Keras functional classification model example

An example of a functional (non-sequential) network used as an OpenML flow.

import keras

import openml.extensions.keras

Define an input layer for the network. In this example we are using the german credit dataset, which contains 20 features, and as such the input shape will be (20,).

inp = keras.layers.Input(shape=(20,))

# Normalize the input data in order to speed up the training process
normalized = keras.layers.BatchNormalization()(inp)

# Fork the input data into two parallel dense layers
# which use ReLU activation.
dense1 = keras.layers.Dense(units=64, activation='relu')(normalized)
dense2 = keras.layers.Dense(units=64, activation='relu')(normalized)

# Merge the results of the two parallel layers into one merge layer.
merged = keras.layers.concatenate([dense1, dense2])

# Introduce an additional Dense layer in combination to a dropout layer.
dense3 = keras.layers.Dense(units=64, activation='sigmoid')(merged)
dropout1 = keras.layers.Dropout(rate=0.25)(dense3)
# Finally, output the probabiltiies in the final dense layer using
# softmax activation.
dense4 = keras.layers.Dense(units=2, activation='softmax')(dropout1)

# Construct this model which uses our functional neural network
# with one input and one output.
model = keras.models.Model(inputs=[inp], outputs=[dense4])

# Compile it using the Adam optimizer while targeting accuracy.
model.compile(optimizer='adam',
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])

Download the OpenML task for the german credit card dataset.

task = openml.tasks.get_task(31)

Run the Keras model on the task (requires an API key).

run = openml.runs.run_model_on_task(model, task, avoid_duplicate_runs=False)
# Publish the experiment on OpenML (optional, requires an API key).
run.publish()

print('URL for run: %s/run/%d' % (openml.config.server, run.run_id))