## ----setup, include = FALSE---------------------------------------------------
## Chunks that call DICEr() or load torch are skipped on CRAN
## (they require the torch system library and ~160 MB of weights).
## Set NOT_CRAN=true locally, or use devtools::check() which sets it automatically.
knitr::opts_chunk$set(
  collapse = TRUE,
  comment  = "#>",
  eval     = identical(Sys.getenv("NOT_CRAN"), "true")
)

## ----install, eval = FALSE----------------------------------------------------
# ## From a local source tarball:
# install.packages(
#   "/path/to/DICErClust_0.1.1.tar.gz",
#   repos = NULL, type = "source"
# )

## ----data-format--------------------------------------------------------------
# ## Build a minimal synthetic dataset ----------------------------------------
# set.seed(42)
# n_train <- 120L; n_test <- 40L; p <- 6L; q <- 3L
# 
# make_rds <- function(n, path) {
#   saveRDS(
#     list(
#       matrix(runif(n * p), n, p),                   # data_x: continuous
#       matrix(as.numeric(rbinom(n * q, 1, 0.5)), n, q), # data_v: binary float
#       rbinom(n, 1, 0.3)                             # data_y: outcome
#     ),
#     path
#   )
# }
# 
# data_dir <- file.path(tempdir(), "dice_intro")
# dir.create(data_dir, showWarnings = FALSE)
# make_rds(n_train, file.path(data_dir, "train.rds"))
# make_rds(n_test,  file.path(data_dir, "test.rds"))
# 
# ## Verify format
# d <- readRDS(file.path(data_dir, "train.rds"))
# cat("data_x:", nrow(d[[1]]), "×", ncol(d[[1]]), " storage:", storage.mode(d[[1]]), "\n")
# cat("data_v:", nrow(d[[2]]), "×", ncol(d[[2]]), " storage:", storage.mode(d[[2]]), "\n")
# cat("data_y: length", length(d[[3]]), " table:", paste(table(d[[3]]), collapse = "/"), "\n")

## ----train, eval = FALSE------------------------------------------------------
# library(DICErClust)
# 
# args <- list(
#   seed              = 42L,
#   input_path        = data_dir,
#   filename_train    = "train.rds",
#   filename_test     = "test.rds",
#   n_input_fea       = p,       # columns in data_x
#   n_hidden_fea      = 3L,      # LSTM latent dimension
#   lstm_layer        = 1L,
#   lstm_dropout      = 0.0,
#   K_clusters        = 2L,      # number of clusters
#   n_dummy_demov_fea = q,       # columns in data_v
#   cuda              = FALSE,   # set TRUE to use GPU
#   lr                = 1e-4,
#   init_AE_epoch     = 5L,      # Stage 1 warm-up epochs
#   iter              = 20L,     # Stage 2 iterations
#   epoch_in_iter     = 2L,
#   lambda_AE         = 1.0,
#   lambda_classifier = 1.0,
#   lambda_outcome    = 1.0,
#   lambda_p_value    = 1.0
# )
# 
# old_wd <- setwd(tempdir())
# DICEr(args)            # writes output to hn_3_K_2/part2_AE_nhidden_3/
# setwd(old_wd)

## ----load-results, eval = FALSE-----------------------------------------------
# part2_dir <- file.path(tempdir(), "hn_3_K_2", "part2_AE_nhidden_3")
# 
# res_train <- readRDS(file.path(part2_dir, "data_train_iter.rds"))
# res_test  <- readRDS(file.path(part2_dir, "data_test_iter.rds"))
# 
# ## Cluster assignments
# ## Training set: use res_train$C   (k-means labels, re-ordered by outcome rate)
# ## Test set:     use res_test$pred_C (nearest-centroid assignments)
# table(res_test$pred_C)

## ----vignette-link, eval = FALSE----------------------------------------------
# vignette("heart-failure-example", package = "DICErClust")