class BootstrapRunner(AbstractRunner):
def __init__(self, runnerArgument):
"""
Parameters
----------
runnerArgument: RunnerArgument
Notes
-----
1. Uses METHOD_LEASTSQ for fitModel iterations.
"""
super().__init__()
#
self.lastErr = ""
self.fitter = runnerArgument.fitter
self.numIteration = runnerArgument.numIteration
self.kwargs = runnerArgument.kwargs
self.synthesizerClass = runnerArgument.synthesizerClass
if "logger" in self.fitter.__dict__.keys():
self.logger = self.fitter.logger
else:
self.logger = Logger()
self._isDone = not self._fitInitial()
self.columns = self.fitter.selectedColumns
# Initializations for bootstrap loop
if not self.isDone:
fittedTS = self.fitter.fittedTS.subsetColumns(self.columns,
isCopy=False)
self.synthesizer = self.synthesizerClass(
observedTS=self.fitter.observedTS.subsetColumns(self.columns,
isCopy=False),
fittedTS=fittedTS,
**self.kwargs)
self.numSuccessIteration = 0
if self.fitter.minimizerResult is None:
self.fitter.fitModel()
self.baseChisq = self.fitter.minimizerResult.redchi
self.curIteration = 0
self.fd = self.logger.getFileDescriptor()
self.baseFittedStatistic = TimeseriesStatistic(
self.fitter.observedTS.subsetColumns(
self.fitter.selectedColumns, isCopy=False))
def report(self, id=None):
if True:
return
if id is None:
self._startTime = time.time()
else:
elapsed = time.time() - self._startTime
print("%s: %2.3f" % (id, elapsed))
@property
def numWorkUnit(self):
return self.numIteration
@property
def isDone(self):
return self._isDone
def run(self):
"""
Runs the bootstrap.
Returns
-------
BootstrapResult
"""
def mkNullResult():
fittedStatistic = TimeseriesStatistic(
self.fitter.observedTS[self.fitter.selectedColumns])
return BootstrapResult(self.fitter, 0, {}, fittedStatistic)
#
if self.isDone:
return
# Set up logging for this run
if self.fd is not None:
sys.stderr = self.fd
sys.stdout = self.fd
isSuccess = False
bootstrapError = 0
self.report()
for _ in range(ITERATION_MULTIPLIER):
newObservedTS = self.synthesizer.calculate()
self.report("newObservedTS")
# Update fitter to use the new observed data
_ = self.fitter._updateObservedTS(newObservedTS, isCheck=False)
self.report("updated fitter")
# Try fitting
try:
self.fitter.fitModel(params=self.fitter.params)
self.report("fitter.fit")
except Exception as err:
# Problem with the fit.
msg = "modelFitterBootstrap. Fit failed on iteration %d." \
% iteration
self.logger.error(msg, err)
bootstrapError += 1
continue
# Verify that there is a result
if self.fitter.minimizerResult is None:
continue
# Check if the fit is of sufficient quality
if self.fitter.minimizerResult.redchi > MAX_CHISQ_MULT * self.baseChisq:
continue
if self.fitter.params is None:
continue
isSuccess = True
self.report("break")
break
# Create the result
if isSuccess:
self.numSuccessIteration += 1
parameterDct = {
k: [v]
for k, v in self.fitter.params.valuesdict().items()
}
fittedStatistic = self.baseFittedStatistic.copy()
fittedStatistic.accumulate(
self.fitter.fittedTS.subsetColumns(self.fitter.selectedColumns,
isCopy=False))
bootstrapResult = BootstrapResult(self.fitter,
self.numSuccessIteration,
parameterDct,
fittedStatistic,
bootstrapError=bootstrapError)
else:
bootstrapResult = mkNullResult()
self._isDone = True
# Close the logging file
if self.fd is not None:
if not self.fd.closed:
self.fd.close()
# See if completed work
if self.numSuccessIteration >= self.numIteration:
self._isDone = True
return bootstrapResult
def _fitInitial(self):
"""
Do the initial fit.
Returns
-------
bool
successful fit
"""
isSuccess = False
for _ in range(MAX_TRIES):
try:
self.fitter.fitModel() # Initialize model
isSuccess = True
break
except Exception as err:
self.lastErr = err
msg = "Could not do initial fit"
self.logger.error(msg, err)
return isSuccess
class Runner(object):
"""Runs tests on biomodels."""
def __init__(self,
firstModel: int = 210,
numModel: int = 2,
pclPath=PCL_FILE,
figPath=FIG_PATH,
useExistingData: bool = False,
isPlot=IS_PLOT,
**kwargDct):
"""
Parameters
----------
firstModel: first model to use
numModel: number of models to use
pclPath: file to which results are saved
useExistingData: use data in existing PCL file
"""
self.useExistingData = useExistingData and os.path.isfile(pclPath)
# Recover previously saved results if desired
if self.useExistingData:
self.restore(pclPath=pclPath)
else:
# Initialize based on type of context variable
for name in CONTEXT:
if name[-1:] == "s":
self.__setattr__(name, [])
elif name[-3:] == "Dct":
self.__setattr__(name, {})
elif name[-4:] == "Path":
self.__setattr__(name, None)
elif name[0:2] == "is":
self.__setattr__(name, False)
else:
self.__setattr__(name, 0)
# Initialize to parameters for this instantiation
self.firstModel = firstModel
self.numModel = numModel
self.pclPath = pclPath
self.figPath = figPath
self.kwargDct = kwargDct
self.isPlot = isPlot
self.useExistingData = useExistingData
#
if LOGGER in kwargDct.keys():
self.logger = kwargDct[LOGGER]
else:
self.logger = Logger()
kwargDct[LOGGER] = self.logger
self.save()
def _isListSame(self, list1, list2):
diff = set(list1).symmetric_difference(list2)
return len(diff) == 0
def equals(self, other):
selfKeys = list(self.__dict__.keys())
otherKeys = list(other.__dict__.keys())
if not self._isListSame(selfKeys, otherKeys):
return False
#
for key, value in self.__dict__.items():
if isinstance(value, list):
isEqual = self._isListSame(value, other.__getattribute__(key))
if not isEqual:
return False
elif any([isinstance(value, t) for t in [int, str, float, bool]]):
if self.__getattribute__(key) != other.__getattribute__(key):
return False
else:
pass
#
return True
def run(self):
"""
Runs the tests. Saves state after each tests.
"""
# Processing models
modelNums = self.firstModel + np.array(range(self.numModel))
for modelNum in modelNums:
if (modelNum in self.processedModels) and self.useExistingData:
continue
else:
self.processedModels.append(modelNum)
input_path = PATH_PAT % modelNum
msg = "Model %s" % input_path
self.logger.activity(msg)
try:
harness = TestHarness(input_path, **self.kwargDct)
if len(harness.parametersToFit) == 0:
self.logger.result("No fitable parameters in model.")
self.save()
continue
harness.evaluate(stdResiduals=1.0,
fractionParameterDeviation=1.0,
relError=2.0)
except Exception as err:
self.erroredModels.append(modelNum)
self.logger.error("TestHarness failed", err)
self.save()
continue
# Parameters for model
self.modelParameterDct[modelNum] = \
list(harness.fitModelResult.parameterRelErrorDct.keys())
# Relative error in initial fit
values = [
v for v in
harness.fitModelResult.parameterRelErrorDct.values()
]
self.fitModelRelerrors.extend(values)
# Relative error in bootstrap
values = [
v for v in
harness.bootstrapResult.parameterRelErrorDct.values()
]
self.bootstrapRelerrors.extend(values)
# Count models without exceptions
self.nonErroredModels.append(modelNum)
self.numNoError = len(self.nonErroredModels)
self.save()
# Check for plot
if self.isPlot:
self.plot()
def save(self):
"""
Saves state. Maintain in sync with self.restore().
