Create GaussianCopula, CTGAN, TVAE and CopulaGAN synthesizer classes

[amontanez24]

Problem Description

Once the BaseSynthesizer class is created, we can add the rest of the synthesizers.

Expected behavior

• Create GaussianCopulaSynthesizer in single_table module

  • Inherits from BaseSynthesizer
  • Has following special __init__ attributes:
    • numerical_distributions (dict): maps a numerical column name to the name of a distribution. This overrides the default distribution.
      • (default) None: No overrides
      • Valid options are: 'gaussian', 'beta', 'truncated_gaussian', 'uniform', 'gamma', 'gaussian_kde'
    • default_distribution: The distribution shape to use by default
      • (default) 'beta': use the beta distribution
      • Valid options: 'gaussian', 'beta', 'uniform', 'gamma', 'gaussian_kde', 'truncated_gaussian'

• Create CTGANSynthesizer in single_table module

  • Inherits from BaseSynthesizer
  • Has following special __init__ parms:
    • embedding_dim
    • generator_dim
    • discriminator_dim
    • generator_lr
    • generator_decay
    • discriminator_lr
    • discriminator_decay
    • batch_size
    • discriminator_steps
    • log_frequency
    • verbose
    • epochs
    • pac
    • cuda

• Create TVAESynthesizer in single_table module

  • Inherits from BaseSynthesizer
  • Has following special __init__ parms:
    • embedding_dim
    • compress_dims
    • decompress_dims
    • l2scale
    • batch_size
    • epochs
    • loss_factor
    • cuda

• Create CopulaGANSynthesizer in single_table module

  • Inherits from CTGANSynthesizer
  • Has following special __init__ attributes:
    • numerical_distributions (dict): maps a numerical column name to the name of a distribution. This overrides the default distribution.
      • (default) None: No overrides
      • Valid options are: 'gaussian', 'beta', 'truncated_gaussian', 'uniform', 'gamma', 'gaussian_kde'
    • default_distribution: The distribution shape to use by default
      • (default) 'beta': use the beta distribution
      • Valid options: 'gaussian', 'beta', 'uniform', 'gamma', 'gaussian_kde', 'truncated_gaussian'
    • embedding_dim
    • generator_dim
    • discriminator_dim
    • generator_lr
    • generator_decay
    • discriminator_lr
    • discriminator_decay
    • batch_size
    • discriminator_steps
    • log_frequency
    • verbose
    • epochs
    • pac
    • cuda

Additional context

• We should use the current single table models as influence for the types of each parameter unless it is specifically specified above.
• There will be private attributes for each of these classes not listed above that we may need. Again, draw influence or just migrate over from the current single table classes