def go(self,
continue_on_failure=True,
compute_covariance=False,
verbose=False,
**options_for_parallel_computation):
# Generate the data frame which will contain all results
self._continue_on_failure = continue_on_failure
self._compute_covariance = compute_covariance
# let's iterate, perform the fit and fill the data frame
if threeML_config["parallel"]["use_parallel"]:
# Parallel computation
with silence_console_log(and_progress_bars=False):
client = ParallelClient(**options_for_parallel_computation)
results = client.execute_with_progress_bar(
self.worker, list(range(self._n_iterations)))
else:
# Serial computation
results = []
with silence_console_log(and_progress_bars=False):
for i in trange(self._n_iterations,
desc="Goodness of fit computation"):
results.append(self.worker(i))
assert len(results) == self._n_iterations, (
"Something went wrong, I have %s results "
"for %s intervals" % (len(results), self._n_iterations))
# Store the results in the data frames
parameter_frames = pd.concat([x[0] for x in results],
keys=list(range(self._n_iterations)))
like_frames = pd.concat([x[1] for x in results],
keys=list(range(self._n_iterations)))
# Store a list with all results (this is a list of lists, each list contains the results for the different
# iterations for the same model)
self._all_results = []
for i in range(self._n_models):
this_model_results = [x[2][i] for x in results]
self._all_results.append(AnalysisResultsSet(this_model_results))
return parameter_frames, like_frames
def go(self, continue_on_failure=True, compute_covariance=False, verbose=False, **options_for_parallel_computation):
# Generate the data frame which will contain all results
if verbose:
log.setLevel(logging.INFO)
self._continue_on_failure = continue_on_failure
self._compute_covariance = compute_covariance
# let's iterate, perform the fit and fill the data frame
if threeML_config['parallel']['use-parallel']:
# Parallel computation
client = ParallelClient(**options_for_parallel_computation)
results = client.execute_with_progress_bar(self.worker, range(self._n_iterations))
else:
# Serial computation
results = []
with progress_bar(self._n_iterations, title='Goodness of fit computation') as p:
for i in range(self._n_iterations):
results.append(self.worker(i))
p.increase()
assert len(results) == self._n_iterations, "Something went wrong, I have %s results " \
"for %s intervals" % (len(results), self._n_iterations)
# Store the results in the data frames
parameter_frames = pd.concat(map(lambda x: x[0], results), keys=range(self._n_iterations))
like_frames = pd.concat(map(lambda x: x[1], results), keys=range(self._n_iterations))
# Store a list with all results (this is a list of lists, each list contains the results for the different
# iterations for the same model)
self._all_results = []
for i in range(self._n_models):
this_model_results = map(lambda x: x[2][i], results)
self._all_results.append(AnalysisResultsSet(this_model_results))
return parameter_frames, like_frames
def go(self):
if is_parallel_computation_active():
client = ParallelClient()
if self._n_decs % client.get_number_of_engines() != 0:
log.warning(
"The number of Dec bands is not a multiple of the number of engine. Make it so for optimal performances.",
RuntimeWarning)
res = client.execute_with_progress_bar(
self.worker,
list(range(len(self._points))),
chunk_size=self._n_ras)
else:
n_points = len(self._points)
p = tqdm(total=n_points)
res = np.zeros(n_points)
for i, point in enumerate(self._points):
res[i] = self.worker(i)
p.update(1)
TS = 2 * (-np.array(res) - self._like0)
#self._debug_map = {k:v for v,k in zip(self._points, TS)}
# Get maximum of TS
idx = TS.argmax()
self._max_ts = (TS[idx], self._points[idx])
log.info("Maximum TS is %.2f at (R.A., Dec) = (%.3f, %.3f)" %
(self._max_ts[0], self._max_ts[1][0], self._max_ts[1][1]))
self._ts_map = TS.reshape(self._n_decs, self._n_ras)
return self._ts_map
def sample(self, quiet=False):
if not self._is_setup:
log.info("You forgot to setup the sampler!")
return
loud = not quiet
self._update_free_parameters()
n_dim = len(list(self._free_parameters.keys()))
# Get starting point
p0 = emcee.State(self._get_starting_points(self._n_walkers))
# Deactivate memoization in astromodels, which is useless in this case since we will never use twice the
# same set of parameters
with use_astromodels_memoization(False):
if threeML_config["parallel"]["use_parallel"]:
c = ParallelClient()
view = c[:]
sampler = emcee.EnsembleSampler(self._n_walkers,
n_dim,
self.get_posterior,
pool=view)
else:
sampler = emcee.EnsembleSampler(self._n_walkers, n_dim,
self.get_posterior)
# If a seed is provided, set the random number seed
if self._seed is not None:
sampler._random.seed(self._seed)
log.debug("Start emcee run")
# Sample the burn-in
if threeML_config.interface.progress_bars:
if is_inside_notebook():
progress = "notebook"
else:
progress = True
else:
progress = False
pos, prob, state = sampler.run_mcmc(initial_state=p0,
nsteps=self._n_burn_in,
progress=progress)
log.debug("Emcee run done")
# Reset sampler
sampler.reset()
state = emcee.State(pos, prob, random_state=state)
# Run the true sampling
_ = sampler.run_mcmc(initial_state=state,
nsteps=self._n_iterations,
progress=progress)
acc = np.mean(sampler.acceptance_fraction)
log.info(f"Mean acceptance fraction: {acc}")
self._sampler = sampler
self._raw_samples = sampler.get_chain(flat=True)
# Compute the corresponding values of the likelihood
# First we need the prior
log_prior = [self._log_prior(x) for x in self._raw_samples]
# Now we get the log posterior and we remove the log prior
self._log_like_values = sampler.get_log_prob(flat=True) - log_prior
# we also want to store the log probability
self._log_probability_values = sampler.get_log_prob(flat=True)
self._marginal_likelihood = None
self._build_samples_dictionary()
self._build_results()
# Display results
if loud:
self._results.display()
return self.samples
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 get_contours(self, param_1, param_1_minimum, param_1_maximum, param_1_n_steps,
param_2=None, param_2_minimum=None, param_2_maximum=None, param_2_n_steps=None,
progress=True, **options):
"""
Generate confidence contours for the given parameters by stepping for the given number of steps between
the given boundaries. Call it specifying only source_1, param_1, param_1_minimum and param_1_maximum to
generate the profile of the likelihood for parameter 1. Specify all parameters to obtain instead a 2d
contour of param_1 vs param_2.