"""
if self.pclPath is not None:
data = [self.__getattribute__(n) for n in CONTEXT]
with (open(self.pclPath, "wb")) as fd:
pickle.dump(data, fd)
def restore(self, pclPath=None):
"""
Restores state. Maintain in sync with self.save().
"""
if pclPath is None:
pclPath = self.pclPath
if os.path.isfile(pclPath):
with (open(pclPath, "rb")) as fd:
data = pickle.load(fd)
[self.__setattr__(n, v) for n, v in zip(CONTEXT, data)]
else:
raise ValueError("***Restart file %s does not exist" %
self.pclPath)
@staticmethod
def _pruneRelativeErrors(relativeErrors, maxError=MAX_RELATIVE_ERROR):
"""
Deletes Nans. Removes very large values.
Parameters
----------
list: relative errors
maxError: maximum relative error considered
Returns
-------
list: pruned errors
float: fraction pruned from non-nan values
"""
noNanErrors = [v for v in relativeErrors if not np.isnan(v)]
prunedErrors = [v for v in noNanErrors if v <= maxError]
prunedFrc = 1 - len(prunedErrors) / len(noNanErrors)
return prunedErrors, prunedFrc
def plot(self):
"""
Does all plots.
"""
_, axes = plt.subplots(1, 2)
prunedModelErrors, modelPrunedFrc = \
self._pruneRelativeErrors(self.fitModelRelerrors)
prunedBootstrapErrors, bootstrapPrunedFrc = \
self._pruneRelativeErrors(self.bootstrapRelerrors)
maxBin1 = self._plotRelativeErrors(axes[0], prunedModelErrors,
FIT_MODEL, modelPrunedFrc)
maxBin2 = self._plotRelativeErrors(axes[1],
prunedBootstrapErrors,
BOOTSTRAP,
bootstrapPrunedFrc,
isYLabel=False)
maxBin = max(maxBin1, maxBin2)
if maxBin > 0:
axes[0].set_ylim([0, maxBin])
axes[1].set_ylim([0, maxBin])
#
if len(self.processedModels) == 0:
frac = 0.0
else:
frac = 1.0 * self.numNoError / len(self.processedModels)
suptitle = "Models %d-%d. Fraction non-errored: %2.3f"
lastModel = self.firstModel + len(self.processedModels) - 1
suptitle = suptitle % (self.firstModel, lastModel, frac)
plt.suptitle(suptitle)
plt.show()
plt.savefig(self.figPath)
def _plotRelativeErrors(self,
ax,
relErrors,
title,
prunedFrc,
isYLabel=True):
"""
Plots histogram of relative errors.
Parameters
----------
ax: Matplotlib.axes
relErrors: list-float
title: str
prunedFrc: float
isYlabel: bool
Returns
-------
float: maximum number in a bin
"""
rr = ax.hist(relErrors)
fullTitle = "%s. Frc Pruned: %2.2f" % (title, prunedFrc)
ax.set_title(fullTitle)
ax.set_xlabel("relative error")
if isYLabel:
ax.set_ylabel("number parameters")
ax.set_xlim([0, 1])
return max(rr[0])
class ModelFitterCore(rpickle.RPickler):
def __init__(
self,
modelSpecification,
observedData,
parametersToFit=None,
selectedColumns=None,
fitterMethods=METHOD_FITTER_DEFAULTS,
numFitRepeat=1,
bootstrapMethods=METHOD_BOOTSTRAP_DEFAULTS,
parameterLowerBound=PARAMETER_LOWER_BOUND,
parameterUpperBound=PARAMETER_UPPER_BOUND,
parameterDct={},
fittedDataTransformDct={},
logger=Logger(),
isPlot=True,
_loggerPrefix="",
# The following must be kept in sync with ModelFitterBootstrap.bootstrap
numIteration: int = 10,
reportInterval: int = 1000,
synthesizerClass=ObservationSynthesizerRandomizedResiduals,
maxProcess: int = None,
serializePath: str = None,
):
"""
Constructs estimates of parameter values.
Parameters
----------
modelSpecification: ExtendedRoadRunner/str
roadrunner model or antimony model
observedData: NamedTimeseries/str
str: path to CSV file
parametersToFit: list-str/None
parameters in the model that you want to fit
if None, no parameters are fit
selectedColumns: list-str
species names you wish use to fit the model
default: all columns in observedData
parameterLowerBound: float
lower bound for the fitting parameters
parameterUpperBound: float
upper bound for the fitting parameters
parameterDct: dict
key: parameter name
value: triple - (lowerVange, startingValue, upperRange)
fittedDataTransformDct: dict
key: column in selectedColumns
value: function of the data in selectedColumns;
input: NamedTimeseries
output: array for the values of the column
logger: Logger
fitterMethods: str/list-str
method used for minimization in fitModel
numFitRepeat: int
number of times fitting is repeated for a method
bootstrapMethods: str/list-str
method used for minimization in bootstrap
numIteration: number of bootstrap iterations
reportInterval: number of iterations between progress reports
synthesizerClass: object that synthesizes new observations
Must subclass ObservationSynthesizer
maxProcess: Maximum number of processes to use. Default: numCPU
serializePath: Where to serialize the fitter after bootstrap
Usage
-----
parameterDct = {
"k1": (1, 5, 10), # name of parameter: low value, initial, high
"k2": (2, 3, 6)}
ftter = ModelFitter(roadrunnerModel, "observed.csv",
parameterDct=parameterDct)
fitter.fitModel() # Do the fit
fitter.bootstrap() # Estimate parameter variance with bootstrap
"""
if modelSpecification is not None:
# Not the default constructor
self._loggerPrefix = _loggerPrefix
self.modelSpecification = modelSpecification
self.parametersToFit = parametersToFit
self.lowerBound = parameterLowerBound
self.upperBound = parameterUpperBound
self.bootstrapKwargs = dict(
numIteration=numIteration,
reportInterval=reportInterval,
maxProcess=maxProcess,
serializePath=serializePath,
)
self.parameterDct = self._updateParameterDct(parameterDct)
self._numFitRepeat = numFitRepeat
if self.parametersToFit is None:
self.parametersToFit = [p for p in self.parameterDct.keys()]
self.observedTS = observedData
if self.observedTS is not None:
self.observedTS = mkNamedTimeseries(observedData)
#
self.fittedDataTransformDct = fittedDataTransformDct
#
if (selectedColumns is None) and (self.observedTS is not None):
selectedColumns = self.observedTS.colnames
self.selectedColumns = selectedColumns
# Construct array of non-nan observed values
self._observedArr = self.observedTS[self.selectedColumns].flatten()
# Other internal state
self._fitterMethods = fitterMethods
if isinstance(self._fitterMethods, str):
if self._fitterMethods == METHOD_BOTH:
self._fitterMethods = METHOD_FITTER_DEFAULTS
else:
self._fitterMethods = [self._fitterMethods]
self._bootstrapMethods = bootstrapMethods
if isinstance(self._bootstrapMethods, str):
self._bootstrapMethods = [self._bootstrapMethods]
self._isPlot = isPlot
self._plotter = tp.TimeseriesPlotter(isPlot=self._isPlot)
self._plotFittedTS = None # Timeseries that is plotted
self.logger = logger
# The following are calculated during fitting
self.roadrunnerModel = None
self.minimizer = None # lmfit.minimizer
self.minimizerResult = None # Results of minimization
self.params = None # params property in lmfit.minimizer
self.fittedTS = self.observedTS.copy(
isInitialize=True) # Initialize
self.residualsTS = None # Residuals for selectedColumns
self.bootstrapResult = None # Result from bootstrapping
# Validation checks
self._validateFittedDataTransformDct()
else:
pass
@classmethod
def rpConstruct(cls):
"""
Overrides rpickler.rpConstruct to create a method that
constructs an instance without arguments.
