Support for Keras 3

README.md

@@ -8,7 +8,7 @@ Scikit-Learn compatible wrappers for Keras Models.
 
 ## Why SciKeras
 
-SciKeras is derived from and API compatible with `tf.keras.wrappers.scikit_learn`. The original TensorFlow (TF) wrappers are not actively maintained,
+SciKeras is derived from and API compatible with `keras.wrappers.scikit_learn`. The original TensorFlow (TF) wrappers are not actively maintained,
 and [will be removed](https://github.com/tensorflow/tensorflow/pull/36137#issuecomment-726271760) in a future release.
 
 An overview of the advantages and differences as compared to the TF wrappers can be found in our
@@ -38,7 +38,7 @@ runtime or SciKeras will throw an error when you try to import it.
 
 The current version of SciKeras depends on `scikit-learn>=1.0.0` and `TensorFlow>=2.7.0`.
 
-### Migrating from `tf.keras.wrappers.scikit_learn`
+### Migrating from `keras.wrappers.scikit_learn`
 
 Please see the [migration](https://www.adriangb.com/scikeras/stable/migration.html) section of our documentation.
 

docs/source/advanced.rst

@@ -112,7 +112,7 @@ offer an easy way to compile and tune compilation parameters. Examples:
 
 .. code:: python
 
-    from tensorflow.keras.optimizers import Adam
+    from keras.optimizers import Adam
 
     def model_build_fn():
         model = Model(...)
@@ -280,7 +280,7 @@ Optimizer
 .. code:: python
 
     from scikeras.wrappers import KerasClassifier
-    from tensorflow import keras
+    import keras
 
     clf = KerasClassifier(
         model=model_build_fn,
@@ -305,7 +305,7 @@ Losses
 
 .. code:: python
 
-    from tensorflow.keras.losses import BinaryCrossentropy, CategoricalCrossentropy
+    from keras.losses import BinaryCrossentropy, CategoricalCrossentropy
 
     clf = KerasClassifier(
         ...,
@@ -322,7 +322,7 @@ Additionally, SciKeras supports routed parameters to each individual loss, or to
 
 .. code:: python
 
-    from tensorflow.keras.losses import BinaryCrossentropy, CategoricalCrossentropy
+    from keras.losses import BinaryCrossentropy, CategoricalCrossentropy
 
     clf = KerasClassifier(
         ...,
@@ -348,7 +348,7 @@ Here are several support use cases:
 
 .. code:: python
 
-    from tensorflow.keras.metrics import BinaryAccuracy, AUC
+    from keras.metrics import BinaryAccuracy, AUC
 
     clf = KerasClassifier(
         ...,
@@ -388,7 +388,7 @@ SciKeras can route parameters to callbacks.
 
     clf = KerasClassifier(
         ...,
-        callbacks=tf.keras.callbacks.EarlyStopping
+        callbacks=keras.callbacks.EarlyStopping
         callbacks__monitor="loss",
     )
 
@@ -399,21 +399,21 @@ Just like metrics and losses, callbacks support several syntaxes to compile them
     # for multiple callbacks using dict syntax
     clf = KerasClassifier(
         ...,
-        callbacks={"bl": tf.keras.callbacks.BaseLogger, "es": tf.keras.callbacks.EarlyStopping}
+        callbacks={"bl": keras.callbacks.BaseLogger, "es": keras.callbacks.EarlyStopping}
         callbacks__es__monitor="loss",
     )
     # or using list sytnax
     clf = KerasClassifier(
         ...,
-        callbacks=[tf.keras.callbacks.BaseLogger, tf.keras.callbacks.EarlyStopping]
+        callbacks=[keras.callbacks.BaseLogger, keras.callbacks.EarlyStopping]
         callbacks__1__monitor="loss",  # EarlyStopping(monitor="loss")
     )
 
 Keras callbacks are event based, and are triggered depending on the methods they implement.
 For example:
 
