def get_isomgchain(lmax_sky, datshape, tol=1e-5, iter_max=np.inf, **kwargs):
assert datshape[0] == datshape[1], datshape
nside_max = datshape[0]
return [[
0, ["diag_cl"], lmax_sky, nside_max, iter_max, tol, cd_solve.tr_cg,
cd_solve.cache_mem()
]]
def solve(self, x, tqu):
self.watch = util.stopwatch()
self.iter_tot = 0
self.prev_eps = None
logger = (lambda iter, eps, stage=0, **kwargs: self.log(
stage, iter, eps, **kwargs))
monitor = cd_monitors.monitor_basic(self.opfilt.dot_op(),
logger=logger,
iter_max=self.iter_max,
eps_min=self.eps_min)
cd_solve.cd_solve(x=x,
b=self.opfilt.calc_prep(tqu, self.s_inv_filt,
self.n_inv_filt),
fwd_op=self.opfilt.fwd_op(self.s_inv_filt,
self.n_inv_filt),
pre_ops=[self.pre_op],
dot_op=self.opfilt.dot_op(),
criterion=monitor,
tr=cd_solve.tr_cg,
cache=cd_solve.cache_mem())
return self.opfilt.calc_fini(x, self.s_inv_filt, self.n_inv_filt)
def __init__(self,
nstages,
opfilt,
s_inv_filt,
n_inv_filt,
plogdepth=0,
eps_min=1.e-6,
stage_iter_max=3):
""" initialize this multigrid chain by constructing the resolution stages.
* nstages = total number of stages in the chain.
* opfilt = module defining dot_op, fwd_op, calc_prep, calc_fini, and pre_op_diag filter functions.
* s_inv_filt = the S^{-1} filter for the highest resolution stage in the chain.
* n_inv_filt = the N^{-1} filter for the highest resolution stage in the chain.
* plogdepth = the lowest stage at which to print convergence information.
* eps_min = the convergence criterion.
* stage_iter_max = the maximum number of iterations in the substages.
"""
self.opfilt = opfilt
self.s_inv_filt = s_inv_filt
self.n_inv_filt = n_inv_filt
self.plogdepth = plogdepth
self.eps_min = eps_min
self.stage_iter_max = stage_iter_max
self.iter_max = np.inf
pre_op = opfilt.pre_op_diag(s_inv_filt, n_inv_filt.degrade(2**nstages))
for i in xrange(0, nstages):
class slog(object):
def __init__(self, id, cobj):
self.id = 1 * id
self.cobj = cobj
def log(self, iter, eps, **kwargs):
self.cobj.log(self.id, iter, eps, **kwargs)
pre_op = pre_op_split(
pre_op_multigrid(opfilt, s_inv_filt,
n_inv_filt.degrade(2**(nstages - i)),
[pre_op],
slog(nstages - i, self).log, cd_solve.tr_cg,
cd_solve.cache_mem(), stage_iter_max, 0.0),
opfilt.pre_op_diag(s_inv_filt,
n_inv_filt.degrade(2**(nstages - i - 1))))
self.pre_op = pre_op
def get_low_res_mgchain(lmax_sky=3000,
lmax_split=2999,
nside=2048,
tol=1e-4,
Nsub=20,
dense_file='',
**kwargs):
#lmax_sky = 3000
#lmax_split = 2999
#nside = 2048
#tol = 1e-4
#Nsub = 20
nside_diag = nside
chain_descr = [[
1, ["diag_cl"], lmax_split, nside_diag, Nsub, 0., cd_solve.tr_cg,
cd_solve.cache_mem()
],
[
0, ["split(stage(1), %s, diag_cl)" % (lmax_split)],
lmax_sky, nside, np.inf, tol, cd_solve.tr_cg,
cd_solve.cache_mem()
]]
