sciapy.regress.statistics._log_prob_pt_samples_mt() - Code Metrics - st-bender/sciapy - Measure and Improve Code Quality continuously with Scrutinizer

_log_prob_pt_samples_mt() A
last analyzed 2023-01-30 12:22 UTC

↳ Parent: sciapy.regress.statistics

Complexity

Conditions

Size

Total Lines	20
Code Lines	14

Duplication

Lines	0
Ratio	0 %

Code Coverage

Tests	1
CRAP Score	10.2065

Importance

Changes

Metric	Value
cc	3
eloc	14
nop	3
dl	0
loc	20
ccs	1
cts	14
cp	0.0714
crap	10.2065
rs	9.7
c	0
b	0
f	0

# -*- coding: utf-8 -*-
# vim:fileencoding=utf-8
#
# Copyright (c) 2017-2019 Stefan Bender
#
# This module is part of sciapy.
# sciapy is free software: you can redistribute it or modify
# it under the terms of the GNU General Public License as published
# by the Free Software Foundation, version 2.
# See accompanying LICENSE file or http://www.gnu.org/licenses/gpl-2.0.html.
"""SCIAMACHY MCMC statistic tools

Statistical functions for MCMC sampled parameters.
"""

import logging

import numpy as np
from scipy import linalg

__all__ = ["mcmc_statistics", "waic_loo"]


def _log_prob(resid, var):
	return -0.5 * (np.log(2 * np.pi * var) + resid**2 / var)


def _log_lh_pt(gp, times, data, errs, s):
	gp.set_parameter_vector(s)
	resid = data - gp.mean.get_value(times)
	ker = gp.get_matrix(times, include_diagonal=True)
	ker[np.diag_indices_from(ker)] += errs**2
	ll, lower = linalg.cho_factor(
			ker, lower=True, check_finite=False, overwrite_a=True)
	linv_r = linalg.solve_triangular(
			ll, resid, lower=True, check_finite=False, overwrite_b=True)
	return -0.5 * (np.log(2. * np.pi) + linv_r**2) - np.log(np.diag(ll))


def _log_pred_pt(gp, train_data, times, data, errs, s):
	gp.set_parameter_vector(s)
	gppred, pvar_gp = gp.predict(train_data, t=times, return_var=True)
	# the predictive covariance should include the data variance
	# if noisy_targets and errs is not None:
	if errs is not None:
		pvar_gp += errs**2
	# residuals
	resid_gp = gppred - data  # GP residuals
	return _log_prob(resid_gp, pvar_gp)


def _log_prob_pt_samples_dask(log_p_pt, samples,
		nthreads=1, cluster=None):
	from dask.distributed import Client, LocalCluster, progress
	import dask.bag as db

	# calculate the point-wise probabilities and stack them together
	if cluster is None:
		# start local dask cluster
		_cl = LocalCluster(n_workers=nthreads, threads_per_worker=1)
	else:
		# use provided dask cluster
		_cl = cluster

	with Client(_cl):
		_log_pred = db.from_sequence(samples).map(log_p_pt)
		progress(_log_pred)
		ret = np.stack(_log_pred.compute())

	if cluster is None:
		_cl.close()

	return ret


def _log_prob_pt_samples_mt(log_p_pt, samples, nthreads=1):
	from multiprocessing import pool
	try:
		from tqdm.autonotebook import tqdm
	except ImportError:
		tqdm = None

	# multiprocessing.pool
	_p = pool.Pool(processes=nthreads)

	_mapped = _p.imap_unordered(log_p_pt, samples)
	if tqdm is not None:
		_mapped = tqdm(_mapped, total=len(samples))

	ret = np.stack(list(_mapped))

	_p.close()
	_p.join()

	return ret


def waic_loo(model, times, data, errs,
		samples,
		method="likelihood",
		train_data=None,
		noisy_targets=True,
		nthreads=1,
		use_dask=False,
		dask_cluster=None,
		):
	"""Watanabe-Akaike information criterion (WAIC) and LOO IC of the (GP) model

	Calculates the WAIC and leave-one-out (LOO) cross validation scores and
	information criteria (IC) from the MCMC samples of the posterior parameter
	distributions. Uses the posterior point-wise (per data point)
	probabilities and the formulae from [1]_ and [2]_.