Returns
-------
Instance of cls
"""
return cls(None, None, None)
def rpRevise(self):
"""
Overrides rpickler.
"""
if not "logger" in self.__dict__.keys():
self.logger = Logger()
def _validateFittedDataTransformDct(self):
if self.fittedDataTransformDct is not None:
keySet = set(self.fittedDataTransformDct.keys())
selectedColumnsSet = self.selectedColumns
if (keySet is not None) and (selectedColumnsSet is not None):
excess = set(keySet).difference(selectedColumnsSet)
if len(excess) > 0:
msg = "Columns not in selectedColumns: %s" % str(excess)
raise ValueError(excess)
def _transformFittedTS(self, data):
"""
Updates the fittedTS taking into account required transformations.
Parameters
----------
data: np.ndarray
Results
----------
NamedTimeseries
"""
colnames = list(self.selectedColumns)
colnames.insert(0, TIME)
fittedTS = NamedTimeseries(array=data[:, :], colnames=colnames)
if self.fittedDataTransformDct is not None:
for column, func in self.fittedDataTransformDct.items():
if func is not None:
fittedTS[column] = func(fittedTS)
return fittedTS
def _updateParameterDct(self, parameterDct):
"""
Handles values that are tuples instead of ParameterSpecification.
"""
dct = dict(parameterDct)
for name, value in parameterDct.items():
if isinstance(value, tuple):
dct[name] = ParameterSpecification(lower=value[0],
upper=value[1],
value=value[2])
return dct
@staticmethod
def addParameter(parameterDct: dict, name: str, lower: float, upper: float,
value: float):
"""
Adds a parameter to a list of parameters.
Parameters
----------
parameterDct: parameter dictionary to agument
name: parameter name
lower: lower range of parameter value
upper: upper range of parameter value
value: initial value
Returns
-------
dict
"""
parameterDct[name] = ParameterSpecification(lower=lower,
upper=upper,
value=value)
def _adjustNames(self, antimonyModel:str, observedTS:NamedTimeseries) \
->typing.Tuple[NamedTimeseries, list]:
"""
Antimony exports can change the names of floating species
by adding a "_" at the end. Check for this and adjust
the names in observedTS.
Return
------
NamedTimeseries: newObservedTS
list: newSelectedColumns
"""
rr = te.loada(antimonyModel)
dataNames = rr.simulate().colnames
names = ["[%s]" % n for n in observedTS.colnames]
missingNames = [n[1:-1] for n in set(names).difference(dataNames)]
newSelectedColumns = list(self.selectedColumns)
if len(missingNames) > 0:
newObservedTS = observedTS.copy()
self.logger.exception("Missing names in antimony export: %s" %
str(missingNames))
for name in observedTS.colnames:
missingName = "%s_" % name
if name in missingNames:
newObservedTS = newObservedTS.rename(name, missingName)
newSelectedColumns.remove(name)
newSelectedColumns.append(missingName)
else:
newObservedTS = observedTS
return newObservedTS, newSelectedColumns
def copy(self, isKeepLogger=False):
"""
Creates a copy of the model fitter.
Preserves the user-specified settings and the results
of bootstrapping.
"""
if not isinstance(self.modelSpecification, str):
try:
modelSpecification = self.modelSpecification.getAntimony()
except Exception as err:
self.logger.error(
"Problem wth conversion to Antimony. Details:", err)
raise ValueError("Cannot proceed.")
observedTS, selectedColumns = self._adjustNames(
modelSpecification, self.observedTS)
else:
modelSpecification = self.modelSpecification
observedTS = self.observedTS.copy()
selectedColumns = self.selectedColumns
#
if isKeepLogger:
logger = self.logger
elif self.logger is not None:
logger = self.logger.copy()
else:
logger = None
newModelFitter = self.__class__(
copy.deepcopy(modelSpecification),
observedTS,
copy.deepcopy(self.parametersToFit),
selectedColumns=selectedColumns,
fitterMethods=self._fitterMethods,
bootstrapMethods=self._bootstrapMethods,
parameterLowerBound=self.lowerBound,
parameterUpperBound=self.upperBound,
parameterDct=copy.deepcopy(self.parameterDct),
fittedDataTransformDct=copy.deepcopy(self.fittedDataTransformDct),
logger=logger,
isPlot=self._isPlot)
if self.bootstrapResult is not None:
newModelFitter.bootstrapResult = self.bootstrapResult.copy()
newModelFitter.params = newModelFitter.bootstrapResult.params
else:
newModelFitter.bootstrapResult = None
newModelFitter.params = self.params
return newModelFitter
def _initializeRoadrunnerModel(self):
"""
Sets self.roadrunnerModel.
"""
if isinstance(self.modelSpecification,
te.roadrunner.extended_roadrunner.ExtendedRoadRunner):
self.roadrunnerModel = self.modelSpecification
elif isinstance(self.modelSpecification, str):
self.roadrunnerModel = te.loada(self.modelSpecification)
else:
msg = 'Invalid model.'
msg = msg + "\nA model must either be a Roadrunner model "
msg = msg + "an Antimony model."
raise ValueError(msg)
def getDefaultParameterValues(self):
"""
Obtain the original values of parameters.
Returns
-------
dict:
key: parameter name
value: value of parameter
"""
dct = {}
self._initializeRoadrunnerModel()
self.roadrunnerModel.reset()
for parameterName in self.parametersToFit:
dct[parameterName] = self.roadrunnerModel.model[parameterName]
return dct
def simulate(self,
params=None,
startTime=None,
endTime=None,
numPoint=None):
"""
Runs a simulation. Defaults to parameter values in the simulation.
Parameters
----------
params: lmfit.Parameters
startTime: float
endTime: float
numPoint: int
Return
------
NamedTimeseries
"""
def set(default, parameter):
# Sets to default if parameter unspecified
if parameter is None:
return default
else:
return parameter
##V
block = Logger.join(self._loggerPrefix, "fitModel.simulate")
guid = self.logger.startBlock(block)
## V
sub1Block = Logger.join(block, "sub1")
sub1Guid = self.logger.startBlock(sub1Block)
startTime = set(self.observedTS.start, startTime)
endTime = set(self.observedTS.end, endTime)
numPoint = set(len(self.observedTS), numPoint)
## V
sub1aBlock = Logger.join(sub1Block, "sub1a")
sub1aGuid = self.logger.startBlock(sub1aBlock)
if self.roadrunnerModel is None:
self._initializeRoadrunnerModel()
self.roadrunnerModel.reset()
## ^
self.logger.endBlock(sub1aGuid)
## V
sub1bBlock = Logger.join(sub1Block, "sub1b")
sub1bGuid = self.logger.startBlock(sub1bBlock)
if params is not None:
# Parameters have been specified
self._setupModel(params)
## ^
self.logger.endBlock(sub1bGuid)
# Do the simulation
selectedColumns = list(self.selectedColumns)
if not TIME in selectedColumns:
selectedColumns.insert(0, TIME)
## ^
self.logger.endBlock(sub1Guid)
## V
roadrunnerBlock = Logger.join(block, "roadrunner")
roadrunnerGuid = self.logger.startBlock(roadrunnerBlock)
data = self.roadrunnerModel.simulate(startTime, endTime, numPoint,
selectedColumns)
self.logger.endBlock(roadrunnerGuid)
## ^
# Select the required columns
## V
sub2Block = Logger.join(block, "sub2")
sub2Guid = self.logger.startBlock(sub2Block)
fittedTS = NamedTimeseries(namedArray=data)
self.logger.endBlock(sub2Guid)
## ^
self.logger.endBlock(guid)
##^
return fittedTS
def updateFittedAndResiduals(self, **kwargs) -> np.ndarray:
"""
Updates values of self.fittedTS and self.residualsTS
based on self.params.
