But, given the neighbourhoodstructure in G, elements of C_k are likely close to that of C_ne(k). Use this to create more efficient starting parameters.
This should speed up computations quite a bit.
Proposal, start from:
C_ne(k) already learnt
average of diagonals upper left
a rolling-average of diagonals upper left. In the limit equal to the one above.
Blunde1
commented
8 months ago
Currently we start from the Identity matrix
https://github.com/equinor/graphite-maps/blob/f0a781d4baee6bd842a3948d15731c5ed5bbf8ed/graphite_maps/precision_estimation.py#L261-L263
But, given the neighbourhoodstructure in
G, elements ofC_kare likely close to that ofC_ne(k). Use this to create more efficient starting parameters.This should speed up computations quite a bit.
Proposal, start from:
C_ne(k)already learnt