class FittingSequence(object):
"""
class to define a sequence of fitting applied, inherit the Fitting class
this is a Workflow manager that allows to update model configurations before executing another step in the modelling
The user can take this module as an example of how to create their own workflows or build their own around the FittingSequence
"""
def __init__(self,
kwargs_data_joint,
kwargs_model,
kwargs_constraints,
kwargs_likelihood,
kwargs_params,
mpi=False,
verbose=True):
"""
:param kwargs_data_joint: keyword argument specifying the data according to LikelihoodModule
:param kwargs_model:
:param kwargs_constraints:
:param kwargs_likelihood:
:param kwargs_params:
:param mpi: MPI option (bool), if True, will launch an MPI Pool job for the steps in the fitting sequence where
possible
:param verbose: bool, if True prints temporary results and indicators of the fitting process
"""
self.kwargs_data_joint = kwargs_data_joint
self.multi_band_list = kwargs_data_joint.get('multi_band_list', [])
self._verbose = verbose
self._mpi = mpi
self._updateManager = MultiBandUpdateManager(kwargs_model,
kwargs_constraints,
kwargs_likelihood,
kwargs_params,
num_bands=len(
self.multi_band_list))
self._mcmc_init_samples = None
def kwargs_fixed(self):
"""
returns the updated kwargs_fixed from the update Manager
:return: list of fixed kwargs, see UpdateManager()
"""
return self._updateManager.fixed_kwargs
def fit_sequence(self, fitting_list):
"""
:param fitting_list: list of [['string', {kwargs}], ..] with 'string being the specific fitting option and kwargs being the arguments passed to this option
:return: fitting results
"""
chain_list = []
for i, fitting in enumerate(fitting_list):
fitting_type = fitting[0]
kwargs = fitting[1]
if fitting_type == 'restart':
self._updateManager.set_init_state()
elif fitting_type == 'update_settings':
self.update_settings(**kwargs)
elif fitting_type == 'set_param_value':
self.set_param_value(**kwargs)
elif fitting_type == 'fix_not_computed':
self.fix_not_computed(**kwargs)
elif fitting_type == 'psf_iteration':
self.psf_iteration(**kwargs)
elif fitting_type == 'align_images':
self.align_images(**kwargs)
elif fitting_type == 'PSO':
kwargs_result, chain, param = self.pso(**kwargs)
self._updateManager.update_param_state(**kwargs_result)
chain_list.append([fitting_type, chain, param])
elif fitting_type == 'SIMPLEX':
kwargs_result = self.simplex(**kwargs)
self._updateManager.update_param_state(**kwargs_result)
chain_list.append([fitting_type, kwargs_result])
elif fitting_type == 'MCMC':
if not 'init_samples' in kwargs:
kwargs['init_samples'] = self._mcmc_init_samples
elif kwargs['init_samples'] is None:
kwargs['init_samples'] = self._mcmc_init_samples
mcmc_output = self.mcmc(**kwargs)
kwargs_result = self._result_from_mcmc(mcmc_output)
self._updateManager.update_param_state(**kwargs_result)
chain_list.append(mcmc_output)
elif fitting_type == 'nested_sampling':
ns_output = self.nested_sampling(**kwargs)
chain_list.append(ns_output)
else:
raise ValueError(
"fitting_sequence %s is not supported. Please use: 'PSO', 'SIMPLEX', 'MCMC', 'psf_iteration', "
"'restart', 'update_settings' or "
"'align_images'" % fitting_type)
return chain_list
def best_fit(self, bijective=False):
"""
:param bijective: bool, if True, the mapping of image2source_plane and the mass_scaling parameterisation are inverted. If you do not use those options, there is no effect.
:return: best fit model of the current state of the FittingSequence class
"""
return self._updateManager.best_fit(bijective=bijective)
def update_state(self, kwargs_update):
"""
updates current best fit state to the input model keywords specified.
:param kwargs_update: format of kwargs_result
:return: None
"""
self._updateManager.update_param_state(**kwargs_update)
@property
def best_fit_likelihood(self):
"""
returns the log likelihood of the best fit model of the current state of this class
:return: log likelihood, float
"""
kwargs_result = self.best_fit(bijective=False)
param_class = self.param_class
likelihoodModule = self.likelihoodModule
logL = likelihoodModule.logL(param_class.kwargs2args(**kwargs_result))
return logL
@property
def bic(self):
"""
returns the bayesian information criterion of the model.