NOTE: if using parallel computation, param_1_n_steps must be an integer multiple of the number of running
engines. If that is not the case, the code will reduce the number of steps to match that requirement, and
issue a warning
:param param_1: fully qualified name of the first parameter or parameter instance
:param param_1_minimum: lower bound for the range for the first parameter
:param param_1_maximum: upper bound for the range for the first parameter
:param param_1_n_steps: number of steps for the first parameter
:param param_2: fully qualified name of the second parameter or parameter instance
:param param_2_minimum: lower bound for the range for the second parameter
:param param_2_maximum: upper bound for the range for the second parameter
:param param_2_n_steps: number of steps for the second parameter
:param progress: (True or False) whether to display progress or not
:param log: by default the steps are taken linearly. With this optional parameter you can provide a tuple of
booleans which specify whether the steps are to be taken logarithmically. For example,
'log=(True,False)' specify that the steps for the first parameter are to be taken logarithmically,
while they are linear for the second parameter. If you are generating the profile for only one
parameter, you can specify 'log=(True,)' or 'log=(False,)' (optional)
:return: a tuple containing an array corresponding to the steps for the first parameter, an array corresponding
to the steps for the second parameter (or None if stepping only in one direction), a matrix of size
param_1_steps x param_2_steps containing the value of the function at the corresponding points in the
grid. If param_2_steps is None (only one parameter), then this reduces to an array of
size param_1_steps.
"""
if hasattr(param_1,"value"):
# Substitute with the name
param_1 = param_1.path
if hasattr(param_2,'value'):
param_2 = param_2.path
# Check that the parameters exist
assert param_1 in self._likelihood_model.free_parameters, "Parameter %s is not a free parameters of the " \
"current model" % param_1
if param_2 is not None:
assert param_2 in self._likelihood_model.free_parameters, "Parameter %s is not a free parameters of the " \
"current model" % param_2
# Check that we have a valid fit
assert self._current_minimum is not None, "You have to run the .fit method before calling get_contours."
# Then restore the best fit
self._minimizer.restore_best_fit()
# Check minimal assumptions about the procedure
assert not (param_1 == param_2), "You have to specify two different parameters"
assert param_1_minimum < param_1_maximum, "Minimum larger than maximum for parameter 1"
min1, max1 = self.likelihood_model[param_1].bounds
if min1 is not None:
assert param_1_minimum >= min1, "Requested low range for parameter %s (%s) " \
"is below parameter minimum (%s)" % (param_1, param_1_minimum, min1)
if max1 is not None:
assert param_1_maximum <= max1, "Requested hi range for parameter %s (%s) " \
"is above parameter maximum (%s)" % (param_1, param_1_maximum, max1)
if param_2 is not None:
min2, max2 = self.likelihood_model[param_2].bounds
if min2 is not None:
assert param_2_minimum >= min2, "Requested low range for parameter %s (%s) " \
"is below parameter minimum (%s)" % (param_2, param_2_minimum, min2)
if max2 is not None:
assert param_2_maximum <= max2, "Requested hi range for parameter %s (%s) " \
"is above parameter maximum (%s)" % (param_2, param_2_maximum, max2)
# Check whether we are parallelizing or not
if not threeML_config['parallel']['use-parallel']:
a, b, cc = self.minimizer.contours(param_1, param_1_minimum, param_1_maximum, param_1_n_steps,
param_2, param_2_minimum, param_2_maximum, param_2_n_steps,
progress, **options)
# Collapse the second dimension of the results if we are doing a 1d contour
if param_2 is None:
cc = cc[:, 0]
else:
# With parallel computation
# In order to distribute fairly the computation, the strategy is to parallelize the computation
# by assigning to the engines one "line" of the grid at the time
# Connect to the engines
client = ParallelClient(**options)
# Get the number of engines
n_engines = client.get_number_of_engines()
# Check whether the number of threads is larger than the number of steps in the first direction
if n_engines > param_1_n_steps:
n_engines = int(param_1_n_steps)
custom_warnings.warn("The number of engines is larger than the number of steps. Using only %s engines."
% n_engines, ReducingNumberOfThreads)
# Check if the number of steps is divisible by the number
# of threads, otherwise issue a warning and make it so
if float(param_1_n_steps) % n_engines != 0:
# Set the number of steps to an integer multiple of the engines
# (note that // is the floor division, also called integer division)
param_1_n_steps = (param_1_n_steps // n_engines) * n_engines
custom_warnings.warn("Number of steps is not a multiple of the number of threads. Reducing steps to %s"
% param_1_n_steps, ReducingNumberOfSteps)
# Compute the number of splits, i.e., how many lines in the grid for each engine.
# (note that this is guaranteed to be an integer number after the previous checks)
p1_split_steps = param_1_n_steps // n_engines
# Prepare arrays for results
if param_2 is None:
# One array
pcc = np.zeros(param_1_n_steps)
pa = np.linspace(param_1_minimum, param_1_maximum, param_1_n_steps)
pb = None
else:
pcc = np.zeros((param_1_n_steps, param_2_n_steps))
# Prepare the two axes of the parameter space
pa = np.linspace(param_1_minimum, param_1_maximum, param_1_n_steps)
pb = np.linspace(param_2_minimum, param_2_maximum, param_2_n_steps)
# Define the parallel worker which will go through the computation
# NOTE: I only divide
# on the first parameter axis so that the different
# threads are more or less well mixed for points close and
# far from the best fit
def worker(start_index):
# Re-create the minimizer
backup_freeParameters = map(lambda x:x.value, self._likelihood_model.free_parameters.values())
this_minimizer = self._get_minimizer(self.minus_log_like_profile,
self._free_parameters)
this_p1min = pa[start_index * p1_split_steps]
this_p1max = pa[(start_index + 1) * p1_split_steps - 1]
# print("From %s to %s" % (this_p1min, this_p1max))
aa, bb, ccc = this_minimizer.contours(param_1, this_p1min, this_p1max, p1_split_steps,
param_2, param_2_minimum, param_2_maximum,
param_2_n_steps,
progress=True, **options)
# Restore best fit values
for val, par in zip(backup_freeParameters, self._likelihood_model.free_parameters.values()):
par.value = val
return ccc
# Now re-assemble the vector of results taking the different parts from the engines
all_results = client.execute_with_progress_bar(worker, range(n_engines), chunk_size=1)
for i, these_results in enumerate(all_results):
if param_2 is None:
pcc[i * p1_split_steps: (i + 1) * p1_split_steps] = these_results[:, 0]
else:
pcc[i * p1_split_steps: (i + 1) * p1_split_steps, :] = these_results
# Give the results the names that the following code expect. These are kept separate for debugging
# purposes
cc = pcc
a = pa
b = pb
# Here we have done the computation, in parallel computation or not. Let's make the plot
# with the contour
if param_2 is not None:
# 2d contour
fig = self._plot_contours("%s" % (param_1), a, "%s" % (param_2,), b, cc)
else:
# 1d contour (i.e., a profile)
fig = self._plot_profile("%s" % (param_1), a, cc)
# Check if we found a better minimum. This shouldn't happen, but in case of very difficult fit
# it might.
if self._current_minimum - cc.min() > 0.1:
if param_2 is not None:
idx = cc.argmin()
aidx, bidx = np.unravel_index(idx, cc.shape)
print("\nFound a better minimum: %s with %s = %s and %s = %s. Run again your fit starting from here."
% (cc.min(), param_1, a[aidx], param_2, b[bidx]))
else:
idx = cc.argmin()
print("Found a better minimum: %s with %s = %s. Run again your fit starting from here."