 .. code:: python
-    from tensorflow import keras
+    import keras
 
     class MyCallback(keras.callbacks.Callback):
 
@@ -433,9 +433,9 @@ simply use the ``fit__`` or ``predict__`` routing prefixes on your callback:
 
     clf = KerasClassifier(
         ...,
-        callbacks=tf.keras.callbacks.Callback,  # called from both fit and predict
-        fit__callbacks=tf.keras.callbacks.Callback,  # called only from fit
-        predict__callbacks=tf.keras.callbacks.Callback,  # called only from predict
+        callbacks=keras.callbacks.Callback,  # called from both fit and predict
+        fit__callbacks=keras.callbacks.Callback,  # called only from fit
+        predict__callbacks=keras.callbacks.Callback,  # called only from predict
     )
 
 Any routed constructor parameters must also use the corresponding prefix to get routed correctly.
@@ -449,7 +449,7 @@ which tells SciKeras to pass that parameter as an positional argument instead of
 
 .. code:: python
 
-   from tensorflow import keras
+   import keras
 
     class Schedule:
         """Exponential decay lr scheduler.

docs/source/index.rst

@@ -3,7 +3,7 @@ Welcome to SciKeras's documentation!
 
 The goal of scikeras is to make it possible to use Keras/TensorFlow with sklearn. 
 This is achieved by providing a wrapper around Keras that has an Scikit-Learn interface.
-SciKeras is the successor to ``tf.keras.wrappers.scikit_learn``, and offers many
+SciKeras is the successor to ``keras.wrappers.scikit_learn``, and offers many
 improvements over the TensorFlow version of the wrappers. See :ref:`Migration<Migration>` for a more details.
 
 SciKeras tries to make things easy for you while staying out of your way.

docs/source/migration.rst

@@ -1,7 +1,7 @@
 .. _Migration:
 
 =================================================
-Migrating from ``tf.keras.wrappers.scikit_learn``
+Migrating from ``keras.wrappers.scikit_learn``
 =================================================
 
 .. contents::
@@ -10,7 +10,7 @@ Migrating from ``tf.keras.wrappers.scikit_learn``
 
 Why switch to SciKeras
 ----------------------
-SciKeras has several advantages over ``tf.keras.wrappers.scikit_learn``:
+SciKeras has several advantages over ``keras.wrappers.scikit_learn``:
 
 * Full compatibility with the Scikit-Learn API, including grid searches, ensembles, transformers, etc.
 * Support for Functional and Subclassed Keras Models.
@@ -29,7 +29,7 @@ SciKeras is largely backwards compatible with the existing wrappers. For most ca
 
 .. code:: diff
 
-   - from tensorflow.keras.wrappers.scikit_learn import KerasClassifier, KerasRegressor
+   - from keras.wrappers.scikit_learn import KerasClassifier, KerasRegressor
    + from scikeras.wrappers import KerasClassifier, KerasRegressor
 
 

docs/source/notebooks/AutoEncoders.md

@@ -53,15 +53,15 @@ warnings.filterwarnings("ignore", message="Setting the random state for TF")
 ```python
 import numpy as np
 from scikeras.wrappers import KerasClassifier, KerasRegressor
-from tensorflow import keras
+import keras
 ```
 
 ## 2. Data
 
 We load the dataset from the Keras tutorial. The dataset consists of images of cats and dogs.
 
 ```python
-from tensorflow.keras.datasets import mnist
+from keras.datasets import mnist
 import numpy as np
 
 

docs/source/notebooks/Basic_Usage.md

@@ -66,7 +66,7 @@ warnings.filterwarnings("ignore", message="Setting the random state for TF")
 ```python
 import numpy as np
 from scikeras.wrappers import KerasClassifier, KerasRegressor
-from tensorflow import keras
+import keras
 ```
 
 ## 2. Training a classifier and making predictions
@@ -100,7 +100,7 @@ do for binary classification. The second option is usually reserved for when
 you have >2 classes.
 
