'trainingOptions' without 'trainNetwork' in deep learning

options = trainingOptions("sgdm", Plots="training-progress")

sets training options. This setting works when training is implemented via:

net = trainNetwork(layers,options)

But if training is implemented via the following instead:

for epoch=1:MaxEpochs
    ...
        [net,velocity] = sgdmupdate(net,grad,velocity);
    ...
end

does 'trainingOptions' still take effect without 'trainNetwork'?

Answer 1

[Walter Roberson]

No. trainNetwork knows to use the options because you pass the options in. sgdmupdate does not permit you to pass in options so there is no way for it to know which option to use.

The options are not stored as part of the network.

[feynman feynman]

thanks so much. Then with 'sgdmupdate' or 'adamupdate', one can only specify training options by normal programming lines of code written by themselves rather than a structured function like 'trainingOptions'?

[Walter Roberson]

sgdmupdate and adamupdate do not have any named options.