def logp(self, pvals=None):
"""
Calculate the log-prior of the system
Parameters
----------
pvals : array-like or refnx.analysis.Parameters
values for the varying or entire set of parameters
Returns
-------
logp : float
log-prior probability
Notes
-----
The log-prior is calculated as:
.. code-block:: python
logp = np.sum(param.logp() for param in
self.varying_parameters())
"""
self.setp(pvals)
logp = np.sum([
param.logp() for param in f_unique(
p for p in flatten(self.parameters) if p.vary)
])
if not np.isfinite(logp):
return -np.inf
return logp
def nvary(self):
"""
Returns
-------
nvary : int
The number of :class:`Parameter` contained in this object that are
allowed to vary.
"""
return len([1 for param in f_unique(flatten(self.data)) if param.vary])
def constrained_parameters(self):
"""
Returns
-------
constrained_parameters : list
A list of unique :class:`Parameter` contained in this object that
have constraints.
"""
return [param for param in f_unique(flatten(self.data))
if param.constraint is not None]
def logp(self):
"""
Calculates logp for all the parameters
Returns
-------
logp : float
Log probability for all the parameters
"""
# logp for all the parameters
return np.sum([param.logp() for param in f_unique(flatten(self.data))
if param.vary])
def varying_parameters(self):
"""
Unique list of varying parameters
Returns
-------
p : list
Unique list of varying parameters
"""
p = [param for param in f_unique(flatten(self.data)) if param.vary]
q = Parameters()
q.data = p
return q
def varying_parameters(self):
"""
Returns
-------
varying_parameters : refnx.analysis.Parameters
The varying Parameter objects allowed to vary during the fit.
"""
# create and return a Parameters object because it has the
# __array__ method, which allows one to quickly get numerical values.
p = Parameters()
p.data = list(f_unique(p for p in flatten(self.parameters) if p.vary))
return p
def pvals(self, pvals):
varying = [param for param in f_unique(flatten(self.data))
if param.vary]
if np.size(pvals) == len(varying):
[setattr(param, 'value', pvals[i]) for i, param
in enumerate(varying)]
return
flattened_parameters = list(flatten(self.data))
if np.size(pvals) == len(flattened_parameters):
[setattr(param, 'value', pvals[i]) for i, param
in enumerate(flattened_parameters)]
return
else:
raise ValueError("You supplied the wrong number of values %d when "
"setting this Parameters.pvals attribute"
% len(pvals))
def flattened(self, unique=False):
"""
A list of all the :class:`Parameter` contained in this object,
including those contained within :class:`Parameters` at any depth.
Parameters
----------
unique : bool
The list will only contain unique objects.
Returns
-------
params : list
A list of :class:`Parameter` contained in this object.
"""
if unique:
return list(f_unique(flatten(self.data)))
else:
return list(flatten(self.data))
def bounds(self):
bounds = list( (p.bounds.lb, p.bounds.ub) for p
in f_unique(flatten(self.objective.parameters)) if p.vary )
self.lowerBounds = np.array(list(bound[0] for bound in bounds))
self.upperBounds = np.array(list(bound[1] for bound in bounds))
def __contains__(self, item):
return id(item) in [id(p) for p in f_unique(flatten(self.data))]
def __contains__(self, item):
"""
Does this instance contain a given :class:`Parameter`
"""
return id(item) in [id(p) for p in f_unique(flatten(self.data))]
def constrained_parameters(self):
"""
List of constrained parameters
"""
return [param for param in f_unique(flatten(self.data))
if param.constraint is not None]
def nvary(self):
return np.sum([1 for param in f_unique(flatten(self.data))
if param.vary])
def flattened(self, unique=False):
if unique:
return list(f_unique(flatten(self.data)))
else:
return list(flatten(self.data))
def fit(self, method="L-BFGS-B", target="nll", verbose=True, **kws):
"""
Obtain the maximum log-likelihood, or log-posterior, estimate (mode)
of the objective. Maximising the log-likelihood is equivalent to
minimising chi2 in a least squares fit.
