I’ve got a turing model where there are a few parameters for each of a couple hundred objects, plus some structural parameters… so let’s call it close to 500 parameters. I have a friend who has a simpler model (similar number of parameters, but expected to be less accurate but easier to fit) which …

[image] dlakelan: I have a friend who has a simpler model (similar number of parameters, but expected to be less accurate but easier to fit) which fits in Stan, and he can get 100,000 samples in less than a couple minutes. (I told him it was crazy overkill but he figured why not, he’s doing thin…

I have no idea what that actually means. What should I be trying? Here’s what I’m doing: Turing.setadbackend(:reversediff) Turing.setrdcache(true) op = optimize(mymodel, MAP(), Optim.ParticleSwarm(n_particles=10), Optim.Options(iterations=2000)) ch = sam…

Ok, here’s some serious comparisons. I implemented exactly the same model as my friend. Literally identical in every way, the priors, the likelihood everything. He runs his model in stan, it has 400+ parameters, it runs in: Chain 1: Iteration: 1600 / 2000 [ 80%] (Sampling) Chain 1: Iteration: 180…

I imagine it will be tricky to resolve without a MWE of code that’s very slow on Turing but (equivalent code is) fast with Stan.

@EvoArt , yeah, I’m going to try to make some minimal thing that will fit fast in stan and slow in Turing. I can’t use this actual model. I’m guessing something as simple as fitting 300 independent normal distributions to 4-5 samples each. Will try to post this weekend. @ToucheSir , thanks I will tr…

Ok I reran with one thread, and got the following: Subroutines, sorted by exclusive sample count Exclusive Inclusive Subroutine 28422 (79.45%) 28422 (79.45%) _methods_by_ftype 1043 (2.92%) 13677 (38.23%) #step#17 905 (2.53%) 4156 (11.62%) GradientTape 768 (2.15%) 14862 (41.54%) mcmcsam…

OK, I discovered that after reading my friend’s CSV file, it has some missing values in some columns… Even though I’m subsetting the non-missing values in a new DataFrame before passing in the data, the TYPE of the column is still Union{Missing,Int64}. Since I’m modeling those columns as normally di…

Stan dev here and I’m quite interested in Turing and what Julia can do. Looking forward to seeing how this gets resolved. Stan has profile capabilities since 2.26. It’s not quite as easy as the above though. See Profiling Stan programs with CmdStanR • cmdstanr

[image] spinkney: Stan has profile capabilities since 2.26 :+1: Glad the Stan devs are over here and glad to hear you can profile Stan these days :slight_smile: Julia has a lot of promise and while i love what Stan can do, the two language problem is real and the ability of Turing to do mix…

Making Turing Fast with large numbers of parameters?

General Usage

ToucheSir October 2, 2021, 6:51pm 6

Another instance of Profiling code reveals that most time is used calling 3 functions in task.jl - #15 by n8xm?

Topic		Replies	Views
Improving performance of item response model in Turing.jl Performance turing	23	2500	January 18, 2023
Turing: help with slow model Probabilistic Programming	46	2978	February 13, 2021
Case study: Speeding up a logistic regression with RHS prior (Turing vs Numpyro) - any tricks I'm missing? Probabilistic Programming	38	3592	February 6, 2023
Speed up multilevel model in Turing.jl Probabilistic Programming turing	15	607	June 17, 2023
Slow Turing.jl sampling compared to python pymc Probabilistic Programming turing	7	300	June 26, 2025

Making Turing Fast with large numbers of parameters?

Related topics