_core.py

import warnings
import itertools
from copy import copy
from functools import partial
from collections.abc import Iterable, Sequence, Mapping
from numbers import Number
from datetime import datetime
from distutils.version import LooseVersion

import numpy as np
import pandas as pd
import matplotlib as mpl

from ._decorators import (
    share_init_params_with_map,
)
from .palettes import (
    QUAL_PALETTES,
    color_palette,
    cubehelix_palette,
    _parse_cubehelix_args,
)
from .utils import (
    get_color_cycle,
    remove_na,
)


class SemanticMapping:
    """Base class for mapping data values to plot attributes."""

    # -- Default attributes that all SemanticMapping subclasses must set

    # Whether the mapping is numeric, categorical, or datetime
    map_type = None

    # Ordered list of unique values in the input data
    levels = None

    # A mapping from the data values to corresponding plot attributes
    lookup_table = None

    def __init__(self, plotter):

        # TODO Putting this here so we can continue to use a lot of the
        # logic that's built into the library, but the idea of this class
        # is to move towards semantic mappings that are agnositic about the
        # kind of plot they're going to be used to draw.
        # Fully achieving that is going to take some thinking.
        self.plotter = plotter

    def map(cls, plotter, *args, **kwargs):
        # This method is assigned the __init__ docstring
        method_name = "_{}_map".format(cls.__name__[:-7].lower())
        setattr(plotter, method_name, cls(plotter, *args, **kwargs))
        return plotter

    def _lookup_single(self, key):
        """Apply the mapping to a single data value."""
        return self.lookup_table[key]

    def __call__(self, key, *args, **kwargs):
        """Get the attribute(s) values for the data key."""
        if isinstance(key, (list, np.ndarray, pd.Series)):
            return [self._lookup_single(k, *args, **kwargs) for k in key]
        else:
            return self._lookup_single(key, *args, **kwargs)


@share_init_params_with_map
class HueMapping(SemanticMapping):
    """Mapping that sets artist colors according to data values."""
    # A specification of the colors that should appear in the plot
    palette = None

    # An object that normalizes data values to [0, 1] range for color mapping
    norm = None

    # A continuous colormap object for interpolating in a numeric context
    cmap = None

    def __init__(
        self, plotter, palette=None, order=None, norm=None,
    ):
        """Map the levels of the `hue` variable to distinct colors.

        Parameters
        ----------
        # TODO add generic parameters

        """
        super().__init__(plotter)

        data = plotter.plot_data.get("hue", pd.Series(dtype=float))

        if data.notna().any():

            map_type = self.infer_map_type(
                palette, norm, plotter.input_format, plotter.var_types["hue"]
            )

            # Our goal is to end up with a dictionary mapping every unique
            # value in `data` to a color. We will also keep track of the
            # metadata about this mapping we will need for, e.g., a legend

            # --- Option 1: numeric mapping with a matplotlib colormap

            if map_type == "numeric":

                data = pd.to_numeric(data)
                levels, lookup_table, norm, cmap = self.numeric_mapping(
                    data, palette, norm,
                )

            # --- Option 2: categorical mapping using seaborn palette

            elif map_type == "categorical":

                cmap = norm = None
                levels, lookup_table = self.categorical_mapping(
                    data, palette, order,
                )

            # --- Option 3: datetime mapping

            else:
                # TODO this needs actual implementation
                cmap = norm = None
                levels, lookup_table = self.categorical_mapping(
                    # Casting data to list to handle differences in the way
                    # pandas and numpy represent datetime64 data
                    list(data), palette, order,
                )

            self.map_type = map_type
            self.lookup_table = lookup_table
            self.palette = palette
            self.levels = levels
            self.norm = norm
            self.cmap = cmap

    def _lookup_single(self, key):
        """Get the color for a single value, using colormap to interpolate."""
        try:
            # Use a value that's in the original data vector
            value = self.lookup_table[key]
        except KeyError:
            # Use the colormap to interpolate between existing datapoints
            # (e.g. in the context of making a continuous legend)
            try:
                normed = self.norm(key)
            except TypeError as err:
                if np.isnan(key):
                    value = (0, 0, 0, 0)
                else:
                    raise err
            else:
                if np.ma.is_masked(normed):
                    normed = np.nan
                value = self.cmap(normed)
        return value

    def infer_map_type(self, palette, norm, input_format, var_type):
        """Determine how to implement the mapping."""
        if palette in QUAL_PALETTES:
            map_type = "categorical"
        elif norm is not None:
            map_type = "numeric"
        elif isinstance(palette, (dict, list)):
            map_type = "categorical"
        elif input_format == "wide":
            map_type = "categorical"
        else:
            map_type = var_type

        return map_type

    def categorical_mapping(self, data, palette, order):
        """Determine colors when the hue mapping is categorical."""
        # -- Identify the order and name of the levels

        levels = categorical_order(data, order)
        n_colors = len(levels)

        # -- Identify the set of colors to use

        if isinstance(palette, dict):

            missing = set(levels) - set(palette)
            if any(missing):
                err = "The palette dictionary is missing keys: {}"
                raise ValueError(err.format(missing))

            lookup_table = palette

        else:

            if palette is None:
                if n_colors <= len(get_color_cycle()):
                    colors = color_palette(None, n_colors)
                else:
                    colors = color_palette("husl", n_colors)
            elif isinstance(palette, list):
                if len(palette) != n_colors:
                    err = "The palette list has the wrong number of colors."
                    raise ValueError(err)
                colors = palette
            else:
                colors = color_palette(palette, n_colors)

            lookup_table = dict(zip(levels, colors))

        return levels, lookup_table

    def numeric_mapping(self, data, palette, norm):
        """Determine colors when the hue variable is quantitative."""
        if isinstance(palette, dict):