return chain_descr
def solve( self, x, tqu ):
self.watch = util.stopwatch()
self.iter_tot = 0
self.prev_eps = None
logger = (lambda iter, eps, stage=0, **kwargs :
self.log(stage, iter, eps, **kwargs))
monitor = cd_monitors.monitor_basic(self.opfilt.dot_op(), logger=logger, iter_max=self.iter_max, eps_min=self.eps_min)
cd_solve.cd_solve( x = x,
b = self.opfilt.calc_prep(tqu, self.s_inv_filt, self.n_inv_filt),
fwd_op = self.opfilt.fwd_op(self.s_inv_filt, self.n_inv_filt),
pre_ops = [self.pre_op], dot_op = self.opfilt.dot_op(),
criterion = monitor, tr=cd_solve.tr_cg, cache=cd_solve.cache_mem() )
return self.opfilt.calc_fini( x, self.s_inv_filt, self.n_inv_filt)
def get_densediagchain(lsides,
lmax_sky,
datshape,
dense_file,
tol=1e-5,
iter_max=np.inf):
assert datshape[0] == datshape[1], datshape
dense_size = 2000
if np.prod(lsides) >= (4. * np.pi) - 0.1:
lmax_dense = 64
else:
lmax_dense = np.sqrt(2. / 2. / np.pi * (2 * np.pi)**2 /
np.prod(lsides) * dense_size)
lmax_dense = int(np.round(min(lmax_dense, 1300)))
print "chain_samples : setting lmax_dense to ", lmax_dense
chain_descr = [[
0,
["split(dense(" + dense_file + "), %s, diag_cl)" % (int(lmax_dense))],
lmax_sky, datshape[0], iter_max, tol, cd_solve.tr_cg,
cd_solve.cache_mem()
]]
return chain_descr
def __init__(self, nstages, opfilt, s_inv_filt, n_inv_filt, plogdepth=0, eps_min=1.e-6, stage_iter_max=3):
""" initialize this multigrid chain by constructing the resolution stages.
* nstages = total number of stages in the chain.
* opfilt = module defining dot_op, fwd_op, calc_prep, calc_fini, and pre_op_diag filter functions.
* s_inv_filt = the S^{-1} filter for the highest resolution stage in the chain.
* n_inv_filt = the N^{-1} filter for the highest resolution stage in the chain.
* plogdepth = the lowest stage at which to print convergence information.
* eps_min = the convergence criterion.
* stage_iter_max = the maximum number of iterations in the substages.
"""
self.opfilt = opfilt
self.s_inv_filt = s_inv_filt
self.n_inv_filt = n_inv_filt
self.plogdepth = plogdepth
self.eps_min = eps_min
self.stage_iter_max = stage_iter_max
self.iter_max = np.inf
pre_op = opfilt.pre_op_diag( s_inv_filt, n_inv_filt.degrade(2**nstages) )
for i in xrange(0, nstages):
class slog(object):
def __init__(self, id, cobj):
self.id = 1*id
self.cobj = cobj
def log(self, iter, eps, **kwargs):
self.cobj.log( self.id, iter, eps, **kwargs )
pre_op = pre_op_split( pre_op_multigrid(opfilt, s_inv_filt, n_inv_filt.degrade(2**(nstages-i)),
[pre_op], slog(nstages-i, self).log, cd_solve.tr_cg,
cd_solve.cache_mem(), stage_iter_max, 0.0 ),
opfilt.pre_op_diag( s_inv_filt, n_inv_filt.degrade(2**(nstages-i-1)) ) )
self.pre_op = pre_op
def get_defaultmgchain(lmax_sky,
lsides,
datshape,
