how to define custom loss function like tf.train.A​​damOptimi​z​er().min​im​ize(los​s) ?

I put X=[4,100] into LSTM and received the last hidden state[4,1]. I use this state vector to estimate Q. I want to use Adam Optimizer to minimize a custom loss function. But options offered by official tool cannot do it. So I want to define a loss function and use Adam Optimizer to min the loss. How can I do it like python:train = tf.train.AdamOptimizer().minimize(loss)?

z=[0.01,0.05,0.1,0.15,0.2,0.25,0.3,0.35,0.4,0.45,0.5,0.55,0.6,0.65,0.7,0.75,0.8,0.85,0.9,0.95,0.99];
z_normal=norminv(z);
numFeatures = 4;
A=4;
X = reshape(X_test.',4,100,[]);
Y = reshape(Y_test.',1,[]);
dlX = dlarray(X,'CBT');
dlY = dlarray(Y,'BT');
numHiddenUnits = 16;
outputdim=4;
H0 = zeros(numHiddenUnits,1);
C0 = zeros(numHiddenUnits,1);
weights = dlarray(randn(4*numHiddenUnits,numFeatures),'CU');
recurrentWeights = dlarray(randn(4*numHiddenUnits,numHiddenUnits),'CU');
bias = dlarray(randn(4*numHiddenUnits,1),'C');
[outputs,hiddenState,cellState] = lstm(dlX,H0,C0,weights,recurrentWeights,bias);
state_h=hiddenState(:,end);
w_out=normrnd(0,1,[outputdim,numHiddenUnits]);
b_out=normrnd(0,1,[outputdim,1]);
out=w_out*state_h+b_out;
params=out+[0;1;1;1];
mu=params(1,:);
sig=params(2,:);
utail=params(3,:);
vtail=params(4,:);
factor1=exp(z_normal.'*utail)/A+1;
factor2=exp(z_normal.'*utail)/A+1;
factor=factor1.*factor2;
Q=factor.*z_normal.'.*sig+mu;
error=dlY-Q;
error1=z.'.*error;
error2=(z.'-1).*error;
loss=mean(max(error1,error2));

[Rik]

You could also have edited the question instead of closing it.

Answer 1

[Sai Bhargav Avula]

Hi,

You can create custom loss function by creating a function of the form loss = myLoss(Y,T), where Y is the network predictions, T are the targets. The loss can be used to update the gradients in the modelGradient function.

This link explains how to create custom layers and loss functions in matlab

https://www.mathworks.com/help/deeplearning/ug/define-custom-training-loops-loss-functions-and-networks.html

Hope this helps!