jonathanschmidt1 / dsky-differentiable-spin-chain-on-skyrmion Goto Github PK

The LearnerND class in the python-adaptive package seems to use a default loss function which only checks for the volume of the simplex. Therefore, we have to implement a good loss function regarding a gradient estimate to define the distance to the real value as the loss.

The adaptive mathematical function calculator is quite GRAM intensive. We can maximally run 32 Processes on the DGX2 with 16 spins. For the 4D parameter space this isn't acceptable because we would like to expand to 20 spins as well to see finite size effects.
Furthermore, the adaptive module doesn't need autograd and therefore we could switch to different gpu implementations like slepc.

the calculation finishes but scalfac and delta parameters are not changing throughout optimization. plotting of parameters does not behave correctly, for example B_ext plot contains two curves instead of one.

When optimizing for localization with a trivial magnetic field (symmetric polar angles), after some hundred optimization steps hte davidson algorithm fails during cholesky decomposition: U(360,360) is zero, singular U.
This happened with a linear weight list [7,6,5,4,3,2,1,14,1,2,3,4,5,6,7]