aqids = args.arrays.split(',')
narrays = len(aqids)
qpairs = []
for qid1 in range(narrays):
for qid2 in range(qid1,narrays):
qpairs.append((aqids[qid1],aqids[qid2]))
"""
We will MPI parallelize over pairs of arrays. This wastes FFTs, but is more memory efficient. (Each job only
holds 2 arrays in memory).
"""
njobs = len(qpairs)
comm,rank,my_tasks = mpi.distribute(njobs)
mask = sints.get_act_mr3_crosslinked_mask(args.region,
version=args.mask_version,
kind='binary_apod')
shape,wcs = mask.shape,mask.wcs
modlmap = mask.modlmap()
aspecs = tutils.ASpecs().get_specs
region = args.region
fbeam = lambda qname,x: tutils.get_kbeam(qname,x,sanitize=not(args.unsanitized_beam),planck_pixwin=True)
nbin_edges = np.arange(20,8000,100)
nbinner = stats.bin2D(modlmap,nbin_edges)
ncents = nbinner.centers
modlmap = enmap.modlmap(shape,wcs)
ngen = {}
ngen['act_mr3'] = actnoise.NoiseGen(args.sim_version,model="act_mr3",extract_region=mask,ncache=0,verbose=True)
ngen['planck_hybrid'] = actnoise.NoiseGen(args.sim_version,model="planck_hybrid",extract_region=mask,ncache=0,verbose=True)
arrays = args.arrays.split(',')
narrays = len(arrays)
nsims = args.nsims
jsim = pipeline.JointSim(arrays,args.fg_res_version+"_"+args.region,
bandpassed=bandpasses,no_act_color_correction=args.no_act_color_correction,
ccor_exp=args.ccor_exp)
comm,rank,my_tasks = mpi.distribute(nsims)
for task in my_tasks:
sim_index = task + args.start_index
print("Rank %d starting task %d at %s..." % (rank,task,str(datetime.now())))
ind_str = str(set_id).zfill(2)+"_"+str(sim_index).zfill(4)
sim_version = "%s_%s" % (args.version,ind_str)
scratch = tutils.get_scratch_path(sim_version,args.region)
try:
os.makedirs(scratch)
except:
pass
"""
MAKE SIMS
def do(ymap, cmap, dmap, mask, ras, decs, wt):
combined = list(zip(ras, decs))
random.shuffle(combined)
ras[:], decs[:] = zip(*combined)
Nrand = 400
njobs = len(ras)
comm, rank, my_tasks = mpi.distribute(njobs)
print("Rank %d starting" % rank)
s = stats.Stats(comm)
i = 0
for task in my_tasks:
ra = ras[task]
dec = decs[task]
mcut, ycut, ccut, dcut, weight = get_cuts(mask, ymap, cmap, dmap, wt,
ra, dec, arcmin, pix)
if mcut is None: continue
if i == 0:
modrmap = np.rad2deg(ycut.modrmap()) * 60.
bin_edges = np.arange(0., 15., 1.0)
binner = stats.bin2D(modrmap, bin_edges)
rras, rdecs = catalogs.random_catalog(ymap.shape,
ymap.wcs,
Nrand,
edge_avoid_deg=4.)
nrej = 0
for rra, rdec in zip(rras, rdecs):
rmcut, rycut, rccut, rdcut, rweight = get_cuts(
mask, ymap, cmap, dmap, wt, rra, rdec, arcmin, pix)
if rmcut is None:
nrej = nrej + 1
continue
cents, ry1d = binner.bin(rycut)
cents, rc1d = binner.bin(rccut)
cents, rd1d = binner.bin(rdcut)
s.add_to_stats("rc1d", rc1d * 1e6)
s.add_to_stats("ry1d", ry1d * 1e6)
s.add_to_stats("rd1d", rd1d * 1e6)
if rank == 0: print(Nrand - nrej, " accepted")
cents, y1d = binner.bin(ycut)
cents, c1d = binner.bin(ccut)
cents, d1d = binner.bin(dcut)
s.add_to_stats("c1d", c1d * 1e6)
s.add_to_stats("y1d", y1d * 1e6)
s.add_to_stats("d1d", d1d * 1e6)
s.add_to_stack("cstack", ccut * 1e6 * weight)
s.add_to_stack("dstack", dcut * 1e6 * weight)
s.add_to_stack("ystack", ycut * 1e6 * weight)
s.add_to_stats("sum", (weight, ))
i = i + 1
if i % 10 == 0 and rank == 0: print(i)
print("Rank %d done " % rank)
s.get_stats()
s.get_stacks()
if rank == 0:
N = s.vectors['sum'].sum()
ystack = s.stacks['ystack'] * N
cstack = s.stacks['cstack'] * N
dstack = s.stacks['dstack'] * N
y1ds = s.vectors['y1d']
c1ds = s.vectors['c1d']
d1ds = s.vectors['d1d']
ry1d = s.stats['ry1d']['mean']
rc1d = s.stats['rc1d']['mean']
rd1d = s.stats['rd1d']['mean']
_, nwcs = enmap.geometry(pos=(0, 0),
shape=ystack.shape,
res=np.deg2rad(0.5 / 60.))