	.. [1] Vehtari, Gelman, and Gabry, Stat Comput (2017) 27:1413–1432,
		doi: 10.1007/s11222-016-9696-4

	.. [2] Vehtari and Gelman, (unpublished)
		http://www.stat.columbia.edu/~gelman/research/unpublished/waic_stan.pdf
		http://www.stat.columbia.edu/~gelman/research/unpublished/loo_stan.pdf

	Parameters
	----------
	model : `celerite.GP`, `george.GP` or `CeleriteModelSet` instance
		The model instance whose parameter distribution was drawn.
	times : (M,) array_like
		The test coordinates to predict or evaluate the model on.
	data : (M,) array_like
		The test data to test the model against.
	errs : (M,) array_like
		The errors (variances) of the test data.
	samples : (K, L) array_like
		The `K` MCMC samples of the `L` parameter distributions.
	method : str ("likelihood" or "predict"), optional
		The method to "predict" the data, the default uses the (log)likelihood
		in the same way as is done when fitting (training) the model.
		"predict" uses the actual GP prediction, might be useful if the IC
		should be estimated for actual test data that was not used to train
		the model.
	train_data : (N,) array_like, optional
		The data on which the model was trained, needed if method="predict" is
		used, otherwise None is the default and the likelihood is used.
	noisy_targets : bool, optional
		Include the given errors when calculating the predictive probability.
	nthreads : int, optional
		Number of threads to distribute the point-wise probability
		calculations to (default: 1).
	use_dask : boot, optional
		Use `dask.distributed` to distribute the point-wise probability
		calculations to `nthreads` workers. The default is to use
		`multiprocessing.pool.Pool()`.
	dask_cluster: str, or `dask.distributed.Cluster` instance, optional
		Will be passed to `dask.distributed.Client()`
		This can be the address of a Scheduler server like a string
		'127.0.0.1:8786' or a cluster object like `dask.distributed.LocalCluster()`.

	Returns
	-------
	waic, waic_se, p_waic, loo_ic, loo_se, p_loo : tuple
		The WAIC and its standard error as well as the
		estimated effective number of parameters, p_waic.
		The LOO IC, its standard error, and the estimated
		effective number of parameters, p_loo.
	"""
	from functools import partial
	from scipy.special import logsumexp

	# the predictive covariance should include the data variance
	# set to a small value if we don't want to account for them
	if not noisy_targets or errs is None:
		errs = 1.123e-12

	# point-wise posterior/predictive probabilities
	_log_p_pt = partial(_log_lh_pt, model, times, data, errs)
	if method == "predict" and train_data is not None:
		_log_p_pt = partial(_log_pred_pt, model, train_data, times, data, errs)

	# calculate the point-wise probabilities and stack them together
	if nthreads > 1 and use_dask:
		log_pred = _log_prob_pt_samples_dask(_log_p_pt, samples,
				nthreads=nthreads, cluster=dask_cluster)
	else:
		log_pred = _log_prob_pt_samples_mt(_log_p_pt, samples,
				nthreads=nthreads)

	lppd_i = logsumexp(log_pred, b=1. / log_pred.shape[0], axis=0)
	p_waic_i = np.nanvar(log_pred, ddof=1, axis=0)
	if np.any(p_waic_i > 0.4):
		logging.warn("""For one or more samples the posterior variance of the
		log predictive densities exceeds 0.4. This could be indication of
		WAIC starting to fail see http://arxiv.org/abs/1507.04544 for details
		""")
	elpd_i = lppd_i - p_waic_i
	waic_i = -2. * elpd_i
	waic_se = np.sqrt(len(waic_i) * np.nanvar(waic_i, ddof=1))
	waic = np.nansum(waic_i)
	p_waic = np.nansum(p_waic_i)
	if 2. * p_waic > len(waic_i):
		logging.warn("""p_waic > n / 2,
		the WAIC approximation is unreliable.
		""")
	logging.info("WAIC: %s, waic_se: %s, p_w: %s", waic, waic_se, p_waic)

	# LOO
	loo_ws = 1. / np.exp(log_pred - np.nanmax(log_pred, axis=0))
	loo_ws_n = loo_ws / np.nanmean(loo_ws, axis=0)
	loo_ws_r = np.clip(loo_ws_n, None, np.sqrt(log_pred.shape[0]))
	elpd_loo_i = logsumexp(log_pred,
			b=loo_ws_r / np.nansum(loo_ws_r, axis=0),
			axis=0)
	p_loo_i = lppd_i - elpd_loo_i
	loo_ic_i = -2 * elpd_loo_i
	loo_ic_se = np.sqrt(len(loo_ic_i) * np.nanvar(loo_ic_i))
	loo_ic = np.nansum(loo_ic_i)
	p_loo = np.nansum(p_loo_i)
	logging.info("loo IC: %s, se: %s, p_loo: %s", loo_ic, loo_ic_se, p_loo)

