None, full_cov=False, fignum=None, ax=None):
def plot_f(self, samples=0, plot_limits=None, which_data='all', which_parts='all', resolution=None, full_cov=False, fignum=None, ax=None,output=None):
"""
Plot the GP's view of the world, where the data is normalized and the
likelihood is Gaussian.
Plot the GP's view of the world, where the data is normalized and the
- In one dimension, the function is plotted with a shaded region identifying two standard deviations.
- In two dimsensions, a contour-plot shows the mean predicted function
- Not implemented in higher dimensions
Plot the posterior of the GP.
- In one dimension, the function is plotted with a shaded region identifying two standard deviations.
- In two dimsensions, a contour-plot shows the mean predicted function
- In higher dimensions, we've no implemented this yet !TODO!
Can plot only part of the data and part of the posterior functions
using which_data and which_functions
:param samples: the number of a posteriori samples to plot
:param plot_limits: The limits of the plot. If 1D [xmin,xmax], if 2D [[xmin,ymin],[xmax,ymax]]. Defaluts to data limits
:param which_data: which if the training data to plot (default all)
:type which_data: 'all' or a slice object to slice self.X, self.Y
:param which_parts: which of the kernel functions to plot (additively)
:type which_parts: 'all', or list of bools
:param resolution: the number of intervals to sample the GP on. Defaults to 200 in 1D and 50 (a 50x50 grid) in 2D
:type resolution: int
:param full_cov:
:type full_cov: bool
:param fignum: figure to plot on.
:type fignum: figure number
:param ax: axes to plot on.
:type ax: axes handle
:param samples: the number of a posteriori samples to plot
:param plot_limits: The limits of the plot. If 1D [xmin,xmax], if 2D [[xmin,ymin],[xmax,ymax]]. Defaluts to data limits
:param which_data: which if the training data to plot (default all)
:type which_data: 'all' or a slice object to slice self.X, self.Y
:param which_parts: which of the kernel functions to plot (additively)
:type which_parts: 'all', or list of bools
:param resolution: the number of intervals to sample the GP on. Defaults to 200 in 1D and 50 (a 50x50 grid) in 2D
:type resolution: int
:param full_cov:
:type full_cov: bool
:param fignum: figure to plot on.
:type fignum: figure number
:param ax: axes to plot on.
:type ax: axes handle
:param output: which output to plot (for multiple output models only)
:type output: integer (first output is 0)
"""
if which_data == 'all':
which_data = slice(None)
@ -93,7 +89,7 @@ class GPBase(Model):
fig = pb.figure(num=fignum)
ax = fig.add_subplot(111)
if self.X.shape[1] == 1:
if self.X.shape[1] == 1 and not hasattr(self,'multioutput'):
Xnew, xmin, xmax = x_frame1D(self.X, plot_limits=plot_limits)
if samples == 0:
m, v = self._raw_predict(Xnew, which_parts=which_parts)
@ -111,7 +107,7 @@ class GPBase(Model):
ymin, ymax = ymin - 0.1 * (ymax - ymin), ymax + 0.1 * (ymax - ymin)
ax.set_ylim(ymin, ymax)
elif self.X.shape[1] == 2:
elif self.X.shape[1] == 2 and not hasattr(self,'multioutput'):
resolution = resolution or 50
Xnew, xmin, xmax, xx, yy = x_frame2D(self.X, plot_limits, resolution)
m, v = self._raw_predict(Xnew, which_parts=which_parts)
@ -120,17 +116,51 @@ class GPBase(Model):
ax.scatter(self.X[:, 0], self.X[:, 1], 40, self.likelihood.Y, linewidth=0, cmap=pb.cm.jet, vmin=m.min(), vmax=m.max()) # @UndefinedVariable
ax.set_xlim(xmin[0], xmax[0])
ax.set_ylim(xmin[1], xmax[1])
elif self.X.shape[1] == 2 and hasattr(self,'multioutput'):
output -= 1
assert self.num_outputs >= output, 'The model has only %s outputs.' %self.num_outputs
Xu = self.X[self.X[:,-1]==output ,0:1]
Xnew, xmin, xmax = x_frame1D(Xu, plot_limits=plot_limits)
if samples == 0:
m, v = self._raw_predict_single_output(Xnew, output=output, which_parts=which_parts)
gpplot(Xnew, m, m - 2 * np.sqrt(v), m + 2 * np.sqrt(v), axes=ax)
ax.plot(Xu[which_data], self.likelihood.Y[self.likelihood.index==output][:,None], 'kx', mew=1.5)
else:
m, v = self._raw_predict_single_output(Xnew, output=output, which_parts=which_parts, full_cov=True)
Ysim = np.random.multivariate_normal(m.flatten(), v, samples)
gpplot(Xnew, m, m - 2 * np.sqrt(np.diag(v)[:, None]), m + 2 * np.sqrt(np.diag(v))[:, None, ], axes=ax)
for i in range(samples):
ax.plot(Xnew, Ysim[i, :], Tango.colorsHex['darkBlue'], linewidth=0.25)
ax.set_xlim(xmin, xmax)
ymin, ymax = min(np.append(self.likelihood.Y, m - 2 * np.sqrt(np.diag(v)[:, None]))), max(np.append(self.likelihood.Y, m + 2 * np.sqrt(np.diag(v)[:, None])))
ymin, ymax = ymin - 0.1 * (ymax - ymin), ymax + 0.1 * (ymax - ymin)
ax.set_ylim(ymin, ymax)
if hasattr(self,'Z'):
Zu = self.Z[self.Z[:,-1]==output,:]
Zu = self.Z * self._Xscale + self._Xoffset
Zu = self.Z[self.Z[:,-1]==output ,0:1] #??
