# This file is part of Patsy # Copyright (C) 2011-2015 Nathaniel Smith # See file LICENSE.txt for license information. # This file defines the main class for storing metadata about a model # design. It also defines a 'value-added' design matrix type -- a subclass of # ndarray that represents a design matrix and holds metadata about its # columns. The intent is that these are useful and usable data structures # even if you're not using *any* of the rest of patsy to actually build # your matrices. # XX TMP TODO: # # - update design_matrix_builders and build_design_matrices docs # - add tests and docs for new design info stuff # - consider renaming design_matrix_builders (and I guess # build_design_matrices too). Ditto for highlevel dbuilder functions. from __future__ import print_function # These are made available in the patsy.* namespace __all__ = ["DesignInfo", "FactorInfo", "SubtermInfo", "DesignMatrix"] import warnings import numbers import six import numpy as np from patsy import PatsyError from patsy.util import atleast_2d_column_default from patsy.compat import OrderedDict from patsy.util import (repr_pretty_delegate, repr_pretty_impl, safe_issubdtype, no_pickling, assert_no_pickling) from patsy.constraint import linear_constraint from patsy.contrasts import ContrastMatrix from patsy.desc import ModelDesc, Term class FactorInfo(object): """A FactorInfo object is a simple class that provides some metadata about the role of a factor within a model. :attr:`DesignInfo.factor_infos` is a dictionary which maps factor objects to FactorInfo objects for each factor in the model. .. versionadded:: 0.4.0 Attributes: .. attribute:: factor The factor object being described. .. attribute:: type The type of the factor -- either the string ``"numerical"`` or the string ``"categorical"``. .. attribute:: state An opaque object which holds the state needed to evaluate this factor on new data (e.g., for prediction). See :meth:`factor_protocol.eval`. .. attribute:: num_columns For numerical factors, the number of columns this factor produces. For categorical factors, this attribute will always be ``None``. .. attribute:: categories For categorical factors, a tuple of the possible categories this factor takes on, in order. For numerical factors, this attribute will always be ``None``. """ def __init__(self, factor, type, state, num_columns=None, categories=None): self.factor = factor self.type = type if self.type not in ["numerical", "categorical"]: raise ValueError("FactorInfo.type must be " "'numerical' or 'categorical', not %r" % (self.type,)) self.state = state if self.type == "numerical": if not isinstance(num_columns, six.integer_types): raise ValueError("For numerical factors, num_columns " "must be an integer") if categories is not None: raise ValueError("For numerical factors, categories " "must be None") else: assert self.type == "categorical" if num_columns is not None: raise ValueError("For categorical factors, num_columns " "must be None") categories = tuple(categories) self.num_columns = num_columns self.categories = categories __repr__ = repr_pretty_delegate def _repr_pretty_(self, p, cycle): assert not cycle class FactorState(object): def __repr__(self): return "" kwlist = [("factor", self.factor), ("type", self.type), # Don't put the state in people's faces, it will # just encourage them to pay attention to the # contents :-). Plus it's a bunch of gobbledygook # they don't care about. They can always look at # self.state if they want to know... ("state", FactorState()), ] if self.type == "numerical": kwlist.append(("num_columns", self.num_columns)) else: kwlist.append(("categories", self.categories)) repr_pretty_impl(p, self, [], kwlist) __getstate__ = no_pickling def test_FactorInfo(): fi1 = FactorInfo("asdf", "numerical", {"a": 1}, num_columns=10) assert fi1.factor == "asdf" assert fi1.state == {"a": 1} assert fi1.type == "numerical" assert fi1.num_columns == 10 assert fi1.categories is None # smoke test repr(fi1) fi2 = FactorInfo("asdf", "categorical", {"a": 2}, categories=["z", "j"]) assert fi2.factor == "asdf" assert fi2.state == {"a": 2} assert fi2.type == "categorical" assert fi2.num_columns is None assert fi2.categories == ("z", "j") # smoke test repr(fi2) import pytest pytest.raises(ValueError, FactorInfo, "asdf", "non-numerical", {}) pytest.raises(ValueError, FactorInfo, "asdf", "numerical", {}) pytest.raises(ValueError, FactorInfo, "asdf", "numerical", {}, num_columns="asdf") pytest.raises(ValueError, FactorInfo, "asdf", "numerical", {}, num_columns=1, categories=1) pytest.raises(TypeError, FactorInfo, "asdf", "categorical", {}) pytest.raises(ValueError, FactorInfo, "asdf", "categorical", {}, num_columns=1) pytest.raises(TypeError, FactorInfo, "asdf", "categorical", {}, categories=1) # Make sure longs are legal for num_columns # (Important on python2+win64, where array shapes are tuples-of-longs) if not six.PY3: fi_long = FactorInfo("asdf", "numerical", {"a": 1}, num_columns=long(10)) assert fi_long.num_columns == 10 class SubtermInfo(object): """A