def run_dynesty(self, kwargs={}):
# dynesty solver
solver = DynamicNestedSampler(self._core_likelihood, self.prior,
len(self._activelist))
solver.run_nested(**kwargs)
results = solver.results
names = sorted(self._activelist)
for i in range(len(names)):
low, high = self.paramrange[names[i]]
results.samples[:, i] = umap(results.samples[:, i], [low, high])
return results
def inference(self, nlive, nworkers=1):
"""
Use Dynamic Nested Sampling to infer the posterior
"""
with Pool(nworkers) as pool:
sampler = DynamicNestedSampler(self.loglikelihood,
self.prior_transfom,
len(self.prior),
pool=pool,
nlive=nlive)
sampler.run_nested()
res = sampler.results
return res
def posterior_sampler(self, t, sample_idx, ndims=2):
nlive = 1024 # number of (initial) live points
bound = 'multi' # use MutliNest algorithm
sample = 'rwalk'
self.X = np.zeros(shape=(self.N, len(sample_idx)))
self.S = np.zeros(shape=(self.N, len(sample_idx)))
for i in range(self.N):
_, sample_trajectories = self.simulate(t, self.action, sample_idx)
self.X[i, :] = sample_trajectories[:, 0]
self.S[i, :] = sample_trajectories[:, 1]
dsampler = DynamicNestedSampler(self.loglikelihood,
self.prior_transform,
ndims,
bound=bound)
dsampler.run_nested(nlive_init=nlive)
res = dsampler.results
weights = np.exp(res['logwt'] - res['logz'][-1])
samples_dynesty = resample_equal(res.samples, weights)
return dsampler
def execute(self):
from dynesty import NestedSampler, DynamicNestedSampler
ndim = self.pipeline.nvaried
sampling_method = None if self.sample == "auto" else self.sample
if self.mode == "static":
sampler = NestedSampler(log_probability_function,
prior_transform,
ndim,
nlive=self.nlive,
bound=self.bound,
sample=sampling_method,
update_interval=self.update_interval,
first_update={
'min_ncall': self.min_ncall,
'min_eff': self.min_eff
},
queue_size=self.queue_size,
pool=self.pool)
sampler.run_nested(dlogz=self.dlogz)
else:
sampler = DynamicNestedSampler(
log_probability_function,
prior_transform,
ndim,
bound=self.bound,
sample=sampling_method,
# update_interval = self.update_interval,
queue_size=self.queue_size,
pool=self.pool)
sampler.run_nested(dlogz_init=self.dlogz)
results = sampler.results
results.summary()
for sample, logwt, logl in zip(results['samples'], results['logwt'],
results['logl']):
self.output.parameters(sample, logwt, logl)
self.output.final("ncall", results['ncall'])
self.output.final("efficiency", results['eff'])
self.output.final("log_z", results['logz'])
self.output.final("log_z_err", results['logzerr'])
self.converged = True
('b0', (0, 4000)),
('b1', (-2, 2)),
('b2', (-2, 2)),
('b3', (-2, 2)),
('b4', (-1500, 1500)),
('A', (0, 10)),
('nu0', (min_nu, max_nu)),
('w', (0, 200)),
('tau', (0, 100)),
('sigma', (0, 10))))
ndim = len(list(free_vars.keys()))
nu_c = (max_nu + min_nu) / 2.0
dsampler = DynamicNestedSampler(log_like, ptform, ndim, sample='rwalk')
dsampler.run_nested(dlogz_init=0.001, nlive_init=1000)
res = dsampler.results
weights = res['logwt']
weights -= np.max(weights)
# plt.plot(xdata, ydata, color='black', lw=1.5)
corner_weights = []
corner_vars = []
rms = []
for si, samp in enumerate(res['samples']):
if weights[si] < -7:
continue
corner_weights.append(np.exp(weights[si]))
def main(args=None):
if args.results == "none":
ndim = len(JOINT_PRIOR)
date = datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
image_positions = pd.read_csv(args.image_positions)
pprint(image_positions)
# A=0, B=1, C=2, D=3
x_image = image_positions["theta_x"].to_numpy()
y_image = image_positions["theta_y"].to_numpy()
quad_model = QuadPseudoNIELensModel(x_image, y_image, args.plate_scale)
time_delays = pd.read_csv(args.time_delays)
pprint(time_delays)
# Expected: DA=0, DB=1, DC=2 (index)
sigma_t = time_delays["sigma"].to_numpy() # units of days
delta_t = time_delays["delta_t"].to_numpy()
joint_loglikelihood = joint_loglikelihood_func(quad_model, delta_t,
sigma_t)
with Pool(args.nworkers) as pool:
sampler = DynamicNestedSampler(joint_loglikelihood,
joint_prior_transform,
ndim,
pool=pool,
nlive=args.nlive,
