1
|
|
|
# -*- coding: utf-8 -*- |
2
|
|
|
# vim:fileencoding=utf-8 |
3
|
|
|
# |
4
|
|
|
# Copyright (c) 2017-2018 Stefan Bender |
5
|
|
|
# |
6
|
|
|
# This module is part of sciapy. |
7
|
|
|
# sciapy is free software: you can redistribute it or modify |
8
|
|
|
# it under the terms of the GNU General Public License as published |
9
|
|
|
# by the Free Software Foundation, version 2. |
10
|
|
|
# See accompanying LICENSE file or http://www.gnu.org/licenses/gpl-2.0.html. |
11
|
1 |
|
"""SCIAMACHY data regression command line interface |
12
|
|
|
|
13
|
|
|
Command line main program for regression analysis of SCIAMACHY |
14
|
|
|
daily zonal mean time series (NO for now). |
15
|
|
|
""" |
16
|
|
|
|
17
|
1 |
|
import ctypes |
18
|
1 |
|
import logging |
19
|
|
|
|
20
|
1 |
|
import numpy as np |
21
|
1 |
|
import scipy.optimize as op |
22
|
1 |
|
from scipy.interpolate import interp1d |
23
|
|
|
|
24
|
1 |
|
import george |
25
|
1 |
|
import celerite |
26
|
|
|
|
27
|
1 |
|
import matplotlib as mpl |
28
|
|
|
# switch off X11 rendering |
29
|
1 |
|
mpl.use("Agg") |
30
|
|
|
|
31
|
1 |
|
from .load_data import load_solar_gm_table, load_scia_dzm |
32
|
1 |
|
from .models_cel import CeleriteModelSet as NOModel |
33
|
1 |
|
from .models_cel import ConstantModel, ProxyModel |
34
|
1 |
|
from .models_cel import HarmonicModelCosineSine, HarmonicModelAmpPhase |
35
|
1 |
|
from .mcmc import mcmc_sample_model |
36
|
1 |
|
from .statistics import mcmc_statistics |
37
|
|
|
|
38
|
1 |
|
from ._gpkernels import (george_solvers, |
39
|
|
|
setup_george_kernel, setup_celerite_terms) |
40
|
1 |
|
from ._plot import (plot_single_sample_and_residuals, |
41
|
|
|
plot_residual, plot_random_samples) |
42
|
1 |
|
from ._options import parser |
43
|
|
|
|
44
|
|
|
|
45
|
1 |
View Code Duplication |
def save_samples_netcdf(filename, model, alt, lat, samples, |
|
|
|
|
46
|
|
|
scale=1e-6, |
47
|
|
|
lnpost=None, compressed=False): |
48
|
|
|
from xarray import Dataset |
49
|
|
|
smpl_ds = Dataset(dict([(pname, (["lat", "alt", "sample"], |
50
|
|
|
samples[..., i].reshape(1, 1, -1))) |
51
|
|
|
for i, pname in enumerate(model.get_parameter_names())] |
52
|
|
|
# + [("lpost", (["lat", "alt", "sample"], lnp.reshape(1, 1, -1)))] |
53
|
|
|
), |
54
|
|
|
coords={"lat": [lat], "alt": [alt]}) |
55
|
|
|
|
56
|
|
|
for modn in model.mean.models: |
57
|
|
|
modl = model.mean.models[modn] |
58
|
|
|
if hasattr(modl, "mean"): |
59
|
|
|
smpl_ds.attrs[modn + ":mean"] = modl.mean |
60
|
|
|
|
61
|
|
|
units = {"kernel": { |
62
|
|
|
"log": "log(10$^{{{0:.0f}}}$ cm$^{{-3}}$)" |
63
|
|
|
.format(-np.log10(scale))}, |
64
|
|
|
"mean": { |
65
|
|
|
"log": "log(10$^{{{0:.0f}}}$ cm$^{{-3}}$)" |
66
|
|
|
.format(-np.log10(scale)), |
67
|
|
|
"val": "10$^{{{0:.0f}}}$ cm$^{{-3}}$".format(-np.log10(scale)), |
68
|
|
|
"amp": "10$^{{{0:.0f}}}$ cm$^{{-3}}$".format(-np.log10(scale)), |
69
|
|
|
"tau": "d"}} |
70
|
|
|
for pname in smpl_ds.data_vars: |
71
|
|
|
_pp = pname.split(':') |
72
|
|
|
for _n, _u in units[_pp[0]].items(): |
73
|
|
|
if _pp[-1].startswith(_n): |
74
|
|
|
logging.debug("units for %s: %s", pname, _u) |
75
|
|
|
smpl_ds[pname].attrs["units"] = _u |
76
|
|
|
|
77
|
|
|