:return: bic value, float
"""
num_data = self.likelihoodModule.num_data
num_param_nonlinear = self.param_class.num_param()[0]
num_param_linear = self.param_class.num_param_linear()
num_param = num_param_nonlinear + num_param_linear
bic = analysis_util.bic_model(self.best_fit_likelihood, num_data,
num_param)
return bic
@property
def param_class(self):
"""
:return: Param() class instance reflecting the current state of Fittingsequence
"""
return self._updateManager.param_class
@property
def likelihoodModule(self):
"""
:return: Likelihood() class instance reflecting the current state of Fittingsequence
"""
kwargs_model = self._updateManager.kwargs_model
kwargs_likelihood = self._updateManager.kwargs_likelihood
likelihoodModule = LikelihoodModule(self.kwargs_data_joint,
kwargs_model, self.param_class,
**kwargs_likelihood)
return likelihoodModule
def simplex(self, n_iterations, method='Nelder-Mead'):
"""
Downhill simplex optimization using the Nelder-Mead algorithm.
:param n_iterations: maximum number of iterations to perform
:param method: the optimization method used, see documentation in scipy.optimize.minimize
:return: result of the best fit
"""
param_class = self.param_class
kwargs_temp = self._updateManager.parameter_state
init_pos = param_class.kwargs2args(**kwargs_temp)
sampler = Sampler(likelihoodModule=self.likelihoodModule)
result = sampler.simplex(init_pos, n_iterations, method)
kwargs_result = param_class.args2kwargs(result, bijective=True)
return kwargs_result
def mcmc(self,
n_burn,
n_run,
walkerRatio,
sigma_scale=1,
threadCount=1,
init_samples=None,
re_use_samples=True,
sampler_type='EMCEE',
progress=True,
backup_filename=None,
start_from_backup=False):
"""
MCMC routine
:param n_burn: number of burn in iterations (will not be saved)
:param n_run: number of MCMC iterations that are saved
:param walkerRatio: ratio of walkers/number of free parameters
:param sigma_scale: scaling of the initial parameter spread relative to the width in the initial settings
:param threadCount: number of CPU threads. If MPI option is set, threadCount=1
:param init_samples: initial sample from where to start the MCMC process
:param re_use_samples: bool, if True, re-uses the samples described in init_samples.nOtherwise starts from scratch.
:param sampler_type: string, which MCMC sampler to be used. Options are: 'EMCEE'
:param progress: boolean, if True shows progress bar in EMCEE
:return: list of output arguments, e.g. MCMC samples, parameter names, logL distances of all samples specified by the specific sampler used
"""
param_class = self.param_class
# run PSO
mcmc_class = Sampler(likelihoodModule=self.likelihoodModule)
kwargs_temp = self._updateManager.parameter_state
mean_start = param_class.kwargs2args(**kwargs_temp)
kwargs_sigma = self._updateManager.sigma_kwargs
sigma_start = np.array(
param_class.kwargs2args(**kwargs_sigma)) * sigma_scale
num_param, param_list = param_class.num_param()
# run MCMC
if not init_samples is None and re_use_samples is True:
num_samples, num_param_prev = np.shape(init_samples)
if num_param_prev == num_param:
print(
"re-using previous samples to initialize the next MCMC run."
)
n_walkers = num_param * walkerRatio
idxs = np.random.choice(len(init_samples), n_walkers)
initpos = init_samples[idxs]
else:
raise ValueError(
"Can not re-use previous MCMC samples as number of parameters have changed!"
