Max_treedepth not adhered to in rstan?

Interfaces RStan
1

I just noticed something funny while running rstan 2.19.2. I have a model that was getting the max treedepth warning, so I ran rstan::sampling(..., control = list(max_treedepth=15)). When the model finished sampling I didn’t get any warning, but when I ran rstan::check_hmc_diagnostics, it once again showed me the treedepth warning. I took a look at the samples csv and it indeed had # max_treedepth=15 in the header, but when I run:

sampler_params <- get_sampler_params(post, inc_warmup = FALSE)
sampler_params_chain1 <- sampler_params[[1]]
max_treedepth_by_chain <- sapply(sampler_params, function(x) max(x[, "treedepth__"]))
print(max_treedepth_by_chain)

It shows that for each chain 10 was the max treedepth. Is this all expected behaviour?

1 Like
2

Oh! If I run:

sampler_params <- get_sampler_params(post, inc_warmup = TRUE)
sampler_params_chain1 <- sampler_params[[1]]
max_treedepth_by_chain <- sapply(sampler_params, function(x) max(x[, "treedepth__"]))
print(max_treedepth_by_chain)

(n.b. inc_warmup = TRUE), it shows max treedepths greater than 10. So does the max_treedepth parameter only apply to the warmup?

3

The limit applies to all the iterations, but you do not get warnings for exceeding the max_treedepth during warmup. Nor do you get divergent transition warnings during warmup. Those almost always happen when you move from one window to the next.

1 Like
4

Ok, then is it not odd that all my treedepths max out at 10 after warmup?

5

Not with control = list(max_treedepth = 15)

6

I had control=list(max_treedepth=15) yet observed that after warmup no iterations went above 10. Before warmup some did, but after warmup none did.

7

Yeah, that is fine. And you could have gotten by with max_treedepth of 11.

1 Like