How to get the gradient with respect to the output of a specific layer in a deep learning model?

3 次查看（过去 30 天）
SC
SC 2019-11-17
回答： Athul Prakash 2019-11-25
Hi,
I have a model myModel which is a simple 2 Layers model (i.e. input size: 3; Layer_1 size: 7; Layer_2 size: 4).
That is, the layer's output is a and the loss is .
From below code, now I have the whose dimention is 60 since there are 60 parameters in the model (i.e. (3+1)*7+(7+1)*4 = 60). Nis the number of data points (i.e. in my code)
But want I want is the gradient with respect to the layer 2's output, instead of the loss. That is, I want instead, whose dimention would be since .
How can I get ? Many thanks!
My code
rng(123); % seed
X_ori=[4,163,80;5,164,75]; % data; #(number) = 2; #(features) = 3;
X=permute(X_ori,[3,4,2,1]);
dlX = dlarray(X, 'SSCB');
Y_ori=[0, 0, 0, 1; 0, 1, 0, 0]; % data labels (i.e. one-hot vectors for 4 classes)
myModel = [
imageInputLayer([1 1 3],'Normalization','none','Name','in')
fullyConnectedLayer(7,'Name','Layer 1')
fullyConnectedLayer(4,'Name','Layer 2')];
MyLGraph = layerGraph(myModel);
myDLnet = dlnetwork(MyLGraph);
gradients = dlfeval(@modelGradients, myDLnet, dlX, Y_ori);
function [gradients] = modelGradients(myModel, modelInput, CorrectLabels)
CorrectLabels_transpose=transpose(CorrectLabels);
[modelOutput,state] = forward(myModel,modelInput);
loss = -sum(sum(CorrectLabels_transpose.*log(sigmoid(modelOutput/100))));
gradients = dlgradient(loss, myModel.Learnables);
end
Thanks

请先登录，再回答此问题。

回答（1 个）

Athul Prakash
Athul Prakash 2019-11-25
You could try looking at the source code for the fullyConnectedLayer that you're using; and then customizing that code to return the gradients you're looking for.
Alternatively, try creating your own custom layer to achieve this. See the doc below for that:
https://www.mathworks.com/help/deeplearning/ug/define-custom-deep-learning-layers.html
Hope it Helps!

类别

在 Help Center 和 File Exchange 中查找有关 Deep Learning Toolbox 的更多信息

标签

产品


版本

R2019b

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by