ENH Adds FeatureHasher support to pypy (#23023)

doc/conftest.py

@@ -126,8 +126,6 @@ def pytest_runtest_setup(item):
         setup_working_with_text_data()
     elif fname.endswith("modules/compose.rst") or is_index:
         setup_compose()
-    elif IS_PYPY and fname.endswith("modules/feature_extraction.rst"):
-        raise SkipTest("FeatureHasher is not compatible with PyPy")
     elif fname.endswith("datasets/loading_other_datasets.rst"):
         setup_loading_other_datasets()
     elif fname.endswith("modules/impute.rst"):

doc/whats_new/v1.1.rst

@@ -474,6 +474,9 @@ Changelog
 :mod:`sklearn.feature_extraction`
 .................................
 
+- |Feature| :class:`feature_extraction.FeatureHasher` now supports PyPy.
+  :pr:`23023` by `Thomas Fan`_.
+
 - |Fix| :class:`feature_extraction.FeatureHasher` now validates input parameters
   in `transform` instead of `__init__`. :pr:`21573` by
   :user:`Hannah Bohle <hhnnhh>` and :user:`Maren Westermann <marenwestermann>`.

sklearn/conftest.py

@@ -118,17 +118,8 @@ def pytest_collection_modifyitems(config, items):
             dataset_fetchers[name]()
 
     for item in items:
-        # FeatureHasher is not compatible with PyPy
-        if (
-            item.name.endswith(("_hash.FeatureHasher", "text.HashingVectorizer"))
-            and platform.python_implementation() == "PyPy"
-        ):
-            marker = pytest.mark.skip(
-                reason="FeatureHasher is not compatible with PyPy"
-            )
-            item.add_marker(marker)
         # Known failure on with GradientBoostingClassifier on ARM64
-        elif (
+        if (
             item.name.endswith("GradientBoostingClassifier")
             and platform.machine() == "aarch64"
         ):

sklearn/feature_extraction/_hash.py

@@ -6,19 +6,8 @@
 import numpy as np
 import scipy.sparse as sp
 
-from ..utils import IS_PYPY
 from ..base import BaseEstimator, TransformerMixin
-
-if not IS_PYPY:
-    from ._hashing_fast import transform as _hashing_transform
-else:
-
-    def _hashing_transform(*args, **kwargs):
-        raise NotImplementedError(
-            "FeatureHasher is not compatible with PyPy (see "
-            "https://github.com/scikit-learn/scikit-learn/issues/11540 "
-            "for the status updates)."
-        )
+from ._hashing_fast import transform as _hashing_transform
 
 
 def _iteritems(d):

sklearn/feature_extraction/_hashing_fast.pyx

@@ -3,13 +3,15 @@
 
 import sys
 import array
-from cpython cimport array
 cimport cython
 from libc.stdlib cimport abs
+from libcpp.vector cimport vector
+
 cimport numpy as np
 import numpy as np
-
+from ..utils._typedefs cimport INT32TYPE_t, INT64TYPE_t
 from ..utils.murmurhash cimport murmurhash3_bytes_s32
+from ..utils._vector_sentinel cimport vector_to_nd_array
 
 np.import_array()
 
@@ -25,19 +27,12 @@ def transform(raw_X, Py_ssize_t n_features, dtype,
         For constructing a scipy.sparse.csr_matrix.
 
     """
-    assert n_features > 0
-
-    cdef np.int32_t h
+    cdef INT32TYPE_t h
     cdef double value
 
-    cdef array.array indices
-    cdef array.array indptr
-    indices = array.array("i")
-    indices_array_dtype = "q"
-    indices_np_dtype = np.longlong
-
-
-    indptr = array.array(indices_array_dtype, [0])
+    cdef vector[INT32TYPE_t] indices
+    cdef vector[INT64TYPE_t] indptr
+    indptr.push_back(0)
 
     # Since Python array does not understand Numpy dtypes, we grow the indices
     # and values arrays ourselves. Use a Py_ssize_t capacity for safety.
@@ -65,13 +60,12 @@ def transform(raw_X, Py_ssize_t n_features, dtype,
 
             h = murmurhash3_bytes_s32(<bytes>f, seed)
 