	# van der Linde, 2005, Statistica Neerlandica, 2005
	# https://doi.org/10.1111/j.1467-9574.2005.00278.x
	hy1 = -np.nanmean(lppd_i)
	hy2 = -np.nanmedian(lppd_i)
	logging.info("H(Y): mean %s, median: %s", hy1, hy2)

	return waic, waic_se, p_waic, loo_ic, loo_ic_se, p_loo


def mcmc_statistics(model, times, data, errs,
		train_times, train_data, train_errs,
		samples, lnp,
		median=False):
	"""Statistics for the (GP) model against the provided data

	Statistical information about the model and the sampled parameter
	distributions with respect to the provided data and its variance.

	Sends the calculated values to the logger, includes the mean
	standardized log loss as described in R&W, 2006, section 2.5, (2.34),
	and some slightly adapted $\\chi^2_{\\text{red}}$ and $R^2$ scores.

	Parameters
	----------
	model : `celerite.GP`, `george.GP` or `CeleriteModelSet` instance
		The model instance whose parameter distribution was drawn.
	times : (M,) array_like
		The test coordinates to predict or evaluate the model on.
	data : (M,) array_like
		The test data to test the model against.
	errs : (M,) array_like
		The errors (variances) of the test data.
	train_times : (N,) array_like
		The coordinates on which the model was trained.
	train_data : (N,) array_like
		The data on which the model was trained.
	train_errs : (N,) array_like
		The errors (variances) of the training data.
	samples : (K, L) array_like
		The `K` MCMC samples of the `L` parameter distributions.
	lnp : (K,) array_like
		The posterior log probabilities of the `K` MCMC samples.
	median : bool, optional
		Whether to use the median of the sampled distributions or
		the maximum posterior sample (the default) to evaluate the
		statistics.

