for the first one racket 20 haskell 10 for the second one both 10

20 what? 10 what?

sorry data-sets

one initialization per data set

I will double check if haskell is doing the same

yes once per data set for 10 datasets and averaged accross all of them

I see, so I would try two things (which can be done in parallel): average more runs, and make sure that Haskell and Racket are computing the same numbers up to rounding error

I verified rkt and hs without categorical, they are exactly same upto 0.0000000000000001

I verified categorical seperatly by running it 10million times and comparing probability, next I will compare the outer loop of gibbs

@rjnw Hmm, so what changed between your figures at 2:17PM and 5:26PM, other than jags being gone? Why is racket at 5:26PM no longer doing worse?

(Thanks for checking numerical equality.)

I will add jags later today as this was done outside docker. There was a bug in normalizing the array before calling categorical in my code.

I had checks for everything except that :persevere:

Yay, so did you do the “divide (i.e., subtract) by max” thing in categorical?

The latest plot is promising.