-            array.resize_smart(indices, len(indices) + 1)
             if h == - 2147483648:
                 # abs(-2**31) is undefined behavior because h is a `np.int32`
                 # The following is defined such that it is equal to: abs(-2**31) % n_features
-                indices[len(indices) - 1] = (2147483647 - (n_features - 1)) % n_features
+                indices.push_back((2147483647 - (n_features - 1)) % n_features)
             else:
-                indices[len(indices) - 1] = abs(h) % n_features
+                indices.push_back(abs(h) % n_features)
             # improve inner product preservation in the hashed space
             if alternate_sign:
                 value *= (h >= 0) * 2 - 1
@@ -84,16 +78,15 @@ def transform(raw_X, Py_ssize_t n_features, dtype,
                 # references to the arrays due to Cython's error checking
                 values = np.resize(values, capacity)
 
-        array.resize_smart(indptr, len(indptr) + 1)
-        indptr[len(indptr) - 1] = size
+        indptr.push_back(size)
 
-    indices_a = np.frombuffer(indices, dtype=np.int32)
-    indptr_a = np.frombuffer(indptr, dtype=indices_np_dtype)
+    indicies_array = vector_to_nd_array(&indices)
+    indptr_array = vector_to_nd_array(&indptr)
 
-    if indptr[len(indptr) - 1] > np.iinfo(np.int32).max:  # = 2**31 - 1
+    if indptr_array[indptr_array.shape[0]-1] > np.iinfo(np.int32).max:  # = 2**31 - 1
         # both indices and indptr have the same dtype in CSR arrays
-        indices_a = indices_a.astype(np.int64, copy=False)
+        indicies_array = indicies_array.astype(np.int64, copy=False)
     else:
-        indptr_a = indptr_a.astype(np.int32, copy=False)
+        indptr_array = indptr_array.astype(np.int32, copy=False)
 
-    return (indices_a, indptr_a, values[:size])
+    return (indicies_array, indptr_array, values[:size])

sklearn/feature_extraction/setup.py

@@ -1,5 +1,4 @@
 import os
-import platform
 
 
 def configuration(parent_package="", top_path=None):
@@ -11,13 +10,13 @@ def configuration(parent_package="", top_path=None):
     if os.name == "posix":
         libraries.append("m")
 
-    if platform.python_implementation() != "PyPy":
-        config.add_extension(
-            "_hashing_fast",
-            sources=["_hashing_fast.pyx"],
-            include_dirs=[numpy.get_include()],
-            libraries=libraries,
-        )
+    config.add_extension(
+        "_hashing_fast",
+        sources=["_hashing_fast.pyx"],
+        include_dirs=[numpy.get_include()],
+        language="c++",
+        libraries=libraries,
+    )
     config.add_subpackage("tests")
 
     return config

sklearn/feature_extraction/tests/test_feature_hasher.py

@@ -3,9 +3,7 @@
 import pytest
 
 from sklearn.feature_extraction import FeatureHasher
-from sklearn.utils._testing import fails_if_pypy
-
-pytestmark = fails_if_pypy
+from sklearn.feature_extraction._hashing_fast import transform as _hashing_transform
 
 
 def test_feature_hasher_dicts():
@@ -47,9 +45,6 @@ def test_feature_hasher_strings():
 
 def test_hashing_transform_seed():
     # check the influence of the seed when computing the hashes
-    # import is here to avoid importing on pypy
-    from sklearn.feature_extraction._hashing_fast import transform as _hashing_transform
-
     raw_X = [
         ["foo", "bar", "baz", "foo".encode("ascii")],
         ["bar".encode("ascii"), "baz", "quux"],

sklearn/utils/__init__.py

@@ -1210,16 +1210,6 @@ def is_abstract(c):
             classes = [
                 (name, est_cls) for name, est_cls in classes if not name.startswith("_")
             ]
-
-            # TODO: Remove when FeatureHasher is implemented in PYPY
-            # Skips FeatureHasher for PYPY
-            if IS_PYPY and "feature_extraction" in modname:
-                classes = [
-                    (name, est_cls)
-                    for name, est_cls in classes
-                    if name == "FeatureHasher"
-                ]
-
             all_classes.extend(classes)
 
     all_classes = set(all_classes)

sklearn/utils/_typedefs.pxd

@@ -7,7 +7,11 @@ ctypedef np.float64_t DTYPE_t  # WARNING: should match DTYPE in typedefs.pyx
 cdef enum:
     DTYPECODE = np.NPY_FLOAT64
     ITYPECODE = np.NPY_INTP
+    INT32TYPECODE = np.NPY_INT32
+    INT64TYPECODE = np.NPY_INT64
 
