CODEWITHSUNDEEP
tensorflowkeraskeras-layer
Manngo
Manngo

Reputation: 829

How to wrap a tensorflow object as Keras layer?

I would like to implement Hierarchical Multiscale LSTM as a Keras layer.
It was published here and implemented in tensorflow here.
My understanding is that there's a way to wrap such a tensorflow object in Keras as a layer. I'm not sure how complicated it is but I think it's feasible. Can you help me how to do it?

Upvotes: 2

Views: 1452

Answers (1)

rvinas
rvinas

Reputation: 11895

This is usually done by implementing a custom Layer. To be more specific, you should inherit from keras.engine.topology.layer and provide a custom implementation for the following methods (and place the TensorFlow code within them):

  • build(input_shape): this is where you will define your weights. This method must set self.built = True, which can be done by calling super([Layer], self).build()
  • call(x): this is where the layer's logic lives. Unless you want your layer to support masking, you only have to care about the first argument passed to call: the input tensor.
  • compute_output_shape(input_shape): in case your layer modifies the shape of its input, you should specify here the shape transformation logic. This allows Keras to do automatic shape inference.

Since you're trying to implement a recurrent layer, it would also be convenient to inherit directly from keras.legacy.layers.recurrent. In this case, you probably do not need to redefine compute_output_shape(input_shape). If your layer needs additional arguments, you can pass them to the __init__ method of your custom layer.

Upvotes: 3

Related Questions