def __init__(self, column_names, term_slices=None, term_name_slices=None, builder=None): self.column_name_indexes = OrderedDict( zip(column_names, range(len(column_names)))) if term_slices is not None: #: An OrderedDict mapping :class:`Term` objects to Python #: func:`slice` objects. May be None, for design matrices which #: were constructed directly rather than by using the patsy #: machinery. If it is not None, then it #: is guaranteed to list the terms in order, and the slices are #: guaranteed to exactly cover all columns with no overlap or #: gaps. self.term_slices = OrderedDict(term_slices) if term_name_slices is not None: raise ValueError("specify only one of term_slices and " "term_name_slices") term_names = [term.name() for term in self.term_slices] #: And OrderedDict mapping term names (as strings) to Python #: :func:`slice` objects. Guaranteed never to be None. Guaranteed #: to list the terms in order, and the slices are #: guaranteed to exactly cover all columns with no overlap or #: gaps. Name overlap is allowed between term names and column #: names, but it is guaranteed that if it occurs, then they refer #: to exactly the same column. self.term_name_slices = OrderedDict( zip(term_names, self.term_slices.values())) else: # term_slices is None self.term_slices = None if term_name_slices is None: # Make up one term per column term_names = column_names slices = [slice(i, i + 1) for i in range(len(column_names))] term_name_slices = zip(term_names, slices) self.term_name_slices = OrderedDict(term_name_slices) self.builder = builder # 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. # If there is any name overlap between terms and columns, they refer # to the same columns. 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())) covered = 0 for slice_ in six.itervalues(self.term_name_slices): start, stop, step = slice_.indices(len(column_names)) if start != covered: raise ValueError("bad term slices") if step != 1: raise ValueError("bad term slices") covered = stop if covered != len(column_names): raise ValueError("bad term indices") 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")
def _make_subterm_infos(terms, num_column_counts, cat_levels_contrasts): # Sort each term into a bucket based on the set of numeric factors it # contains: term_buckets = OrderedDict() bucket_ordering = [] for term in terms: num_factors = [] for factor in term.factors: if factor in num_column_counts: num_factors.append(factor) bucket = frozenset(num_factors) if bucket not in term_buckets: bucket_ordering.append(bucket) term_buckets.setdefault(bucket, []).append(term) # Special rule: if there is a no-numerics bucket, then it always comes # first: if frozenset() in term_buckets: bucket_ordering.remove(frozenset()) bucket_ordering.insert(0, frozenset()) term_to_subterm_infos = OrderedDict() new_term_order = [] # Then within each bucket, work out which sort of contrasts we want to use # for each term to avoid redundancy for bucket in bucket_ordering: bucket_terms = term_buckets[bucket] # Sort by degree of interaction bucket_terms.sort(key=lambda t: len(t.factors)) new_term_order += bucket_terms used_subterms = set() for term in bucket_terms: subterm_infos = [] factor_codings = pick_contrasts_for_term(term, num_column_counts, used_subterms) # Construct one SubtermInfo for each subterm for factor_coding in factor_codings: subterm_factors = [] contrast_matrices = {} subterm_columns = 1 # In order to preserve factor ordering information, the # coding_for_term just returns dicts, and we refer to # the original factors to figure out which are included in # each subterm, and in what order for factor in term.factors: # Numeric factors are included in every subterm if factor in num_column_counts: subterm_factors.append(factor) subterm_columns *= num_column_counts[factor] elif factor in factor_coding: subterm_factors.append(factor) levels, contrast = cat_levels_contrasts[factor] # This is where the default coding is set to # Treatment: coded = code_contrast_matrix(factor_coding[factor], levels, contrast, default=Treatment) contrast_matrices[factor] = coded subterm_columns *= coded.matrix.shape[1] subterm_infos.append(SubtermInfo(subterm_factors, contrast_matrices, subterm_columns)) term_to_subterm_infos[term] = subterm_infos assert new_term_order == list(term_to_subterm_infos) return term_to_subterm_infos