 # Index/integer type.
 #  WARNING: ITYPE_t must be a signed integer type or you will have a bad time!
 ctypedef np.intp_t ITYPE_t  # WARNING: should match ITYPE in typedefs.pyx
+ctypedef np.int32_t INT32TYPE_t  # WARNING: should match INT32TYPE in typedefs.pyx
+ctypedef np.int64_t INT64TYPE_t  # WARNING: should match INT32TYPE in typedefs.pyx

sklearn/utils/_typedefs.pyx

@@ -14,6 +14,8 @@ np.import_array()
 #cdef ITYPE_t[:] idummy_view = <ITYPE_t[:1]> &idummy
 #ITYPE = np.asarray(idummy_view).dtype
 ITYPE = np.intp  # WARNING: this should match ITYPE_t in typedefs.pxd
+INT32TYPE = np.int32 # WARNING: should match INT32TYPE_t in typedefs.pyx
+INT64TYPE = np.int64 # WARNING: this should match INT64TYPE_t in typedefs.pxd
 
 #cdef DTYPE_t ddummy
 #cdef DTYPE_t[:] ddummy_view = <DTYPE_t[:1]> &ddummy

sklearn/utils/_vector_sentinel.pxd

@@ -1,10 +1,12 @@
 cimport numpy as np
 
 from libcpp.vector cimport vector
-from ..utils._typedefs cimport ITYPE_t, DTYPE_t
+from ..utils._typedefs cimport ITYPE_t, DTYPE_t, INT32TYPE_t, INT64TYPE_t
 
 ctypedef fused vector_typed:
     vector[DTYPE_t]
     vector[ITYPE_t]
+    vector[INT32TYPE_t]
+    vector[INT64TYPE_t]
 
 cdef np.ndarray vector_to_nd_array(vector_typed * vect_ptr)

sklearn/utils/_vector_sentinel.pyx

@@ -2,14 +2,18 @@ from cython.operator cimport dereference as deref
 from cpython.ref cimport Py_INCREF
 cimport numpy as np
 
-from ._typedefs cimport DTYPECODE, ITYPECODE
+from ._typedefs cimport DTYPECODE, ITYPECODE, INT32TYPECODE, INT64TYPECODE
 
 np.import_array()
 
 
 cdef StdVectorSentinel _create_sentinel(vector_typed * vect_ptr):
     if vector_typed is vector[DTYPE_t]:
         return StdVectorSentinelFloat64.create_for(vect_ptr)
+    elif vector_typed is vector[INT32TYPE_t]:
+        return StdVectorSentinelInt32.create_for(vect_ptr)
+    elif vector_typed is vector[INT64TYPE_t]:
+        return StdVectorSentinelInt64.create_for(vect_ptr)
     else:
         return StdVectorSentinelIntP.create_for(vect_ptr)
 
@@ -64,6 +68,42 @@ cdef class StdVectorSentinelIntP(StdVectorSentinel):
         return ITYPECODE
 
 
+cdef class StdVectorSentinelInt32(StdVectorSentinel):
+    cdef vector[INT32TYPE_t] vec
+
+    @staticmethod
+    cdef StdVectorSentinel create_for(vector[INT32TYPE_t] * vect_ptr):
+        # This initializes the object directly without calling __init__
+        # See: https://cython.readthedocs.io/en/latest/src/userguide/extension_types.html#instantiation-from-existing-c-c-pointers # noqa
+        cdef StdVectorSentinelInt32 sentinel = StdVectorSentinelInt32.__new__(StdVectorSentinelInt32)
+        sentinel.vec.swap(deref(vect_ptr))
+        return sentinel
+
+    cdef void* get_data(self):
+        return self.vec.data()
+
+    cdef int get_typenum(self):
+        return INT32TYPECODE
+
+
+cdef class StdVectorSentinelInt64(StdVectorSentinel):
+    cdef vector[INT64TYPE_t] vec
+
+    @staticmethod
+    cdef StdVectorSentinel create_for(vector[INT64TYPE_t] * vect_ptr):
+        # This initializes the object directly without calling __init__
+        # See: https://cython.readthedocs.io/en/latest/src/userguide/extension_types.html#instantiation-from-existing-c-c-pointers # noqa
+        cdef StdVectorSentinelInt64 sentinel = StdVectorSentinelInt64.__new__(StdVectorSentinelInt64)
+        sentinel.vec.swap(deref(vect_ptr))
+        return sentinel
+
+    cdef void* get_data(self):
+        return self.vec.data()
+
+    cdef int get_typenum(self):
+        return INT64TYPECODE
+
+
 cdef np.ndarray vector_to_nd_array(vector_typed * vect_ptr):
     cdef:
         np.npy_intp size = deref(vect_ptr).size()