"""Base class for optimization/retrieval."""
import math
import typing as t
from dataclasses import dataclass
import numpy as np
import numpy.typing as npt
from taurex import OutputSize
from taurex.cache import GlobalCache
from taurex.core import DerivedType, FittingType
from taurex.core.priors import LogUniform, Prior, PriorMode, Uniform
from taurex.data.citation import Citable
from taurex.log import Logger, disableLogging, enableLogging
from taurex.model import ForwardModel
from taurex.output import OutputGroup
from taurex.spectrum import BaseSpectrum
from taurex.types import AnyValType
SQRTPI = np.sqrt(2 * np.pi)
[docs]
class FitParamOutput(t.TypedDict):
"""Dictionary for storing fit parameter output."""
mean: float
sigma: float
value: float
sigma_m: float
sigma_p: float
trace: float
map: float
[docs]
@dataclass
class FitParam:
"""Holds information about a fitting parameter."""
name: str
latex: str
fget: t.Callable[[], float]
fset: t.Callable[[float], None]
mode: t.Literal["linear", "log"]
to_fit: bool
bounds: t.Tuple[float, float]
prior: Prior = None
@property
def fit_prior(self):
"""Get prior for fitting."""
if self.prior is not None:
return self.prior
if self.mode == "linear":
return Uniform(bounds=self.bounds)
return LogUniform(lin_bounds=self.bounds)
@property
def value(self) -> float:
"""Get value of parameter."""
return self.fget()
@value.setter
def value(self, value: float) -> None:
"""Set value of parameter.
Parameters
----------
value : float
Value to set parameter to
"""
self.fset(self.fit_prior.prior(value))
@property
def fit_name(self) -> str:
"""Name for fitting."""
return (
self.name
if self.fit_prior.priorMode is PriorMode.LINEAR
else f"log_{self.name}"
)
@property
def fit_latex(self) -> str:
"""Latex for fitting."""
return (
self.latex
if self.fit_prior.priorMode is PriorMode.LINEAR
else f"log({self.latex})"
)
@property
def fit_value(self) -> float:
"""Get value of parameter considering its mode."""
import math
return (
self.value
if self.fit_prior.priorMode is PriorMode.LINEAR
else math.log10(self.value)
)
@property
def boundaries(self):
"""Get boundaries of parameter."""
return self.fit_prior.boundaries()
[docs]
@dataclass
class DerivedParam:
"""Holds information about a derived parameter."""
name: str
latex: str
fget: t.Callable[[], float]
compute: bool
[docs]
def compile_params(
fitparams: t.Dict[str, FittingType],
driveparams: t.Dict[str, DerivedType],
fit_priors: t.Dict[str, Prior] = None,
):
"""Compile fitting and derived parameters."""
fitting_parameters = [
FitParam(*params, prior=fit_priors.get(params[0]))
for params in fitparams.values()
]
derived_parameters = [DerivedParam(*params) for params in driveparams.values()]
return fitting_parameters, derived_parameters
[docs]
class Optimizer(Logger, Citable):
"""A base class that handles fitting and optimization of forward models.
The class handles the compiling and management of fitting parameters in
forward models, in its current form it cannot fit and requires a class
derived from it to implement the :func:`compute_fit` function.
"""
def __init__(
self,
name: str,
observed: t.Optional[BaseSpectrum] = None,
model: t.Optional[ForwardModel] = None,
sigma_fraction: t.Optional[float] = 0.1,
) -> None:
"""Initilize optimizer.
Parameters
----------
name:
Name of optimizer for logging
observed:
Observed spectrum
model:
Model to be optimized
sigma_fraction:
Fraction of weights to use in computing the error. (Default: 10%)
"""
super().__init__(name)
self._model_callback = None
self._sigma_fraction = sigma_fraction
self._fit_priors = {}
self.avail_fit_parameters: t.Dict[str, FitParam] = {}
self.avail_derived_parameters: t.Dict[str, DerivedParam] = {}
self._model: ForwardModel = None
self._observed: BaseSpectrum = None
self.set_model(model)
self.set_observed(observed)
[docs]
def set_model(self, model: ForwardModel) -> None:
"""Sets the model to be optimized/fit
Parameters
----------
model:
The forward model we wish to optimize
"""
self._model: ForwardModel = model
self._compile_params()
[docs]
def set_observed(self, observed: BaseSpectrum) -> None:
"""Sets the observation to optimize the model to.