SubtermInfo object is a simple metadata container describing a single primitive interaction and how it is coded in our design matrix. Our final design matrix is produced by coding each primitive interaction in order from left to right, and then stacking the resulting columns. For each :class:`Term`, we have one or more of these objects which describe how that term is encoded. :attr:`DesignInfo.term_codings` is a dictionary which maps term objects to lists of SubtermInfo objects. To code a primitive interaction, the following steps are performed: * Evaluate each factor on the provided data. * Encode each factor into one or more proto-columns. For numerical factors, these proto-columns are identical to whatever the factor evaluates to; for categorical factors, they are encoded using a specified contrast matrix. * Form all pairwise, elementwise products between proto-columns generated by different factors. (For example, if factor 1 generated proto-columns A and B, and factor 2 generated proto-columns C and D, then our final columns are ``A * C``, ``B * C``, ``A * D``, ``B * D``.) * The resulting columns are stored directly into the final design matrix. Sometimes multiple primitive interactions are needed to encode a single term; this occurs, for example, in the formula ``"1 + a:b"`` when ``a`` and ``b`` are categorical. See :ref:`formulas-building` for full details. .. versionadded:: 0.4.0 Attributes: .. attribute:: factors The factors which appear in this subterm's interaction. .. attribute:: contrast_matrices A dict mapping factor objects to :class:`ContrastMatrix` objects, describing how each categorical factor in this interaction is coded. .. attribute:: num_columns The number of design matrix columns which this interaction generates. """ def __init__(self, factors, contrast_matrices, num_columns): self.factors = tuple(factors) factor_set = frozenset(factors) if not isinstance(contrast_matrices, dict): raise ValueError("contrast_matrices must be dict") for factor, contrast_matrix in six.iteritems(contrast_matrices): if factor not in factor_set: raise ValueError("Unexpected factor in contrast_matrices dict") if not isinstance(contrast_matrix, ContrastMatrix): raise ValueError("Expected a ContrastMatrix, not %r" % (contrast_matrix,)) self.contrast_matrices = contrast_matrices if not isinstance(num_columns, six.integer_types): raise ValueError("num_columns must be an integer") self.num_columns = num_columns __repr__ = repr_pretty_delegate def _repr_pretty_(self, p, cycle): assert not cycle repr_pretty_impl(p, self, [], [("factors", self.factors), ("contrast_matrices", self.contrast_matrices), ("num_columns", self.num_columns)]) __getstate__ = no_pickling def test_SubtermInfo(): cm = ContrastMatrix(np.ones((2, 2)), ["[1]", "[2]"]) s = SubtermInfo(["a", "x"], {"a": cm}, 4) assert s.factors == ("a", "x") assert s.contrast_matrices == {"a": cm} assert s.num_columns == 4 # Make sure longs are accepted for num_columns if not six.PY3: s = SubtermInfo(["a", "x"], {"a": cm}, long(4)) assert s.num_columns == 4 # smoke test repr(s) import pytest pytest.raises(TypeError, SubtermInfo, 1, {}, 1) pytest.raises(ValueError, SubtermInfo, ["a", "x"], 1, 1) pytest.raises(ValueError, SubtermInfo, ["a", "x"], {"z": cm}, 1) pytest.raises(ValueError, SubtermInfo, ["a", "x"], {"a": 1}, 1) pytest.raises(ValueError, SubtermInfo, ["a", "x"], {}, 1.5) class DesignInfo(object): """A DesignInfo object holds metadata about a design matrix. This is the main object that Patsy uses to pass metadata about a design matrix to statistical libraries, in order to allow further downstream processing like intelligent tests, prediction on new data, etc. Usually encountered as the `.design_info` attribute on design matrices. """ def __init__(self, column_names, factor_infos=None, term_codings=None): self.column_name_indexes = OrderedDict(zip(column_names, range(len(column_names)))) if (factor_infos is None) != (term_codings is None): raise ValueError("Must specify either both or neither of " "factor_infos= and term_codings=") self.factor_infos = factor_infos self.term_codings = term_codings # factor_infos is a dict containing one entry for every factor # mentioned in our terms # and mapping each to FactorInfo object if self.factor_infos is not None: if not isinstance(self.factor_infos, dict): raise ValueError("factor_infos should be a dict") if not isinstance(self.term_codings, OrderedDict): raise ValueError("term_codings must be an OrderedDict") for term, subterms in six.iteritems(self.term_codings): if not isinstance(term, Term): raise ValueError("expected a Term, not %r" % (term,)) if not isinstance(subterms, list): raise ValueError("term_codings must contain lists") term_factors = set(term.factors) for subterm in subterms: if not isinstance(subterm, SubtermInfo): raise ValueError("expected SubtermInfo, " "not %r" % (subterm,)) if not term_factors.issuperset(subterm.factors): raise ValueError("unexpected factors in subterm") all_factors = set() for term in self.term_codings: all_factors.update(term.factors) if