MNT: Update repo URL of sklearn nightly build, fix sparse matrix persistence issue

.github/workflows/build-test.yml

@@ -63,7 +63,7 @@ jobs:
         pip install black=="22.6.0" isort=="5.10.1" mypy=="1.0.0"
         pip uninstall --yes scikit-learn
         if [ ${{ matrix.sklearn_version }} == "nightly" ];
-          then pip install --pre --extra-index https://pypi.anaconda.org/scipy-wheels-nightly/simple scikit-learn;
+          then pip install --pre --extra-index https://pypi.anaconda.org/scientific-python-nightly-wheels/simple scikit-learn;
           else pip install "scikit-learn~=${{ matrix.sklearn_version }}";
         fi
         if [ ${{ matrix.os }} == "ubuntu-latest" ];

skops/io/tests/_utils.py

@@ -133,10 +133,25 @@ def _assert_vals_equal(val1, val2):
         _assert_generic_objects_equal(val1, val2)
 
 
+def _clean_params(params):
+    # this function deals with cleaning special parameters that for one reason
+    # or another should be removed or modified.
+    params = params.copy()
+
+    # see #375
+    keys_to_remove = ["_has_canonical_format", "_has_sorted_indices"]
+    for key in keys_to_remove:
+        params.pop(key, None)
+
+    return params
+
+
 def assert_params_equal(params1, params2):
     # helper function to compare estimator dictionaries of parameters
     if params1 is None and params2 is None:
         return
+
+    params1, params2 = _clean_params(params1), _clean_params(params2)
     assert len(params1) == len(params2)
     assert set(params1.keys()) == set(params2.keys())
     for key in params1:

skops/io/tests/test_persist.py

@@ -1047,3 +1047,21 @@ def test_compression_level():
     dumped_compressed = dumps(model, compression=ZIP_DEFLATED, compresslevel=9)
     # This reduces the size substantially
     assert len(dumped_raw) > 5 * len(dumped_compressed)
+
+
+@pytest.mark.parametrize("call_has_canonical_format", [False, True])
+def test_sparse_matrix(call_has_canonical_format):
+    # see https://github.com/skops-dev/skops/pull/375
+
+    # note: this behavior is already implicitly tested by sklearn estimators
+    # that use sparse matrices under the hood (tfidf) but it is better to check
+    # the behavior explicitly
+    x = sparse.csr_matrix((3, 4))
+    if call_has_canonical_format:
+        x.has_canonical_format
+
+    dumped = dumps(x)
+    untrusted_types = get_untrusted_types(data=dumped)
+    y = loads(dumped, trusted=untrusted_types)
+
+    assert_params_equal(x.__dict__, y.__dict__)