From the doc page of custom layers: (https://www.mathworks.com/help/deeplearning/ug/define-custom-deep-learning-layers.html):
Defining the backward function is optional. If you do not specify a backward function, and the layer forward functions support dlarray objects, then the software automatically determines the backward function using automatic differentiation. For a list of functions that support dlarray objects, see List of Functions with dlarray Support. Define a custom backward function when you want to:
- Use a specific algorithm to compute the derivatives.
- Use operations in the forward functions that do not support dlarray objects.
Hence, writing a backward function might not be necessary in your case.
I understood that you would like to connect 2 networks between each other. In order to connect layers, the layerGraph API is the suggested approach (https://www.mathworks.com/help/deeplearning/ref/nnet.cnn.layergraph.html). You should:
- create the layerGraph that you want to attach in front (which may contain custom layers with or without backward function)
- extract the layerGraph from the pretrainedNet (using layerGraph function)
- connect the two graphs using the connectLayers function (here the doc: https://www.mathworks.com/help/deeplearning/ref/connectlayers.html)
- train the network with trainNetwork or with custom traning loops, depending on the network and the flexibility that you need during training.
Hope this helps!
