Migrating from tf.keras.wrappers.scikit_learn

Why switch to SciKeras

SciKeras has several advantages over tf.keras.wrappers.scikit_learn:

  • Full compatibility with the Scikit-Learn API, including grid searches, ensembles, transformers, etc.

  • Support for Functional and Subclassed Keras Models.

  • Support for pre-trained models.

  • Support for dynamically set Keras parameters depending on inputs (e.g. input shape).

  • Support for hyperparameter tuning of optimizers and losses.

  • Support for multi-input and multi-ouput Keras models.

  • Functional random_state for reproducible training.

  • Many more that you will discover as you use SciKeras!

Changes to your code

SciKeras is largely backwards compatible with the existing wrappers. For most cases, you can just change your import statement from:

- from tensorflow.keras.wrappers.scikit_learn import KerasClassifier, KerasRegressor
+ from scikeras.wrappers import KerasClassifier, KerasRegressor

SciKeras does however have some backward incompatible changes:

Fit returns self

In TensorFlow calling fit on wrappers returned a Keras History object, which held an attribute called history that is a dictionary with loss and metric names as keys and lists of recorded values for each epoch as the dictionary values. However, in SciKeras, fit now returns and instance of the estimator itself in order to conform to the Scikit-Learn API. Instead, the history is saved in the history_ attribute. Calling fit resets this attribute, calling partial_fit on the other hand extends it.

def get_model():
   ...
   model.compile(loss="mse", metrics=["mae"])
   return model

clf = KerasClassifier(get_model)
- hist = clf.fit(...).history
- losses = hist["mae"]
+ hist = clf.fit(...).history_
+ losses = hist["mean_absolute_error"]

Note

Unlike the TensorFlow wrappers, SciKeras normalizes the names of the keys, so that if you use metrics=[“mae”] you will get a key named “mean_absolute_error”.

One-hot encoding of targets for categorical crossentropy losses

SciKeras will not longer implicitly inspect your Model’s loss function to determine if it needs to one-hot encode your target to match the loss function. Instead, you must explicitly pass your loss function to the constructor:

clf = KerasClassifier(loss="categorical_crossentropy")

Variable keyword arguments in fit and predict

Keras supports a variable keyword arguments (commonly referred to as **kwargs) for fit and predict. Scikit-Learn on the other hand does not support these arguments, and using them is largely incompatible with the Scikit-Learn ecosystem. As a compromise, SciKeras supports these arguments, but we recommended that you set parameters using the constructor or set_params for first-class SciKeras support.

For example, to declare batch_size in the constructor:

- clf = KerasClassifier(...)
- clf.fit(..., batch_size=32)
+ clf = KerasClassifier(..., batch_size=32)
+ clf.fit(...)

Or to declare separate values for fit and predict:

clf = KerasClassifier(..., fit__batch_size=32, predict__batch_size=10000)

If you want to change the parameters on a live instance, you can do:

clf = KerasClassifier(...)
clf.set_params(fit__batch_size=32, predict__batch_size=10000)
clf.fit(...)

Functionally, this is the same as passing these parameters to fit, just with one more function call. This is much more compatible with the Scikit-Learn API. In fact, this is what Scikti-Learn does in the background for hyperparameter tuning.

Renaming of build_fn to model

SciKeras renamed the constructor argument build_fn to model. In a future release, passing build_fn as a _keyword_ argument will raise a TypeError. Passing it as a positional argument remains unchanged. You can make the following change to future proof your code:

- clf = KerasClassifier(build_fn=...)
+ clf = KerasClassifier(model=...)

Default arguments in build_fn/model

SciKeras will no longer introspect your callable model for user defined parameters (the behavior for parameters like optimizer is unchanged). You must now “declare” them as keyword arguments to the constructor if you want them to be tunable parameters (i.e. settable via set_params):

- def get_model(my_param=123):
+ def get_model(my_param):  # You can optionally remove the default here
   ...
   return model

- clf = KerasClassifier(get_model)
+ clf = KerasClassifier(get_model, my_param=123)  # option 1
+ clf = KerasClassifier(get_model, model__my_param=123)  # option 2

That said, if you do not need them to work with set_params (which is only really necessary if you are doing hyperparameter tuning), you do not need to make any changes.