Skip to content

ENH store per-transformer index into feature space in FeatureUnion #1952

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
wants to merge 1 commit into from
Closed

ENH store per-transformer index into feature space in FeatureUnion #1952

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 5 additions & 3 deletions doc/modules/pipeline.rst
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -109,9 +109,8 @@ convenience and joint parameter estimation and validation.
create complex models.

(A :class:`FeatureUnion` has no way of checking whether two transformers
might produce identical features. It only produces a union when the
feature sets are disjoint, and making sure they are is the caller's
responsibility.)
might produce identical features. Making sure the features are disjoint
is the caller's responsibility.)


Usage
Expand All @@ -135,6 +134,9 @@ and ``value`` is an estimator object::
n_components=None, remove_zero_eig=False, tol=0))],
transformer_weights=None)

After ``fit_transform`` is called, ``FeatureUnion`` will store a `feature_ptr_`
attribute indicating which slices of the transfiormed matrix's features
correspond to which constituent transformers.


.. topic:: Examples:
Expand Down
9 changes: 9 additions & 0 deletions sklearn/pipeline.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -252,6 +252,13 @@ class FeatureUnion(BaseEstimator, TransformerMixin):
Multiplicative weights for features per transformer.
Keys are transformer names, values the weights.

Attributes
----------
feature_ptr_: array of shape (len(transformers) + 1)
Stores the feature slice corresponding to each transformer.
Transformer `i` generates feature columns `k` where
`feature_ptr_[i] <= k < feature_ptr_[i + 1]`.
Only available if `fit_transform` is used.
"""
def __init__(self, transformer_list, n_jobs=1, transformer_weights=None):
self.transformer_list = transformer_list
Expand Down Expand Up @@ -307,6 +314,8 @@ def fit_transform(self, X, y=None, **fit_params):
delayed(_fit_transform_one)(trans, name, X, y,
self.transformer_weights, **fit_params)
for name, trans in self.transformer_list)
print(len(Xs), len(self.transformer_list), [f.shape for f in Xs])
self.feature_ptr_ = np.hstack([0, np.cumsum([f.shape[1] for f in Xs])])
if any(sparse.issparse(f) for f in Xs):
Xs = sparse.hstack(Xs).tocsr()
else:
Expand Down
22 changes: 18 additions & 4 deletions sklearn/tests/test_pipeline.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -196,25 +196,39 @@ def test_feature_union():
assert_equal(X_transformed.shape, (X.shape[0], 3))

# check if it does the expected thing
assert_array_almost_equal(X_transformed[:, :-1], pca.fit_transform(X))
assert_array_equal(X_transformed[:, -1],
select.fit_transform(X, y).ravel())
pca_expected = pca.fit_transform(X)
select_expected = select.fit_transform(X, y)
assert_array_almost_equal(X_transformed[:, :-1], pca_expected)
assert_array_equal(X_transformed[:, -1:], select_expected)

# use fit_transform as an alternative, providing feature_ptr_
X_transformed = fs.fit_transform(X, y)
assert_equal(len(fs.feature_ptr_), len(fs.transformer_list) + 1)
pca_slice = slice(*fs.feature_ptr_[0:0 + 2])
select_slice = slice(*fs.feature_ptr_[1:1 + 2])
print(fs.feature_ptr_, pca_slice, select_slice)
assert_array_almost_equal(X_transformed[:, pca_slice], pca_expected)
assert_array_equal(X_transformed[:, select_slice], select_expected)

# test if it also works for sparse input
# We use a different pca object to control the random_state stream
fs = FeatureUnion([("pca", pca), ("select", select)])
X_sp = sparse.csr_matrix(X)
X_sp_transformed = fs.fit_transform(X_sp, y)
assert_array_almost_equal(X_transformed, X_sp_transformed.toarray())
assert_array_almost_equal(X_transformed,
X_sp_transformed.tocsr().todense())

# test setting parameters
fs.set_params(select__k=2)
assert_equal(fs.fit_transform(X, y).shape, (X.shape[0], 4))
assert_equal(2, fs.feature_ptr_[2] - fs.feature_ptr_[1])

# test it works with transformers missing fit_transform
fs = FeatureUnion([("mock", TransfT()), ("pca", pca), ("select", select)])
X_transformed = fs.fit_transform(X, y)
assert_equal(X_transformed.shape, (X.shape[0], 8))
assert_array_equal(fs.feature_ptr_, [0, 4, 6, 8])
assert_equal(len(fs.feature_ptr_), len(fs.transformer_list) + 1)


def test_pipeline_transform():
Expand Down