ax.plot(Zu, np.zeros_like(Zu) + ax.get_ylim()[0], 'r|', mew=1.5, markersize=12)
elif self.X.shape[1] == 3 and hasattr(self,'multioutput'):
raise NotImplementedError, "Plots not implemented for multioutput models with 2D inputs...yet"
output -= 1
assert self.num_outputs >= output, 'The model has only %s outputs.' %self.num_outputs
else:
raise NotImplementedError, "Cannot define a frame with more than two input dimensions"
def plot(self, plot_limits=None, which_data='all', which_parts='all', resolution=None, levels=20, samples=0, fignum=None, ax=None, fixed_inputs=[], linecol=Tango.colorsHex['darkBlue'],fillcol=Tango.colorsHex['lightBlue']):
def plot(self, plot_limits=None, which_data='all', which_parts='all', resolution=None, levels=20, samples=0, fignum=None, ax=None, output=None, fixed_inputs=[], linecol=Tango.colorsHex['darkBlue'],fillcol=Tango.colorsHex['lightBlue']):
"""
Plot the GP with noise where the likelihood is Gaussian.
Plot the posterior of the GP.
- In one dimension, the function is plotted with a shaded region identifying two standard deviations.
- In two dimsensions, a contour-plot shows the mean predicted function
- In higher dimensions, we've no implemented this yet !TODO!
- Not implemented in higher dimensions
Can plot only part of the data and part of the posterior functions
using which_data and which_functions
@ -151,15 +181,13 @@ class GPBase(Model):
:type fignum: figure number
:param ax: axes to plot on.
:type ax: axes handle
:type output: integer (first output is 0)
:param fixed_inputs: a list of tuple [(i,v), (i,v)...], specifying that input index i should be set to value v.
:type fixed_inputs: a list of tuples
:param linecol: color of line to plot.
:type linecol:
:param fillcol: color of fill
:type fillcol:
:param levels: for 2D plotting, the number of contour levels to use
is ax is None, create a new figure
:param levels: for 2D plotting, the number of contour levels to use is ax is None, create a new figure
"""
# TODO include samples
if which_data == 'all':
@ -169,41 +197,81 @@ class GPBase(Model):
fig = pb.figure(num=fignum)
ax = fig.add_subplot(111)
plotdims = self.input_dim - len(fixed_inputs)
if not hasattr(self,'multioutput'):
if plotdims == 1:
plotdims = self.input_dim - len(fixed_inputs)
if plotdims == 1:
resolution = resolution or 200
Xu = self.X * self._Xscale + self._Xoffset # NOTE self.X are the normalized values now
Xu = self.X * self._Xscale + self._Xoffset # NOTE self.X are the normalized values now
fixed_dims = np.array([i for i,v in fixed_inputs])
freedim = np.setdiff1d(np.arange(self.input_dim),fixed_dims)
fixed_dims = np.array([i for i,v in fixed_inputs])
freedim = np.setdiff1d(np.arange(self.input_dim),fixed_dims)
Xnew, xmin, xmax = x_frame1D(Xu[:,freedim], plot_limits=plot_limits)
Xgrid = np.empty((Xnew.shape[0],self.input_dim))
Xgrid[:,freedim] = Xnew
for i,v in fixed_inputs:
Xgrid[:,i] = v
Xnew, xmin, xmax = x_frame1D(Xu[:,freedim], plot_limits=plot_limits)
Xgrid = np.empty((Xnew.shape[0],self.input_dim))
