Training Network stopping automatically after 3 iteration without showing any error.

tspan = 0:0.001:10;
y0 = 0;
[t,y] = ode45(@(t,y) t^2+2, tspan, y0);
T=t(1:0.9*end)
Y=y(1:0.9*end)
x=t(0.9*end+1:end)
v=y(1+0.9*end:end)

layer = functionLayer((@(X) X./(1 -X^2)))
layers = [
    sequenceInputLayer(1)
    fullyConnectedLayer(1)
    tanhLayer
    functionLayer(((@(t) t./(1 -t.^2))),Description="softsign")
    fullyConnectedLayer(1)
    tanhLayer
    functionLayer(((@(t) t./(1 -t.^2))),Description="softsign")
    regressionLayer]
options = trainingOptions('adam', ...
    'LearnRateSchedule','piecewise', ...
    'LearnRateDropFactor',0.2, ...
    'LearnRateDropPeriod',5, ...
    'miniBatchSize',20,.....
    'VerboseFrequency',1,...
    'ValidationPatience',Inf,...
    'MaxEpochs',100, ...
    'Plots','training-progress')

net = trainNetwork(T',Y',layers,options);
ypre=predict(net,tspan);
plot(ypre)
plot(y)

Answer 1

[Prateek Rai]

Hi,

Training of the network stopped because training loss is NaN. This implies that the predictions using the output network might contain NaN values.

On analyzing network, I found that size of the all the layers is 1*1*1 which is why NaN values are coming.

You might want to recheck the dimension of the layers of the network using:

analyzeNetwork(layers)

[Image Analyst]

I get the same error trying to train on 448 images and my layers are not 1*1*1 -- they're 227x227x3