Keras – Customized Layer

Keras allows to create our own customized layer. Once a new layer is created, it can be used in any model without any restriction. Let us learn how to create new layer in this chapter.
Keras provides a base layer class, Layer which can sub-classed to create our own customized layer. Let us create a simple layer which will find weight based on normal distribution and then do the basic computation of finding the summation of the product of input and its weight during training.

Step 1: Import the necessary module

First, let us import the necessary modules −

from keras import backend as K
from keras.layers import Layer

Here,

• backend is used to access the dot function.
• Layer is the base class and we will be sub-classing it to create our layer
  

  Step 2: Define a layer class

  Let us create a new class, MyCustomLayer by sub-classing Layer class −

  class MyCustomLayer(Layer):
  ...

  Step 3: Initialize the layer class

  Let us initialize our new class as specified below −

  def __init__(self, output_dim, **kwargs):
  self.output_dim = output_dim
  super(MyCustomLayer, self).__init__(**kwargs)

  Here,

• Line 2 sets the output dimension.
• Line 3 calls the base or super layer’s init function.
  

  Step 4: Implement build method

  build is the main method and its only purpose is to build the layer properly. It can do anything related to the inner working of the layer. Once the custom functionality is done, we can call the base class build function. Our custom build function is as follows −

  def build(self, input_shape):
  self.kernel = self.add_weight(name = 'kernel',
  shape = (input_shape[1], self.output_dim),
  initializer = 'normal', trainable = True)
  super(MyCustomLayer, self).build(input_shape)

  Here,

• Line 1 defines the build method with one argument, input_shape. Shape of the input data is referred by input_shape.
• Line 2 creates the weight corresponding to input shape and set it in the kernel. It is our custom functionality of the layer. It creates the weight using ‘normal’ initializer.
• Line 6 calls the base class, build method.
  

  Step 5: Implement call method

  call method does the exact working of the layer during training process.
  Our custom call method is as follows

  def call(self, input_data):
  return K.dot(input_data, self.kernel)

  Here,

• Line 1 defines the call method with one argument, input_data. input_data is the input data for our layer.
• Line 2 return the dot product of the input data, input_data and our layer’s kernel, self.kernel
  

  Step 6: Implementing compute_output_shape method

  def compute_output_shape(self, input_shape): return (input_shape[0], self.output_dim)

  Here,

• Line 1 defines compute_output_shape method with one argument input_shape
• Line 2 computes the output shape using shape of input data and output dimension set while initializing the layer.
  Implementing the build, call and compute_output_shape completes the creating a customized layer. The final and complete code is as follows

  from keras import backend as K from keras.layers import Layer
  class MyCustomLayer(Layer):
  def __init__(self, output_dim, **kwargs):
  self.output_dim = output_dim
  super(MyCustomLayer, self).__init__(**kwargs)
  def build(self, input_shape): self.kernel =
  self.add_weight(name = 'kernel',
  shape = (input_shape[1], self.output_dim),
  initializer = 'normal', trainable = True)
  super(MyCustomLayer, self).build(input_shape) #
  Be sure to call this at the end
  def call(self, input_data): return K.dot(input_data, self.kernel)
  def compute_output_shape(self, input_shape): return (input_shape[0], self.output_dim)

  Using our customized layer

  Let us create a simple model using our customized layer as specified below −

  from keras.models import Sequential
  from keras.layers import Dense
  my_model = Sequential()
  my_model.add(MyCustomLayer(32, input_shape = (16,)))
  my_model.add(Dense(8, activation = 'softmax')) my_model.summary()

  Here,

• Our MyCustomLayer is added to the model using 32 units and (16,) as input shape
  Running the application will print the model summary as below −

  Model: "sequential_1"
  _________________________________________________________________
  Layer (type) Output Shape Param
  #================================================================
  my_custom_layer_1 (MyCustomL (None, 32) 512
  _________________________________________________________________
  dense_1 (Dense) (None, 8) 264
  =================================================================
  Total params: 776
  Trainable params: 776
  Non-trainable params: 0
  _________________________________________________________________