all_factors != set(self.factor_infos): raise ValueError("Provided Term objects and factor_infos " "do not match") for factor, factor_info in six.iteritems(self.factor_infos): if not isinstance(factor_info, FactorInfo): raise ValueError("expected FactorInfo object, not %r" % (factor_info,)) if factor != factor_info.factor: raise ValueError("mismatched factor_info.factor") for term, subterms in six.iteritems(self.term_codings): for subterm in subterms: exp_cols = 1 cat_factors = set() for factor in subterm.factors: fi = self.factor_infos[factor] if fi.type == "numerical": exp_cols *= fi.num_columns else: assert fi.type == "categorical" cm = subterm.contrast_matrices[factor].matrix if cm.shape[0] != len(fi.categories): raise ValueError("Mismatched contrast matrix " "for factor %r" % (factor,)) cat_factors.add(factor) exp_cols *= cm.shape[1] if cat_factors != set(subterm.contrast_matrices): raise ValueError("Mismatch between contrast_matrices " "and categorical factors") if exp_cols != subterm.num_columns: raise ValueError("Unexpected num_columns") if term_codings is None: # Need to invent term information self.term_slices = None # We invent one term per column, with the same name as the column term_names = column_names slices = [slice(i, i + 1) for i in range(len(column_names))] self.term_name_slices = OrderedDict(zip(term_names, slices)) else: # Need to derive term information from term_codings self.term_slices = OrderedDict() idx = 0 for term, subterm_infos in six.iteritems(self.term_codings): term_columns = 0 for subterm_info in subterm_infos: term_columns += subterm_info.num_columns self.term_slices[term] = slice(idx, idx + term_columns) idx += term_columns if idx != len(self.column_names): raise ValueError("mismatch between column_names and columns " "coded by given terms") self.term_name_slices = OrderedDict( [(term.name(), slice_) for (term, slice_) in six.iteritems(self.term_slices)]) # Guarantees: # term_name_slices is never None # The slices in term_name_slices are in order and exactly cover the # whole range of columns. # term_slices may be None # If term_slices is not None, then its slices match the ones in # term_name_slices. assert self.term_name_slices is not None if self.term_slices is not None: assert (list(self.term_slices.values()) == list(self.term_name_slices.values())) # These checks probably aren't necessary anymore now that we always # generate the slices ourselves, but we'll leave them in just to be # safe. covered = 0 for slice_ in six.itervalues(self.term_name_slices): start, stop, step = slice_.indices(len(column_names)) assert start == covered assert step == 1 covered = stop assert covered == len(column_names) # If there is any name overlap between terms and columns, they refer # to the same columns. for column_name, index in six.iteritems(self.column_name_indexes): if column_name in self.term_name_slices: slice_ = self.term_name_slices[column_name] if slice_ != slice(index, index + 1): raise ValueError("term/column name collision") __repr__ = repr_pretty_delegate def _repr_pretty_(self, p, cycle): assert not cycle repr_pretty_impl(p, self, [self.column_names], [("factor_infos", self.factor_infos), ("term_codings", self.term_codings)]) @property def column_names(self): "A list of the column names, in order." return list(self.column_name_indexes) @property def terms(self): "A list of :class:`Terms`, in order, or else None." if self.term_slices is None: return None return list(self.term_slices) @property def term_names(self): "A list of terms, in order." return list(self.term_name_slices) @property def builder(self): ".. deprecated:: 0.4.0" warnings.warn(DeprecationWarning( "The DesignInfo.builder attribute is deprecated starting in " "patsy v0.4.0; distinct builder objects have been eliminated " "and design_info.builder is now just a long-winded way of " "writing 'design_info' (i.e. the .builder attribute just " "returns self)"), stacklevel=2) return self @property def design_info(self): ".. deprecated:: 0.4.0" warnings.warn(DeprecationWarning( "Starting in patsy v0.4.0, the DesignMatrixBuilder class has " "been merged into the DesignInfo class. So there's no need to " "use builder.design_info to access the DesignInfo; 'builder' " "already *is* a DesignInfo."), stacklevel=2) return self def slice(self, columns_specifier): """Locate a subset of design matrix columns, specified symbolically. A patsy design matrix has two levels of structure: the individual columns (which are named), and the :ref:`terms ` in the formula that generated those columns. This is a one-to-many relationship: a single term may span several columns. This method provides a user-friendly API for locating those columns. (While we talk about columns here, this is probably most useful for indexing into other arrays that are derived from the design matrix, such as regression coefficients or covariance matrices.) The `columns_specifier` argument can take a number of forms: * A term name * A column name * A :class:`Term` object * An integer giving a raw index * A raw slice