Source code for astropy.modeling.tabular

# Licensed under a 3-clause BSD style license - see LICENSE.rst

"""
Tabular models.

Tabular models of any dimension can be created using `tabular_model`.
For convenience `Tabular1D` and `Tabular2D` are provided.

Examples
--------
>>> table = np.array([[ 3.,  0.,  0.],
...                  [ 0.,  2.,  0.],
...                  [ 0.,  0.,  0.]])
>>> points = ([1, 2, 3], [1, 2, 3])
>>> t2 = Tabular2D(points, lookup_table=table, bounds_error=False,
...                fill_value=None, method='nearest')

"""
# pylint: disable=invalid-name

import numpy as np

from astropy import units as u

from .core import Model

try:
    from scipy.interpolate import interpn

    has_scipy = True
except ImportError:
    has_scipy = False

__all__ = ["tabular_model", "Tabular1D", "Tabular2D"]

__doctest_requires__ = {"tabular_model": ["scipy"]}


class _Tabular(Model):
    """
    Returns an interpolated lookup table value.

    Parameters
    ----------
    points : tuple of ndarray of float, optional
        The points defining the regular grid in n dimensions.
        ndarray must have shapes (m1, ), ..., (mn, ),
    lookup_table : array-like
        The data on a regular grid in n dimensions.
        Must have shapes (m1, ..., mn, ...)
    method : str, optional
        The method of interpolation to perform. Supported are "linear" and
        "nearest", and "splinef2d". "splinef2d" is only supported for
        2-dimensional data. Default is "linear".
    bounds_error : bool, optional
        If True, when interpolated values are requested outside of the
        domain of the input data, a ValueError is raised.
        If False, then ``fill_value`` is used.
    fill_value : float or `~astropy.units.Quantity`, optional
        If provided, the value to use for points outside of the
        interpolation domain. If None, values outside
        the domain are extrapolated.  Extrapolation is not supported by method
        "splinef2d". If Quantity is given, it will be converted to the unit of
        ``lookup_table``, if applicable.

    Returns
    -------
    value : ndarray
        Interpolated values at input coordinates.

    Raises
    ------
    ImportError
        Scipy is not installed.

    Notes
    -----
    Uses `scipy.interpolate.interpn`.

    """

    linear = False
    fittable = False

    standard_broadcasting = False

    _is_dynamic = True

    _id = 0

    def __init__(
        self,
        points=None,
        lookup_table=None,
        method="linear",
        bounds_error=True,
        fill_value=np.nan,
        **kwargs,
    ):
        n_models = kwargs.get("n_models", 1)
        if n_models > 1:
            raise NotImplementedError("Only n_models=1 is supported.")
        super().__init__(**kwargs)
        self.outputs = ("y",)
        if lookup_table is None:
            raise ValueError("Must provide a lookup table.")

        if not isinstance(lookup_table, u.Quantity):
            lookup_table = np.asarray(lookup_table)

        if self.lookup_table.ndim != lookup_table.ndim:
            raise ValueError(
                "lookup_table should be an array with "
                f"{self.lookup_table.ndim} dimensions."
            )

        if points is None:
            points = tuple(np.arange(x, dtype=float) for x in lookup_table.shape)
        else:
            if lookup_table.ndim == 1 and not isinstance(points, tuple):
                points = (points,)
            npts = len(points)
            if npts != lookup_table.ndim:
                raise ValueError(
                    "Expected grid points in "
                    f"{lookup_table.ndim} directions, got {npts}."
                )
            if (
                npts > 1
                and isinstance(points[0], u.Quantity)
                and len({getattr(p, "unit", None) for p in points}) > 1
            ):
                raise ValueError("points must all have the same unit.")

        if isinstance(fill_value, u.Quantity):
            if not isinstance(lookup_table, u.Quantity):
                raise ValueError(
                    f"fill value is in {fill_value.unit} but expected to be unitless."
                )
            fill_value = fill_value.to(lookup_table.unit).value

        self.points = points
        self.lookup_table = lookup_table
        self.bounds_error = bounds_error
        self.method = method
        self.fill_value = fill_value

    def __repr__(self):
        return (
            f"<{self.__class__.__name__}(points={self.points}, "
            f"lookup_table={self.lookup_table})>"
        )

    def __str__(self):
        default_keywords = [
            ("Model", self.__class__.__name__),
            ("Name", self.name),
            ("N_inputs", self.n_inputs),
            ("N_outputs", self.n_outputs),
            ("Parameters", ""),
            ("  points", self.points),
            ("  lookup_table", self.lookup_table),
            ("  method", self.method),
            ("  fill_value", self.fill_value),
            ("  bounds_error", self.bounds_error),
        ]

        parts = [
            f"{keyword}: {value}"
            for keyword, value in default_keywords
            if value is not None
        ]

        return "\n".join(parts)

    @property
    def input_units(self):
        pts = self.points[0]
        if not isinstance(pts, u.Quantity):
            return None
        return {x: pts.unit for x in self.inputs}

    @property
    def return_units(self):
        if not isinstance(self.lookup_table, u.Quantity):
            return None
        return {self.outputs[0]: self.lookup_table.unit}

    @property
    def bounding_box(self):
        """
        Tuple defining the default ``bounding_box`` limits,
        ``(points_low, points_high)``.

