mirror of
https://github.com/SheffieldML/GPy.git
synced 2026-05-15 06:52:39 +02:00
428 lines
18 KiB
Python
428 lines
18 KiB
Python
# Copyright (c) 2012, James Hensman and Ricardo Andrade
|
|
# Licensed under the BSD 3-clause license (see LICENSE.txt)
|
|
|
|
from kernpart import Kernpart
|
|
import numpy as np
|
|
from GPy.util.linalg import mdot, pdinv
|
|
from GPy.util import ln_diff_erfs
|
|
import pdb
|
|
from scipy import weave
|
|
|
|
class Ode1(Kernpart):
|
|
"""
|
|
Covariance function for first order differential equation driven by an exponetiated quadratic covariance.
|
|
|
|
This kernel has the form
|
|
.. math::
|
|
|
|
:param num_outputs: number of outputs driven by latent function.
|
|
:type num_outputs: int
|
|
|
|
.. Note: see first order differential equation examples in GPy.examples.regression for some usage.
|
|
"""
|
|
def __init__(self,num_outputs, W=None, rank=1, delay=None, kappa=None):
|
|
self.rank = rank
|
|
self.input_dim = 1
|
|
self.name = 'ode1'
|
|
self.num_outputs = num_outputs
|
|
self.num_params = self.num_outputs*(1. + self.rank) + 1
|
|
if kappa is not None:
|
|
self.num_params+=num_outputs
|
|
if delay is not None:
|
|
self.num_params+=num_outputs
|
|
self.rank = rank
|
|
if W is None:
|
|
self.W = 0.5*np.random.randn(self.num_outputs,self.rank)/np.sqrt(self.rank)
|
|
else:
|
|
assert W.shape==(self.num_outputs,self.rank)
|
|
self.W = W
|
|
|
|
if kappa is not None:
|
|
assert kappa.shape==(self.num_outputs,)
|
|
self.kappa = kappa
|
|
|
|
if delay is not None:
|
|
assert delay.shape==(self.num_outputs,)
|
|
self.delay = delay
|
|
self._set_params(self._get_params())
|
|
|
|
def _get_params(self):
|
|
param_list = [self.W.flatten()]
|
|
if self.kappa is not None:
|
|
param_list.append(self.kappa)
|
|
param_list.append(self.decay)
|
|
if self.delay is not None:
|
|
param_list.append(self.delay)
|
|
param_list.append(self.length_scale)
|
|
return np.hstack(param_list)
|
|
|
|
def _set_params(self,x):
|
|
assert x.size == self.num_params
|
|
end = self.num_outputs*self.rank
|
|
self.W = x[:end].reshape(self.num_outputs,self.rank)
|
|
start = end
|
|
self.B = np.dot(self.W,self.W.T)
|
|
if self.kappa is not None:
|
|
end+=self.num_outputs
|
|
self.kappa = x[start:end]
|
|
self.B += np.diag(self.kappa)
|
|
start=end
|
|
end+=num_outputs
|
|
self.decay = x[start:end]
|
|
start=end
|
|
if self.delay is not None:
|
|
end+=num_outputs
|
|
self.delay = x[start:end]
|
|
start=end
|
|
end+=1
|
|
self.length_scale = x[start]
|
|
self.sigma = np.sqrt(2)*self.length_scale
|
|
|
|
|
|
def _get_param_names(self):
|
|
param_names = sum([['W%i_%i'%(i,j) for j in range(self.rank)] for i in range(self.num_outputs)],[])
|
|
if self.kappa is not None:
|
|
param_names += ['kappa_%i'%i for i in range(self.num_outputs)]
|
|
param_names += ['decay_%i'%i for i in range(self.num_outputs)]
|
|
if self.delay is not None:
|
|
param_names += ['delay_%i'%i for i in range(self.num_outputs)]
|
|
param_names+= ['length_scale']
|
|
return param_names
|
|
|
|
def K(self,X,X2,target):
|
|
|
|
if X.shape[1] > 2:
|
|
raise ValueError('Input matrix for ode1 covariance should have at most two columns, one containing times, the other output indices')
|
|
|
|
self._K_computations()
|
|
target += self._scales*self._dK_dvar
|
|
|
|
if self.gaussian_initial:
|
|
# Add covariance associated with initial condition.
|
|
t1_mat = self._t[self._rorder, None]
|
|
t2_mat = self._t2[None, self._rorder2]
|
|
target+=self.initial_variance * np.exp(- self.decay * (t1_mat + t2_mat))
|
|
|
|
|
|
|
|
def Kdiag(self,index,target):
|
|
#target += np.diag(self.B)[np.asarray(index,dtype=np.int).flatten()]
|
|
pass
|
|
|
|
def dK_dtheta(self,dL_dK,index,index2,target):
|
|
pass
|
|
|
|
def dKdiag_dtheta(self,dL_dKdiag,index,target):
|
|
pass
|
|
|
|
def dK_dX(self,dL_dK,X,X2,target):
|
|
pass
|
|
|
|
def _extract_t_indices(X, X2=None):
|
|
"""Extract times and output indices from the input matrix X. Times are ordered according to their index for convenience of computation, this ordering is stored in self._order and self.order2. These orderings are then mapped back to the original ordering (in X) using self._rorder and self._rorder2. """
|
|
|
|
# TODO: some fast checking here to see if this needs recomputing?
|
|
self._t = X[:, 0]
|
|
if X.shape[1]==1:
|
|
# No index passed, assume single output of ode model.
|
|
self._index = np.ones_like(X[:, 1],dtype=np.int)
|
|
self._index = np.asarray(X[:, 1],dtype=np.int)
|
|
# Sort indices so that outputs are in blocks for computational
|
|
# convenience.
|
|
self._order = self._index.argsort()
|
|
self._index = self._index[self._order]
|
|
self._t = self._t[self._order]
|
|
self._rorder = self._order.argsort() # rorder is for reversing the order
|
|
|
|
if X2 is None:
|
|
self._t2 = None
|
|
self._index2 = None
|
|
self._rorder2 = self._rorder
|
|
else:
|
|
if X2.shape[1] > 2:
|
|
raise ValueError('Input matrix for ode1 covariance should have at most two columns, one containing times, the other output indices')
|
|
self._t2 = X2[:, 0]
|
|
if X.shape[1]==1:
|
|
# No index passed, assume single output of ode model.
|
|
self._index2 = np.ones_like(X2[:, 1],dtype=np.int)
|
|
self._index2 = np.asarray(X2[:, 1],dtype=np.int)
|
|
self._order2 = self._index2.argsort()
|
|
slef._index2 = self._index2[self._order2]
|
|
self._t2 = self._t2[self._order2]
|
|
self._rorder2 = self._order2.argsort() # rorder2 is for reversing order
|
|
|
|
def _K_computations(X, X2):
|
|
"""Perform main body of computations for the ode1 covariance function."""
|
|
# First extract times and indices.
|
|
self._extract_t_indices(X, X2)
|
|
|
|
self._K_compute_eq()
|
|
self._K_compute_eq_x_ode()
|
|
if X2 is None:
|
|
self._K_ode_eq = self._K_ode_eq.T
|
|
else:
|
|
self._K_compute_eq_x_ode(transpose=True)
|
|
self._K_compute_ode()
|
|
|
|
# Reorder values of blocks for placing back into _K_dvar.
|
|
self._K_dvar[self._rorder, :] = np.vstack((
|
|
np.hstack((self._K_eq, self._Keq_ode))
|
|
np.hstack((self._K_ode_eq, self.K_ode))))[:, self._rorder2]
|
|
|
|
|
|
def _K_compute_eq():
|
|
"""Compute covariance for latent covariance."""
|
|
t_eq = self._t[self._index==0]
|
|
if t_eq.shape[0]==0:
|
|
self._K_eq = np.zeros((0, 0))
|
|
return
|
|
|
|
if self._t2 is None:
|
|
self._dist2 = np.square(t_eq[:, None] - t_eq[None, :])
|
|
else:
|
|
t2_eq = self._t2[self._index2==0]
|
|
if t2_eq.shape[0]==0:
|
|
self._K_eq_eq = np.zeros((0, 0))
|
|
return
|
|
self._dist2 = np.square(t_eq[:, None] - t2_eq[None, :])
|
|
|
|
self._K_eq = np.exp(-self._dist2/(2*self.length_scale*self.length_scale))
|
|
if self.is_normalise:
|
|
self._K_eq/=(np.sqrt(2*np.pi)*self.length_scale)
|
|
|
|
def _K_compute_ode_eq(transpose=False):
|
|
"""Compute the cross covariances between latent exponentiated quadratic and observed ordinary differential equations.