Parameters
----------
method : str
which method to use for the optimisation. One of:
- `'least_squares'`: :func:`scipy.optimize.least_squares`.
- `'L-BFGS-B'`: L-BFGS-B.
- `'differential_evolution'`:
:func:`scipy.optimize.differential_evolution`
- `'dual_annealing'`:
:func:`scipy.optimize.dual_annealing` (SciPy >= 1.2.0)
- `'shgo'`: :func:`scipy.optimize.shgo` (SciPy >= 1.2.0)
You can also choose many of the minimizers from
:func:`scipy.optimize.minimize`.
target : {'nll', 'nlpost'}, optional
Minimize the negative log-likelihood (`'nll'`) or the negative
log-posterior (`'nlpost'`). This is equivalent to maximising the
likelihood or posterior probabilities respectively.
Maximising the likelihood is equivalent to minimising chi^2 in a
least-squares fit.
This option only applies to the `differential_evolution`, `shgo`,
`dual_annealing` or `L-BFGS-B` methods.
These optimisers require lower and upper (box) bounds for each
parameter. If the `Bounds` on a parameter are not an `Interval`,
but a `PDF` specifying a statistical distribution, then the lower
and upper bounds are approximated as
``PDF.rv.ppf([0.005, 0.995])``, covering 99 % of the statistical
distribution.
verbose : bool, optional
Gives fitting progress. To see a progress bar tqdm has to be
installed.
kws : dict
Additional arguments are passed to the underlying minimization
method.
Returns
-------
result, covar : :class:`scipy.optimize.OptimizeResult`, np.ndarray
`result.x` contains the best fit parameters
`result.covar` is the covariance matrix for the fit.
`result.stderr` is the uncertainties on each of the fit parameters.
Notes
-----
If the `objective` supplies a `residuals` method then `least_squares`
can be used. Otherwise the `nll` method of the `objective` is
minimised. Use this method just before a sampling run.
If `self.objective.parameters` is a `Parameters` instance, then each
of the varying parameters has its value updated by the fit, and each
`Parameter` has a `stderr` attribute which represents the uncertainty
on the fit parameter.
The use of `dual annealing` and `shgo` requires that `scipy >= 1.2.0`
be installed.
"""
_varying_parameters = self.objective.varying_parameters()
init_pars = np.array(_varying_parameters)
_min_kws = {}
_min_kws.update(kws)
_bounds = bounds_list(self.objective.varying_parameters())
_min_kws["bounds"] = _bounds
# setup callback default
_min_kws.setdefault("callback", None)
cost = self.objective.nll
if target == "nlpost":
cost = self.objective.nlpost
# a decorator for the progress bar updater
def _callback_wrapper(callback_func, pbar):
def callback(*args, **kwds):
pbar.update(1)
if callback_func is None:
return None
else:
return callback_func(*args, **kwds)
return callback
# least_squares Trust Region Reflective by default
if method == "least_squares":
b = np.array(_bounds)
_min_kws["bounds"] = (b[..., 0], b[..., 1])
# least_squares doesn't have a callback
_min_kws.pop("callback", None)
res = least_squares(self.objective.residuals, init_pars,
**_min_kws)
# differential_evolution, dual_annealing, shgo require lower and upper
# bounds
elif method in ["differential_evolution", "dual_annealing", "shgo"]:
mini = getattr(sciopt, method)
with get_progress_bar(verbose, None) as pbar:
_min_kws["callback"] = _callback_wrapper(
_min_kws["callback"], pbar)
res = mini(cost, **_min_kws)
else:
# otherwise stick it to minimizer. Default being L-BFGS-B
_min_kws["method"] = method
_min_kws["bounds"] = _bounds
with get_progress_bar(verbose, None) as pbar:
_min_kws["callback"] = _callback_wrapper(
_min_kws["callback"], pbar)
res = minimize(cost, init_pars, **_min_kws)
# OptimizeResult.success may not be present (dual annealing)
if hasattr(res, "success") and res.success:
self.objective.setp(res.x)
# Covariance matrix estimation
covar = self.objective.covar()
errors = np.sqrt(np.diag(covar))