Parameters
----------
kwargs: dict
arguments for simulation
Instance Variables Updated
--------------------------
self.fittedTS
self.residualsTS
Returns
-------
1-d ndarray of residuals
"""
self.fittedTS = self.simulate(**kwargs) # Updates self.fittedTS
cols = self.selectedColumns
if self.residualsTS is None:
self.residualsTS = self.observedTS.subsetColumns(cols)
self.residualsTS[cols] = self.observedTS[cols] - self.fittedTS[cols]
for col in cols:
self.residualsTS[col] = np.nan_to_num(self.residualsTS[col])
def _residuals(self, params) -> np.ndarray:
"""
Compute the residuals between objective and experimental data
Handle nan values in observedTS. This internal-only method
is implemented to maximize efficieency.
Parameters
----------
kwargs: dict
arguments for simulation
Instance Variables Updated
--------------------------
self.residualsTS
Returns
-------
1-d ndarray of residuals
"""
block = Logger.join(self._loggerPrefix, "fitModel._residuals")
guid = self.logger.startBlock(block)
##V
self.roadrunnerModel.reset()
self._setupModel(params)
#
roadrunnerBlock = Logger.join(block, "roadrunner")
roadrunnerGuid = self.logger.startBlock(roadrunnerBlock)
## V
#
data = self.roadrunnerModel.simulate(self.observedTS.start,
self.observedTS.end,
len(self.observedTS),
self.selectedColumns)
## ^
self.logger.endBlock(roadrunnerGuid)
#
tailBlock = Logger.join(block, "tail")
tailGuid = self.logger.startBlock(tailBlock)
## V
residualsArr = self._observedArr - data.flatten()
residualsArr = np.nan_to_num(residualsArr)
## ^
self.logger.endBlock(tailGuid)
##^
self.logger.endBlock(guid)
#
# Used for detailed debugging
if False:
self.logger.details("_residuals/std(residuals): %f" %
np.std(residualsArr))
self.logger.details("_residuals/params: %s" % str(params))
return residualsArr
def fitModel(self, params: lmfit.Parameters = None, max_nfev: int = 100):
"""
Fits the model by adjusting values of parameters based on
differences between simulated and provided values of
floating species.
Parameters
----------
params: starting values of parameters
max_nfev: maximum number of function evaluations
Example
-------
f.fitModel()
"""
ParameterDescriptor = collections.namedtuple(
"ParameterDescriptor",
"params method std minimizer minimizerResult")
block = Logger.join(self._loggerPrefix, "fitModel")
guid = self.logger.startBlock(block)
self._initializeRoadrunnerModel()
if self.parametersToFit is None:
# Compute fit and residuals for base model
self.params = None
else:
if params is None:
params = self.mkParams()
# Fit the model to the data using one or more methods.
# Choose the result with the lowest residual standard deviation
paramDct = {}
for method in self._fitterMethods:
for _ in range(self._numFitRepeat):
minimizer = lmfit.Minimizer(self._residuals,
params,
max_nfev=max_nfev)
try:
minimizerResult = minimizer.minimize(method=method,
max_nfev=max_nfev)
except Exception as excp:
msg = "Error minimizing for method: %s" % method
self.logger.error(msg, excp)
continue
params = minimizerResult.params
std = np.std(self._residuals(params))
if method in paramDct.keys():
if std >= paramDct[method].std:
continue
paramDct[method] = ParameterDescriptor(
params=params.copy(),
method=method,
std=std,
minimizer=minimizer,
minimizerResult=minimizerResult,
)
if len(paramDct) == 0:
msg = "*** Minimizer failed for this model and data."
raise ValueError(msg)
# Select the result that has the smallest residuals
sortedMethods = sorted(paramDct.keys(),
key=lambda m: paramDct[m].std)
bestMethod = sortedMethods[0]
self.params = paramDct[bestMethod].params
self.minimizer = paramDct[bestMethod].minimizer
self.minimizerResult = paramDct[bestMethod].minimizerResult
# Ensure that residualsTS and fittedTS match the parameters
self.updateFittedAndResiduals(params=self.params)
self.logger.endBlock(guid)
def getFittedModel(self):
"""
Provides the roadrunner model with fitted parameters
Returns
-------
ExtendedRoadrunner
"""
self._checkFit()
self.roadrunnerModel.reset()
self._setupModel(self.params)
return self.roadrunnerModel
def _setupModel(self, params):
"""
Sets up the model for use based on the parameter parameters
Parameters
----------
params: lmfit.Parameters
"""
pp = params.valuesdict()
for parameter in self.parametersToFit:
try:
self.roadrunnerModel.model[parameter] = pp[parameter]
except Exception as err:
msg = "_modelFitterCore/_setupModel: Could not set value for %s" \
% parameter
self.logger.error(msg, err)
def mkParams(self, parameterDct: dict = None) -> lmfit.Parameters:
"""
Constructs lmfit parameters based on specifications.
Parameters
----------
parameterDct: key=name, value=ParameterSpecification
Returns
-------
lmfit.Parameters
"""
def get(value, base_value, multiplier):
if value is not None:
return value
return base_value * multiplier
#
if parameterDct is None:
parameterDct = self.parameterDct
params = lmfit.Parameters()
for parameterName in self.parametersToFit:
if parameterName in parameterDct.keys():
specification = parameterDct[parameterName]
value = get(specification.value, specification.value, 1.0)
if value > 0:
lower_factor = LOWER_PARAMETER_MULT
upper_factor = UPPER_PARAMETER_MULT
else:
upper_factor = UPPER_PARAMETER_MULT
lower_factor = LOWER_PARAMETER_MULT
lower = get(specification.lower, specification.value,
lower_factor)
upper = get(specification.upper, specification.value,
upper_factor)
if np.isclose(lower - upper, 0):
upper = 0.0001
try:
params.add(parameterName,
value=value,
min=lower,
max=upper)
except Exception as err:
msg = "modelFitterCore/mkParams parameterName %s" \
% parameterName
self.logger.error(msg, err)
else:
value = np.mean([self.lowerBound, self.upperBound])
params.add(parameterName,
value=value,
min=self.lowerBound,
max=self.upperBound)
return params
def _checkFit(self):
if self.params is None:
raise ValueError("Must use fitModel before using this method.")
def serialize(self, path):
"""
Serialize the model to a path.
Parameters
----------
path: str
File path
"""
newModelFitter = self.copy()
with open(path, "wb") as fd:
rpickle.dump(newModelFitter, fd)
@classmethod
def deserialize(cls, path):
"""
Deserialize the model from a path.
Parameters
----------
path: str
File path
Return
------
ModelFitter
Model is initialized.
"""
with open(path, "rb") as fd:
fitter = rpickle.load(fd)
fitter._initializeRoadrunnerModel()
return fitter
class ModelFitterCore(rpickle.RPickler):
# Subclasses used in interface
class OptimizerMethod():
def __init__(self, method, kwargs):
self.method = method
self.kwargs = kwargs
def __init__(
self,
modelSpecification,
observedData,
parametersToFit=None,
selectedColumns=None,
fitterMethods=None,
numFitRepeat=1,
bootstrapMethods=None,
parameterLowerBound=PARAMETER_LOWER_BOUND,
parameterUpperBound=PARAMETER_UPPER_BOUND,
parameterDct=None,
fittedDataTransformDct=None,
logger=Logger(),
isPlot=True,
_loggerPrefix="",
# The following must be kept in sync with ModelFitterBootstrap.bootstrap
numIteration: int = 10,
reportInterval: int = 1000,
maxProcess: int = None,
serializePath: str = None,
):
"""
Constructs estimates of parameter values.
