Parameter not understood in nn.Linear(?, self.width) about neuraloperator HOT 3 CLOSED

Hi @florianwechsung ,
Thanks for the catch! We will update it in the paper. I guess we didn't add the location (x) in the earlier version of the implementation, but later we found it helpful to capture location-wise and non-periodic information.
v_0(x) = P(x, a(x)) is correct. It's also included in the previous paper GNO, Eq 7.

from neuraloperator.

For 1d, the input v_0(x) = (x, a(x)), the location (1d) and the input function.
For 2d, the input v_0(x) = (x, y, a(x,y)), the location (2d) and the input function.

For 2d_rnn, the input v_0(x, t) = (x, y, u(x,y,t-10), u(x,y,t-9), ..., u(x,y,t-1)), the location (2d) and the input function of the first 10 time-steps.
For 3d, the input v_0(x, t) = (x, y, t, u(x,y,1), u(x,y,2), ..., u(x,y,10)), the location (3d) and the input function of the first 10 time-steps.

Hope it helps!

from neuraloperator.

Hi @zongyi-li ,

this definition of the input first layer seems to be different to what you claim to do in the paper, doesn't it? There you write (~line 4 of Section 3), that v_0(x) = P(a(x)). If I understand your comment above correctly, that should actually be
v_0(x) = P(x, a(x)) for 1D example
v_0(x) = P(x, y, a(x, y)) for the 2D example
etc.. Is that correct?

from neuraloperator.