Parameters
----------
observed:
Observed spectrum we will optimize to
"""
self._observed: BaseSpectrum = observed
if observed is not None:
self._binner = observed.create_binner()
self._compile_params()
[docs]
def compile_params(self) -> None:
"""Dummy, does nothing and will be depcreated"""
import warnings
warnings.warn(
"compile_params is deprecated and will be removed in future versions. ",
DeprecationWarning,
stacklevel=2,
)
def _compile_params(self) -> None:
"""Compile parameters for fitting from model and observation."""
self.info("Initializing parameters")
model_fit, model_derive = [], []
if self._model is not None:
(
model_fit,
model_derive,
) = compile_params(
self._model.fittingParameters,
self._model.derivedParameters,
self._fit_priors,
)
obs_fit, obs_deriv = [], []
if self._observed is not None:
obs_fit, obs_deriv = compile_params(
self._observed.fittingParameters,
self._observed.derivedParameters,
self._fit_priors,
)
fitting_params = model_fit + obs_fit
derived_params = model_derive + obs_deriv
# self.fitting_priors.extend(obs_prior)
self.avail_fit_parameters = {p.name: p for p in fitting_params}
self.avail_derived_parameters = {p.name: p for p in derived_params}
self.info("-------FITTING---------------")
self.info("Fitting Parameters available:")
for params in self.avail_fit_parameters.values():
self.info(
"%s, Value: %s Type: %s, Params: %s",
params.name,
params.fget(),
params.fit_prior.__class__.__name__,
params.fit_prior.params(),
)
self.info("-------DERIVED---------------")
self.info("Derived Parameters available:")
for params in self.avail_derived_parameters.values():
self.info("%s, Value: %s", params.name, params.fget())
@property
def fitting_parameters(self) -> t.List[FitParam]:
"""Returns the list of fitting parameters."""
return [f for f in self.avail_fit_parameters.values() if f.to_fit]
@property
def derived_parameters(self) -> t.List[DerivedParam]:
"""Returns the list of derived parameters."""
return [f for f in self.avail_derived_parameters.values() if f.compute]
@property
def fitting_priors(self) -> t.List[Prior]:
"""Returns the list of fitting priors.
This is mostly for compatibility with the old code.
"""
return [f.fit_prior for f in self.fitting_parameters]
[docs]
def log_likelihood(self, parameters: npt.ArrayLike) -> float:
r"""Log likelihood function.
Computed as:
.. math::
\log \mathcal{L} = -\sum_i \sigma_i \sqrt{2\pi}) - \frac{1}{2} \chi^2
Employ mixins to override this method.
"""
parameters = np.asarray(parameters)
self.update_model(parameters)
data = self._observed.spectrum
datastd = self._observed.errorBar
chi_t = self.chisq_trans(parameters, data, datastd)
loglike = -np.sum(np.log(datastd * SQRTPI)) - 0.5 * chi_t
return loglike
[docs]
def update_model(self, fit_params: t.List[float]) -> None:
"""Updates the model with new parameters
Parameters
----------
fit_params : :obj:`list`
A list of new values to apply to the model. The list of values are
assumed to be in the same order as the parameters given by
:func:`fit_names`
"""
if len(fit_params) != len(self.fitting_parameters):
self.error(
"Trying to update model with more fitting parameters" " than enabled"
)
self.error(
f"No. enabled parameters:{len(self.fitting_parameters)}"
f" Update length: {len(fit_params)}"
)
raise ValueError(
"Trying to update model with more fitting" " parameters than enabled"
)
for value, param in zip(fit_params, self.fitting_parameters):
param.value = value
@property
def fit_values_nomode(self) -> t.List[float]:
"""Returns a list of the current values of a fitting parameter.
Regardless of the ``mode`` setting
Returns
-------
:obj:`list`:
List of each value of a fitting parameter
"""
return [c.value for c in self.fitting_parameters]
@property
def fit_values(self) -> t.List[float]:
"""Returns a list of the current values of a fitting parameter.
This respects the ``mode`` setting
Returns
-------
:obj:`list`:
List of each value of a fitting parameter
"""
return [c.fit_value for c in self.fitting_parameters]
@property
def fit_boundaries(self):
"""
Returns the fitting boundaries of the parameter
Returns
-------
:obj:`list`:
List of boundaries for each fitting parameter. It takes the form of
a python :obj:`tuple` with the form
( ``bound_min`` , ``bound_max`` )
"""
return [
c[-1] if c[4] == "linear" else (math.log10(c[-1][0]), math.log10(c[-1][1]))
for c in self.fitting_parameters
]
@property
def fit_names(self) -> t.List[str]:
"""Returns the names of the model parameters we will be fitting
Returns
-------
:obj:`list`:
List of names of parameters that will be fit
"""
return [c.fit_name for c in self.fitting_parameters]
@property
def fit_latex(self) -> t.List[str]:
"""Returns the names of the parameters in LaTeX format."""