Parameters
----------
modelSpecification: ExtendedRoadRunner/str
roadrunner model or antimony model
observedData: NamedTimeseries/str
str: path to CSV file
parametersToFit: list-str/None
parameters in the model that you want to fit
if None, no parameters are fit
selectedColumns: list-str
species names you wish use to fit the model
default: all columns in observedData
parameterLowerBound: float
lower bound for the fitting parameters
parameterUpperBound: float
upper bound for the fitting parameters
parameterDct: dict
key: parameter name
value: triple - (lowerVange, startingValue, upperRange)
fittedDataTransformDct: dict
key: column in selectedColumns
value: function of the data in selectedColumns;
input: NamedTimeseries
output: array for the values of the column
logger: Logger
fitterMethods: str/list-str/list-OptimizerMethod
method used for minimization in fitModel
numFitRepeat: int
number of times fitting is repeated for a method
bootstrapMethods: str/list-str/list-OptimizerMethod
method used for minimization in bootstrap
numIteration: number of bootstrap iterations
reportInterval: number of iterations between progress reports
maxProcess: Maximum number of processes to use. Default: numCPU
serializePath: Where to serialize the fitter after bootstrap
Usage
-----
parameterDct = {
"k1": (1, 5, 10), # name of parameter: low value, initial, high
"k2": (2, 3, 6)}
ftter = ModelFitter(roadrunnerModel, "observed.csv",
parameterDct=parameterDct)
fitter.fitModel() # Do the fit
fitter.bootstrap() # Estimate parameter variance with bootstrap
"""
if modelSpecification is not None:
# Not the default constructor
self._loggerPrefix = _loggerPrefix
self.modelSpecification = modelSpecification
self.parametersToFit = parametersToFit
self.lowerBound = parameterLowerBound
self.upperBound = parameterUpperBound
self.bootstrapKwargs = dict(
numIteration=numIteration,
reportInterval=reportInterval,
maxProcess=maxProcess,
serializePath=serializePath,
)
self.parameterDct = ModelFitterCore._updateParameterDct(
parameterDct)
self._numFitRepeat = numFitRepeat
if self.parametersToFit is None:
self.parametersToFit = list(self.parameterDct.keys())
self.observedTS = observedData
if self.observedTS is not None:
self.observedTS = mkNamedTimeseries(observedData)
#
self.fittedDataTransformDct = fittedDataTransformDct
#
if (selectedColumns is None) and (self.observedTS is not None):
selectedColumns = self.observedTS.colnames
self.selectedColumns = selectedColumns
if self.observedTS is not None:
self._observedArr = self.observedTS[
self.selectedColumns].flatten()
else:
self._observedArr = None
# Other internal state
self._fitterMethods = self._makeMethods(fitterMethods,
METHOD_FITTER_DEFAULTS)
self._bootstrapMethods = self._makeMethods(
bootstrapMethods, METHOD_BOOTSTRAP_DEFAULTS)
if isinstance(self._bootstrapMethods, str):
self._bootstrapMethods = [self._bootstrapMethods]
self._isPlot = isPlot
self._plotter = tp.TimeseriesPlotter(isPlot=self._isPlot)
self._plotFittedTS = None # Timeseries that is plotted
self.logger = logger
# The following are calculated during fitting
self.roadrunnerModel = None
self.minimizer = None # lmfit.minimizer
self.minimizerResult = None # Results of minimization
self.params = None # params property in lmfit.minimizer
self.fittedTS = self.observedTS.copy(
isInitialize=True) # Initialize
self.residualsTS = None # Residuals for selectedColumns
self.bootstrapResult = None # Result from bootstrapping
# Validation checks
self._validateFittedDataTransformDct()
self._bestParameters = _BestParameters(rssq=None, params=None)
else:
pass
def _makeMethods(self, methods, default):
"""
Creates a method dictionary.
Parameters
----------
methods: str/list-str/dict
method used for minimization in fitModel
dict: key-method, value-optional parameters
Returns
-------
list-OptimizerMethod
key: method name
value: dict of optional parameters
"""
if methods is None:
methods = default
if isinstance(methods, str):
if methods == METHOD_BOTH:
methods = METHOD_FITTER_DEFAULTS
else:
methods = [methods]
if isinstance(methods, list):
if isinstance(methods[0], str):
results = [
ModelFitterCore.OptimizerMethod(method=m, kwargs={})
for m in methods
]
else:
results = methods
else:
raise RuntimeError("Must be a list")
trues = [
isinstance(m, ModelFitterCore.OptimizerMethod) for m in results
]
if not all(trues):
raise ValueError("Invalid methods: %s" % str(methods))
return results
@classmethod
def mkParameters(
cls,
parameterDct: dict = None,
parametersToFit: list = None,
logger: Logger = Logger(),
lowerBound: float = PARAMETER_LOWER_BOUND,
upperBound: float = PARAMETER_UPPER_BOUND) -> lmfit.Parameters:
"""
Constructs lmfit parameters based on specifications.
Parameters
----------
parameterDct: key=name, value=ParameterSpecification
parametersToFit: list of parameters to fit
logger: error logger
lowerBound: lower value of range for parameters
upperBound: upper value of range for parameters
Returns
-------
lmfit.Parameters
"""
def get(value, base_value, multiplier):
if value is not None:
return value
return base_value * multiplier
#
if (parametersToFit is None) and (parameterDct is None):
raise RuntimeError("Must specify one of these parameters.")
if parameterDct is None:
parameterDct = {}
if parametersToFit is None:
parametersToFit = parameterDct.keys()
if logger is None:
logger = logger()
params = lmfit.Parameters()
for parameterName in parametersToFit:
if parameterName in parameterDct.keys():
specification = parameterDct[parameterName]
value = get(specification.value, specification.value, 1.0)
if value > 0:
lower_factor = LOWER_PARAMETER_MULT
upper_factor = UPPER_PARAMETER_MULT
else:
upper_factor = UPPER_PARAMETER_MULT
lower_factor = LOWER_PARAMETER_MULT
lower = get(specification.lower, specification.value,
lower_factor)
upper = get(specification.upper, specification.value,
upper_factor)
if np.isclose(lower - upper, 0):
upper = 0.0001
try:
params.add(parameterName,
value=value,
min=lower,
max=upper)
except Exception as err:
msg = "modelFitterCore/mkParameters parameterName %s" \
% parameterName
logger.error(msg, err)
else:
value = np.mean([lowerBound, upperBound])
params.add(parameterName,
value=value,
min=lowerBound,
max=upperBound)
return params
@classmethod
def initializeRoadrunnerModel(cls, modelSpecification):
"""
Sets self.roadrunnerModel.
Parameters
----------
modelSpecification: ExtendedRoadRunner/str
Returns
-------
ExtendedRoadRunner
"""
if isinstance(modelSpecification,
te.roadrunner.extended_roadrunner.ExtendedRoadRunner):
roadrunnerModel = modelSpecification
elif isinstance(modelSpecification, str):
roadrunnerModel = te.loada(modelSpecification)
else:
msg = 'Invalid model.'
msg = msg + "\nA model must either be a Roadrunner model "
msg = msg + "an Antimony model."
raise ValueError(msg)
return roadrunnerModel
@classmethod
def setupModel(cls, roadrunner, parameters, logger=Logger()):
"""
Sets up the model for use based on the parameter parameters
Parameters
----------
roadrunner: ExtendedRoadRunner
parameters: lmfit.Parameters
logger Logger
"""
pp = parameters.valuesdict()
for parameter in pp.keys():
try:
roadrunner.model[parameter] = pp[parameter]
except Exception as err:
msg = "_modelFitterCore.setupModel: Could not set value for %s" \
% parameter
logger.error(msg, err)
@classmethod
def runSimulation(
cls,
parameters=None,
roadrunner=None,
startTime=0,
endTime=5,
numPoint=30,
selectedColumns=None,
returnDataFrame=True,
_logger=Logger(),
_loggerPrefix="",
):
"""
Runs a simulation. Defaults to parameter values in the simulation.
