Open mdymczyk opened 7 years ago
@mdymczyk @pseudotensor We did discuss about implementing KF as one of the solvers inside h2o4gpu quite sometime back. The more we read and gather info from the experts (in HPC/ML alike), the more we see KF as a generic algorithm.
Let me explain.
Everyone has their own way of data assimilation and as such, there are a myriad of variants of KF’s (Wavelet KF, linear KF, LETKF, etc…) The only common part of all these KF variants is the linear solver which happens to be solved using QR factorization, Cholesky decomp, or probably even explicitly! However, even the decision as to which solver to use for this is very specific to the problem set at hand.
Hence, it sounds like it’ll always be very customer/problem specific. This makes creating a generic “KF library” is almost impossible.
Better way to provide support would be to improve the gpu-enabled solver libraries to better support these use-cases.
Do you guys agree with my understanding so far? Or do you think I’m missing something important?
A question specifically to H2O: when KF was put in H2O’s roadmap, were there any specific use-cases or customers being targeted?
Regards, Thejaswi
@teju85 thanks for the research!
@arnocandel @pseudotensor @IzuiT do you think we could use KF in any capacity for DAI?
@mdymczyk @arnocandel @pseudotensor @IzuiT We currently have a single GPU version of linear KF implemented by one of our interns. He has closely followed filterpy's APIs to implement this, but currently in C++.
Do you guys have any more inputs for us on KF, that can help us further concentrate our efforts?
Awesome. They would make a great way to deal with some time series type of target encoding to see if the filtered version is a good feature, for example.
Being worked on by NVidia.