smpl_ds["alt"].attrs = {"long_name": "altitude", "units": "km"} |
78
|
|
|
smpl_ds["lat"].attrs = {"long_name": "latitude", "units": "degrees_north"} |
79
|
|
|
|
80
|
|
|
_encoding = None |
81
|
|
|
if compressed: |
82
|
|
|
_encoding = {var: {"zlib": True, "complevel": 1} |
83
|
|
|
for var in smpl_ds.data_vars} |
84
|
|
|
smpl_ds.to_netcdf(filename, encoding=_encoding) |
85
|
|
|
smpl_ds.close() |
86
|
|
|
|
87
|
|
|
|
88
|
1 |
View Code Duplication |
def _train_test_split(times, data, errs, train_frac, |
|
|
|
|
89
|
|
|
test_frac, randomize): |
90
|
|
|
# split the data into training and test subsets according to the |
91
|
|
|
# fraction given (default is 1, i.e. no splitting) |
92
|
|
|
ndata = len(times) |
93
|
|
|
train_size = int(ndata * train_frac) |
94
|
|
|
test_size = min(ndata - train_size, int(ndata * test_frac)) |
95
|
|
|
# randomize if requested |
96
|
|
|
if randomize: |
97
|
|
|
permut_idx = np.random.permutation(np.arange(ndata)) |
98
|
|
|
else: |
99
|
|
|
permut_idx = np.arange(ndata) |
100
|
|
|
train_idx = np.sort(permut_idx[:train_size]) |
101
|
|
|
test_idx = np.sort(permut_idx[train_size:train_size + test_size]) |
102
|
|
|
times_train = times[train_idx] |
103
|
|
|
data_train = data[train_idx] |
104
|
|
|
errs_train = errs[train_idx] |
105
|
|
|
if test_size > 0: |
106
|
|
|
times_test = times[test_idx] |
107
|
|
|
data_test = data[test_idx] |
108
|
|
|
errs_test = errs[test_idx] |
109
|
|
|
else: |
110
|
|
|
times_test = times |
111
|
|
|
data_test = data |
112
|
|
|
errs_test = errs |
113
|
|
|
logging.info("using %s of %s samples for training.", len(times_train), ndata) |
114
|
|
|
logging.info("using %s of %s samples for testing.", len(times_test), ndata) |
115
|
|
|
return (times_train, data_train, errs_train, |
116
|
|
|
times_test, data_test, errs_test) |
117
|
|
|
|
118
|
|
|
|
119
|
1 |
View Code Duplication |
def _r_sun_earth(time, tfmt="jyear"): |
|
|
|
|
120
|
|
|
"""First order approximation of the Sun-Earth distance |
121
|
|
|
|
122
|
|
|
The Sun-to-Earth distance can be used to (un-)normalize proxies |
123
|
|
|
to the actual distance to the Sun instead of 1 AU. |
124
|
|
|
|
125
|
|
|
Parameters |
126
|
|
|
---------- |
127
|
|
|
time : float |
128
|
|
|
Time value in the units given by 'tfmt'. |
129
|
|
|
tfmt : str, optional |
130
|
|
|
The units of 'time' as supported by the |
131
|
|
|
astropy.time time formats. Default: 'jyear'. |
132
|
|
|
|
133
|
|
|
Returns |
134
|
|
|
------- |
135
|
|
|
dist : float |
136
|
|
|
The Sun-Earth distance at the given day of year in AU. |
137
|
|
|
""" |
138
|
|
|
from astropy.time import Time |
139
|
|
|
tdoy = Time(time, format=tfmt) |
140
|
|
|
tdoy.format = "yday" |
141
|
|
|
doy = int(tdoy.value.split(':')[1]) |
142
|
|
|
return 1 - 0.01672 * np.cos(2 * np.pi / 365.256363 * (doy - 4)) |
143
|
|
|
|
144
|
|
|
|
145
|
1 |
View Code Duplication |
def main(): |
|
|
|
|
146
|
1 |
|
logging.basicConfig(level=logging.WARNING, |
147
|
|
|
format="[%(levelname)-8s] (%(asctime)s) " |
148
|
|
|
"%(filename)s:%(lineno)d %(message)s", |
149
|
|
|
datefmt="%Y-%m-%d %H:%M:%S %z") |
150
|
|
|
|
151
|
1 |
|
args = parser.parse_args() |
152
|
|
|
|
153
|
|
|
logging.info("command line arguments: %s", args) |