Parameters
----------
roadrunner: ExtendedRoadRunner/str
Roadrunner model
parameters: lmfit.Parameters
lmfit parameters
startTime: float
start time for the simulation
endTime: float
end time for the simulation
numPoint: int
number of points in the simulation
selectedColumns: list-str
output columns in simulation
returnDataFrame: bool
return a DataFrame
_logger: Logger
_loggerPrefix: str
Return
------
NamedTimeseries (or None if fail to converge)
"""
if isinstance(roadrunner, str):
roadrunner = cls.initializeRoadrunnerModel(roadrunner)
else:
roadrunner.reset()
if parameters is not None:
# Parameters have been specified
cls.setupModel(roadrunner, parameters, logger=_logger)
# Do the simulation
if selectedColumns is not None:
newSelectedColumns = list(selectedColumns)
if TIME not in newSelectedColumns:
newSelectedColumns.insert(0, TIME)
try:
data = roadrunner.simulate(startTime, endTime, numPoint,
newSelectedColumns)
except Exception as err:
_logger.error("Roadrunner exception: ", err)
data = None
else:
try:
data = roadrunner.simulate(startTime, endTime, numPoint)
except Exception as err:
_logger.exception("Roadrunner exception: %s", err)
data = None
if data is None:
return data
fittedTS = NamedTimeseries(namedArray=data)
if returnDataFrame:
result = fittedTS.to_dataframe()
else:
result = fittedTS
return result
@classmethod
def rpConstruct(cls):
"""
Overrides rpickler.rpConstruct to create a method that
constructs an instance without arguments.
Returns
-------
Instance of cls
"""
return cls(None, None, None)
def rpRevise(self):
"""
Overrides rpickler.
"""
if "logger" not in self.__dict__.keys():
self.logger = Logger()
def _validateFittedDataTransformDct(self):
if self.fittedDataTransformDct is not None:
keySet = set(self.fittedDataTransformDct.keys())
selectedColumnsSet = self.selectedColumns
if (keySet is not None) and (selectedColumnsSet is not None):
excess = set(keySet).difference(selectedColumnsSet)
if len(excess) > 0:
msg = "Columns not in selectedColumns: %s" % str(excess)
raise ValueError(msg)
def _transformFittedTS(self, data):
"""
Updates the fittedTS taking into account required transformations.
Parameters
----------
data: np.ndarray
Results
----------
NamedTimeseries
"""
colnames = list(self.selectedColumns)
colnames.insert(0, TIME)
fittedTS = NamedTimeseries(array=data[:, :], colnames=colnames)
if self.fittedDataTransformDct is not None:
for column, func in self.fittedDataTransformDct.items():
if func is not None:
fittedTS[column] = func(fittedTS)
return fittedTS
@staticmethod
def _updateParameterDct(parameterDct):
"""
Handles values that are tuples instead of ParameterSpecification.
"""
if parameterDct is None:
parameterDct = {}
dct = dict(parameterDct)
for name, value in parameterDct.items():
if isinstance(value, tuple):
dct[name] = ParameterSpecification(lower=value[0],
upper=value[1],
value=value[2])
return dct
@staticmethod
def addParameter(parameterDct: dict, name: str, lower: float, upper: float,
value: float):
"""
Adds a parameter to a list of parameters.
Parameters
----------
parameterDct: parameter dictionary to agument
name: parameter name
lower: lower range of parameter value
upper: upper range of parameter value
value: initial value
Returns
-------
dict
"""
parameterDct[name] = ParameterSpecification(lower=lower,
upper=upper,
value=value)
def _adjustNames(self, antimonyModel:str, observedTS:NamedTimeseries) \
->typing.Tuple[NamedTimeseries, list]:
"""
Antimony exports can change the names of floating species
by adding a "_" at the end. Check for this and adjust
the names in observedTS.
Return
------
NamedTimeseries: newObservedTS
list: newSelectedColumns
"""
rr = te.loada(antimonyModel)
dataNames = rr.simulate().colnames
names = ["[%s]" % n for n in observedTS.colnames]
missingNames = [n[1:-1] for n in set(names).difference(dataNames)]
newSelectedColumns = list(self.selectedColumns)
if len(missingNames) > 0:
newObservedTS = observedTS.copy()
self.logger.exception("Missing names in antimony export: %s" %
str(missingNames))
for name in observedTS.colnames:
missingName = "%s_" % name
if name in missingNames:
newObservedTS = newObservedTS.rename(name, missingName)
newSelectedColumns.remove(name)
newSelectedColumns.append(missingName)
else:
newObservedTS = observedTS
return newObservedTS, newSelectedColumns
def copy(self, isKeepLogger=False):
"""
Creates a copy of the model fitter.
Preserves the user-specified settings and the results
of bootstrapping.
"""
if not isinstance(self.modelSpecification, str):
try:
modelSpecification = self.modelSpecification.getAntimony()
except Exception as err:
self.logger.error(
"Problem wth conversion to Antimony. Details:", err)
raise ValueError("Cannot proceed.")
observedTS, selectedColumns = self._adjustNames(
modelSpecification, self.observedTS)
else:
modelSpecification = self.modelSpecification
observedTS = self.observedTS.copy()
selectedColumns = self.selectedColumns
#
if isKeepLogger:
logger = self.logger
elif self.logger is not None:
logger = self.logger.copy()
else:
logger = None
newModelFitter = self.__class__(
copy.deepcopy(modelSpecification),
observedTS,
copy.deepcopy(self.parametersToFit),
selectedColumns=selectedColumns,
fitterMethods=self._fitterMethods,
bootstrapMethods=self._bootstrapMethods,
parameterLowerBound=self.lowerBound,
parameterUpperBound=self.upperBound,
parameterDct=copy.deepcopy(self.parameterDct),
fittedDataTransformDct=copy.deepcopy(self.fittedDataTransformDct),
logger=logger,
isPlot=self._isPlot)
if self.bootstrapResult is not None:
newModelFitter.bootstrapResult = self.bootstrapResult.copy()
newModelFitter.params = newModelFitter.bootstrapResult.params
else:
newModelFitter.bootstrapResult = None
newModelFitter.params = self.params
return newModelFitter
def initializeRoadRunnerModel(self):
"""
Sets self.roadrunnerModel.
"""
self.roadrunnerModel = ModelFitterCore.initializeRoadrunnerModel(
self.modelSpecification)
def getDefaultParameterValues(self):
"""
Obtain the original values of parameters.
Returns
-------
dict:
key: parameter name
value: value of parameter
"""
dct = {}
self.initializeRoadRunnerModel()
self.roadrunnerModel.reset()
for parameterName in self.parametersToFit:
dct[parameterName] = self.roadrunnerModel.model[parameterName]
return dct
def simulate(self,
params=None,
startTime=None,
endTime=None,
numPoint=None):
"""
Runs a simulation. Defaults to parameter values in the simulation.
Parameters
----------
params: lmfit.Parameters
startTime: float
endTime: float
numPoint: int
Return
------
NamedTimeseries
"""
def setValue(default, parameter):
# Sets to default if parameter unspecified
if parameter is None:
return default
return parameter
#
startTime = setValue(self.observedTS.start, startTime)
endTime = setValue(self.observedTS.end, endTime)
numPoint = setValue(len(self.observedTS), numPoint)
#
if self.roadrunnerModel is None:
self.initializeRoadRunnerModel()
#
return ModelFitterCore.runSimulation(
parameters=params,
roadrunner=self.roadrunnerModel,
startTime=startTime,
endTime=endTime,
numPoint=numPoint,
selectedColumns=self.selectedColumns,
_logger=self.logger,
_loggerPrefix=self._loggerPrefix,
returnDataFrame=False)
def updateFittedAndResiduals(self, **kwargs) -> np.ndarray:
"""
Updates values of self.fittedTS and self.residualsTS
based on self.params.
