Add discrete_lognormal and discrete_normal

DESCRIPTION

@@ -44,7 +44,7 @@ Encoding: UTF-8
 Language: en-GB
 LazyData: true
 Roxygen: list(markdown = TRUE)
-RoxygenNote: 7.2.0
+RoxygenNote: 7.2.1
 SystemRequirements: Python (>= 2.7.0) with header files and shared
     library; TensorFlow (v1.14; https://www.tensorflow.org/); TensorFlow
     Probability (v0.7.0; https://www.tensorflow.org/probability/)

NAMESPACE

@@ -1,4 +1,8 @@
 # Generated by roxygen2: do not edit by hand
 
+export(discrete_lognormal)
+export(discrete_normal)
+import(R6)
+importFrom(R6,R6Class)
 importFrom(greta,.internals)
 importFrom(tensorflow,tf)

R/checks.R

@@ -0,0 +1,17 @@
+# check functions
+
+
+check_break_length <- function(breaks){
+  if (length(breaks) <= 1) {
+    msg <- cli::format_error(
+      c(
+        "{.var breaks} must be a vector with at least two break points",
+        "but {.var breaks} has length {length(breaks)}"
+      )
+    ) 
+    stop(
+      msg,
+      call. = FALSE
+    )
+  }
+}

R/conda_greta_env.R

@@ -0,0 +1,18 @@
+use_greta_conda_env <- function() {
+  tryCatch(
+    expr = reticulate::use_condaenv("greta-env", required = TRUE),
+    error = function(e) NULL
+  )
+}
+
+using_greta_conda_env <- function() {
+  config <- reticulate::py_discover_config()
+  grepl("greta-env", config$python)
+}
+
+have_greta_conda_env <- function(){
+  tryCatch(
+    expr = "greta-env" %in% reticulate::conda_list()$name,
+    error = function(e) FALSE
+  )
+}

R/discrete_lognormal.R

@@ -0,0 +1,107 @@
+#' @name discrete_lognormal
+#' @title Discrete lognormal distribution
+#'
+#' @description a discretised lognormal distribution (i.e. sampled by applying 
+#'   the round operation to samples from a lognormal). Due to the numerical 
+#'   instability of integrating across the distribution, a vector of breaks 
+#'   must be defined and the observations will be treated as censored 
+#'   within those breaks
+#'
+#' @param meanlog unconstrained parameters giving the mean of the distribution 
+#'   on the log scale
+#' @param sdlog unconstrained parameters giving the standard deviation of the 
+#'   distribution on the log scale
+#' @param breaks a vector of breaks; observations will be treated as censored 
+#'   within those breaks
+#' @param dim a scalar giving the number of rows in the resulting greta array
+#'
+#' @importFrom R6 R6Class
+#' @export
+
+discrete_lognormal <- function(meanlog, sdlog, breaks, dim = NULL) {
+  distrib("discrete_lognormal", meanlog, sdlog, breaks, dim)
+}
+
+# define the discrete lognormal distribution
+discrete_lognormal_distribution <- R6Class(
+  classname = "discrete_lognormal_distribution",
+  inherit = distribution_node,
+  public = list(
+
+    breaks = NA,
+    lower_bounds = NA,
+    upper_bounds = NA,
+
+    initialize = function(meanlog, sdlog, breaks, dim) {
+
+      meanlog <- as.greta_array(meanlog)
+      sdlog   <- as.greta_array(sdlog)
+
+      # check length of breaks
+      check_break_length(breaks)
+
+      # handle gradient issue between sdlog and 0s
+      breaks <- pmax(breaks, .Machine$double.eps)
+      self$breaks <- breaks
+      self$lower_bounds <- breaks[-length(breaks)]
+      self$upper_bounds <- breaks[-1]
+
+      # add the nodes as parents and parameters
+      dim <- check_dims(meanlog, sdlog, target_dim = dim)
+      super$initialize("discrete_lognormal", dim, discrete = TRUE)
+      self$add_parameter(meanlog, "meanlog")
+      self$add_parameter(sdlog, "sdlog")
+
+    },
+
+    tf_distrib = function(parameters, dag) {
+
+      meanlog <- parameters$meanlog
+      sdlog <- parameters$sdlog
+
+      tf_breaks <- fl(self$breaks)
+      tf_lower_bounds <- fl(self$lower_bounds)
+      tf_upper_bounds <- fl(self$upper_bounds)
+
+      log_prob <- function(x) {
+
+        # build distribution object
+        d <- tfp$distributions$LogNormal(
+          loc = meanlog,
+          scale = sdlog
+        )
+
+        # for those lumped into groups,
+        # compute the bounds of the observed groups
+        # and get tensors for the bounds in the format expected by TFP
+        x_safe <- tf$math$maximum(x, fl(.Machine$double.eps))
+        tf_idx <- tfp$stats$find_bins(x_safe, tf_breaks)
+        tf_idx_int <- tf_as_integer(tf_idx)
+        tf_lower_vec <- tf$gather(tf_lower_bounds, tf_idx_int)
+        tf_upper_vec <- tf$gather(tf_upper_bounds, tf_idx_int)
+
+        # compute the density over the observed groups
+        low <- tf_safe_cdf(tf_lower_vec, d)
+        up <- tf_safe_cdf(tf_upper_vec, d)
+        log_density <- log(up - low)
+
+      }
+
+      sample <- function(seed) {
+
+        d <- tfp$distributions$LogNormal(
+          loc = meanlog,
+          scale = sdlog
+        )
+        continuous <- d$sample(seed = seed)
+        # tf$floor(continuous)
+        tf$round(continuous)
+
+      }
+
+      list(log_prob = log_prob, sample = sample)
+
+    }
+
+  )
+)

