> I noticed for example that the sqlite package does this. The author forked 
> the library, modified it so that its easier to integrate, and then bound J to 
> it leaving code on both sides to maintain.

Right, though we didn't modify the sqlite code itself, just extended
it to allow J to read and write array arguments to sqlite. Without
that extension, J would have to loop, e.g. record at a time, so this
would slow things down. The only maintenance done is to recompile from
a later sqlite source.

On Fri, Feb 26, 2021 at 12:04 AM Emir U <[email protected]> wrote:
>
> Hey Devon, I think J's stat libraries are a basic start. the other day I 
> wrote an interquartile range function because there wasn't one or a quantile 
> function. Please compare the GNU Scientific list or an ok library like 
> https://www.statsmodels.org/stable/api.html for gaps. PCA, PLS, KDE, MCMC, ME 
> est., multivariate distributions, are bread and butter for my work but none 
> of these things come with J. I personally think that binding a different 
> language (e.g. R) and then calling stats functions from it is at best a stop 
> gap but not a solution in itself. That's for example something that Julia 
> does whilst they hustle to cover all basis. I.e. worst case scenario, call 
> Python from Julia. What happens in practice is that you just end up writing 
> stuff in both languages in order to facilitate the integration which I don't 
> like for maintenance. I noticed for example that the sqlite package does 
> this. The author forked the library, modified it so that its easier to 
> integrate, and then bound J to it leaving code on both sides to maintain.
>
> Hey Bill, I've also used Stan a lot in the last 1.5 years. I've used both the 
> R and Python packages extensively, and tbh I find the integrations pretty 
> clunky. I think part of the reason is that Stan has a DSL so you end up 
> either writing the code in a file and referencing that or having big 
> multi-line strings in which your model is embedded. I personally prefer PyMC3 
> or Nimble type integrations. Just a thought but I think a better way to grow 
> would be to integrate the SGLD and NUTS routines from Tensorflow probability, 
> and then specify likelihoods directly in J if that's possible. However, it 
> means having much better support for commonplace uni/multi variate 
> distributions at the least.
>
> Emir
> ----------------------------------------------------------------------
> For information about J forums see http://www.jsoftware.com/forums.htm
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm

Reply via email to