How to create a dynamic number of layers in Tensorflow?

user_7619808 0
Viewed 506

In Keras, I would do the following to dynamically create a model's layers:

for i in range(number_dense_layers):
        model.add(layers.Dense(units=units, input_dim=input_dim,
                  kernel_initializer='normal', activation='relu'))

however, in the case of Tensorflow, I have the following:

class generic_vns_function(tf.keras.Model):
    def __init__(self, num_layers, num_class=10): 
        super().__init__() 
        # Convolutional layers and MaxPools

        self.conv1 = tf.keras.layers.Conv2D(64, 3, activation="relu") 
        self.conv2 = tf.keras.layers.Conv2D(64, 3, activation="relu")

where I would want to do something like:

for i in range(num_layers):
            self.add(tf.keras.layers.Conv2D(64, 3, activation="relu"))

but I am unsure how to dynamically create this layer since the add function does not work in this context as it did in Keras.

keras tensorflow
4 Answers

It looks like the following is what is needed:

for i in range(num_layers):
    self.layers.append(tf.keras.layers.Conv2D(64, 3, activation="relu"))

where we append in a specific layer to the models layers attribute.

user_76198080

You can append first and stack them later.

Here is a rough example:

import tensorflow as tf

class generic_vns_function(tf.keras.Model):
    def __init__(self, num_layers, num_class=10): 
        super().__init__() 
        self.convolutions = []
        ...
        for i in range(num_layers):
          self.convolutions.append(tf.keras.layers.Conv2D(64, 3, activation="relu"))

    def call(self, inputs):
      ...
      for layer in self.convolutions:
        x = layer(x)
      ...
      return x
user_137266680

Inspired by Frightera, If you want dynamic number of layers with different function, try dict. Dict is ordered after python 3.6.

user_94551300

Instead of using for-loops to define multiple layers I prefere list comprehension:

import tensorflow as tf

class generic_vns_function(tf.keras.Model):
    def __init__(self, num_layers, num_class=10): 
        super().__init__() 
        ...
        self.convolutions = [tf.keras.layers.Conv2D(64, 3, activation="relu") for _ in range(num_layers)]

def call(self, inputs):
  ...
  for layer in self.convolutions:
    x = layer(x)
  ...
  return x
user_179057640
Related