R/discrete_normal.R

@@ -0,0 +1,142 @@
+#' @name discrete_normal
+#' @title Discrete normal distribution
+#'
+#' @description a discretised normal distribution (i.e. sampled by applying 
+#'   the round operation to samples from a normal). Due to the numerical 
+#'   instability of integrating across the distribution, a vector of breaks 
+#'   must be defined and the observations will be treated as censored 
+#'   within those breaks
+#'
+#' @param mean unconstrained parameters giving the mean of the distribution
+#' @param sd unconstrained parameters giving the standard deviation of the 
+#'   distribution
+#' @param breaks a vector of breaks; observations will be treated as censored 
+#'   within those breaks
+#' @param edges a vector of edges; length needs to be length(breaks)+1;
+#'   observations between two consecutive edges will be discretised to the 
+#'   break value between the corresponding edges
+#' @param dim a scalar giving the number of rows in the resulting greta array
+#'
+#' @importFrom R6 R6Class
+#' @export
+
+discrete_normal <- function(mean, sd, breaks, edges, dim = NULL) {
+  distrib("discrete_normal", mean, sd, breaks, edges, dim)
+}
+
+# define the discrete normal distribution
+discrete_normal_distribution <- R6Class(
+  classname = "discrete_normal_distribution",
+  inherit = distribution_node,
+  public = list(
+
+    breaks = NA,
+    edges = NA,
+    lower_bounds = NA,
+    upper_bounds = NA,
+    lower_bound = NA,
+    upper_bound = NA,
+
+    initialize = function(mean, sd, breaks, edges, dim) {
+
+      mean <- as.greta_array(mean)
+      sd   <- as.greta_array(sd)
+
+      # check length of breaks
+      check_break_length(breaks)
+
+      # add breaks, vector of lower and upper bounds, and the lower and upper
+      # bound of supported values
+      # self$breaks <- breaks
+      # self$lower_bounds <- breaks[-length(breaks)]
+      # self$upper_bounds <- breaks[-1]
+      # self$lower_bound <- min(breaks)
+      # self$upper_bound <- max(breaks)
+
+      # EXPERIMENTAL:
+      # convert breaks to edges, which
+      # will be used to gather samples at the breaks and convert them to 
+      # rounded values 
+      # while avoiding the use of round or floor, which assumes that 
+      # the breaks are all the integers
+      # perhaps more crucially, we use edges, not breaks, to integrate the CDFs
+      # because we say that the likelihood of observing a break value is the sum
+      # of probability from the lower to upper edges surrounding a break
+      # first, create the lower and upper bounds that will bin a continuous
+      # variable into breaks
+      # using midpoint between breaks for now (from diff)
+      # generate edges using midpoint, -Inf, and Inf
+      # breaks_diff <- diff(breaks)
+      # edges <- c(-Inf, breaks[-length(breaks)] + breaks_diff/2, Inf)
+      self$edges <- edges
+      self$lower_bounds <- edges[-length(edges)]
+      self$upper_bounds <- edges[-1]
+      self$lower_bound <- min(edges)
+      self$upper_bound <- max(edges)
+
+      # add the nodes as parents and parameters
+      dim <- check_dims(mean, sd, target_dim = dim)
+      super$initialize("discrete_normal", dim, discrete = TRUE)
+      self$add_parameter(mean, "mean")
+      self$add_parameter(sd, "sd")
+
+    },
+
+    tf_distrib = function(parameters, dag) {
+
+      mean <- parameters$mean
+      sd <- parameters$sd
+
+      tf_breaks <- fl(self$breaks)
+      tf_edges <- fl(self$edges)
+      tf_lower_bounds <- fl(self$lower_bounds)
+      tf_upper_bounds <- fl(self$upper_bounds)
+      tf_lower_bound <- fl(self$lower_bound)
+      tf_upper_bound <- fl(self$upper_bound)
+
+      log_prob <- function(x) {
+
+        # build distribution object
+        d <- tfp$distributions$Normal(
+          loc = mean,
+          scale = sd
+        )
+
+        # for those lumped into groups,
+        # compute the bounds of the observed groups
+        # and get tensors for the bounds in the format expected by TFP
+        tf_idx <- tfp$stats$find_bins(x, tf_edges)
+        tf_idx_int <- tf_as_integer(tf_idx)
+        tf_lower_vec <- tf$gather(tf_lower_bounds, tf_idx_int)
+        tf_upper_vec <- tf$gather(tf_upper_bounds, tf_idx_int)
+
+        # compute the density over the observed groups
+        # note-to-self: this looks like https://mc-stan.org/docs/2_29/stan-users-guide/bayesian-measurement-error-model.html#rounding
+        low <- tf_safe_cdf(tf_lower_vec, d, tf_lower_bound, tf_upper_bound)
+        up <- tf_safe_cdf(tf_upper_vec, d, tf_lower_bound, tf_upper_bound)
+        log_density <- log(up - low)
+
+      }
+
+      sample <- function(seed) {
+
+        d <- tfp$distributions$Normal(
+          loc = mean,
+          scale = sd
+        )
+        continuous <- d$sample(seed = seed)
+
+        # gather samples at the breaks to convert them to rounded values
+        # ditto from what we did to breaks above
+        tf_edges_idx <- tfp$stats$find_bins(continuous, tf_edges)
+        tf_edges_idx_int <- tf_as_integer(tf_edges_idx)
+        tf$gather(tf_breaks, tf_edges_idx_int)
+
+      }
+
+      list(log_prob = log_prob, sample = sample)
+
+    }
+
+  )
+)