            # The presence of a norm object overrides a dictionary of hues
            # in specifying a numeric mapping, so we need to process it here.
            levels = list(sorted(palette))
            colors = [palette[k] for k in sorted(palette)]
            cmap = mpl.colors.ListedColormap(colors)
            lookup_table = palette.copy()

        else:

            # The levels are the sorted unique values in the data
            levels = list(np.sort(remove_na(data.unique())))

            # --- Sort out the colormap to use from the palette argument

            # Default numeric palette is our default cubehelix palette
            # TODO do we want to do something complicated to ensure contrast?
            palette = "ch:" if palette is None else palette

            if isinstance(palette, mpl.colors.Colormap):
                cmap = palette
            elif str(palette).startswith("ch:"):
                args, kwargs = _parse_cubehelix_args(palette)
                cmap = cubehelix_palette(0, *args, as_cmap=True, **kwargs)
            else:
                try:
                    cmap = mpl.cm.get_cmap(palette)
                except (ValueError, TypeError):
                    err = f"Palette {palette} not understood"
                    raise ValueError(err)

            # Now sort out the data normalization
            if norm is None:
                norm = mpl.colors.Normalize()
            elif isinstance(norm, tuple):
                norm = mpl.colors.Normalize(*norm)
            elif not isinstance(norm, mpl.colors.Normalize):
                err = "``hue_norm`` must be None, tuple, or Normalize object."
                raise ValueError(err)

            if not norm.scaled():
                norm(np.asarray(data.dropna()))

            lookup_table = dict(zip(levels, cmap(norm(levels))))

        return levels, lookup_table, norm, cmap


@share_init_params_with_map
class SizeMapping(SemanticMapping):
    """Mapping that sets artist sizes according to data values."""
    # An object that normalizes data values to [0, 1] range
    norm = None

    def __init__(
        self, plotter, sizes=None, order=None, norm=None,
    ):
        """Map the levels of the `size` variable to distinct values.

        Parameters
        ----------
        # TODO add generic parameters

        """
        super().__init__(plotter)

        data = plotter.plot_data.get("size", pd.Series(dtype=float))

        if data.notna().any():

            map_type = self.infer_map_type(
                norm, sizes, plotter.var_types["size"]
            )

            # --- Option 1: numeric mapping

            if map_type == "numeric":

                levels, lookup_table, norm = self.numeric_mapping(
                    data, sizes, norm,
                )

            # --- Option 2: categorical mapping

            elif map_type == "categorical":

                levels, lookup_table = self.categorical_mapping(
                    data, sizes, order,
                )

            # --- Option 3: datetime mapping

            # TODO this needs an actual implementation
            else:

                levels, lookup_table = self.categorical_mapping(
                    # Casting data to list to handle differences in the way
                    # pandas and numpy represent datetime64 data
                    list(data), sizes, order,
                )

            self.map_type = map_type
            self.levels = levels
            self.norm = norm
            self.sizes = sizes
            self.lookup_table = lookup_table

    def infer_map_type(self, norm, sizes, var_type):

        if norm is not None:
            map_type = "numeric"
        elif isinstance(sizes, (dict, list)):
            map_type = "categorical"
        else:
            map_type = var_type

        return map_type

    def _lookup_single(self, key):

        try:
            value = self.lookup_table[key]
        except KeyError:
            normed = self.norm(key)
            if np.ma.is_masked(normed):
                normed = np.nan
            size_values = self.lookup_table.values()
            size_range = min(size_values), max(size_values)
            value = size_range[0] + normed * np.ptp(size_range)
        return value

    def categorical_mapping(self, data, sizes, order):

        levels = categorical_order(data, order)

        if isinstance(sizes, dict):

            # Dict inputs map existing data values to the size attribute
            missing = set(levels) - set(sizes)
            if any(missing):
                err = f"Missing sizes for the following levels: {missing}"
                raise ValueError(err)
            lookup_table = sizes.copy()

        elif isinstance(sizes, list):

            # List inputs give size values in the same order as the levels
            if len(sizes) != len(levels):
                err = "The `sizes` list has the wrong number of values."
                raise ValueError(err)

            lookup_table = dict(zip(levels, sizes))

        else:

            if isinstance(sizes, tuple):

                # Tuple input sets the min, max size values
                if len(sizes) != 2:
                    err = "A `sizes` tuple must have only 2 values"
                    raise ValueError(err)

            elif sizes is not None:

                err = f"Value for `sizes` not understood: {sizes}"
                raise ValueError(err)

            else:

                # Otherwise, we need to get the min, max size values from
                # the plotter object we are attached to.

                # TODO this is going to cause us trouble later, because we
                # want to restructure things so that the plotter is generic
                # across the visual representation of the data. But at this
                # point, we don't know the visual representation. Likely we
                # want to change the logic of this Mapping so that it gives
                # points on a nornalized range that then gets unnormalized
                # when we know what we're drawing. But given the way the
                # package works now, this way is cleanest.
                sizes = self.plotter._default_size_range