154
|
|
|
if args.quiet: |
155
|
|
|
logging.getLogger().setLevel(logging.ERROR) |
156
|
|
|
elif args.verbose: |
157
|
|
|
logging.getLogger().setLevel(logging.INFO) |
158
|
|
|
else: |
159
|
|
|
logging.getLogger().setLevel(args.loglevel) |
160
|
|
|
|
161
|
|
|
from numpy.distutils.system_info import get_info |
162
|
|
|
for oblas_path in get_info("openblas")["library_dirs"]: |
163
|
|
|
oblas_name = "{0}/libopenblas.so".format(oblas_path) |
164
|
|
|
logging.info("Trying %s", oblas_name) |
165
|
|
|
try: |
166
|
|
|
oblas_lib = ctypes.cdll.LoadLibrary(oblas_name) |
167
|
|
|
oblas_cores = oblas_lib.openblas_get_num_threads() |
168
|
|
|
oblas_lib.openblas_set_num_threads(args.openblas_threads) |
169
|
|
|
logging.info("Using %s/%s Openblas thread(s).", |
170
|
|
|
oblas_lib.openblas_get_num_threads(), oblas_cores) |
171
|
|
|
except: |
172
|
|
|
logging.info("Setting number of openblas threads failed.") |
173
|
|
|
|
174
|
|
|
if args.random_seed is not None: |
175
|
|
|
np.random.seed(args.random_seed) |
176
|
|
|
|
177
|
|
|
if args.proxies: |
178
|
|
|
proxies = args.proxies.split(',') |
179
|
|
|
proxy_dict = dict(_p.split(':') for _p in proxies) |
180
|
|
|
else: |
181
|
|
|
proxy_dict = {} |
182
|
|
|
lag_dict = {pn: 0 for pn in proxy_dict.keys()} |
183
|
|
|
|
184
|
|
|
# Post-processing of arguments... |
185
|
|
|
# List of proxy lag fits from csv |
186
|
|
|
fit_lags = args.fit_lags.split(',') |
187
|
|
|
# List of proxy lifetime fits from csv |
188
|
|
|
fit_lifetimes = args.fit_lifetimes.split(',') |
189
|
|
|
fit_annlifetimes = args.fit_annlifetimes.split(',') |
190
|
|
|
# List of proxy lag times from csv |
191
|
|
|
lag_dict.update(dict(_ls.split(':') for _ls in args.lag_times.split(','))) |
192
|
|
|
# List of cycles (frequencies in 1/year) from argument list (csv) |
193
|
|
|
try: |
194
|
|
|
freqs = list(map(float, args.freqs.split(','))) |
195
|
|
|
except ValueError: |
196
|
|
|
freqs = [] |
197
|
|
|
# List of initial parameter values |
198
|
|
|
initial = None |
199
|
|
|
if args.initial is not None: |
200
|
|
|
try: |
201
|
|
|
initial = list(map(float, args.initial.split(','))) |
202
|
|
|
except ValueError: |
203
|
|
|
pass |
204
|
|
|
# List of GP kernels from argument list (csv) |
205
|
|
|
kernls = args.kernels.split(',') |
206
|
|
|
|
207
|
|
|
lat = args.latitude |
208
|
|
|
alt = args.altitude |
209
|
|
|
logging.info("location: %.0f°N %.0f km", lat, alt) |
210
|
|
|
|
211
|
|
|
no_ys, no_dens, no_errs, no_szas = load_scia_dzm(args.file, alt, lat, |
212
|
|
|
tfmt=args.time_format, |
213
|
|
|
scale=args.scale, |
214
|
|
|
#subsample_factor=args.random_subsample, |
215
|
|
|
#subsample_method="random", |
216
|
|
|
akd_threshold=args.akd_threshold, |
217
|
|
|
cnt_threshold=args.cnt_threshold, |
218
|
|
|
center=args.center_data, |
219
|
|
|
season=args.season, |
220
|
|
|
SPEs=args.exclude_spe) |
221
|
|
|
|
222
|
|
|
(no_ys_train, no_dens_train, no_errs_train, |
223
|
|
|
no_ys_test, no_dens_test, no_errs_test) = _train_test_split( |
224
|
|
|
no_ys, no_dens, no_errs, args.train_fraction, |
225
|
|
|
args.test_fraction, args.random_train_test) |
226
|
|
|
|
227
|
|
|
sza_intp = interp1d(no_ys, no_szas, fill_value="extrapolate") |