object In all cases, a Python :func:`slice` object is returned, which can be used directly for indexing. Example:: y, X = dmatrices("y ~ a", demo_data("y", "a", nlevels=3)) betas = np.linalg.lstsq(X, y)[0] a_betas = betas[X.design_info.slice("a")] (If you want to look up a single individual column by name, use ``design_info.column_name_indexes[name]``.) """ if isinstance(columns_specifier, slice): return columns_specifier if np.issubdtype(type(columns_specifier), np.integer): return slice(columns_specifier, columns_specifier + 1) if (self.term_slices is not None and columns_specifier in self.term_slices): return self.term_slices[columns_specifier] if columns_specifier in self.term_name_slices: return self.term_name_slices[columns_specifier] if columns_specifier in self.column_name_indexes: idx = self.column_name_indexes[columns_specifier] return slice(idx, idx + 1) raise PatsyError("unknown column specified '%s'" % (columns_specifier,)) def linear_constraint(self, constraint_likes): """Construct a linear constraint in matrix form from a (possibly symbolic) description. Possible inputs: * A dictionary which is taken as a set of equality constraint. Keys can be either string column names, or integer column indexes. * A string giving a arithmetic expression referring to the matrix columns by name. * A list of such strings which are ANDed together. * A tuple (A, b) where A and b are array_likes, and the constraint is Ax = b. If necessary, these will be coerced to the proper dimensionality by appending dimensions with size 1. The string-based language has the standard arithmetic operators, / * + - and parentheses, plus "=" is used for equality and "," is used to AND together multiple constraint equations within a string. You can If no = appears in some expression, then that expression is assumed to be equal to zero. Division is always float-based, even if ``__future__.true_division`` isn't in effect. Returns a :class:`LinearConstraint` object. Examples:: di = DesignInfo(["x1", "x2", "x3"]) # Equivalent ways to write x1 == 0: di.linear_constraint({"x1": 0}) # by name di.linear_constraint({0: 0}) # by index di.linear_constraint("x1 = 0") # string based di.linear_constraint("x1") # can leave out "= 0" di.linear_constraint("2 * x1 = (x1 + 2 * x1) / 3") di.linear_constraint(([1, 0, 0], 0)) # constraint matrices # Equivalent ways to write x1 == 0 and x3 == 10 di.linear_constraint({"x1": 0, "x3": 10}) di.linear_constraint({0: 0, 2: 10}) di.linear_constraint({0: 0, "x3": 10}) di.linear_constraint("x1 = 0, x3 = 10") di.linear_constraint("x1, x3 = 10") di.linear_constraint(["x1", "x3 = 0"]) # list of strings di.linear_constraint("x1 = 0, x3 - 10 = x1") di.linear_constraint([[1, 0, 0], [0, 0, 1]], [0, 10]) # You can also chain together equalities, just like Python: di.linear_constraint("x1 = x2 = 3") """ return linear_constraint(constraint_likes, self.column_names) def describe(self): """Returns a human-readable string describing this design info. Example: .. ipython:: In [1]: y, X = dmatrices("y ~ x1 + x2", demo_data("y", "x1", "x2")) In [2]: y.design_info.describe() Out[2]: 'y' In [3]: X.design_info.describe() Out[3]: '1 + x1 + x2' .. warning:: There is no guarantee that the strings returned by this function can be parsed as formulas, or that if they can be parsed as a formula that they will produce a model equivalent to the one you started with. This function produces a best-effort description intended for humans to read. """ names = [] for name in self.term_names: if name == "Intercept": names.append("1") else: names.append(name) return " + ".join(names) def subset(self, which_terms): """Create a new :class:`DesignInfo` for design matrices that contain a subset of the terms that the current :class:`DesignInfo` does. For example, if ``design_info`` has terms ``x``, ``y``, and ``z``, then:: design_info2 = design_info.subset(["x", "z"]) will return a new DesignInfo that can be used to construct design matrices with only the columns corresponding to the terms ``x`` and ``z``. After we do this, then in general these two expressions will return the same thing (here we assume that ``x``, ``y``, and ``z`` each generate a single column of the output):: build_design_matrix([design_info], data)[0][:, [0, 2]] build_design_matrix([design_info2], data)[0] However, a critical difference is that in the second case, ``data`` need not contain any values for ``y``. This is very useful when doing prediction using a subset of a model, in which situation R usually forces you to specify dummy values for ``y``. If using a formula to specify the terms to include, remember that like any formula, the intercept term will be included by default, so use ``0`` or ``-1`` in your formula if you want to avoid this. This method can also be used to reorder the terms in your design matrix, in case you want to do that for some reason. I can't think of any. Note that this method will generally *not* produce the same result as creating a new model directly. Consider these DesignInfo objects:: design1 = dmatrix("1 + C(a)", data) design2 = design1.subset("0 + C(a)") design3 = dmatrix("0 + C(a)", data) Here ``design2`` and ``design3`` will both produce design matrices that contain an encoding of ``C(a)`` without any intercept term. But ``design3`` uses a full-rank encoding for the categorical term ``C(a)``, while ``design2`` uses the same reduced-rank encoding as ``design1``. :arg which_terms: The terms which should be kept in the new :class:`DesignMatrixBuilder`. If this is a string, then it is parsed as a formula, and then the names of the resulting terms are taken as the terms to keep. If it is a list, then it can contain a mixture of term names (as strings) and :class:`Term` objects. .. versionadded: 0.2.0 New method on the class DesignMatrixBuilder. .. versionchanged: 0.4.0 Moved from DesignMatrixBuilder to DesignInfo, as part of the removal of DesignMatrixBuilder. """ if isinstance(which_terms, str): desc = ModelDesc.from_formula(which_terms) if desc.lhs_termlist: raise PatsyError("right-hand-side-only formula required") which_terms = [term.name() for term in desc.rhs_termlist] if self.term_codings is None: # This is a minimal DesignInfo # If the name is unknown we just let the KeyError escape new_names = [] for t in which_terms: new_names += self.column_names[self.term_name_slices[t]] return DesignInfo(new_names) else: term_name_to_term = {} for term in self.term_codings: term_name_to_term[term.name()] = term new_column_names = [] new_factor_infos = {} new_term_codings = OrderedDict() for name_or_term in which_terms: term = term_name_to_term.get(name_or_term, name_or_term) # If the name is unknown we just let the KeyError escape s = self.term_slices[term] new_column_names += self.column_names[s] for f in term.factors: new_factor_infos[f] = self.factor_infos[f] new_term_codings[term] = self.term_codings[term] return DesignInfo(new_column_names, factor_infos=new_factor_infos, term_codings=new_term_codings) @classmethod def from_array(cls, array_like, default_column_prefix="column"): """Find or construct a DesignInfo appropriate for a given array_like. If the input `array_like` already has a ``.design_info`` attribute, then it will be returned. Otherwise, a new DesignInfo object will be constructed, using names either taken from the `array_like` (e.g., for a pandas DataFrame with named columns), or constructed using `default_column_prefix`. This is how :func:`dmatrix` (for example) creates a DesignInfo object if an arbitrary matrix is passed in. :arg array_like: An ndarray or pandas container. :arg default_column_prefix: If it's necessary to invent column names, then this will be used to construct them. :returns: a DesignInfo object """ if hasattr(array_like, "design_info") and isinstance(array_like.design_info, cls): return array_like.design_info arr = atleast_2d_column_default(array_like, preserve_pandas=True) if arr.ndim > 2: raise ValueError("design matrix can't have >2 dimensions") columns = getattr(arr, "columns", range(arr.shape[1])) if (hasattr(columns, "dtype") and not safe_issubdtype(columns.dtype, np.integer)): column_names = [str(obj) for obj in columns] else: column_names = ["%s%s" % (default_column_prefix, i) for i in columns] return DesignInfo(column_names) __getstate__ = no_pickling def test_DesignInfo(): import pytest class _MockFactor(object): def __init__(self, name): self._name = name def name(self): return self._name f_x = _MockFactor("x") f_y = _MockFactor("y") t_x = Term([f_x]) t_y = Term([f_y]) factor_infos = {f_x: FactorInfo(f_x, "numerical", {}, num_columns=3), f_y: FactorInfo(f_y, "numerical", {}, num_columns=1), } term_codings = OrderedDict([(t_x, [SubtermInfo([f_x], {}, 3)]), (t_y, [SubtermInfo([f_y], {}, 1)])]) di = DesignInfo(["x1", "x2", "x3", "y"], factor_infos, term_codings) assert di.column_names == ["x1", "x2", "x3", "y"] assert di.term_names == ["x", "y"] assert di.terms == [t_x, t_y] assert di.column_name_indexes == {"x1": 0, "x2": 1, "x3": 2, "y": 3} assert di.term_name_slices == {"x": slice(0, 3), "y": slice(3, 4)} assert di.term_slices == {t_x: slice(0, 3), t_y: slice(3, 4)} assert di.describe() == "x + y" assert di.slice(1) == slice(1, 2) assert di.slice("x1") == slice(0, 1) assert di.slice("x2") == slice(1, 2) assert di.slice("x3") == slice(2, 3) assert di.slice("x") == slice(0, 3) assert di.slice(t_x) == slice(0, 3) assert di.slice("y") == slice(3, 4) assert di.slice(t_y) == slice(3, 4) assert di.slice(slice(2, 4)) == slice(2, 4) pytest.raises(PatsyError, di.slice, "asdf") # smoke test repr(di) assert_no_pickling(di) # One without term objects di = DesignInfo(["a1", "a2", "a3", "b"]) assert di.column_names == ["a1", "a2", "a3", "b"] assert di.term_names == ["a1", "a2", "a3", "b"] assert di.terms is None assert di.column_name_indexes == {"a1": 0, "a2": 1, "a3": 2, "b": 3} assert di.term_name_slices == {"a1": slice(0, 1), "a2": slice(1, 2), "a3": slice(2, 3), "b": slice(3, 4)} assert di.term_slices is None assert di.describe() == "a1 + a2 + a3 + b" assert di.slice(1) == slice(1, 2) assert di.slice("a1") == slice(0, 1) assert