            # For categorical sizes, use regularly-spaced linear steps
            # between the minimum and maximum sizes. Then reverse the
            # ramp so that the largest value is used for the first entry
            # in size_order, etc. This is because "ordered" categoricals
            # are often though to go in decreasing priority.
            sizes = np.linspace(*sizes, len(levels))[::-1]
            lookup_table = dict(zip(levels, sizes))

        return levels, lookup_table

    def numeric_mapping(self, data, sizes, norm):

        if isinstance(sizes, dict):
            # The presence of a norm object overrides a dictionary of sizes
            # in specifying a numeric mapping, so we need to process it
            # dictionary here
            levels = list(np.sort(list(sizes)))
            size_values = sizes.values()
            size_range = min(size_values), max(size_values)

        else:

            # The levels here will be the unique values in the data
            levels = list(np.sort(remove_na(data.unique())))

            if isinstance(sizes, tuple):

                # For numeric inputs, the size can be parametrized by
                # the minimum and maximum artist values to map to. The
                # norm object that gets set up next specifies how to
                # do the mapping.

                if len(sizes) != 2:
                    err = "A `sizes` tuple must have only 2 values"
                    raise ValueError(err)

                size_range = sizes

            elif sizes is not None:

                err = f"Value for `sizes` not understood: {sizes}"
                raise ValueError(err)

            else:

                # When not provided, we get the size range from the plotter
                # object we are attached to. See the note in the categorical
                # method about how this is suboptimal for future development.:
                size_range = self.plotter._default_size_range

        # Now that we know the minimum and maximum sizes that will get drawn,
        # we need to map the data values that we have into that range. We will
        # use a matplotlib Normalize class, which is typically used for numeric
        # color mapping but works fine here too. It takes data values and maps
        # them into a [0, 1] interval, potentially nonlinear-ly.

        if norm is None:
            # Default is a linear function between the min and max data values
            norm = mpl.colors.Normalize()
        elif isinstance(norm, tuple):
            # It is also possible to give different limits in data space
            norm = mpl.colors.Normalize(*norm)
        elif not isinstance(norm, mpl.colors.Normalize):
            err = f"Value for size `norm` parameter not understood: {norm}"
            raise ValueError(err)
        else:
            # If provided with Normalize object, copy it so we can modify
            norm = copy(norm)

        # Set the mapping so all output values are in [0, 1]
        norm.clip = True

        # If the input range is not set, use the full range of the data
        if not norm.scaled():
            norm(levels)

        # Map from data values to [0, 1] range
        sizes_scaled = norm(levels)

        # Now map from the scaled range into the artist units
        if isinstance(sizes, dict):
            lookup_table = sizes
        else:
            lo, hi = size_range
            sizes = lo + sizes_scaled * (hi - lo)
            lookup_table = dict(zip(levels, sizes))

        return levels, lookup_table, norm


@share_init_params_with_map
class StyleMapping(SemanticMapping):
    """Mapping that sets artist style according to data values."""

    # Style mapping is always treated as categorical
    map_type = "categorical"

    def __init__(
        self, plotter, markers=None, dashes=None, order=None,
    ):
        """Map the levels of the `style` variable to distinct values.

        Parameters
        ----------
        # TODO add generic parameters

        """
        super().__init__(plotter)

        data = plotter.plot_data.get("style", pd.Series(dtype=float))

        if data.notna().any():

            # Cast to list to handle numpy/pandas datetime quirks
            if variable_type(data) == "datetime":
                data = list(data)

            # Find ordered unique values
            levels = categorical_order(data, order)

            markers = self._map_attributes(
                markers, levels, unique_markers(len(levels)), "markers",
            )
            dashes = self._map_attributes(
                dashes, levels, unique_dashes(len(levels)), "dashes",
            )

            # Build the paths matplotlib will use to draw the markers
            paths = {}
            filled_markers = []
            for k, m in markers.items():
                if not isinstance(m, mpl.markers.MarkerStyle):
                    m = mpl.markers.MarkerStyle(m)
                paths[k] = m.get_path().transformed(m.get_transform())
                filled_markers.append(m.is_filled())

            # Mixture of filled and unfilled markers will show line art markers
            # in the edge color, which defaults to white. This can be handled,
            # but there would be additional complexity with specifying the
            # weight of the line art markers without overwhelming the filled
            # ones with the edges. So for now, we will disallow mixtures.
            if any(filled_markers) and not all(filled_markers):
                err = "Filled and line art markers cannot be mixed"
                raise ValueError(err)

            lookup_table = {}
            for key in levels:
                lookup_table[key] = {}
                if markers:
                    lookup_table[key]["marker"] = markers[key]
                    lookup_table[key]["path"] = paths[key]
                if dashes:
                    lookup_table[key]["dashes"] = dashes[key]

            self.levels = levels
            self.lookup_table = lookup_table

    def _lookup_single(self, key, attr=None):
        """Get attribute(s) for a given data point."""
        if attr is None:
            value = self.lookup_table[key]
        else:
            value = self.lookup_table[key][attr]
        return value

    def _map_attributes(self, arg, levels, defaults, attr):
        """Handle the specification for a given style attribute."""
        if arg is True:
            lookup_table = dict(zip(levels, defaults))
        elif isinstance(arg, dict):
            missing = set(levels) - set(arg)
            if missing:
                err = f"These `{attr}` levels are missing values: {missing}"
                raise ValueError(err)
            lookup_table = arg
        elif isinstance(arg, Sequence):
            if len(levels) != len(arg):
                err = f"The `{attr}` argument has the wrong number of values"
                raise ValueError(err)
            lookup_table = dict(zip(levels, arg))
        elif arg:
            err = f"This `{attr}` argument was not understood: {arg}"
            raise ValueError(err)
        else:
            lookup_table = {}

        return lookup_table


# =========================================================================== #


class VectorPlotter:
    """Base class for objects underlying *plot functions."""

