Run in Google Colab

MLPClassifier and MLPRegressor in SciKeras

SciKeras is a bridge between Keras and Scikit-Learn. As such, one of SciKeras’ design goals is to be able to create a Scikit-Learn style estimator backed by Keras.

This notebook implements an estimator that is analogous to sklearn.neural_network.MLPClassifier using Keras. This estimator should (for the most part) work as a drop-in replacement for MLPClassifier!

Table of contents

• 1. Setup

• 2. Defining the Keras Model

  • 2.1 Inputs

  • 2.2 Hidden layers

  • 2.3 Output layers

  • 2.4 Losses and optimizer

  • 2.5 Wrapping with SciKeras

• 3. Testing our classifier

• 4. Self contained MLPClassifier

  • 4.1 Subclassing

• 5. MLPRegressor

1. Setup

try:
    import scikeras
except ImportError:
    !python -m pip install scikeras

Silence TensorFlow logging to keep output succinct.

import warnings
from tensorflow import get_logger
get_logger().setLevel('ERROR')
warnings.filterwarnings("ignore", message="Setting the random state for TF")

import numpy as np
from scikeras.wrappers import KerasClassifier, KerasRegressor
from tensorflow import keras

2. Defining the Keras Model

First, we outline our model building function, using a Sequential Model:

def get_clf_model():
    model = keras.Sequential()
    return model

2.1 Inputs

We need to define an input layer for Keras. SciKeras allows you to dynamically determine the input size based on the features (X). To do this, you need to add the meta parameter to get_clf_model’s parameters. meta will be a dictionary with all of the meta attributes that KerasClassifier generates during the fit call, including n_features_in_, which we will use to dynamically size the input layer.

from typing import Dict, Iterable, Any


def get_clf_model(meta: Dict[str, Any]):
    model = keras.Sequential()
    inp = keras.layers.Input(shape=(meta["n_features_in_"]))
    model.add(inp)
    return model

2.2 Hidden Layers

Multilayer perceptrons are generally composed of an input layer, an output layer and 0 or more hidden layers. The size of the hidden layers is specified via the hidden_layer_sizes parameter in MLClassifier, where the the ith element represents the number of neurons in the ith hidden layer. Let’s add that parameter:

def get_clf_model(hidden_layer_sizes: Iterable[int], meta: Dict[str, Any]):
    model = keras.Sequential()
    inp = keras.layers.Input(shape=(meta["n_features_in_"]))
    model.add(inp)
    for hidden_layer_size in hidden_layer_sizes:
        layer = keras.layers.Dense(hidden_layer_size, activation="relu")
        model.add(layer)
    return model

2.3 Output layers

The output layer needs to reflect the type of classification task being performed. Here, we will handle 2 cases:

• binary classification: single output unit with sigmoid activation

• multiclass classification: one output unit for each class, with softmax activation The main complication arises from determining which one to use. Like with the input features, SciKeras provides useful information on the target within the meta parameter. Specifically, we will use the n_classes_ and target_type_ attributes to determine the number of output units and activation function.

def get_clf_model(hidden_layer_sizes: Iterable[int], meta: Dict[str, Any]):
    model = keras.Sequential()
    inp = keras.layers.Input(shape=(meta["n_features_in_"]))
    model.add(inp)
    for hidden_layer_size in hidden_layer_sizes:
        layer = keras.layers.Dense(hidden_layer_size, activation="relu")
        model.add(layer)
    if meta["target_type_"] == "binary":
        n_output_units = 1
        output_activation = "sigmoid"
    elif meta["target_type_"] == "multiclass":
        n_output_units = meta["n_classes_"]
        output_activation = "softmax"
    else:
        raise NotImplementedError(f"Unsupported task type: {meta['target_type_']}")
    out = keras.layers.Dense(n_output_units, activation=output_activation)
    model.add(out)
    return model

For now, we raise a NotImplementedError for other target types. For an example handling multi-output target types, see the Multi Output notebook.

