Open jgabry opened 1 year ago
Yes
Agreed. We can remove them from both v2.26 & the experimental branch. Though, not sure if any dependency is using them right now. This won't affect the release of StanHeaders
v2.26.
I think there may be a few dependencies using them, but not many. How about just a deprecation warning at first and then removing them entirely at some point in the future? If that sounds good I can make a PR against the experimental branch adding the deprecation warnings.
@andrjohns Can you work on this for the experimental branch? I'll backport to v2.26 and run reverse dependency checks.
I think there may be a few dependencies using them, but not many. How about just a deprecation warning at first and then removing them entirely at some point in the future? If that sounds good I can make a PR against the experimental branch adding the deprecation warnings.
That works, but we could also depend on bayesplot
or create aliases (if bayesplot
is installed use it; otherwise, print a deprecation warning).
I think there may be a few dependencies using them
Alternatively, we could remove them completely and contact the maintainers to switch to bayesplot
.
I vote for removing them completely, if the number of dependencies is small. We can easily fix it for them.
That's true, although I think there are a lot of users who got accustomed to using the plotting functions in RStan and giving them one release with a message pointing them to bayesplot would be nice (instead of them just getting an error from R that the function doesn't exist). @bgoodri and @andrjohns what do you guys think?
If it's going to require fixing other package dependencies, I'd prefer to wait until 2.31 is on CRAN before removing. At the moment the submission path for 2.31 is relatively clear, with all dependencies having patches either submitted or released to CRAN. I'd rather not add an additional source of delay/breakage on that side of things
@andrjohns That makes sense. Are you ok if I add deprecation warnings? That way everything continues to work (no breaking dependencies) but we warn them of the change coming in the future?
I'm on board with deprecation warnings
I can make a PR. Is the right place to do that the experimental
branch?
I can make a PR. Is the right place to do that the
experimental
branch?
Yes. That's the target for v2.32 after releasing StanHeaders
v2.26; we can jump directly to the latest version of rstan
since StanHeaders
now includes a compatible version of stanc3
.
It occurs to me that the one thing we lose by removing these functions from rstan is the ability to automatically unconstrain parameters if the user wants (I think this is why we haven't already deprecated them, if I remember previous discussion accurately). bayesplot can't unconstrain the parameters automatically, so the user would have to unconstrain and then pass to bayesplot if they want plots of unconstrained parameters.
Are we ok with that or is that sufficient reason to keep plots in rstan itself?
Could the rstan
functions instead just be a wrapper that performs the unconstraining and then calls bayesplot
? It could also be something that gets built into bayesplot
The idea being not to remove existing functionality, in case it's needed in a workflow somewhere
Could the
rstan
functions instead just be a wrapper that performs the unconstraining and then callsbayesplot
? It could also be something that gets built intobayesplot
Maybe the first option. I think if we can keep bayesplot as separate as possible and not call rstan::unconstrain_pars inside bayesplot that would be ideal.
This is only really an issue for complicated transformations. For a standard deviation, for example, it's easy to get bayesplot to do the transformation:
bayesplot::mcmc_hist(x, transformations = list(sigma = "log"))
but you have to know the function to specify. For more complicated transformations it would be tough for the user to do this without relying on unconstrain_pars. So yeah maybe we should do it in rstan and then call bayesplot instead of deprecating the rstan functions. I'll look into that.
@andrjohns What's the best way for me to build the experimental branch locally? Seems like I need StanHeaders 2.31.0, right?
I either just devtools::install()
the StanHeaders
subdir and then the rstan/rstan
subdir, or use the pre-built package source from any of the recent actions runs on the experimental branch
(and yeah I normally use StanHeaders 2.31 with it as well)
Ok thanks. I’ll give that a try later (got to hop on a few zoom calls first)
On Thu, Mar 16, 2023 at 1:00 PM Andrew Johnson @.***> wrote:
(and yeah I normally use StanHeaders 2.31 with it as well)
— Reply to this email directly, view it on GitHub https://github.com/stan-dev/rstan/issues/1051#issuecomment-1472586858, or unsubscribe https://github.com/notifications/unsubscribe-auth/AB3PQQ5AMRVBP3JPKCQELBDW4NPNPANCNFSM6AAAAAAV5RFEJQ . You are receiving this because you authored the thread.Message ID: @.***>
@andrjohns I successfully built the experimental branch, and it's a good thing I tried it because I think the existing code for plotting unconstrained parameters breaks, potentially due to changes in the generated C++ code in newer versions of Stan.
