How can Tensorflow be used to compose layers using Python?

Tensorflow can be used to compose layers by defining a class that inherits from ‘ResnetIdentityBlock’. This is used to define a block which can be used to compose the layers.

Read More: What is TensorFlow and how Keras work with TensorFlow to create Neural Networks?

A neural network that contains at least one layer is known as a convolutional layer. We can use the Convolutional Neural Network to build learning model. 

TensorFlow Hub is a repository that contains pre-trained TensorFlow models. TensorFlow can be used to fine-tune learning models. We will understand how to use models from TensorFlow Hub with tf.keras, use an image classification model from TensorFlow Hub.  Once this is done, transfer learning can be performed to fine-tune a model for customized image classes. This is done by using a pretrained classifier model to take an image and predict what it is. This can be done without needing any training.  

We are using the Google Colaboratory to run the below code. Google Colab or Colaboratory helps run Python code over the browser and requires zero configuration and free access to GPUs (Graphical Processing Units). Colaboratory has been built on top of Jupyter Notebook.

Example

print("Composing layers")
class ResnetIdentityBlock(tf.keras.Model):
   def __init__(self, kernel_size, filters):
      super(ResnetIdentityBlock, self).__init__(name='')
      filters1, filters2, filters3 = filters
      self.conv2a = tf.keras.layers.Conv2D(filters1, (1, 1))
      self.bn2a = tf.keras.layers.BatchNormalization()
      self.conv2b = tf.keras.layers.Conv2D(filters2, kernel_size, padding='same')
      self.bn2b = tf.keras.layers.BatchNormalization()
      self.conv2c = tf.keras.layers.Conv2D(filters3, (1, 1))
      self.bn2c = tf.keras.layers.BatchNormalization()
   def call(self, input_tensor, training=False):
      x = self.conv2a(input_tensor)
      x = self.bn2a(x, training=training)
      x = tf.nn.relu(x)
      x = self.conv2b(x)
      x = self.bn2b(x, training=training)
      x = tf.nn.relu(x)
      x = self.conv2c(x)
      x = self.bn2c(x, training=training)
      x += input_tensor
      return tf.nn.relu(x)
print("The layer is called")
block = ResnetIdentityBlock(1, [1, 2, 3])
_ = block(tf.zeros([1, 2, 3, 3]))
block.layers
len(block.variables)
block.summary()

Code credit −https://www.tensorflow.org/tutorials/customization/custom_layers

Output

Composing layers
The layer is called
Model: "resnet_identity_block"
_________________________________________________________________
Layer (type)        Output Shape      Param #
=================================================================
conv2d (Conv2D)       multiple          4
_________________________________________________________________
batch_normalization (BatchNo multiple   4
_________________________________________________________________
conv2d_1 (Conv2D)      multiple        4
_________________________________________________________________
batch_normalization_1 (Batch multiple  8
_________________________________________________________________
conv2d_2 (Conv2D)     multiple         9
_________________________________________________________________
batch_normalization_2 (Batch multiple  12
=================================================================
Total params: 41
Trainable params: 29
Non-trainable params: 12

Explanation

• Every residual block in a resnet is composed of convolutions, batch normalizations, and a shortcut.

• Layers can be nested inside other layers too.

• When we need model methods such as Model.fit,Model.evaluate, and Model.save, it can be inherited from keras.Model.

• The keras.Model is used instead of keras.layers.Layer, which helps in tracking variables.

• A keras.Model tracks its internal layers, thereby making it easier to inspect the layers