return rank, enmap.enmap(
ystack, nwcs), enmap.enmap(cstack, nwcs), enmap.enmap(
dstack, nwcs), N, cents, y1ds, c1ds, d1ds, ry1d, rc1d, rd1d
else:
return rank, None, None, None, None, None, None, None, None, None, None, None
self.mlmax)
self.cache[key] = alms
return alms
def get_kmap(self, X, seed):
s_i, s_set, _ = solenspipe.convert_seeds(seed)
key = (s_i, s_set)
if key in self.cache:
alms = self.cache[key]
else:
alms = self.load_alms(s_i, s_set)
return alms
nsims = 40
comm, _, _ = mpi.distribute(nsims, verbose=True)
power = lambda x, y: hp.alm2cl(x, y)
mcache = MapCacher(mlmax)
get_kmap = mcache.get_kmap
cl_n1 = bias.mcn1(0,
polcomb,
polcomb,
qfunc,
get_kmap,
comm,
power,
nsims,
verbose=True)
ls = np.arange(len(cl_n1))
cents, cn1 = binner.bin(ls, cl_n1)
def mcrdn0(icov,
get_kmap,
power,
nsims,
qfunc1,
qfunc2=None,
Xdat=None,
use_mpi=True,
verbose=True,
skip_rd=False):
"""
Using Monte Carlo sims, this function calculates
the anisotropic Gaussian N0 bias
between two quadratic estimators. It returns both the
data realization-dependent version (RDN0) and the pure-simulation
version (MCN0).
e.g.
>> mcrdn0(0,qfunc,get_kmap,comm,power)
Parameters
----------
icov: int
The index of the realization passed to get_kmap if performing
a covariance calculation - otherwise, set to zero.
get_kmap: function
Function for getting the filtered a_lms of data and simulation
maps. See notes at top of module.
power: function
Returns C(l) from two maps x,y, as power(x,y).
nsims: int
Number of sims
qfunc1: function
Function for reconstructing lensing from maps x and y,
called as e.g. qfunc(x, y). See e.g. SoLensPipe.qfunc.
The x and y arguments accept a [T_alm,E_alm,B_alm] tuple.
The function should return an (N,...) array where N is typically
two components for the gradient and curl.
qfunc2: function, optional
Same as above, for the third and fourth legs of the 4-point
RDN0.
comm: object, optional
MPI communicator
verbose: bool, optional
Whether to show progress
skip_rd: bool, optional
Whether to skip the RDN0 terms. The first returned component
is then None.
Returns
-------
rdn0: (N*(N+1)/2,...) array
Estimate of the RDN0 bias. If N=2 for gradient and curl,
the three components correspond to the gradient RDN0, the
curl RDN0 and the gradient x curl RDN0. None is returned
if skip_rd is True.
mcn0: (N*(N+1)/2,...) array
Estimate of the MCN0 bias. If N=2 for gradient and curl,
the three components correspond to the gradient RDN0, the
curl RDN0 and the gradient x curl RDN0.