)
else:
initpos = None
if sampler_type is 'EMCEE':
n_walkers = num_param * walkerRatio
samples, dist = mcmc_class.mcmc_emcee(
n_walkers,
n_run,
n_burn,
mean_start,
sigma_start,
mpi=self._mpi,
threadCount=threadCount,
progress=progress,
initpos=initpos,
backup_filename=backup_filename,
start_from_backup=start_from_backup)
output = [sampler_type, samples, param_list, dist]
else:
raise ValueError('sampler_type %s not supported!' % sampler_type)
self._mcmc_init_samples = samples # overwrites previous samples to continue from there in the next MCMC run
return output
def pso(self,
n_particles,
n_iterations,
sigma_scale=1,
print_key='PSO',
threadCount=1):
"""
Particle Swarm Optimization
:param n_particles: number of particles in the Particle Swarm Optimization
:param n_iterations: number of iterations in the optimization process
:param sigma_scale: scaling of the initial parameter spread relative to the width in the initial settings
:param print_key: string, printed text when executing this routine
:param threadCount: number of CPU threads. If MPI option is set, threadCount=1
:return: result of the best fit, the chain of the best fit parameter after each iteration, list of parameters in same order
"""
param_class = self.param_class
kwargs_temp = self._updateManager.parameter_state
init_pos = param_class.kwargs2args(**kwargs_temp)
kwargs_sigma = self._updateManager.sigma_kwargs
sigma_start = param_class.kwargs2args(**kwargs_sigma)
lowerLimit = np.array(init_pos) - np.array(sigma_start) * sigma_scale
upperLimit = np.array(init_pos) + np.array(sigma_start) * sigma_scale
num_param, param_list = param_class.num_param()
# run PSO
sampler = Sampler(likelihoodModule=self.likelihoodModule)
result, chain = sampler.pso(n_particles,
n_iterations,
lowerLimit,
upperLimit,
init_pos=init_pos,
threadCount=threadCount,
mpi=self._mpi,
print_key=print_key)
kwargs_result = param_class.args2kwargs(result, bijective=True)
return kwargs_result, chain, param_list
def nested_sampling(self,
sampler_type='MULTINEST',
kwargs_run={},
prior_type='uniform',
width_scale=1,
sigma_scale=1,
output_basename='chain',
remove_output_dir=True,
dypolychord_dynamic_goal=0.8,
polychord_settings={},
dypolychord_seed_increment=200,
output_dir="nested_sampling_chains",
dynesty_bound='multi',
dynesty_sample='auto'):
"""
Run (Dynamic) Nested Sampling algorithms, depending on the type of algorithm.
:param sampler_type: 'MULTINEST', 'DYPOLYCHORD', 'DYNESTY'
:param kwargs_run: keywords passed to the core sampling method
:param prior_type: 'uniform' of 'gaussian', for converting the unit hypercube to param cube
:param width_scale: scale the width (lower/upper limits) of the parameters space by this factor
:param sigma_scale: if prior_type is 'gaussian', scale the gaussian sigma by this factor
:param output_basename: name of the folder in which the core MultiNest/PolyChord code will save output files
:param remove_output_dir: if True, the above folder is removed after completion
:param dypolychord_dynamic_goal: dynamic goal for DyPolyChord (trade-off between evidence (0) and posterior (1) computation)
:param polychord_settings: settings dictionary to send to pypolychord. Check dypolychord documentation for details.
:param dypolychord_seed_increment: seed increment for dypolychord with MPI. Check dypolychord documentation for details.
