Michael Krabbe Borregaard
Michael Krabbe Borregaard
I guess that's hard to discuss without a sample distribution and a criterion for what diagnostic aspect of the benchmarking runs we are trying to isolate.
FWIW I still think the arithmetic scale is more diagnostic than the log one in @timholy 's example
I like the natural units because I can pronounce them to myself and they make sense to me - "ah, microseconds, I know what those are", whereas as a non-engineer...
Looks good like that, though. Might still consider microseconds, as that would be easier to talk about ("half a microsecond" or "26 microseconds").
Yes, sounds very sensible. Ah, and what's the current policy for predict? I believe I should not have to worry about it for https://github.com/JuliaStats/DataFrames.jl/pull/1160.
Cool. Yes, e.g. `df = DataFrame(x = x, y = y); mod = lm(y~x, df); plot(df, :y, predict(mod))` would be nice to be able to do even in the presence...
It is a bit tricky perhaps to get the `skipnull = true` and `pairwise_complete=true` combination to be intuitive (would setting `pairwise_complete=true` lead to fewer or more observations to be included?...
I actually recently put a generic `pairwise` function in VectorizedRoutines.jl if that's relevant https://github.com/ChrisRackauckas/VectorizedRoutines.jl/pull/9/files . You'd be welcome to copy over any elements of it (given that I'm pretty sure...
That is a good idea - the function has always been intended to be extended to multiple dimensions. Which design do you think would be preferable - one where `f`...
This looks nice