Extending SrFit

The Examples give an overview of how to use SrFit and extend it with various custom-made objects. Many pieces of SrFit that are not covered in the examples are discussed here.

Plugging Other Objects into SrFit

Much of the power of SrFit comes from being able to plug existing python codes into the framework. For example, external forward calculators can be wrapped up inside ProfileGenerators without modifying the calculator. This is demonstrated in Examples. Structure adapters defined in the diffpy.srfit.structure module are also built around this principle. These adapters are hierarchical ParameterSets (found in diffpy.srfit.fitbase.parameterset) that encapsulate the different pieces of a structure. For example, the DiffpyStructureParSet structure adapter in diffpy.srfit.structure.diffpyparset contains DiffpyLatticeParSet, which encapsulates the lattice data and one DiffpyAtomParSet per atom. These each contain parameters for what they encapsulate, such as lattice parameters or atom positions.

Fundamentally, it is the adjustable parameters of a structure container, forward calculator or other object that needs to be adapted so that SrFit can manipulate the underlying data object. These adapted parameters can then be organized into ParameterSets, as in the case of a structure adapter. The ParameterAdapter class found in diffpy.srfit.fitbase.parameter is designed for this purpose. ParameterAdapter is a Parameter that defers to another object when setting or retrieving its value.

class diffpy.srfit.fitbase.parameter.ParameterAdapter(name, obj, getter=None, setter=None, attr=None)

An adapter for parameter-like objects.

This class wraps an object as a Parameter. The getValue and setValue methods defer to the data of the wrapped object.

The name argument is used to give attribute access to the ParameterAdapter instance when it is added to a ParameterSet or similar object. The obj argument is the parameter-like object to be adapted. It must provide some form of access to its data. If it provides a getter and setter, these can be specified with the getter and setter arguments. If the getter and setter require an attribute name, this is specified with the attr argument. If the data can be retrieved as an attribute, then the name of this attribute can be passed in the attr argument.

Here is a simple example of using ParameterAdapter to adapt a hypothetical atom object called SimpleAtom that has attributes x, y and z.

class SimpleAtom(object):
    """Simple class holding x, y and z coordinates of an atom."""

    def __init__(self, x, y, z):
        self.x = x
        self.y = y
        self.z = z
        return

# End class SimpleAtom

class SimpleAtomParSet(ParameterSet):
    """Class adapting the x, y and z attributes of SimpleAtom as Parameters."""

    def __init__(self, atom, name):
        ParameterSet.__init__(self, name)
        # Store the atom, we might need it later
        self.atom = atom

        # Create a ParameterAdapter for the x, y and z attributes of atom
        xpar = ParameterAdapter("x", atom, attr = "x")
        ypar = ParameterAdapter("y", atom, attr = "y")
        zpar = ParameterAdapter("z", atom, attr = "z")

        # Add these to the parameter set
        self.addParameter(xpar)
        self.addParameter(ypar)
        self.addParameter(zpar)

        return

# End class SimpleAtomParSet

The x, y and z attributes (specified by the attr keyword argument of ParameterAdapter) of a SimpleAtom are wrapped as ParameterAdapter objects named x, y, and z. They are then added to the SimpleAtomParSet using the addParameter method, which makes them accessible as attributes.

If SimpleAtom did not have an attribute named x, but rather accessor methods named getX and setX, then the ParameterAdapter would be used as:

xpar = ParameterAdapter("x", atom, getter = SimpleAtom.getX,
    setter = SimpleAtom.setX)

Note that the unbound methods are used. The names getter and setter describe how the accessor attributes are used to access the value of the parameter. When xpar.getValue() is called, it redirects to SimpleAtom.getX(atom).

If instead SimpleAtom had methods called get and set that take as the second argument the name of the attribute to retrieve or modify, then this can be adapted as:

xpar = ParameterAdapter("x", atom, getter = SimpleAtom.get,
        setter = SimpleAtom.set, attr = "x")

Thus, when xpar.getValue() is called, it in turn calls SimpleAtom.get(atom, "x"). xpar.setValue(value) calls SimpleAtom.set(atom, "x", value).

If the attributes of an object cannot be accessed in one of these three ways, then you must write external accessor methods that can be set as the getter and setter of the ParameterAdapter. For example, if the x, y and z values were held in a list called xyz, then you would have to write the functions getX and setX that would manipulate this list, and use these functions as in the second example.

Extending Profile Parsers

The ProfileParser class is located in the diffpy.srfit.fitbase.parser module. The purpose of this class is to read data and metadata from a file or string and pass those data and metadata to a Profile instance. The Profile in turn will pass this information to a ProfileGenerator.

The simplest way to extend the ProfileParser is to derive a new class from ProfileParser and overload the parseString method. By default, the parseFile method can read an ASCII file and passes the loaded string to the parseString method. For non-ASCII data one should overload both of these methods. An example of a customized ProfileParser is the PDFParser class in the diffpy.srfit.pdf.pdfparser module.