:param dynesty_bound: see https://dynesty.readthedocs.io for details
:param dynesty_sample: see https://dynesty.readthedocs.io for details
:return: list of output arguments : samples, mean inferred values, log-likelihood, log-evidence, error on log-evidence for each sample
"""
mean_start, sigma_start = self._prepare_sampling(prior_type)
if sampler_type == 'MULTINEST':
sampler = MultiNestSampler(self.likelihoodModule,
prior_type=prior_type,
prior_means=mean_start,
prior_sigmas=sigma_start,
width_scale=width_scale,
sigma_scale=sigma_scale,
output_dir=output_dir,
output_basename=output_basename,
remove_output_dir=remove_output_dir,
use_mpi=self._mpi)
samples, means, logZ, logZ_err, logL, results_object = sampler.run(
kwargs_run)
elif sampler_type == 'DYPOLYCHORD':
if 'resume_dyn_run' in kwargs_run and kwargs_run[
'resume_dyn_run'] is True:
resume_dyn_run = True
else:
resume_dyn_run = False
sampler = DyPolyChordSampler(self.likelihoodModule,
prior_type=prior_type,
prior_means=mean_start,
prior_sigmas=sigma_start,
width_scale=width_scale,
sigma_scale=sigma_scale,
output_dir=output_dir,
output_basename=output_basename,
polychord_settings=polychord_settings,
remove_output_dir=remove_output_dir,
resume_dyn_run=resume_dyn_run,
use_mpi=self._mpi)
samples, means, logZ, logZ_err, logL, results_object = sampler.run(
dypolychord_dynamic_goal, kwargs_run)
elif sampler_type == 'DYNESTY':
sampler = DynestySampler(self.likelihoodModule,
prior_type=prior_type,
prior_means=mean_start,
prior_sigmas=sigma_start,
width_scale=width_scale,
sigma_scale=sigma_scale,
bound=dynesty_bound,
sample=dynesty_sample,
use_mpi=self._mpi)
samples, means, logZ, logZ_err, logL, results_object = sampler.run(
kwargs_run)
else:
raise ValueError('Sampler type %s not supported.' % sampler_type)
# update current best fit values
self._update_state(samples[-1])
output = [
sampler_type, samples, sampler.param_names, logL, logZ, logZ_err,
results_object
]
return output
def psf_iteration(self,
num_iter=10,
no_break=True,
stacking_method='median',
block_center_neighbour=0,
keep_psf_error_map=True,
psf_symmetry=1,
psf_iter_factor=1,
verbose=True,
compute_bands=None):
"""
iterative PSF reconstruction
:param num_iter: number of iterations in the process
:param no_break: bool, if False will break the process as soon as one step lead to a wors reconstruction then the previous step
:param stacking_method: string, 'median' and 'mean' supported
:param block_center_neighbour: radius of neighbouring point source to be blocked in the reconstruction
:param keep_psf_error_map: bool, whether or not to keep the previous psf_error_map
:param psf_symmetry: int, number of invariant rotations in the reconstructed PSF
:param psf_iter_factor: factor of new estimated PSF relative to the old one PSF_updated = (1-psf_iter_factor) * PSF_old + psf_iter_factor*PSF_new
:param verbose: bool, print statements
:param compute_bands: bool list, if multiple bands, this process can be limited to a subset of bands
:return: 0, updated PSF is stored in self.mult_iband_list
"""
kwargs_model = self._updateManager.kwargs_model
kwargs_likelihood = self._updateManager.kwargs_likelihood
likelihood_mask_list = kwargs_likelihood.get(
'image_likelihood_mask_list', None)
#param_class = self.param_class
kwargs_temp = self.best_fit(bijective=False)
if compute_bands is None:
compute_bands = [True] * len(self.multi_band_list)
for band_index in range(len(self.multi_band_list)):
if compute_bands[band_index] is True:
kwargs_psf = self.multi_band_list[band_index][1]
image_model = SingleBandMultiModel(
self.multi_band_list,
kwargs_model,
likelihood_mask_list=likelihood_mask_list,
band_index=band_index)
psf_iter = PsfFitting(image_model_class=image_model)
kwargs_psf = psf_iter.update_iterative(
kwargs_psf,
kwargs_params=kwargs_temp,
num_iter=num_iter,
no_break=no_break,
stacking_method=stacking_method,
block_center_neighbour=block_center_neighbour,
keep_psf_error_map=keep_psf_error_map,
psf_symmetry=psf_symmetry,
psf_iter_factor=psf_iter_factor,
verbose=verbose)
self.multi_band_list[band_index][1] = kwargs_psf
return 0
def align_images(self,