	Returns
	-------
	nothing
	"""
	ndat = len(times)
	ndim = len(model.get_parameter_vector())
	mdim = len(model.mean.get_parameter_vector())
	samples_max_lp = np.max(lnp)
	if median:
		sample_pos = np.nanmedian(samples, axis=0)
	else:
		sample_pos = samples[np.argmax(lnp)]
	model.set_parameter_vector(sample_pos)
	# calculate the GP predicted values and covariance
	gppred, gpcov = model.predict(train_data, t=times, return_cov=True)
	# the predictive covariance should include the data variance
	gpcov[np.diag_indices_from(gpcov)] += errs**2
	# residuals
	resid_mod = model.mean.get_value(times) - data  # GP mean model
	resid_gp = gppred - data  # GP prediction
	resid_triv = np.nanmean(train_data) - data  # trivial model
	_const = ndat * np.log(2.0 * np.pi)
	test_logpred = -0.5 * (resid_gp.dot(linalg.solve(gpcov, resid_gp))
			+ np.trace(np.log(gpcov))
			+ _const)
	# MSLL -- mean standardized log loss
	# as described in R&W, 2006, section 2.5, (2.34)
	var_mod = np.nanvar(resid_mod, ddof=mdim)  # mean model variance
	var_gp = np.nanvar(resid_gp, ddof=ndim)  # gp model variance
	var_triv = np.nanvar(train_data, ddof=1)  # trivial model variance
	logpred_mod = _log_prob(resid_mod, var_mod)
	logpred_gp = _log_prob(resid_gp, var_gp)
	logpred_triv = _log_prob(resid_triv, var_triv)
	logging.info("MSLL mean: %s", np.nanmean(-logpred_mod + logpred_triv))
	logging.info("MSLL gp: %s", np.nanmean(-logpred_gp + logpred_triv))
	# predictive variances
	logpred_mod = _log_prob(resid_mod, var_mod + errs**2)
	logpred_gp = _log_prob(resid_gp, var_gp + errs**2)
	logpred_triv = _log_prob(resid_triv, var_triv + errs**2)
	logging.info("pred MSLL mean: %s", np.nanmean(-logpred_mod + logpred_triv))
	logging.info("pred MSLL gp: %s", np.nanmean(-logpred_gp + logpred_triv))
	# cost values
	cost_mod = np.sum(resid_mod**2)
	cost_triv = np.sum(resid_triv**2)
	cost_gp = np.sum(resid_gp**2)
	# chi^2 (variance corrected costs)
	chisq_mod_ye = np.sum((resid_mod / errs)**2)
	chisq_triv = np.sum((resid_triv / errs)**2)
	chisq_gpcov = resid_mod.dot(linalg.solve(gpcov, resid_mod))
	# adjust for degrees of freedom
	cost_gp_dof = cost_gp / (ndat - ndim)
	cost_mod_dof = cost_mod / (ndat - mdim)
	cost_triv_dof = cost_triv / (ndat - 1)
	# reduced chi^2
	chisq_red_mod_ye = chisq_mod_ye / (ndat - mdim)
	chisq_red_triv = chisq_triv / (ndat - 1)
	chisq_red_gpcov = chisq_gpcov / (ndat - ndim)
	# "generalized" R^2
	logp_triv1 = np.sum(_log_prob(resid_triv, errs**2))
	logp_triv2 = np.sum(_log_prob(resid_triv, var_triv))
	logp_triv3 = np.sum(_log_prob(resid_triv, var_triv + errs**2))
	log_lambda1 = test_logpred - logp_triv1
	log_lambda2 = test_logpred - logp_triv2
	log_lambda3 = test_logpred - logp_triv3
	gen_rsq1a = 1 - np.exp(-2 * log_lambda1 / ndat)
	gen_rsq1b = 1 - np.exp(-2 * log_lambda1 / (ndat - ndim))
	gen_rsq2a = 1 - np.exp(-2 * log_lambda2 / ndat)
	gen_rsq2b = 1 - np.exp(-2 * log_lambda2 / (ndat - ndim))
	gen_rsq3a = 1 - np.exp(-2 * log_lambda3 / ndat)
	gen_rsq3b = 1 - np.exp(-2 * log_lambda3 / (ndat - ndim))
	# sent to the logger
	logging.info("train max logpost: %s", samples_max_lp)
	logging.info("test log_pred: %s", test_logpred)
	logging.info("1a cost mean model: %s, dof adj: %s", cost_mod, cost_mod_dof)
	logging.debug("1c cost gp predict: %s, dof adj: %s", cost_gp, cost_gp_dof)
	logging.debug("1b cost triv model: %s, dof adj: %s", cost_triv, cost_triv_dof)
	logging.info("1d var resid mean model: %s, gp model: %s, triv: %s",
			var_mod, var_gp, var_triv)
	logging.info("2a adjR2 mean model: %s, adjR2 gp predict: %s",
			1 - cost_mod_dof / cost_triv_dof, 1 - cost_gp_dof / cost_triv_dof)
	logging.info("2b red chi^2 mod: %s / triv: %s = %s",
			chisq_red_mod_ye, chisq_red_triv, chisq_red_mod_ye / chisq_red_triv)
	logging.info("2c red chi^2 mod (gp cov): %s / triv: %s = %s",
			chisq_red_gpcov, chisq_red_triv, chisq_red_gpcov / chisq_red_triv)
	logging.info("3a stand. red chi^2: %s", chisq_red_gpcov / chisq_red_triv)
	logging.info("3b 1 - stand. red chi^2: %s",
			1 - chisq_red_gpcov / chisq_red_triv)
	logging.info("5a generalized R^2: 1a: %s, 1b: %s",
			gen_rsq1a, gen_rsq1b)
	logging.info("5b generalized R^2: 2a: %s, 2b: %s",
			gen_rsq2a, gen_rsq2b)
	logging.info("5c generalized R^2: 3a: %s, 3b: %s",
			gen_rsq3a, gen_rsq3b)
	try:
		# celerite
		logdet = model.solver.log_determinant()
	except TypeError:
		# george
		logdet = model.solver.log_determinant
	logging.debug("5 logdet: %s, const 2: %s", logdet, _const)