 ```python
-from tensorflow import keras
+import keras
 
 
 def get_clf(meta, hidden_layer_sizes, dropout):
@@ -355,7 +355,7 @@ This is exactly the same logic that allows to access estimator parameters in `sk
 
 This feature is useful in several ways. For one, it allows to set those parameters in the model definition. Furthermore, it allows you to set parameters in an `sklearn GridSearchCV` as shown below.
 
-To differentiate paramters like `callbacks` which are accepted by both `tf.keras.Model.fit` and `tf.keras.Model.predict` you can add a `fit__` or `predict__` routing suffix respectively. Similar, the `model__` prefix may be used to specify that a paramter is destined only for `get_clf`/`get_reg` (or whatever callable you pass as your `model` argument).
+To differentiate paramters like `callbacks` which are accepted by both `kerasModel.fit` and `kerasModel.predict` you can add a `fit__` or `predict__` routing suffix respectively. Similar, the `model__` prefix may be used to specify that a paramter is destined only for `get_clf`/`get_reg` (or whatever callable you pass as your `model` argument).
 
 For more information on parameter routing with special prefixes, see the [Advanced Usage Docs](https://www.adriangb.com/scikeras/stable/advanced.html#routed-parameters)
 

docs/source/notebooks/Benchmarks.md

@@ -50,7 +50,7 @@ warnings.filterwarnings("ignore", message="Setting the random state for TF")
 ```python
 import numpy as np
 from scikeras.wrappers import KerasClassifier, KerasRegressor
-from tensorflow import keras
+import keras
 ```
 
 ## 2. Dataset

docs/source/notebooks/DataTransformers.md

@@ -65,7 +65,7 @@ warnings.filterwarnings("ignore", message="Setting the random state for TF")
 ```python
 import numpy as np
 from scikeras.wrappers import KerasClassifier, KerasRegressor
-from tensorflow import keras
+import keras
 ```
 
 ## 2. Data transformer interface

docs/source/notebooks/MLPClassifier_MLPRegressor.md

@@ -58,7 +58,7 @@ warnings.filterwarnings("ignore", message="Setting the random state for TF")
 ```python
 import numpy as np
 from scikeras.wrappers import KerasClassifier, KerasRegressor
-from tensorflow import keras
+import keras
 ```
 
 ## 2. Defining the Keras Model

docs/source/notebooks/Meta_Estimators.md

@@ -50,7 +50,7 @@ warnings.filterwarnings("ignore", message="Setting the random state for TF")
 ```python
 import numpy as np
 from scikeras.wrappers import KerasClassifier, KerasRegressor
-from tensorflow import keras
+import keras
 ```
 
 ## 2. Defining the Keras Model

docs/source/notebooks/sparse.md

@@ -55,7 +55,7 @@ import numpy as np
 from scikeras.wrappers import KerasRegressor
 from sklearn.preprocessing import OneHotEncoder
 from sklearn.pipeline import Pipeline
-from tensorflow import keras
+import keras
 ```
 
 ## Data
@@ -89,19 +89,19 @@ def get_clf(meta) -> keras.Model:
 ## Pipelines
 
 Here is where it gets interesting.
-We make two Scikit-Learn pipelines that use `OneHotEncoder`: one that uses `sparse=False` to force a dense matrix as the output and another that uses `sparse=True` (the default).
+We make two Scikit-Learn pipelines that use `OneHotEncoder`: one that uses `sparse_output=False` to force a dense matrix as the output and another that uses `sparse_output=True` (the default).
 
