Custom Layer Learnable Parameters Initialization.

Hi!

I created a Custom Layer with Learnable Parameters (https://www.mathworks.com/help/deeplearning/ug/define-custom-deep-learning-layer.html)

However, to initialize these parameters I need to know the outputsize of the previous layer. I noticed that some pre-defined layers such as FullyConnected automaticly initialize its own weights using information from previous layers. Is there a way to automatize the weights initialization on custom layers at network initialization?

David

[Jose Cortes-Briones]

I asked a similar question a while ago and never got a response :/

Answer 1

[Joss Knight]

This is now supported via the optional method initialize: https://www-integ2.mathworks.com/help/deeplearning/ug/define-custom-deep-learning-intermediate-layers.html#mw_87362c7c-37d3-40c2-b5c8-c45ded304bcb

[Jose Cortes-Briones]

Great! Thanks!

Answer 2

[John Smith]

Unfortunately, at least as of version 2021a, Matlab doesn't provide staright forward capabilities to automatically initialize learnable parameters of a custom layer similar to its built-in layers.

There's a workaround though using a shadow built-in network nested inside your custom layer (see here). Matlab will automatically initialize this layer from which learnable parameters you can pick up the values that you need.

It's a kludge, but it works (I've implemented such a scheme myself.

[Jose Cortes-Briones]

Thanks!