"""
qa = qfunc1
qb = qfunc2
mcn0evals = []
if not (skip_rd):
assert Xdat is not None # Data
rdn0evals = []
if use_mpi:
comm, rank, my_tasks = mpi.distribute(nsims)
else:
comm, rank, my_tasks = FakeCommunicator(), 0, range(nsims)
for i in my_tasks:
if rank == 0 and verbose:
print("MCRDN0: Rank %d doing task %d" % (rank, i))
Xs = get_kmap((icov, 0, i))
if not (skip_rd):
qaXXs = qa(Xdat, Xs)
qbXXs = qb(Xdat, Xs) if qb is not None else qaXXs
qaXsX = qa(Xs, Xdat)
qbXsX = qb(Xs, Xdat) if qb is not None else qaXsX
rdn0_only_term = power(qaXXs,qbXXs) + power(qaXsX,qbXXs) \
+ power(qaXsX,qbXsX) + power(qaXXs,qbXsX)
Xsp = get_kmap((icov, 1, i))
qaXsXsp = qa(Xs, Xsp)
qbXsXsp = qb(Xs, Xsp) if qb is not None else qaXsXsp
qbXspXs = qb(Xsp, Xs) if qb is not None else qa(Xsp, Xs)
mcn0_term = (power(qaXsXsp, qbXsXsp) + power(qaXsXsp, qbXspXs))
mcn0evals.append(mcn0_term.copy())
if not (skip_rd): rdn0evals.append(rdn0_only_term - mcn0_term)
if not (skip_rd):
avgrdn0 = utils.allgatherv(rdn0evals, comm)
else:
avgrdn0 = None
avgmcn0 = utils.allgatherv(mcn0evals, comm)
return avgrdn0, avgmcn0
def mcn1(icov,
get_kmap,
power,
nsims,
qfunc1,
qfunc2=None,
comm=None,
verbose=True):
"""
MCN1 for alpha=XY cross beta=AB
qfunc(x,y) returns QE reconstruction minus mean-field in fourier space
Parameters
----------
icov: int
The index of the realization passed to get_kmap if performing
a covariance calculation - otherwise, set to zero.
get_kmap: function
Function for getting the filtered a_lms of data and simulation
maps. See notes at top of module.
power: function
Returns C(l) from two maps x,y, as power(x,y).
nsims: int
Number of sims
qfunc1: function
Function for reconstructing lensing from maps x and y,
called as e.g. qfunc(x, y). See e.g. SoLensPipe.qfunc.
The x and y arguments accept a [T_alm,E_alm,B_alm] tuple.
The function should return an (N,...) array where N is typically
two components for the gradient and curl.
qfunc2: function, optional
Same as above, for the third and fourth legs of the 4-point
MCN1.
comm: object, optional
MPI communicator
verbose: bool, optional
Whether to show progress
Returns
-------
mcn1: (N*(N+1)/2,...) array
Estimate of the MCN1 bias. If N=2 for gradient and curl,
the three components correspond to the gradient MCN1, the
curl MCN1 and the gradient x curl MCN1.
Estimate of the MCN1 bias
"""
qa = qfunc1
qb = qfunc2
comm, rank, my_tasks = mpi.distribute(nsims)
n1evals = []
for i in my_tasks:
if rank == 0 and verbose:
print("MCN1: Rank %d doing task %d" % (comm.rank, i))
Xs = get_kmap((icov, 0, i)) # S
Ysp = get_kmap((icov, 1, i)) # S'
Xsk = get_kmap((icov, 2, i)) # Sphi
Yskp = get_kmap((icov, 3, i)) # Sphi'
qa_Xsk_Yskp = qa(Xsk, Yskp)
qb_Xsk_Yskp = qb(Xsk, Yskp) if qb is not None else qa_Xsk_Yskp
qb_Yskp_Xsk = qb(Yskp, Xsk) if qb is not None else qa(Yskp, Xsk)
qa_Xs_Ysp = qa(Xs, Ysp)
qb_Xs_Ysp = qb(Xs, Ysp) if qb is not None else qa_Xs_Ysp
qb_Ysp_Xs = qb(Ysp, Xs) if qb is not None else qa(Ysp, Xs)
qb_Yskp_Xsk = qb(Yskp, Xsk) if qb is not None else qa(Yskp, Xsk)
term = power(qa_Xsk_Yskp,qb_Xsk_Yskp) + power(qa_Xsk_Yskp,qb_Yskp_Xsk) \
- power(qa_Xs_Ysp,qb_Xs_Ysp) - power(qa_Xs_Ysp,qb_Ysp_Xs)
n1evals.append(term.copy())
n1s = utils.allgatherv(n1evals, comm)
return n1s
def do(ymap, cmap, mask, ras, decs, wt):
combined = list(zip(ras, decs))
random.shuffle(combined)
ras[:], decs[:] = zip(*combined)
njobs = len(ras)
comm, rank, my_tasks = mpi.distribute(njobs)
print("Rank %d starting" % rank)
s = stats.Stats(comm)
i = 0
for task in my_tasks:
ra = ras[task]
dec = decs[task]
# mcut = reproject.cutout(mask, ra=np.deg2rad(ra), dec=np.deg2rad(dec),npix=int(arcmin/pix))
mcut = reproject.postage_stamp(mask, np.deg2rad(ra), np.deg2rad(dec),
arcmin, pix)
if mcut is None:
continue
if np.any(mcut) <= 0:
continue
# ycut = reproject.cutout(ymap, ra=np.deg2rad(ra), dec=np.deg2rad(dec),npix=int(arcmin/pix))
# ccut = reproject.cutout(cmap, ra=np.deg2rad(ra), dec=np.deg2rad(dec),npix=int(arcmin/pix))
# wcut = reproject.cutout(wt, ra=np.deg2rad(ra), dec=np.deg2rad(dec),npix=int(arcmin/pix))
ycut = reproject.postage_stamp(ymap, np.deg2rad(ra), np.deg2rad(dec),
arcmin, pix)
ccut = reproject.postage_stamp(cmap, np.deg2rad(ra), np.deg2rad(dec),
arcmin, pix)
wcut = reproject.postage_stamp(wt, np.deg2rad(ra), np.deg2rad(dec),
arcmin, pix)
weight = wcut.mean()