|
|
|
|
:param transpose: if set to false the exponentiated quadratic is on the rows of the matrix and is computed according to self._t, if set to true it is on the columns and is computed according to self._t2 (default=False).
|
|
:type transpose: bool"""
|
|
|
|
if transpose:
|
|
t_ode = self._t2[self._index>0]
|
|
index_ode = self._index2[self._index>0]-1
|
|
if t_ode.shape[0]==0:
|
|
self._K_ode = np.zeros((0, 0))
|
|
return
|
|
|
|
if self._t2 is not None:
|
|
t2_ode = self._t2[self._index2>0]
|
|
index2_ode = self._index2[self._index2>0]-1
|
|
if t2_eq.shape[0]==0:
|
|
self._K_ode = np.zeros((0, 0))
|
|
return
|
|
else:
|
|
|
|
# Matrix giving scales of each output
|
|
self._scale = np.zeros((t_ode.shape[0], t_eq.shape[0]))
|
|
code="""
|
|
for(int i=0;i<N; i++){
|
|
for(int j=0; j<N2; j++){
|
|
scale_mat[i+j*N] = self.W[index_sim[i]+index_eq[j]*num_outputs];
|
|
}
|
|
}
|
|
"""
|
|
scale_mat, B = self._scale, self._B
|
|
N, N2, num_outputs = index_ode.size, index_eq.size, self.num_outputs
|
|
weave.inline(code,['index_ode', 'index_eq',
|
|
'scale_mat', 'B',
|
|
'N', 'N2', 'num_outputs'])
|
|
else:
|
|
self._scale = np.zeros((t_ode.shape[0], t2_ode.shape[0]))
|
|
code = """
|
|
for(int i=0; i<N; i++){
|
|
for(int j=0; j<N2; j++){
|
|
scale_mat[i+j*N] = B[index_ode[i]+num_outputs*index2_ode[j]]
|
|
}
|
|
}
|
|
"""
|
|
scale_mat, B = self._scale, self._B
|
|
N, N2, num_outputs = index_ode.size, index2.size, self.num_outputs
|
|
weave.inline(code, ['index_ode', 'index2_ode',
|
|
'scale_mat', 'B',
|
|
'N', 'N2', 'num_outputs'])
|
|
if transpose:
|
|
t_ode = t2 - self.delay
|
|
t_eq_mat = t1[None, :]
|
|
else:
|
|
t_ode = t1 - self.delay
|
|
t_eq_mat = t2[None, :]
|
|
t_ode_mat = t_ode[:, None]
|
|
diff_t = (t_ode_mat - t_eq_mat)
|
|
sigma = sqrt(2/self.inverseWidth)
|
|
|
|
invSigmaDiffT = 1/sigma*diff_t
|
|
halfSigmaD_i = 0.5*sigma*self.decay
|
|
|
|
if self.isStationary == false
|
|
[ln_part, signs] = ln_diff_erfs(halfSigmaD_i + t2Mat/sigma, halfSigmaD_i - invSigmaDiffT)
|
|
else
|
|
[ln_part, signs] = ln_diff_erfs(inf, halfSigmaD_i - invSigmaDiffT)
|
|
end
|
|
sK = signs .* exp(halfSigmaD_i*halfSigmaD_i - self.decay*diff_t + ln_part)
|
|
|
|
sK *= 0.5
|
|
|
|
if not self.is_normalised:
|
|
sK *= sqrt(pi)*self.sigma
|
|
|
|
|
|
if transpose:
|
|
self._K_eq_ode =
|
|
else:
|
|
self._K_ode_eq = sK
|
|
return K
|
|
|
|
def _K_compute_ode():
