def setbeta(self, ep_beta):
sz = self.size_pars()
if not helpers.check_vecsize(ep_beta, sz):
raise TypeError('EP_BETA must be vector of size {0}'.format(sz))
if self.ep_beta is None:
self.ep_beta = np.empty(sz)
self.ep_beta[:] = ep_beta
def setpi(self, ep_pi):
sz = self.size_pars()
if not helpers.check_vecsize(ep_pi, sz):
raise TypeError('EP_PI must be vector of size {0}'.format(sz))
if self.ep_pi is None:
self.ep_pi = np.empty(sz)
self.ep_pi[:] = ep_pi
def mat_btdb(self,v,out=None):
if not helpers.check_vecsize(v,self.m):
raise TypeError('V wrong')
if out is None:
return np.diag(v)
else:
self._matbtdb_setdgout(out,v)
return out
def mat_btdb(self, v, out=None):
if not helpers.check_vecsize(v, self.m):
raise TypeError('V wrong')
if out is None:
return np.diag(v)
else:
self._matbtdb_setdgout(out, v)
return out
def predict(self, pbfact, pmeans, pvars=None):
if not isinstance(pbfact, cf.MatFactorizedInf):
raise TypeError('PBFACT must be apbsint.MatFactorizedInf')
pm, n = pbfact.shape()
if n != self.bfact.shape(1):
raise TypeError('PBFACT has wrong size')
if not (helpers.check_vecsize(pmeans, pm) and
(pvars is None or helpers.check_vecsize(pvars, pm))):
raise TypeError('PMEANS or PVARS wrong')
tvec = 1. / self.marg_pi
if pvars is not None:
# 'pbfact.b2fact' is B_test**2
pvars[:] = pbfact.b2fact.dot(tvec)
tvec *= self.marg_beta
if pmeans.flags['C_CONTIGUOUS']:
pbfact.mvm(tvec, pmeans)
else:
pmeans[:] = pbfact.mvm(tvec)
def _mvm_checkin(self, v, out):
"""
Tests input arguments to 'mvm'. Returns buffer vector for result,
which is 'out' if given, otherwise a new vector.
"""
m, n = self.shape()
if not helpers.check_vecsize(v, n):
raise TypeError('V wrong')
return self._check_resultvec(out, m)
def _mvm_checkin(self,v,out):
"""
Tests input arguments to 'mvm'. Returns buffer vector for result,
which is 'out' if given, otherwise a new vector.
"""
m, n = self.shape()
if not helpers.check_vecsize(v,n):
raise TypeError('V wrong')
return self._check_resultvec(out,m)
def _check_resultvec(self,out,sz):
"""
If 'out' is not None, checks that it is a contiguous vector of
size 'sz'. Otherwise, a vector of this size is created. In any
case, the vector is returned.
"""
if out is None:
out = np.empty(sz)
elif not (helpers.check_vecsize(out,sz) and
out.flags['C_CONTIGUOUS']):
raise TypeError('OUT argument wrong (must be contiguous)')
return out
def __init__(self,dg):
if isinstance(dg,MatDiag):
Mat.__init__(self,dg)
self.dg = dg.dg
self.sqdg = dg.sqdg
else:
if not helpers.check_vecsize(dg) or dg.shape[0] == 0:
raise TypeError('DG wrong type or size')
n = dg.shape[0]
Mat.__init__(self,n,n)
self.dg = dg
self.sqdg = dg**2
def __init__(self, dg):
if isinstance(dg, MatDiag):
Mat.__init__(self, dg)
self.dg = dg.dg
self.sqdg = dg.sqdg
else:
if not helpers.check_vecsize(dg) or dg.shape[0] == 0:
raise TypeError('DG wrong type or size')
n = dg.shape[0]
Mat.__init__(self, n, n)
self.dg = dg
self.sqdg = dg**2
def _check_resultvec(self, out, sz):
"""
If 'out' is not None, checks that it is a contiguous vector of
size 'sz'. Otherwise, a vector of this size is created. In any
case, the vector is returned.
"""
if out is None:
out = np.empty(sz)
elif not (helpers.check_vecsize(out, sz)
and out.flags['C_CONTIGUOUS']):
raise TypeError('OUT argument wrong (must be contiguous)')
return out
def predict(self, pbfact, pmeans, pvars=None, use_cov=False):
"""
Compute predictive means (and variances, optional), given test set
coupling factor B_p (in 'pbfact').
Predictive variances require A^-1. If 'post_cov' is defined and
'use_cov'==True, A^-1 is taken from 'post_cov'. Otherwise, it is
computed here (and written into 'post_cov').
NOTE: Use 'use_cov'=True if 'refresh' with 'keep_margs'=True has
been called just before.