    _semantic_mappings = {
        "hue": HueMapping,
        "size": SizeMapping,
        "style": StyleMapping,
    }

    # TODO units is another example of a non-mapping "semantic"
    # we need a general name for this and separate handling
    semantics = "x", "y", "hue", "size", "style", "units"
    wide_structure = {
        "x": "@index", "y": "@values", "hue": "@columns", "style": "@columns",
    }
    flat_structure = {"x": "@index", "y": "@values"}

    _default_size_range = 1, 2  # Unused but needed in tests, ugh

    def __init__(self, data=None, variables={}):

        self.assign_variables(data, variables)

        for var, cls in self._semantic_mappings.items():

            # Create the mapping function
            map_func = partial(cls.map, plotter=self)
            setattr(self, f"map_{var}", map_func)

            # Call the mapping function to initialize with default values
            getattr(self, f"map_{var}")()

        self._var_levels = {}

    @classmethod
    def get_semantics(cls, kwargs, semantics=None):
        """Subset a dictionary` arguments with known semantic variables."""
        # TODO this should be get_variables since we have included x and y
        if semantics is None:
            semantics = cls.semantics
        variables = {}
        for key, val in kwargs.items():
            if key in semantics and val is not None:
                variables[key] = val
        return variables

    @property
    def has_xy_data(self):
        """Return True at least one of x or y is defined."""
        return bool({"x", "y"} & set(self.variables))

    @property
    def var_levels(self):
        """Property interface to ordered list of variables levels.

        Each time it's accessed, it updates the var_levels dictionary with the
        list of levels in the current semantic mappers. But it also allows the
        dictionary to persist, so it can be used to set levels by a key. This is
        used to track the list of col/row levels using an attached FacetGrid
        object, but it's kind of messy and ideally fixed by improving the
        faceting logic so it interfaces better with the modern approach to
        tracking plot variables.

        """
        for var in self.variables:
            try:
                map_obj = getattr(self, f"_{var}_map")
                self._var_levels[var] = map_obj.levels
            except AttributeError:
                pass
        return self._var_levels

    def assign_variables(self, data=None, variables={}):
        """Define plot variables, optionally using lookup from `data`."""
        x = variables.get("x", None)
        y = variables.get("y", None)

        if x is None and y is None:
            self.input_format = "wide"
            plot_data, variables = self._assign_variables_wideform(
                data, **variables,
            )
        else:
            self.input_format = "long"
            plot_data, variables = self._assign_variables_longform(
                data, **variables,
            )

        self.plot_data = plot_data
        self.variables = variables
        self.var_types = {
            v: variable_type(
                plot_data[v],
                boolean_type="numeric" if v in "xy" else "categorical"
            )
            for v in variables
        }

        return self

    def _assign_variables_wideform(self, data=None, **kwargs):
        """Define plot variables given wide-form data.

        Parameters
        ----------
        data : flat vector or collection of vectors
            Data can be a vector or mapping that is coerceable to a Series
            or a sequence- or mapping-based collection of such vectors, or a
            rectangular numpy array, or a Pandas DataFrame.
        kwargs : variable -> data mappings
            Behavior with keyword arguments is currently undefined.

        Returns
        -------
        plot_data : :class:`pandas.DataFrame`
            Long-form data object mapping seaborn variables (x, y, hue, ...)
            to data vectors.
        variables : dict
            Keys are defined seaborn variables; values are names inferred from
            the inputs (or None when no name can be determined).

        """
        # Raise if semantic or other variables are assigned in wide-form mode
        assigned = [k for k, v in kwargs.items() if v is not None]
        if any(assigned):
            s = "s" if len(assigned) > 1 else ""
            err = f"The following variable{s} cannot be assigned with wide-form data: "
            err += ", ".join(f"`{v}`" for v in assigned)
            raise ValueError(err)

        # Determine if the data object actually has any data in it
        empty = data is None or not len(data)

        # Then, determine if we have "flat" data (a single vector)
        if isinstance(data, dict):
            values = data.values()
        else:
            values = np.atleast_1d(np.asarray(data, dtype=object))
        flat = not any(
            isinstance(v, Iterable) and not isinstance(v, (str, bytes))
            for v in values
        )

        if empty:

            # Make an object with the structure of plot_data, but empty
            plot_data = pd.DataFrame()
            variables = {}

        elif flat:

            # Handle flat data by converting to pandas Series and using the
            # index and/or values to define x and/or y
            # (Could be accomplished with a more general to_series() interface)
            flat_data = pd.Series(data).copy()
            names = {
                "@values": flat_data.name,
                "@index": flat_data.index.name
            }

            plot_data = {}
            variables = {}

            for var in ["x", "y"]:
                if var in self.flat_structure:
                    attr = self.flat_structure[var]
                    plot_data[var] = getattr(flat_data, attr[1:])
                    variables[var] = names[self.flat_structure[var]]

            plot_data = pd.DataFrame(plot_data)

        else:

            # Otherwise assume we have some collection of vectors.

            # Handle Python sequences such that entries end up in the columns,
            # not in the rows, of the intermediate wide DataFrame.
            # One way to accomplish this is to convert to a dict of Series.
            if isinstance(data, Sequence):
                data_dict = {}
                for i, var in enumerate(data):
                    key = getattr(var, "name", i)
                    # TODO is there a safer/more generic way to ensure Series?
                    # sort of like np.asarray, but for pandas?
                    data_dict[key] = pd.Series(var)

                data = data_dict

            # Pandas requires that dict values either be Series objects
            # or all have the same length, but we want to allow "ragged" inputs
            if isinstance(data, Mapping):
                data = {key: pd.Series(val) for key, val in data.items()}

            # Otherwise, delegate to the pandas DataFrame constructor
            # This is where we'd prefer to use a general interface that says
            # "give me this data as a pandas DataFrame", so we can accept
            # DataFrame objects from other libraries
            wide_data = pd.DataFrame(data, copy=True)

            # At this point we should reduce the dataframe to numeric cols
            numeric_cols = wide_data.apply(variable_type) == "numeric"
            wide_data = wide_data.loc[:, numeric_cols]

            # Now melt the data to long form
            melt_kws = {"var_name": "@columns", "value_name": "@values"}
            use_index = "@index" in self.wide_structure.values()
            if use_index:
                melt_kws["id_vars"] = "@index"
                try:
                    orig_categories = wide_data.columns.categories
                    orig_ordered = wide_data.columns.ordered
                    wide_data.columns = wide_data.columns.add_categories("@index")
                except AttributeError:
                    category_columns = False
                else:
                    category_columns = True
                wide_data["@index"] = wide_data.index.to_series()

            plot_data = wide_data.melt(**melt_kws)

            if use_index and category_columns:
                plot_data["@columns"] = pd.Categorical(plot_data["@columns"],
                                                       orig_categories,
                                                       orig_ordered)

            # Assign names corresponding to plot semantics
            for var, attr in self.wide_structure.items():
                plot_data[var] = plot_data[attr]

            # Define the variable names
            variables = {}
            for var, attr in self.wide_structure.items():
                obj = getattr(wide_data, attr[1:])
                variables[var] = getattr(obj, "name", None)

            # Remove redundant columns from plot_data
            plot_data = plot_data[list(variables)]

        return plot_data, variables

    def _assign_variables_longform(self, data=None, **kwargs):
        """Define plot variables given long-form data and/or vector inputs.

        Parameters
        ----------
        data : dict-like collection of vectors
            Input data where variable names map to vector values.
        kwargs : variable -> data mappings
            Keys are seaborn variables (x, y, hue, ...) and values are vectors
            in any format that can construct a :class:`pandas.DataFrame` or
            names of columns or index levels in ``data``.

        Returns
        -------
        plot_data : :class:`pandas.DataFrame`
            Long-form data object mapping seaborn variables (x, y, hue, ...)
            to data vectors.
        variables : dict
            Keys are defined seaborn variables; values are names inferred from
            the inputs (or None when no name can be determined).

        Raises
        ------
        ValueError
            When variables are strings that don't appear in ``data``.

        """
        plot_data = {}
        variables = {}

        # Data is optional; all variables can be defined as vectors
        if data is None:
            data = {}

        # TODO should we try a data.to_dict() or similar here to more
        # generally accept objects with that interface?
        # Note that dict(df) also works for pandas, and gives us what we
        # want, whereas DataFrame.to_dict() gives a nested dict instead of
        # a dict of series.

        # Variables can also be extraced from the index attribute
        # TODO is this the most general way to enable it?
        # There is no index.to_dict on multiindex, unfortunately
        try:
            index = data.index.to_frame()
        except AttributeError:
            index = {}

        # The caller will determine the order of variables in plot_data
        for key, val in kwargs.items():

            if isinstance(val, (str, bytes)):
                # String inputs trigger __getitem__
                if val in data:
                    # First try to get an entry in the data object
                    plot_data[key] = data[val]
                    variables[key] = val
                elif val in index:
                    # Failing that, try to get an entry in the index object
                    plot_data[key] = index[val]
                    variables[key] = val
                else:
                    # We don't know what this name means
                    err = f"Could not interpret value `{val}` for parameter `{key}`"
                    raise ValueError(err)

            else:

                # Otherwise, assume the value is itself a vector of data

                # Raise when data is present and a vector can't be combined with it
                if isinstance(data, pd.DataFrame) and not isinstance(val, pd.Series):
                    if val is not None and len(data) != len(val):
                        val_cls = val.__class__.__name__
                        err = (
                            f"Length of {val_cls} vectors must match length of `data`"
                            f" when both are used, but `data` has length {len(data)}"
                            f" and the vector passed to `{key}` has length {len(val)}."
                        )
                        raise ValueError(err)

                plot_data[key] = val

                # Try to infer the name of the variable
                variables[key] = getattr(val, "name", None)

        # Construct a tidy plot DataFrame. This will convert a number of
        # types automatically, aligning on index in case of pandas objects
        plot_data = pd.DataFrame(plot_data)

        # Reduce the variables dictionary to fields with valid data
        variables = {
            var: name
            for var, name in variables.items()
            if plot_data[var].notnull().any()
        }

        return plot_data, variables

    def iter_data(
        self, grouping_vars=None, reverse=False, from_comp_data=False,
    ):
        """Generator for getting subsets of data defined by semantic variables.

