The gradient of mini batches

Hi there.

I need your confimation or rejection for this question...

In following code, if the minibatch size is h,

[grad,loss] = dlfeval(@modelGradients,dlnet,dlX_miniBatch,Y_miniBatch);

the grad is the average of gradients of loss over h samples? Does it calculate dradients automatically and at the end with:

grad = 1/h * sum_i=1:h (\nabla loss(y_i,yHat_i)) ??

Following this question, for computing the total loss and geadient (for a full batch), does we should take avarage of losses and averages of gradients (averaging with the number of batches, say 1000 batches each with h size)??

Answer 1

[Mahesh Taparia]

Hi

The function dlfeval evaluate the custom deep learning models. The loss are calculated based on what has been defined in modelGradients function. So if you are calculating the average loss in this function, then it will return the averaged one. For example, consider this modelGradient function, it is calculating the average cross entropy loss, so it will return the average loss. The gradients are calculated with respect to the loss function defined in for the network.

[MAHSA YOUSEFI]

In the example you mentioned, there is a mistake.

function [gradients, loss] = modelGradients(parameters, dlX, T)

    % Forward data through the model function.
    dlY = model(parameters,dlX);

    % Compute loss.
    loss = crossentropy(dlX,T);

    % Compute gradients.
    gradients = dlgradient(loss,parameters);

end

dlY must be feed to crossentropy!

[Mahesh Taparia]

Yeah, crossentropy loss will be calculated between dlY and T. The documentation page will be updated.