Xgrid[:,freedim] = Xnew
for i,v in fixed_inputs:
Xgrid[:,i] = v
m, _, lower, upper = self.predict(Xgrid, which_parts=which_parts)
for d in range(m.shape[1]):
gpplot(Xnew, m[:, d], lower[:, d], upper[:, d], axes=ax, edgecol=linecol, fillcol=fillcol)
ax.plot(Xu[which_data,freedim], self.likelihood.data[which_data, d], 'kx', mew=1.5)
ymin, ymax = min(np.append(self.likelihood.data, lower)), max(np.append(self.likelihood.data, upper))
ymin, ymax = ymin - 0.1 * (ymax - ymin), ymax + 0.1 * (ymax - ymin)
ax.set_xlim(xmin, xmax)
ax.set_ylim(ymin, ymax)
m, _, lower, upper = self.predict(Xgrid, which_parts=which_parts)
for d in range(m.shape[1]):
gpplot(Xnew, m[:, d], lower[:, d], upper[:, d], axes=ax, edgecol=linecol, fillcol=fillcol)
ax.plot(Xu[which_data,freedim], self.likelihood.data[which_data, d], 'kx', mew=1.5)
ymin, ymax = min(np.append(self.likelihood.data, lower)), max(np.append(self.likelihood.data, upper))
ymin, ymax = ymin - 0.1 * (ymax - ymin), ymax + 0.1 * (ymax - ymin)
ax.set_xlim(xmin, xmax)
ax.set_ylim(ymin, ymax)
elif self.X.shape[1] == 2: # FIXME
resolution = resolution or 50
Xnew, _, _, xmin, xmax = x_frame2D(self.X, plot_limits, resolution)
x, y = np.linspace(xmin[0], xmax[0], resolution), np.linspace(xmin[1], xmax[1], resolution)
m, _, lower, upper = self.predict(Xnew, which_parts=which_parts)
m = m.reshape(resolution, resolution).T
ax.contour(x, y, m, levels, vmin=m.min(), vmax=m.max(), cmap=pb.cm.jet) # @UndefinedVariable
Yf = self.likelihood.data.flatten()
ax.scatter(self.X[:, 0], self.X[:, 1], 40, Yf, cmap=pb.cm.jet, vmin=m.min(), vmax=m.max(), linewidth=0.) # @UndefinedVariable
ax.set_xlim(xmin[0], xmax[0])
ax.set_ylim(xmin[1], xmax[1])
Xnew, xmin, xmax = x_frame1D(Xu, plot_limits=plot_limits,resolution=resolution)
m, _, lower, upper = self.predict(Xnew, which_parts=which_parts)
for d in range(m.shape[1]):
gpplot(Xnew, m[:, d], lower[:, d], upper[:, d], axes=ax)
ax.plot(Xu[which_data], self.likelihood.data[which_data, d], 'kx', mew=1.5)
ymin, ymax = min(np.append(self.likelihood.data, lower)), max(np.append(self.likelihood.data, upper))
ymin, ymax = ymin - 0.1 * (ymax - ymin), ymax + 0.1 * (ymax - ymin)
ax.set_xlim(xmin, xmax)
ax.set_ylim(ymin, ymax)
elif self.X.shape[1] == 2:
resolution = resolution or 50
Xnew, _, _, xmin, xmax = x_frame2D(self.X, plot_limits, resolution)
x, y = np.linspace(xmin[0], xmax[0], resolution), np.linspace(xmin[1], xmax[1], resolution)
m, _, lower, upper = self.predict(Xnew, which_parts=which_parts)
m = m.reshape(resolution, resolution).T
ax.contour(x, y, m, levels, vmin=m.min(), vmax=m.max(), cmap=pb.cm.jet) # @UndefinedVariable
Yf = self.likelihood.Y.flatten()
ax.scatter(self.X[:, 0], self.X[:, 1], 40, Yf, cmap=pb.cm.jet, vmin=m.min(), vmax=m.max(), linewidth=0.) # @UndefinedVariable
ax.set_xlim(xmin[0], xmax[0])
ax.set_ylim(xmin[1], xmax[1])
else:
raise NotImplementedError, "Cannot define a frame with more than two input dimensions"
else:
raise NotImplementedError, "Cannot define a frame with more than two input dimensions"