        Also injects "col" and "row" into grouping semantics.

        Parameters
        ----------
        grouping_vars : string or list of strings
            Semantic variables that define the subsets of data.
        reverse : bool, optional
            If True, reverse the order of iteration.
        from_comp_data : bool, optional
            If True, use self.comp_data rather than self.plot_data

        Yields
        ------
        sub_vars : dict
            Keys are semantic names, values are the level of that semantic.
        sub_data : :class:`pandas.DataFrame`
            Subset of ``plot_data`` for this combination of semantic values.

        """
        # TODO should this default to using all (non x/y?) semantics?
        # or define groupping vars somewhere?
        if grouping_vars is None:
            grouping_vars = []
        elif isinstance(grouping_vars, str):
            grouping_vars = [grouping_vars]
        elif isinstance(grouping_vars, tuple):
            grouping_vars = list(grouping_vars)

        # Always insert faceting variables
        facet_vars = {"col", "row"}
        grouping_vars.extend(
            facet_vars & set(self.variables) - set(grouping_vars)
        )

        # Reduce to the semantics used in this plot
        grouping_vars = [
            var for var in grouping_vars if var in self.variables
        ]

        if from_comp_data:
            data = self.comp_data
        else:
            data = self.plot_data

        if grouping_vars:

            grouped_data = data.groupby(
                grouping_vars, sort=False, as_index=False
            )

            grouping_keys = []
            for var in grouping_vars:
                grouping_keys.append(self.var_levels.get(var, []))

            iter_keys = itertools.product(*grouping_keys)
            if reverse:
                iter_keys = reversed(list(iter_keys))

            for key in iter_keys:

                # Pandas fails with singleton tuple inputs
                pd_key = key[0] if len(key) == 1 else key

                try:
                    data_subset = grouped_data.get_group(pd_key)
                except KeyError:
                    continue

                sub_vars = dict(zip(grouping_vars, key))

                yield sub_vars, data_subset

        else:

            yield {}, data

    @property
    def comp_data(self):
        """Dataframe with numeric x and y, after unit conversion and log scaling."""
        if not hasattr(self, "ax"):
            # Probably a good idea, but will need a bunch of tests updated
            # Most of these tests should just use the external interface
            # Then this can be re-enabled.
            # raise AttributeError("No Axes attached to plotter")
            return self.plot_data

        if not hasattr(self, "_comp_data"):

            comp_data = (
                self.plot_data
                .copy(deep=False)
                .drop(["x", "y"], axis=1, errors="ignore")
            )
            for var in "yx":
                if var not in self.variables:
                    continue

                # Get a corresponding axis object so that we can convert the units
                # to matplotlib's numeric representation, which we can compute on
                # This is messy and it would probably be better for VectorPlotter
                # to manage its own converters (using the matplotlib tools).
                # XXX Currently does not support unshared categorical axes!
                # (But see comment in _attach about how those don't exist)
                if self.ax is None:
                    ax = self.facets.axes.flat[0]
                else:
                    ax = self.ax
                axis = getattr(ax, f"{var}axis")

                comp_var = axis.convert_units(self.plot_data[var])
                if axis.get_scale() == "log":
                    comp_var = np.log10(comp_var)
                comp_data.insert(0, var, comp_var)

            self._comp_data = comp_data

        return self._comp_data

    def _get_axes(self, sub_vars):
        """Return an Axes object based on existence of row/col variables."""
        row = sub_vars.get("row", None)
        col = sub_vars.get("col", None)
        if row is not None and col is not None:
            return self.facets.axes_dict[(row, col)]
        elif row is not None:
            return self.facets.axes_dict[row]
        elif col is not None:
            return self.facets.axes_dict[col]
        elif self.ax is None:
            return self.facets.ax
        else:
            return self.ax

    def _attach(self, obj, allowed_types=None, log_scale=None):
        """Associate the plotter with an Axes manager and initialize its units.

        Parameters
        ----------
        obj : :class:`matplotlib.axes.Axes` or :class:'FacetGrid`
            Structural object that we will eventually plot onto.
        allowed_types : str or list of str
            If provided, raise when either the x or y variable does not have
            one of the declared seaborn types.
        log_scale : bool, number, or pair of bools or numbers
            If not False, set the axes to use log scaling, with the given
            base or defaulting to 10. If a tuple, interpreted as separate
            arguments for the x and y axes.

        """
        from .axisgrid import FacetGrid
        if isinstance(obj, FacetGrid):
            self.ax = None
            self.facets = obj
            ax_list = obj.axes.flatten()
            if obj.col_names is not None:
                self.var_levels["col"] = obj.col_names
            if obj.row_names is not None:
                self.var_levels["row"] = obj.row_names
        else:
            self.ax = obj
            self.facets = None
            ax_list = [obj]

        if allowed_types is None:
            allowed_types = ["numeric", "datetime", "categorical"]
        elif isinstance(allowed_types, str):
            allowed_types = [allowed_types]

        for var in set("xy").intersection(self.variables):
            # Check types of x/y variables
            var_type = self.var_types[var]
            if var_type not in allowed_types:
                err = (
                    f"The {var} variable is {var_type}, but one of "
                    f"{allowed_types} is required"
                )
                raise TypeError(err)

