core code complete

CoryMcCartan · Apr 23, 2020 · 9d60a4b · 9d60a4b
1 parent d4e60f7
commit 9d60a4b
Show file tree

Hide file tree

Showing 31 changed files with 522 additions and 85 deletions.
diff --git a/NAMESPACE b/NAMESPACE
@@ -8,6 +8,8 @@ S3method(pull,adjustr_weighted)
 S3method(rename,adjustr_spec)
 S3method(select,adjustr_spec)
 S3method(slice,adjustr_spec)
+S3method(summarise,adjustr_weighted)
+S3method(summarize,adjustr_weighted)
 export(adjust_weights)
 export(extract_samp_stmts)
 export(get_resampling_idxs)

diff --git a/R/adjust_weights.R b/R/adjust_weights.R
@@ -76,6 +76,8 @@ adjust_weights = function(spec, object, data=NULL, keep_bad=F) {
     if (!keep_bad)
         adjust_obj$.weights[adjust_obj$.pareto_k > 0.7] = list(NA_real_)
     attr(adjust_obj, "draws") = rstan::extract(object)
+    attr(adjust_obj, "data") = data
+    attr(adjust_obj, "iter") = object@sim$chains * (object@sim$iter - object@sim$warmup)
     adjust_obj
 }
 

diff --git a/R/adjustr-package.R b/R/adjustr-package.R
@@ -15,6 +15,7 @@
 #' \itemize{
 #'     \item \code{\link{make_spec}}
 #'     \item \code{\link{adjust_weights}}
+#'     \item \code{\link{summarize}}
 #' }
 #'
 #' @importFrom methods is

diff --git a/R/logprob.R b/R/logprob.R
@@ -126,4 +126,4 @@ if (requireNamespace("extraDistr", quietly=T)) {
 }
 # Turn mapping into an environment suitable for metaprogramming,
 # and turn each density into its curried form (see `make_dens` above)
-distr_env = new_environment(purrr::map(distrs, make_dens), parent=empty_env())
+distr_env = new_environment(purrr::map(distrs, make_dens))
diff --git a/R/mockup.R b/R/mockup.R
@@ -32,6 +32,7 @@ eightschools_m = stan(model_code=model_code, chains=2, data=model_d, warmup=500,
                       iter=510, save_dso=F, save_warmup=F)
 eightschools_m@stanmodel@dso = new("cxxdso")
 save(eightschools_m, file="tests/test_model.rda")
+load("tests/test_model.rda")
 
 #slot(eightschools_m@stanmodel, "dso", F) = NULL
 #draws = extract(eightschools_m)

diff --git a/R/parsing.R b/R/parsing.R
@@ -10,11 +10,11 @@ get_parser = function() { # nocov start
     }
 } # nocov end
 
-# Parse Satan `model_code` into a data frame which represents the parsing tree
+# Parse Stan `model_code` into a data frame which represents the parsing tree
 parse_model = function(model_code) {
     parser_output = utils::capture.output(
     get_parser()(model_code, function(name, value, pos, depth) {
-        cat(stringr::str_glue('"{name}","{value}",{pos},{depth}\n\n'))
+        cat('"', name, '","', value, '",', pos, ',', depth, '\n', sep="")
     })
     )
     parser_csv = paste0("name,value,pos,depth\n",

