Problem in automatic gradient calculation using Deep Learning Toolbox

Question

GAURAV YADAV 2021-9-8

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/1449384-problem-in-automatic-gradient-calculation-using-deep-learning-toolbox

评论： GAURAV YADAV 2021-9-9

Hello,

Could anyone please tell me why the dlgradient function is throwing the error? It seems to me that somehow the loss is not getting traced to the input weights? Could anyone suggest why it is happening?

x0 = dlarray([1,2,3],'BC');
y0 = dlarray(14,'BC');
fcnn_graph = layerGraph;
layers = [featureInputLayer(3,"Normalization","none","Name","InputLayer")
          fullyConnectedLayer(5,"WeightsInitializer","glorot","BiasInitializer","ones","Name","fc1")
          tanhLayer("Name","active_1")
          fullyConnectedLayer(5,"WeightsInitializer","glorot","BiasInitializer","ones","Name","fc2")
          tanhLayer("Name","active_2")
          fullyConnectedLayer(1,"WeightsInitializer","glorot","BiasInitializer","ones","Name","fc3")
          ];
fcnn_graph = addLayers(fcnn_graph,layers);
dlnet = dlnetwork(fcnn_graph);
[gradients,state,loss] = modelGradients(dlnet,x0,y0);
function [gradients,state,loss] = modelGradients(dlnet,X,Y)
    [YPred,state] = forward(dlnet,X);
    loss = crossentropy(YPred,Y);
    gradients = dlgradient(loss,dlnet.Learnables);
    loss = double(gather(extractdata(loss)));
end