|
|
# Compute covariances between outputs of the ODE models.
|
|
|
|
t_ode = self._t[self._index>0]
|
|
index_ode = self._index[self._index>0]-1
|
|
if t_ode.shape[0]==0:
|
|
self._K_ode = np.zeros((0, 0))
|
|
return
|
|
|
|
if self._t2 is not None:
|
|
t2_ode = self._t2[self._index2>0]
|
|
index2_ode = self._index2[self._index2>0]-1
|
|
if t2_eq.shape[0]==0:
|
|
self._K_ode = np.zeros((0, 0))
|
|
return
|
|
|
|
if self._index2 is None:
|
|
# Matrix giving scales of each output
|
|
self._scale = np.zeros((t_ode.shape[0], t_ode.shape[0]))
|
|
code="""
|
|
for(int i=0;i<N; i++){
|
|
scale_mat[i+i*N] = B[index_ode[i]+num_outputs*(index_ode[i])];
|
|
for(int j=0; j<i; j++){
|
|
scale_mat[j+i*N] = B[index_ode[i]+num_outputs*index_ode[j]];
|
|
scale_mat[i+j*N] = scale_mat[j+i*N];
|
|
}
|
|
}
|
|
"""
|
|
scale_mat, B = self._scale, self._B
|
|
N, num_outputs = index_ode.size, self.num_outputs
|
|
weave.inline(code,['index_ode',
|
|
'scale_mat', 'B',
|
|
'N', 'num_outputs'])
|
|
else:
|
|
self._scale = np.zeros((t_ode.shape[0], t2_ode.shape[0]))
|
|
code = """
|
|
for(int i=0; i<N; i++){
|
|
for(int j=0; j<N2; j++){
|
|
scale_mat[i+j*N] = B[index_ode[i]+num_outputs*index2_ode[j]]
|
|
}
|
|
}
|
|
"""
|
|
scale_mat, B = self._scale, self._B
|
|
N, N2, num_outputs = index_ode.size, index2.size, self.num_outputs
|
|
weave.inline(code, ['index_ode', 'index2_ode',
|
|
'scale_mat', 'B',
|
|
'N', 'N2', 'num_outputs'])
|
|
|
|
# When index is identical
|
|
if self.is_stationary:
|
|
h = self._compute_H_stat(t_ode, index_ode, t2_ode, index2_ode)
|
|
else:
|
|
h = self._compute_H(t_ode, index_ode, t2_ode, index2_ode)
|
|
|
|
if t2 is None:
|
|
self._K_ode = 0.5 * (h + h.T)
|
|
else:
|
|
if self.is_stationary:
|
|
h2 = self._compute_H_stat(t2, index2, t, index)
|
|
else:
|
|
h2 = self._compute_H(t2, index2, t, index)
|
|
self._K_ode += 0.5 * (h + h2.T)
|
|
|
|
if not self.is_normalized:
|
|
self._K_ode *= np.sqrt(np.pi)*sigma
|
|
|
|
def _compute_H(t, index, t2, index2, update_derivatives=False):
|
|
"""Helper function for computing part of the ode1 covariance function.
|
|
|
|
:param t: first time input.
|
|
:type t: array
|
|
:param index: Indices of first output.
|
|
:type index: array of int
|
|
:param t2: second time input.
|
|
:type t2: array
|
|
:param index2: Indices of second output.
|
|
:type index2: array of int
|
|
:param update_derivatives: whether to update derivatives (default is False)
|
|
:return h : result of this subcomponent of the kernel for the given values.
|
|
:rtype: ndarray
|
|
"""
|
|
|
|
if t.shape[1] > 1 or t2.shape[1] > 1:
|
|
raise error('Input can only have one column')