228
|
|
|
|
229
|
|
|
max_amp = 1e10 * args.scale |
230
|
|
|
max_days = 100 |
231
|
|
|
|
232
|
|
|
harmonic_models = [] |
233
|
|
|
for freq in freqs: |
234
|
|
|
if not args.fit_phase: |
235
|
|
|
harmonic_models.append(("f{0:.0f}".format(freq), |
236
|
|
|
HarmonicModelCosineSine(freq=freq, |
237
|
|
|
cos=0, sin=0, |
238
|
|
|
bounds=dict([ |
239
|
|
|
("cos", [-max_amp, max_amp]), |
240
|
|
|
("sin", [-max_amp, max_amp])]) |
241
|
|
|
))) |
242
|
|
|
else: |
243
|
|
|
harmonic_models.append(("f{0:.0f}".format(freq), |
244
|
|
|
HarmonicModelAmpPhase(freq=freq, |
245
|
|
|
amp=0, phase=0, |
246
|
|
|
bounds=dict([ |
247
|
|
|
# ("amp", [-max_amp, max_amp]), |
248
|
|
|
("amp", [0, max_amp]), |
249
|
|
|
("phase", [-np.pi, np.pi])]) |
250
|
|
|
))) |
251
|
|
|
proxy_models = [] |
252
|
|
|
for pn, pf in proxy_dict.items(): |
253
|
|
|
pt, pp = load_solar_gm_table(pf, cols=[0, 1], names=["time", pn], tfmt=args.time_format) |
254
|
|
|
pv = np.log(pp[pn]) if pn in args.log_proxies.split(',') else pp[pn] |
255
|
|
|
if pn in args.norm_proxies_distSEsq: |
256
|
|
|
rad_sun_earth = np.vectorize(_r_sun_earth)(pt, tfmt=args.time_format) |
257
|
|
|
pv /= rad_sun_earth**2 |
258
|
|
|
if pn in args.norm_proxies_SZA: |
259
|
|
|
pv *= np.cos(np.radians(sza_intp(pt))) |
260
|
|
|
proxy_models.append((pn, |
261
|
|
|
ProxyModel(pt, pv, |
262
|
|
|
center=pn in args.center_proxies.split(','), |
263
|
|
|
sza_intp=sza_intp if args.use_sza else None, |
264
|
|
|
fit_phase=args.fit_phase, |
265
|
|
|
lifetime_prior=args.lifetime_prior, |
266
|
|
|
lifetime_metric=args.lifetime_metric, |
267
|
|
|
days_per_time_unit=1 if args.time_format.endswith("d") else 365.25, |
268
|
|
|
amp=0., |
269
|
|
|
lag=float(lag_dict[pn]), |
270
|
|
|
tau0=0, |
271
|
|
|
taucos1=0, tausin1=0, |
272
|
|
|
taucos2=0, tausin2=0, |
273
|
|
|
ltscan=args.lifetime_scan, |
274
|
|
|
bounds=dict([ |
275
|
|
|
("amp", |
276
|
|
|
[0, max_amp] if pn in args.positive_proxies.split(',') |
277
|
|
|
else [-max_amp, max_amp]), |
278
|
|
|
("lag", [0, max_days]), |
279
|
|
|
("tau0", [0, max_days]), |
280
|
|
|
("taucos1", [0, max_days] if args.fit_phase else [-max_days, max_days]), |
281
|
|
|
("tausin1", [-np.pi, np.pi] if args.fit_phase else [-max_days, max_days]), |
282
|
|
|
# semi-annual cycles for the life time |
283
|
|
|
("taucos2", [0, max_days] if args.fit_phase else [-max_days, max_days]), |
284
|
|
|
("tausin2", [-np.pi, np.pi] if args.fit_phase else [-max_days, max_days]), |
285
|
|
|
("ltscan", [0, 200])]) |
286
|
|
|
))) |
287
|
|
|
logging.info("%s mean: %s", pn, proxy_models[-1][1].mean) |
288
|
|
|
offset_model = [("offset", |
289
|
|
|
ConstantModel(value=0., |
290
|
|
|
bounds={"value": [-max_amp, max_amp]}))] |
291
|
|
|
|
292
|
|
|
model = NOModel(offset_model + harmonic_models + proxy_models) |
293
|
|
|
|
294
|
|
|
logging.debug("model dict: %s", model.get_parameter_dict()) |
295
|
|
|
model.freeze_all_parameters() |
296
|
|
|
# thaw parameters according to requested fits |
297
|
|
|
for pn in proxy_dict.keys(): |
298
|
|
|
model.thaw_parameter("{0}:amp".format(pn)) |
299
|
|
|
if pn in fit_lags: |
300
|
|
|
model.thaw_parameter("{0}:lag".format(pn)) |
301
|
|
|