Here is a simple example demonstrating how to extract (x,y) data from a two-column string.

def parseString(self, datastring):

    xvals = []
    yvals = []
    dxvals = None
    dyvals = None

    for line in datastring.splitlines():

        sline = line.split()
        x, y = map(float, sline)
        xvals.append(x)
        yvals.append(y)

    self._banks.append([xvals, yvals, dxvals, dyvals])
    return

The self._banks.append line puts the data arrays into the _banks list. This list is for collecting multiple data sets that may be present within a single file. The dxvals and dyvals are the uncertainty values on the xvals and yvals. In this simple example they are not present, and so are set to None.

In general, the data string may contain metadata. The ProfileParser has a dictionary attribute named _meta. The parser can put any information into this dictionary. It is up to a ProfileGenerator that may use the parsed data to define and retrieve usable metadata.

If the data are not in a form that can be stored in a Profile then it is the responsibility of the parser to convert this data to a usable form.

Extending Profiles

Even with the ability to customize ProfileParsers, it may be necessary to create custom Profile objects for different types of data. This is useful when adapting an external data container to the SrFit interface. For example, the external container may need to be retained so it can be used within an external program before or after interfacing with SrFit. An example of a customized Profile is the SASProfile class in the diffpy.srfit.sas.sasprofile module:

class SASProfile(Profile):
    """Observed and calculated profile container for SAS data.

    This wraps a sas DataInfo object as a Profile object. Use this when you
    want to use and manipulate a DataProfile before using it with SrFit.
    Otherwise, use the SASParser class and load the data into a base Profile
    object.

    Attributes

    _xobs   --  A numpy array of the observed independent variable (default
                None)
    xobs    --  Read-only property of _xobs.
    _yobs   --  A numpy array of the observed signal (default None)
    yobs    --  Read-only property of _yobs.
    _dyobs  --  A numpy array of the uncertainty of the observed signal (default
                None, optional).
    dyobs   --  Read-only property of _dyobs.
    x       --  A numpy array of the calculated independent variable (default
                None, property for xpar accessors).
    y       --  The profile over the calculation range (default None, property
                for ypar accessors).
    dy      --  The uncertainty in the profile over the calculation range
                (default None, property for dypar accessors).
    ycalc   --  A numpy array of the calculated signal (default None).
    xpar    --  A ProfileParameter that stores x (named "x").
    ypar    --  A ProfileParameter that stores y (named "y").
    dypar   --  A ProfileParameter that stores dy (named "dy").
    meta    --  A dictionary of metadata. This is only set if provided by a
                parser.

    _datainfo   --  The DataInfo object this wraps.

    """

    def __init__(self, datainfo):
        """Initialize the attributes.

        datainfo   --  The DataInfo object this wraps.

        """
        self._datainfo = datainfo
        Profile.__init__(self)

        self._xobs = self._datainfo.x
        self._yobs = self._datainfo.y
        if self._datainfo.dy is None or 0 == len(self._datainfo.dy):
            self._dyobs = ones_like(self.xobs)
        else:
            self._dyobs = self._datainfo.dy
        return

    def setObservedProfile(self, xobs, yobs, dyobs = None):
        """Set the observed profile.

        This is overloaded to change the value within the datainfo object.

        Arguments
        xobs    --  Numpy array of the independent variable
        yobs    --  Numpy array of the observed signal.
        dyobs   --  Numpy array of the uncertainty in the observed signal. If
                    dyobs is None (default), it will be set to 1 at each
                    observed xobs.

        Raises ValueError if len(yobs) != len(xobs)
        Raises ValueError if dyobs != None and len(dyobs) != len(xobs)

        """
        Profile.setObservedProfile(self, xobs, yobs, dyobs)
        # Copy the arrays to the _datainfo attribute.
        self._datainfo.x = self._xobs
        self._datainfo.y = self._yobs
        self._datainfo.dy = self._dyobs
        return

The __init__ method sets the xobs, yobs and dyobs attributes of the SASProfile to the associated arrays within the _datainfo attribute. The setObservedProfile method is overloaded to modify the _datainfo arrays when their corresponding attributes are modified. This keeps the arrays in sync.

Custom Restraints

Restraints in SrFit are one way to include known information about a system into a fit recipe. When customizing SrFit for a specific purpose, one may want to create restraints. One example of this is in the SrRealParSet base class in diffpy.srfit.structure.srrealparset. SrReal provides many real-space structure utilities for compatible structures, such as a PDF calculator and a bond-valence sum (BVS) calculator. The PDF calculator works very well as a ProfileGenerator (see the Examples), but the BVS calculator is better suited as a restraint. This makes it very easy to keep the BVS constrained during a PDF fit or some other refinement.

Creating a custom restraint is a two-step process. First, a class must be derived from diffpy.srfit.fitbase.restraint.Restraint that can calculate the restraint cost. This requires the penalty method to be overloaded. This method has the following signature

Restraint.penalty(w=1.0)

Calculate the penalty of the restraint.

w – The point-average chi^2 which is optionally used to scale the
penalty (default 1.0).

Returns the penalty as a float

The w factor is optionally used to scale the restraint cost. Its purpose is to keep the restraint cost comparable to the residual of a single data point.

