fit.options¶

Bases: StrEnum

Strategy for fitting a set of fit experiments.

ALL : fit all experiments together, i.e., one parameter set describes every experiment.

SINGLE : fit every experiment on its own, i.e., one parameter set per experiment. These are the individual parameters, see sbmlsim.fit.metrics.

Bases: Enum

Type of optimization.

least square : Least square is a local optimization method and works well in combination with many start values, i.e., many repeats of the optimization problem. See https://docs.scipy.org/doc/scipy/reference/generated/scipy.optimize.least_squares.html for more information.

differential evolution : Differential evolution is a global optimization method and normally is run with a limited number of repeats. See https://docs.scipy.org/doc/scipy/reference/generated/scipy.optimize.differential_evolution.html#scipy.optimize.differential_evolution for more information.

Bases: Enum

Handling of the residuals.

How are the residuals calculated? Are the absolute residuals used, or are the residuals normalized based on the data points, i.e., relative residuals.

absolute (default) : uses the absolute data values for calculation of the residuals::

r(i,k) = y(i,k) - f(xi,k)

normalized : normalizing residuals of curve with 1/mean of the timecourse::

r(i,k) = (y(i,k) - f(xi,k))/mean(y(k))

This allows to use time courses with very different absolute values in a single optimization problem.

absolute_to_baseline (experimental) : uses the absolute changes to baseline with baseline being the first data point. This requires appropriate pre-simulations for the model to reach a baseline. Data and fits have to be checked carefully. Residuals are calculated as::

r(i,k) = (y(i,k) - ybase(k)) - (f(xi,k) - fbase(k))

normalized changes baseline (experimental): Uses the normalized changes to baseline with baseline being the first data point. This requires appropriate pre-simulations for the model to reach a baseline. Data and fits have to be checked carefully. The residuals are calculated as::

r(i,k) = (y(i,k) - ybase(k)) - (f(xi,k) - fbase(k))/mean(y(k))

Bases: Enum

Determines the loss function.

minimize F(x) = 0.5 * sum(rho(residuals_weighted(x)**2)

The following loss functions are supported are allowed:

‘linear’ (default) : rho(z) = z. Gives a standard least-squares problem.

‘soft_l1’ : rho(z) = 2 * ((1 + z)**0.5 - 1). The smooth approximation of l1 (absolute value) loss. Usually a good choice for robust least squares.

‘cauchy’ : rho(z) = ln(1 + z). Severely weakens outliers influence, but may cause difficulties in optimization process.

‘arctan’ : rho(z) = arctan(z). Limits a maximum loss on a single residual, has properties similar to ‘cauchy’.

Bases: Enum

Weighting w_{k} of the curves k.

Users can provide set of weightings for the individual curves. By default no weightings are applied, i.e. all curves are weighted equally if no weighting option is provided::

w_{k} = 1.0

mapping : curves k are weighted with the provided user weights in the fit mappings, e.g., counts::

w_{k} = wu_{k}

points : weighting with the number of data points. Often time courses contain different number of data points. The residuals should contribute equally per data point::

w_{k} = 1.0/count{k}

The various options can be combined, e.g. mapping and points results in::

w_{k} = wu_{k}/count{k}/mean{y(k)}

Bases: Enum

Space the optimizer searches the parameters in.

A parameter of a model is a positive quantity which often spans orders of magnitude, e.g. a rate constant, and an optimizer which searches it on the linear scale spends its steps on the largest of them. Searching the logarithm makes the steps relative, which is what such a parameter needs.

The scale is a property of the optimization and not of the model or of the data: PEtab v2 removed the parameterScale of its parameter table for that reason, so it is part of the FitSettings here and is stored with them.

The bounds and the start value of a FitParameter are always on the linear scale, i.e. in the units of the model, and so are the fitted parameters a fit reports; only the search happens in the scaled space.

is_log

to_scale(x)

from_scale(x)

Bases: Enum

Weighting w_{i,k} of the data points i within a single fit mapping k.

This decides how the data points within a single fit mapping are weighted.

no weighting (default) : all data points are weighted equally::

w_{i,k} = 1.0

error_weighting: data points are weighted as ~1/error

# FIXME: update documentation, These must probably be normalized also. if yerr{i,k}: w_{i,k} = 1.0/yerr{i,k} else: w_{i,k} = 1.0/yerr{i,k}

Settings of a parameter fit.

The settings decide how the residuals of an optimization problem are calculated, so the same settings are needed to run a fit and to report it afterwards. They are stored with the result of a fit and read back by the report, see sbmlsim.fit.report.FitReport.

Attributes:

to_dict()

from_dict(d)