if pn in fit_lifetimes: |
302
|
|
|
model.set_parameter("{0}:tau0".format(pn), 1e-3) |
303
|
|
|
model.thaw_parameter("{0}:tau0".format(pn)) |
304
|
|
|
if pn in fit_annlifetimes: |
305
|
|
|
model.thaw_parameter("{0}:taucos1".format(pn)) |
306
|
|
|
model.thaw_parameter("{0}:tausin1".format(pn)) |
307
|
|
|
for freq in freqs: |
308
|
|
|
if not args.fit_phase: |
309
|
|
|
model.thaw_parameter("f{0:.0f}:cos".format(freq)) |
310
|
|
|
model.thaw_parameter("f{0:.0f}:sin".format(freq)) |
311
|
|
|
else: |
312
|
|
|
model.thaw_parameter("f{0:.0f}:amp".format(freq)) |
313
|
|
|
model.thaw_parameter("f{0:.0f}:phase".format(freq)) |
314
|
|
|
if args.fit_offset: |
315
|
|
|
#model.set_parameter("offset:value", -100.) |
316
|
|
|
#model.set_parameter("offset:value", 0) |
317
|
|
|
model.thaw_parameter("offset:value") |
318
|
|
|
|
319
|
|
|
if initial is not None: |
320
|
|
|
model.set_parameter_vector(initial) |
321
|
|
|
# model.thaw_parameter("GM:ltscan") |
322
|
|
|
logging.debug("params: %s", model.get_parameter_dict()) |
323
|
|
|
logging.debug("param names: %s", model.get_parameter_names()) |
324
|
|
|
logging.debug("param vector: %s", model.get_parameter_vector()) |
325
|
|
|
logging.debug("param bounds: %s", model.get_parameter_bounds()) |
326
|
|
|
#logging.debug("model value: %s", model.get_value(no_ys)) |
327
|
|
|
#logging.debug("default log likelihood: %s", model.log_likelihood(model.vector)) |
328
|
|
|
|
329
|
|
|
# setup the Gaussian Process kernel |
330
|
|
|
kernel_base = (1e7 * args.scale)**2 |
331
|
|
|
ksub = args.name_suffix |
332
|
|
|
|
333
|
|
|
solver = "basic" |
334
|
|
|
skwargs = {} |
335
|
|
|
if args.HODLR_Solver: |
336
|
|
|
solver = "HODLR" |
337
|
|
|
#skwargs = {"tol": 1e-3} |
338
|
|
|
|
339
|
|
|
if args.george: |
340
|
|
|
gpname, kernel = setup_george_kernel(kernls, |
341
|
|
|
kernel_base=kernel_base, fit_bias=args.fit_bias) |
342
|
|
|
gpmodel = george.GP(kernel, mean=model, |
343
|
|
|
white_noise=1.e-25, fit_white_noise=args.fit_white, |
344
|
|
|
solver=george_solvers[solver], **skwargs) |
345
|
|
|
# the george interface does not allow setting the bounds in |
346
|
|
|
# the kernel initialization so we prepare simple default bounds |
347
|
|
|
kernel_bounds = [(-0.3 * max_amp, 0.3 * max_amp) |
348
|
|
|
for _ in gpmodel.kernel.get_parameter_names()] |
349
|
|
|
bounds = gpmodel.mean.get_parameter_bounds() + kernel_bounds |
350
|
|
|
else: |
351
|
|
|
gpname, cel_terms = setup_celerite_terms(kernls, |
352
|
|
|
fit_bias=args.fit_bias, fit_white=args.fit_white) |
353
|
|
|
gpmodel = celerite.GP(cel_terms, mean=model, |
354
|
|
|
fit_white_noise=args.fit_white, |
355
|
|
|
fit_mean=True) |
356
|
|
|
bounds = gpmodel.get_parameter_bounds() |
357
|
|
|
gpmodel.compute(no_ys_train, no_errs_train) |
358
|
|
|
logging.debug("gpmodel params: %s", gpmodel.get_parameter_dict()) |
359
|
|
|
logging.debug("gpmodel bounds: %s", bounds) |
360
|
|
|
logging.debug("initial log likelihood: %s", gpmodel.log_likelihood(no_dens_train)) |
361
|
|
|
if isinstance(gpmodel, celerite.GP): |
362
|
|
|
logging.info("(GP) jitter: %s", gpmodel.kernel.jitter) |
363
|
|
|
model_name = "_".join(gpmodel.mean.get_parameter_names()).replace(':', '') |
364
|
|
|
