"case.weights" take very long about ranger

[mnwright]

I just released a version (0.4.2) based on the new toolchain. As reported, the problem is solved there. In addition, multithreading is finally working! This version can also be installed on the current R version by using the binary, see https://github.com/imbs-hl/ranger/releases.

I hope it's solved with R-3.3.0!

[mnwright]

No this is not as expected. I can reproduce the issue on Windows but not on Mac or Linux. I will check the code for some Windows-specific problems.

[mnwright]

The problem seems to be std::discrete_distribution<> with gcc 4.6.3. I tried with the new 4.9.3 toolchain and R-devel and it was fast.

Any idea how to solve this instead of waiting for a newer gcc?

[khotilov]

Using boost::random::discrete_distribution as a replacement helps:
before:

> system.time(fit.1 <- ranger(y ~ x)) 
   user  system elapsed 
   9.27    0.13    9.41 
> system.time(fit.3 <- ranger(y ~ x, case.weights = rep(1, times = n))) 
   user  system elapsed 
  93.02    0.07   93.19 

after:

> system.time(fit.1 <- ranger(y ~ x)) 
   user  system elapsed 
   8.76    0.16    8.96 
> system.time(fit.3 <- ranger(y ~ x, case.weights = rep(1, times = n))) 
   user  system elapsed 
   8.98    0.09    9.09 

[mnwright]

Thanks! However I'm reluctant to merge it in the master because of the Boost dependency... ;)

[khotilov]

That is a temporary simple solution while waiting for a newer gcc. I didn't do extensive testing, but a quick check showed very similar model performance (see below). That should make it at least feasible for me to run some prototyping with ranger on my windows laptop, as I frequently need to use weights. And the real dependency is only for the windows R version, which is already a neglected child with no multithreading :)

# with the original std::discrete_distribution
set.seed(111)
fit_std <- ranger(y ~ x, case.weights = rep(1, times = n), write.forest=T)
pr_std <- predict(fit_std, data.frame(x = x))

# with boost::random::discrete_distribution
set.seed(111)
fit_boost <- ranger(y ~ x, case.weights = rep(1, times = n), write.forest=T)
pr_boost <- predict(fit_boost, data.frame(x = x))

cor(pr_std$predictions, pr_boost$predictions)
[1] 0.9979446

The gcc's <random> was based on boost. But some over-engineering resulted in overheads and worse speed - I've seen a few discussions about that in the past. It wasn't just the discrete_distribution, but some other distributions too were several times slower. Maybe things did significantly improve in this regard in the latest releases (I didn't really follow), but I personally had more trust in boost::random.

It's your choice in the end. I'm just telling you what I know. I'm glad I've noticed this discussion, since my initial observations didn't agree with the claims of ranger being very fast, so I didn't even try it on a linux server.

[mayer79]

This is brilliant, thank you very much for these investigations. Even on the current R version, the issue seems to be fixed with ranger 0.4.2. Wow!