In the past this (admittedly hacky) helper function was able to get the names of the variables declared in the parameters block:
(I think it would get data + parameters and then drop the data by intersecting with model_pars
)
But now it seems that grepping for context__.vals_r
in the C++ code only returns variables in the data block not the parameters block. This basically breaks the plotting code for unconstrained parameters (we were using this to provide a way to unconstrained parameters automatically for the user, for which we needed to know if the parameter the user asked to plot was in the parameters block -- if it was in generated quantities, for example, we couldn't unconstrain it).
Do you know of a way to get the names of variables just in the parameters block that I could use instead of this?
Ooh good question. Wouldn't it be easier to grep the Stan code in the stanfit object for the entries of the parameters
block? Then it wouldn't be dependent on the C++
I think there was some reason we didn’t do that before but I can’t remember and now I can’t think of a good reason not to do that, so maybe I’ll give that a try.
On Thu, Mar 16, 2023 at 3:15 PM Andrew Johnson @.***> wrote:
Ooh good question. Wouldn't it be easier to grep the Stan code in the stanfit object for the entries of the parameters block? Then it wouldn't be dependent on the C++
— Reply to this email directly, view it on GitHub https://github.com/stan-dev/rstan/issues/1051#issuecomment-1472750998, or unsubscribe https://github.com/notifications/unsubscribe-auth/AB3PQQ5UBOAHJKOQEPT6QPDW4N7IPANCNFSM6AAAAAAV5RFEJQ . You are receiving this because you authored the thread.Message ID: @.***>
Is there a way to run stanc3's auto formatter (I don't see it as an option to stanc
)? I think one reason we didn't grep the Stan code is that we'd need to find the parameters block but the user could have the word "parameters" in comments. The block doesn't even have to start on a new line due to Stan's whitespace rules, so we can't just search for "parameters" starting a line. There may be other issues along those lines too. (I think the auto formatter removes comments if I recall correctly).
Is there a way to run stanc3's auto formatter (I don't see it as an option to
stanc
)?
The auto-formatting option generates Stan code while stanc()
generates C++ code. To run the formatter internally, use auto_format = TRUE
in stanc_process
or set stanc.auto_format
option to TRUE
.
Is there a way to run stanc3's auto formatter (I don't see it as an option to
stanc
)?The auto-formatting option generates Stan code while
stanc()
generates C++ code. To run the formatter internally, useauto_format = TRUE
instanc_process
or setstanc.auto_format
option toTRUE
.
Thanks! I'll try stanc_process
. I think I'll need to take the user's stanfit object that they pass to the plotting function, grab the stan code, and feed it into stanc_process
with auto_format = TRUE
. That seems like the only way to get an auto formatted version of the user's existing Stan program?
The easiest way is to set the option:
options(stanc.auto_format = TRUE)
We could enable it by default (which would help brms
; see https://github.com/paul-buerkner/brms/issues/1376) or pass auto_format
argument to stanc()
, if needed.
But actually I guess auto formatting doesn't remove user code comments (so the word "parameters" can appear anywhere in the Stan file even after auto formatting). I need a way to figure out which variables are declared in the parameters block (as opposed to transformed parameters and generated quantities). @andrjohns suggested grepping the Stan code itself but that seems potentially error prone if the word "parameters" can appear anywhere in user comments.
You could use a C++ preprocessor to remove the comments; something like $CXX -fpreprocessed -dD -E <filename.cpp>
.
For Stan code, could we request this in stanc3
?
Is there a way to use stanc3's existing --info
option in the experimental rstan branch? I think that would actually give the information I need.
Sure. We can pass any argument to stanc3
JS, if it's accepted. I'd suggest that we create a new helper function though since stanc()
won't need to use this argument. Try the following:
model_info <- stanc_ctx$call("stanc", model_cppname, model_code, as.array("info"))
stanc_ctx
is defined internally in rstan
on load.
Awesome, thanks. That works. I can get the following now:
{
"inputs": {},
"parameters": {
"theta": { "type": "real", "dimensions": 0 },
"beta": { "type": "real", "dimensions": 1 }
},
"transformed parameters": { "omega": { "type": "real", "dimensions": 1 } },
"generated quantities": {},
"functions": [],
"distributions": [],
"included_files": []
}
Awesome, thanks. That works.
Great! I suggest that you wrap it with try()
and fallback to another approach or return and error message:
model_info <- try(stanc_ctx$call("stanc", model_cppname,
model_code, as.array("info")),
silent = TRUE)
if (!inherits(model_info , "try-error")) {
# model_info <- ?
stop(<error message>)
}
Good idea
Also, if the info is in a standard format (yaml
?), you can interpret it directly in R without grep
as text.
jsonlite::fromJSON(model_info$result)
seems to work really nicely to get this into a R list, but that requires the jsonlite package that rstan doesn't currently depend on. Do you know anything in base R that could handle it?
Do you know anything in base R that could handle it?
