R/categorical.R

In BayesForge: Bayesian Inference using 'numpyro' and 'XLA'

#' @title Sample from a Categorical distribution.
#'
#' @description The Categorical distribution, also known as the multinomial distribution,
#' describes the probability of different outcomes from a finite set of possibilities.
#' It is commonly used to model discrete choices or classifications.
#'
#' \deqn{P(k) = \frac{e^{\log(p_k)}}{\sum_{j=1}^{K} e^{\log(p_j)}}}
#'
#' where \deqn{p_k} is the probability of outcome \deqn{k}, and the sum is over all possible outcomes.
#'
#' @param probs A numeric vector of probabilities for each category. Must sum to 1.
#' @param logits A numeric vector of  Log-odds of each category.
#' @param shape A numeric vector specifying the shape. When \code{sample=FALSE} (model building),
#'   this is used with `.expand(shape)` to set the distribution's batch shape.
#'   When \code{sample=TRUE} (direct sampling), this is used as `sample_shape` to draw a raw JAX array
#'   of the given shape.
#' @param event The number of batch dimensions to reinterpret as event dimensions (used in model building).
#' @param mask An optional boolean vector to mask observations.
#' @param create_obj Logical; If `TRUE`, returns the raw BI distribution object instead of creating a sample site.
#' @param validate_args Logical: Whether to validate parameter values.  Defaults to `reticulate::py_none()`.
#' @param sample A logical value that controls the function's behavior. If `TRUE`,
#'   the function will directly draw samples from the distribution. If `FALSE`,
#'   it will create a random variable within a model. Defaults to `FALSE`.
#' @param seed An integer used to set the random seed for reproducibility when
#'   `sample = TRUE`. This argument has no effect when `sample = FALSE`, as
#'   randomness is handled by the model's inference engine. Defaults to 0.
#' @param obs A numeric vector or array of observed values. If provided, the
#'   random variable is conditioned on these values. If `NULL`, the variable is
#'   treated as a latent (unobserved) variable. Defaults to `NULL`.
#' @param name A character string representing the name of the random variable
#'   within a model. This is used to uniquely identify the variable. Defaults to 'x'.
#' @param to_jax Boolean. Indicates whether to return a JAX array or not.
#'
#' @return
#'  - When \code{sample=FALSE}, a BI Categorical distribution object (for model building).
#'
#'  - When \code{sample=TRUE}, a JAX array of samples drawn from the Categorical distribution (for direct sampling).
#'
#'  - When \code{create_obj=TRUE}, the raw BI distribution object (for advanced use cases).
#'
#' @examples
#' \donttest{
#' library(BayesForge)
#' m=importBF(platform='cpu')
#' bf.dist.categorical(probs = c(0.5,0.5), sample = TRUE, shape = c(3))
#' }
#' @export
bf.dist.categorical=function(probs=py_none(), logits=py_none(), validate_args=py_none(), name='x', obs=py_none(), mask=py_none(), sample=FALSE, seed = py_none(), shape=c(), event=0, create_obj=FALSE, to_jax = TRUE) {
     shape=do.call(tuple, as.list(as.integer(shape)))
     event=as.integer(event)
     if (!.BF_env$.py$is_none(seed)){seed=as.integer(seed);}
     reticulate::py_run_string("def is_none(x): return x is None")
     if(!.BF_env$.py$is_none(logits)){logits = .BF_env$jnp$array(logits)}
     if(!.BF_env$.py$is_none(probs)){probs = .BF_env$jnp$array(probs)}
     .BF_env$.bf_instance$dist$categorical(
       probs = probs,
       logits = logits,
       validate_args= validate_args,  name= name,  obs= obs,  mask= mask,  sample= sample,  seed= seed,  shape= shape,  event= event,  create_obj= create_obj,   to_jax = to_jax)
}