def mergeCovarianceToGF(self):
"""
merge data covariance (cd) with green function
"""
# get references for data covariance
cd_inv = self.dataobs.gcd_inv
# get a reference for green's function
green = self.gGF
# copy from CPU
green.copy_from_host(source=self.GF)
# check whether cd is a constant or a matrix
if isinstance(cd_inv, float):
green *= cd_inv
elif isinstance(cd_inv, altar.cuda.matrix):
# (obsxobs) x (obsxparameters) = (obsxparameters)
cublas.trmm(cd_inv,
green,
out=green,
side=cublas.SideLeft,
uplo=cublas.FillModeUpper,
transa=cublas.OpNoTrans,
diag=cublas.DiagNonUnit,
alpha=1.0,
handle=self.cublas_handle)
# all done
return
def displace(self, displacement):
"""
Construct a set of displacement vectors for the random walk from a distribution with zero
mean and my covariance
"""
# get my decomposed covariance
Σ_chol = self.gsigma_chol
# generate gaussian random numbers (samples x parameters)
altar.cuda.curand.gaussian(out=displacement)
# multiply by the sigma_cholesky
cublas.trmm(Σ_chol, displacement, out=displacement,
alpha=1.0, uplo=cublas.FillModeUpper, side=cublas.SideRight,
transa = cublas.OpNoTrans, diag=cublas.DiagNonUnit)
# and return
return displacement
def cuEvalLikelihood(self,
data,
constant=0.0,
out=None,
batch=None,
cdinv=None):
"""
Compute the L2 norm data likelihood of the given data const - ||x||^2/2
Arguments:
data - matrix (samples x observations)
batch - number of samples to be computed (first rows)
constant - normalization constant
cdinv - inverse covariance matrix (observations x observations) in its Cholesky decomposed form (Upper Triangle)
Return:
out - data likelihood vector (samples)
"""
samples = data.shape[0]
# no batch provided, compute all samples
if batch is None or batch > samples:
batch = samples
# if output is not pre-allocated, allocate it
if out is None:
out = altar.cuda.vector(shape=samples)
# if covariance matrix is provided
if cdinv is not None:
# data [samples][obs] = data[samples][obs] x cdinv[obs][obs]
cublas.trmm(cdinv,
data,
out=data,
alpha=1.0,
uplo=cublas.FillModeUpper,
side=cublas.SideRight,
transa=cublas.OpNoTrans,
diag=cublas.DiagNonUnit)
# compute the norm
libcudaaltar.cudaL2_normllk(data.data, out.data, batch, constant)
# return the result
return out
def prepareGF(self):
"""
copy green function to gpu and merge cd with green function
"""
# make a gpu copy
self.gGF = altar.cuda.matrix(source=self.GF, dtype=self.precision)
# merge cd with Green's function
cd_inv = self.dataobs.gcd_inv
green = self.gGF
# (obsxobs) x (obsxparameters) = (obsxparameters)
cublas.trmm(cd_inv,
green,
out=green,
side=cublas.SideLeft,
uplo=cublas.FillModeUpper,
transa=cublas.OpNoTrans,
diag=cublas.DiagNonUnit,
alpha=1.0,
handle=self.cublas_handle)
#cublas.gemm(cd_inv, green, out=green)
return