lambda layer function definition without tf.keras.backend (Python Keras Package)

Question

tf.keras.layers.Lambda documentation explains how a function can be defined in a lambda layer. That document provides the following function as an example,

def antirectifier(x):

    x -= K.mean(x, axis=1, keepdims=True)
    x = K.l2_normalize(x, axis=1)

    pos = K.relu(x)
    neg = K.relu(-x)

    return K.concatenate([pos, neg], axis=1)

model.add(Lambda(antirectifier))

But according to that, tf.keras.backend must be used to conduct operations on the input Tensor object.

Is there any way we can use default python packages and user-defined function to define the steps of a lambda function.

If it's possible, please be kind enough to provide some examples.

Daniel Möller · Accepted Answer · 2018-08-15 14:21:47Z

2

If you're not using import tensorflow and its functions, there is absolutely no problem.

The code is perfect and that's it.

Just import keras.backend as K

Example rounded = K.round(x)

This is Keras independent documentation: https://keras.io/layers/core/#lambda

edited Aug 15, 2018 at 14:21

answered Aug 15, 2018 at 14:00

Daniel Möller

86.8k24 gold badges202 silver badges222 bronze badges

Sign up to request clarification or add additional context in comments.

3 Comments

cuser Over a year ago

But, for example, I simply tried to use round(x/3) in the body of the lamda function. Then I get "type Tensor doesn't define round method".

Daniel Möller Over a year ago

Yes... you can't use "python" functions in tensors. You must use backend functions: keras.io/backend ---------- Keras does not exist without a backend, no matter which you choose, there is always one. (K.round(x))

cuser Over a year ago

Thanks for your answer.

Leo K · Accepted Answer · 2018-08-15 14:21:07Z

1

There is no requirement to use tf.keras.backend as such. The function that you give to keras.layers.Lambda() can do anything you want with the input tensor.

However, most 'default python packages' are simply not able to work with a Tensor object (e.g., as you noted in your comment, round(x) won't work on it).

You can use most of the Python language constructs on a tensor x (such as x*2, or x+y, etc) - the tensor objects implement those. Likewise, you can call some 'regular' python libraries, as long as they use only operations that are supported by tensors. For everything else, you will need to use either methods of the tensor object or appropriate Keras library functions - or write your own.

answered Aug 15, 2018 at 14:21

Leo K

5,3843 gold badges14 silver badges27 bronze badges

1 Comment

cuser Over a year ago

Thanks for your answer.

Collectives™ on Stack Overflow

lambda layer function definition without tf.keras.backend (Python Keras Package)

2 Answers 2

3 Comments

1 Comment

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

2 Answers 2

3 Comments

1 Comment

Your Answer

Sign up or log in

Post as a guest

Related