assert self.num_outputs > output, 'The model has only %s outputs.' %self.num_outputs
if self.X.shape[1] == 2:
resolution = resolution or 200
Xu = self.X[self.X[:,-1]==output,:] #keep the output of interest
Xu = self.X * self._Xscale + self._Xoffset
Xu = self.X[self.X[:,-1]==output ,0:1] #get rid of the index column
Xnew, xmin, xmax = x_frame1D(Xu, plot_limits=plot_limits)
m, _, lower, upper = self.predict_single_output(Xnew, which_parts=which_parts,output=output)
for d in range(m.shape[1]):
gpplot(Xnew, m[:, d], lower[:, d], upper[:, d], axes=ax)
ax.plot(Xu[which_data], self.likelihood.noise_model_list[output].data, 'kx', mew=1.5)
ymin, ymax = min(np.append(self.likelihood.data, lower)), max(np.append(self.likelihood.data, upper))
ymin, ymax = ymin - 0.1 * (ymax - ymin), ymax + 0.1 * (ymax - ymin)
ax.set_xlim(xmin, xmax)
ax.set_ylim(ymin, ymax)
elif self.X.shape[1] == 3:
raise NotImplementedError, "Plots not yet implemented for multioutput models with 2D inputs"
resolution = resolution or 50
else:
raise NotImplementedError, "Cannot define a frame with more than two input dimensions"

4
GPy/core/model.py
View file

@ -458,7 +458,7 @@ class Model(Parameterized):
numerical_gradient = (f1 - f2) / (2 * dx)
global_ratio = (f1 - f2) / (2 * np.dot(dx, np.where(gradient==0, 1e-32, gradient)))
return (np.abs(1. - global_ratio) < tolerance) or (np.abs(gradient - numerical_gradient).mean() - 1) < tolerance
return (np.abs(1. - global_ratio) < tolerance) or (np.abs(gradient - numerical_gradient).mean() < tolerance)
else:
# check the gradient of each parameter individually, and do some pretty printing
try:
@ -549,7 +549,7 @@ class Model(Parameterized):
:type optimzer: string TODO: valid strings?
"""
assert isinstance(self.likelihood, likelihoods.EP), "pseudo_EM is only available for EP likelihoods"
assert isinstance(self.likelihood, likelihoods.EP) or isinstance(self.likelihood, likelihoods.EP_Mixed_Noise), "pseudo_EM is only available for EP likelihoods"
ll_change = epsilon + 1.
iteration = 0
last_ll = -np.inf

138
GPy/core/sparse_gp.py
View file

@ -5,7 +5,7 @@ import numpy as np
import pylab as pb
from ..util.linalg import mdot, jitchol, tdot, symmetrify, backsub_both_sides, chol_inv, dtrtrs, dpotrs, dpotri
from scipy import linalg
from ..likelihoods import Gaussian
from ..likelihoods import Gaussian, EP,EP_Mixed_Noise
from gp_base import GPBase
class SparseGP(GPBase):
@ -109,7 +109,6 @@ class SparseGP(GPBase):
tmp, _ = dtrtrs(self._Lm, np.asfortranarray(tmp.T), lower=1)
self._A = tdot(tmp)
# factor B
self.B = np.eye(self.num_inducing) + self._A
self.LB = jitchol(self.B)
@ -139,6 +138,7 @@ class SparseGP(GPBase):
dL_dpsi2_beta = 0.5 * backsub_both_sides(self._Lm, self.output_dim * np.eye(self.num_inducing) - self.DBi_plus_BiPBi)
if self.likelihood.is_heteroscedastic:
if self.has_uncertain_inputs:
self.dL_dpsi2 = self.likelihood.precision.flatten()[:, None, None] * dL_dpsi2_beta[None, :, :]
else:
@ -160,9 +160,23 @@ class SparseGP(GPBase):