queue_size=args.nworkers)
sampler.run_nested()
res = sampler.results
# save results
with open(f"../results/joint_result_{date}.p", "wb") as f:
pickle.dump(res, f)
else: # we just want to plot the result from an older run
with open(args.results, "rb") as f:
res = pickle.load(f)
ndim = res.samples.shape[1]
if args.plot_results:
# trace plot
fig, axs = dyplot.traceplot(
res,
show_titles=True,
trace_cmap='plasma',
connect=True,
connect_highlight=range(5),
labels=LABELS,
)
fig.tight_layout(pad=2.0)
fig.savefig("../figures/joint_inference_trace_plot.png",
bbox_inches="tight")
# corner points plot
fig, axes = plt.subplots(ndim - 1, ndim - 1, figsize=(15, 15))
axes.reshape([ndim - 1, ndim - 1])
fg, ax = dyplot.cornerpoints(res,
cmap='plasma',
kde=False,
fig=(fig, axes),
labels=LABELS)
fg.savefig("../figures/joint_inference_cornerpoints.png",
bbox_inches="tight")
# corner plot
fig, axes = plt.subplots(ndim, ndim, figsize=(15, 15))
axes.reshape([ndim, ndim])
fg, ax = dyplot.cornerplot(res,
fig=(fig, axes),
color="b",
labels=LABELS,
show_titles=True)
fg.savefig("../figures/joint_inference_corner_plot.png",
bbox_inches="tight")
#### marginalized posterior #####
Ddt = res.samples[:, 0] # histogram of Ddt
weights = np.exp(
res['logwt'] -
res['logz'][-1]) # posterior probability (Bayes theorem)
# eliminate outliers with low and high + estimate confidance interval
low, fifth, median, ninety_fifth, high = weighted_quantile(
Ddt, [0.0001, 0.05, 0.5, 0.95, 0.9999], weights)
error_plus = ninety_fifth - median # error estimate with 5th percentile
error_minus = median - fifth # error estimate with 95th percentile
good = (Ddt > low) & (Ddt < high) # remove outliers
Ddt = Ddt[good]
weights = weights[good]
plt.figure(figsize=(8, 8))
plt.hist(Ddt, bins=100, weights=weights)
plt.title(r"$D_{\Delta t}$=%.2f$^{+%.2f}_{-%.2f}$" %
(median, error_plus, error_minus))
plt.xlabel(r"$D_{\Delta t}$")
plt.savefig("../figures/marginalized_posterior_Ddt.png")
# assume a flat LambdCDM model (with negligible radiation)
# We need to model kappa_ext for this step
def integrand(z):
return 1 / np.sqrt(args.omega_m * (1 + z)**3 + args.omega_l)
Dd = quad(integrand, 0, args.z_lens)[0] / (1 + args.z_lens)
Ds = quad(integrand, 0, args.z_source)[0] / (1 + args.z_source)
Dds = quad(integrand, args.z_lens,
args.z_source)[0] / (1 + args.z_source)
factor = (1 + args.z_lens) * Ds * Dd / Dds
H0 = (c * factor / Ddt / u.Mpc).to(u.km / u.s / u.Mpc).value
plt.figure(figsize=(8, 8))
fifth, median, ninety_fifth = weighted_quantile(
H0, [0.05, 0.5, 0.95], weights)
error_plus = ninety_fifth - median
error_minus = median - fifth
plt.hist(H0, bins=100, weights=weights)
plt.title(r"$H_0$=%.2f$^{+%.2f}_{-%.2f}$" %
(median, error_plus, error_minus))
plt.xlabel(r"$H_0$ [km s$^{-1}$ Mpc$^{-1}$]")
plt.savefig("../figures/marginalized_posterior_H0.png")
def sample(self, quiet=False):
"""
sample using the UltraNest numerical integration method
:rtype:
:returns:
"""
if not self._is_setup:
log.info("You forgot to setup the sampler!")
return
loud = not quiet
self._update_free_parameters()
param_names = list(self._free_parameters.keys())
ndim = len(param_names)
self._kwargs["ndim"] = ndim
loglike, dynesty_prior = self._construct_unitcube_posterior(return_copy=True)
# check if we are doing to do things in parallel
if threeML_config["parallel"]["use_parallel"]:
c = ParallelClient()
view = c[:]
self._kwargs["pool"] = view
self._kwargs["queue_size"] = len(view)
sampler = DynamicNestedSampler(loglike, dynesty_prior, **self._kwargs)
self._sampler_kwargs["print_progress"] = loud
with use_astromodels_memoization(False):
log.debug("Start dynestsy run")
sampler.run_nested(**self._sampler_kwargs)
log.debug("Dynesty run done")
self._sampler = sampler
results = self._sampler.results
# draw posterior samples
weights = np.exp(results["logwt"] - results["logz"][-1])
SQRTEPS = math.sqrt(float(np.finfo(np.float64).eps))
rstate = np.random
if abs(np.sum(weights) - 1.0) > SQRTEPS: # same tol as in np.random.choice.
raise ValueError("Weights do not sum to 1.")