gpmodel_name = model_name + gpname |
365
|
|
|
logging.info("GP model name: %s", gpmodel_name) |
366
|
|
|
|
367
|
|
|
pre_opt = False |
368
|
|
|
if args.optimize > 0: |
369
|
|
|
def gpmodel_mean(x, *p): |
370
|
|
|
gpmodel.set_parameter_vector(p) |
371
|
|
|
return gpmodel.mean.get_value(x) |
372
|
|
|
|
373
|
|
|
def gpmodel_res(x, *p): |
374
|
|
|
gpmodel.set_parameter_vector(p) |
375
|
|
|
return (gpmodel.mean.get_value(x) - no_dens_train) / no_errs_train |
376
|
|
|
|
377
|
|
|
def lpost(p, y, gp): |
378
|
|
|
gp.set_parameter_vector(p) |
379
|
|
|
return gp.log_likelihood(y, quiet=True) + gp.log_prior() |
380
|
|
|
|
381
|
|
|
def nlpost(p, y, gp): |
382
|
|
|
lp = lpost(p, y, gp) |
383
|
|
|
return -lp if np.isfinite(lp) else 1e25 |
384
|
|
|
|
385
|
|
|
def grad_nlpost(p, y, gp): |
386
|
|
|
gp.set_parameter_vector(p) |
387
|
|
|
grad_ll = gp.grad_log_likelihood(y) |
388
|
|
|
if isinstance(grad_ll, tuple): |
389
|
|
|
# celerite |
390
|
|
|
return -grad_ll[1] |
391
|
|
|
# george |
392
|
|
|
return -grad_ll |
393
|
|
|
|
394
|
|
|
if args.optimize == 1: |
395
|
|
|
resop_gp = op.minimize( |
396
|
|
|
nlpost, |
397
|
|
|
gpmodel.get_parameter_vector(), |
398
|
|
|
args=(no_dens_train, gpmodel), |
399
|
|
|
bounds=bounds, |
400
|
|
|
# method="l-bfgs-b", options=dict(disp=True, maxcor=100, eps=1e-9, ftol=2e-15, gtol=1e-8)) |
401
|
|
|
method="l-bfgs-b", jac=grad_nlpost) |
402
|
|
|
# method="tnc", options=dict(disp=True, maxiter=500, xtol=1e-12)) |
403
|
|
|
# method="nelder-mead", options=dict(disp=True, maxfev=100000, fatol=1.49012e-8, xatol=1.49012e-8)) |
404
|
|
|
# method="Powell", options=dict(ftol=1.49012e-08, xtol=1.49012e-08)) |
405
|
|
|
if args.optimize == 2: |
406
|
|
|
resop_gp = op.differential_evolution( |
407
|
|
|
nlpost, |
408
|
|
|
bounds=bounds, |
409
|
|
|
args=(no_dens_train, gpmodel), |
410
|
|
|
popsize=2 * args.walkers, tol=0.01) |
411
|
|
|
if args.optimize == 3: |
412
|
|
|
resop_bh = op.basinhopping( |
413
|
|
|
nlpost, |
414
|
|
|
gpmodel.get_parameter_vector(), |
415
|
|
|
niter=200, |
416
|
|
|
minimizer_kwargs=dict( |
417
|
|
|
args=(no_dens_train, gpmodel), |
418
|
|
|
bounds=bounds, |
419
|
|
|
# method="tnc")) |
420
|
|
|
# method="l-bfgs-b", options=dict(maxcor=100))) |
421
|
|
|
method="l-bfgs-b", jac=grad_nlpost)) |
422
|
|
|
# method="Nelder-Mead")) |
423
|
|
|
# method="BFGS")) |
424
|
|
|
# method="Powell", options=dict(ftol=1.49012e-08, xtol=1.49012e-08))) |
425
|
|
|
logging.debug("optimization result: %s", resop_bh) |
426
|
|
|
resop_gp = resop_bh.lowest_optimization_result |
427
|
|
|
if args.optimize == 4: |
428
|
|
|
resop_gp, cov_gp = op.curve_fit( |
429
|
|
|
gpmodel_mean, |
430
|
|
|
no_ys_train, no_dens_train, gpmodel.get_parameter_vector(), |
431
|
|
|
bounds=tuple(np.array(bounds).T), |
432
|
|
|
# method='lm', |
433
|
|
|
# absolute_sigma=True, |
434
|
|
|
sigma=no_errs_train) |
435
|
|
|
print(resop_gp, np.sqrt(np.diag(cov_gp))) |
436
|
|
|
logging.info("%s", resop_gp.message) |
437
|
|
|
logging.debug("optimization result: %s", resop_gp) |
438
|
|
|
logging.info("gpmodel dict: %s", gpmodel.get_parameter_dict()) |
439
|
|
|
logging.info("log posterior trained: %s", lpost(gpmodel.get_parameter_vector(), no_dens_train, gpmodel)) |