res["covar"] = covar
res["stderr"] = errors
# check if the parameters are all Parameter instances.
flat_params = list(f_unique(flatten(self.objective.parameters)))
if np.all([is_parameter(param) for param in flat_params]):
# zero out all the old parameter stderrs
for param in flat_params:
param.stderr = None
param.chain = None
for i, param in enumerate(_varying_parameters):
param.stderr = errors[i]
# need to touch up the output to check we leave
# parameters as we found them
self.objective.setp(res.x)
return res
def nDim(self):
if self.nDimensions is None:
self.nDimensions = len(list(p for p
in f_unique(flatten(self.objective.parameters)) if p.vary ))
return self.nDimensions
def sample(self, steps, nthin=1, random_state=None, f=None, callback=None,
verbose=True, pool=0):
"""
Performs sampling from the objective.
Parameters
----------
steps : int
Collect `steps` samples into the chain. The sampler will run a
total of `steps * nthin` moves.
nthin : int, optional
Each chain sample is separated by `nthin` iterations.
random_state : int or `np.random.RandomState`, optional
If `random_state` is an int, a new `np.random.RandomState` instance
is used, seeded with `random_state`.
If `random_state` is already a `np.random.RandomState` instance,
then that `np.random.RandomState` instance is used. Specify
`random_state` for repeatable sampling
f : file-like or str
File to incrementally save chain progress to. Each row in the file
is a flattened array of size `(nwalkers, ndim)` or
`(ntemps, nwalkers, ndim)`. There are `steps` rows in the
file.
callback : callable
callback function to be called at each iteration step
verbose : bool, optional
Gives updates on the sampling progress
pool : int or map-like object, optional
If `pool` is an `int` then it specifies the number of threads to
use for parallelization. If `pool == 0`, then all CPU's are used.
If pool is an object with a map method that follows the same
calling sequence as the built-in map function, then this pool is
used for parallelisation.
Notes
-----
Please see :class:`emcee.EnsembleSampler` for its detailed behaviour.
>>> # we'll burn the first 500 steps
>>> fitter.sample(500)
>>> # after you've run those, then discard them by resetting the
>>> # sampler.
>>> fitter.sampler.reset()
>>> # Now collect 40 steps, each step separated by 50 sampler
>>> # generations.
>>> fitter.sample(40, nthin=50)
One can also burn and thin in `Curvefitter.process_chain`.
"""
self._check_vars_unchanged()
if self._state is None:
self.initialise()
self.__pt_iterations = 0
if isinstance(self.sampler, PTSampler):
steps *= nthin
# for saving progress to file
def _callback_wrapper(state, h=None):
if callback is not None:
callback(state.coords, state.log_prob)
if h is not None:
# if you're parallel tempering, then you only
# want to save every nthin
if isinstance(self.sampler, PTSampler):
self.__pt_iterations += 1
if self.__pt_iterations % nthin:
return None
h.write(' '.join(map(str, state.coords.ravel())))
h.write('\n')
# set the random state of the sampler
# normally one could give this as an argument to the sample method
# but PTSampler didn't historically accept that...
if random_state is not None:
rstate0 = check_random_state(random_state).get_state()
self._state.random_state = rstate0
if isinstance(self.sampler, PTSampler):
self.sampler._random = rstate0
# remove chains from each of the parameters because they slow down
# pickling but only if they are parameter objects.
flat_params = f_unique(flatten(self.objective.parameters))
flat_params = [param for param in flat_params if is_parameter(param)]
# zero out all the old parameter stderrs
for param in flat_params:
param.stderr = None
param.chain = None
# make sure the checkpoint file exists
if f is not None:
with possibly_open_file(f, 'w') as h:
# write the shape of each step of the chain
h.write('# ')
shape = self._state.coords.shape
h.write(', '.join(map(str, shape)))
h.write('\n')
# using context manager means we kill off zombie pool objects
# but does mean that the pool has to be specified each time.
with possibly_create_pool(pool) as g, possibly_open_file(f, 'a') as h:
# if you're not creating more than 1 thread, then don't bother with
# a pool.
if pool == 1:
self.sampler.pool = None
else:
self.sampler.pool = g
# these kwargs are provided to the sampler.sample method
kwargs = {'iterations': steps,
'thin': nthin}
# new emcee arguments
sampler_args = getargspec(self.sampler.sample).args
if 'progress' in sampler_args and verbose:
kwargs['progress'] = True
verbose = False
if 'thin_by' in sampler_args:
kwargs['thin_by'] = nthin
kwargs.pop('thin', 0)
# ptemcee returns coords, lnprob
# emcee returns a State object
if isinstance(self.sampler, PTSampler):
for result in self.sampler.sample(self._state.coords,
**kwargs):
self._state = State(result[0],
log_prob=result[1] + result[2],
random_state=self.sampler._random)
_callback_wrapper(self._state, h=h)
else:
for state in self.sampler.sample(self._state,
**kwargs):
self._state = state
_callback_wrapper(state, h=h)
self.sampler.pool = None
# finish off the progress bar
if verbose:
sys.stdout.write("\n")
# sets parameter value and stderr
return process_chain(self.objective, self.chain)
def fit(self, method='L-BFGS-B', **kws):
"""
Obtain the maximum log-likelihood estimate (mode) of the objective. For
a least-squares objective this would correspond to lowest chi2.
Parameters
----------
method : str
which method to use for the optimisation. One of:
- `'least_squares'`: `scipy.optimize.least_squares`.
- `'L-BFGS-B'`: L-BFGS-B
- `'differential_evolution'`: differential evolution
You can also choose many of the minimizers from
``scipy.optimize.minimize``.
kws : dict
Additional arguments are passed to the underlying minimization
method.
Returns
-------
result, covar : OptimizeResult, np.ndarray
`result.x` contains the best fit parameters
`result.covar` is the covariance matrix for the fit.
`result.stderr` is the uncertainties on each of the fit parameters.
Notes
-----
If the `objective` supplies a `residuals` method then `least_squares`
can be used. Otherwise the `nll` method of the `objective` is
minimised. Use this method just before a sampling run.
If `self.objective.parameters` is a `Parameters` instance, then each
of the varying parameters has its value updated by the fit, and each
`Parameter` has a `stderr` attribute which represents the uncertainty
on the fit parameter.
"""
_varying_parameters = self.objective.varying_parameters()
init_pars = np.array(_varying_parameters)
_min_kws = {}
_min_kws.update(kws)
_bounds = bounds_list(self.objective.varying_parameters())
_min_kws['bounds'] = _bounds
# least_squares Trust Region Reflective by default
if method == 'least_squares':
b = np.array(_bounds)
_min_kws['bounds'] = (b[..., 0], b[..., 1])
res = least_squares(self.objective.residuals,
init_pars,
**_min_kws)
# differential_evolution requires lower and upper bounds
elif method == 'differential_evolution':
res = differential_evolution(self.objective.nll,
**_min_kws)
else:
# otherwise stick it to minimizer. Default being L-BFGS-B
_min_kws['method'] = method
_min_kws['bounds'] = _bounds
res = minimize(self.objective.nll, init_pars, **_min_kws)
if res.success:
self.objective.setp(res.x)
# Covariance matrix estimation
covar = self.objective.covar()
errors = np.sqrt(np.diag(covar))
res['covar'] = covar
res['stderr'] = errors
# check if the parameters are all Parameter instances.
flat_params = list(f_unique(flatten(
self.objective.parameters)))
if np.all([is_parameter(param) for param in flat_params]):
# zero out all the old parameter stderrs
for param in flat_params:
param.stderr = None
param.chain = None
for i, param in enumerate(_varying_parameters):
param.stderr = errors[i]
# need to touch up the output to check we leave
# parameters as we found them
self.objective.setp(res.x)
return res
def sample(
self,
steps,
nthin=1,
random_state=None,
f=None,
callback=None,
verbose=True,
pool=-1,
):
"""
Performs sampling from the objective.
Parameters
----------
steps : int
Collect `steps` samples into the chain. The sampler will run a
total of `steps * nthin` moves.
nthin : int, optional
Each chain sample is separated by `nthin` iterations.
random_state : {int, `np.random.RandomState`, `np.random.Generator`}
If `random_state` is not specified the `~np.random.RandomState`
singleton is used.