% (cc.min(), param_1, a[idx]))
return a, b, cc, fig
def _unbinned_fit_polynomials(self, bayes=False):
self._poly_fit_exists = True
# Select all the events that are in the background regions
# and make a mask
all_bkg_masks = []
total_duration = 0.0
poly_exposure = 0
for selection in self._poly_intervals:
total_duration += selection.duration
poly_exposure += self.exposure_over_interval(
selection.start_time, selection.stop_time)
all_bkg_masks.append(
np.logical_and(
self._arrival_times >= selection.start_time,
self._arrival_times <= selection.stop_time,
))
poly_mask = all_bkg_masks[0]
# If there are multiple masks:
if len(all_bkg_masks) > 1:
for mask in all_bkg_masks[1:]:
poly_mask = np.logical_or(poly_mask, mask)
# Select the all the events in the poly selections
# We only need to do this once
total_poly_events = self._arrival_times[poly_mask]
# For the channel energies we will need to down select again.
# We can go ahead and do this to avoid repeated computations
total_poly_energies = self._measurement[poly_mask]
# Now we will find the the best poly order unless the use specified one
# The total cnts (over channels) is binned to .1 sec intervals
if self._user_poly_order == -1:
self._optimal_polynomial_grade = (
self._unbinned_fit_global_and_determine_optimum_grade(
total_poly_events, poly_exposure, bayes=bayes))
log.info("Auto-determined polynomial order: %d" %
self._optimal_polynomial_grade)
else:
self._optimal_polynomial_grade = self._user_poly_order
channels = list(
range(self._first_channel, self._n_channels + self._first_channel))
# Check whether we are parallelizing or not
t_start = self._poly_intervals.start_times
t_stop = self._poly_intervals.stop_times
if threeML_config["parallel"]["use_parallel"]:
def worker(channel):
channel_mask = total_poly_energies == channel
# Mask background events and current channel
# poly_chan_mask = np.logical_and(poly_mask, channel_mask)
# Select the masked events
current_events = total_poly_events[channel_mask]
polynomial, _ = unbinned_polyfit(
current_events,
self._optimal_polynomial_grade,
t_start,
t_stop,
poly_exposure,
bayes=bayes)
return polynomial
client = ParallelClient()
polynomials = client.execute_with_progress_bar(
worker,
channels,
name=f"Fitting {self._instrument} background")
else:
polynomials = []
for channel in tqdm(channels,
desc=f"Fitting {self._instrument} background"):
channel_mask = total_poly_energies == channel
# Mask background events and current channel
# poly_chan_mask = np.logical_and(poly_mask, channel_mask)
# Select the masked events
current_events = total_poly_events[channel_mask]
polynomial, _ = unbinned_polyfit(
current_events,
self._optimal_polynomial_grade,
t_start,
t_stop,
poly_exposure,
bayes=bayes)
polynomials.append(polynomial)
# We are now ready to return the polynomials
self._polynomials = polynomials
def sample(self, quiet=False):
assert self._is_setup, "You forgot to setup the sampler!"
loud = not quiet
self._update_free_parameters()
n_dim = len(list(self._free_parameters.keys()))
# Get starting point
p0 = self._get_starting_points(1)[0]
print(p0)
# Deactivate memoization in astromodels, which is useless in this case since we will never use twice the
# same set of parameters
with use_astromodels_memoization(False):
if threeML_config["parallel"]["use-parallel"]:
c = ParallelClient()
view = c[:]
pool = view
else:
pool = None
def logp(theta):
return self.get_posterior(theta)
def grad(theta):
return numerical_grad(theta, self.get_posterior)
nuts_fn = nuts.NutsSampler_fn_wrapper(self.get_posterior, grad)
samples, lnprob, epsilon = nuts.nuts6(nuts_fn, self._n_iterations, self._n_adapt, p0)
# sampler = nuts.NUTSSampler(n_dim, self.get_posterior, gradfn=None)
# # if a seed is provided, set the random number seed
# if self._seed is not None:
# sampler._random.seed(self._seed)
# # Run the true sampling
# samples = sampler.run_mcmc(
# initial_state=p0,
# M=self._n_iterations,
# Madapt=self._n_adapt,
# delta=self._delta,
# progress=loud,
# )
self._sampler = None
self._raw_samples = samples
# Compute the corresponding values of the likelihood
self._test = lnprob
# First we need the prior
log_prior = np.array([self._log_prior(x) for x in self._raw_samples])
# Now we get the log posterior and we remove the log prior
self._log_like_values = np.array([self._log_like(x) for x in self._raw_samples])
# we also want to store the log probability
self._log_probability_values = log_prior + self._log_like_values
self._marginal_likelihood = None
self._build_samples_dictionary()
self._build_results()
# Display results
if loud:
self._results.display()
return self.samples
def sample(self, n_walkers, burn_in, n_samples, quiet=False, seed=None):
"""
Sample the posterior with the Goodman & Weare's Affine Invariant Markov chain Monte Carlo
:param n_walkers:
:param burn_in:
:param n_samples:
:param quiet: if False, do not print results
:param seed: if provided, it is used to seed the random numbers generator before the MCMC
:return: MCMC samples
"""
self._update_free_parameters()
n_dim = len(self._free_parameters.keys())
# Get starting point
p0 = self._get_starting_points(n_walkers)
sampling_procedure = sample_with_progress
# Deactivate memoization in astromodels, which is useless in this case since we will never use twice the
# same set of parameters
with use_astromodels_memoization(False):
if threeML_config['parallel']['use-parallel']:
c = ParallelClient()
view = c[:]
sampler = emcee.EnsembleSampler(n_walkers,
n_dim,
self.get_posterior,
pool=view)
# Sampling with progress in parallel is super-slow, so let's
# use the non-interactive one
sampling_procedure = sample_without_progress
else:
sampler = emcee.EnsembleSampler(n_walkers, n_dim,
self.get_posterior)
# If a seed is provided, set the random number seed
if seed is not None:
sampler._random.seed(seed)
# Sample the burn-in
pos, prob, state = sampling_procedure(title="Burn-in",
p0=p0,
sampler=sampler,
n_samples=burn_in)
# Reset sampler
sampler.reset()
# Run the true sampling
_ = sampling_procedure(title="Sampling",
p0=pos,
sampler=sampler,
n_samples=n_samples,
rstate0=state)
acc = np.mean(sampler.acceptance_fraction)
print("\nMean acceptance fraction: %s\n" % acc)
self._sampler = sampler
self._raw_samples = sampler.flatchain
# Compute the corresponding values of the likelihood
# First we need the prior
log_prior = map(lambda x: self._log_prior(x), self._raw_samples)
# Now we get the log posterior and we remove the log prior
self._log_like_values = sampler.flatlnprobability - log_prior
# we also want to store the log probability
self._log_probability_values = sampler.flatlnprobability
self._marginal_likelihood = None
self._build_samples_dictionary()
self._build_results()
# Display results
if not quiet:
self._results.display()
return self.samples
def sample(self, quiet=False):
if not self._is_setup:
log.info("You forgot to setup the sampler!")
