Source code for gammapy.utils.fitting.parameter

# Licensed under a 3-clause BSD style license - see LICENSE.rst
"""Model parameter classes."""
import copy
import numpy as np
from astropy import units as u
from astropy.units.core import UnitConversionError
from astropy.table import Table
from ..array import check_type

__all__ = ["Parameter", "Parameters"]


[docs]class Parameter: """A model parameter. Note that the parameter value has been split into a factor and scale like this:: value = factor x scale Users should interact with the ``value``, ``quantity`` or ``min`` and ``max`` properties and consider the fact that there is a ``factor``` and ``scale`` an implementation detail. That was introduced for numerical stability in parameter and error estimation methods, only in the Gammapy optimiser interface do we interact with the ``factor``, ``factor_min`` and ``factor_max`` properties, i.e. the optimiser "sees" the well-scaled problem. Parameters ---------- name : str Name factor : float or `~astropy.units.Quantity` Factor scale : float, optional Scale (sometimes used in fitting) unit : `~astropy.units.Unit` or str, optional Unit min : float, optional Minimum (sometimes used in fitting) max : float, optional Maximum (sometimes used in fitting) frozen : bool, optional Frozen? (used in fitting) """ __slots__ = ["_name", "_factor", "_scale", "_unit", "_min", "_max", "_frozen"] def __init__( self, name, factor, unit="", scale=1, min=np.nan, max=np.nan, frozen=False ): self.name = name self.scale = scale if isinstance(factor, u.Quantity) or isinstance(factor, str): val = u.Quantity(factor) self.value = val.value self.unit = val.unit else: self.factor = factor self.unit = unit self.min = min self.max = max self.frozen = frozen @property def name(self): """Name (str).""" return self._name @name.setter def name(self, val): self._name = check_type(val, "str") @property def factor(self): """Factor (float).""" return self._factor @factor.setter def factor(self, val): self._factor = float(val) @property def scale(self): """Scale (float).""" return self._scale @scale.setter def scale(self, val): self._scale = float(val) @property def unit(self): """Unit (`~astropy.units.Unit`).""" return self._unit @unit.setter def unit(self, val): self._unit = u.Unit(val) @property def min(self): """Minimum (float).""" return self._min @min.setter def min(self, val): self._min = float(val) @property def factor_min(self): """Factor min (float). This ``factor_min = min / scale`` is for the optimizer interface. """ return self.min / self.scale @property def max(self): """Maximum (float).""" return self._max @max.setter def max(self, val): self._max = float(val) @property def factor_max(self): """Factor max (float). This ``factor_max = max / scale`` is for the optimizer interface. """ return self.max / self.scale @property def frozen(self): """Frozen? (used in fitting) (bool).""" return self._frozen @frozen.setter def frozen(self, val): self._frozen = check_type(val, "bool") @property def value(self): """Value = factor x scale (float).""" return self._factor * self._scale @value.setter def value(self, val): self._factor = float(val) / self._scale @property def quantity(self): """Value times unit (`~astropy.units.Quantity`).""" return self.value * self.unit @quantity.setter def quantity(self, val): val = u.Quantity(val) try: val.to(self.unit) self.value = val.value self.unit = val.unit except UnitConversionError: raise UnitConversionError( "{0} parameter must have units homogeneous with {1}".format( self.name, self.unit ) ) def __repr__(self): return ( "Parameter(name={name!r}, value={value!r}, " "factor={factor!r}, scale={scale!r}, unit={unit!r}, " "min={min!r}, max={max!r}, frozen={frozen!r})" ).format(**self.to_dict())
[docs] def to_dict(self, selection="all"): """Convert to dict. Parameters ----------- selection : {"all", "simple"} Selection of information to include """ if selection == "simple": return dict( name=self.name, value=self.value, unit=self.unit.to_string("fits"), frozen=self.frozen, ) elif selection == "all": return dict( name=self.name, value=self.value, factor=self.factor, scale=self.scale, unit=self.unit.to_string("fits"), min=self.min, max=self.max, frozen=self.frozen, ) else: raise ValueError("Invalid selection: {!r}".format(selection))
[docs] def autoscale(self, method="scale10"): """Autoscale the parameters. Set ``factor`` and ``scale`` according to ``method`` Available methods: * ``scale10`` sets ``scale`` to power of 10, so that abs(factor) is in the range 1 to 10 * ``factor1`` sets ``factor, scale = 1, value`` In both cases the sign of value is stored in ``factor``, i.e. the ``scale`` is always positive. Parameters ---------- method : {'factor1', 'scale10'} Method to apply """ if method == "scale10": value = self.value if value != 0: exponent = np.floor(np.log10(np.abs(value))) scale = np.power(10.0, exponent) self.factor = value / scale self.scale = scale elif method == "factor1": self.factor, self.scale = 1, self.value else: raise ValueError("Invalid method: {}".format(method))