No. I'd create a simple helper function to convert the JSON info, if it's safe to do so. Otherwise, adding the dependency could be helpful depending on the value of the ability to automatically unconstraint the parameters.
We could also use the info in other places, replacing a lot of grep
code.
I'll see how much of a pain it is to parse the JSON without adding a dependency, but yeah it could be worth adding the dependency if we can use this info elsewhere.
if the info is in a standard format (
yaml
?)
the info should be in standards-compliant JSON. Do ping me if there is anything you wish --info
printed out but doesn't, and we can see if it's feasible
Do ping me if there is anything you wish
--info
printed out but doesn't, and we can see if it's feasible
Thanks @WardBrian! That would be great. We could rely on stanc3 --info
instead of dealing with the code as text. I'll look into this ASAP, to list the missing info we may need.
After playing around with all of this for a while, I think the fundamental issue here is that unconstrain_pars
is just really hard to use. If the user could easily unconstrain all the posterior draws for a specified subset of parameters then we could just get rid of all the plotting code in RStan, which would be great.
Basically what we need is something like
unconstrain_draws(stanfit, pars = "tau")
which returns all of the posterior draws for tau but unconstrained. If we had this then we wouldn't need to maintain any plotting code in RStan. The user could just do something like this
stanfit %>%
unconstrain_draws(pars = "tau") %>%
bayesplot::mcmc_trace()
@andrjohns You know more about unconstrain_pars than I do (or I assume so since you figured it out for CmdStanR and I haven't spent much time with it). How hard do you think this would be to do for RStan?
That's a good-looking syntax!
Yeah it wouldn't be too hard to implement, the main thing is making a "sane" structure for the return (I'm not overly happy with how cmdstanr
returns them now - a list of vectors per chain). A discussion I had with @n-kall a while back was to implement the structured format for unconstrained pars in the draws_*
format, so that we can easily handle multiple chains in a syntax/structure consistent with how users interact with the constrained draws.
I'll be getting back to cmdstanr
development later this week (now that there's not much more do for rstan
for a bit), so I can implement that in cmdstanr
and then port the approach over to rstan
. It will be easier to do in cmdstanr
first since the plumbing for unconstraining everything is already in place, it's just a structure/format change
Cool, sounds good!
On Mon, Mar 20, 2023 at 12:41 PM Andrew Johnson @.***> wrote:
That's a good-looking syntax!
Yeah it wouldn't be too hard to implement, the main thing is making a "sane" structure for the return (I'm not overly happy with how cmdstanr returns them now - a list of vectors per chain). A discussion I had with @n-kall https://github.com/n-kall a while back was to implement the structured format for unconstrained pars https://github.com/stan-dev/cmdstanr/issues/730 in the draws_* format, so that we can easily handle multiple chains in a syntax/structure consistent with how users interact with the constrained draws.
I'll be getting back to cmdstanr development later this week (now that there's not much more do for rstan for a bit), so I can implement that in cmdstanr and then port the approach over to rstan. It will be easier to do in cmdstanr first since the plumbing for unconstraining everything is already in place, it's just a structure/format change
— Reply to this email directly, view it on GitHub https://github.com/stan-dev/rstan/issues/1051#issuecomment-1476752023, or unsubscribe https://github.com/notifications/unsubscribe-auth/AB3PQQYW35SV67GDMEX2ENDW5CQHFANCNFSM6AAAAAAV5RFEJQ . You are receiving this because you authored the thread.Message ID: @.***>
This may be an inappropriate place to ask, but since all the major stakeholders seem to be participating here: Is there a good place to follow along with the general status of RStan 2.26, CRAN updates, etc?
This may be an inappropriate place to ask, but since all the major stakeholders seem to be participating here: Is there a good place to follow along with the general status of RStan 2.26, CRAN updates, etc?
For 2.26 you can follow #1034, that's tracking the progress of the downstream packages and their patching, as well as the submission decision
For 2.31+ you can follow #1046
This may be an inappropriate place to ask, but since all the major stakeholders seem to be participating here: Is there a good place to follow along with the general status of RStan 2.26, CRAN updates, etc?
The next step is to release StanHeaders
v2.26, which should be ready. The reverse dependency checks are available here. The failing dependencies are temporary and can be passed by communicating with CRAN.
Does that mean it could be a matter of days (rather than weeks, months, etc.) for RStan 2.26? Or would you expect it to still be further off
@bgoodri how have CRAN normally been for rstan
releases? Much of a battle?
@bgoodri @andrjohns @hsbadr Should we deprecate all the ggplot functions in rstan? Maybe in conjunction with the release of 2.26 or another future version? The ggplot functions in rstan were added before bayesplot existed and I'm not sure if there's any reason to keep supporting both. We could deprecate and point users to bayesplot. Thoughts?