n_particles=10,
n_iterations=10,
lowerLimit=-0.2,
upperLimit=0.2,
threadCount=1,
compute_bands=None):
"""
aligns the coordinate systems of different exposures within a fixed model parameterisation by executing a PSO
with relative coordinate shifts as free parameters
:param n_particles: number of particles in the Particle Swarm Optimization
:param n_iterations: number of iterations in the optimization process
:param lowerLimit: lower limit of relative shift
:param upperLimit: upper limit of relative shift
:param verbose: bool, print statements
:param compute_bands: bool list, if multiple bands, this process can be limited to a subset of bands
:return:
"""
kwargs_model = self._updateManager.kwargs_model
kwargs_likelihood = self._updateManager.kwargs_likelihood
likelihood_mask_list = kwargs_likelihood.get(
'image_likelihood_mask_list', None)
#param_class = self.param_class
#lens_temp, source_temp, lens_light_temp, ps_temp, cosmo_temp = self._updateManager.parameter_state
#lens_updated = param_class.update_lens_scaling(cosmo_temp, lens_temp)
#source_updated = param_class.image2source_plane(source_temp, lens_updated)
kwargs_temp = self.best_fit(bijective=False)
if compute_bands is None:
compute_bands = [True] * len(self.multi_band_list)
for i in range(len(self.multi_band_list)):
if compute_bands[i] is True:
alignmentFitting = AlignmentFitting(
self.multi_band_list,
kwargs_model,
kwargs_temp,
band_index=i,
likelihood_mask_list=likelihood_mask_list)
kwargs_data, chain = alignmentFitting.pso(
n_particles=n_particles,
n_iterations=n_iterations,
lowerLimit=lowerLimit,
upperLimit=upperLimit,
threadCount=threadCount,
mpi=self._mpi,
print_key='Alignment fitting for band %s ...' % i)
print('Align completed for band %s.' % i)
print('ra_shift: %s, dec_shift: %s' %
(kwargs_data['ra_shift'], kwargs_data['dec_shift']))
self.multi_band_list[i][0] = kwargs_data
return 0
def update_settings(self,
kwargs_model={},
kwargs_constraints={},
kwargs_likelihood={},
lens_add_fixed=[],
source_add_fixed=[],
lens_light_add_fixed=[],
ps_add_fixed=[],
cosmo_add_fixed=[],
lens_remove_fixed=[],
source_remove_fixed=[],
lens_light_remove_fixed=[],
ps_remove_fixed=[],
cosmo_remove_fixed=[],
change_source_lower_limit=None,
change_source_upper_limit=None):
"""
updates lenstronomy settings "on the fly"
:param kwargs_model: kwargs, specified keyword arguments overwrite the existing ones
:param kwargs_constraints: kwargs, specified keyword arguments overwrite the existing ones
:param kwargs_likelihood: kwargs, specified keyword arguments overwrite the existing ones
:param lens_add_fixed: [[i_model, ['param1', 'param2',...], [...]]
:param source_add_fixed: [[i_model, ['param1', 'param2',...], [...]]
:param lens_light_add_fixed: [[i_model, ['param1', 'param2',...], [...]]
:param ps_add_fixed: [[i_model, ['param1', 'param2',...], [...]]
:param cosmo_add_fixed: ['param1', 'param2',...]
:param lens_remove_fixed: [[i_model, ['param1', 'param2',...], [...]]
:param source_remove_fixed: [[i_model, ['param1', 'param2',...], [...]]
:param lens_light_remove_fixed: [[i_model, ['param1', 'param2',...], [...]]
:param ps_remove_fixed: [[i_model, ['param1', 'param2',...], [...]]
:param cosmo_remove_fixed: ['param1', 'param2',...]
:return: 0, the settings are overwritten for the next fitting step to come
"""
self._updateManager.update_options(kwargs_model, kwargs_constraints,
kwargs_likelihood)
self._updateManager.update_fixed(lens_add_fixed, source_add_fixed,
lens_light_add_fixed, ps_add_fixed,
cosmo_add_fixed, lens_remove_fixed,
source_remove_fixed,
lens_light_remove_fixed,
ps_remove_fixed, cosmo_remove_fixed)
self._updateManager.update_limits(change_source_lower_limit,
change_source_upper_limit)
return 0
def set_param_value(self, **kwargs):
"""
Set a parameter to a specific value. `kwargs` are below.
:param lens: [[i_model, ['param1', 'param2',...], [...]]
:type lens:
:param source: [[i_model, ['param1', 'param2',...], [...]]
:type source:
:param lens_light: [[i_model, ['param1', 'param2',...], [...]]
:type lens_light:
:param ps: [[i_model, ['param1', 'param2',...], [...]]
