What is target in the losses function?

[cam1681]

Hello, thanks for your wonderful code, especially the efforts for PyTorch!

What is the variable "target" in the "losses" function? I tested the Possion_Lshape.py and print the "targets" in the losses function and I found that it is always None. So, how does the target plays in the losses? If I want the residual not the loss, i.e., the vector of [residualPDE, boundary_difference], how can I modify the code?

Thanks again!

[cam1681]

It seems that the answer is buried in the data/pde.py where losses function is defined. error_f and error on the boundary give the residual vector and the difference on the boundary. But why the size of the tensor in the inputs seems the num_domain+ 2*num_doundary?

def losses(self, targets, outputs, loss, model):
f = []
if self.pde is not None:
if get_num_args(self.pde) == 2:
# can be optimized by not putting boundary points into pde
f = self.pde(model.net.inputs, outputs)
elif get_num_args(self.pde) == 3:
if self.auxiliary_var_fn is None:
raise ValueError("Auxiliary variable function not defined.")
f = self.pde(model.net.inputs, outputs, model.net.auxiliary_vars)
if not isinstance(f, (list, tuple)):
f = [f]

[lululxvi]

The points on the boundary are used twice: once of the PDE residual, and one for the boundary loss. see #39

[cam1681]

The points on the boundary are used twice: once of the PDE residual, and one for the boundary loss.

Thanks! I have another question. Are the training points randomly chosen at each iteration when using L-BFGS or Adam? If not, how could I modify the code to randomly sampling training data in the domain?

[lululxvi]

You can use dde.callbacks.PDEResidualResampler for point resampling in each every iterations, see https://github.com/lululxvi/deepxde/blob/master/examples/diffusion_1d_resample.py

[cam1681]

Thank you very much for your reply, I have got what I want.