TabularPTGAN

Guide & worked example · Source

TabularPTGAN(
    hidden_dims=(128, 128, 128),
    critic_hidden_dims=None,
    noise_dim=None,
    batch_size=128,
    max_steps=1000,
    critic_steps=1,
    lr=0.0001,
    betas=(0.0, 0.9),
    optimizer='adam',
    temperature_ratio=0.5,
    coherency_weight=100.0,
    gp_weight=0.0,
    interpolate_noise=True,
    generator_dropout=0.1,
    critic_dropout=0.0,
    binary_dims=(),
    lower_bounds=None,
    upper_bounds=None,
    seed=None,
    device=None,
)

Parallelly tempered Wasserstein GAN for multimodal tabular data.

This implements Algorithm 3 of Sohn and Song (2025), arXiv:2411.11786v2. At each step it learns the joint family of convexly tempered targets alpha * X_1 + (1 - alpha) * X_2 and applies their coherency penalty to synchronize the critic across temperatures. Sampling at alpha=1 returns draws from the original target distribution.

Parameters are mostly shared with :class:TabularWGAN. The PTGAN-specific parameters are temperature_ratio (the point mass on alpha=1), coherency_weight, and interpolate_noise.

Attributes

Name	Description
batch_size
betas
binary_dims
coherency_weight
critic_dropout
critic_hidden_dims
critic_steps
device
generator_dropout
gp_weight
hidden_dims
history
interpolate_noise
lower_bounds
lr
max_steps
noise_dim
optimizer
params
seed
temperature_ratio
upper_bounds

Methods

Name	Description
fit	Train across convexly tempered target distributions.
sample	Draw rows at a requested tempering coordinate.
summary	Print a summary of the model results.
to	Move all model parameters to specified device.

fit

TabularPTGAN.fit(X, context=None, callback=None)

Train across convexly tempered target distributions.

Parameters

Name	Type	Description	Default
X	array-like of shape (n_rows, n_features)	Numeric training rows, converted to float32. Paired rows define the convexly tempered training targets.	required
context	array-like of shape (n_rows, n_context)	Aligned conditioning rows. Their interpolation follows the paired-data construction during training.	None
callback	callable	Called with the step and model during training.	None

Returns

Name	Type	Description
	TabularPTGAN	Self, with fitted networks and loss/penalty histories.

Notes

This is a GAN objective over tempered laws, not replica-exchange MCMC. The coherency penalty is a squared critic directional derivative along paired data differences. Optional gradient penalties are one-sided.

sample

TabularPTGAN.sample(n, context=None, alpha=1.0)

Draw rows at a requested tempering coordinate.

Parameters

Name	Type	Description	Default
n	int	Number of requested rows.	required
context	array-like of shape (n, n_context)	One conditioning row per draw.	None
alpha	float or array - like	Tempering coordinate in [0, 1], scalar or one value per row. The endpoint 1 targets the original data law; interior coordinates target convex combinations of paired rows.	1.0

Returns

Name	Type	Description
	torch.Tensor	Detached CPU tensor of shape (n, n_features), on training scale.

summary

TabularPTGAN.summary()

Print a summary of the model results.

to

TabularPTGAN.to(device)

Move all model parameters to specified device.

Parameters

Name	Type	Description	Default
device	torch.device or str	Target device (‘cuda’, ‘cpu’, etc.)	required

Returns

Name	Type	Description
	BaseEstimator	Self for method chaining