R/package.R

@@ -7,6 +7,7 @@
 #' 
 #' @importFrom tensorflow tf
 #' @importFrom greta .internals
+#' @import R6
 #' 
 #' @examples
 #' 

R/tf_functions.R

@@ -0,0 +1,25 @@
+# tensorflow functions
+
+# CDF of the provided distribution, handling 0s and Infs
+tf_safe_cdf <- function(x, distribution, lower_bound, upper_bound) {
+
+  # prepare to handle values outside the supported range
+  too_low <- tf$less(x, lower_bound)
+  too_high <- tf$greater_equal(x, upper_bound)
+  supported <- !too_low & !too_high
+  ones <- tf$ones_like(x)
+  zeros <- tf$zeros_like(x)
+
+  # run cdf on supported values, and fill others in with the appropriate value
+  x_clean <- tf$where(supported, x, ones)
+  cdf_clean <- distribution$cdf(x_clean)
+  mask <- tf$where(supported, ones, zeros)
+  add <- tf$where(too_high, ones, zeros)
+  cdf_clean * mask + add
+
+}
+
+# cast to integer
+tf_as_integer <- function(x) {
+  tf$cast(x, tf$int32)
+}

R/zzz.R

@@ -0,0 +1,31 @@
+# load tf probability
+tfp <- reticulate::import("tensorflow_probability", delay_load = TRUE)
+
+# crate the node list object whenever the package is loaded
+.onLoad <- function(libname, pkgname) { # nolint
+
+  # unset reticulate python environment, for more details, see:
+  # https://github.com/greta-dev/greta/issues/444
+  Sys.unsetenv("RETICULATE_PYTHON")
+
+  if (have_greta_conda_env()) {
+    use_greta_conda_env()
+  }
+
+  # silence TF's CPU instructions message
+  Sys.setenv(TF_CPP_MIN_LOG_LEVEL = 2)
+
+  # silence messages about deprecation etc.
+  # disable_tensorflow_logging()
+
+  # warn if TF version is bad
+  # check_tf_version("startup")
+
+  # switch back to 0-based extraction in tensorflow, and don't warn about
+  # indexing with tensors
+  options(tensorflow.one_based_extract = FALSE)
+  options(tensorflow.extract.warn_tensors_passed_asis = FALSE)
+
+  # default float type
+  options(greta_tf_float = "float64")
+}