Discriminator

The number of columns in the data set

data_dim

A list of the number of neurons per layer, the length of the list determines the number of hidden layers

hidden_units

The dropout rate for each hidden layer

dropout_rate

Switch between a sigmoid and linear output layer (the sigmoid is needed for the original GAN value function)

sigmoid

Provides a torch::nn_module with a simple fully connected neural
net, for use as the default architecture for tabular data in RGAN.

An easy way to get started with Generative Adversarial Nets (GAN) in R. The GAN algorithm was initially
described by Goodfellow et al. 2014 <https://proceedings.neurips.cc/paper/2014/file/5ca3e9b122f61f8f06494c97b1afccf3-Paper.pdf>. A GAN can be used to learn the joint distribution of complex data by
comparison. A GAN consists of two neural networks a Generator and a Discriminator, where the two
neural networks play an adversarial minimax game.
Built-in GAN models make the training of GANs in R possible in one line and make it easy to
experiment with different design choices (e.g. different network architectures, value functions, optimizers).
The built-in GAN models work with tabular data (e.g. to produce synthetic data) and image data.
Methods to post-process the output of GAN models to enhance the quality of samples are available.

Discriminator: Discriminator

Description

Usage

Arguments

Value