Custom Layers

English Version tensorflow
1

https://d2l.ai/chapter_builders-guide/custom-layer.html

2

Hello,

As a start, thanks for this nice book.

I have defined the following layer which adds 2:

class add_2(Model):
    def __init__(self):
        #initiate class
        super().__init__()
    
    def call(self, X):
        return X+2

I then defined this layer which adds takes a parameter K as input and adds it:

class add_K(Model):
    def __init__(self, K):
        #initiate class
        super().__init__()
        self.K = K

    def call(self, X):
        return X+self.K

Now, assume I realise I always want to add 2 to layers, I can do that:

class add_2_to_layer(Model):
    def __init__(self, layer):
        #initiate class
        super().__init__()
        self.layer = layer
    
    def call(self, X):
        return self.layer(X)+2

If I do add_2_to_layer(add_K(K=3)) it works fine.

Now, assume I always want to add 2 on add_K, implicitly, without calling add_2_to_layer, then I think I need some kind of ‘‘decorator’’ (though I am not familiar with them). I have tried:

class add_2_to_layer_class(Model):
    def __init__(self, layer):
        #initiate class
        super().__init__()
        self.layer = layer
    
    def call(self, X):
        return self.layer(X)+2

def add_2_to_layer_func(layer):
    return add_2_to_layer_class(layer)
    
@add_2_to_layer_func
class add_K(Model):
    def __init__(self, K):
        #initiate class
        super().__init__()
        self.K = K

    def call(self, X):
        return X+self.K

and

class add_2_to_layer(Model):
    def __init__(self, layer):
        #initiate class
        super().__init__()
        self.layer = layer
    
    def call(self, X):
        return self.layer(X)+2

@add_2_to_layer
class add_K(Model):
    def __init__(self, K):
        #initiate class
        super().__init__()
        self.K = K

    def call(self, X):
        return X+self.K

but none of them work.

I could define a new class:

class add_K_with_2(Model):
    def __init__(self, K):
        #initiate class
        super().__init__()
        self.K = K

        self.layer = add_2_to_layer(add_K(K=self.K))

    def call(self, X):
        return self.layer(X)

but it’s not very elegant nor versatile (if I want to add 2 to another layer type, I have to write another piece of code). Any idea?

1 Like
3

import tensorflow as tf
import tensorflow.keras as keras
import numpy as np

class ReducedDense(keras.Model):
def init(self, units, k):
super().init()
self.units = units
self.no_of_weights = k

def build(self, input):
self.weight = tf.Variable(np.random.uniform(0, 1, size=(self.no_of_weights, self.units , input[-1])), dtype=tf.float32)
print(self.weight.shape)

def call(self, X ):
out = tf.matmul(tf.transpose(X), X)
y = self.weight[2]
return [tf.matmul(self.weight[i] , out ) for i in range(self.no_of_weights)]

dense = ReducedDense(3, 4 )
dense(tf.random.uniform((2, 5)))

a layer that takes an input and computes a tensor reduction

4

I think there is a mistake in the code of 6.5.1
Y = net(tf.random.uniform((4, 8)))
tf.reduce_mean(Y)
this code gives reduce_mean for all the distribution where we need to print for each row in tf.random.uniform
instead it can be
for row in Y:
…print(tf.reduce_mean(row))