2.4 Losses and optimizer

Like the output layer, the loss must match the type of classification task. Generally, it is easier and safet to allow SciKeras to compile your model for you by passing the loss to KerasClassifier directly (KerasClassifier(loss="binary_crossentropy")). However, in order to implement custom logic around the choice of loss function, we compile the model ourselves within get_clf_model; SciKeras will not re-compile the model.

def get_clf_model(hidden_layer_sizes: Iterable[int], meta: Dict[str, Any]):
    model = keras.Sequential()
    inp = keras.layers.Input(shape=(meta["n_features_in_"]))
    model.add(inp)
    for hidden_layer_size in hidden_layer_sizes:
        layer = keras.layers.Dense(hidden_layer_size, activation="relu")
        model.add(layer)
    if meta["target_type_"] == "binary":
        n_output_units = 1
        output_activation = "sigmoid"
        loss = "binary_crossentropy"
    elif meta["target_type_"] == "multiclass":
        n_output_units = meta["n_classes_"]
        output_activation = "softmax"
        loss = "sparse_categorical_crossentropy"
    else:
        raise NotImplementedError(f"Unsupported task type: {meta['target_type_']}")
    out = keras.layers.Dense(n_output_units, activation=output_activation)
    model.add(out)
    model.compile(loss=loss)
    return model

At this point, we have a valid, compiled model. However if we want to be able to tune the optimizer, we should accept compile_kwargs as a parameter in get_clf_model. compile_kwargs will be a dictionary containing valid kwargs for Model.compile, so we can unpack it directly like model.compile(**compile_kwargs). In this case however, we will only be taking the optimizer kwarg.

def get_clf_model(hidden_layer_sizes: Iterable[int], meta: Dict[str, Any], compile_kwargs: Dict[str, Any]):
    model = keras.Sequential()
    inp = keras.layers.Input(shape=(meta["n_features_in_"]))
    model.add(inp)
    for hidden_layer_size in hidden_layer_sizes:
        layer = keras.layers.Dense(hidden_layer_size, activation="relu")
        model.add(layer)
    if meta["target_type_"] == "binary":
        n_output_units = 1
        output_activation = "sigmoid"
        loss = "binary_crossentropy"
    elif meta["target_type_"] == "multiclass":
        n_output_units = meta["n_classes_"]
        output_activation = "softmax"
        loss = "sparse_categorical_crossentropy"
    else:
        raise NotImplementedError(f"Unsupported task type: {meta['target_type_']}")
    out = keras.layers.Dense(n_output_units, activation=output_activation)
    model.add(out)
    model.compile(loss=loss, optimizer=compile_kwargs["optimizer"])
    return model

2.5 Wrapping with SciKeras

Our last step in defining our model is to wrap it with SciKeras. A couple of things to note are: - Every user-defined parameter in model/get_clf_model (in our case just hidden_layer_sizes) must be defined as a keyword argument to KerasClassifier with a default value. - Keras defaults to "rmsprop" for optimizer. We set it to "adam" to mimic MLPClassifier. - We set the learning rate for the optimizer to 0.001, again to mimic MLPClassifier. We set this parameter using parameter routing. - Other parameters, such as activation, can be added similar to hidden_layer_sizes, but we omit them here for simplicity.

clf = KerasClassifier(
    model=get_clf_model,
    hidden_layer_sizes=(100, ),
    optimizer="adam",
    optimizer__learning_rate=0.001,
    epochs=50,
    verbose=0,
)

3. Testing our classifier

Before continouing, we will run a small test to make sure we get somewhat reasonable results.

from sklearn.datasets import make_classification


X, y = make_classification()

# check that fit works
clf.fit(X, y)
# check score
print(clf.score(X, y))

0.96

We get a score above 0.7, which is reasonable and indicates that our classifier is generally working.

4. Self contained MLPClassifier

You will have noticed that up until now, we define our Keras model in a function and pass that function to KerasClassifier via the model argument.

