So I'm finding the Mcdm only field performs incredibly well, even when only on 1k maps. Perhaps because the filters are able to optimise to that specific field. Currently when we extend to multifield analysis, we are using the same wavelets for each. Allowing a different set of filters for each field within a single model could therefore boost performance whilst only increasing the overall number of parameters by a very small amount. It would also be interesting to see how the filters for each field evolve differently.
One consideration though would be memory, I think the filters are fairly memory intensive, as currently our scattering models take up significantly more memory than the CNN. I guess this is due to each filter being stored as a 256x256 grid, so would have to consider this if we go this route though.
So I'm finding the Mcdm only field performs incredibly well, even when only on 1k maps. Perhaps because the filters are able to optimise to that specific field. Currently when we extend to multifield analysis, we are using the same wavelets for each. Allowing a different set of filters for each field within a single model could therefore boost performance whilst only increasing the overall number of parameters by a very small amount. It would also be interesting to see how the filters for each field evolve differently.
One consideration though would be memory, I think the filters are fairly memory intensive, as currently our scattering models take up significantly more memory than the CNN. I guess this is due to each filter being stored as a 256x256 grid, so would have to consider this if we go this route though.