How can I define a custom loss function using trainnet?

Hello,

I am trying to define a custom loss function using trainnet. The documentation says:

If the trainnet function does not provide the loss function that you need for your task, then you can specify a custom loss function to the trainnet as a function handle. The function must have the syntax loss = f(Y,T), where Y and T are the predictions and targets, respectively.

However, I am not sure how the predictions and targets are defined here. I am currently using trainnet as follows:

trainedNet = trainnet(dsTrain,layers,"mse",options);

dsTrain is a datastore containing the input and target images for the regression problem. But I would like change the loss to a custom function involving ssim. I would like something similar to the following, although, I know this isn't quite right:

trainedNet = trainnet(dsTrain,layers,@(Y,targets) 1-ssim(Y,targets),options);

I get the following errror message:

Error using trainnet

Value to differentiate is non-scalar. It must be a traced real dlarray scalar.

Thanks!

Answer 1

[Matt J]

If you have multichannel output, the loss function will give you an SSIM per channel, e..g,

loss = @(Y,targets) 1-ssim(Y,targets);

[Y,T]=deal(dlarray(rand(5,4,8),'SSC'));

L=loss(Y,T);

whos L

  Name      Size             Bytes  Class      Attributes

  L         1x1x8               70  dlarray              

You need to decide how you want this reduced to a single value.