1		# -- coding: utf-8 --
2		# vim:fileencoding=utf-8
3		#
4		# Copyright (c) 2017-2019 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 MCMC statistic tools
12
13		Statistical functions for MCMC sampled parameters.
14		"""
15
16	1	import logging
17
18	1	import numpy as np
19	1	from scipy import linalg
20
21	1	__all__ = ["mcmc_statistics", "waic_loo"]
22
23
24	1	def _log_prob(resid, var):
25	1	return -0.5 * (np.log(2 * np.pi * var) + resid**2 / var)
26
27
28	1	def _log_lh_pt(gp, times, data, errs, s):
29		gp.set_parameter_vector(s)
30		resid = data - gp.mean.get_value(times)
31		ker = gp.get_matrix(times, include_diagonal=True)
32		ker[np.diag_indices_from(ker)] += errs**2
33		ll, lower = linalg.cho_factor(
34		ker, lower=True, check_finite=False, overwrite_a=True)
35		linv_r = linalg.solve_triangular(
36		ll, resid, lower=True, check_finite=False, overwrite_b=True)
37		return -0.5 * (np.log(2. * np.pi) + linv_r**2) - np.log(np.diag(ll))
38
39
40	1	def _log_pred_pt(gp, train_data, times, data, errs, s):
41		gp.set_parameter_vector(s)
42		gppred, pvar_gp = gp.predict(train_data, t=times, return_var=True)
43		# the predictive covariance should include the data variance
44		# if noisy_targets and errs is not None:
45		if errs is not None:
46		pvar_gp += errs**2
47		# residuals
48		resid_gp = gppred - data # GP residuals
49		return _log_prob(resid_gp, pvar_gp)
50
51
52	1	def _log_prob_pt_samples_dask(log_p_pt, samples,
53		nthreads=1, cluster=None):
54		from dask.distributed import Client, LocalCluster, progress
55		import dask.bag as db
56
57		# calculate the point-wise probabilities and stack them together
58		if cluster is None:
59		# start local dask cluster
60		_cl = LocalCluster(n_workers=nthreads, threads_per_worker=1)
61		else:
62		# use provided dask cluster
63		_cl = cluster
64
65		with Client(_cl):
66		_log_pred = db.from_sequence(samples).map(log_p_pt)
67		progress(_log_pred)
68		ret = np.stack(_log_pred.compute())
69
70		if cluster is None:
71		_cl.close()
72
73		return ret
74
75
76	1	def _log_prob_pt_samples_mt(log_p_pt, samples, nthreads=1):
77		from multiprocessing import pool
78		try:
79		from tqdm.autonotebook import tqdm
80		except ImportError:
81		tqdm = None
82
83		# multiprocessing.pool
84		_p = pool.Pool(processes=nthreads)
85
86		_mapped = _p.imap_unordered(log_p_pt, samples)
87		if tqdm is not None:
88		_mapped = tqdm(_mapped, total=len(samples))
89
90		ret = np.stack(list(_mapped))
91
92		_p.close()
93		_p.join()
94
95		return ret
96
97
98	1	def waic_loo(model, times, data, errs,
99		samples,
100		method="likelihood",
101		train_data=None,
102		noisy_targets=True,
103		nthreads=1,
104		use_dask=False,
105		dask_cluster=None,
106		):
107		"""Watanabe-Akaike information criterion (WAIC) and LOO IC of the (GP) model
108
109		Calculates the WAIC and leave-one-out (LOO) cross validation scores and
110		information criteria (IC) from the MCMC samples of the posterior parameter
111		distributions. Uses the posterior point-wise (per data point)
112		probabilities and the formulae from [1]_ and [2]_.
113
114		.. [1] Vehtari, Gelman, and Gabry, Stat Comput (2017) 27:1413–1432,
115		doi: 10.1007/s11222-016-9696-4
116
117		.. [2] Vehtari and Gelman, (unpublished)
118		http://www.stat.columbia.edu/~gelman/research/unpublished/waic_stan.pdf
119		http://www.stat.columbia.edu/~gelman/research/unpublished/loo_stan.pdf
120
121		Parameters
122		----------
123		model : `celerite.GP`, `george.GP` or `CeleriteModelSet` instance
124		The model instance whose parameter distribution was drawn.
125		times : (M,) array_like
126		The test coordinates to predict or evaluate the model on.
127		data : (M,) array_like
