Auxiliary function for dae.params — dae

Auxiliary function for setting up the default dae-related hyperparameters for midae.

Usage

dae_default(
  shuffle = TRUE,
  drop.last = FALSE,
  input.dropout = 0.2,
  hidden.dropout = 0.5,
  optimizer = "adamW",
  learning.rate = 0.001,
  weight.decay = 0.01,
  momentum = 0,
  dampening = 0,
  eps = 1e-08,
  rho = 0.9,
  alpha = 0.99,
  learning.rate.decay = 0,
  encoder.structure = c(256, 128, 64),
  latent.dim = 8,
  decoder.structure = c(64, 128, 256),
  act = "elu",
  init.weight = "he.normal.elu.dropout",
  scaler = "standard",
  initial.imp = "sample",
  lower = 0.25,
  upper = 0.75
)

Arguments

shuffle: Whether or not to shuffle training data. Default: TRUE.
drop.last: Whether or not to drop the last batch. Default: FALSE.
input.dropout: The dropout probability of the input layer. Default: 0.2.
hidden.dropout: The dropout probability of the hidden layers. Default: 0.5.
optimizer: The name of the optimizer. Options are : "adamW" (default), "adam", "adadelta", "adagrad", "rmsprop", or "sgd".
learning.rate: The learning rate. Default: 0.001.
weight.decay: Weight decay (L2 penalty). Default: 0.01.
momentum: Parameter for the "sgd" optimizer (default: 0). It is used for accelerating SGD in the relevant direction and dampens oscillations.
dampening: Dampening for momentum (default: 0) used for the "sgd" optimizer.
eps: A small positive value (default: 1e-08) used to prevent division by zero for optimizers "adamW", "adam", "adadelta","adagrad" and "rmsprop".
rho: Parameter for the "adadelta" optimizer (default: 0.9). A coefficient used for computing a running average of squared gradients.
alpha: Smoothing constant (default: 0.99) for the "rmsprop" optimizer.
learning.rate.decay: Learning rate decay (default: 0) for the "adagrad" optimizer.
encoder.structure: A vector indicating the structure of encoder. Default: c(256, 128, 64)
latent.dim: Size of the latent layer. Default: 8.
decoder.structure: A vector indicating the structure of decoder. Default: c(64, 128, 256)
act: The name of activation function. Can be: "relu", "elu" (default), "leaky.relu", "tanh", "sigmoid" and "identity".
init.weight: The distribution for weight initialization. Can be "he.normal", "he.uniform", "xavier.uniform", "xavier.normal", "he.normal.dropout", "he.normal.elu", "he.normal.elu.dropout" (default), "he.normal.selu" or "he.normal.leaky.relu".
scaler: The name of the scaler used for transforming numeric features. Can be "standard" (default), "minmax" , "decile", "robust" or "none".
initial.imp: The method for initial imputation. Can be "mean", "median" or "sample" (default).
lower: The lower quantile (0.25 by default) for scaler = "robust".
upper: The upper quantile (0.75 by default) for scaler = "robust".