R/d_H2OAE.R

In egenn/rtemis: Machine Learning and Visualization

# d_H2OAE.R
# ::rtemis::
# 2017 E.D. Gennatas www.lambdamd.org

#' Autoencoder using H2O
#'
#' Train an Autoencoder using `h2o::h2o.deeplearning`
#' Check out the H2O Flow at `[ip]:[port]`, Default IP:port is "localhost:54321"
#' e.g. if running on localhost, point your web browser to `localhost:54321`
#'
#' @inheritParams s_H2ODL
#' @param learning.rate Float: Learning rate. Default = .005
#' @param learning.rate.annealing Float: Learning rate annealing. Default = 1e-06
#' @param stopping.rounds Integer: Stop if simple moving average of length `stopping.rounds` of the
#' `stopping.metric` does not improve. Set to 0 to disable. Default = 50
#' @param stopping.metric Character: Stopping metric to use: "AUTO", "deviance", "logloss", "MSE", "RMSE", "MAE", "RMSLE",
#' "AUC", "lift_top_group", "misclassification", "mean_per_class_error". Default = "AUTO" ("logloss" for Classification,
#' "deviance" for Regression)
#' @param scale Logical: If TRUE, scale input before training autoencoder. Default = TRUE
#' @param center Logical: If TRUE, center input before training autoencoder. Default = TRUE
#' @param extract.layer Integer: Which layer to extract. For regular autoencoder, this is the middle layer.
#' Default = `ceiling(length(n.hidden.nodes)/2)`
#' @param epochs Integer: How many times to iterate through the dataset. Default = 5000
#' @param activation Character: Activation function to use: "Tanh" (Default), "TanhWithDropout", "Rectifier", "RectifierWithDropout",
#' "Maxout", "MaxoutWithDropout"
#' @param loss Character: "Automatic" (Default), "CrossEntropy", "Quadratic", "Huber", "Absolute"
#' @param n.cores Integer: Number of cores to use
#' @param ... Additional arguments to pass to `h2p::h2o.deeplearning`
#' @return `rtDecom` object
#' @author E.D. Gennatas
#' @seealso [decom]
#' @family Decomposition
#' @family Deep Learning
#' @export

d_H2OAE <- function(x,
                    x.test = NULL,
                    x.valid = NULL,
                    ip = "localhost",
                    port = 54321,
                    n.hidden.nodes = c(ncol(x), 3, ncol(x)),
                    extract.layer = ceiling(length(n.hidden.nodes) / 2),
                    epochs = 5000,
                    activation = "Tanh",
                    loss = "Automatic",
                    input.dropout.ratio = 0,
                    hidden.dropout.ratios = rep(0, length(n.hidden.nodes)),
                    learning.rate = .005,
                    learning.rate.annealing = 1e-06,
                    l1 = 0,
                    l2 = 0,
                    stopping.rounds = 50,
                    stopping.metric = "AUTO",
                    scale = TRUE,
                    center = TRUE,
                    n.cores = rtCores,
                    verbose = TRUE,
                    save.mod = FALSE,
                    outdir = NULL, ...) {
  # Intro ----
  if (!is.null(outdir)) outdir <- normalizePath(outdir, mustWork = FALSE)
  logFile <- if (!is.null(outdir)) {
    paste0(outdir, "/", sys.calls()[[1]][[1]], ".", format(Sys.time(), "%Y%m%d.%H%M%S"), ".log")
  } else {
    NULL
  }
  start.time <- intro(verbose = verbose, logFile = logFile)

  # Dependencies ----
  dependency_check("h2o")

  # Intro ----
  decom.name <- "H2OAE"

  # Arguments ----
  verbose <- verbose | !is.null(logFile)
  if (save.mod && is.null(outdir)) outdir <- paste0("./s.", decom.name)
  if (!is.null(outdir)) outdir <- paste0(normalizePath(outdir, mustWork = FALSE), "/")

  # Data ----
  x <- as.data.frame(x)
  if (!is.null(x.test)) x.test <- as.data.frame(x.test)
  n <- NROW(x)
  p <- NCOL(x)
  if (verbose) cat("||| Input has dimensions ", n, " rows by ", p, " columns,\n", sep = "")
  if (verbose) cat("    interpreted as", n, "cases with", p, "features.\n")
  if (is.null(colnames(x))) colnames(x) <- paste0("Feature_", seq_len(NCOL(x)))
  xnames <- colnames(x)
  if (!is.null(x.test)) colnames(x.test) <- xnames
  if (scale) {
    x <- scale(x, center = center)
    if (!is.null(x.test)) x.test <- scale(x.test, center = center)
  }

  # h2o Frames
  if (verbose) msg2("Connecting to H2O server...")
  h2o::h2o.init(ip = ip, port = port, nthreads = n.cores)
  if (verbose) msg2("Creating H2O frames...")
  df.train <- h2o::as.h2o(x, "df_train")
  if (!is.null(x.test)) {
    df.test <- h2o::as.h2o(x.test, "df_test")
  } else {
    df.test <- NULL
  }

  # H2OAE ----
  if (verbose) msg2("Training H2O Autoencoder...")
  params <- list(
    x = seq_len(ncol(x)), training_frame = df.train,
    model_id = paste0("rtemis.H2OAE.", format(Sys.time(), "%b%d.%H:%M:%S.%Y")),
    validation_frame = df.test,
    hidden = n.hidden.nodes,
    epochs = epochs,
    activation = activation,
    loss = loss,
    rate = learning.rate,
    rate_annealing = learning.rate.annealing,
    input_dropout_ratio = input.dropout.ratio,
    stopping_rounds = stopping.rounds,
    stopping_metric = stopping.metric,
    # nfolds = nfolds, # not supported for Autoencoder
    autoencoder = TRUE, ...
  )
  if (sum(hidden.dropout.ratios) > 0) params$hidden_dropout_ratios <- hidden.dropout.ratios

  mod <- do.call(h2o::h2o.deeplearning, params)

  if (verbose) print(summary(mod))

  # Projections ----
  if (verbose) msg2("Extracting Deep Features...")
  projections.train <- as.data.frame(h2o::h2o.deepfeatures(mod, df.train, layer = extract.layer))
  if (!is.null(x.test)) {
    projections.test <- as.data.frame(h2o::h2o.deepfeatures(mod, df.test, layer = extract.layer))
  } else {
    projections.test <- NULL
  }

  # Outro ----
  extra <- list()
  rt <- rtDecom$new(
    decom.name = decom.name,
    decom = mod,
    xnames = xnames,
    projections.train = projections.train,
    projections.test = projections.test,
    parameters = list(
      n.hidden.nodes = n.hidden.nodes,
      extract.layer = extract.layer,
      epochs = epochs,
      activation = activation,
      loss = loss,
      input.dropout.ratio = input.dropout.ratio,
      hidden.dropout.ratios = hidden.dropout.ratios,
      learning.rate = learning.rate,
      learning.rate.annealing = learning.rate.annealing,
      l1 = l1,
      l2 = l2,
      stopping.rounds = stopping.rounds,
      stopping.metric = stopping.metric,
      scale = scale,
      center = center
    ),
    extra = extra
  )
  if (verbose) msg2("Access H2O Flow at ", ip, ":", port, " in your browser", sep = "")
  outro(start.time, verbose = verbose)
  rt
} # rtemis::d_H2OAE