DLNETWORK STATE IS ALWAYS A 0 TABLE.

I am implemeting neural network, but when I define my Dlnetwork the State is always a 0 table, also during the training process in fact after the training my nerual structure isn't learning anything:

Below I show only a part of my code

%GENERATOR

lgraph_G = layerGraph;

Generator = [

sequenceInputLayer(30,"Name","sequence")

fullyConnectedLayer(512,"Name","DENSE_G1")

reluLayer("Name","relu_1")

fullyConnectedLayer(512,"Name","DENSE_G2")

reluLayer("Name","relu_2")

fullyConnectedLayer(512,"Name","DENSE_G4")

reluLayer("Name","relu_3")

fullyConnectedLayer(1,"Name","DENSE_G5")

];

lgraph_G = addLayers(lgraph_G,Generator);

netG = dlnetwork(Generator);

%DISCRIMINATORE

Discriminator = layerGraph(sequenceInputLayer(2,"Name","sequence"));

combined_D = [

fullyConnectedLayer(100,"Name","fc")

leakyReluLayer(0.2)

fullyConnectedLayer(100,"Name","fc_1")

leakyReluLayer(0.2)

fullyConnectedLayer(1,"Name","fc_2")

softplusLayer("Name","softplus")

];

Discriminator = addLayers(Discriminator,combined_D);

Discriminator = connectLayers(Discriminator,'sequence','fc');

Discriminator = dlnetwork(Discriminator);

combinde_D = layerGraph(combined_D);

lgraph_D = layerGraph(Discriminator);

netD = dlnetwork(lgraph_D);

Answer 1

[Ben]

This network does not have any layers with state parameters. The learnable parameters are in the netG.Learnables and netD.Learnables properties.

Some examples of layers with state are lstmLayer (which needs to pass its state from each sequence element to the next) or batchNormalizationLayer (which needs to pass its state from one minibatch to another).

You can define custom layers with state if you need to pass state information from the layer from one forward pass to another, see the custom layer documentation.