return
loud = not quiet
self._update_free_parameters()
n_dim = len(list(self._free_parameters.keys()))
# Get starting point
p0 = self._get_starting_points(self._n_walkers)
# Deactivate memoization in astromodels, which is useless in this case since we will never use twice the
# same set of parameters
with use_astromodels_memoization(False):
if using_mpi:
with MPIPoolExecutor() as executor:
sampler = zeus.sampler(
logprob_fn=self.get_posterior,
nwalkers=self._n_walkers,
ndim=n_dim,
pool=executor,
)
# if self._seed is not None:
# sampler._random.seed(self._seed)
# Run the true sampling
log.debug("Start zeus run")
_ = sampler.run(
p0,
self._n_iterations + self._n_burn_in,
progress=loud,
)
log.debug("Zeus run done")
elif threeML_config["parallel"]["use_parallel"]:
c = ParallelClient()
view = c[:]
sampler = zeus.sampler(
logprob_fn=self.get_posterior,
nwalkers=self._n_walkers,
ndim=n_dim,
pool=view,
)
else:
sampler = zeus.sampler(logprob_fn=self.get_posterior,
nwalkers=self._n_walkers,
ndim=n_dim)
# If a seed is provided, set the random number seed
# if self._seed is not None:
# sampler._random.seed(self._seed)
# Sample the burn-in
if not using_mpi:
log.debug("Start zeus run")
_ = sampler.run(p0,
self._n_iterations + self._n_burn_in,
progress=loud)
log.debug("Zeus run done")
self._sampler = sampler
self._raw_samples = sampler.get_chain(flat=True,
discard=self._n_burn_in)
# Compute the corresponding values of the likelihood
# First we need the prior
log_prior = np.array([self._log_prior(x) for x in self._raw_samples])
self._log_probability_values = sampler.get_log_prob(
flat=True, discard=self._n_burn_in)
# np.array(
# [self.get_posterior(x) for x in self._raw_samples]
# )
# Now we get the log posterior and we remove the log prior
self._log_like_values = self._log_probability_values - log_prior
# we also want to store the log probability
self._marginal_likelihood = None
self._build_samples_dictionary()
self._build_results()
# Display results
if loud:
print(self._sampler.summary)
self._results.display()
return self.samples
def contours(
self,
param_1,
param_1_minimum,
param_1_maximum,
param_1_n_steps,
param_2=None,
param_2_minimum=None,
param_2_maximum=None,
param_2_n_steps=None,
progress=True,
**options
):
"""
Generate confidence contours for the given parameters by stepping for the given number of steps between
the given boundaries. Call it specifying only source_1, param_1, param_1_minimum and param_1_maximum to
generate the profile of the likelihood for parameter 1. Specify all parameters to obtain instead a 2d
contour of param_1 vs param_2
:param param_1: name of the first parameter
:param param_1_minimum: lower bound for the range for the first parameter
:param param_1_maximum: upper bound for the range for the first parameter
:param param_1_n_steps: number of steps for the first parameter
:param param_2: name of the second parameter
:param param_2_minimum: lower bound for the range for the second parameter
:param param_2_maximum: upper bound for the range for the second parameter
:param param_2_n_steps: number of steps for the second parameter
:param progress: (True or False) whether to display progress or not
:param log: by default the steps are taken linearly. With this optional parameter you can provide a tuple of
booleans which specify whether the steps are to be taken logarithmically. For example,
'log=(True,False)' specify that the steps for the first parameter are to be taken logarithmically, while they
are linear for the second parameter. If you are generating the profile for only one parameter, you can specify
'log=(True,)' or 'log=(False,)' (optional)
:param: parallel: whether to use or not parallel computation (default:False)
:return: a : an array corresponding to the steps for the first parameter
b : an array corresponding to the steps for the second parameter (or None if stepping only in one
direction)
contour : a matrix of size param_1_steps x param_2_steps containing the value of the function at the
corresponding points in the grid. If param_2_steps is None (only one parameter), then this reduces to
an array of size param_1_steps.
"""
# Figure out if we are making a 1d or a 2d contour
if param_2 is None:
n_dimensions = 1
else:
n_dimensions = 2
# Check the options
p1log = False
p2log = False
parallel = False
if "log" in options.keys():
assert len(options["log"]) == n_dimensions, (
"When specifying the 'log' option you have to provide a "
+ "boolean for each dimension you are stepping on."
)
p1log = bool(options["log"][0])
if param_2 is not None:
p2log = bool(options["log"][1])
if "parallel" in options.keys():
parallel = bool(options["parallel"])
# Generate the steps
if p1log:
param_1_steps = numpy.logspace(math.log10(param_1_minimum), math.log10(param_1_maximum), param_1_n_steps)
else:
param_1_steps = numpy.linspace(param_1_minimum, param_1_maximum, param_1_n_steps)
if n_dimensions == 2:
if p2log:
param_2_steps = numpy.logspace(
math.log10(param_2_minimum), math.log10(param_2_maximum), param_2_n_steps
)
else:
param_2_steps = numpy.linspace(param_2_minimum, param_2_maximum, param_2_n_steps)
else:
# Only one parameter to step through
# Put param_2_steps as nan so that the worker can realize that it does not have
# to step through it
param_2_steps = numpy.array([numpy.nan])
# Generate the grid
grid = cartesian([param_1_steps, param_2_steps])
# Define the worker which will compute the value of the function at a given point in the grid
# Restore best fit
self._restore_best_fit()
# Duplicate the options used for the original minimizer
new_args = dict(self.minuit.fitarg)
# Get the minuit names for the parameters
minuit_param_1 = self._parameter_name_to_minuit_name(param_1)
if param_2 is None:
minuit_param_2 = None
else:
minuit_param_2 = self._parameter_name_to_minuit_name(param_2)
# Instance the worker
contour_worker = ContourWorker(
self._f, self.minuit.values, new_args, minuit_param_1, minuit_param_2, self.name_to_position
)
# We are finally ready to do the computation
# Serial and parallel computation are slightly different, so check whether we are in one case
# or the other
if not parallel:
# Serial computation
if progress:
# Computation with progress bar
progress_bar = ProgressBar(grid.shape[0])
# Define a wrapper which will increase the progress before as well as run the actual computation
def wrap(args):
results = contour_worker(args)
progress_bar.increase()
return results
# Do the computation
results = map(wrap, grid)
else:
# Computation without the progress bar
results = map(contour_worker, grid)
else:
# Parallel computation
# Connect to the engines
client = ParallelClient(**options)
# Get a balanced view of the engines
load_balance_view = client.load_balanced_view()
# Distribute the work among the engines and start it, but return immediately the control
# to the main thread
amr = load_balance_view.map_async(contour_worker, grid)
# print progress
n_points = grid.flatten().shape[0]
progress = ProgressBar(n_points)
# This loop will check from time to time the status of the computation, which is happening on
# different threads, and update the progress bar
while not amr.ready():
# Check and report the status of the computation every second
time.sleep(1)
# if (debug):
# stdouts = amr.stdout
#
# # clear_output doesn't do much in terminal environments
# for stdout, stderr in zip(amr.stdout, amr.stderr):
# if stdout:
# print "%s" % (stdout[-1000:])
# if stderr:
# print "%s" % (stderr[-1000:])
# sys.stdout.flush()
progress.animate(amr.progress - 1)
# If there have been problems, here is where they will be raised
results = amr.get()
# Always display 100% at the end
progress.animate(n_points)
# Add a new line after the progress bar
print("\n")
# Return results
return (
param_1_steps,
param_2_steps,
numpy.array(results).reshape((param_1_steps.shape[0], param_2_steps.shape[0])),
)
def _unbinned_fit_global_and_determine_optimum_grade(
self, events, exposure, bayes=False):
"""
Provides the ability to find the optimum polynomial grade for *unbinned* events by fitting the
total (all channels) to 0-2 order polynomials and then comparing them via a likelihood ratio test.