return [c.fit_latex for c in self.fitting_parameters]
@property
def derived_names(self) -> t.List[str]:
"""Return names for derived parameters."""
return [c.name for c in self.derived_parameters]
@property
def derived_latex(self) -> t.List[str]:
"""Returns a list of the current values of a fitting parameter."""
return [c.latex for c in self.derived_parameters]
@property
def derived_values(self) -> t.List[float]:
"""Returns current values of derived parameters."""
return [c.fget() for c in self.derived_parameters]
[docs]
def enable_fit(self, parameter: str) -> None:
"""Enables fitting of the parameter"""
self.avail_fit_parameters[parameter].to_fit = True
[docs]
def enable_derived(self, parameter: str) -> None:
"""Enables computation of derived parameter."""
self.avail_derived_parameters[parameter].compute = True
[docs]
def disable_fit(self, parameter: str) -> None:
"""Disables fitting of the parameter."""
self.avail_fit_parameters[parameter].to_fit = False
[docs]
def disable_derived(self, parameter: str) -> None:
"""Disables computation of derived parameter."""
self.avail_derived_parameters[parameter].compute = False
[docs]
def set_boundary(self, parameter: str, new_boundaries: t.List[float]) -> None:
"""Sets the boundary of the parameter."""
self.avail_fit_parameters[parameter].bounds = new_boundaries
[docs]
def set_factor_boundary(self, parameter: str, factors: t.List[float]) -> None:
"""Sets the boundary of the parameter based on a factor of value."""
param = self.avail_fit_parameters[parameter]
bounds = (param.value * factors[0], param.value * factors[1])
param.bounds = (min(bounds), max(bounds))
[docs]
def set_mode(self, parameter: str, new_mode: t.Literal["linear", "log"]) -> None:
"""Sets the fitting mode of a parameter."""
self.avail_fit_parameters[parameter].mode = new_mode
[docs]
def set_prior(self, parameter: str, prior: Prior) -> None:
"""Sets the prior of a parameter."""
self.avail_fit_parameters[parameter].prior = prior
[docs]
def chisq_trans(
self,
fit_params: npt.ArrayLike,
data: npt.NDArray[np.float64],
datastd: npt.NDArray[np.float64],
) -> float:
"""Computes the Chi-Squared of model and observation
Computes the Chi-Squared between the forward model and
observation. The steps taken are:
1. Forward model (FM) is updated with :func:`update_model`
2. FM is then computed at its native grid then binned.
3. Chi-squared between FM and observation is computed
"""
from taurex.exceptions import InvalidModelException
#self.update_model(fit_params)
obs_bins = self._observed.wavenumberGrid
try:
_, final_model, _, _ = self._binner.bin_model(
self._model.model(wngrid=obs_bins)
)
except InvalidModelException:
return np.nan
res = (data.ravel() - final_model.ravel()) / datastd.ravel()
reject_nan = GlobalCache()["reject_spectral_nan"]
if reject_nan and np.any(np.isnan(res)):
return np.nan
res = np.nansum(res * res)
if res == 0:
res = np.nan
return res
[docs]
def compute_fit(self) -> t.Any:
"""Main compute fit function.
Unimplemented. When inheriting this should be overwritten
to perform the actual fit.
Raises
------
NotImplementedError
Raised when a derived class does override this function
"""
raise NotImplementedError
[docs]
def fit(self, output_size=OutputSize.heavy) -> t.Dict[str, AnyValType]:
"""Performs retrieval."""
import time
from tabulate import tabulate
fit_names = self.fit_names
prior_type = [p.__class__.__name__ for p in self.fitting_priors]
args = [p.params() for p in self.fitting_priors]
fit_values = self.fit_values
self.info("")
self.info("-------------------------------------")
self.info("------Retrieval Parameters-----------")
self.info("-------------------------------------")
self.info("")
self.info("Dimensionality of fit: %s", len(fit_names))
self.info("")
output = tabulate(
zip(fit_names, fit_values, prior_type, args),
headers=["Param", "Value", "Type", "Args"],
)
self.info("\n%s\n\n", output)
self.info("")
start_time = time.time()
disableLogging()
# Compute fit here
self.compute_fit()
enableLogging()
end_time = time.time()
self.info("Sampling time %s s", end_time - start_time)
solution = self.generate_solution(output_size=output_size)
self.info("")
self.info("-------------------------------------")
self.info("------Final results------------------")
self.info("-------------------------------------")
self.info("")
self.info("Dimensionality of fit: %s", len(fit_names))
self.info("")
for idx, optimized_map, optimized_median, _ in self.get_solution():
self.info("\n%s", f"---Solution {idx}------")
output = tabulate(
zip(fit_names, optimized_map, optimized_median),
headers=["Param", "MAP", "Median"],
)
self.info("\n%s\n\n", output)
return solution
[docs]
def write_optimizer(self, output: OutputGroup) -> OutputGroup:
"""Writes optimizer settings under the 'Optimizer' heading in an output file.
