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