# Make N subdivisions and choose positions with a consistent random offset.
nsamples = len(weights)
positions = (rstate.random() + np.arange(nsamples)) / nsamples
# Resample the data.
idx = np.zeros(nsamples, dtype=np.int)
cumulative_sum = np.cumsum(weights)
i, j = 0, 0
while i < nsamples:
if positions[i] < cumulative_sum[j]:
idx[i] = j
i += 1
else:
j += 1
samples_dynesty = results["samples"][idx]
self._raw_samples = samples_dynesty
# now do the same for the log likes
logl_dynesty = results["logl"][idx]
self._log_like_values = logl_dynesty
self._log_probability_values = self._log_like_values + np.array(
[self._log_prior(samples) for samples in self._raw_samples]
)
self._marginal_likelihood = self._sampler.results["logz"][-1] / np.log(10.0)
self._build_results()
# Display results
if loud:
self._results.display()
# now get the marginal likelihood
return self.samples
def run(self, walker_data):
"""Use nested sampling to determine posteriors."""
from dynesty import DynamicNestedSampler
from dynesty.dynamicsampler import stopping_function, weight_function
from mosfit.fitter import ln_likelihood, draw_from_icdf
prt = self._printer
if len(walker_data):
prt.message('nester_not_use_walkers', warning=True)
ndim = self._model._num_free_parameters
if self._num_walkers:
self._nwalkers = self._num_walkers
else:
self._nwalkers = 2 * ndim
self._nlive = 20 * ndim
self._lnprob = None
self._lnlike = None
prt.message('nmeas_nfree', [self._model._num_measurements, ndim])
nested_dlogz_init = self._cc
post_thresh = self._cc
max_iter = self._iterations if self._ct is None else np.inf
if max_iter <= 0:
return
s_exception = None
iter_denom = None if self._ct is not None else self._iterations
# Save a few things from the dynesty run for diagnostic purposes.
scales = []
try:
sampler = DynamicNestedSampler(ln_likelihood,
draw_from_icdf,
ndim,
pool=self._pool,
sample='rwalk',
queue_size=max(self._pool.size, 1))
# Perform initial sample.
ncall = sampler.ncall
self._niter = sampler.it - 1
for li, res in enumerate(
sampler.sample_initial(dlogz=nested_dlogz_init,
nlive=self._nlive)):
ncall0 = ncall
(worst, ustar, vstar, loglstar, logvol, logwt, self._logz,
logzvar, h, nc, worst_it, propidx, propiter, eff,
delta_logz) = res
ncall += nc
self._niter += 1
max_iter -= 1
if max_iter < 0:
break
if (self._fitter._maximum_walltime is not False and
self.time_running() > self._fitter._maximum_walltime):
prt.message('exceeded_walltime', warning=True)
break
self._results = sampler.results
scales.append(sampler.results.scale)
kmat = self._get_best_kmat()
# The above added 1 call.
ncall += 1
self._e_logz = np.sqrt(logzvar)
prt.status(self,
'baseline',
kmat=kmat,
iterations=[self._niter, iter_denom],
nc=ncall - ncall0,
ncall=ncall,
eff=eff,
logz=[
self._logz, self._e_logz, delta_logz,
nested_dlogz_init
],
loglstar=[loglstar],
time_running=self.time_running(),
maximum_walltime=self._fitter._maximum_walltime)
if max_iter >= 0:
prt.status(self,
'starting_batches',
kmat=kmat,
iterations=[self._niter, iter_denom],
nc=ncall - ncall0,
ncall=ncall,
eff=eff,
logz=[
self._logz, self._e_logz, delta_logz,
nested_dlogz_init
],
loglstar=[loglstar],
time_running=self.time_running(),
maximum_walltime=self._fitter._maximum_walltime)
n = 0
while max_iter >= 0:
n += 1
if (self._fitter._maximum_walltime is not False and
self.time_running() > self._fitter._maximum_walltime):
prt.message('exceeded_walltime', warning=True)
break
self._results = sampler.results
scales.append(sampler.results.scale)
stop, stop_vals = stopping_function(
self._results,
return_vals=True,
args={'post_thresh': post_thresh})
stop_post, stop_evid, stop_val = stop_vals
if not stop:
logl_bounds = weight_function(self._results)
self._logz, self._e_logz = self._results.logz[
-1], self._results.logzerr[-1]
for res in sampler.sample_batch(
logl_bounds=logl_bounds,
nlive_new=int(np.ceil(self._nlive / 2))):
(worst, ustar, vstar, loglstar, nc, worst_it, propidx,
propiter, eff) = res
ncall0 = ncall
ncall += nc
self._niter += 1
max_iter -= 1
self._results = sampler.results
kmat = self._get_best_kmat()
