autotab (version 0.1.2)

mog_prior: Mixture-of-Gaussians (MoG) prior in AutoTab

Description

AutoTab allows the encoder prior to be either a single Gaussian (prior = "single_gaussian") or a mixture of Gaussians (prior = "mixture_gaussian"). When using a MoG prior, the user may optionally specify the component means, variances, and mixture weights. The user may also indicate if the means, variances, and mixture weights can be learned or not using learnable_mog with a logical TRUE/FALSE.

Prior options in <code>VAE_train()</code>

• prior: character, one of "single_gaussian" or "mixture_gaussian".

• K: integer, number of mixture components when prior = "mixture_gaussian".

• learnable_mog: logical; if TRUE, the MoG parameters (means, log-variances, and mixture weights) are learned during training.

• mog_means: optional numeric matrix of size K x latent_dim, giving the initial means for each mixture component in the latent space.

• mog_log_vars: optional numeric matrix of size K x latent_dim, giving initial log-variances for each component.

• mog_weights: optional numeric vector of length K, giving initial mixture weights that should sum to 1.

Shape of <code>mog_means</code>

For a latent dimension latent_dim and K mixture components, mog_means must be a numeric matrix with:

• nrow(mog_means) == K

• ncol(mog_means) == latent_dim

Each row corresponds to the mean vector of one mixture component in the latent space.

Details

If prior = "single_gaussian", the prior is a standard Normal in the latent space and the MoG-related arguments (K, mog_means, mog_log_vars, mog_weights, learnable_mog) are ignored.

When prior = "mixture_gaussian":

• If learnable_mog = FALSE, then mog_means, mog_log_vars, and mog_weights must be supplied and are treated as fixed.

• If learnable_mog = TRUE, any of mog_means, mog_log_vars, or mog_weights that are provided are used as initial values and are updated during training. If they are omitted, AutoTab initializes them internally (e.g., Normal or zero-centered initializations).

See Also

VAE_train()

Examples

# Examples of a Mixture-of-Gaussians (MoG) prior in AutoTab

# These examples illustrate:
# 1) learnable_mog = FALSE with fixed MoG parameters
# 2) learnable_mog = TRUE with preset means/variances/weights
# 3) learnable_mog = TRUE with all MoG parameters learned

# Required packages for the full example:
# - AutoTab (this package)
# - keras
# - caret (for dummyVars)

# \donttest{
if (requireNamespace("caret", quietly = TRUE) &&
    reticulate::py_module_available("tensorflow")) {

  # -------------------------------
  # Data simulation and preparation
  # -------------------------------
  set.seed(123)
  age        <- rnorm(100, mean = 45, sd = 12)
  income     <- rnorm(100, mean = 60000, sd = 15000)
  bmi        <- rnorm(100, mean = 25, sd = 4)
  smoker     <- rbinom(100, 1, 0.25)
  exercise   <- rbinom(100, 1, 0.6)
  diabetic   <- rbinom(100, 1, 0.15)
  education  <- sample(
    c("HighSchool", "College", "Graduate"),
    100, replace = TRUE,
    prob = c(0.4, 0.4, 0.2)
  )
  marital    <- sample(
    c("Single", "Married", "Divorced"),
    100, replace = TRUE
  )
  occupation <- sample(
    c("Clerical", "Technical", "Professional", "Other"),
    100, replace = TRUE
  )

  data_final <- data.frame(
    age, income, bmi,
    smoker, exercise, diabetic,
    education, marital, occupation
  )

  # One-hot encode categorical variables
  encoded_data  <- caret::dummyVars(~ education + marital + occupation,
                                    data = data_final)
  one_hot_coded <- as.data.frame(predict(encoded_data, newdata = data_final))

  data_cont <- subset(data_final, select = c(age, income, bmi))
  Continuous_MinMaxScaled <- as.data.frame(
    lapply(data_cont, min_max_scale)  # min_max_scale is an AutoTab function
  )
  data_bin <- subset(data_final, select = c(smoker, exercise, diabetic))

  # Bind all data together
  data <- cbind(Continuous_MinMaxScaled, data_bin, one_hot_coded)

  # Step 1: Extract and set feature distributions
  feat_dist   <- feat_reorder(extracting_distribution(data_final), data)
  rownames(feat_dist) <- NULL
  set_feat_dist(feat_dist)

  # Step 2: Define encoder / decoder architectures and MoG parameters
  encoder_info <- list(
    list("dense", 25, "relu"),
    list("dense", 50, "relu")
  )

  decoder_info <- list(
    list("dense", 50, "relu"),
    list("dense", 25, "relu")
  )

  mog_means <- matrix(
    c(rep(-5, 5), rep(0, 5), rep(5, 5)),
    nrow = 3, byrow = TRUE
  )
  mog_log_vars <- matrix(log(0.5), nrow = 3, ncol = 5)
  mog_weights  <- c(0.3, 0.4, 0.3)

  # ------------------------------------------------------------
  # Example 1: learnable_mog = FALSE (fixed MoG)
  # ------------------------------------------------------------
  reset_seeds(1234)

  training <- VAE_train(
    data         = data,
    encoder_info = encoder_info,
    decoder_info = decoder_info,
    Lip_en       = 0,
    pi_enc       = 0,
    lip_dec      = 0,
    pi_dec       = 0,
    latent_dim   = 5,
    epoch        = 200,
    beta         = 0.01,
    kl_warm      = TRUE,
    beta_epoch   = 20,
    temperature  = 0.5,
    batchsize    = 16,
    wait         = 20,
    lr           = 0.001,
    K            = 3,
    mog_means    = mog_means,
    mog_log_vars = mog_log_vars,
    mog_weights  = mog_weights,
    prior        = "mixture_gaussian",
    learnable_mog = FALSE
  )


  # -------------------------------------------------------------------
  # Example 2: learnable_mog = TRUE with preset MoG params
  # -------------------------------------------------------------------
  reset_seeds(1234)

  training <- VAE_train(
    data         = data,
    encoder_info = encoder_info,
    decoder_info = decoder_info,
    Lip_en       = 0,
    pi_enc       = 0,
    lip_dec      = 0,
    pi_dec       = 0,
    latent_dim   = 5,
    epoch        = 200,
    beta         = 0.01,
    kl_warm      = TRUE,
    beta_epoch   = 20,
    temperature  = 0.5,
    batchsize    = 16,
    wait         = 20,
    lr           = 0.001,
    K            = 3,
    mog_means    = mog_means,
    mog_log_vars = mog_log_vars,
    mog_weights  = mog_weights,
    prior        = "mixture_gaussian",
    learnable_mog = TRUE
  )


  # -----------------------------------------------------------------------
  # Example 3: learnable_mog = TRUE with all MoG params learned
  #           (mog_means, mog_log_vars, mog_weights = NULL)
  # -----------------------------------------------------------------------
  reset_seeds(1234)

  training <- VAE_train(
    data         = data,
    encoder_info = encoder_info,
    decoder_info = decoder_info,
    Lip_en       = 0,
    pi_enc       = 0,
    lip_dec      = 0,
    pi_dec       = 0,
    latent_dim   = 5,
    epoch        = 200,
    beta         = 0.01,
    kl_warm      = TRUE,
    beta_epoch   = 20,
    temperature  = 0.5,
    batchsize    = 16,
    wait         = 20,
    lr           = 0.001,
    K            = 3,
    mog_means    = NULL,
    mog_log_vars = NULL,
    mog_weights  = NULL,
    prior        = "mixture_gaussian",
    learnable_mog = TRUE
  )


}
# }