--mlp-width should have a sensible default

rs-station / careless

Merge X-ray diffraction data with Wilson's priors, variational inference, and metadata

MIT License

16 stars 6 forks source link

--mlp-width should have a sensible default #142

Closed kmdalton closed 9 months ago

kmdalton commented 9 months ago

Automatically setting mlp-width leads to excessive memory when combined with positional encoding. It also makes crossvalidation between different metadata scenarios invalid. I should just change the default --mlp-width to be a sensible value like 10.

kmdalton commented 9 months ago

resolved by #144

DHekstra commented 8 months ago

At the risk of overextrapolating from two examples (lysozyme and DHFR laue anomalous), perhaps the default --mlp-width should be larger, at least until we revisit this?