:param events: an event list
:param exposure: the exposure per event
:return: polynomial grade
"""
# Fit the sum of all the channels to determine the optimal polynomial
# grade
min_grade = 0
max_grade = 2
log_likelihoods = []
t_start = self._poly_intervals.start_times
t_stop = self._poly_intervals.stop_times
log.debug("attempting to find best fit poly with unbinned")
if threeML_config["parallel"]["use_parallel"]:
def worker(grade):
polynomial, log_like = unbinned_polyfit(events,
grade,
t_start,
t_stop,
exposure,
bayes=bayes)
return log_like
client = ParallelClient()
log_likelihoods = client.execute_with_progress_bar(
worker,
list(range(min_grade, max_grade + 1)),
name="Finding best polynomial Order")
else:
for grade in trange(min_grade,
max_grade + 1,
desc="Finding best polynomial Order"):
polynomial, log_like = unbinned_polyfit(events,
grade,
t_start,
t_stop,
exposure,
bayes=bayes)
log_likelihoods.append(log_like)
# Found the best one
delta_loglike = np.array([
2 * (x[0] - x[1])
for x in zip(log_likelihoods[:-1], log_likelihoods[1:])
])
log.debug(f"log likes {log_likelihoods}")
log.debug(f" delta loglikes {delta_loglike}")
delta_threshold = 9.0
mask = delta_loglike >= delta_threshold
if len(mask.nonzero()[0]) == 0:
# best grade is zero!
best_grade = 0
else:
best_grade = mask.nonzero()[0][-1] + 1
return best_grade
def _fit_global_and_determine_optimum_grade(self,
cnts,
bins,
exposure,
bayes=False):
"""
Provides the ability to find the optimum polynomial grade for *binned* counts by fitting the
total (all channels) to 0-4 order polynomials and then comparing them via a likelihood ratio test.
:param cnts: counts per bin
:param bins: the bins used
:param exposure: exposure per bin
:param bayes:
:return: polynomial grade
"""
min_grade = 0
max_grade = 4
log_likelihoods = []
log.debug("attempting to find best poly with binned data")
if threeML_config["parallel"]["use_parallel"]:
def worker(grade):
polynomial, log_like = polyfit(bins,
cnts,
grade,
exposure,
bayes=bayes)
return log_like
client = ParallelClient()
log_likelihoods = client.execute_with_progress_bar(
worker,
list(range(min_grade, max_grade + 1)),
name="Finding best polynomial Order")
else:
for grade in trange(min_grade,
max_grade + 1,
desc="Finding best polynomial Order"):
polynomial, log_like = polyfit(bins,
cnts,
grade,
exposure,
bayes=bayes)
log_likelihoods.append(log_like)
# Found the best one
delta_loglike = np.array([
2 * (x[0] - x[1])
for x in zip(log_likelihoods[:-1], log_likelihoods[1:])
])
log.debug(f"log likes {log_likelihoods}")
log.debug(f" delta loglikes {delta_loglike}")
delta_threshold = 9.0
mask = delta_loglike >= delta_threshold
if len(mask.nonzero()[0]) == 0:
# best grade is zero!
best_grade = 0
else:
best_grade = mask.nonzero()[0][-1] + 1
return best_grade
def _fit_polynomials(self, bayes=False):
"""
fits a polynomial to all channels over the input time intervals
:param fit_intervals: str input intervals
:return:
"""
# mark that we have fit a poly now
self._poly_fit_exists = True
# we need to adjust the selection to the true intervals of the time-binned spectra
tmp_poly_intervals = self._poly_intervals
poly_intervals = self._adjust_to_true_intervals(tmp_poly_intervals)
self._poly_intervals = poly_intervals
# now lets get all the counts, exposure and midpoints for the
# selection
selected_counts = []
selected_exposure = []
selected_midpoints = []
for selection in poly_intervals:
# get the mask of these bins
mask = self._select_bins(selection.start_time, selection.stop_time)
# the counts will be (time, channel) here,
# so the mask is selecting time.
# a sum along axis=0 is a sum in time, while axis=1 is a sum in energy
selected_counts.extend(
self._binned_spectrum_set.counts_per_bin[mask])
selected_exposure.extend(
self._binned_spectrum_set.exposure_per_bin[mask])
selected_midpoints.extend(
self._binned_spectrum_set.time_intervals.mid_points[mask]
)
selected_counts = np.array(selected_counts)
selected_midpoints = np.array(selected_midpoints)
selected_exposure = np.array(selected_exposure)
# Now we will find the the best poly order unless the use specified one
# The total cnts (over channels) is binned
if self._user_poly_order == -1:
self._optimal_polynomial_grade = (
self._fit_global_and_determine_optimum_grade(
selected_counts.sum(axis=1),
selected_midpoints,
selected_exposure,
bayes=bayes,
)
)
log.info(
"Auto-determined polynomial order: %d"
% self._optimal_polynomial_grade
)
else:
self._optimal_polynomial_grade = self._user_poly_order
if threeML_config["parallel"]["use_parallel"]:
def worker(counts):
with silence_console_log():
polynomial, _ = polyfit(
selected_midpoints,
counts,
self._optimal_polynomial_grade,
selected_exposure,
bayes=bayes,
)
return polynomial
client = ParallelClient()
polynomials = client.execute_with_progress_bar(
worker, selected_counts.T, name=f"Fitting {self._instrument} background")
else:
polynomials = []
# now fit the light curve of each channel
# and save the estimated polynomial
for counts in tqdm(
selected_counts.T, desc=f"Fitting {self._instrument} background"
):
with silence_console_log():
polynomial, _ = polyfit(
selected_midpoints,
counts,
self._optimal_polynomial_grade,
selected_exposure,
bayes=bayes,
)
polynomials.append(polynomial)
self._polynomials = polynomials
def _fit_polynomials(self, bayes=False):
"""
Binned fit to each channel. Sets the polynomial array that will be used to compute
counts over an interval
:return:
"""
self._poly_fit_exists = True
# Select all the events that are in the background regions
# and make a mask
all_bkg_masks = []
for selection in self._poly_intervals:
all_bkg_masks.append(
np.logical_and(
self._arrival_times >= selection.start_time,
self._arrival_times <= selection.stop_time,
))
poly_mask = all_bkg_masks[0]
# If there are multiple masks:
if len(all_bkg_masks) > 1:
for mask in all_bkg_masks[1:]:
poly_mask = np.logical_or(poly_mask, mask)
# Select the all the events in the poly selections
# We only need to do this once
total_poly_events = self._arrival_times[poly_mask]
