3 Ways to build a Keras Model
The scope of this story is to build a Keras model using different ways: the sequential model, the functional model, and a custom model via sub-classing.
The sequential model:
First, building a model by passing a list of layers to it.
model = tf.keras.Sequential([
tf.keras.layers.Flatten(input_shape=(28, 28)),
tf.keras.layers.Dense(128, activation='relu'),
tf.keras.layers.Dense(10)
])Or by adding layers via add method.
model = tf.keras.Sequential()
model.add(tf.keras.layers.Flatten(input_shape=(28, 28)))
model.add(tf.keras.layers.Dense(128, activation='relu'))
model.add(tf.keras.layers.Dense(10))This approach only works when you have a single input and a single output.
The functional API:
Second, this approach adds arguments using parentheses; for example, the function f(x, y) becomes f(x)(y). This style is called currying in functional programming.
input_ = tf.keras.layers.Input(shape=(28, 28))
flat = tf.keras.layers.Flatten()(input_)
hidden = tf.keras.layers.Dense(128, activation='relu')(flat)
output_ = tf.keras.layers.Dense(10)(hidden)
model = tf.keras.Model(inputs=[input_], outputs=[output_])This approach allows for creating complex models with: multiple inputs, auxiliary outputs, and non-sequential paths.
A custom model via sub-classing:
Finally, this approach is for research use. It allows researchers to create complex models such as: adding custom layers, skipping connections, etc.
class CustomModel(tf.keras.Model):
def __init__(self, units=128, activation='relu', **kwargs):
super().__init__(**kwargs)
self.flat = tf.keras.layers.Flatten()
self.hidden = tf.keras.layers.Dense(units,
activation=activation)
self.output_ = tf.keras.layers.Dense(10)def call(self, inputs):
flat = self.flat(inputs)
hidden = self.hidden(flat)
return self.output_(hidden)model = CustomModel()
Albeit, when it comes to serialization, this approach requires additional work.
The complete code is in the link below.
https://github.com/booletic/medium/blob/98d3023ddf044ad5cc318e5b29ebcc252331fb74/keras_models.ipynb