|
|
|
|
# Vector of decays and delays associated with each output.
|
|
Decay = np.zeros(t.shape[0])
|
|
Delay = np.zeros(t.shape[0])
|
|
Decay2 = np.zeros(t2.shape[0])
|
|
Delay2 = np.zeros(t2.shape[0])
|
|
code="""
|
|
for(int i=0;i<N; i++){
|
|
Delay[i] = decays[index[i]];
|
|
Decay[i] = delays[index[i]];
|
|
}
|
|
for(int i=0; i<N2; i++){
|
|
Delay2[i] = decays[index2[i]];
|
|
Decay2[i] = delays[index2[i]];
|
|
}
|
|
"""
|
|
delays, decays = self.delays, self.decays
|
|
N, N2 = index.size, index2.size
|
|
weave.inline(code,['index_ode',
|
|
'Delay', 'Decay',
|
|
'Delay2', 'Decay2',
|
|
'delays', 'decays',
|
|
'N', 'N2'])
|
|
|
|
t_mat = t[:, None]-Delay[:, None]
|
|
t2_mat = t2[None, :]-Delay2[None, :]
|
|
diff_t = (t_mat - t2_mat)
|
|
inv_sigma_diff_t = 1./self.sigma*diff_t
|
|
half_sigma_decay_i = 0.5*self.sigma*Decay[:, None]
|
|
|
|
ln_part_1, sign1 = ln_diff_erfs(half_sigma_decay_i + t2_mat/self.sigma,
|
|
half_sigma_decay_i - inv_sigma_diff_t)
|
|
ln_part_2, sign2 = ln_diff_erfs(half_sigma_decay_i,
|
|
half_sigma_decay_i - t_mat/self.sigma)
|
|
|
|
h = sign1*np.exp(half_sigma_decay_i
|
|
*half_sigma_decay_i
|
|
-Decay[:, None]*diff_t+ln_part_1
|
|
-np.log(Decay[:, None] + Decay2[None, :]))
|
|
h -= sign2*np.exp(half_sigma_decay_i*half_sigma_decay_i
|
|
-Decay[:, None]*t_mat-Decay2[None, :]*t2_mat+ln_part_2
|
|
-np.log(Decay[:, None] + Decay2[None, :]))
|
|
|
|
|
|
# if update_derivatives:
|
|
# sigma2 = self.sigma*self.sigma
|
|
# # Update ith decay gradient
|
|
# dh_ddecay += (0.5*self.decay[i]*sigma2*(self.decay[i] + decay[j])-1)*h
|
|
# + (-diff_t*sign1*np.exp(half_sigma_decay_i*half_sigma_decay_i-self.decay[i]*diff_t+ln_part_1)
|
|
# +t_mat*sign2*np.exp(half_sigma_decay_i*half_sigma_decay_i-self.decay[i]*t_mat - decay[j]*t2_mat+ln_part_2)) ...
|
|
# +self.sigma/sqrt(pi)*(-np.exp(-diff_t*diff_t/sigma2)
|
|
# +np.exp(-t2_mat*t2_mat/sigma2-self.decay[i]*t_mat)
|
|
# +np.exp(-t_mat*t_mat/sigma2-decay[j]*t2_mat) ...
|
|
# -np.exp(-(self.decay[i]*t_mat + decay[j]*t2_mat)))
|
|
# self._dh_ddecay[i] += real(dh_ddecay/(self.decay[i]+decay[j]))
|
|
|
|
# # Update jth decay gradient
|
|
# dh_ddecay = t2_mat*sign2*np.exp(half_sigma_decay_i*half_sigma_decay_i-(self.decay[i]*t_mat + decay[j]*t2_mat)+ln_part_2)-h
|
|
# self._dh_ddecay[j] += real(dh_ddecay/(self.decay[i] + decay[j]))
|
|
|
|
# # Update sigma gradient
|
|
# self._dh_dsigma += 0.5*self.decay[i]*self.decay[i]*self.sigma*h + 2/(np.sqrt(np.pi)*(self.decay[i]+decay[j]))*((-diff_t/sigma2-self.decay[i]/
|
|
# 2)*np.exp(-diff_t*
|
|
# diff_t/sigma2)
|
|
# + (-t2_mat/sigma2+self.decay[i]/2)
|
|
# *np.exp(-t2_mat*t2_mat/sigma2
|
|
# -self.decay[i]*t_mat)
|
|
# - (-t_mat/sigma2-self.decay[i]/2)
|
|
# *np.exp(-t_mat*t_mat/sigma2-decay[j]*t2_mat)
|
|
# - self.decay[i]/2*np.exp(-(self.decay[i]*t_mat+decay[j]*t2_mat)))
|
|
|