Parameters
----------
output:
Group (or root) in output file to write to
"""
output.write_string("optimizer", self.__class__.__name__)
output.write_string_array("fit_parameter_names", self.fit_names)
output.write_string_array("fit_parameter_latex", self.fit_latex)
output.write_array(
"fit_boundary_low",
np.array([x.boundaries()[0] for x in self.fitting_priors]),
)
output.write_array(
"fit_boundary_high",
np.array([x.boundaries()[1] for x in self.fitting_priors]),
)
if len(self.derived_names) > 0:
output.write_string_array("derived_parameter_names", self.derived_names)
output.write_string_array("derived_parameter_latex", self.derived_latex)
return output
[docs]
def write_fit(self, output: OutputGroup) -> OutputGroup:
"""Writes basic fitting parameters into output.
Parameters
----------
output:
Group (or root) in output file to write to
"""
fit = output.create_group("FitParams")
fit.write_string("fit_format", self.__class__.__name__)
fit.write_string_array("fit_parameter_names", self.fit_names)
fit.write_string_array("fit_parameter_latex", self.fit_latex)
fit.write_array(
"fit_boundary_low", np.array([x[0] for x in self.fit_boundaries])
)
fit.write_array(
"fit_boundary_high", np.array([x[1] for x in self.fit_boundaries])
)
# This is the last sampled value ... should not be recorded to avoid confusion.
# fit.write_list('fit_parameter_values',self.fit_values)
# fit.write_list('fit_parameter_values_nomode',self.fit_values_nomode)
return output
[docs]
def generate_profiles(
self, solution: int, binning: npt.NDArray[np.float64]
) -> t.Tuple[
t.Dict[str, npt.NDArray[np.float64]], t.Dict[str, npt.NDArray[np.float64]]
]:
"""Generates sigma plots for profiles"""
from taurex import mpi
sample_list: t.List[float] = []
if mpi.get_rank() == 0:
sample_list = list(self.sample_parameters(solution))
sample_list = mpi.broadcast(sample_list)
self.debug("We all got %s", sample_list)
self.info("------------Variance generation step------------------")
self.info("We are sampling %s points for the profiles", len(sample_list))
rank = mpi.get_rank()
size = mpi.nprocs()
enableLogging()
self.info(
"I will only iterate through partitioned %s "
"points (the rest is in parallel)",
len(sample_list) // size,
)
disableLogging()
def sample_iter():
count = 0
for parameters, weight in sample_list[rank::size]:
self.update_model(parameters)
enableLogging()
if rank == 0 and count % 10 == 0 and count > 0:
self.info(
"Progress %.3f", count * 100.0 / (len(sample_list) / size)
)
disableLogging()
yield weight
count += 1
return self._model.compute_error(
sample_iter, wngrid=binning, binner=self._binner
)
[docs]
def generate_solution( # noqa: C901
self, output_size=OutputSize.heavy
) -> t.Dict[str, AnyValType]:
"""Generates a dictionary with all solutions and other useful parameters."""
from taurex.util.output import store_contributions
solution_dict = {}
self.info("Post-processing - Generating spectra and profiles")
# Loop through each solution, grab optimized parameters and anything
# else we want to store
for solution, optimized_map, optimized_median, values in self.get_solution():
enableLogging()
self.info("Computing solution %s", solution)
sol_values = {}
# Include extra stuff we might want to store (provided by the child)
for k, v in values:
sol_values[k] = v
# Update the model with optimized map values
self.update_model(optimized_map)
opt_result = self._model.model(cutoff_grid=False) # Run the model
sol_values["Spectra"] = self._binner.generate_spectrum_output(
opt_result, output_size=output_size
)
try:
sol_values["Spectra"]["Contributions"] = store_contributions(
self._binner, self._model, output_size=output_size - 3
)
except Exception as e:
self.warning("Not bothering to store contributions since its broken")
self.warning("%s ", str(e))
# Update with the optimized median
self.update_model(optimized_median)
self._model.model(cutoff_grid=False)
# Store profiles here
sol_values["Profiles"] = self._model.generate_profiles()
profile_dict, spectrum_dict = self.generate_profiles(
solution, self._observed.wavenumberGrid
)
for k, v in profile_dict.items():
if k in sol_values["Profiles"]:
sol_values["Profiles"][k].update(v)
else:
sol_values["Profiles"][k] = v
for k, v in spectrum_dict.items():
if k in sol_values["Spectra"]:
sol_values["Spectra"][k].update(v)
else:
sol_values["Spectra"][k] = v
solution_dict[f"solution{solution}"] = sol_values
if len(self.derived_names) > 0:
# solution_dict[f'solution{solution}']['derived_params'] = {}
# Compute derived
for (
solution,
_,
_,
_,
) in self.get_solution():
solution_dict[f"solution{solution}"]["derived_params"] = {}
derived_dict = self.compute_derived_trace(solution)
if derived_dict is None:
continue
solution_dict[f"solution{solution}"]["derived_params"].update(
derived_dict
)
enableLogging()
self.info("Post-processing - Complete")
return solution_dict
[docs]
def compute_derived_trace(self, solution: int) -> t.Dict[str, AnyValType]:
"""Computes derived parameters from traces."""