"""
if not isinstance(pbfact, cf.Mat):
raise TypeError('PBFACT must be instance of apbsint.Mat')
pm, n = pbfact.shape()
if n != self.bfact.shape(1):
raise TypeError('PBFACT has wrong size')
if not (helpers.check_vecsize(pmeans, pm) and
(pvars is None or helpers.check_vecsize(pvars, pm))):
raise TypeError('PMEANS or PVARS wrong')
# Predictive means
pbfact.mvm(
sla.solve_triangular(self.lfact, self.cvec, lower=True, trans='T'),
pmeans)
if pvars is not None:
# Predictive variances: Need inverse A^-1
try:
if self.post_cov.shape != (n, n):
raise TypeError(
'Internal error: POST_COV attribute has wrong size')
except AttributeError:
if use_cov:
raise ValueError('POST_COV is not defined')
self.post_cov = np.empty((n, n))
amat = self.post_cov
if not use_cov:
self._comp_inva(amat)
pbfact.diag_bsbt(amat, pvars)
def mat_btdb(self,v,out=None):
if self.transp:
raise NotImplementedError("NOT IMPLEMENTED")
if not helpers.check_vecsize(v,self.m):
raise TypeError('V wrong')
off = 0
m, n = self.shape()
if out is None:
out = np.empty((n,n))
else:
self._matbtdb_checkout(out)
out.fill(0.)
# 'buffmat1' is allocated once, then reused
try:
self.buffmat1.resize((n,n),refcheck=False)
except AttributeError:
self.buffmat1 = np.empty((n,n))
for chd in self.child:
sz = chd.shape(0)
chd.mat_btdb(v[off:off+sz],self.buffmat1)
out += self.buffmat1
off += sz
return out
def mat_btdb(self, v, out=None):
if self.transp:
raise NotImplementedError("NOT IMPLEMENTED")
if not helpers.check_vecsize(v, self.m):
raise TypeError('V wrong')
off = 0
m, n = self.shape()
if out is None:
out = np.empty((n, n))
else:
self._matbtdb_checkout(out)
out.fill(0.)
# 'buffmat1' is allocated once, then reused
try:
self.buffmat1.resize((n, n), refcheck=False)
except AttributeError:
self.buffmat1 = np.empty((n, n))
for chd in self.child:
sz = chd.shape(0)
chd.mat_btdb(v[off:off + sz], self.buffmat1)
out += self.buffmat1
off += sz
return out
def get_marg(self, j, vvec=None):
"""
Returns (mu, rho), mu marginal mean, rho marginal variance at potential
j (Gaussian marginal, not tilted marginal).
If 'vvec' is given, L^-1 B[j,:] is written there. In this case, the
marginal is always computed from scratch. Otherwise, if
'keep_margs'==True, we use 'marg_XXX'.
If (mu, rho) are computed from scratch and 'keep_margs'==True, the
corr. entries of 'marg_XXX' are refreshed.
"""
bfact = self.bfact
m, n = bfact.shape()
if not (isinstance(j, numbers.Integral) and j >= 0 and j < m):
raise ValueError('J wrong')
if vvec is None:
if self.keep_margs:
return (self.marg_means[j], self.marg_vars[j])
vvec = np.empty(n)
else:
if not helpers.check_vecsize(vvec, n):
raise TypeError('VVEC wrong')
try:
self.us_bvec.resize(n, refcheck=False)
except AttributeError:
self.us_bvec = np.empty(n)
bfact.T().getcol(j, self.us_bvec)
vvec[:] = sla.solve_triangular(self.lfact,
self.us_bvec,
lower=True,
trans='N')
mu = np.inner(vvec, self.cvec)
rho = np.inner(vvec, vvec)
if self.keep_margs:
# Refresh entries
self.marg_means[j] = mu
self.marg_vars[j] = rho
return (mu, rho)
def mat_btdb(self,v,out=None):
"""
Returns matrix B^T (diag v) B.
If 'out' is given, the result is written there directly. 'out' must
have exactly the right size and must be C contiguous, otherwise an
exception is thrown.
"""
m, n = self.shape()
if not helpers.check_vecsize(v,m):
raise TypeError('V wrong')
if out is None:
out = np.empty((n,n))
else:
self._matbtdb_checkout(out)
bt = self.T()
tv1 = np.zeros(n)
tv2 = np.empty(m)
for i in xrange(n):
tv1[i] = 1.
self.mvm(tv1,tv2)
tv1[i] = 0.
tv2 *= v
bt.mvm(tv2,out[i])
return out
def mat_btdb(self, v, out=None):
"""
Returns matrix B^T (diag v) B.
If 'out' is given, the result is written there directly. 'out' must
have exactly the right size and must be C contiguous, otherwise an
exception is thrown.
"""
m, n = self.shape()
if not helpers.check_vecsize(v, m):
raise TypeError('V wrong')
if out is None:
out = np.empty((n, n))
else:
self._matbtdb_checkout(out)
bt = self.T()
tv1 = np.zeros(n)
tv2 = np.empty(m)
for i in xrange(n):
tv1[i] = 1.
self.mvm(tv1, tv2)
tv1[i] = 0.
tv2 *= v
bt.mvm(tv2, out[i])
return out
def seldamp_reset(self, numk, subind=None, subexcl=False):
"""
Initializes or resets the selective damping (SD) representation. SD
ensures that cavity marginals are well-defined after each EP update.
