dlgradients returning zeros, loss dependant of other already trained network

Question

Marcos 2024-1-31

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/2076566-dlgradients-returning-zeros-loss-dependant-of-other-already-trained-network

评论： Marcos 2024-2-12

Hello,

I'm facing a problem that might be similar to the one here https://uk.mathworks.com/matlabcentral/answers/1844083-why-happens-all-the-gradients-of-the-generator-are-zero-from-the-beginning-to-the-end-when-traini, but I don't understand why.

The loss I use depends on another already trained shallow network. Then I use dlgradient and the calculated loss to find the gradients, but this only returns zeros...

Here's the loss,gradients function:

function [loss,gradients] = modelLoss(dlnet,X,net2)
% Forward data through network.
[Y] = forward(dlnet,X);
% Data through trained network2.
X_2 = [extractdata(X(1:2,:));extractdata(Y)];
X_pred = net2(X_2);
% Convert to dlarray.
dlX_pred = dlarray(X_pred,'CB');
% Calculate loss.
loss = mean(mean((dlX_pred - X(3:end,:)).^2));
% Calculate gradients of loss with respect to learnable parameters.
gradients = dlgradient(loss,dlnet.Learnables);
end

And here is the use of the dlfeval:

[loss,gradients] = dlfeval(@modelLoss,dlnet,dLXMiniBatch,net2);

Any idea on what's missing?

Thanks!

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Avadhoot 2024-2-12

1
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/2076566-dlgradients-returning-zeros-loss-dependant-of-other-already-trained-network#answer_1406866

Hi Marcos,

I see that you are calculating gradients for your custom loss function for your model. As mentioned in the example, you have correctly included all the calculations involved in the loss computation inside the loss function, including the pretrained network. Still there could be a few issues which might cause the gradient function to return 0. Below are the probable fixes for this issue:

Make sure that you are not updating the pretrained network weights in calculating the loss, as this can cause a problem in the computation.
The gradients might not get calculated if the pretrained network includes a non-differentiable step.
The dlnet.Learnables should contain all the parameters that you want to update. Check if all the relevant parameters are included in it.
X and Y need to be "dlarray" objects for the gradient function to work. If not, please convert them to dlarray objects before passing them to the loss function.

As you mentioned that you are using a shallow network, the vanishing gradient problem should not trouble you. Please check on all the above factors to determine the cause of the problem.

I hope this helps.

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

Marcos 2024-2-12

Hi Avadhoot,

Thanks for your response.

I managed to solve my problem by also using a dlnetwork approach for net2 instead of using feedforwardnet. Y was a dlarray and I tried converting the X_pred to dlarray in the loss function. I guess that was the problem then.

Anyway, having two dlnetwork objects (a pretrained one and the one being trained) isn't a problem, so I'm using that and it seems to be working :)

Thanks again!

请先登录，再进行评论。

dlgradients returning zeros, loss dependant of other already trained network

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

采纳的回答

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

更多回答（0 个）

另请参阅

类别

标签

产品

版本

Community Treasure Hunt

dlgradients returning zeros, loss dependant of other already trained network

0 个评论 显示 -2更早的评论隐藏 -2更早的评论

采纳的回答

1 个评论 显示 -1更早的评论隐藏 -1更早的评论

更多回答（0 个）

另请参阅

类别

标签

产品

版本

Community Treasure Hunt

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

1 个评论
显示 -1更早的评论隐藏 -1更早的评论