# save computation here.
self.partial_for_likelihood = None
elif self.likelihood.is_heteroscedastic:
raise NotImplementedError, "heteroscedatic derivates not implemented"
if self.has_uncertain_inputs:
raise NotImplementedError, "heteroscedatic derivates with uncertain inputs not implemented"
else:
Lmi_psi1, nil = dtrtrs(self._Lm, np.asfortranarray(self.psi1.T), lower=1, trans=0)
_LBi_Lmi_psi1, _ = dtrtrs(self.LB, np.asfortranarray(Lmi_psi1), lower=1, trans=0)
_Bi_Lmi_psi1, _ = dtrtrs(self.LB.T, np.asfortranarray(_LBi_Lmi_psi1), lower=1, trans=0)
self.partial_for_likelihood = -0.5 * self.likelihood.precision + 0.5 * self.likelihood.V**2
self.partial_for_likelihood += 0.5 * self.output_dim * (self.psi0 - np.sum(Lmi_psi1**2,0))[:,None] * self.likelihood.precision**2
self.partial_for_likelihood += 0.5*np.sum(_Bi_Lmi_psi1*Lmi_psi1,0)[:,None]*self.likelihood.precision**2 #NOTE this term has numerical issues
self.partial_for_likelihood += -np.dot(self._LBi_Lmi_psi1Vf.T,_LBi_Lmi_psi1).T * self.likelihood.Y * self.likelihood.precision**2
self.partial_for_likelihood += 0.5*np.dot(self._LBi_Lmi_psi1Vf.T,_LBi_Lmi_psi1).T**2 * self.likelihood.precision**2
else:
# likelihood is not heterscedatic
# likelihood is not heteroscedatic
self.partial_for_likelihood = -0.5 * self.num_data * self.output_dim * self.likelihood.precision + 0.5 * self.likelihood.trYYT * self.likelihood.precision ** 2
self.partial_for_likelihood += 0.5 * self.output_dim * (self.psi0.sum() * self.likelihood.precision ** 2 - np.trace(self._A) * self.likelihood.precision)
self.partial_for_likelihood += self.likelihood.precision * (0.5 * np.sum(self._A * self.DBi_plus_BiPBi) - self.data_fit)
@ -298,7 +312,7 @@ class SparseGP(GPBase):
:type X_variance_new: np.ndarray, Nnew x self.input_dim
:param which_parts: specifies which outputs kernel(s) to use in prediction
:type which_parts: ('all', list of bools)
:param full_cov: whether to return the folll covariance matrix, or just the diagonal
:param full_cov: whether to return the full covariance matrix, or just the diagonal
:type full_cov: bool
:rtype: posterior mean, a Numpy array, Nnew x self.input_dim
:rtype: posterior variance, a Numpy array, Nnew x 1 if full_cov=False, Nnew x Nnew otherwise
@ -322,18 +336,15 @@ class SparseGP(GPBase):
return mean, var, _025pm, _975pm
def plot(self, samples=0, plot_limits=None, which_data='all', which_parts='all', resolution=None, levels=20, fignum=None, ax=None):
def plot(self, samples=0, plot_limits=None, which_data='all', which_parts='all', resolution=None, levels=20, fignum=None, ax=None, output=None):
if ax is None:
fig = pb.figure(num=fignum)
ax = fig.add_subplot(111)
if which_data is 'all':
which_data = slice(None)
GPBase.plot(self, samples=0, plot_limits=None, which_data='all', which_parts='all', resolution=None, levels=20, ax=ax)
# add the inducing inputs and some errorbars
if self.X.shape[1] == 1:
GPBase.plot(self, samples=0, plot_limits=plot_limits, which_data='all', which_parts='all', resolution=None, levels=20, ax=ax, output=output)
if self.X.shape[1] == 1 and not hasattr(self,'multioutput'):
if self.has_uncertain_inputs:
Xu = self.X * self._Xscale + self._Xoffset # NOTE self.X are the normalized values now
ax.errorbar(Xu[which_data, 0], self.likelihood.data[which_data, 0],
@ -342,6 +353,109 @@ class SparseGP(GPBase):
Zu = self.Z * self._Xscale + self._Xoffset
ax.plot(Zu, np.zeros_like(Zu) + ax.get_ylim()[0], 'r|', mew=1.5, markersize=12)
elif self.X.shape[1] == 2:
elif self.X.shape[1] == 2 and not hasattr(self,'multioutput'):
Zu = self.Z * self._Xscale + self._Xoffset
ax.plot(Zu[:, 0], Zu[:, 1], 'wo')
elif self.X.shape[1] == 2 and hasattr(self,'multioutput'):
Xu = self.X[self.X[:,-1]==output,:]
if self.has_uncertain_inputs:
Xu = self.X * self._Xscale + self._Xoffset # NOTE self.X are the normalized values now
Xu = self.X[self.X[:,-1]==output ,0:1] #??
ax.errorbar(Xu[which_data, 0], self.likelihood.data[which_data, 0],
xerr=2 * np.sqrt(self.X_variance[which_data, 0]),
ecolor='k', fmt=None, elinewidth=.5, alpha=.5)
Zu = self.Z[self.Z[:,-1]==output,:]
Zu = self.Z * self._Xscale + self._Xoffset
Zu = self.Z[self.Z[:,-1]==output ,0:1] #??
ax.plot(Zu, np.zeros_like(Zu) + ax.get_ylim()[0], 'r|', mew=1.5, markersize=12)
#ax.set_ylim(ax.get_ylim()[0],)
else:
raise NotImplementedError, "Cannot define a frame with more than two input dimensions"
def predict_single_output(self, Xnew, output=0, which_parts='all', full_cov=False):
"""
For a specific output, predict the function at the new point(s) Xnew.
