enforcing positive solution using sigmoid fails.

[yoanncheny]

Dear DeepXDE users,

I'm solving the BCRE equations with deepxde. The solutions of this problem are under the form h(x,t) and R(x,t) which are positive or zero.
To help the training and avoiding unrelevant solutions I aply a sigmoid function for the output layer :

`net = dde.nn.FNN([2] + 4 * [50] + [2], ["tanh","tanh","tanh","tanh","sigmoid"], "Glorot normal")

model = dde.Model(data, net)

model.compile("adam", lr=1e-3)
losshistory, train_state =model.train(iterations=500)`

However after training the solution has negative values up to -0.3.

Has anyone already faced such an issue ?
Regards,
Yoann

[haison19952013]

Have you tried apply_output_transform() ?, I suggest using Softplus instead of sigmoid.

[yoanncheny]

Thx for your answer ! Indeed, by using apply_output_transform() the solution remains positive.
However I'm a bit confused : why did I get negative values by applying the sigmoid function to the output layer ? Is this some sort of bug ?

[yoanncheny]

The solution proposed by @haison19952013 allows me to get a positive solution. However the solution is totally wrong. Previously it was rather ok except for the unphysical negative values. Has anyone got some hints for understanding this behaviour ?