di.slice("a2") == slice(1, 2) assert di.slice("a3") == slice(2, 3) assert di.slice("b") == slice(3, 4) # Check intercept handling in describe() assert DesignInfo(["Intercept", "a", "b"]).describe() == "1 + a + b" # Failure modes # must specify either both or neither of factor_infos and term_codings: pytest.raises(ValueError, DesignInfo, ["x1", "x2", "x3", "y"], factor_infos=factor_infos) pytest.raises(ValueError, DesignInfo, ["x1", "x2", "x3", "y"], term_codings=term_codings) # factor_infos must be a dict pytest.raises(ValueError, DesignInfo, ["x1", "x2", "x3", "y"], list(factor_infos), term_codings) # wrong number of column names: pytest.raises(ValueError, DesignInfo, ["x1", "x2", "x3", "y1", "y2"], factor_infos, term_codings) pytest.raises(ValueError, DesignInfo, ["x1", "x2", "x3"], factor_infos, term_codings) # name overlap problems pytest.raises(ValueError, DesignInfo, ["x1", "x2", "y", "y2"], factor_infos, term_codings) # duplicate name pytest.raises(ValueError, DesignInfo, ["x1", "x1", "x1", "y"], factor_infos, term_codings) # f_y is in factor_infos, but not mentioned in any term term_codings_x_only = OrderedDict(term_codings) del term_codings_x_only[t_y] pytest.raises(ValueError, DesignInfo, ["x1", "x2", "x3"], factor_infos, term_codings_x_only) # f_a is in a term, but not in factor_infos f_a = _MockFactor("a") t_a = Term([f_a]) term_codings_with_a = OrderedDict(term_codings) term_codings_with_a[t_a] = [SubtermInfo([f_a], {}, 1)] pytest.raises(ValueError, DesignInfo, ["x1", "x2", "x3", "y", "a"], factor_infos, term_codings_with_a) # bad factor_infos not_factor_infos = dict(factor_infos) not_factor_infos[f_x] = "what is this I don't even" pytest.raises(ValueError, DesignInfo, ["x1", "x2", "x3", "y"], not_factor_infos, term_codings) mismatch_factor_infos = dict(factor_infos) mismatch_factor_infos[f_x] = FactorInfo(f_a, "numerical", {}, num_columns=3) pytest.raises(ValueError, DesignInfo, ["x1", "x2", "x3", "y"], mismatch_factor_infos, term_codings) # bad term_codings pytest.raises(ValueError, DesignInfo, ["x1", "x2", "x3", "y"], factor_infos, dict(term_codings)) not_term_codings = OrderedDict(term_codings) not_term_codings["this is a string"] = term_codings[t_x] pytest.raises(ValueError, DesignInfo, ["x1", "x2", "x3", "y"], factor_infos, not_term_codings) non_list_term_codings = OrderedDict(term_codings) non_list_term_codings[t_y] = tuple(term_codings[t_y]) pytest.raises(ValueError, DesignInfo, ["x1", "x2", "x3", "y"], factor_infos, non_list_term_codings) non_subterm_term_codings = OrderedDict(term_codings) non_subterm_term_codings[t_y][0] = "not a SubtermInfo" pytest.raises(ValueError, DesignInfo, ["x1", "x2", "x3", "y"], factor_infos, non_subterm_term_codings) bad_subterm = OrderedDict(term_codings) # f_x is a factor in this model, but it is not a factor in t_y term_codings[t_y][0] = SubtermInfo([f_x], {}, 1) pytest.raises(ValueError, DesignInfo, ["x1", "x2", "x3", "y"], factor_infos, bad_subterm) # contrast matrix has wrong number of rows factor_codings_a = {f_a: FactorInfo(f_a, "categorical", {}, categories=["a1", "a2"])} term_codings_a_bad_rows = OrderedDict([ (t_a, [SubtermInfo([f_a], {f_a: ContrastMatrix(np.ones((3, 2)), ["[1]", "[2]"])}, 2)])]) pytest.raises(ValueError, DesignInfo, ["a[1]", "a[2]"], factor_codings_a, term_codings_a_bad_rows) # have a contrast matrix for a non-categorical factor t_ax = Term([f_a, f_x]) factor_codings_ax = {f_a: FactorInfo(f_a, "categorical", {}, categories=["a1", "a2"]), f_x: FactorInfo(f_x, "numerical", {}, num_columns=2)} term_codings_ax_extra_cm = OrderedDict([ (t_ax, [SubtermInfo([f_a, f_x], {f_a: ContrastMatrix(np.ones((2, 2)), ["[1]", "[2]"]), f_x: ContrastMatrix(np.ones((2, 2)), ["[1]", "[2]"])}, 4)])]) pytest.raises(ValueError, DesignInfo, ["a[1]:x[1]", "a[2]:x[1]", "a[1]:x[2]", "a[2]:x[2]"], factor_codings_ax, term_codings_ax_extra_cm) # no contrast matrix for a categorical factor term_codings_ax_missing_cm = OrderedDict([ (t_ax, [SubtermInfo([f_a, f_x], {}, 4)])]) # This actually fails before it hits the relevant check with a KeyError, # but that's okay... the previous test still exercises the check. pytest.raises((ValueError, KeyError), DesignInfo, ["a[1]:x[1]", "a[2]:x[1]", "a[1]:x[2]", "a[2]:x[2]"], factor_codings_ax, term_codings_ax_missing_cm) # subterm num_columns doesn't match the value computed from the individual # factors term_codings_ax_wrong_subterm_columns = OrderedDict([ (t_ax, [SubtermInfo([f_a, f_x], {f_a: ContrastMatrix(np.ones((2, 3)), ["[1]", "[2]", "[3]"])}, # should be 2 * 3 = 6 5)])]) pytest.raises(ValueError, DesignInfo, ["a[1]:x[1]", "a[2]:x[1]", "a[3]:x[1]", "a[1]:x[2]", "a[2]:x[2]", "a[3]:x[2]"], factor_codings_ax, term_codings_ax_wrong_subterm_columns) def test_DesignInfo_from_array(): di = DesignInfo.from_array([1, 2, 3]) assert di.column_names == ["column0"] di2 = DesignInfo.from_array([[1, 2], [2, 3], [3, 4]]) assert di2.column_names == ["column0", "column1"] di3 = DesignInfo.from_array([1, 2, 3], default_column_prefix="x") assert di3.column_names == ["x0"] di4 = DesignInfo.from_array([[1, 