Arguments
---------
:param Xnew: The points at which to make a prediction
:type Xnew: np.ndarray, Nnew x self.input_dim
:param output: output to predict
:type output: integer in {0,..., num_outputs-1}
:param which_parts: specifies which outputs kernel(s) to use in prediction
:type which_parts: ('all', list of bools)
:param full_cov: whether to return the full covariance matrix, or just the diagonal
:type full_cov: bool
:rtype: posterior mean, a Numpy array, Nnew x self.input_dim
:rtype: posterior variance, a Numpy array, Nnew x 1 if full_cov=False, Nnew x Nnew otherwise
:rtype: lower and upper boundaries of the 95% confidence intervals, Numpy arrays, Nnew x self.input_dim
.. Note:: For multiple output models only
"""
assert hasattr(self,'multioutput')
index = np.ones_like(Xnew)*output
Xnew = np.hstack((Xnew,index))
# normalize X values
Xnew = (Xnew.copy() - self._Xoffset) / self._Xscale
mu, var = self._raw_predict(Xnew, full_cov=full_cov, which_parts=which_parts)
# now push through likelihood
mean, var, _025pm, _975pm = self.likelihood.predictive_values(mu, var, full_cov, noise_model = output)
return mean, var, _025pm, _975pm
def _raw_predict_single_output(self, _Xnew, output=0, X_variance_new=None, which_parts='all', full_cov=False,stop=False):
"""
Internal helper function for making predictions for a specific output,
does not account for normalization or likelihood
---------
:param Xnew: The points at which to make a prediction
:type Xnew: np.ndarray, Nnew x self.input_dim
:param output: output to predict
:type output: integer in {0,..., num_outputs-1}
:param which_parts: specifies which outputs kernel(s) to use in prediction
:type which_parts: ('all', list of bools)
:param full_cov: whether to return the full covariance matrix, or just the diagonal
.. Note:: For multiple output models only
"""
Bi, _ = dpotri(self.LB, lower=0) # WTH? this lower switch should be 1, but that doesn't work!
symmetrify(Bi)
Kmmi_LmiBLmi = backsub_both_sides(self._Lm, np.eye(self.num_inducing) - Bi)
if self.Cpsi1V is None:
psi1V = np.dot(self.psi1.T,self.likelihood.V)
tmp, _ = dtrtrs(self._Lm, np.asfortranarray(psi1V), lower=1, trans=0)
tmp, _ = dpotrs(self.LB, tmp, lower=1)
self.Cpsi1V, _ = dtrtrs(self._Lm, tmp, lower=1, trans=1)
assert hasattr(self,'multioutput')
index = np.ones_like(_Xnew)*output
_Xnew = np.hstack((_Xnew,index))
if X_variance_new is None:
Kx = self.kern.K(self.Z, _Xnew, which_parts=which_parts)
mu = np.dot(Kx.T, self.Cpsi1V)
if full_cov:
Kxx = self.kern.K(_Xnew, which_parts=which_parts)
var = Kxx - mdot(Kx.T, Kmmi_LmiBLmi, Kx) # NOTE this won't work for plotting
else:
Kxx = self.kern.Kdiag(_Xnew, which_parts=which_parts)
var = Kxx - np.sum(Kx * np.dot(Kmmi_LmiBLmi, Kx), 0)
else:
Kx = self.kern.psi1(self.Z, _Xnew, X_variance_new)
mu = np.dot(Kx, self.Cpsi1V)
if full_cov:
raise NotImplementedError, "TODO"
else:
Kxx = self.kern.psi0(self.Z, _Xnew, X_variance_new)
psi2 = self.kern.psi2(self.Z, _Xnew, X_variance_new)
var = Kxx - np.sum(np.sum(psi2 * Kmmi_LmiBLmi[None, :, :], 1), 1)
return mu, var[:, None]