            # Register with the matplotlib unit conversion machinery
            # Perhaps cleaner to manage our own transform objects?
            # XXX Currently this does not allow "unshared" categorical axes
            # We could add metadata to a FacetGrid and set units based on that.
            # See also comment in comp_data, which only uses a single axes to do
            # its mapping, meaning that it won't handle unshared axes well either.
            for ax in ax_list:
                axis = getattr(ax, f"{var}axis")
                seed_data = self.plot_data[var]
                if var_type == "categorical":
                    seed_data = categorical_order(seed_data)
                axis.update_units(seed_data)

        # For categorical y, we want the "first" level to be at the top of the axis
        if self.var_types.get("y", None) == "categorical":
            for ax in ax_list:
                try:
                    ax.yaxis.set_inverted(True)
                except AttributeError:  # mpl < 3.1
                    if not ax.yaxis_inverted():
                        ax.invert_yaxis()

        # Possibly log-scale one or both axes
        if log_scale is not None:
            # Allow single value or x, y tuple
            try:
                scalex, scaley = log_scale
            except TypeError:
                scalex = log_scale if "x" in self.variables else False
                scaley = log_scale if "y" in self.variables else False

            for axis, scale in zip("xy", (scalex, scaley)):
                if scale:
                    for ax in ax_list:
                        set_scale = getattr(ax, f"set_{axis}scale")
                        if scale is True:
                            set_scale("log")
                        else:
                            if LooseVersion(mpl.__version__) >= "3.3":
                                set_scale("log", base=scale)
                            else:
                                set_scale("log", **{f"base{axis}": scale})

    def _log_scaled(self, axis):
        """Return True if specified axis is log scaled on all attached axes."""
        if self.ax is None:
            axes_list = self.facets.axes.flatten()
        else:
            axes_list = [self.ax]

        log_scaled = []
        for ax in axes_list:
            data_axis = getattr(ax, f"{axis}axis")
            log_scaled.append(data_axis.get_scale() == "log")

        if any(log_scaled) and not all(log_scaled):
            raise RuntimeError("Axis scaling is not consistent")

        return any(log_scaled)

    def _add_axis_labels(self, ax, default_x="", default_y=""):
        """Add axis labels if not present, set visibility to match ticklabels."""
        # TODO ax could default to None and use attached axes if present
        # but what to do about the case of facets? Currently using FacetGrid's
        # set_axis_labels method, which doesn't add labels to the interior even
        # when the axes are not shared. Maybe that makes sense?
        if not ax.get_xlabel():
            x_visible = any(t.get_visible() for t in ax.get_xticklabels())
            ax.set_xlabel(self.variables.get("x", default_x), visible=x_visible)
        if not ax.get_ylabel():
            y_visible = any(t.get_visible() for t in ax.get_yticklabels())
            ax.set_ylabel(self.variables.get("y", default_y), visible=y_visible)


def variable_type(vector, boolean_type="numeric"):
    """Determine whether a vector contains numeric, categorical, or dateime data.

    This function differs from the pandas typing API in two ways:

    - Python sequences or object-typed PyData objects are considered numeric if
      all of their entries are numeric.
    - String or mixed-type data are considered categorical even if not
      explicitly represented as a :class:pandas.api.types.CategoricalDtype`.

    Parameters
    ----------
    vector : :func:`pandas.Series`, :func:`numpy.ndarray`, or Python sequence
        Input data to test.
    binary_type : 'numeric' or 'categorical'
        Type to use for vectors containing only 0s and 1s (and NAs).

    Returns
    -------
    var_type : 'numeric', 'categorical', or 'datetime'
        Name identifying the type of data in the vector.

    """
    # Special-case all-na data, which is always "numeric"
    if pd.isna(vector).all():
        return "numeric"

    # Special-case binary/boolean data, allow caller to determine
    # This triggers a numpy warning when vector has strings/objects
    # https://github.com/numpy/numpy/issues/6784
    # Because we reduce with .all(), we are agnostic about whether the
    # comparison returns a scalar or vector, so we will ignore the warning.
    # It triggers a separate DeprecationWarning when the vector has datetimes:
    # https://github.com/numpy/numpy/issues/13548
    # This is considered a bug by numpy and will likely go away.
    with warnings.catch_warnings():
        warnings.simplefilter(
            action='ignore', category=(FutureWarning, DeprecationWarning)
        )
        if np.isin(vector, [0, 1, np.nan]).all():
            return boolean_type

    # Defer to positive pandas tests
    if pd.api.types.is_numeric_dtype(vector):
        return "numeric"

    if pd.api.types.is_categorical_dtype(vector):
        return "categorical"

    if pd.api.types.is_datetime64_dtype(vector):
        return "datetime"

    # --- If we get to here, we need to check the entries

    # Check for a collection where everything is a number

    def all_numeric(x):
        for x_i in x:
            if not isinstance(x_i, Number):
                return False
        return True

    if all_numeric(vector):
        return "numeric"

    # Check for a collection where everything is a datetime

    def all_datetime(x):
        for x_i in x:
            if not isinstance(x_i, (datetime, np.datetime64)):
                return False
        return True

    if all_datetime(vector):
        return "datetime"

    # Otherwise, our final fallback is to consider things categorical

    return "categorical"


def infer_orient(x=None, y=None, orient=None, require_numeric=True):
    """Determine how the plot should be oriented based on the data.

