Defining Models in Functional vs Sequential

Summary

• Defining a model using functional pattern
• Defining a model using sequential pattern

Content

Instead of using Sequential to create layers, functional approach can be used to create the model. From the results, you can see, evaluation shows the accuracy is similar in both models.

Functional

import tensorflow as tf

base_model = tf.keras.applications.EfficientNetB0(include_top=False)
base_model.trainable = False

inputs = tf.keras.layers.Input((256, 256, 3), name="input_layer")

x = base_model(inputs)
x = tf.keras.layers.GlobalAveragePooling2D(name="pooling_layer")(x)
outputs = tf.keras.layers.Dense(10, activation="softmax", name="output_layers")(x)

model_0 = tf.keras.Model(inputs, outputs)

model_0.compile(
    loss=tf.keras.losses.SparseCategoricalCrossentropy(),
    optimizer=tf.keras.optimizers.Adam(),
    metrics=["accuracy"],
)

model_0.fit(train_data, validation_data=test_data, epochs=4)
"""
loss: 1.8878 - accuracy : 0.3880 - val_loss: 1.3062 - val_accuracy: 0.7228
loss: 1.1279 - accuracy : 0.7467 - val_loss: 0.8900 - val_accuracy: 0.8208
loss: 0.8099 - accuracy : 0.8240 - val_loss: 0.7078 - val_accuracy: 0.8412
loss: 0.6601 - accuracy : 0.8467 - val_loss: 0.6165 - val_accuracy: 0.8564
"""

Sequential

import tensorflow as tf

base_model = tf.keras.applications.EfficientNetB0(include_top=False)
base_model.trainable = False

model = tf.keras.Sequential(
    [
        tf.keras.layers.Input((256, 256, 3)),
        base_model,
        tf.keras.layers.GlobalAveragePooling2D(),
        tf.keras.layers.Dense(10, activation="softmax"),
    ]
)

model.compile(
    loss=tf.keras.losses.SparseCategoricalCrossentropy(),
    optimizer=tf.keras.optimizers.Adam(),
    metrics=["accuracy"],
)

model.fit(train_data, validation_data=test_data, epochs=4)

"""
loss: 1.9393 - accuracy : 0.3893 - val_loss: 1.3565 - val_accuracy: 0.7188
loss: 1.1338 - accuracy : 0.7640 - val_loss: 0.8972 - val_accuracy: 0.8192
loss: 0.8184 - accuracy: 0.8213 - val_loss: 0.7134 - val_accuracy: 0.8416
loss: 0.6624 - accuracy: 0.8547 - val_loss: 0.6223 - val_accuracy: 0.8532
""""