If `random_state` is an int, a new ``RandomState`` instance is
used, seeded with random_state.
If `random_state` is already a ``RandomState`` or a ``Generator``
instance, then that object is used.
Specify `random_state` for repeatable minimizations.
f : file-like or str
File to incrementally save chain progress to. Each row in the file
is a flattened array of size `(nwalkers, ndim)` or
`(ntemps, nwalkers, ndim)`. There are `steps` rows in the
file.
callback : callable
callback function to be called at each iteration step. Has the
signature `callback(coords, logprob)`.
verbose : bool, optional
Gives updates on the sampling progress
pool : int or map-like object, optional
If `pool` is an `int` then it specifies the number of threads to
use for parallelization. If `pool == -1`, then all CPU's are used.
If pool is a map-like callable that follows the same calling
sequence as the built-in map function, then this pool is used for
parallelisation.
Notes
-----
Please see :class:`emcee.EnsembleSampler` for its detailed behaviour.
>>> # we'll burn the first 500 steps
>>> fitter.sample(500)
>>> # after you've run those, then discard them by resetting the
>>> # sampler.
>>> fitter.sampler.reset()
>>> # Now collect 40 steps, each step separated by 50 sampler
>>> # generations.
>>> fitter.sample(40, nthin=50)
One can also burn and thin in `Curvefitter.process_chain`.
"""
self._check_vars_unchanged()
# setup a random number generator
rng = check_random_state(random_state)
if self._state is None:
self.initialise(random_state=rng)
# for saving progress to file
def _callback_wrapper(state, h=None):
if callback is not None:
callback(state.coords, state.log_prob)
if h is not None:
h.write(" ".join(map(str, state.coords.ravel())))
h.write("\n")
# remove chains from each of the parameters because they slow down
# pickling but only if they are parameter objects.
flat_params = f_unique(flatten(self.objective.parameters))
flat_params = [param for param in flat_params if is_parameter(param)]
# zero out all the old parameter stderrs
for param in flat_params:
param.stderr = None
param.chain = None
# make sure the checkpoint file exists
if f is not None:
with possibly_open_file(f, "w") as h:
# write the shape of each step of the chain
h.write("# ")
shape = self._state.coords.shape
h.write(", ".join(map(str, shape)))
h.write("\n")
# set the random state of the sampler
# normally one could give this as an argument to the sample method
# but PTSampler didn't historically accept that...
if isinstance(rng, np.random.RandomState):
rstate0 = rng.get_state()
self._state.random_state = rstate0
self.sampler.random_state = rstate0
# using context manager means we kill off zombie pool objects
# but does mean that the pool has to be specified each time.
with MapWrapper(pool) as g, possibly_open_file(f, "a") as h:
# these kwargs are provided to the sampler.sample method
kwargs = {"iterations": steps, "thin": nthin}
# if you're not creating more than 1 thread, then don't bother with
# a pool.
if isinstance(self.sampler, emcee.EnsembleSampler):
if pool == 1:
self.sampler.pool = None
else:
self.sampler.pool = g
else:
kwargs["mapper"] = g
# new emcee arguments
sampler_args = getargspec(self.sampler.sample).args
if "progress" in sampler_args and verbose:
kwargs["progress"] = True
verbose = False
if "thin_by" in sampler_args:
kwargs["thin_by"] = nthin
kwargs.pop("thin", 0)
# perform the sampling
for state in self.sampler.sample(self._state, **kwargs):
self._state = state
_callback_wrapper(state, h=h)
if isinstance(self.sampler, emcee.EnsembleSampler):
self.sampler.pool = None
# sets parameter value and stderr
return process_chain(self.objective, self.chain)
def logp(self):
# logp for all the parameters
return np.sum([param.logp() for param in f_unique(flatten(self.data))
if param.vary])
def process_chain(objective, chain, nburn=0, nthin=1, flatchain=False):
"""
Process the chain produced by a sampler for a given Objective
Parameters
----------
objective : refnx.analysis.Objective
The Objective function that the Posterior was sampled for
chain : array
The MCMC chain
nburn : int, optional
discard this many steps from the start of the chain
nthin : int, optional
only accept every `nthin` samples from the chain
flatchain : bool, optional
collapse the walkers down into a single dimension.