440
|
|
|
gpmodel.compute(no_ys_test, no_errs_test) |
441
|
|
|
logging.info("log posterior test: %s", lpost(gpmodel.get_parameter_vector(), no_dens_test, gpmodel)) |
442
|
|
|
gpmodel.compute(no_ys, no_errs) |
443
|
|
|
logging.info("log posterior all: %s", lpost(gpmodel.get_parameter_vector(), no_dens, gpmodel)) |
444
|
|
|
# cross check to make sure that the gpmodel parameter vector is really |
445
|
|
|
# set to the fitted parameters |
446
|
|
|
logging.info("opt. model vector: %s", resop_gp.x) |
447
|
|
|
gpmodel.compute(no_ys_train, no_errs_train) |
448
|
|
|
logging.debug("opt. log posterior trained 1: %s", lpost(resop_gp.x, no_dens_train, gpmodel)) |
449
|
|
|
gpmodel.compute(no_ys_test, no_errs_test) |
450
|
|
|
logging.debug("opt. log posterior test 1: %s", lpost(resop_gp.x, no_dens_test, gpmodel)) |
451
|
|
|
gpmodel.compute(no_ys, no_errs) |
452
|
|
|
logging.debug("opt. log posterior all 1: %s", lpost(resop_gp.x, no_dens, gpmodel)) |
453
|
|
|
logging.debug("opt. model vector: %s", gpmodel.get_parameter_vector()) |
454
|
|
|
gpmodel.compute(no_ys_train, no_errs_train) |
455
|
|
|
logging.debug("opt. log posterior trained 2: %s", lpost(gpmodel.get_parameter_vector(), no_dens_train, gpmodel)) |
456
|
|
|
gpmodel.compute(no_ys_test, no_errs_test) |
457
|
|
|
logging.debug("opt. log posterior test 2: %s", lpost(gpmodel.get_parameter_vector(), no_dens_test, gpmodel)) |
458
|
|
|
gpmodel.compute(no_ys, no_errs) |
459
|
|
|
logging.debug("opt. log posterior all 2: %s", lpost(gpmodel.get_parameter_vector(), no_dens, gpmodel)) |
460
|
|
|
pre_opt = resop_gp.success |
461
|
|
|
try: |
462
|
|
|
logging.info("GM lt: %s", gpmodel.get_parameter("mean:GM:tau0")) |
463
|
|
|
except ValueError: |
464
|
|
|
pass |
465
|
|
|
logging.info("(GP) model: %s", gpmodel.kernel) |
466
|
|
|
if isinstance(gpmodel, celerite.GP): |
467
|
|
|
logging.info("(GP) jitter: %s", gpmodel.kernel.jitter) |
468
|
|
|
|
469
|
|
|
bestfit = gpmodel.get_parameter_vector() |
470
|
|
|
filename_base = ("NO_regress_fit_{0}_{1:.0f}_{2:.0f}_{{0}}_{3}" |
471
|
|
|
.format(gpmodel_name, lat * 10, alt, ksub)) |
472
|
|
|
|
473
|
|
|
if args.mcmc: |
474
|
|
|
gpmodel.compute(no_ys_train, no_errs_train) |
475
|
|
|
samples, lnp = mcmc_sample_model(gpmodel, |
476
|
|
|
no_dens_train, |
477
|
|
|
beta=1.0, |
478
|
|
|
nwalkers=args.walkers, nburnin=args.burn_in, |
479
|
|
|
nprod=args.production, nthreads=args.threads, |
480
|
|
|
show_progress=args.progress, |
481
|
|
|
optimized=pre_opt, bounds=bounds, return_logpost=True) |
482
|
|
|
|
483
|
|
|
if args.train_fraction < 1. or args.test_fraction < 1.: |
484
|
|
|
logging.info("Statistics for the test samples") |
485
|
|
|
mcmc_statistics(gpmodel, |
486
|
|
|
no_ys_test, no_dens_test, no_errs_test, |
487
|
|
|
no_ys_train, no_dens_train, no_errs_train, |
488
|
|
|
samples, lnp, |
489
|
|
|
) |
490
|
|
|
logging.info("Statistics for all samples") |
491
|
|
|
mcmc_statistics(gpmodel, |
492
|
|
|
no_ys, no_dens, no_errs, |
493
|
|
|
no_ys_train, no_dens_train, no_errs_train, |
494
|
|
|
samples, lnp, |
495
|
|
|
) |
496
|
|
|
|
497
|
|
|
sampl_percs = np.percentile(samples, [2.5, 50, 97.5], axis=0) |
498
|
|
|
if args.plot_corner: |
