CODEWITHSUNDEEP
pythontensorflowkeras
Nick-H
Nick-H

Reputation: 398

How do I create repeated structures in a Keras network?

I am trying to create an autoencoder network in Keras to learn the dynamics of a system I'm analyzing but am not sure how to do it. As you can see below, x is encoded by phi into a lower dimensional z-space. The A network computes the derivative of z and then psi decodes the derivative back up into the full dimensional space. In addition to having psi decode z_dot to x_dot, I would like to add the constraint that psi also decodes z back into x. You can see in the diagram below that this can be done by repeating the decoder structure.

enter image description here

I know you can create Keras networks with multiple input/output tensors, but I am not sure if it's possible to repeat my decoder structure like I have shown? In this sense the decoder will learn from two data sets simultaneously. Is this possible? If so, how could I implement this?

Thanks!

Upvotes: 0

Views: 55

Answers (1)

Ayush Goel
Ayush Goel

Reputation: 375

For this, you'd want to create an architecture called psi. Then instantiate it once, and reuse that as a layer in your network. The idea is to share the weights (use the exact same layer twice). So you might have:

def create_psi():
    # define your psi architecture here
    return psi

def decoder():
    psi = create_psi()
    X_input = Input(dims) #dims are the dimensions of z
    Z_dot = A(X_input) #Using whatever layer A is
    X_dot = psi(Z_dot)
    X = psi(X_input)
    decoder = Model(inputs=X_input, outputs=[X_dot, X])
    return decoder

# Instantiate decoder model here, compile with appropriate loss, etc.

Upvotes: 1

Related Questions