The SD representation tracks max_k pi_{k,i} for each variable, by
storing the 'numk' largest pi values for each i. The larger 'numk',
the less often maxima have to be recomputed.
If 'subind' is given, max_k runs only over this index (if
'subexcl'==False) or over its complement (if 'subexcl'==True).
"""
bf = self.bfact
m, n = bf.shape()
if not isinstance(numk, numbers.Integral) or numk < 2:
raise TypeError('NUMK must be integer > 1')
if not (subind is None or
(helpers.check_vecsize(subind) and subind.dtype == np.int32)):
raise TypeError(
'SUBIND must be numpy.ndarray with dtype numpy.int32')
(self.sd_numvalid, self.sd_topind, self.sd_topval) \
= epx.fact_compmaxpi(n,m,bf.rowind,bf.colind,bf.bvals,self.ep_pi,
self.ep_beta,numk,subind,subexcl)
self.sd_subind = subind
self.sd_subexcl = subexcl
self.sd_numk = numk
def update_single(self, j, delpi, delbeta, vvec=None):
"""
Change of EP parameters:
ep_pi[j] += delpi; ep_beta[j] += delbeta
The representation is updated accordingly. In particular, the
Cholesky factor 'lfact' is updated ('delpi'>0) or downdated
('delpi'<0). If 'keep_margs'==True, the marginal moments are
updated as well.
In 'vvec', the vector L^-1 B[j,:] can be passed. If not, it is
recomputed here.
NOTE: 'post_cov' (if given) is not updated!
"""
bfact = self.bfact
m, n = bfact.shape()
if not (isinstance(j, numbers.Integral) and j >= 0 and j < m
and isinstance(delpi, numbers.Real)
and isinstance(delbeta, numbers.Real)):
raise ValueError('J, DELPI or DELBETA wrong')
if not (vvec is None or helpers.check_vecsize(vvec, n)):
raise TypeError('VVEC wrong')
# Scratch variables. We keep them as members, to avoid having to
# allocate them in every call
try:
self.cup_c.resize(n, refcheck=False)
self.cup_s.resize(n, refcheck=False)
self.cup_wk.resize(n, refcheck=False)
self.cup_z.resize((1, n), refcheck=False)
self.us_bvec.resize(n, refcheck=False)
if self.keep_margs:
self.us_wvec.resize(m, refcheck=False)
self.us_w2vec.resize(m, refcheck=False)
except AttributeError:
self.cup_c = np.empty(n)
self.cup_s = np.empty(n)
self.cup_wk = np.empty(n)
self.cup_z = np.empty((1, n), order='F')
self.us_bvec = np.empty(n)
if self.keep_margs:
self.us_wvec = np.empty(m)
self.us_w2vec = np.empty(m)
bvec = self.us_bvec
if self.keep_margs:
wvec = self.us_wvec
w2vec = self.us_w2vec
if delpi > 0.:
# Cholesky update
tscal = np.sqrt(delpi)
bfact.T().getcol(j, bvec)
if self.keep_margs:
if vvec is None:
# Need 'vvec' below, so compute it here
vvec = sla.solve_triangular(self.lfact,
bvec,
lower=True,
trans='N')
mu = np.inner(vvec, self.cvec)
rho = np.inner(vvec, vvec)
bvec *= tscal
yscal = np.empty(1)
yscal[0] = delbeta / tscal
self.cup_z[0] = self.cvec
stat = epx.choluprk1(self.lfact, 'L', bvec, self.cup_c, self.cup_s,
self.cup_wk, self.cup_z, yscal)
if stat != 0:
raise sla.LinAlgError(
"Numerical error in 'choluprk1' (external)")
self.cvec[:] = self.cup_z.ravel()
else:
# Cholesky downdate
tscal = np.sqrt(-delpi)
if vvec is None:
bfact.T().getcol(j, bvec)
vvec = sla.solve_triangular(self.lfact,
bvec,
lower=True,
trans='N')
if self.keep_margs:
mu = np.inner(vvec, self.cvec)
rho = np.inner(vvec, vvec)
yscal = np.empty(1)
yscal[0] = -delbeta / tscal
self.cup_z[0] = self.cvec
bvec[:] = vvec
bvec *= tscal
stat = epx.choldnrk1(self.lfact, 'L', bvec, self.cup_c, self.cup_s,
self.cup_wk, self.cup_z, yscal)
if stat != 0:
raise sla.LinAlgError(
"Numerical error in 'choldnrk1' (external)")
self.cvec[:] = self.cup_z.ravel()
self.ep_pi[j] += delpi
self.ep_beta[j] += delbeta
if self.keep_margs:
# Update marginal moments
assert vvec is not None
bfact.mvm(
sla.solve_triangular(self.lfact, vvec, lower=True, trans='T'),
wvec)
tscal = 1. / (delpi * rho + 1.)
w2vec[:] = wvec
w2vec *= ((delbeta - delpi * mu) * tscal)
self.marg_means += w2vec
wvec *= wvec
wvec *= (delpi * tscal)
self.marg_vars -= wvec