128		The test data to test the model against.
129		errs : (M,) array_like
130		The errors (variances) of the test data.
131		samples : (K, L) array_like
132		The `K` MCMC samples of the `L` parameter distributions.
133		method : str ("likelihood" or "predict"), optional
134		The method to "predict" the data, the default uses the (log)likelihood
135		in the same way as is done when fitting (training) the model.
136		"predict" uses the actual GP prediction, might be useful if the IC
137		should be estimated for actual test data that was not used to train
138		the model.
139		train_data : (N,) array_like, optional
140		The data on which the model was trained, needed if method="predict" is
141		used, otherwise None is the default and the likelihood is used.
142		noisy_targets : bool, optional
143		Include the given errors when calculating the predictive probability.
144		nthreads : int, optional
145		Number of threads to distribute the point-wise probability
146		calculations to (default: 1).
147		use_dask : boot, optional
148		Use `dask.distributed` to distribute the point-wise probability
149		calculations to `nthreads` workers. The default is to use
150		`multiprocessing.pool.Pool()`.
151		dask_cluster: str, or `dask.distributed.Cluster` instance, optional
152		Will be passed to `dask.distributed.Client()`
153		This can be the address of a Scheduler server like a string
154		'127.0.0.1:8786' or a cluster object like `dask.distributed.LocalCluster()`.
155
156		Returns
157		-------
158		waic, waic_se, p_waic, loo_ic, loo_se, p_loo : tuple
159		The WAIC and its standard error as well as the
160		estimated effective number of parameters, p_waic.
161		The LOO IC, its standard error, and the estimated
162		effective number of parameters, p_loo.
163		"""
164		from functools import partial
165		from scipy.special import logsumexp
166
167		# the predictive covariance should include the data variance
168		# set to a small value if we don't want to account for them
169		if not noisy_targets or errs is None:
170		errs = 1.123e-12
171
172		# point-wise posterior/predictive probabilities
173		_log_p_pt = partial(_log_lh_pt, model, times, data, errs)
174		if method == "predict" and train_data is not None:
175		_log_p_pt = partial(_log_pred_pt, model, train_data, times, data, errs)
176
177		# calculate the point-wise probabilities and stack them together
178		if nthreads > 1 and use_dask:
179		log_pred = _log_prob_pt_samples_dask(_log_p_pt, samples,
180		nthreads=nthreads, cluster=dask_cluster)
181		else:
182		log_pred = _log_prob_pt_samples_mt(_log_p_pt, samples,
183		nthreads=nthreads)
184
185		lppd_i = logsumexp(log_pred, b=1. / log_pred.shape[0], axis=0)
186		p_waic_i = np.nanvar(log_pred, ddof=1, axis=0)
187		if np.any(p_waic_i > 0.4):
188		logging.warn("""For one or more samples the posterior variance of the
189		log predictive densities exceeds 0.4. This could be indication of
190		WAIC starting to fail see http://arxiv.org/abs/1507.04544 for details
191		""")
192		elpd_i = lppd_i - p_waic_i
193		waic_i = -2. * elpd_i
194		waic_se = np.sqrt(len(waic_i) * np.nanvar(waic_i, ddof=1))
195		waic = np.nansum(waic_i)
196		p_waic = np.nansum(p_waic_i)
197		if 2. * p_waic > len(waic_i):
198		logging.warn("""p_waic > n / 2,
199		the WAIC approximation is unreliable.
200		""")
201		logging.info("WAIC: %s, waic_se: %s, p_w: %s", waic, waic_se, p_waic)
202
203		# LOO
204		loo_ws = 1. / np.exp(log_pred - np.nanmax(log_pred, axis=0))
205		loo_ws_n = loo_ws / np.nanmean(loo_ws, axis=0)
206		loo_ws_r = np.clip(loo_ws_n, None, np.sqrt(log_pred.shape[0]))
207		elpd_loo_i = logsumexp(log_pred,
208		b=loo_ws_r / np.nansum(loo_ws_r, axis=0),