        Examples
        --------
        >>> from astropy.modeling.models import Tabular1D, Tabular2D
        >>> t1 = Tabular1D(points=[1, 2, 3], lookup_table=[10, 20, 30])
        >>> t1.bounding_box
        ModelBoundingBox(
            intervals={
                x: Interval(lower=1, upper=3)
            }
            model=Tabular1D(inputs=('x',))
            order='C'
        )
        >>> t2 = Tabular2D(points=[[1, 2, 3], [2, 3, 4]],
        ...                lookup_table=[[10, 20, 30], [20, 30, 40]])
        >>> t2.bounding_box
        ModelBoundingBox(
            intervals={
                x: Interval(lower=1, upper=3)
                y: Interval(lower=2, upper=4)
            }
            model=Tabular2D(inputs=('x', 'y'))
            order='C'
        )

        """
        bbox = [(min(p), max(p)) for p in self.points][::-1]
        if len(bbox) == 1:
            bbox = bbox[0]
        return bbox

    def evaluate(self, *inputs):
        """
        Return the interpolated values at the input coordinates.

        Parameters
        ----------
        inputs : list of scalar or list of ndarray
            Input coordinates. The number of inputs must be equal
            to the dimensions of the lookup table.
        """
        inputs = np.broadcast_arrays(*inputs)

        shape = inputs[0].shape
        inputs = [inp.flatten() for inp in inputs[: self.n_inputs]]
        inputs = np.array(inputs).T
        if not has_scipy:  # pragma: no cover
            raise ImportError("Tabular model requires scipy.")
        result = interpn(
            self.points,
            self.lookup_table,
            inputs,
            method=self.method,
            bounds_error=self.bounds_error,
            fill_value=self.fill_value,
        )

        # return_units not respected when points has no units
        if isinstance(self.lookup_table, u.Quantity) and not isinstance(
            self.points[0], u.Quantity
        ):
            result = result * self.lookup_table.unit

        if self.n_outputs == 1:
            result = result.reshape(shape)
        else:
            result = [r.reshape(shape) for r in result]
        return result

    @property
    def inverse(self):
        if self.n_inputs == 1:
            # If the wavelength array is descending instead of ascending, both
            # points and lookup_table need to be reversed in the inverse transform
            # for scipy.interpolate to work properly
            if np.all(np.diff(self.lookup_table) > 0):
                # ascending case
                points = self.lookup_table
                lookup_table = self.points[0]
            elif np.all(np.diff(self.lookup_table) < 0):
                # descending case, reverse order
                points = self.lookup_table[::-1]
                lookup_table = self.points[0][::-1]
            else:
                # equal-valued or double-valued lookup_table
                raise NotImplementedError
            return Tabular1D(
                points=points,
                lookup_table=lookup_table,
                method=self.method,
                bounds_error=self.bounds_error,
                fill_value=self.fill_value,
            )
        raise NotImplementedError(
            "An analytical inverse transform has not been implemented for this model."
        )


[docs] def tabular_model(dim, name=None): """ Make a ``Tabular`` model where ``n_inputs`` is based on the dimension of the lookup_table. This model has to be further initialized and when evaluated returns the interpolated values. Parameters ---------- dim : int Dimensions of the lookup table. name : str Name for the class. Examples -------- >>> table = np.array([[3., 0., 0.], ... [0., 2., 0.], ... [0., 0., 0.]]) >>> tab = tabular_model(2, name='Tabular2D') >>> print(tab) <class 'astropy.modeling.tabular.Tabular2D'> Name: Tabular2D N_inputs: 2 N_outputs: 1 >>> points = ([1, 2, 3], [1, 2, 3]) Setting fill_value to None, allows extrapolation. >>> m = tab(points, lookup_table=table, name='my_table', ... bounds_error=False, fill_value=None, method='nearest') >>> xinterp = [0, 1, 1.5, 2.72, 3.14] >>> m(xinterp, xinterp) # doctest: +FLOAT_CMP array([3., 3., 3., 0., 0.]) """ if dim < 1: raise ValueError("Lookup table must have at least one dimension.") table = np.zeros([2] * dim) members = {"lookup_table": table, "n_inputs": dim, "n_outputs": 1} if dim == 1: members["_separable"] = True else: members["_separable"] = False if name is None: model_id = _Tabular._id _Tabular._id += 1 name = f"Tabular{model_id}" model_class = type(str(name), (_Tabular,), members) model_class.__module__ = "astropy.modeling.tabular" return model_class
Tabular1D = tabular_model(1, name="Tabular1D") Tabular2D = tabular_model(2, name="Tabular2D") _tab_docs = """ method : str, optional The method of interpolation to perform. Supported are "linear" and "nearest", and "splinef2d". "splinef2d" is only supported for 2-dimensional data. Default is "linear". bounds_error : bool, optional If True, when interpolated values are requested outside of the domain of the input data, a ValueError is raised. If False, then ``fill_value`` is used. fill_value : float, optional If provided, the value to use for points outside of the interpolation domain. If None, values outside the domain are extrapolated. Extrapolation is not supported by method "splinef2d". Returns ------- value : ndarray Interpolated values at input coordinates. Raises ------ ImportError Scipy is not installed. Notes ----- Uses `scipy.interpolate.interpn`. """ Tabular1D.__doc__ = ( """ Tabular model in 1D. Returns an interpolated lookup table value. Parameters ---------- points : array-like of float of ndim=1. The points defining the regular grid in n dimensions. lookup_table : array-like, of ndim=1. The data in one dimensions. """ + _tab_docs ) Tabular2D.__doc__ = ( """ Tabular model in 2D. Returns an interpolated lookup table value. Parameters ---------- points : tuple of ndarray of float, optional The points defining the regular grid in n dimensions. ndarray with shapes (m1, m2). lookup_table : array-like The data on a regular grid in 2 dimensions. Shape (m1, m2). """ + _tab_docs )