keras

Wrapper class you can use to make any Keras model compatible with CleanLearning and sklearn. Use KerasWrapperModel to wrap existing functional API code for keras.Model objects, and KerasWrapperSequential to wrap existing tf.keras.models.Sequential objects. Most of the instance methods of this class work the same as the ones for the wrapped Keras model, see the Keras documentation for details.

This is a good example of making any bespoke neural network compatible with cleanlab.

You must have Tensorflow 2 installed (only compatible with Python versions >= 3.7).

Tips:

• If this class lacks certain functionality, you can alternatively try scikeras.

• Unlike scikeras, our KerasWrapper classes can operate directly on tensorflow.data.Dataset objects (like regular Keras models).

• To call fit() on a tensorflow Dataset object with a Keras model, the Dataset should already be batched.

• Check out our example using this class: huggingface_keras_imdb

• Our unit tests also provide basic usage examples.

Classes:

KerasWrapperModel(model[, model_kwargs, ...])	Takes in a callable function to instantiate a Keras Model (using Keras functional API) that is compatible with CleanLearning and sklearn.
KerasWrapperSequential([layers, name, ...])	Makes any tf.keras.models.Sequential object compatible with CleanLearning and sklearn.

class cleanlab.experimental.keras.KerasWrapperModel(model, model_kwargs={}, compile_kwargs={'loss': <keras.losses.SparseCategoricalCrossentropy object>})

Bases: object

Takes in a callable function to instantiate a Keras Model (using Keras functional API) that is compatible with CleanLearning and sklearn.

The instance methods of this class work in the same way as those of any keras.Model object, see the Keras documentation for details. For using Keras sequential instead of functional API, see the KerasWrapperSequential class.

Parameters:

• model (Callable) –

  A callable function to construct the Keras Model (using functional API). Pass in the function here, not the constructed model!

  For example:

  def model(num_features, num_classes):
      inputs = tf.keras.Input(shape=(num_features,))
      outputs = tf.keras.layers.Dense(num_classes)(inputs)
      return tf.keras.Model(inputs=inputs, outputs=outputs, name="my_keras_model")
  

• model_kwargs (dict, default = {}) – Dict of optional keyword arguments to pass into model() when instantiating the keras.Model.

• compile_kwargs (dict, default = {"loss": tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)}) – Dict of optional keyword arguments to pass into model.compile() for declaring loss, metrics, optimizer, etc.

Methods:

fit(X[, y])	Note that X dataset object must already contain the labels as is required for standard Keras fit.
get_params([deep])
predict(X, **kwargs)
predict_proba(X, *[, apply_softmax])	Set extra argument apply_softmax to True to indicate your network only outputs logits not probabilities.
summary(**kwargs)

fit(X, y=None, **kwargs)

Note that X dataset object must already contain the labels as is required for standard Keras fit. You can optionally provide the labels again here as argument y to be compatible with sklearn, but they are ignored.

get_params(deep=True)

predict(X, **kwargs)

predict_proba(X, *, apply_softmax=True, **kwargs)

Set extra argument apply_softmax to True to indicate your network only outputs logits not probabilities.

summary(**kwargs)

class cleanlab.experimental.keras.KerasWrapperSequential(layers=None, name=None, compile_kwargs={'loss': <keras.losses.SparseCategoricalCrossentropy object>})

Bases: object

Makes any tf.keras.models.Sequential object compatible with CleanLearning and sklearn.

KerasWrapperSequential is instantiated in the same way as a keras Sequential object, except for optional extra compile_kwargs argument. Just instantiate this object in the same way as your tf.keras.models.Sequential object (rather than passing in an existing Sequential object). The instance methods of this class work in the same way as those of any keras Sequential object, see the Keras documentation for details.

Parameters:

• layers (list) – A list containing the layers to add to the keras Sequential model (same as for tf.keras.models.Sequential).

• name (str, default = None) – Name for the Keras model (same as for tf.keras.models.Sequential).

• compile_kwargs (dict, default = {"loss": tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)}) – Dict of optional keyword arguments to pass into model.compile() for declaring loss, metrics, optimizer, etc.

Methods:

fit(X[, y])	Note that X dataset object must already contain the labels as is required for standard Keras fit.
get_params([deep])
predict(X, **kwargs)
predict_proba(X, *[, apply_softmax])	Set extra argument apply_softmax to True to indicate your network only outputs logits not probabilities.
summary(**kwargs)

fit(X, y=None, **kwargs)

Note that X dataset object must already contain the labels as is required for standard Keras fit. You can optionally provide the labels again here as argument y to be compatible with sklearn, but they are ignored.

get_params(deep=True)

predict(X, **kwargs)

predict_proba(X, *, apply_softmax=True, **kwargs)

Set extra argument apply_softmax to True to indicate your network only outputs logits not probabilities.

summary(**kwargs)