loss_weights order in model.compile

[aoguedao]

Hi everyone!

First of all, thank you very much for this package! I've been spend the last weeks studying it and it is really amazing.

I have a question that maybe could help more people in the near future. The documentation about model.compile doesn't say anything how is the order or how many elements need to have the list received as argument of loss_weights. Intuitively, I would say first should be the boundary conditions in the same order as defined and then the elements associated to the PDE residual loss. Am I right?

For example, the forward second order ODE utilizes loss_weights=[0.01,1,1] since the function 'ode' just return a single element and it has two boundary conditions.

Same idea, the inverse Poisson equation has loss_weights=[1, 100, 1000], does this mean the first boundary condition has a weight of 1, the second boundary condition (observer) has a weight of 100 and the pde loss a weight of 1000?

Thanks in advance for your help.

[forxltk]

No, the boundary loss is always behind the pde loss.

[lululxvi]

Yes, it was wrong. See the updated doc.

[aoguedao]

Thank you so much. Everything is clear now!

[jmtucny]

Has the doc being updated ? I have the same question, and I still see the following text on the website for the forward Second order ODE :

"Now, we have the PDE problem and the network. We build a Model, choose the optimizer, set the learning rate to 0.001, and train the network for 15000 iterations. We set the weight of the ODE loss to 0.01, and the weights of the two ICs to 1. We also compute the L2
relative error as a metric during training."

With the following code:

model = dde.Model(data, net)
model.compile("adam", lr=.001, loss_weights=[0.01, 1, 1], metrics=["l2 relative error"])
losshistory, train_state = model.train(iterations=15000)

Thanks in advance,

[lululxvi]

It is updated.

[jmtucny]

Yes, I was confused while reading the documentation. Thank you !