[docs]class Parameters: """List of `Parameter`. Holds covariance matrix. Parameters ---------- parameters : list of `Parameter` List of parameters covariance : `~numpy.ndarray`, optional Parameters covariance matrix. Order of values as specified by `parameters`. apply_autoscale : bool, optional Flag for optimizers, if True parameters are autoscaled before the fit, see `~gammapy.utils.modeling.Parameter.autoscale` """ def __init__(self, parameters=None, covariance=None, apply_autoscale=True): if parameters is None: parameters = [] self._parameters = self._filter_unique_parameters(parameters) self.covariance = covariance self.apply_autoscale = apply_autoscale @staticmethod def _filter_unique_parameters(parameters): """Filter unique parameters from a list of parameters""" unique_parameters = [] for par in parameters: if par not in unique_parameters: unique_parameters.append(par) return unique_parameters def _init_covariance(self): if self.covariance is None: shape = (len(self.parameters), len(self.parameters)) self.covariance = np.zeros(shape)
[docs] def copy(self): """A deep copy""" return copy.deepcopy(self)
@property def parameters(self): """List of `Parameter`.""" return self._parameters @property def free_parameters(self): """List of free parameters""" return [par for par in self.parameters if not par.frozen] # TODO: replace this with a better API to update parameters @parameters.setter def parameters(self, vals): self._parameters = vals @property def names(self): """List of parameter names""" return [par.name for par in self.parameters] def __str__(self): str_ = self.__class__.__name__ + "\n\n" for par in self.parameters: if par.name == "amplitude": line = "\t{:12} {:11}: {:.2e} {} {}\n" else: line = "\t{:12} {:11}: {:.3f} {} {}\n" frozen = "(frozen)" if par.frozen else "" try: error = "+/- {:.2f}".format(self.get_error(par)) except AttributeError: error = "" str_ += line.format(par.name, frozen, par.value, error, par.unit) return str_ def _get_idx(self, val): """Get position index for a given parameter. The input can be a parameter object, parameter name (str) or if a parameter index (int) is passed in, it is simply returned. """ if isinstance(val, int): return val elif isinstance(val, Parameter): return self.parameters.index(val) elif isinstance(val, str): for idx, par in enumerate(self.parameters): if val == par.name: return idx raise IndexError("No parameter: {!r}".format(val)) else: raise TypeError("Invalid type: {!r}".format(type(val))) def __getitem__(self, name): """Access parameter by name or index""" idx = self._get_idx(name) return self.parameters[idx]
[docs] def to_dict(self, selection="all"): data = dict(parameters=[], covariance=None) for par in self.parameters: data["parameters"].append(par.to_dict(selection)) if self.covariance is not None: data["covariance"] = self.covariance.tolist() return data
[docs] def to_table(self): """Convert parameter attributes to `~astropy.table.Table`.""" t = Table() t["name"] = [p.name for p in self.parameters] t["value"] = [p.value for p in self.parameters] if self.covariance is None: t["error"] = np.nan else: t["error"] = [self.error(idx) for idx in range(len(self.parameters))] t["unit"] = [p.unit.to_string("fits") for p in self.parameters] t["min"] = [p.min for p in self.parameters] t["max"] = [p.max for p in self.parameters] t["frozen"] = [p.frozen for p in self.parameters] for name in ["value", "error", "min", "max"]: t[name].format = ".3e" return t
[docs] @classmethod def from_dict(cls, data): parameters = [] for par in data["parameters"]: parameter = Parameter( name=par["name"], factor=float(par["value"]), unit=par.get("unit", ""), min=float(par.get("min", np.nan)), max=float(par.get("max", np.nan)), frozen=par.get("frozen", False), ) parameters.append(parameter) try: covariance = np.array(data["covariance"]) except KeyError: covariance = None return cls(parameters=parameters, covariance=covariance)
[docs] def covariance_to_table(self): """Convert covariance matrix to `~astropy.table.Table`.""" if self.covariance is None: raise ValueError("No covariance available") table = Table() table["name"] = self.names for idx, par in enumerate(self.parameters): vals = self.covariance[idx] table[par.name] = vals table[par.name].format = ".3e" return table
@property def _ufloats(self): """Return dict of ufloats with covariance.""" from uncertainties import correlated_values values = [_.value for _ in self.parameters] try: # convert existing parameters to ufloats uarray = correlated_values(values, self.covariance) except np.linalg.LinAlgError: raise ValueError("Covariance matrix not set.") upars = {} for par, upar in zip(self.parameters, uarray): upars[par.name] = upar return upars # TODO: deprecate or remove this?