    For historical reasons, the convention is to call a plot "horizontally"
    or "vertically" oriented based on the axis representing its dependent
    variable. Practically, this is used when determining the axis for
    numerical aggregation.

    Paramters
    ---------
    x, y : Vector data or None
        Positional data vectors for the plot.
    orient : string or None
        Specified orientation, which must start with "v" or "h" if not None.
    require_numeric : bool
        If set, raise when the implied dependent variable is not numeric.

    Returns
    -------
    orient : "v" or "h"

    Raises
    ------
    ValueError: When `orient` is not None and does not start with "h" or "v"
    TypeError: When dependant variable is not numeric, with `require_numeric`

    """

    x_type = None if x is None else variable_type(x)
    y_type = None if y is None else variable_type(y)

    nonnumeric_dv_error = "{} orientation requires numeric `{}` variable."
    single_var_warning = "{} orientation ignored with only `{}` specified."

    if x is None:
        if str(orient).startswith("h"):
            warnings.warn(single_var_warning.format("Horizontal", "y"))
        if require_numeric and y_type != "numeric":
            raise TypeError(nonnumeric_dv_error.format("Vertical", "y"))
        return "v"

    elif y is None:
        if str(orient).startswith("v"):
            warnings.warn(single_var_warning.format("Vertical", "x"))
        if require_numeric and x_type != "numeric":
            raise TypeError(nonnumeric_dv_error.format("Horizontal", "x"))
        return "h"

    elif str(orient).startswith("v"):
        if require_numeric and y_type != "numeric":
            raise TypeError(nonnumeric_dv_error.format("Vertical", "y"))
        return "v"

    elif str(orient).startswith("h"):
        if require_numeric and x_type != "numeric":
            raise TypeError(nonnumeric_dv_error.format("Horizontal", "x"))
        return "h"

    elif orient is not None:
        raise ValueError(f"Value for `orient` not understood: {orient}")

    elif x_type != "numeric" and y_type == "numeric":
        return "v"

    elif x_type == "numeric" and y_type != "numeric":
        return "h"

    elif require_numeric and "numeric" not in (x_type, y_type):
        err = "Neither the `x` nor `y` variable appears to be numeric."
        raise TypeError(err)

    else:
        return "v"


def unique_dashes(n):
    """Build an arbitrarily long list of unique dash styles for lines.

    Parameters
    ----------
    n : int
        Number of unique dash specs to generate.

    Returns
    -------
    dashes : list of strings or tuples
        Valid arguments for the ``dashes`` parameter on
        :class:`matplotlib.lines.Line2D`. The first spec is a solid
        line (``""``), the remainder are sequences of long and short
        dashes.

    """
    # Start with dash specs that are well distinguishable
    dashes = [
        "",
        (4, 1.5),
        (1, 1),
        (3, 1.25, 1.5, 1.25),
        (5, 1, 1, 1),
    ]

    # Now programatically build as many as we need
    p = 3
    while len(dashes) < n:

        # Take combinations of long and short dashes
        a = itertools.combinations_with_replacement([3, 1.25], p)
        b = itertools.combinations_with_replacement([4, 1], p)

        # Interleave the combinations, reversing one of the streams
        segment_list = itertools.chain(*zip(
            list(a)[1:-1][::-1],
            list(b)[1:-1]
        ))

        # Now insert the gaps
        for segments in segment_list:
            gap = min(segments)
            spec = tuple(itertools.chain(*((seg, gap) for seg in segments)))
            dashes.append(spec)

        p += 1

    return dashes[:n]


def unique_markers(n):
    """Build an arbitrarily long list of unique marker styles for points.

    Parameters
    ----------
    n : int
        Number of unique marker specs to generate.

    Returns
    -------
    markers : list of string or tuples
        Values for defining :class:`matplotlib.markers.MarkerStyle` objects.
        All markers will be filled.

    """
    # Start with marker specs that are well distinguishable
    markers = [
        "o",
        "X",
        (4, 0, 45),
        "P",
        (4, 0, 0),
        (4, 1, 0),
        "^",
        (4, 1, 45),
        "v",
    ]

    # Now generate more from regular polygons of increasing order
    s = 5
    while len(markers) < n:
        a = 360 / (s + 1) / 2
        markers.extend([
            (s + 1, 1, a),
            (s + 1, 0, a),
            (s, 1, 0),
            (s, 0, 0),
        ])
        s += 1

    # Convert to MarkerStyle object, using only exactly what we need
    # markers = [mpl.markers.MarkerStyle(m) for m in markers[:n]]

    return markers[:n]


def categorical_order(vector, order=None):
    """Return a list of unique data values.

    Determine an ordered list of levels in ``values``.

    Parameters
    ----------
    vector : list, array, Categorical, or Series
        Vector of "categorical" values
    order : list-like, optional
        Desired order of category levels to override the order determined
        from the ``values`` object.

    Returns
    -------
    order : list
        Ordered list of category levels not including null values.

    """
    if order is None:
        if hasattr(vector, "categories"):
            order = vector.categories
        else:
            try:
                order = vector.cat.categories
            except (TypeError, AttributeError):

                try:
                    order = vector.unique()
                except AttributeError:
                    order = pd.unique(vector)

                if variable_type(vector) == "numeric":
                    order = np.sort(order)

        order = filter(pd.notnull, order)
    return list(order)