209		axis=0)
210		p_loo_i = lppd_i - elpd_loo_i
211		loo_ic_i = -2 * elpd_loo_i
212		loo_ic_se = np.sqrt(len(loo_ic_i) * np.nanvar(loo_ic_i))
213		loo_ic = np.nansum(loo_ic_i)
214		p_loo = np.nansum(p_loo_i)
215		logging.info("loo IC: %s, se: %s, p_loo: %s", loo_ic, loo_ic_se, p_loo)
216
217		# van der Linde, 2005, Statistica Neerlandica, 2005
218		# https://doi.org/10.1111/j.1467-9574.2005.00278.x
219		hy1 = -np.nanmean(lppd_i)
220		hy2 = -np.nanmedian(lppd_i)
221		logging.info("H(Y): mean %s, median: %s", hy1, hy2)
222
223		return waic, waic_se, p_waic, loo_ic, loo_ic_se, p_loo
224
225
226	1	def mcmc_statistics(model, times, data, errs,
227		train_times, train_data, train_errs,
228		samples, lnp,
229		median=False):
230		"""Statistics for the (GP) model against the provided data
231
232		Statistical information about the model and the sampled parameter
233		distributions with respect to the provided data and its variance.
234
235		Sends the calculated values to the logger, includes the mean
236		standardized log loss as described in R&W, 2006, section 2.5, (2.34),
237		and some slightly adapted $\\chi^2_{\\text{red}}$ and $R^2$ scores.
238
239		Parameters
240		----------
241		model : `celerite.GP`, `george.GP` or `CeleriteModelSet` instance
242		The model instance whose parameter distribution was drawn.
243		times : (M,) array_like
244		The test coordinates to predict or evaluate the model on.
245		data : (M,) array_like
246		The test data to test the model against.
247		errs : (M,) array_like
248		The errors (variances) of the test data.
249		train_times : (N,) array_like
250		The coordinates on which the model was trained.
251		train_data : (N,) array_like
252		The data on which the model was trained.
253		train_errs : (N,) array_like
254		The errors (variances) of the training data.
255		samples : (K, L) array_like
256		The `K` MCMC samples of the `L` parameter distributions.
257		lnp : (K,) array_like
258		The posterior log probabilities of the `K` MCMC samples.
259		median : bool, optional
260		Whether to use the median of the sampled distributions or
261		the maximum posterior sample (the default) to evaluate the
262		statistics.
263
264		Returns
265		-------
266		nothing
267		"""
268	1	ndat = len(times)
269	1	ndim = len(model.get_parameter_vector())
270	1	mdim = len(model.mean.get_parameter_vector())
271	1	samples_max_lp = np.max(lnp)
272	1	if median:
273		sample_pos = np.nanmedian(samples, axis=0)
274		else:
275	1	sample_pos = samples[np.argmax(lnp)]
276	1	model.set_parameter_vector(sample_pos)
277		# calculate the GP predicted values and covariance
278	1	gppred, gpcov = model.predict(train_data, t=times, return_cov=True)
279		# the predictive covariance should include the data variance
280	1	gpcov[np.diag_indices_from(gpcov)] += errs**2
281		# residuals
282	1	resid_mod = model.mean.get_value(times) - data # GP mean model
283	1	resid_gp = gppred - data # GP prediction
284	1	resid_triv = np.nanmean(train_data) - data # trivial model
285	1	_const = ndat * np.log(2.0 * np.pi)
286	1	test_logpred = -0.5 * (resid_gp.dot(linalg.solve(gpcov, resid_gp))
287		+ np.trace(np.log(gpcov))
288		+ _const)
289		# MSLL -- mean standardized log loss
290		# as described in R&W, 2006, section 2.5, (2.34)
291	1	var_mod = np.nanvar(resid_mod, ddof=mdim) # mean model variance
292	1	var_gp = np.nanvar(resid_gp, ddof=ndim) # gp model variance
293	1	var_triv = np.nanvar(train_data, ddof=1) # trivial model variance
294	1	logpred_mod = _log_prob(resid_mod, var_mod)
295	1	logpred_gp = _log_prob(resid_gp, var_gp)
296	1	logpred_triv = _log_prob(resid_triv, var_triv)
297	1	logging.info("MSLL mean: %s", np.nanmean(-logpred_mod + logpred_triv))