2], [2, 3], [3, 4]], default_column_prefix="x") assert di4.column_names == ["x0", "x1"] m = DesignMatrix([1, 2, 3], di3) assert DesignInfo.from_array(m) is di3 # But weird objects are ignored m.design_info = "asdf" di_weird = DesignInfo.from_array(m) assert di_weird.column_names == ["column0"] import pytest pytest.raises(ValueError, DesignInfo.from_array, np.ones((2, 2, 2))) from patsy.util import have_pandas if have_pandas: import pandas # with named columns di5 = DesignInfo.from_array(pandas.DataFrame([[1, 2]], columns=["a", "b"])) assert di5.column_names == ["a", "b"] # with irregularly numbered columns di6 = DesignInfo.from_array(pandas.DataFrame([[1, 2]], columns=[0, 10])) assert di6.column_names == ["column0", "column10"] # with .design_info attr df = pandas.DataFrame([[1, 2]]) df.design_info = di6 assert DesignInfo.from_array(df) is di6 def test_DesignInfo_linear_constraint(): di = DesignInfo(["a1", "a2", "a3", "b"]) con = di.linear_constraint(["2 * a1 = b + 1", "a3"]) assert con.variable_names == ["a1", "a2", "a3", "b"] assert np.all(con.coefs == [[2, 0, 0, -1], [0, 0, 1, 0]]) assert np.all(con.constants == [[1], [0]]) def test_DesignInfo_deprecated_attributes(): d = DesignInfo(["a1", "a2"]) def check(attr): with warnings.catch_warnings(record=True) as w: warnings.simplefilter("always") assert getattr(d, attr) is d assert len(w) == 1 assert w[0].category is DeprecationWarning check("builder") check("design_info") # Idea: format with a reasonable amount of precision, then if that turns out # to be higher than necessary, remove as many zeros as we can. But only do # this while we can do it to *all* the ordinarily-formatted numbers, to keep # decimal points aligned. def _format_float_column(precision, col): format_str = "%." + str(precision) + "f" assert col.ndim == 1 # We don't want to look at numbers like "1e-5" or "nan" when stripping. simple_float_chars = set("+-0123456789.") col_strs = np.array([format_str % (x,) for x in col], dtype=object) # Really every item should have a decimal, but just in case, we don't want # to strip zeros off the end of "10" or something like that. mask = np.array([simple_float_chars.issuperset(col_str) and "." in col_str for col_str in col_strs]) mask_idxes = np.nonzero(mask)[0] strip_char = "0" if np.any(mask): while True: if np.all([s.endswith(strip_char) for s in col_strs[mask]]): for idx in mask_idxes: col_strs[idx] = col_strs[idx][:-1] else: if strip_char == "0": strip_char = "." else: break return col_strs def test__format_float_column(): def t(precision, numbers, expected): got = _format_float_column(precision, np.asarray(numbers)) print(got, expected) assert np.array_equal(got, expected) # This acts weird on old python versions (e.g. it can be "-nan"), so don't # hardcode it: nan_string = "%.3f" % (np.nan,) t(3, [1, 2.1234, 2.1239, np.nan], ["1.000", "2.123", "2.124", nan_string]) t(3, [1, 2, 3, np.nan], ["1", "2", "3", nan_string]) t(3, [1.0001, 2, 3, np.nan], ["1", "2", "3", nan_string]) t(4, [1.0001, 2, 3, np.nan], ["1.0001", "2.0000", "3.0000", nan_string]) # http://docs.scipy.org/doc/numpy/user/basics.subclassing.html#slightly-more-realistic-example-attribute-added-to-existing-array class DesignMatrix(np.ndarray): """A simple numpy array subclass that carries design matrix metadata. .. attribute:: design_info A :class:`DesignInfo` object containing metadata about this design matrix. This class also defines a fancy __repr__ method with labeled columns. Otherwise it is identical to a regular numpy ndarray. .. warning:: You should never check for this class using :func:`isinstance`. Limitations of the numpy API mean that it is impossible to prevent the creation of numpy arrays that have type DesignMatrix, but that are not actually design matrices (and such objects will behave like regular ndarrays in every way). Instead, check for the presence of a ``.design_info`` attribute -- this will be present only on "real" DesignMatrix objects. """ def __new__(cls, input_array, design_info=None, default_column_prefix="column"): """Create a DesignMatrix, or cast an existing matrix to a DesignMatrix. A call like:: DesignMatrix(my_array) will convert an arbitrary array_like object into a DesignMatrix. The return from this function is guaranteed to be a two-dimensional ndarray with a real-valued floating point dtype, and a ``.design_info`` attribute which matches its shape. If the `design_info` argument is not given, then one is created via :meth:`DesignInfo.from_array` using the given `default_column_prefix`. Depending on the input array, it is possible this will pass through its input unchanged, or create a view. """ # Pass through existing DesignMatrixes. The design_info check is # necessary because numpy is sort of annoying and cannot be stopped # from turning non-design-matrix arrays into DesignMatrix # instances. (E.g., my_dm.diagonal() will return a DesignMatrix # object, but one without a design_info attribute.) if (isinstance(input_array, DesignMatrix) and hasattr(input_array, "design_info")): return input_array self = atleast_2d_column_default(input_array).view(cls) # Upcast integer to floating point if safe_issubdtype(self.dtype, np.integer): self = np.asarray(self, dtype=float).view(cls) if self.ndim > 2: raise ValueError("DesignMatrix must be 2d") assert self.ndim == 2 if design_info is None: design_info = DesignInfo.from_array(self, default_column_prefix) if len(design_info.column_names) != self.shape[1]: raise ValueError("wrong number of column names for design matrix " "(got %s, wanted %s)" % (len(design_info.column_names), self.shape[1])) self.design_info = design_info if not safe_issubdtype(self.dtype, np.floating): raise ValueError("design matrix must be real-valued floating point") return self __repr__ = repr_pretty_delegate def _repr_pretty_(self, p, cycle): if not hasattr(self, "design_info"): # Not a real DesignMatrix p.pretty(np.asarray(self)) return assert not cycle # XX: could try calculating width of the current terminal window: # http://stackoverflow.com/questions/566746/how-to-get-console-window-width-in-python # sadly it looks like ipython does not actually pass this information # in, even if we use _repr_pretty_ -- the pretty-printer object has a # fixed width it always uses. (As of IPython 0.12.) MAX_TOTAL_WIDTH = 78 SEP = 2 INDENT = 2 MAX_ROWS = 30 PRECISION = 5 names = self.design_info.column_names column_name_widths = [len(name) for name in names] min_total_width = (INDENT + SEP * (self.shape[1] - 1) + np.sum(column_name_widths)) if min_total_width <= MAX_TOTAL_WIDTH: printable_part = np.asarray(self)[:MAX_ROWS, :] formatted_cols = [_format_float_column(PRECISION, printable_part[:, i]) for i in range(self.shape[1])] def max_width(col): assert col.ndim == 1 if not col.shape[0]: return 0 else: return max([len(s) for s in col]) column_num_widths = [max_width(col) for col in formatted_cols] column_widths = [max(name_width, num_width) for (name_width, num_width) in zip(column_name_widths, column_num_widths)] total_width = (INDENT + SEP * (self.shape[1] - 1) + np.sum(column_widths)) print_numbers = (total_width < MAX_TOTAL_WIDTH) else: print_numbers = False p.begin_group(INDENT, "DesignMatrix with shape %s" % (self.shape,)) p.breakable("\n" + " " * p.indentation) if print_numbers: # We can fit the numbers on the screen sep = " " * SEP # list() is for Py3 compatibility for row in [names] + list(zip(*formatted_cols)): cells = [cell.rjust(width) for (width, cell) in zip(column_widths, row)] p.text(sep.join(cells)) p.text("\n" + " " * p.indentation) if MAX_ROWS < self.shape[0]: p.text("[%s rows omitted]" % (self.shape[0] - MAX_ROWS,)) p.text("\n" + " " * p.indentation) else: p.begin_group(2, "Columns:") p.breakable("\n" + " " * p.indentation) p.pretty(names) p.end_group(2, "") p.breakable("\n" + " " * p.indentation) p.begin_group(2, "Terms:") p.breakable("\n" + " " * p.indentation) for term_name, span in six.iteritems(self.design_info.term_name_slices): if span.start != 0: p.breakable(", ") p.pretty(term_name) if span.stop - span.start == 1: coltext = "column %s" % (span.start,) else: coltext = "columns %s:%s" % (span.start, span.stop) p.text(" (%s)" % (coltext,)) p.end_group(2, "") if not print_numbers or self.shape[0] > MAX_ROWS: # some data was not shown p.breakable("\n" + " " * p.indentation) p.text("(to view full data, use np.asarray(this_obj))") p.end_group(INDENT, "") # No __array_finalize__ method, because we don't want slices of this # object to keep the design_info (they may have different columns!), or # anything fancy like that. __reduce__ = no_pickling def test_design_matrix(): import pytest di = DesignInfo(["a1", "a2", "a3", "b"]) mm = DesignMatrix([[12, 14, 16, 18]], di) assert mm.design_info.column_names == ["a1", "a2", "a3", "b"] bad_di = DesignInfo(["a1"]) pytest.raises(ValueError, DesignMatrix, [[12, 14, 16, 18]], bad_di) mm2 = DesignMatrix([[12, 14, 16, 18]]) assert mm2.design_info.column_names == ["column0", "column1", "column2", "column3"] mm3 = DesignMatrix([12, 14, 16, 18]) assert mm3.shape == (4, 1) # DesignMatrix always has exactly 2 dimensions pytest.raises(ValueError, DesignMatrix, [[[1]]]) # DesignMatrix constructor passes through existing DesignMatrixes mm4 = DesignMatrix(mm) assert mm4 is mm # But not if they are really slices: mm5 = DesignMatrix(mm.diagonal()) assert mm5 is not mm mm6 = DesignMatrix([[12, 14, 16, 18]], default_column_prefix="x") assert mm6.design_info.column_names == ["x0", "x1", "x2", "x3"] assert_no_pickling(mm6) # Only real-valued matrices can be DesignMatrixs pytest.raises(ValueError, DesignMatrix, [1, 2, 3j]) pytest.raises(ValueError, DesignMatrix, ["a", "b", "c"]) pytest.raises(ValueError, DesignMatrix, [1, 2, object()]) # Just smoke tests repr(mm) repr(DesignMatrix(np.arange(100))) repr(DesignMatrix(np.arange(100) * 2.0)) repr(mm[1:, :]) repr(DesignMatrix(np.arange(100).reshape((1, 100)))) repr(DesignMatrix([np.nan, np.inf])) repr(DesignMatrix([np.nan, 0, 1e20, 20.5])) # handling of zero-size matrices repr(DesignMatrix(np.zeros((1, 0)))) repr(DesignMatrix(np.zeros((0, 1)))) repr(DesignMatrix(np.zeros((0, 0))))