[docs] def set_parameter_errors(self, errors): """Set uncorrelated parameters errors. Parameters ---------- errors : dict of `~astropy.units.Quantity` Dict of parameter errors. """ diag = [] for par in self.parameters: error = errors.get(par.name, 0) error = u.Quantity(error, par.unit).value diag.append(error) self.covariance = np.diag(np.power(diag, 2))
# TODO: this is a temporary solution until we have a better way # to handle covariance matrices via a class
[docs] def error(self, parname): """Get parameter error. Parameters ---------- parname : str, int Parameter name or index """ if self.covariance is None: raise ValueError("Covariance matrix not set.") idx = self._get_idx(parname) return np.sqrt(self.covariance[idx, idx])
# TODO: this is a temporary solution until we have a better way # to handle covariance matrices via a class
[docs] def set_error(self, parname, err): """Set parameter error. Parameters ---------- parname : str, int Parameter name or index err : float or Quantity Parameter error """ self._init_covariance() idx = self._get_idx(parname) err = u.Quantity(err, self[idx].unit).value self.covariance[idx, idx] = err ** 2
@property def correlation(self): r"""Correlation matrix (`numpy.ndarray`). Correlation :math:`C` is related to covariance :math:`\Sigma` via: .. math:: C_{ij} = \frac{ \Sigma_{ij} }{ \sqrt{\Sigma_{ii} \Sigma_{jj}} } """ err = np.sqrt(np.diag(self.covariance)) return self.covariance / np.outer(err, err)
[docs] def set_parameter_factors(self, factors): """Set factor of all parameters. Used in the optimizer interface. """ idx = 0 for parameter in self.parameters: if not parameter.frozen: parameter.factor = factors[idx] idx += 1
@property def _scale_matrix(self): scales = [par.scale for par in self.parameters] return np.outer(scales, scales) def _expand_factor_matrix(self, matrix): """Expand covariance matrix with zeros for frozen parameters""" shape = (len(self.parameters), len(self.parameters)) matrix_expanded = np.zeros(shape) mask = np.array([par.frozen for par in self.parameters]) free_parameters = ~(mask | mask[:, np.newaxis]) matrix_expanded[free_parameters] = matrix.ravel() return matrix_expanded
[docs] def set_covariance_factors(self, matrix): """Set covariance from factor covariance matrix. Used in the optimizer interface. """ if not np.sqrt(matrix.size) == len(self.parameters): matrix = self._expand_factor_matrix(matrix) self.covariance = self._scale_matrix * matrix
[docs] def autoscale(self, method="scale10"): """Autoscale all parameters. See :func:`~gammapy.utils.modelling.Parameter.autoscale` Parameters ---------- method : {'factor1', 'scale10'} Method to apply """ for par in self.parameters: par.autoscale(method)
@property def restore_values(self): """Context manager to restore values. A copy of the values is made on enter, and those values are restored on exit. Examples -------- :: from gammapy.spectrum.models import PowerLaw pwl = PowerLaw(index=2) with pwl.parameters.restore_values: pwl.parameters["index"].value = 3 print(pwl.parameters["index"].value) """ return restore_parameters_values(self)
[docs] def freeze_all(self): """Freeze all parameters""" for par in self.parameters: par.frozen = True
class restore_parameters_values: def __init__(self, parameters): self.parameters = parameters self.values = [_.value for _ in parameters] self.frozen = [_.frozen for _ in parameters] def __enter__(self): pass def __exit__(self, type, value, traceback): for value, par, frozen in zip(self.values, self.parameters, self.frozen): par.value = value par.frozen = frozen