# The above added 1 call.
ncall += 1
prt.status(
self,
'batching',
kmat=kmat,
iterations=[self._niter, iter_denom],
batch=n,
nc=ncall - ncall0,
ncall=ncall,
eff=eff,
logz=[self._logz, self._e_logz],
loglstar=[
logl_bounds[0], loglstar, logl_bounds[1]
],
stop=stop_val,
time_running=self.time_running(),
maximum_walltime=self._fitter._maximum_walltime)
if max_iter < 0:
break
sampler.combine_runs()
else:
break
# self._results.summary()
# prt.nester_status(self, desc='sampling')
except (KeyboardInterrupt, SystemExit):
prt.message('ctrl_c', error=True, prefix=False, color='!r')
s_exception = sys.exc_info()
except Exception:
print('Scale history:')
print(scales)
pickle.dump(
sampler.results,
open(self._fitter._event_name + '-dynesty.pickle', 'wb'))
self._pool.close()
raise
if max_iter < 0:
prt.message('max_iter')
if s_exception is not None:
self._pool.close()
if (not prt.prompt('mc_interrupted')):
sys.exit()
sampler.reset()
gc.collect()
norm = -0.5 * M * LN2PI - M * LNSIGMA
# chi-squared (data, sigma and x are global variables defined early on in this notebook)
chisq = np.sum(((data - straight_line(x, m, c)) / sigma)**2)
return norm - 0.5 * chisq
nlive = 1024 # number of (initial) live points
bound = 'multi' # use MutliNest algorithm
sample = 'rwalk' # use the random walk to draw new samples
ndims = 2 # two parameters
dsampler = DynamicNestedSampler(loglikelihood_dynesty,
prior_transform,
ndims,
bound=bound,
sample=sample)
dsampler.run_nested(nlive_init=nlive)
dres = dsampler.results
dlogZdynesty = dres.logz[-1] # value of logZ
dlogZerrdynesty = dres.logzerr[
-1] # estimate of the statistcal uncertainty on logZ
# output marginal likelihood
print('Marginalised evidence (using dynamic sampler) is {} ± {}'.format(
dlogZdynesty, dlogZerrdynesty))
# get the posterior samples
dweights = np.exp(dres['logwt'] - dres['logz'][-1])
def extended_dynamic_nested_sampler(*args, **kwargs):
return DynamicNestedSampler(*args, **kwargs)
log_gamma_bg=log_bg,
C_bounds=[1., 10.],
log_C=False)
# =====================================================================
# if __name__ == '__main__':
# sample from the target distribution
t0 = time.time()
npool = 7
with ProcessPoolExecutor(max_workers=npool) as executor:
sampler = DynamicNestedSampler(infer.lnlike,
infer.prior_transform,
ndim,
pool=executor,
queue_size=npool,
bound='multi',
sample='unif')
sampler.run_nested(
dlogz_init=0.05,
nlive_init=1000,
nlive_batch=100,
maxiter=maxiter,
use_stop=False,
# wt_kwargs={'pfrac': 1.0}
)
res = sampler.results
pickle.dump(res, open(chain_file, "wb"))
m, c = theta # unpack the parameters
# normalisation
norm = -0.5*M*LN2PI - M*LNSIGMA
# chi-squared (data, sigma and x are global variables defined early on in this notebook)
chisq = np.sum(((data-straight_line(x, m, c))/sigma)**2)
return norm - 0.5*chisq
nlive = 1024 # number of (initial) live points
bound = 'multi' # use MutliNest algorithm
sample = 'rwalk' # use the random walk to draw new samples
ndims = 2 # two parameters
dsampler = DynamicNestedSampler(loglikelihood_dynesty, prior_transform, ndims,
bound=bound, sample=sample)
dsampler.run_nested(nlive_init=nlive)
dres = dsampler.results
dlogZdynesty = dres.logz[-1] # value of logZ
dlogZerrdynesty = dres.logzerr[-1] # estimate of the statistcal uncertainty on logZ
# output marginal likelihood
print('Marginalised evidence (using dynamic sampler) is {} ± {}'.format(dlogZdynesty, dlogZerrdynesty))
# get the posterior samples
dweights = np.exp(dres['logwt'] - dres['logz'][-1])
dpostsamples = resample_equal(dres.samples, dweights)
print('Number of posterior samples (using dynamic sampler) is {}'.format(dpostsamples.shape[0]))