# For the channel energies we will need to down select again.
# We can go ahead and do this to avoid repeated computations
total_poly_energies = self._measurement[poly_mask]
# This calculation removes the unselected portion of the light curve
# so that we are not fitting zero counts. It will be used in the channel calculations
# as well
bin_width = 1.0 # seconds
these_bins = np.arange(self._start_time, self._stop_time, bin_width)
cnts, bins = np.histogram(total_poly_events, bins=these_bins)
# Find the mean time of the bins and calculate the exposure in each bin
mean_time = []
exposure_per_bin = []
for i in range(len(bins) - 1):
m = np.mean((bins[i], bins[i + 1]))
mean_time.append(m)
exposure_per_bin.append(
self.exposure_over_interval(bins[i], bins[i + 1]))
mean_time = np.array(mean_time)
exposure_per_bin = np.array(exposure_per_bin)
# Remove bins with zero counts
all_non_zero_mask = []
for selection in self._poly_intervals:
all_non_zero_mask.append(
np.logical_and(mean_time >= selection.start_time,
mean_time <= selection.stop_time))
non_zero_mask = all_non_zero_mask[0]
if len(all_non_zero_mask) > 1:
for mask in all_non_zero_mask[1:]:
non_zero_mask = np.logical_or(mask, non_zero_mask)
# Now we will find the the best poly order unless the use specified one
# The total cnts (over channels) is binned to .1 sec intervals
if self._user_poly_order == -1:
self._optimal_polynomial_grade = (
self._fit_global_and_determine_optimum_grade(
cnts[non_zero_mask],
mean_time[non_zero_mask],
exposure_per_bin[non_zero_mask],
bayes=bayes))
log.info("Auto-determined polynomial order: %d" %
self._optimal_polynomial_grade)
else:
self._optimal_polynomial_grade = self._user_poly_order
channels = list(
range(self._first_channel, self._n_channels + self._first_channel))
if threeML_config["parallel"]["use_parallel"]:
def worker(channel):
channel_mask = total_poly_energies == channel
# Mask background events and current channel
# poly_chan_mask = np.logical_and(poly_mask, channel_mask)
# Select the masked events
current_events = total_poly_events[channel_mask]
cnts, bins = np.histogram(current_events, bins=these_bins)
polynomial, _ = polyfit(mean_time[non_zero_mask],
cnts[non_zero_mask],
self._optimal_polynomial_grade,
exposure_per_bin[non_zero_mask],
bayes=bayes)
return polynomial
client = ParallelClient()
polynomials = client.execute_with_progress_bar(
worker,
channels,
name=f"Fitting {self._instrument} background")
else:
polynomials = []
for channel in tqdm(channels,
desc=f"Fitting {self._instrument} background"):
channel_mask = total_poly_energies == channel
# Mask background events and current channel
# poly_chan_mask = np.logical_and(poly_mask, channel_mask)
# Select the masked events
current_events = total_poly_events[channel_mask]
# now bin the selected channel counts
cnts, bins = np.histogram(current_events, bins=these_bins)
# Put data to fit in an x vector and y vector
polynomial, _ = polyfit(mean_time[non_zero_mask],
cnts[non_zero_mask],
self._optimal_polynomial_grade,
exposure_per_bin[non_zero_mask],
bayes=bayes)
polynomials.append(polynomial)
# We are now ready to return the polynomials
self._polynomials = polynomials
def get_contours(self,
param_1,
param_1_minimum,
param_1_maximum,
param_1_n_steps,
param_2=None,
param_2_minimum=None,
param_2_maximum=None,
param_2_n_steps=None,
progress=True,
**options):
"""
Generate confidence contours for the given parameters by stepping for the given number of steps between
the given boundaries. Call it specifying only source_1, param_1, param_1_minimum and param_1_maximum to
generate the profile of the likelihood for parameter 1. Specify all parameters to obtain instead a 2d
contour of param_1 vs param_2.
NOTE: if using parallel computation, param_1_n_steps must be an integer multiple of the number of running
engines. If that is not the case, the code will reduce the number of steps to match that requirement, and
issue a warning
:param param_1: fully qualified name of the first parameter or parameter instance
:param param_1_minimum: lower bound for the range for the first parameter
:param param_1_maximum: upper bound for the range for the first parameter
:param param_1_n_steps: number of steps for the first parameter
:param param_2: fully qualified name of the second parameter or parameter instance
:param param_2_minimum: lower bound for the range for the second parameter
:param param_2_maximum: upper bound for the range for the second parameter
:param param_2_n_steps: number of steps for the second parameter
:param progress: (True or False) whether to display progress or not
:param log: by default the steps are taken linearly. With this optional parameter you can provide a tuple of
booleans which specify whether the steps are to be taken logarithmically. For example,
'log=(True,False)' specify that the steps for the first parameter are to be taken logarithmically,
while they are linear for the second parameter. If you are generating the profile for only one
parameter, you can specify 'log=(True,)' or 'log=(False,)' (optional)
:return: a tuple containing an array corresponding to the steps for the first parameter, an array corresponding
to the steps for the second parameter (or None if stepping only in one direction), a matrix of size
param_1_steps x param_2_steps containing the value of the function at the corresponding points in the
grid. If param_2_steps is None (only one parameter), then this reduces to an array of
size param_1_steps.
"""
if hasattr(param_1, "value"):
# Substitute with the name
param_1 = param_1.path
if hasattr(param_2, "value"):
param_2 = param_2.path
# Check that the parameters exist
assert param_1 in self._likelihood_model.free_parameters, (
"Parameter %s is not a free parameters of the "
"current model" % param_1)
if param_2 is not None:
assert param_2 in self._likelihood_model.free_parameters, (
"Parameter %s is not a free parameters of the "
"current model" % param_2)
# Check that we have a valid fit
assert (
self._current_minimum is not None
), "You have to run the .fit method before calling get_contours."