:type ps:
:return: 0, the value of the param is overwritten
:rtype:
"""
self._updateManager.update_param_value(**kwargs)
def fix_not_computed(self, free_bands):
"""
fixes lens model parameters of imaging bands/frames that are not computed and frees the parameters of the other
lens models to the initial kwargs_fixed options
:param free_bands: bool list of length of imaging bands in order of imaging bands, if False: set fixed lens model
:return: None
"""
self._updateManager.fix_not_computed(free_bands=free_bands)
def _prepare_sampling(self, prior_type):
if prior_type == 'gaussian':
mean_start = self.param_class.kwargs2args(
**self._updateManager.parameter_state)
sigma_start = self.param_class.kwargs2args(
**self._updateManager.sigma_kwargs)
mean_start = np.array(mean_start)
sigma_start = np.array(sigma_start)
else:
mean_start, sigma_start = None, None
return mean_start, sigma_start
def _update_state(self, result):
"""
:param result: array of parameters being sampled (e.g. result of MCMC chain)
:return: None, updates the parameter state of the class instance
"""
kwargs_result = self.param_class.args2kwargs(result, bijective=True)
self._updateManager.update_param_state(**kwargs_result)
def _result_from_mcmc(self, mcmc_output):
"""
:param mcmc_output: list returned by self.mcmc()
:return: kwargs_result like returned by self.pso(), from best logL MCMC sample
"""
_, samples, _, logL_values = mcmc_output
# get index of best logL sample
bestfit_idx = np.argmax(logL_values)
bestfit_sample = samples[bestfit_idx, :]
bestfit_result = bestfit_sample.tolist()
# get corresponding kwargs
kwargs_result = self.param_class.args2kwargs(bestfit_result,
bijective=True)
return kwargs_result
class TestUpdateManager(object):
def setup(self):
kwargs_model = {
'lens_model_list': ['SHEAR', 'SHEAR'],
'source_light_model_list': ['UNIFORM'],
'index_lens_model_list': [[0], [1]]
}
kwargs_constraints = {}
kwargs_likelihood = {}
kwargs_params = {}
lens_init = [{
'e1': 0,
'e2': 0,
'ra_0': 0,
'dec_0': 0
}, {
'e1': 0,
'e2': 0,
'ra_0': 0,
'dec_0': 0
}]
lens_sigma = [{'e1': 0.1, 'e2': 0.1}, {'e1': 0.1, 'e2': 0.1}]
lens_fixed = [{'ra_0': 0, 'dec_0': 0}, {'ra_0': 0, 'dec_0': 0}]
lens_lower = [{'e1': -1, 'e2': -1}, {'e1': -1, 'e2': -1}]
lens_upper = [{'e1': 1, 'e2': 1}, {'e1': 1, 'e2': 1}]
kwargs_params['lens_model'] = [
lens_init, lens_sigma, lens_fixed, lens_lower, lens_upper
]
kwargs_params['source_model'] = [[{}], [{}], [{}], [{}], [{}]]
kwargs_params['special'] = [{
'special1': 1
}, {
'special1': 1
}, {
'special1': 0.1
}, {
'special1': 0
}, {
'special1': 1
}]
self.manager = MultiBandUpdateManager(kwargs_model,
kwargs_constraints,
kwargs_likelihood,
kwargs_params,
num_bands=2)
def test_none_mamager(self):
manager = MultiBandUpdateManager(kwargs_model={},
kwargs_constraints={},
kwargs_likelihood={},
kwargs_params={},
num_bands=0)
results = manager.best_fit()
assert len(results['kwargs_lens']) == 0
def test_keep_frame_fixed(self):
frame_list_fixed = [0]
assert 'e1' not in self.manager._lens_fixed[0]
self.manager.keep_frame_fixed(frame_list_fixed)
assert 'e1' in self.manager._lens_fixed[0]
self.manager.undo_frame_fixed(frame_list=[0])
assert 'e1' not in self.manager._lens_fixed[0]
assert 'ra_0' in self.manager._lens_fixed[0]
def test_fix_not_computed(self):
self.manager.fix_not_computed(free_bands=[False, True])
print(self.manager._lens_fixed)
assert 'e1' in self.manager._lens_fixed[0]
assert 'ra_0' in self.manager._lens_fixed[0]
assert 'e1' not in self.manager._lens_fixed[1]