 ```python
 dense_pipeline = Pipeline(
     [
-        ("encoder", OneHotEncoder(sparse=False)),
+        ("encoder", OneHotEncoder(sparse_output=False)),
         ("model", KerasRegressor(get_clf, loss="mse", epochs=5, verbose=False))
     ]
 )
 
 sparse_pipeline = Pipeline(
     [
-        ("encoder", OneHotEncoder(sparse=True)),
+        ("encoder", OneHotEncoder(sparse_output=True)),
         ("model", KerasRegressor(get_clf, loss="mse", epochs=5, verbose=False))
     ]
 )
@@ -153,7 +153,7 @@ You might be able to save even more memory by changing the output dtype of `OneH
 ```python
 sparse_pipline_uint8 = Pipeline(
     [
-        ("encoder", OneHotEncoder(sparse=True, dtype=np.uint8)),
+        ("encoder", OneHotEncoder(sparse_output=True, dtype=np.uint8)),
         ("model", KerasRegressor(get_clf, loss="mse", epochs=5, verbose=False))
     ]
 )

docs/source/quickstart.rst

@@ -13,7 +13,7 @@ it on a toy classification dataset using SciKeras
 
     import numpy as np
     from sklearn.datasets import make_classification
-    from tensorflow import keras
+    import keras
 
     from scikeras.wrappers import KerasClassifier
 

pyproject.toml

@@ -31,24 +31,18 @@ version = "0.12.0"
 
 [tool.poetry.dependencies]
 importlib-metadata = {version = ">=3", python = "<3.8"}
-python = ">=3.8.0,<3.12.0"
-scikit-learn = ">=1.0.0"
-packaging = ">=0.21"
-tensorflow = {version = ">=2.12.0,<2.13.0", optional = true}
-tensorflow-cpu = {version = ">=2.12.0,<2.13.0", optional = true}
-tensorflow-metal = {markers = "sys_platform == \"darwin\" and platform_machine == \"arm64\"", version = "^1.1.0"}
+python = ">=3.9.0,<4"
+scikit-learn = ">=1.4.1.post1"
+keras = { git = "https://github.com/keras-team/keras.git", rev = "master" }
+tensorflow = { version = ">=2.16.0", optional = true }
+tensorflow-cpu = { version = ">=2.16.0", optional = true }
 
 [tool.poetry.extras]
-tensorflow = ["tensorflow"]
 tensorflow-cpu = ["tensorflow-cpu"]
-
-[tool.poetry.dependencies.tensorflow-io-gcs-filesystem]
-# see https://github.com/tensorflow/tensorflow/issues/60202
-version = ">=0.23.1,<0.32"
-markers = "sys_platform == 'win32'"
+tensorflow = ["tensorflow"]
+test = ["tensorflow"]
 
 [tool.poetry.dev-dependencies]
-tensorflow = ">=2.12.0,<2.13.0"
 coverage = {extras = ["toml"], version = ">=6.4.2"}
 insipid-sphinx-theme = ">=0.3.2"
 ipykernel = ">=6.15.1"

scikeras/__init__.py

@@ -5,27 +5,13 @@
 try:
     import importlib.metadata as importlib_metadata
 except ModuleNotFoundError:
-    import importlib_metadata  # python <3.8
+    import importlib_metadata  # type: ignore  # python <3.8
 
-__version__ = importlib_metadata.version("scikeras")
+__version__ = importlib_metadata.version("scikeras")  # type: ignore
 
+import keras as _keras
 
-MIN_TF_VERSION = "2.7.0"
-TF_VERSION_ERR = f"SciKeras requires TensorFlow >= {MIN_TF_VERSION}."
-
-from packaging import version  # noqa: E402
-
-try:
-    from tensorflow import __version__ as tf_version
-except ImportError:  # pragma: no cover
-    raise ImportError("TensorFlow is not installed. " + TF_VERSION_ERR) from None
-else:
-    if version.parse(tf_version) < version.parse(MIN_TF_VERSION):  # pragma: no cover
-        raise ImportError(TF_VERSION_ERR) from None
-
-import tensorflow.keras as _keras  # noqa: E402
-
-from scikeras import _saving_utils  # noqa: E402
+from scikeras import _saving_utils
 
 _keras.Model.__reduce__ = _saving_utils.pack_keras_model
 _keras.Model.__deepcopy__ = _saving_utils.deepcopy_model