As discussed, the simple (3,3) Conv stencil to mimick the second-order 2D laplacian won't produce the desired stencil coefficients, regardless of optimizer, training strategy and training duration.
Normally, with an amount of parameters this small, getting the coefficients shouldn't be much of an issue. However, the optimizer doesn't seem to be able to update the stencil coefficients at all (in a meaningful manner), given the setup of the input data.
As this is the baseline for further experiments, this needs to be fixed.
As discussed, the simple
(3,3)Conv stencil to mimick the second-order 2D laplacian won't produce the desired stencil coefficients, regardless of optimizer, training strategy and training duration.Normally, with an amount of parameters this small, getting the coefficients shouldn't be much of an issue. However, the optimizer doesn't seem to be able to update the stencil coefficients at all (in a meaningful manner), given the setup of the input data.
As this is the baseline for further experiments, this needs to be fixed.