298	1	logging.info("MSLL gp: %s", np.nanmean(-logpred_gp + logpred_triv))
299		# predictive variances
300	1	logpred_mod = _log_prob(resid_mod, var_mod + errs**2)
301	1	logpred_gp = _log_prob(resid_gp, var_gp + errs**2)
302	1	logpred_triv = _log_prob(resid_triv, var_triv + errs**2)
303	1	logging.info("pred MSLL mean: %s", np.nanmean(-logpred_mod + logpred_triv))
304	1	logging.info("pred MSLL gp: %s", np.nanmean(-logpred_gp + logpred_triv))
305		# cost values
306	1	cost_mod = np.sum(resid_mod**2)
307	1	cost_triv = np.sum(resid_triv**2)
308	1	cost_gp = np.sum(resid_gp**2)
309		# chi^2 (variance corrected costs)
310	1	chisq_mod_ye = np.sum((resid_mod / errs)**2)
311	1	chisq_triv = np.sum((resid_triv / errs)**2)
312	1	chisq_gpcov = resid_mod.dot(linalg.solve(gpcov, resid_mod))
313		# adjust for degrees of freedom
314	1	cost_gp_dof = cost_gp / (ndat - ndim)
315	1	cost_mod_dof = cost_mod / (ndat - mdim)
316	1	cost_triv_dof = cost_triv / (ndat - 1)
317		# reduced chi^2
318	1	chisq_red_mod_ye = chisq_mod_ye / (ndat - mdim)
319	1	chisq_red_triv = chisq_triv / (ndat - 1)
320	1	chisq_red_gpcov = chisq_gpcov / (ndat - ndim)
321		# "generalized" R^2
322	1	logp_triv1 = np.sum(_log_prob(resid_triv, errs**2))
323	1	logp_triv2 = np.sum(_log_prob(resid_triv, var_triv))
324	1	logp_triv3 = np.sum(_log_prob(resid_triv, var_triv + errs**2))
325	1	log_lambda1 = test_logpred - logp_triv1
326	1	log_lambda2 = test_logpred - logp_triv2
327	1	log_lambda3 = test_logpred - logp_triv3
328	1	gen_rsq1a = 1 - np.exp(-2 * log_lambda1 / ndat)
329	1	gen_rsq1b = 1 - np.exp(-2 * log_lambda1 / (ndat - ndim))
330	1	gen_rsq2a = 1 - np.exp(-2 * log_lambda2 / ndat)
331	1	gen_rsq2b = 1 - np.exp(-2 * log_lambda2 / (ndat - ndim))
332	1	gen_rsq3a = 1 - np.exp(-2 * log_lambda3 / ndat)
333	1	gen_rsq3b = 1 - np.exp(-2 * log_lambda3 / (ndat - ndim))
334		# sent to the logger
335	1	logging.info("train max logpost: %s", samples_max_lp)
336	1	logging.info("test log_pred: %s", test_logpred)
337	1	logging.info("1a cost mean model: %s, dof adj: %s", cost_mod, cost_mod_dof)
338	1	logging.debug("1c cost gp predict: %s, dof adj: %s", cost_gp, cost_gp_dof)
339	1	logging.debug("1b cost triv model: %s, dof adj: %s", cost_triv, cost_triv_dof)
340	1	logging.info("1d var resid mean model: %s, gp model: %s, triv: %s",
341		var_mod, var_gp, var_triv)
342	1	logging.info("2a adjR2 mean model: %s, adjR2 gp predict: %s",
343		1 - cost_mod_dof / cost_triv_dof, 1 - cost_gp_dof / cost_triv_dof)
344	1	logging.info("2b red chi^2 mod: %s / triv: %s = %s",
345		chisq_red_mod_ye, chisq_red_triv, chisq_red_mod_ye / chisq_red_triv)
346	1	logging.info("2c red chi^2 mod (gp cov): %s / triv: %s = %s",
347		chisq_red_gpcov, chisq_red_triv, chisq_red_gpcov / chisq_red_triv)
348	1	logging.info("3a stand. red chi^2: %s", chisq_red_gpcov / chisq_red_triv)
349	1	logging.info("3b 1 - stand. red chi^2: %s",
350		1 - chisq_red_gpcov / chisq_red_triv)
351	1	logging.info("5a generalized R^2: 1a: %s, 1b: %s",
352		gen_rsq1a, gen_rsq1b)
353	1	logging.info("5b generalized R^2: 2a: %s, 2b: %s",
354		gen_rsq2a, gen_rsq2b)
355	1	logging.info("5c generalized R^2: 3a: %s, 3b: %s",
356		gen_rsq3a, gen_rsq3b)
357	1	try:
358		# celerite
359	1	logdet = model.solver.log_determinant()
360	1	except TypeError:
361		# george
362	1	logdet = model.solver.log_determinant
363		logging.debug("5 logdet: %s, const 2: %s", logdet, _const)
364

st-bender / sciapy

_log_prob_pt_samples_mt() A last analyzed 2023-01-30 12:22 UTC

Complexity

Size

Duplication

Code Coverage

Importance

Duplication Side-by-Side

Filter issues like

_log_prob_pt_samples_mt() A
last analyzed 2023-01-30 12:22 UTC