ENH - Gets SciKeras script working

docs/changes.rst

@@ -18,6 +18,8 @@ v0.9
   :pr:`386` by :user:`Reid Johnson <reidjohnson>`.
 - :func:`skops.hub_utils.get_model_output` and :func:`skops.hub_utils.push` are
   deprecated and will be removed in version 0.10. :pr:`396` by `Adrin Jalali`_.
+- Fix dumping Scikeras model failing because of maximum recursion depth. :pr:`388`
+  by :user:`Thomas Lazarus <lazarust>`.
 
 v0.8
 ----

skops/_min_dependencies.py

@@ -36,6 +36,7 @@
     "catboost": ("1.0", "tests", None),
     "fairlearn": ("0.7.0", "docs, tests", None),
     "rich": ("12", "tests, rich", None),
+    "scikeras": ("0.4.0", "tests", None),
 }
 
 

skops/io/_general.py

@@ -4,6 +4,7 @@
 import json
 import operator
 import uuid
+import weakref
 from functools import partial
 from reprlib import Repr
 from types import FunctionType, MethodType
@@ -46,10 +47,12 @@ def dict_get_state(obj: Any, save_context: SaveContext) -> dict[str, Any]:
     for key, value in obj.items():
         if isinstance(value, property):
             continue
+        if isinstance(key, weakref.ref):
+            key = getattr(key(), "_name")
         if np.isscalar(key) and hasattr(key, "item"):
             # convert numpy value to python object
             key = key.item()  # type: ignore
-        content[key] = get_state(value, save_context)
+        content[str(key)] = get_state(value, save_context)
     res["content"] = content
     res["key_types"] = key_types
     return res

skops/io/_utils.py

@@ -157,14 +157,13 @@ def get_state(value, save_context: SaveContext) -> dict[str, Any]:
     # fails with `get_state`, we try with json.dumps, if that fails, we raise
     # the original error alongside the json error.
 
-    # TODO: This should help with fixing recursive references.
-    # if id(value) in save_context.memo:
-    #     return {
-    #         "__module__": None,
-    #         "__class__": None,
-    #         "__id__": id(value),
-    #         "__loader__": "CachedNode",
-    #     }
+    if id(value) in save_context.memo:
+        return {
+            "__module__": None,
+            "__class__": None,
+            "__id__": id(value),
+            "__loader__": "CachedNode",
+        }
 
     __id__ = save_context.memoize(obj=value)
 

skops/io/tests/test_external.py

@@ -16,6 +16,7 @@
 
 import pytest
 from sklearn.datasets import make_classification, make_regression
+from sklearn.pipeline import Pipeline
 
 from skops.io import dumps, loads, visualize
 from skops.io.tests._utils import assert_method_outputs_equal, assert_params_equal
@@ -427,3 +428,56 @@ def test_quantile_forest(self, quantile_forest, regr_data, trusted, tree_method)
         assert_method_outputs_equal(estimator, loaded, X)
 
         visualize(dumped, trusted=trusted)
+
+
+class TestSciKeras:
+    """Tests for SciKerasRegressor and SciKerasClassifier"""
+
+    @pytest.fixture(autouse=True)
+    def capture_stdout(self):
+        # Mock print and rich.print so that running these tests with pytest -s
+        # does not spam stdout. Other, more common methods of suppressing
+        # printing to stdout don't seem to work, perhaps because of pytest.
+        with patch("builtins.print", Mock()), patch("rich.print", Mock()):
+            yield
+
+    @pytest.fixture(autouse=True)
+    def keras(self):
+        scikeras = pytest.importorskip("scikeras")
+        return scikeras
+
+    @pytest.fixture
+    def trusted(self):
+        return [
+            "scikeras.wrappers.KerasClassifier",
+            "keras.models.Sequential",
+            "keras.layers.core.Dense",
+            "keras.layers.core.Input",
+        ]
+
+    @pytest.fixture
+    def test_dumping_model(self, keras, trusted):
+        from scikeras.wrappers import KerasClassifier
+
+        # This simplifies the basic usage tutorial from https://adriangb.com/scikeras/stable/notebooks/Basic_Usage.html
+
+        def get_clf(meta):
+            n_features_in_ = meta["n_features_in_"]
+            model = keras.models.Sequential()
+            model.add(keras.layers.Input(shape=(n_features_in_,)))
+            model.add(keras.layers.Dense(1, activation="sigmoid"))
+            return model
+
+        clf = KerasClassifier(model=get_clf, loss="binary_crossentropy")
+
+        pipeline = Pipeline([("classifier", clf)])
+
+        dumps(clf, "keras-test.skops")
+        dumps(pipeline, "keras-test.skops")
+
+        X, y = make_classification(1000, 20, n_informative=10, random_state=0)
+        clf.fit(X, y)
+        dumped = dumps(clf, "keras-test.skops")
+
+        loaded = loads(dumped, trusted=trusted)
+        assert_method_outputs_equal(clf, loaded, X)

skops/io/tests/test_persist.py

@@ -605,6 +605,11 @@ def fit(self, X, y=None, **fit_params):
     for key, val_expected in expected.items():
         for state in states:
             val_state = state[key]
+
+            # skipping all the state values that are a cached node
+            if val_state["__loader__"] == "CachedNode":
+                continue
+
             # check presence of "content"/"file" but not exact values
             assert ("content" in val_state) or ("file" in val_state)
             assert val_state["__class__"] == val_expected["__class__"]