499
|
|
|
import corner |
500
|
|
|
# Corner plot of the sampled parameters |
501
|
|
|
fig = corner.corner(samples, |
502
|
|
|
quantiles=[0.025, 0.5, 0.975], |
503
|
|
|
show_titles=True, |
504
|
|
|
labels=gpmodel.get_parameter_names(), |
505
|
|
|
truths=bestfit, |
506
|
|
|
hist_args=dict(normed=True)) |
507
|
|
|
fig.savefig(filename_base.format("corner") + ".pdf", transparent=True) |
508
|
|
|
|
509
|
|
|
if args.save_samples: |
510
|
|
|
if args.samples_format in ["npz"]: |
511
|
|
|
# save the samples compressed to save space. |
512
|
|
|
np.savez_compressed(filename_base.format("sampls") + ".npz", |
513
|
|
|
samples=samples) |
514
|
|
|
if args.samples_format in ["nc", "netcdf4"]: |
515
|
|
|
save_samples_netcdf(filename_base.format("sampls") + ".nc", |
516
|
|
|
gpmodel, alt, lat, samples, scale=args.scale, compressed=True) |
517
|
|
|
if args.samples_format in ["h5", "hdf5"]: |
518
|
|
|
save_samples_netcdf(filename_base.format("sampls") + ".h5", |
519
|
|
|
gpmodel, alt, lat, samples, scale=args.scale, compressed=True) |
520
|
|
|
# MCMC finished here |
521
|
|
|
|
522
|
|
|
# set the model times and errors to use the full data set for plotting |
523
|
|
|
gpmodel.compute(no_ys, no_errs) |
524
|
|
|
if args.save_model: |
525
|
|
|
try: |
526
|
|
|
# python 2 |
527
|
|
|
import cPickle as pickle |
528
|
|
|
except ImportError: |
529
|
|
|
# python 3 |
530
|
|
|
import pickle |
531
|
|
|
# pickle and save the model |
532
|
|
|
with open(filename_base.format("model") + ".pkl", "wb") as f: |
533
|
|
|
pickle.dump((gpmodel), f, -1) |
534
|
|
|
|
535
|
|
|
if args.plot_samples and args.mcmc: |
536
|
|
|
plot_random_samples(gpmodel, no_ys, no_dens, no_errs, |
537
|
|
|
samples, args.scale, |
538
|
|
|
filename_base.format("sampls") + ".pdf", |
539
|
|
|
size=4, extra_years=[4, 2]) |
540
|
|
|
|
541
|
|
|
if args.plot_median: |
542
|
|
|
plot_single_sample_and_residuals(gpmodel, no_ys, no_dens, no_errs, |
543
|
|
|
sampl_percs[1], |
544
|
|
|
filename_base.format("median") + ".pdf") |
545
|
|
|
if args.plot_residuals: |
546
|
|
|
plot_residual(gpmodel, no_ys, no_dens, no_errs, |
547
|
|
|
sampl_percs[1], args.scale, |
548
|
|
|
filename_base.format("medres") + ".pdf") |
549
|
|
|
if args.plot_maxlnp: |
550
|
|
|
plot_single_sample_and_residuals(gpmodel, no_ys, no_dens, no_errs, |
551
|
|
|
samples[np.argmax(lnp)], |
552
|
|
|
filename_base.format("maxlnp") + ".pdf") |
553
|
|
|
if args.plot_maxlnpres: |
554
|
|
|
plot_residual(gpmodel, no_ys, no_dens, no_errs, |
555
|
|
|
samples[np.argmax(lnp)], args.scale, |
556
|
|
|
filename_base.format("mlpres") + ".pdf") |
557
|
|
|
|
558
|
|
|
labels = gpmodel.get_parameter_names() |
559
|
|
|
logging.info("param percentiles [2.5, 50, 97.5]:") |
560
|
|
|
for pc, label in zip(sampl_percs.T, labels): |
561
|
|
|
median = pc[1] |
562
|
|
|
pc_minus = median - pc[0] |
563
|
|
|
pc_plus = pc[2] - median |
564
|
|
|
logging.debug("%s: %s", label, pc) |
565
|
|
|
logging.info("%s: %.6f (- %.6f) (+ %.6f)", label, |
566
|
|
|
median, pc_minus, pc_plus) |
567
|
|
|
|
568
|
|
|
logging.info("Finished successfully.") |
569
|
|
|
|
570
|
|
|
|
571
|
1 |
|
if __name__ == "__main__": |
572
|
|
|
main() |
573
|
|
|
|