diff --git a/R/use_weights.R b/R/use_weights.R
@@ -34,3 +34,93 @@ get_resampling_idxs = function(x, frac=1, replace=T) {
         get_idxs(x)
     }
 }
+
+#' Summarize Posterior Distributions Under Alternative Model Specifications
+#'
+#' Uses weights computed in \code{\link{adjust_weights}} to compute posterior
+#' summary statistics. These statistics can be compared against their reference
+#' values to quantify the sensitivity of the model to aspects of its
+#' specification.
+#'
+#' @param .data An \code{adjustr_weighted} object.
+#' @param ... Name-value pairs of expressions. The name of each argument will be
+#'   the name of a new variable, and the value will be computed for the
+#'   posterior distribution of eight alternative specification. For example,
+#'   a value of \code{mean(theta)} will compute the posterior mean of
+#'   \code{theta} for each alternative specification.
+#' @param .resampling Wether to compute summary statistics by first resampling
+#'  the data according to the weights. Defaults to \code{FALSE}, but will be
+#'  used for any summary statistic that is not \code{mean}, \code{var} or
+#'  \code{sd}.
+#' @param .model_data Stan model data, if not provided in the earlier call to
+#'   \code{\link{adjust_weights}}.
+#'
+#' @return An \code{adjustr_weighted} object, wth the new columns specified in
+#' \code{...} added.
+#'
+#' @rdname summarize.adjustr_weighted
+#' @export
+summarise.adjustr_weighted = function(.data, ..., .resampling=F, .model_data=NULL) {
+    stopifnot(is.adjustr_weighted(.data)) # just in case called manually
+    args = enexprs(...)
+
+    broadcast = function(x) {
+        dims = c(dim(as.array(x)), iter)
+        x = array(rep(x, iter), dim=dims)
+        aperm(x, c(length(dims), 2:length(dims) - 1))
+    }
+    iter = attr(.data, "iter")
+    if (!is_null(.model_data)) attr(.data, "data") = .model_data
+    data = append(attr(.data, "draws"), map(attr(.data, "data"), broadcast))
+
+    n_args = length(args)
+    for (i in seq_along(args)) {
+        name = names(args)[i]
+        if (name == "") name = expr_name(args[[i]])
+
+        call = args[[i]]
+        if (!.resampling && exists(call_name(call), funs_env)) {
+            fun = call_fn(call, funs_env)
+        } else {
+            fun = function(x, ...) apply(x, 2, call_fn(call), ...)
+            .resampling = T
+        }
+
+        expr = expr_deparse(call_args(call)[[1]])
+        expr = stringr::str_replace_all(expr, "\\[(\\d)", "[,\\1")
+        expr = stringr::str_replace_all(expr, "(?<![a-zA-Z0-9._])mean\\(", "rowMeans(")
+        expr = stringr::str_replace_all(expr, "(?<![a-zA-Z0-9._])sum\\(", "rowSum(")
+        computed = as.array(eval_tidy(parse_expr(expr), data))
+        if (length(dim(computed)) == 1) dim(computed) = c(dim(computed), 1)
+
+        if (!.resampling) {
+            new_col = map(.data$.weights, ~ fun(computed, .))
+        } else {
+            idxs = map(.data$.weights, ~ sample.int(iter, iter, replace=T, prob=.))
+            new_col = map(idxs, function(idx) {
+                comp = as.array(computed[idx,])
+                if (length(dim(comp)) == 1) dim(comp) = c(dim(comp), 1)
+                exec(fun, comp, !!!map(call_args(call)[-1], eval_tidy))
+            })
+        }
+        if (length(new_col[[1]]) == 1 && is.numeric(new_col[[1]]))
+            new_col = as.numeric(new_col)
+        .data[[name]] = new_col
+    }
+
+    .data
+}
+#' @rdname summarize.adjustr_weighted
+#' @export
+summarize.adjustr_weighted = summarise.adjustr_weighted
+
+# Weighted summary functions that work on arrays
+wtd_array_mean = function(arr, wgt) colSums(as.array(arr)*wgt) / sum(wgt)
+wtd_array_var = function(arr, wgt)  wtd_array_mean((arr - wtd_array_mean(arr, wgt))^2, wgt)
+wtd_array_sd = function(arr, wgt) sqrt(wtd_array_var(arr, wgt))
+
+funs_env = new_environment(list(
+    mean = wtd_array_mean,
+    var = wtd_array_var,
+    sd = wtd_array_sd
+))
diff --git a/README.md b/README.md
@@ -10,18 +10,18 @@
 sensitivity of a Bayesian model (fitted with [Stan](https://mc-stan.org)) to the
 specification of its likelihood and priors. Users provide a series of alternate
 sampling specifications, and the package uses Pareto-smoothed importance
-sampling to estimate posterior quantities of interest under each specification.
-The package also provides functions to summarize and plot how these quantities
+sampling to estimate the posterior under each specification. The package also
+provides functions to summarize and plot how posterior quantities quantities
 change across specifications.
 
 The package aims to provide simple interface that makes it as easy as possible
 for modellers to try out various adjustments to their Stan models, without
 needing to write any specific Stan code or even recompile or rerun their model.
 
 The package works by parsing Stan model code, so everything works best if the
-model was written by the user. The complexity of **rstanarm** and **brms**
-models makes it harder to make interpretable adjustments, but in principle 
-these are possible too.
+model was written by the user. Models made using **brms** may in principle be
+used as well. Models made using **rstanarm** are constructed using more 
+complex model templates, and cannot be used.
 
 ## Getting Started
 

diff --git a/_pkgdown.yml b/_pkgdown.yml
@@ -41,7 +41,7 @@ navbar:
    - icon: fa-twitter
      href: https://twitter.com/mcmc_stan
    - icon: fa-github
-     href: https://github.com/stan-dev/bayesplot
+     href: https://github.com/CoryMcCartan/adjustr/
    - icon: fa-users
      href: https://discourse.mc-stan.org/
 
@@ -72,17 +72,14 @@ reference:
     contents:
       - make_spec
       - adjust_weights
+      - summarise.adjustr_weighted
   - title: "Helper Functions"
     desc: >
       Various helper functions for examining a model or building sampling
       specifications.
     contents:
       - extract_samp_stmts
+      - as.data.frame.adjustr_spec
       - dplyr.adjustr_spec
       - get_resampling_idxs
       - pull.adjustr_weighted
-  - title: "Plotting Functions"
-    desc: >
-      Functions to visualize changes to a model's posterior across
-      specifications.
-    contents:
diff --git a/docs/404.html b/docs/404.html
diff --git a/docs/LICENSE-text.html b/docs/LICENSE-text.html
diff --git a/docs/authors.html b/docs/authors.html
diff --git a/docs/index.html b/docs/index.html
diff --git a/docs/pkgdown.yml b/docs/pkgdown.yml
@@ -2,7 +2,7 @@ pandoc: 2.3.1
 pkgdown: 1.5.1
 pkgdown_sha: ~
 articles: []
-last_built: 2020-04-19T01:34Z
+last_built: 2020-04-23T01:51Z
 urls:
   reference: https://corymccartan.github.io/adjustr//reference
   article: https://corymccartan.github.io/adjustr//articles

diff --git a/docs/reference/adjust_weights.html b/docs/reference/adjust_weights.html
diff --git a/docs/reference/adjustr.html b/docs/reference/adjustr.html
diff --git a/docs/reference/as.data.frame.adjustr_spec.html b/docs/reference/as.data.frame.adjustr_spec.html
diff --git a/docs/reference/dplyr.adjustr_spec.html b/docs/reference/dplyr.adjustr_spec.html
diff --git a/docs/reference/extract_samp_stmts.html b/docs/reference/extract_samp_stmts.html
diff --git a/docs/reference/get_resampling_idxs.html b/docs/reference/get_resampling_idxs.html