# if i==0:
# modrmap = np.rad2deg(ycut.modrmap())*60.
# bin_edges = np.arange(0.,15.,1.0)
# binner = stats.bin2D(modrmap,bin_edges)
# cents,y1d = binner.bin(ycut)
# cents,c1d = binner.bin(ccut)
# s.add_to_stats("c1d",c1d*1e6)
# s.add_to_stats("y1d",y1d*1e6)
s.add_to_stack("cstack", ccut * 1e6 * weight)
s.add_to_stack("ystack", ycut * 1e6 * weight)
s.add_to_stats("sum", (weight, ))
i = i + 1
if i % 10 == 0 and rank == 0: print(i)
print("Rank %d done " % rank)
s.get_stats()
s.get_stacks()
if rank == 0:
N = s.vectors['sum'].sum()
ystack = s.stacks['ystack'] * N
cstack = s.stacks['cstack'] * N
_, nwcs = enmap.geometry(pos=(0, 0),
shape=ystack.shape,
res=np.deg2rad(0.5 / 60.))
return rank, enmap.enmap(ystack, nwcs), enmap.enmap(cstack, nwcs), N
else:
return rank, None, None, None
#qids = 'boss_03,boss_04'.split(',')
#qids = 'd56_01,d56_02,d56_03,d56_04,d56_05,d56_06'.split(',')
#qids = 'd56_02,d56_03,d56_04'.split(',')
#qids = 'd56_04,p01,p02,p03,p04,p05,p06,p07,p08'.split(',')
mask = sints.get_act_mr3_crosslinked_mask(region,
kind='binary_apod')
modlmap = mask.modlmap()
jsim = pipeline.JointSim(qids,fg_res_version+"_"+region,bandpassed=True)
jsim0 = pipeline.JointSim(qids,None,bandpassed=True)
comm,rank,my_tasks = mpi.distribute(len(qids))
bin_edges = np.arange(20,8000,20)
binner = stats.bin2D(modlmap,bin_edges)
jsim.update_signal_index(0,set_idx=0)
jsim0.update_signal_index(0,set_idx=0)
for task in my_tasks:
qid = qids[task]
with bench.show("signal"):
signal = jsim.compute_map(mask.shape,mask.wcs,qid,
include_cmb=True,include_tsz=True,
include_fgres=True,sht_beam=True) # !!!
lmin = args.lmin
lmax = args.lmax
polcomb = args.polcomb
config = io.config_from_yaml("../input/config.yml")
"""
nside=2048
smooth_deg=4.
ch=SOChannel('LA',145)
polcomb='TT'
lmin=100
lmax=3000
config = io.config_from_yaml("../input/config.yml")
"""
comm,rank,my_tasks = mpi.distribute(args.nsims,verbose=True) #tasks are the number of simulations
mask = initialize_mask(nside,smooth_deg)
solint = SOLensInterface(mask)
thloc = "../data/" + config['theory_root']
theory = cosmology.loadTheorySpectraFromCAMB(thloc,get_dimensionless=False)
# norm dict
Als = {}
with bench.show("norm"):
ls,Als['TT'],Als['EE'],Als['EB'],Als['TE'],Als['TB'],al_mv_pol,al_mv,Al_te_hdv = initialize_norm(solint,ch,lmin,lmax)
Als['mv'] = al_mv
Als['mvpol'] = al_mv_pol
al_mv = Als[polcomb]