BVSRestraint from diffpy.srfit.structure.bvsrestraint is a custom Restraint whose penalty is the root-mean-square deviation from the expected and calculated BVS of a structure.

class BVSRestraint(Restraint):
    """Wrapping of BVSCalculator.bvmsdiff as a Restraint.

    The restraint penalty is the root-mean-square deviation of the theoretical
    and calculated bond-valence sum of a structure.

    Attributes:
    _calc   --  The SrReal BVSCalculator instance.
    _parset --  The SrRealParSet that created this BVSRestraint.
    sig     --  The uncertainty on the BVS (default 1).
    scaled  --  A flag indicating if the restraint is scaled (multiplied)
                by the unrestrained point-average chi^2 (chi^2/numpoints)
                (default False).

    """

    def __init__(self, parset, sig = 1, scaled = False):
        """Initialize the Restraint.

        parset  --  SrRealParSet that creates this BVSRestraint.
        sig     --  The uncertainty on the BVS (default 1).
        scaled  --  A flag indicating if the restraint is scaled
                    (multiplied) by the unrestrained point-average chi^2
                    (chi^2/numpoints) (bool, default False).

        """
        from diffpy.srreal.bvscalculator import BVSCalculator
        self._calc = BVSCalculator()
        self._parset = parset
        self.sig = float(sig)
        self.scaled = bool(scaled)
        return

    def penalty(self, w = 1.0):
        """Calculate the penalty of the restraint.

        w   --  The point-average chi^2 which is optionally used to scale the
                penalty (float, default 1.0).

        """
        # Get the bvms from the BVSCalculator
        stru = self._parset._getSrRealStructure()
        self._calc.eval(stru)
        penalty = self._calc.bvmsdiff

        # Scale by the prefactor
        penalty /= self.sig**2

        # Optionally scale by w
        if self.scaled: penalty *= w

        return penalty

    def _validate(self):
        """This evaluates the calculator.

        Raises SrFitError if validation fails.

        """
        from numpy import nan
        p = self.penalty()
        if p is None or p is nan:
            raise SrFitError("Cannot evaluate penalty")
        v = self._calc.value
        if len(v) > 1 and not v.any():
            emsg = ("Bond valence sums are all zero.  Check atom symbols in "
                    "the structure or define custom bond-valence parameters.")
            raise SrFitError(emsg)
        return

Note that the penalty scaling is optional (selected by the scaled flag) and uncertainty on the result (sig) may be applied. These two options are recommended with any custom Restraint.

The second part of a custom restraint is to allow it to be created from a restrainable object. A BVSRestraint is used to restrain a SrRealParSet, which is a ParameterSet wrapper base class for SrReal-compatible structures. The restraint is applied with the restrainBVS method.

    def restrainBVS(self, sig = 1, scaled = False):
        """Restrain the bond-valence sum to zero.

        This adds a penalty to the cost function equal to
        bvmsdiff / sig**2
        where bvmsdiff is the mean-squared difference between the calculated
        and expected bond valence sums for the structure. If scaled is True,
        this is also scaled by the current point-averaged chi^2 value so the
        restraint is roughly equally weighted in the fit.

        sig     --  The uncertainty on the BVS (default 1).
        scaled  --  A flag indicating if the restraint is scaled
                    (multiplied) by the unrestrained point-average chi^2
                    (chi^2/numpoints) (default False).

        Returns the BVSRestraint object for use with the 'unrestrain' method.

        """

        # Create the Restraint object
        res = BVSRestraint(self, sig, scaled)
        # Add it to the _restraints set
        self._restraints.add(res)
        # Our configuration changed. Notify observers.
        self._updateConfiguration()
        # Return the Restraint object
        return res

The purpose of the method is to create the custom Restraint object, configure it and store it. Note that the optional sig and scaled flag are passed as part of this method. Both _restraints and _updateConfiguration come from ParameterSet, from which SrRealParSet is derived. The _restraints attribute is a set of Restraints on the object. The _updateConfiguration method makes any object containing the SrRealParSet aware of the configuration change. This gets propagated to the top-level FitRecipe, if there is one. The restraint object is returned by the method so that it may be later removed.

For more examples of custom restraints can be found in the diffpy.srfit.structure.objcrystparset module.

Custom FitHooks

The FitHook class is used by a FitRecipe to report fit progress to a user. FitHook can be found in the diffpy.srfit.fitbase.fithook module. FitHook can be customized to provide customized fit output, such as a live plot of the output. The FitHook class has three methods that one can overload.

  • FitHook.reset(recipe)

    Reset the hook data.

    This is called whenever FitRecipe._prepare is called, which is whenever a configurational change to the fit hierarchy takes place, such as adding a new ParameterSet, constraint or restraint.

  • FitHook.precall(recipe)

    This is called within FitRecipe.residual, before the calculation.

    recipe – The FitRecipe instance

  • FitHook.postcall(recipe, chiv)

    This is called within FitRecipe.residual, after the calculation.

    recipe – The FitRecipe instance chiv – The residual vector

To use a custom FitHook, assign an instance to a FitRecipe using the pushFitHook method. All FitHook instances held by a FitRecipe will be used in sequence during a call to FitRecipe.residual.