tol=1e-5,
iter_max=np.inf,
dense_file='',
**kwargs):
# FIXME :
assert datshape[0] == datshape[1], datshape
nside_max = datshape[0]
if lmax_sky > 4000:
dense_size = 2000
if np.prod(lsides) >= (4. * np.pi) - 0.1:
lmax_dense = 64
else:
lmax_dense = np.sqrt(2. / 2. / np.pi * (2 * np.pi)**2 /
np.prod(lsides) * dense_size)
lmax_dense = int(np.round(min(lmax_dense, 1300)))
print "chain_samples : setting lmax_dense to ", lmax_dense
chain_descr = [[
3,
[
"split(dense(" + dense_file + "), %s, diag_cl)" %
(int(lmax_dense))
], 1400, nside_max / 4, 3, 0., cd_solve.tr_cg,
cd_solve.cache_mem()
],
[
2, ["split(stage(3), %s, diag_cl)" % (1400)], 3000,
nside_max / 2, 3, 0., cd_solve.tr_cg,
cd_solve.cache_mem()
],
[
1, ["split(stage(2), %s, diag_cl)" % (3000)], 4000,
nside_max / 2, 3, 0., cd_solve.tr_cg,
cd_solve.cache_mem()
],
[
0, ["split(stage(1), %s, diag_cl)" % (4000)],
lmax_sky, nside_max, iter_max, tol, cd_solve.tr_cg,
cd_solve.cache_mem()
]]
elif lmax_sky > 3000:
dense_size = 2000
lmax_dense = np.sqrt(2. / 2. / np.pi * (2 * np.pi)**2 /
np.prod(lsides) * dense_size)
lmax_dense = int(np.round(min(lmax_dense, 1300)))
print "chain_samples : setting lmax_dense to ", lmax_dense
chain_descr = [[
2,
[
"split(dense(" + dense_file + "), %s, diag_cl)" %
(int(lmax_dense))
], 1400, nside_max / 4, 3, 0., cd_solve.tr_cg,
cd_solve.cache_mem()
],
[
1, ["split(stage(2), %s, diag_cl)" % (1400)], 3000,
nside_max / 2, 3, 0., cd_solve.tr_cg,
cd_solve.cache_mem()
],
[
0, ["split(stage(1), %s, diag_cl)" % (3000)],
lmax_sky, nside_max / 2, iter_max, tol,
cd_solve.tr_cg,
cd_solve.cache_mem()
]]
else:
# Same as PL2015 pipeline :
# chain_descr = [[3, ["split(dense(" + pcf + "), 64, diag_cl)"], 256, 128, 3, 0.0, qcinv.cd_solve.tr_cg,
# qcinv.cd_solve.cache_mem()],
# [2, ["split(stage(3), 256, diag_cl)"], 512, 256, 3, 0.0, qcinv.cd_solve.tr_cg,
# qcinv.cd_solve.cache_mem()],
# [1, ["split(stage(2), 512, diag_cl)"], 1024, 512, 3, 0.0, qcinv.cd_solve.tr_cg,
# qcinv.cd_solve.cache_mem()],
# [0, ["split(stage(1), 1024, diag_cl)"], lmax, nside, np.inf, 1.0e-5, qcinv.cd_solve.tr_cg,
# qcinv.cd_solve.cache_mem()]]
# On the full flat sky with lmax_sky 2048 it solves the thing to 1e-5 in 8 min or so on the laptop.
res = lambda fac: max(10, nside_max / fac)
chain_descr = [[
3, ["split(dense(" + dense_file + "), %s, diag_cl)" % (64)], 256,
res(16), 3, 0., cd_solve.tr_cg,
cd_solve.cache_mem()
],
[
2, ["split(stage(3), %s, diag_cl)" % (256)], 512,
res(8), 3, 0., cd_solve.tr_cg,
cd_solve.cache_mem()
],
[
1, ["split(stage(2), %s, diag_cl)" % (512)], 1024,
res(4), 3, 0., cd_solve.tr_cg,
cd_solve.cache_mem()
],
[
0, ["split(stage(1), %s, diag_cl)" % (1024)],
lmax_sky, nside_max, iter_max, tol, cd_solve.tr_cg,
cd_solve.cache_mem()
]]
return chain_descr
