A step of the momento_native implementation. No other function in the package calls it. See the file header for the source the module follows. the source it follows.
Examples
cat(momento_cheatsheet())
#> momento: masked time-series pretraining. Mask patches with ZEROS and reconstruct; the loss counts the MASKED positions only, since scoring visible ones rewards copying. The hard part is multi-dataset pretraining: series differ in resolution, channel count, length and amplitude, so harmonise per-series and keep channels independent. Mask rate is a real knob -- too low is interpolation, too high leaves no context. Task changes only WHERE the mask goes: tail for forecasting, interior for imputation.