from taurex import mpi
from taurex.util import quantile_corner
enableLogging()
samples = self.get_samples(solution)
weights = self.get_weights(solution)
len_samples = len(samples)
rank = mpi.get_rank()
num_procs = mpi.nprocs()
count = 0
derived_param = {p: ([], []) for p in self.derived_names}
if len(self.derived_names) == 0:
return
self.info("Computing derived parameters......")
disableLogging()
for idx in range(rank, len_samples, num_procs):
enableLogging()
if rank == 0 and count % 10 == 0 and count > 0:
self.info(f"Progress {idx * 100.0 / len_samples}%")
disableLogging()
parameters = samples[idx]
weight = weights[idx]
self.update_model(parameters)
self._model.initialize_profiles()
for p, v in zip(self.derived_names, self.derived_values):
derived_param[p][0].append(v)
derived_param[p][1].append(weight)
result_dict = {}
sorted_weights = weights.argsort()
for param, (trace, w) in derived_param.items():
# I cant remember why this works
all_trace = np.array(mpi.allreduce(trace, op="SUM"))
# I cant remember why this works
all_weight = np.array(mpi.allreduce(w, op="SUM"))
all_weight_sort = all_weight.argsort()
# Sort them into the right order
all_weight[sorted_weights] = all_weight[all_weight_sort]
all_trace[sorted_weights] = all_trace[all_weight_sort]
q_16, q_50, q_84 = quantile_corner(
np.array(all_trace), [0.16, 0.5, 0.84], weights=np.array(all_weight)
)
mean = np.average(all_trace, weights=all_weight, axis=0)
derived = {
"value": q_50,
"sigma_m": q_50 - q_16,
"sigma_p": q_84 - q_50,
"trace": all_trace,
"mean": mean,
}
result_dict[f"{param}_derived"] = derived
return result_dict
[docs]
def sample_parameters(
self, solution: int
) -> t.Generator[t.Tuple[npt.NDArray[np.float64], float], None, None]:
"""Read traces and weights and return a random ``sigma_fraction`` sample
Parameters
----------
solution:
a solution output from sampler
Yields
------
traces: :obj:`array`
Traces of a particular sample
weight: float
Weight of sample
"""
from taurex.util import random_int_iter
samples = self.get_samples(solution)
weights = self.get_weights(solution)
iterator = random_int_iter(samples.shape[0], self._sigma_fraction)
for x in iterator:
w = weights[x] + 1e-300
yield samples[x, :], w
[docs]
def get_solution(
self,
) -> t.Generator[
t.Tuple[
int,
npt.NDArray[np.float64],
npt.NDArray[np.float64],
t.Dict[str, AnyValType],
],
None,
None,
]:
"""Generator for solutions.
** Requires implementation **
Generator for solutions and their
median and MAP values
Yields
------
solution_no:
Solution number
map:
Map values
median:
Median values
extra:
List of tuples of extra information to store.
Must be of form ``(name, data)``
"""
raise NotImplementedError
[docs]
def get_samples(self, solution_id: int) -> npt.NDArray[np.float64]:
"""Get traces for a particular solution."""
raise NotImplementedError
[docs]
def get_weights(self, solution_id: int) -> npt.NDArray[np.float64]:
"""Get weights for a particular solution."""
raise NotImplementedError
[docs]
def write(self, output: OutputGroup) -> None:
"""Creates 'Optimizer' and writes output
Parameters
----------
output :
Group (or root) in output file to write to
"""
opt = output.create_group("Optimizer")
self.write_optimizer(opt)