Parameters
----------
kwargs: dict
arguments for simulation
Instance Variables Updated
--------------------------
self.fittedTS
self.residualsTS
Returns
-------
1-d ndarray of residuals
"""
self.fittedTS = self.simulate(**kwargs) # Updates self.fittedTS
residualsArr = self._residuals(self.params)
numRow = len(self.fittedTS)
numCol = len(residualsArr) // numRow
residualsArr = np.reshape(residualsArr, (numRow, numCol))
cols = self.selectedColumns
if self.residualsTS is None:
self.residualsTS = self.observedTS.subsetColumns(cols)
self.residualsTS[cols] = residualsArr
def _residuals(self, params) -> np.ndarray:
"""
Compute the residuals between objective and experimental data
Handle nan values in observedTS. This internal-only method
is implemented to maximize efficieency.
Parameters
----------
kwargs: dict
arguments for simulation
Returns
-------
1-d ndarray of residuals
"""
data = ModelFitterCore.runSimulation(
parameters=params,
roadrunner=self.roadrunnerModel,
startTime=self.observedTS.start,
endTime=self.observedTS.end,
numPoint=len(self.observedTS),
selectedColumns=self.selectedColumns,
_logger=self.logger,
_loggerPrefix=self._loggerPrefix,
returnDataFrame=False)
if data is None:
residualsArr = np.repeat(LARGE_RESIDUAL, len(self._observedArr))
else:
residualsArr = self._observedArr - data.flatten()
residualsArr = np.nan_to_num(residualsArr)
rssq = sum(residualsArr**2)
if (self._bestParameters.rssq is None) \
or (rssq < self._bestParameters.rssq):
self._bestParameters = _BestParameters(params=params.copy(),
rssq=rssq)
return residualsArr
def fitModel(self, params: lmfit.Parameters = None, max_nfev=100):
"""
Fits the model by adjusting values of parameters based on
differences between simulated and provided values of
floating species.
Parameters
----------
params: starting values of parameters
Example
-------
f.fitModel()
"""
ParameterDescriptor = collections.namedtuple(
"ParameterDescriptor",
"params method rssq kwargs minimizer minimizerResult")
MAX_NFEV = "max_nfev"
block = Logger.join(self._loggerPrefix, "fitModel")
guid = self.logger.startBlock(block)
self.initializeRoadRunnerModel()
self.params = None
if self.parametersToFit is not None:
if params is None:
params = self.mkParams()
# Fit the model to the data using one or more methods.
# Choose the result with the lowest residual standard deviation
paramResults = []
lastExcp = None
for idx, optimizerMethod in enumerate(self._fitterMethods):
method = optimizerMethod.method
kwargs = optimizerMethod.kwargs
if MAX_NFEV not in kwargs:
kwargs[MAX_NFEV] = max_nfev
for _ in range(self._numFitRepeat):
self._bestParameters = _BestParameters(params=None,
rssq=None)
minimizer = lmfit.Minimizer(self._residuals, params)
try:
minimizerResult = minimizer.minimize(method=method,
**kwargs)
except Exception as excp:
lastExcp = excp
msg = "Error minimizing for method: %s" % method
self.logger.error(msg, excp)
continue
params = self._bestParameters.params.copy()
rssq = np.sum(self._residuals(params)**2)
if len(paramResults) > idx:
if rssq >= paramResults[idx].rssq:
continue
parameterDescriptor = ParameterDescriptor(
params=params,
method=method,
rssq=rssq,
kwargs=dict(kwargs),
minimizer=minimizer,
minimizerResult=minimizerResult,
)
paramResults.append(parameterDescriptor)
if len(paramResults) == 0:
msg = "*** Minimizer failed for this model and data."
self.logger.error(msg, lastExcp)
else:
# Select the result that has the smallest residuals
sortedMethods = sorted(paramResults, key=lambda r: r.rssq)
bestMethod = sortedMethods[0]
self.params = bestMethod.params
self.minimizer = bestMethod.minimizer
self.minimizerResult = bestMethod.minimizerResult
# Ensure that residualsTS and fittedTS match the parameters
self.updateFittedAndResiduals(params=self.params)
self.logger.endBlock(guid)
def getFittedModel(self):
"""
Provides the roadrunner model with fitted parameters
Returns
-------
ExtendedRoadrunner
"""
self._checkFit()
self.roadrunnerModel.reset()
self._setupModel(self.params)
return self.roadrunnerModel
def _setupModel(self, parameters):
"""
Sets up the model for use based on the parameter parameters
Parameters
----------
parameters: lmfit.Parameters
"""
ModelFitterCore.setupModel(self.roadrunnerModel,
parameters,
logger=self.logger)
def mkParams(self, parameterDct: dict = None) -> lmfit.Parameters:
"""
Constructs lmfit parameters based on specifications.
Parameters
----------
parameterDct: key=name, value=ParameterSpecification
Returns
-------
lmfit.Parameters
"""
if parameterDct is None:
parameterDct = self.parameterDct
return ModelFitterCore.mkParameters(
parameterDct,
parametersToFit=self.parametersToFit,
logger=self.logger,
lowerBound=self.lowerBound,
upperBound=self.upperBound)
def _checkFit(self):
if self.params is None:
raise ValueError("Must use fitModel before using this method.")
def serialize(self, path):
"""
Serialize the model to a path.
Parameters
----------
path: str
File path
"""
newModelFitter = self.copy()
with open(path, "wb") as fd:
rpickle.dump(newModelFitter, fd)
@classmethod
def deserialize(cls, path):
"""
Deserialize the model from a path.
Parameters
----------
path: str
File path
Return
------
ModelFitter
Model is initialized.
"""
with open(path, "rb") as fd:
fitter = rpickle.load(fd)
fitter.initializeRoadRunnerModel()
return fitter
def mkParameters(
cls,
parameterDct: dict = None,
parametersToFit: list = None,
logger: Logger = Logger(),
lowerBound: float = PARAMETER_LOWER_BOUND,
upperBound: float = PARAMETER_UPPER_BOUND) -> lmfit.Parameters:
"""
Constructs lmfit parameters based on specifications.
Parameters
----------
parameterDct: key=name, value=ParameterSpecification
parametersToFit: list of parameters to fit
logger: error logger
lowerBound: lower value of range for parameters
upperBound: upper value of range for parameters
Returns
-------
lmfit.Parameters
"""
def get(value, base_value, multiplier):
if value is not None:
return value
return base_value * multiplier
#
if (parametersToFit is None) and (parameterDct is None):
raise RuntimeError("Must specify one of these parameters.")
if parameterDct is None:
parameterDct = {}
if parametersToFit is None:
parametersToFit = parameterDct.keys()
if logger is None:
logger = logger()
params = lmfit.Parameters()
for parameterName in parametersToFit:
if parameterName in parameterDct.keys():
specification = parameterDct[parameterName]
value = get(specification.value, specification.value, 1.0)
if value > 0:
lower_factor = LOWER_PARAMETER_MULT
upper_factor = UPPER_PARAMETER_MULT
else:
upper_factor = UPPER_PARAMETER_MULT
lower_factor = LOWER_PARAMETER_MULT
lower = get(specification.lower, specification.value,
lower_factor)
upper = get(specification.upper, specification.value,
upper_factor)
if np.isclose(lower - upper, 0):
upper = 0.0001
try:
params.add(parameterName,
value=value,
min=lower,
max=upper)
except Exception as err:
msg = "modelFitterCore/mkParameters parameterName %s" \
% parameterName
logger.error(msg, err)
else:
value = np.mean([lowerBound, upperBound])
params.add(parameterName,
value=value,
min=lowerBound,
max=upperBound)
return params
class Optimizer():
"""
Implements an interface to optimizers with abstractions
for multiple methods and performance reporting.
The class also handles an oddity with lmfit that the final parameters
returned may not be the best.
