ENH - Gets SciKeras script working

docs/changes.rst

@@ -14,6 +14,8 @@ v0.10
 - Removes Pythn 3.8 support and adds Python 3.12 Support :pr:`418` by :user:`Thomas Lazarus <lazarust>`.
 - Removes a shortcut to add `sklearn-intelex` as a not dependency.
   :pr:`420` by :user:`Thomas Lazarus < lazarust > `.
+- Fix dumping Scikeras model failing because of maximum recursion depth. :pr:`388`
+  by :user:`Thomas Lazarus <lazarust>`.
 
 v0.9
 ----

pyproject.toml

@@ -25,6 +25,9 @@ filterwarnings = [
     "ignore:DataFrameGroupBy.apply operated on the grouping columns.:DeprecationWarning",
     # Ignore Pandas 2.2 warning on PyArrow. It might be reverted in a later release.
     "ignore:\\s*Pyarrow will become a required dependency of pandas.*:DeprecationWarning",
+    # Ignore Google protobuf deprecation warnings. This should be reverted once tensorflow supports protobuf 5.0+
+    "ignore:Type google._upb._message.MessageMapContainer uses PyType_Spec with a metaclass that has custom tp_new. This is deprecated and will no longer be allowed in Python 3.14.:DeprecationWarning:importlib.*:",
+    "ignore:Type google._upb._message.ScalarMapContainer uses PyType_Spec with a metaclass that has custom tp_new. This is deprecated and will no longer be allowed in Python 3.14.:DeprecationWarning:importlib.*:",
 ]
 markers = [
     "network: marks tests as requiring internet (deselect with '-m \"not network\"')",

skops/_min_dependencies.py

@@ -36,6 +36,8 @@
     "catboost": ("1.0", "tests", None),
     "fairlearn": ("0.7.0", "docs, tests", None),
     "rich": ("12", "tests, rich", None),
+    "scikeras": ("0.12.0", "docs, tests", None),
+    "tensorflow": ("2.12.0", "docs, tests", None),
 }
 
 

skops/card/_model_card.py

@@ -1227,7 +1227,7 @@ def add_permutation_importances(
         _, ax = plt.subplots()
         ax.boxplot(
             x=permutation_importances.importances[sorted_importances_idx].T,
-            labels=columns[sorted_importances_idx],
+            tick_labels=columns[sorted_importances_idx],
             vert=False,
         )
         ax.set_title(plot_name)

skops/io/_persist.py

@@ -15,7 +15,14 @@
 # We load the dispatch functions from the corresponding modules and register
 # them. Old protocols are found in the 'old/' directory, with the protocol
 # version appended to the corresponding module name.
-modules = ["._general", "._numpy", "._scipy", "._sklearn", "._quantile_forest"]
+modules = [
+    "._general",
+    "._numpy",
+    "._scikeras",
+    "._scipy",
+    "._sklearn",
+    "._quantile_forest",
+]
 modules.extend([".old._general_v0", ".old._numpy_v0"])
 for module_name in modules:
     # register exposed functions for get_state and get_tree

skops/io/_scikeras.py

@@ -0,0 +1,63 @@
+from __future__ import annotations
+
+import io
+import os
+import tempfile
+from typing import Sequence, Type
+
+import tensorflow as tf
+from scikeras.wrappers import KerasClassifier, KerasRegressor
+
+from ._audit import Node
+from ._protocol import PROTOCOL
+from ._utils import Any, LoadContext, SaveContext, get_module
+
+
+def scikeras_get_state(obj: Any, save_context: SaveContext) -> dict[str, Any]:
+    res = {
+        "__class__": obj.__class__.__name__,
+        "__module__": get_module(type(obj)),
+        "__loader__": "SciKerasNode",
+    }
+
+    obj_id = save_context.memoize(obj)
+    f_name = f"{obj_id}.keras"
+
+    with tempfile.TemporaryDirectory() as temp_dir:
+        file_name = os.path.join(temp_dir, "model.keras")
+        obj.model.save(file_name)
+        save_context.zip_file.write(file_name, f_name)
+
+    res.update(type="scikeras", file=f_name)
+    return res
+
+
+class SciKerasNode(Node):
+    def __init__(
+        self,
+        state: dict[str, Any],
+        load_context: LoadContext,
+        trusted: bool | Sequence[str] = False,
+    ) -> None:
+        super().__init__(state, load_context, trusted)
+        self.trusted = self._get_trusted(trusted, default=[])
+
+        self.children = {"content": io.BytesIO(load_context.src.read(state["file"]))}
+
+    def _construct(self):
+        with tempfile.TemporaryDirectory() as temp_dir:
+            file_path = os.path.join(temp_dir, "model.keras")
+            with open(file_path, "wb") as f:
+                f.write(self.children["content"].getbuffer())
+            model = tf.keras.models.load_model(file_path, compile=False)
+        return model
+
+
+GET_STATE_DISPATCH_FUNCTIONS = [
+    (KerasClassifier, scikeras_get_state),
+    (KerasRegressor, scikeras_get_state),
+]
+
+NODE_TYPE_MAPPING: dict[tuple[str, int], Type[Node]] = {
+    ("SciKerasNode", PROTOCOL): SciKerasNode
+}

skops/io/tests/test_external.py

@@ -11,7 +11,6 @@
 with a range of hyperparameters.
 
 """
-
 from unittest.mock import Mock, patch
 
 import pytest
@@ -427,3 +426,51 @@ 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
+    def trusted(self):
+        return ["scikeras.wrappers.KerasClassifier"]
+
+    @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 tensorflow(self):
+        tensorflow = pytest.importorskip("tensorflow")
+        return tensorflow
+
+    def test_dumping_model(self, tensorflow, trusted):
+        # This simplifies the basic usage tutorial from https://adriangb.com/scikeras/stable/notebooks/Basic_Usage.html
+
+        n_features_in_ = 20
+        model = tensorflow.keras.models.Sequential()
+        model.add(tensorflow.keras.layers.Input(shape=(n_features_in_,)))
+        model.add(tensorflow.keras.layers.Dense(1, activation="sigmoid"))
+
+        from scikeras.wrappers import KerasClassifier
+
+        clf = KerasClassifier(model=model, loss="binary_crossentropy")
+
+        X, y = make_classification(1000, 20, n_informative=10, random_state=0)
+        clf.fit(X, y)
+
+        predictions = clf.predict(X)
+
+        dumped = dumps(clf)
+
+        # Loads returns the Keras model so we need to initialize it as a SciKeras model
+        new_clf_model = loads(dumped, trusted=trusted)
+
+        clf_new = KerasClassifier(new_clf_model)
+        clf_new.initialize(X, y)
+        new_preidctions = clf_new.predict(X)
+        assert all(new_preidctions == predictions)