This is convenient, but it does not give us a self-contained class that we could package within a module for users to instantiate. To do that, we need to subclass KerasClassifier.

4.1 Subclassing

By subclassing KerasClassifier, you can embed your Keras model into directly into your estimator class. We start by inheriting from KerasClassifier and defining an __init__ method with all of our parameters.

class MLPClassifier(KerasClassifier):

    def __init__(
        self,
        hidden_layer_sizes=(100, ),
        optimizer="adam",
        optimizer__learning_rate=0.001,
        epochs=200,
        verbose=0,
        **kwargs
    ):
        super().__init__(**kwargs)
        self.hidden_layer_sizes = hidden_layer_sizes
        self.optimizer = optimizer
        self.epochs = epochs
        self.verbose = verbose

Next, we will embed our model into _keras_build_fn, which takes the place of get_clf_model. Note that since this is now an part of the model, we no longer need to accept the any parameters in the function signature. We still accept compile_kwargs because we use it to get the optimizer initialized with all of it’s parameters.

class MLPClassifier(KerasClassifier):

    def __init__(
        self,
        hidden_layer_sizes=(100, ),
        optimizer="adam",
        optimizer__learning_rate=0.001,
        epochs=200,
        verbose=0,
        **kwargs,
    ):
        super().__init__(**kwargs)
        self.hidden_layer_sizes = hidden_layer_sizes
        self.optimizer = optimizer
        self.epochs = epochs
        self.verbose = verbose

    def _keras_build_fn(self, compile_kwargs: Dict[str, Any]):
        model = keras.Sequential()
        inp = keras.layers.Input(shape=(self.n_features_in_))
        model.add(inp)
        for hidden_layer_size in self.hidden_layer_sizes:
            layer = keras.layers.Dense(hidden_layer_size, activation="relu")
            model.add(layer)
        if self.target_type_ == "binary":
            n_output_units = 1
            output_activation = "sigmoid"
            loss = "binary_crossentropy"
        elif self.target_type_ == "multiclass":
            n_output_units = self.n_classes_
            output_activation = "softmax"
            loss = "sparse_categorical_crossentropy"
        else:
            raise NotImplementedError(f"Unsupported task type: {self.target_type_}")
        out = keras.layers.Dense(n_output_units, activation=output_activation)
        model.add(out)
        model.compile(loss=loss, optimizer=compile_kwargs["optimizer"])
        return model

Let’s check that our subclassed model works:

clf = MLPClassifier(epochs=20)  # for notebook execution time

# check score
print(clf.fit(X, y).score(X, y))

0.96

5. MLPRegressor

The process for MLPRegressor is similar, we only change the loss function and output layers.

class MLPRegressor(KerasRegressor):

    def __init__(
        self,
        hidden_layer_sizes=(100, ),
        optimizer="adam",
        optimizer__learning_rate=0.001,
        epochs=200,
        verbose=0,
        **kwargs,
    ):
        super().__init__(**kwargs)
        self.hidden_layer_sizes = hidden_layer_sizes
        self.optimizer = optimizer
        self.epochs = epochs
        self.verbose = verbose

    def _keras_build_fn(self, compile_kwargs: Dict[str, Any]):
        model = keras.Sequential()
        inp = keras.layers.Input(shape=(self.n_features_in_))
        model.add(inp)
        for hidden_layer_size in self.hidden_layer_sizes:
            layer = keras.layers.Dense(hidden_layer_size, activation="relu")
            model.add(layer)
        out = keras.layers.Dense(1)
        model.add(out)
        model.compile(loss="mse", optimizer=compile_kwargs["optimizer"])
        return model

from sklearn.datasets import make_regression


reg = MLPRegressor(epochs=20)  # for notebook execution time

# Define a simple linear relationship
y = np.arange(100)
X = (y/2).reshape(-1, 1)

# check score
reg.fit(X, y)
print(reg.score(X, y))

-0.2058464795630428