Usage
-----
optimizer = Optimizer(calcResiduals, params, [cn.METHOD_LEASTSQ])
optimizer.execute()
"""
def __init__(self,
function,
initialParams,
methods,
logger=None,
isCollect=False):
"""
Parameters
----------
function: Funtion
Arguments
lmfit.parameters
isInitialze (bool). True on first call the
isGetBest (bool). True to retrieve best parameters
returns residuals (if bool arguments are false)
initialParams: lmfit.parameters
methods: list-_helpers.OptimizerMethod
isCollect: bool
Collects performance statistcs
"""
self._function = function
self._methods = methods
self._initialParams = initialParams
self._isCollect = isCollect
self.logger = logger
if self.logger is None:
self.logger = Logger()
# Outputs
self.performanceStats = [] # list of performance results
self.qualityStats = [] # relative rssq
self.params = None
self.minimizerResult = None
self.rssq = None
def copyResults(self):
"""
Copies of the results of the optimization.
Returns
-------
Optimizer
"""
newOptimizer = Optimizer(self._function,
self._initialParams.copy(),
self._methods,
logger=self.logger,
isCollect=self._isCollect)
newOptimizer._function = None # Not serializable
#
newOptimizer.performanceStats = copy.deepcopy(self.performanceStats)
newOptimizer.qualityStats = copy.deepcopy(self.qualityStats)
newOptimizer.minimizerResult = copy.deepcopy(self.minimizerResult)
newOptimizer.params = None
if self.params is not None:
newOptimizer.params = self.params.copy()
newOptimizer.rssq = self.rssq
return newOptimizer
@staticmethod
def _setRandomValue(params):
"""
Sets value to a uniformly distributed random number between min and max.
Parameters
----------
params: lmfit.Parameters
Returns
-------
lmfit.Parameters
"""
newParameters = lmfit.Parameters()
for name, parameter in params.items():
newValue = np.random.uniform(parameter.min, parameter.max)
newParameters.add(name,
min=parameter.min,
max=parameter.max,
value=newValue)
return newParameters
def execute(self):
"""
Performs the optimization on the function.
Result is self.params
"""
lastExcp = None
self.params = self._initialParams.copy()
minimizer = None
for optimizerMethod in self._methods:
method = optimizerMethod.method
kwargs = optimizerMethod.kwargs
wrapperFunction = _FunctionWrapper(self._function,
isCollect=self._isCollect)
minimizer = lmfit.Minimizer(wrapperFunction.execute, self.params)
try:
self.minimizerResult = minimizer.minimize(method=method,
**kwargs)
except Exception as excp:
lastExcp = excp
msg = "Error minimizing for method: %s" % method
self.logger.error(msg, excp)
continue
# Update the parameters
if wrapperFunction.bestParamDct is not None:
_helpers.updateParameterValues(self.params,
wrapperFunction.bestParamDct)
# Update other statistics
self.rssq = wrapperFunction.rssq
self.performanceStats.append(list(wrapperFunction.perfStatistics))
self.qualityStats.append(list(wrapperFunction.rssqStatistics))
if minimizer is None:
msg = "*** Optimization failed."
self.logger.error(msg, lastExcp)
def report(self):
"""
Reports the result of an optimization.
Returns
-------
str
"""
VARIABLE_STG = "[[Variables]]"
CORRELATION_STG = "[[Correlations]]"
if self.minimizerResult is None:
raise ValueError("Must do fitModel before reportFit.")
valuesDct = self.params.valuesdict()
valuesStg = _helpers.ppDict(dict(valuesDct), indent=4)
reportSplit = str(lmfit.fit_report(self.minimizerResult)).split("\n")
# Eliminate Variables section
inVariableSection = False
trimmedReportSplit = []
for line in reportSplit:
if VARIABLE_STG in line:
inVariableSection = True
if CORRELATION_STG in line:
inVariableSection = False
if inVariableSection:
continue
trimmedReportSplit.append(line)
# Construct the report
newReportSplit = [VARIABLE_STG]
newReportSplit.extend(valuesStg.split("\n"))
newReportSplit.extend(trimmedReportSplit)
return "\n".join(newReportSplit)
def plotPerformance(self, isPlot=True):
"""
Plots the statistics for running the objective function.
"""
if not self._isCollect:
msg = "Must construct with isCollect = True "
msg += "to get performance plot."
raise ValueError(msg)
# Compute statistics
TOT = "Tot"
CNT = "Cnt"
AVG = "Avg"
IDX = "Idx"
totalTimes = [sum(v) for v in self.performanceStats]
counts = [len(v) for v in self.performanceStats]
averages = [np.mean(v) for v in self.performanceStats]
df = pd.DataFrame({
IDX: range(len(self.performanceStats)),
TOT: totalTimes,
CNT: counts,
AVG: averages,
})
#
_, axes = plt.subplots(1, 3)
df.plot.bar(x=IDX,
y=TOT,
ax=axes[0],
title="Total time",
xlabel="method")
df.plot.bar(x=IDX,
y=AVG,
ax=axes[1],
title="Average time",
xlabel="method")
df.plot.bar(x=IDX,
y=CNT,
ax=axes[2],
title="Number calls",
xlabel="method")
if isPlot:
plt.show()
def plotQuality(self, isPlot=True):
"""
Plots the quality results
"""
if not self._isCollect:
msg = "Must construct with isCollect = True "
msg += "to get quality plots."
raise ValueError(msg)
ITERATION = "iteration"
_, axes = plt.subplots(len(self._methods))
minLength = min([len(v) for v in self.qualityStats])
# Compute statistics
dct = {
self._methods[i].method: self.qualityStats[i][:minLength]
for i in range(len(self._methods))
}
df = pd.DataFrame(dct)
df[ITERATION] = range(minLength)
#
for idx, method in enumerate(self._methods):
if "AxesSubplot" in str(type(axes)):
ax = axes
else:
ax = axes[idx]
df.plot.line(x=ITERATION, y=method.method, ax=ax, xlabel="")
ax.set_ylabel("SSQ")
if idx == len(self._methods) - 1:
ax.set_xlabel(ITERATION)
if isPlot:
plt.show()
@staticmethod
def mkOptimizerMethod(methodNames=None,
methodKwargs=None,
maxFev=cn.MAX_NFEV_DFT):
"""
Constructs an OptimizerMethod
Parameters
----------
methodNames: list-str/str
methodKwargs: list-dict/dict
Returns
-------
list-OptimizerMethod
"""
if methodNames is None:
methodNames = [cn.METHOD_LEASTSQ]
if isinstance(methodNames, str):
methodNames = [methodNames]
if methodKwargs is None:
methodKwargs = {}
# Ensure that there is a limit of function evaluations
newMethodKwargs = dict(methodKwargs)
if cn.MAX_NFEV not in newMethodKwargs.keys():
newMethodKwargs[cn.MAX_NFEV] = maxFev
elif maxFev is None:
del newMethodKwargs[cn.MAX_NFEV]
methodKwargs = np.repeat(newMethodKwargs, len(methodNames))
#
result = [_helpers.OptimizerMethod(n, k) for n, k \
in zip(methodNames, methodKwargs)]
return result
@classmethod
def optimize(cls,
function,
initialParams,
methods,
numRestart=0,
**kwargs):
"""
Parameters
----------
function: Funtion
Arguments
lmfit.parameters
isInitialze (bool). True on first call the
isGetBest (bool). True to retrieve best parameters
returns residuals (if bool arguments are false)
initialParams: lmfit.parameters
methods: list-_helpers.OptimizerMethod
numRestart: int
Number of restarts with randomly chosen initial values
Returns
-------
Optimizer
"""
bestOptimizer = cls(function, initialParams, methods, **kwargs)
bestOptimizer.execute()
#
for _ in range(numRestart):
newInitialParams = Optimizer._setRandomValue(initialParams)
newOptimizer = cls(function, newInitialParams, methods, **kwargs)
newOptimizer.execute()
if newOptimizer.rssq < bestOptimizer.rssq:
bestOptimizer = newOptimizer
return bestOptimizer