-
Notifications
You must be signed in to change notification settings - Fork 7
/
messenger.py
236 lines (223 loc) · 7.9 KB
/
messenger.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
from __future__ import division, print_function
import numpy as np, os, h5py
from enlib import enmap, pmat, utils, scan, cg, bench, nmat, config, mpi, errors, array_ops
from enact import actdata, filedb, actscan
from pixell import fft
parser = config.ArgumentParser(os.environ["HOME"]+"/.enkirc")
parser.add_argument("area")
parser.add_argument("sel")
parser.add_argument("odir")
parser.add_argument("-n", "--nstep", type=int, default=50)
parser.add_argument("-m", "--method",type=str, default="messenger")
parser.add_argument( "--ndet", type=int, default=None)
parser.add_argument("-p", "--precompute", action="store_true")
parser.add_argument("-o", "--ostep", type=int, default=10)
args = parser.parse_args()
print("Setting fft engine to numpy for now, until I can fix my installation")
fft.engine = "numpy"
print(fft.engine)
utils.mkdir(args.odir)
comm = mpi.COMM_WORLD
dtype = np.float32
ncomp = 3
area = enmap.read_map(args.area)
area = enmap.zeros((ncomp,)+area.shape[-2:],area.wcs,dtype)
Tscale = 0.9
nstep = args.nstep
downsample = config.get("downsample")
filedb.init()
ids = filedb.scans[args.sel]
# Was 1e7
cooldown = sum([[10**j]*5 for j in range(6,0,-1)],[])+[1]
# Read my scans
njunk_tot = 0
cg_rhs = area*0
cg_rjunk = []
if args.precompute:
prec_NNmap = {lam: area*0 for lam in np.unique(cooldown)}
prec_NNjunk = {lam: [] for lam in np.unique(cooldown)}
scans = []
for ind in range(comm.rank, len(ids), comm.size):
id = ids[ind]
entry = filedb.data[id]
try:
scan = actscan.ACTScan(entry)
if scan.ndet == 0 or scan.nsamp == 0:
raise errors.DataMissing("No samples in scan")
if args.ndet:
scan = scan[:args.ndet]
if downsample > 1:
scan = scan[:,::downsample]
scan.pmap = pmat.PmatMap(scan, area)
scan.pcut = pmat.PmatCut(scan)
# Build the noise model
tod = scan.get_samples()
tod -= np.mean(tod,1)[:,None]
tod = tod.astype(dtype)
scan.noise = scan.noise.update(tod, scan.srate)
scan.T = np.min(scan.noise.D)*Tscale
scan.noise_bar = nmat.NmatDetvecs(
scan.noise.D-scan.T, scan.noise.V, scan.noise.E,
scan.noise.bins, scan.noise.ebins, scan.noise.dets)
# Set up cuts
scan.cut_range = [njunk_tot,njunk_tot+scan.pcut.njunk]
njunk_tot += scan.pcut.njunk
# Prepare our filtered data. We do this one of two ways.
# Either store Nb"d for each TOD, which can end up taking
# up a lot of memory, or precompute P'(Nb"+(lT)")"Nbd"d for
# each value of lambda. This saves memory if the maps aren't
# too big and if the number of lambdas is reasonably small.
# For 6 lambdas and deep56 size, we get 240 MB * 6 = 1.4 GB.
# That corresponds to storing 4 downsampled tods.
if args.precompute:
iNbd = scan.noise_bar.apply(tod.copy())
for lam in np.unique(cooldown):
# Could cache this too, but it's fast to compute
iNbt = nmat.NmatDetvecs(
scan.noise_bar.iD + 1/(lam*scan.T), scan.noise_bar.iV,
scan.noise_bar.iE, scan.noise_bar.bins,
scan.noise_bar.ebins, scan.noise_bar.dets)
work = iNbt.apply(iNbd.copy())
work/= scan.T
pjunk= np.zeros(scan.pcut.njunk, dtype)
scan.pcut.backward(work, pjunk)
scan.pmap.backward(work, prec_NNmap[lam])
prec_NNjunk[lam].append(pjunk)
del iNbd
else:
# Compute Nbd, which we need to store
scan.Nbd = scan.noise_bar.apply(tod.copy())
if args.method == "cg":
scan.noise.apply(tod)
tmp = np.zeros(scan.pcut.njunk,dtype)
scan.pcut.backward(tod, tmp)
scan.pmap.backward(tod, cg_rhs)
cg_rjunk.append(tmp)
except errors.DataMissing as e:
print("Skipping %s (%s)" % (id, str(e)))
continue
print("Read %s" % id)
scans.append(scan)
if args.precompute:
for lam in prec_NNjunk:
prec_NNmap[lam] = utils.allreduce(prec_NNmap[lam], comm)
prec_NNjunk[lam] = np.concatenate(prec_NNjunk[lam])
if args.method == "cg":
cg_rhs = utils.allreduce(cg_rhs, comm)
cg_rjunk = np.concatenate(cg_rjunk)
if comm.rank == 0:
enmap.write_map(args.odir + "/map_rhs.fits", cg_rhs)
with h5py.File(args.odir + "/cut_rhs_%02d.hdf" % comm.rank, "w") as hfile:
hfile["data"] = cg_rjunk
# Build div, which we need in both cases
div = enmap.zeros((ncomp,)+area.shape,area.wcs,dtype)
for i in range(ncomp):
work = div[0]*0
work[i] = 1
for scan in scans:
tod = np.zeros((scan.ndet,scan.nsamp),dtype)
scan.pmap.forward(tod, work)
if args.method == "cg":
scan.noise.white(tod)
else: tod /= scan.T
scan.pcut.backward(tod, np.zeros(scan.pcut.njunk,dtype))
scan.pmap.backward(tod, div[i])
div = utils.allreduce(div, comm)
#idiv = utils.eigpow(div,-1,[0,1])
idiv = array_ops.eigpow(div,-1,[0,1], lim=1e-6)
if comm.rank == 0:
enmap.write_map(args.odir + "/map_div.fits", div)
enmap.write_map(args.odir + "/map_idiv.fits", idiv)
del work, div
# And the same for junk
jdiv = np.full(njunk_tot, 1.0, dtype)
for scan in scans:
tod = np.zeros((scan.ndet,scan.nsamp),dtype)
scan.pcut.forward(tod, jdiv[scan.cut_range[0]:scan.cut_range[1]])
if args.method == "cg":
scan.noise.white(tod)
else: tod /= scan.T
scan.pcut.backward(tod, jdiv[scan.cut_range[0]:scan.cut_range[1]])
del tod
if args.method == "cg":
with h5py.File(args.odir + "/cut_div_%02d.hdf" % comm.rank, "w") as hfile:
hfile["data"] = jdiv
if args.method == "messenger":
print(cooldown)
#cooldown = [1e8]*1 + [1e6]*2 + [1e5]*5 + [1e4]*7 + [1e3]*7 + [1e2]*10 + [1e1] * 10
map = area*0
junk = np.zeros(njunk_tot, dtype)
plam = 0
for i in range(nstep):
lam = cooldown[i] if i < len(cooldown) else 1
#lam = max(1,10**(6-i*0.4))
for scan in scans:
# Precompute Nb"+(lT)". Nb" = iD + iV iE iV'. since
# T is diagonal, we can just add it to iD directly.
# Could there be a fourier space unit issue, though?
scan.iNbT = nmat.NmatDetvecs(
scan.noise_bar.iD + 1/(lam*scan.T), scan.noise_bar.iV,
scan.noise_bar.iE, scan.noise_bar.bins,
scan.noise_bar.ebins, scan.noise_bar.dets)
plam = lam
# solve for t. We only use t[si] once, so we don't
# actually need to store it separately like I do here.
rhs = area*0
for si, scan in enumerate(scans):
t = np.zeros([scan.ndet,scan.nsamp],dtype)
scan.pmap.forward(t, (lam*scan.T)**-1*map)
scan.pcut.forward(t, (lam*scan.T)**-1*junk[scan.cut_range[0]:scan.cut_range[1]])
if not args.precompute:
t += scan.Nbd
t = scan.iNbT.apply(t)
t /= scan.T
scan.pcut.backward(t, junk[scan.cut_range[0]:scan.cut_range[1]])
scan.pmap.backward(t, rhs)
rhs = utils.allreduce(rhs, comm)
if args.precompute:
rhs += prec_NNmap[lam]
junk += prec_NNjunk[lam]
junk /= jdiv
map[:] = enmap.map_mul(idiv, rhs)
if comm.rank == 0:
print("%4d %15.7e %8.1f" % (i+1, np.std(map), lam))
if (i+1) % args.ostep == 0:
enmap.write_map(args.odir + "/map%04d.fits" % (i+1), map)
elif args.method == "cg":
def A(x):
map = x[:area.size].reshape(area.shape)
junk = x[area.size:]
omap = map*0
ojunk= junk*0
for scan in scans:
tod = np.zeros([scan.ndet,scan.nsamp],dtype)
scan.pmap.forward(tod, map)
scan.pcut.forward(tod, junk[scan.cut_range[0]:scan.cut_range[1]])
scan.noise.apply(tod)
scan.pcut.backward(tod, ojunk[scan.cut_range[0]:scan.cut_range[1]])
scan.pmap.backward(tod, omap)
del tod
omap = utils.allreduce(omap, comm)
return np.concatenate([omap.reshape(-1),ojunk],0)
def M(x):
map = x[:area.size].reshape(area.shape)
junk = x[area.size:]
omap = map*0
omap[:] = enmap.map_mul(idiv, map)
ojunk= junk/jdiv
return np.concatenate([omap.reshape(-1),ojunk],0)
def dot(x,y):
mprod = np.sum(x[:area.size]*y[:area.size])
jprod = np.sum(x[area.size:]*y[area.size:])
return mprod + comm.allreduce(jprod)
bin = enmap.map_mul(idiv, cg_rhs)
enmap.write_map(args.odir + "/map_bin.fits", bin)
b = np.concatenate([cg_rhs.reshape(-1),cg_rjunk],0)
solver = cg.CG(A, b, M=M, dot=dot)
for i in range(nstep):
solver.step()
if comm.rank == 0:
print("%5d %15.7e" % (solver.i, solver.err))
if solver.i % args.ostep == 0:
map = enmap.samewcs(solver.x[:area.size].reshape(area.shape),area)
enmap.write_map(args.odir + "/map%04d.fits" % solver.i, map)