# Then restore the best fit
self._minimizer.restore_best_fit()
# Check minimal assumptions about the procedure
assert not (param_1
== param_2), "You have to specify two different parameters"
assert (param_1_minimum <
param_1_maximum), "Minimum larger than maximum for parameter 1"
min1, max1 = self.likelihood_model[param_1].bounds
if min1 is not None:
assert param_1_minimum >= min1, (
"Requested low range for parameter %s (%s) "
"is below parameter minimum (%s)" %
(param_1, param_1_minimum, min1))
if max1 is not None:
assert param_1_maximum <= max1, (
"Requested hi range for parameter %s (%s) "
"is above parameter maximum (%s)" %
(param_1, param_1_maximum, max1))
if param_2 is not None:
min2, max2 = self.likelihood_model[param_2].bounds
if min2 is not None:
assert param_2_minimum >= min2, (
"Requested low range for parameter %s (%s) "
"is below parameter minimum (%s)" %
(param_2, param_2_minimum, min2))
if max2 is not None:
assert param_2_maximum <= max2, (
"Requested hi range for parameter %s (%s) "
"is above parameter maximum (%s)" %
(param_2, param_2_maximum, max2))
# Check whether we are parallelizing or not
if not threeML_config["parallel"]["use-parallel"]:
a, b, cc = self.minimizer.contours(
param_1, param_1_minimum, param_1_maximum, param_1_n_steps,
param_2, param_2_minimum, param_2_maximum, param_2_n_steps,
progress, **options)
# Collapse the second dimension of the results if we are doing a 1d contour
if param_2 is None:
cc = cc[:, 0]
else:
# With parallel computation
# In order to distribute fairly the computation, the strategy is to parallelize the computation
# by assigning to the engines one "line" of the grid at the time
# Connect to the engines
client = ParallelClient(**options)
# Get the number of engines
n_engines = client.get_number_of_engines()
# Check whether the number of threads is larger than the number of steps in the first direction
if n_engines > param_1_n_steps:
n_engines = int(param_1_n_steps)
custom_warnings.warn(
"The number of engines is larger than the number of steps. Using only %s engines."
% n_engines,
ReducingNumberOfThreads,
)
# Check if the number of steps is divisible by the number
# of threads, otherwise issue a warning and make it so
if float(param_1_n_steps) % n_engines != 0:
# Set the number of steps to an integer multiple of the engines
# (note that // is the floor division, also called integer division)
param_1_n_steps = (param_1_n_steps // n_engines) * n_engines
custom_warnings.warn(
"Number of steps is not a multiple of the number of threads. Reducing steps to %s"
% param_1_n_steps,
ReducingNumberOfSteps,
)
# Compute the number of splits, i.e., how many lines in the grid for each engine.
# (note that this is guaranteed to be an integer number after the previous checks)
p1_split_steps = param_1_n_steps // n_engines
# Prepare arrays for results
if param_2 is None:
# One array
pcc = np.zeros(param_1_n_steps)
pa = np.linspace(param_1_minimum, param_1_maximum,
param_1_n_steps)
pb = None
else:
pcc = np.zeros((param_1_n_steps, param_2_n_steps))
# Prepare the two axes of the parameter space
pa = np.linspace(param_1_minimum, param_1_maximum,
param_1_n_steps)
pb = np.linspace(param_2_minimum, param_2_maximum,
param_2_n_steps)
# Define the parallel worker which will go through the computation
# NOTE: I only divide
# on the first parameter axis so that the different
# threads are more or less well mixed for points close and
# far from the best fit
def worker(start_index):
# Re-create the minimizer
backup_freeParameters = [
x.value for x in list(
self._likelihood_model.free_parameters.values())
]
this_minimizer = self._get_minimizer(
self.minus_log_like_profile, self._free_parameters)
this_p1min = pa[start_index * p1_split_steps]
this_p1max = pa[(start_index + 1) * p1_split_steps - 1]
# print("From %s to %s" % (this_p1min, this_p1max))
aa, bb, ccc = this_minimizer.contours(param_1,
this_p1min,
this_p1max,
p1_split_steps,
param_2,
param_2_minimum,
param_2_maximum,
param_2_n_steps,
progress=True,
**options)
# Restore best fit values
for val, par in zip(
backup_freeParameters,
list(self._likelihood_model.free_parameters.values()),
):
par.value = val
return ccc
# Now re-assemble the vector of results taking the different parts from the engines
all_results = client.execute_with_progress_bar(
worker, list(range(n_engines)), chunk_size=1)
for i, these_results in enumerate(all_results):
if param_2 is None:
pcc[i * p1_split_steps:(i + 1) *
p1_split_steps] = these_results[:, 0]
else:
pcc[i * p1_split_steps:(i + 1) *
p1_split_steps, :] = these_results
# Give the results the names that the following code expect. These are kept separate for debugging
# purposes
cc = pcc
a = pa
b = pb
# Here we have done the computation, in parallel computation or not. Let's make the plot
# with the contour
if param_2 is not None:
# 2d contour
fig = self._plot_contours("%s" % (param_1), a, "%s" % (param_2, ),
b, cc)
else:
# 1d contour (i.e., a profile)
fig = self._plot_profile("%s" % (param_1), a, cc)
# Check if we found a better minimum. This shouldn't happen, but in case of very difficult fit
# it might.
if self._current_minimum - cc.min() > 0.1:
if param_2 is not None:
idx = cc.argmin()
aidx, bidx = np.unravel_index(idx, cc.shape)
print(
"\nFound a better minimum: %s with %s = %s and %s = %s. Run again your fit starting from here."
% (cc.min(), param_1, a[aidx], param_2, b[bidx]))
else:
idx = cc.argmin()
print(
"Found a better minimum: %s with %s = %s. Run again your fit starting from here."
% (cc.min(), param_1, a[idx]))
return a, b, cc, fig
def get_contours(
self,
param_1,
param_1_minimum,
param_1_maximum,
param_1_n_steps,
param_2=None,
param_2_minimum=None,
param_2_maximum=None,
param_2_n_steps=None,
progress=True,
**options
):
"""
Generate confidence contours for the given parameters by stepping for the given number of steps between
the given boundaries. Call it specifying only source_1, param_1, param_1_minimum and param_1_maximum to
generate the profile of the likelihood for parameter 1. Specify all parameters to obtain instead a 2d
contour of param_1 vs param_2.
NOTE: if using parallel computation, param_1_n_steps must be an integer multiple of the number of running
engines. If that is not the case, the code will reduce the number of steps to match that requirement, and
issue a warning
:param param_1: fully qualified name of the first parameter or parameter instance
:param param_1_minimum: lower bound for the range for the first parameter
:param param_1_maximum: upper bound for the range for the first parameter
:param param_1_n_steps: number of steps for the first parameter
:param param_2: fully qualified name of the second parameter or parameter instance
:param param_2_minimum: lower bound for the range for the second parameter
:param param_2_maximum: upper bound for the range for the second parameter
:param param_2_n_steps: number of steps for the second parameter
:param progress: (True or False) whether to display progress or not
:param log: by default the steps are taken linearly. With this optional parameter you can provide a tuple of
booleans which specify whether the steps are to be taken logarithmically. For example,
'log=(True,False)' specify that the steps for the first parameter are to be taken logarithmically,
while they are linear for the second parameter. If you are generating the profile for only one
parameter, you can specify 'log=(True,)' or 'log=(False,)' (optional)
:return: a tuple containing an array corresponding to the steps for the first parameter, an array corresponding
to the steps for the second parameter (or None if stepping only in one direction), a matrix of size
param_1_steps x param_2_steps containing the value of the function at the corresponding points in the
grid. If param_2_steps is None (only one parameter), then this reduces to an array of
size param_1_steps.
