Fix #579: re-enable approx inference vignette evaluation (#591)

Co-authored-by: Sam Abbott <contact@samabbott.co.uk>

Author: seabbs-bot

DESCRIPTION

@@ -70,6 +70,7 @@ Suggests:
     tidybayes,
     usethis
 Remotes:
+    paul-buerkner/brms,
     stan-dev/cmdstanr
 Config/Needs/website:
     r-lib/pkgdown,

NEWS.md

@@ -7,6 +7,8 @@
   the marginal model. See #583.
 - Added `primarycensored (>= 1.4.0)` version bound to DESCRIPTION.
 - Updated test expectations for changed primarycensored error handling.
+- Re-enabled approximate inference vignette evaluation using dev brms with
+  pathfinder path fix. See #579.
 
 ## Package
 

vignettes/approx-inference.Rmd

@@ -29,7 +29,7 @@ knitr::opts_chunk$set(
   message = FALSE,
   warning = FALSE,
   error = FALSE,
-  eval = FALSE
+  eval = TRUE
 )
 ```
 
@@ -138,7 +138,8 @@ obs_cens_trunc_samp <- simulate_gillespie(seed = 101) |>
 
 </details>
 
-We now prepare the data for fitting with the marginal model, and perform inference with HMC:
+We now prepare the data for fitting with the marginal model.
+We first pre-compile the Stan model so that compilation time is excluded from the timing comparisons below.
 
 ```{r results='hide'}
 linelist_data <- as_epidist_linelist_data(
@@ -151,8 +152,19 @@ linelist_data <- as_epidist_linelist_data(
 
 data <- as_epidist_marginal_model(linelist_data)
 
+# Pre-compile the model so compilation time is not included
+# in the timing comparisons
+fit_compile <- epidist(
+  data = data, backend = "cmdstanr",
+  chains = 1, iter = 5
+)
+```
+
+We now perform inference with HMC:
+
+```{r results='hide'}
 t <- proc.time()
-fit_hmc <- epidist(data = data, algorithm = "sampling", backend = "cmdstanr")
+fit_hmc <- update(fit_compile, chains = 4, iter = 2000)
 time_hmc <- proc.time() - t
 ```
 
@@ -163,27 +175,25 @@ To match the four Markov chains of length 1000 in HMC above, we then draw 4000 s
 
 ```{r results='hide'}
 t <- proc.time()
-fit_laplace <- epidist(
-  data = data, algorithm = "laplace", draws = 4000, backend = "cmdstanr"
+fit_laplace <- update(
+  fit_compile, algorithm = "laplace", draws = 4000
 )
 time_laplace <- proc.time() - t
 
 t <- proc.time()
-fit_advi <- epidist(
-  data = data, algorithm = "meanfield", draws = 4000, backend = "cmdstanr"
+fit_advi <- update(
+  fit_compile, algorithm = "meanfield", draws = 4000
 )
 time_advi <- proc.time() - t
 ```
 
 For the Pathfinder algorithm we will set `num_paths = 1`.
-Although both the Laplace and ADVI methods ran without problems in all cases during testing, we found that Pathfinder often produced the error message "Error evaluating model log probability: Non-finite gradient."
-Although a `save_single_paths` option is available, which may have allowed recovery of individual Pathfinder paths (and therefore removing faulty paths), it does not appear to be working currently^[See https://github.com/stan-dev/cmdstanr/issues/878].
 
-```{r}
+```{r results='hide'}
 t <- proc.time()
-fit_pathfinder <- epidist(
-  data = data, algorithm = "pathfinder", draws = 4000, chains = 1,
-  backend = "cmdstanr"
+fit_pathfinder <- update(
+  fit_compile, algorithm = "pathfinder", draws = 4000,
+  chains = 1
 )
 time_pathfinder <- proc.time() - t
 ```
@@ -307,7 +317,6 @@ times
 
 The range of alternative approximation algorithms available, and their ease of use, is an attractive feature of `brms`.
 We found that these algorithms do produce reasonable approximations in far less time than HMC.
-Of course, this vignette only includes one example, and a more thorough investigation would be required to make specific recommendations.
-That said, currently we do not recommend use of the Pathfinder algorithm due to its unstable performance in our testing and early stage software implementation.
+Of course, this vignette only includes one example, so results may differ for other models and datasets.
 
 ## References {-}