def _set_results(self, phi, value, error):
fit_parameters = np.zeros(self.model.n_pars)
fit_errors = np.zeros(self.model.n_pars)
for ipar in range(self.model.n_pars):
fit_parameters[ipar] = np.average(self.samples[self.burn_in:,ipar])
fit_errors[ipar] = np.std(self.samples[self.burn_in:,ipar])
loss_value = loss.chi2(value,
self.model.update_and_evaluate(phi, fit_parameters),
error)
pred = self.model.update_and_evaluate(phi, fit_parameters)
ndf = len(phi)
quality = 1-chi2pdf.cdf(loss.chi2(value, pred, error), ndf)
return FitResult(fit_parameters=fit_parameters,
fit_errors=fit_errors,
loss=loss_value,
quality=quality,
ndf=ndf,
samples=self.samples[self.burn_in:,:])
def fit(self, phi, value, error, x0=None):
func = lambda p: self.loss_function(value, self.model.update_and_evaluate(phi, p), error) + self.penalty * np.sum(p**2)
if x0 is not None:
self.model.pars = x0
else:
if self.bounds is None:
self.model.initialize_parameters()
else:
self.model.pars = utils.random_search(func,
self.bounds,
1000)
bad_fit = True
while bad_fit:
if self.bounds is not None:
result = minimize(func, x0=self.model.pars, bounds=self.bounds)
identity = np.identity(self.model.n_pars)
err = np.sqrt(np.abs(np.array(np.matrix(result.hess_inv * identity).diagonal())))
err = err[0]
else:
result = minimize(func, x0=self.model.pars)
err = np.sqrt(result.hess_inv.diagonal())
bad_fit = not result.success
fit_parameters = result.x
fit_errors = err
loss_value = self.loss_function(value,
self.model.update_and_evaluate(phi, fit_parameters),
error)
pred = self.model.update_and_evaluate(phi, fit_parameters)
ndf = len(phi)
quality = 1-chi2pdf.cdf(loss.chi2(value, pred, error), ndf)
return FitResult(fit_parameters=fit_parameters,
fit_errors=fit_errors,
loss=loss_value,
quality=quality,
ndf=ndf)
def _set_result(self, result, phi, value, error):
z = result[0].mean
quality = result.Q
fit_parameters = np.array([result[1].mean/z,
result[2].mean/z,
result[3].mean/z])
fit_errors = np.array([result[4].mean/z - fit_parameters[0]**2,
result[5].mean/z - fit_parameters[1]**2,
result[6].mean/z - fit_parameters[2]**2])
fit_errors = np.sqrt(fit_errors)
loss_value = loss.chi2(value,
self.model.update_and_evaluate(phi, fit_parameters),
error)
return FitResult(fit_parameters=fit_parameters,
fit_errors=fit_errors,
loss=loss_value,
quality=quality,
z=z)
def _set_results(self, phi, value, error):
self.fit_container = np.array(self.fit_container, dtype=np.float32)
fit_parameters = np.zeros(self.model.n_pars)
fit_errors = np.zeros(self.model.n_pars)
for ipar in range(self.model.n_pars):
fit_parameters[ipar] = np.average(self.fit_container[:,ipar])
fit_errors[ipar] = np.std(self.fit_container[:,ipar])
loss_value = self.loss_function(value,
self.model.update_and_evaluate(phi, fit_parameters),
error)
pred = self.model.update_and_evaluate(phi, fit_parameters)
ndf = len(phi)
quality = 1-chi2pdf.cdf(loss.chi2(value, pred, error), ndf)
return FitResult(fit_parameters=fit_parameters,
fit_errors=fit_errors,
loss=loss_value,
quality=quality,
ndf=ndf,
replicas=self.fit_container)