"""
if hasattr(param_1, "value"):
# Substitute with the name
param_1 = param_1.path
if hasattr(param_2, "value"):
param_2 = param_2.path
# Check that the parameters exist
assert param_1 in self._likelihood_model.free_parameters, (
"Parameter %s is not a free parameters of the " "current model" % param_1
)
if param_2 is not None:
assert param_2 in self._likelihood_model.free_parameters, (
"Parameter %s is not a free parameters of the " "current model" % param_2
)
# Check that we have a valid fit
assert self._current_minimum is not None, "You have to run the .fit method before calling get_contours."
# Then restore the best fit
self._minimizer._restore_best_fit()
# Check minimal assumptions about the procedure
assert not (param_1 == param_2), "You have to specify two different parameters"
assert param_1_minimum < param_1_maximum, "Minimum larger than maximum for parameter 1"
if param_2 is not None:
assert param_2_minimum < param_2_maximum, "Minimum larger than maximum for parameter 2"
# Check whether we are parallelizing or not
if not threeML_config["parallel"]["use-parallel"]:
a, b, cc = self.minimizer.contours(
param_1,
param_1_minimum,
param_1_maximum,
param_1_n_steps,
param_2,
param_2_minimum,
param_2_maximum,
param_2_n_steps,
progress,
**options
)
# Collapse the second dimension of the results if we are doing a 1d contour
if param_2 is None:
cc = cc[:, 0]
else:
# With parallel computation
# In order to distribute fairly the computation, the strategy is to parallelize the computation
# by assigning to the engines one "line" of the grid at the time
# Connect to the engines
client = ParallelClient(**options)
# Get the number of engines
n_engines = client.get_number_of_engines()
# Check whether the number of threads is larger than the number of steps in the first direction
if n_engines > param_1_n_steps:
n_engines = int(param_1_n_steps)
custom_warnings.warn(
"The number of engines is larger than the number of steps. Using only %s engines." % n_engines,
ReducingNumberOfThreads,
)
# Check if the number of steps is divisible by the number
# of threads, otherwise issue a warning and make it so
if float(param_1_n_steps) % n_engines != 0:
# Set the number of steps to an integer multiple of the engines
# (note that // is the floor division, also called integer division)
param_1_n_steps = (param_1_n_steps // n_engines) * n_engines
custom_warnings.warn(
"Number of steps is not a multiple of the number of threads. Reducing steps to %s"
% param_1_n_steps,
ReducingNumberOfSteps,
)
# Compute the number of splits, i.e., how many lines in the grid for each engine.
# (note that this is guaranteed to be an integer number after the previous checks)
p1_split_steps = param_1_n_steps // n_engines
# Prepare arrays for results
if param_2 is None:
# One array
pcc = numpy.zeros(param_1_n_steps)
pa = numpy.linspace(param_1_minimum, param_1_maximum, param_1_n_steps)
pb = None
else:
pcc = numpy.zeros((param_1_n_steps, param_2_n_steps))
# Prepare the two axes of the parameter space
pa = numpy.linspace(param_1_minimum, param_1_maximum, param_1_n_steps)
pb = numpy.linspace(param_2_minimum, param_2_maximum, param_2_n_steps)
# Define the parallel worker which will go through the computation
# NOTE: I only divide
# on the first parameter axis so that the different
# threads are more or less well mixed for points close and
# far from the best fit
def worker(start_index):
# Re-create the minimizer
# backup_freeParameters = copy.deepcopy(self.freeParameters)
this_minimizer = self.Minimizer(self.minus_log_like_profile, self._free_parameters)
this_p1min = pa[start_index * p1_split_steps]
this_p1max = pa[(start_index + 1) * p1_split_steps - 1]
# print("From %s to %s" % (this_p1min, this_p1max))
aa, bb, ccc = this_minimizer.contours(
param_1,
this_p1min,
this_p1max,
p1_split_steps,
param_2,
param_2_minimum,
param_2_maximum,
param_2_n_steps,
False,
**options
)
# self.freeParameters = backup_freeParameters
return ccc
# Get a balanced view of the engines
lview = client.load_balanced_view()
# lview.block = True
# Distribute the work among the engines and start it, but return immediately the control
# to the main thread
amr = lview.map_async(worker, range(n_engines))
# print progress
progress = ProgressBar(n_engines)
# This loop will check from time to time the status of the computation, which is happening on
# different threads, and update the progress bar
while not amr.ready():
# Check and report the status of the computation every second
time.sleep(1 + np.random.uniform(0, 1))
# if (debug):
# stdouts = amr.stdout
#
# # clear_output doesn't do much in terminal environments
# for stdout, stderr in zip(amr.stdout, amr.stderr):
# if stdout:
# print "%s" % (stdout[-1000:])
# if stderr:
# print "%s" % (stderr[-1000:])
# sys.stdout.flush()
progress.animate(amr.progress - 1)
# Always display 100% at the end
progress.animate(n_engines - 1)
# Add a new line after the progress bar
print("\n")
# print("Serial time: %1.f (speed-up: %.1f)" %(amr.serial_time, float(amr.serial_time) / amr.wall_time))
# Get the results. This will raise exceptions if something wrong happened during the computation.
# We don't catch it so that the user will be aware of that
res = amr.get()
# Now re-assemble the vector of results taking the different parts from the engines
for i in range(n_engines):
if param_2 is None:
pcc[i * p1_split_steps : (i + 1) * p1_split_steps] = res[i][:, 0]
else:
pcc[i * p1_split_steps : (i + 1) * p1_split_steps, :] = res[i]
# Give the results the names that the following code expect. These are kept separate for debugging
# purposes
cc = pcc
a = pa
b = pb
# Here we have done the computation, in parallel computation or not. Let's make the plot
# with the contour
if param_2 is not None:
# 2d contour
fig = self._plot_contours("%s" % (param_1), a, "%s" % (param_2,), b, cc)
else:
# 1d contour (i.e., a profile)
fig = self._plot_profile("%s" % (param_1), a, cc)
# Check if we found a better minimum. This shouldn't happen, but in case of very difficult fit
# it might.
if self._current_minimum - cc.min() > 0.1:
if param_2 is not None:
idx = cc.argmin()
aidx, bidx = numpy.unravel_index(idx, cc.shape)
print(
"\nFound a better minimum: %s with %s = %s and %s = %s. Run again your fit starting from here."
% (cc.min(), param_1, a[aidx], param_2, b[bidx])
)
else:
idx = cc.argmin()
print(
"Found a better minimum: %s with %s = %s. Run again your fit starting from here."
% (cc.min(), param_1, a[idx])
)
return a, b, cc, fig