Returns
-------
[(param, stderr, chain)] : list
List of (param, stderr, chain) tuples.
If `isinstance(objective.parameters, Parameters)` then `param` is a
`Parameter` instance. `param.value`, `param.stderr` and
`param.chain` will contain the median, stderr and chain samples,
respectively. Otherwise `param` will be a float representing the
median of the chain samples.
`stderr` is the half width of the [15.87, 84.13] spread (similar to
standard deviation) and `chain` is an array containing the MCMC
samples for that parameter.
Notes
-----
The chain should have the shape `(iterations, nwalkers, nvary)` or
`(iterations, ntemps, nwalkers, nvary)` if parallel tempering was
employed.
The burned and thinned chain is created via:
`chain[nburn::nthin]`.
Note, if parallel tempering is employed, then only the lowest temperature
of the parallel tempering chain is processed and returned as it
corresponds to the (lowest energy) target distribution.
If `flatten is True` then the burned/thinned chain is reshaped and
`arr.reshape(-1, nvary)` is returned.
This function has the effect of setting the parameter stderr's.
"""
chain = chain[nburn::nthin]
shape = chain.shape
nvary = shape[-1]
# nwalkers = shape[1]
if len(shape) == 4:
ntemps = shape[1]
elif len(shape) == 3:
ntemps = -1
if ntemps != -1:
# PTSampler, we require the target distribution in the first row.
chain = chain[:, 0]
_flatchain = chain.reshape((-1, nvary))
if flatchain:
chain = _flatchain
flat_params = list(f_unique(flatten(objective.parameters)))
varying_parameters = objective.varying_parameters()
# set the stderr of each of the Parameters
result_list = []
if np.all([is_parameter(param) for param in flat_params]):
# zero out all the old parameter stderrs
for param in flat_params:
param.stderr = None
param.chain = None
# do the error calcn for the varying parameters and set the chain
quantiles = np.percentile(_flatchain, [15.87, 50, 84.13], axis=0)
for i, param in enumerate(varying_parameters):
std_l, median, std_u = quantiles[:, i]
param.value = median
param.stderr = 0.5 * (std_u - std_l)
# copy in the chain
param.chain = np.copy(chain[..., i])
res = MCMCResult(
name=param.name,
param=param,
median=param.value,
stderr=param.stderr,
chain=param.chain,
)
result_list.append(res)
fitted_values = np.array(varying_parameters)
# give each constrained param a chain (to be reshaped later)
constrained_params = [
param for param in flat_params if param.constraint is not None
]
for constrain_param in constrained_params:
constrain_param.chain = np.empty(chain.shape[:-1], float)
# now iterate through the varying parameters, set the values, thereby
# setting the constraint value
if len(constrained_params):
for index in np.ndindex(chain.shape[:-1]):
# iterate over parameter vectors
pvals = chain[index]
objective.setp(pvals)
for constrain_param in constrained_params:
constrain_param.chain[index] = constrain_param.value
for constrain_param in constrained_params:
quantiles = np.percentile(constrain_param.chain,
[15.87, 50, 84.13])
std_l, median, std_u = quantiles
constrain_param.value = median
constrain_param.stderr = 0.5 * (std_u - std_l)
# now reset fitted parameter values (they would've been changed by
# constraints calculations
objective.setp(fitted_values)
# the parameter set are not Parameter objects, an array was probably
# being used with BaseObjective.
else:
for i in range(nvary):
c = np.copy(chain[..., i])
median, stderr = uncertainty_from_chain(c)
res = MCMCResult(name="",
param=median,
median=median,
stderr=stderr,
chain=c)
result_list.append(res)
return result_list
def lnprob(self):
# lnprob for all the parameters
return np.sum([param.lnprob() for param in f_unique(flatten(self.data))
if param.vary])