def test_linear_constraint(): from nose.tools import assert_raises from patsy.compat import OrderedDict t = _check_lincon t(LinearConstraint(["a", "b"], [2, 3]), ["a", "b"], [[2, 3]], [[0]]) assert_raises(ValueError, linear_constraint, LinearConstraint(["b", "a"], [2, 3]), ["a", "b"]) t({"a": 2}, ["a", "b"], [[1, 0]], [[2]]) t(OrderedDict([("a", 2), ("b", 3)]), ["a", "b"], [[1, 0], [0, 1]], [[2], [3]]) t(OrderedDict([("a", 2), ("b", 3)]), ["b", "a"], [[0, 1], [1, 0]], [[2], [3]]) t({0: 2}, ["a", "b"], [[1, 0]], [[2]]) t(OrderedDict([(0, 2), (1, 3)]), ["a", "b"], [[1, 0], [0, 1]], [[2], [3]]) t(OrderedDict([("a", 2), (1, 3)]), ["a", "b"], [[1, 0], [0, 1]], [[2], [3]]) assert_raises(ValueError, linear_constraint, {"q": 1}, ["a", "b"]) assert_raises(ValueError, linear_constraint, {"a": 1, 0: 2}, ["a", "b"]) t(np.array([2, 3]), ["a", "b"], [[2, 3]], [[0]]) t(np.array([[2, 3], [4, 5]]), ["a", "b"], [[2, 3], [4, 5]], [[0], [0]]) t("a = 2", ["a", "b"], [[1, 0]], [[2]]) t("a - 2", ["a", "b"], [[1, 0]], [[2]]) t("a + 1 = 3", ["a", "b"], [[1, 0]], [[2]]) t("a + b = 3", ["a", "b"], [[1, 1]], [[3]]) t("a = 2, b = 3", ["a", "b"], [[1, 0], [0, 1]], [[2], [3]]) t("b = 3, a = 2", ["a", "b"], [[0, 1], [1, 0]], [[3], [2]]) t(["a = 2", "b = 3"], ["a", "b"], [[1, 0], [0, 1]], [[2], [3]]) assert_raises(ValueError, linear_constraint, ["a", {"b": 0}], ["a", "b"]) # Actual evaluator tests t("2 * (a + b/3) + b + 2*3/4 = 1 + 2*3", ["a", "b"], [[2, 2.0/3 + 1]], [[7 - 6.0/4]]) t("+2 * -a", ["a", "b"], [[-2, 0]], [[0]]) t("a - b, a + b = 2", ["a", "b"], [[1, -1], [1, 1]], [[0], [2]]) t("a = 1, a = 2, a = 3", ["a", "b"], [[1, 0], [1, 0], [1, 0]], [[1], [2], [3]]) t("a * 2", ["a", "b"], [[2, 0]], [[0]]) t("-a = 1", ["a", "b"], [[-1, 0]], [[1]]) t("(2 + a - a) * b", ["a", "b"], [[0, 2]], [[0]]) t("a = 1 = b", ["a", "b"], [[1, 0], [0, -1]], [[1], [-1]]) t("a = (1 = b)", ["a", "b"], [[0, -1], [1, 0]], [[-1], [1]]) t("a = 1, a = b = c", ["a", "b", "c"], [[1, 0, 0], [1, -1, 0], [0, 1, -1]], [[1], [0], [0]]) # One should never do this of course, but test that it works anyway... t("a + 1 = 2", ["a", "a + 1"], [[0, 1]], [[2]]) t(([10, 20], [30]), ["a", "b"], [[10, 20]], [[30]]) t(([[10, 20], [20, 40]], [[30], [35]]), ["a", "b"], [[10, 20], [20, 40]], [[30], [35]]) # wrong-length tuple assert_raises(ValueError, linear_constraint, ([1, 0], [0], [0]), ["a", "b"]) assert_raises(ValueError, linear_constraint, ([1, 0],), ["a", "b"]) t([10, 20], ["a", "b"], [[10, 20]], [[0]]) t([[10, 20], [20, 40]], ["a", "b"], [[10, 20], [20, 40]], [[0], [0]]) t(np.array([10, 20]), ["a", "b"], [[10, 20]], [[0]]) t(np.array([[10, 20], [20, 40]]), ["a", "b"], [[10, 20], [20, 40]], [[0], [0]]) # unknown object type assert_raises(ValueError, linear_constraint, None, ["a", "b"])
def test_DesignInfo(): from nose.tools import assert_raises 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) assert_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: assert_raises(ValueError, DesignInfo, ["x1", "x2", "x3", "y"], factor_infos=factor_infos) assert_raises(ValueError, DesignInfo, ["x1", "x2", "x3", "y"], term_codings=term_codings) # factor_infos must be a dict assert_raises(ValueError, DesignInfo, ["x1", "x2", "x3", "y"], list(factor_infos), term_codings) # wrong number of column names: assert_raises(ValueError, DesignInfo, ["x1", "x2", "x3", "y1", "y2"], factor_infos, term_codings) assert_raises(ValueError, DesignInfo, ["x1", "x2", "x3"], factor_infos, term_codings) # name overlap problems assert_raises(ValueError, DesignInfo, ["x1", "x2", "y", "y2"], factor_infos, term_codings) # duplicate name assert_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] assert_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)] assert_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" assert_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) assert_raises(ValueError, DesignInfo, ["x1", "x2", "x3", "y"], mismatch_factor_infos, term_codings) # bad term_codings assert_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] assert_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]) assert_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" assert_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) assert_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)])]) assert_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)])]) assert_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. assert_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)])]) assert_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 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)
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")