Compare real vs synthetic marginals from a trained generator

For continuous variables, overlays kernel densities of the real training variables against draws from the trained generator. For binary (0/1) variables – detected at training time and stored in the iconic_gan object – uses side-by-side bar charts of the proportion of 1s, which is the appropriate visualisation for discrete data. One panel per variable.

Usage

plot_gan_diagnostics(trained_gan, real_data, n_draw = NULL, M_matrix = NULL)

Arguments

trained_gan

An iconic_gan from train_gan_on_real_data().

real_data

The data frame the generator was trained on.

n_draw

Synthetic rows to draw for the overlay. Default: number of real rows.

M_matrix

Optional mediator panel (features x samples) for feature-level marginal comparison. When supplied and the GAN carries a feature_texture, an additional panel of density overlays compares real vs copula-sampled mediator marginals.

Value

Called for its side effects; invisibly returns NULL.

Examples

if (check_torch_setup()) {
  dat <- load_real_input_data(example = TRUE)
  gan <- train_gan_on_real_data(dat$gan_training_data, epochs = 5,
    verbose = FALSE)
  plot_gan_diagnostics(gan, dat$gan_training_data)
}