Implementation:Gretelai Gretel synthetics ACTGAN Init

Knowledge Sources	gretel-synthetics
Domains	Synthetic_Data, GAN, Tabular_Data
Last Updated	2026-02-14 19:00 GMT

Overview

Concrete tool for constructing and configuring an ACTGAN model instance provided by the gretel-synthetics library.

Description

The ACTGAN.__init__ method creates a fully configured ACTGAN model by accepting metadata-level parameters (field names, field types, field transformers, primary key, constraints, table metadata), architecture parameters (embedding dimension, generator and discriminator layer sizes), optimizer parameters (learning rates, weight decay), training parameters (batch size, epochs, PAC size, discriminator steps), conditional generation parameters (conditional vector type, column selection probability, reconstruction loss coefficient, force conditioning), and data transformation parameters (binary encoder cutoff, cluster-based normalizer sample size, automatic datetime detection). It delegates metadata handling to the SDV BaseTabularModel superclass and stores all GAN-specific hyperparameters in the _model_kwargs dictionary for later use when building the underlying ACTGANSynthesizer.

Usage

Call ACTGAN(...) with desired hyperparameters before calling fit(). Sensible defaults are provided for all parameters: embedding_dim=128, generator_dim=(256,256), discriminator_dim=(256,256), batch_size=500, epochs=300, pac=10.

Code Reference

Source Location

Repository: gretel-synthetics
File: src/gretel_synthetics/actgan/actgan_wrapper.py
Lines: 338-413

Signature

class ACTGAN(_ACTGANModel):
    _MODEL_CLASS = ACTGANSynthesizer

    def __init__(
        self,
        field_names: Optional[List[str]] = None,
        field_types: Optional[Dict[str, dict]] = None,
        field_transformers: Optional[Dict[str, Union[BaseTransformer, str]]] = None,
        auto_transform_datetimes: bool = False,
        anonymize_fields: Optional[Dict[str, str]] = None,
        primary_key: Optional[str] = None,
        constraints: Optional[Union[List[Constraint], List[dict]]] = None,
        table_metadata: Optional[Union[Metadata, dict]] = None,
        embedding_dim: int = 128,
        generator_dim: Sequence[int] = (256, 256),
        discriminator_dim: Sequence[int] = (256, 256),
        generator_lr: float = 2e-4,
        generator_decay: float = 1e-6,
        discriminator_lr: float = 2e-4,
        discriminator_decay: float = 1e-6,
        batch_size: int = 500,
        discriminator_steps: int = 1,
        binary_encoder_cutoff: int = 500,
        binary_encoder_nan_handler: Optional[str] = None,
        cbn_sample_size: Optional[int] = 250_000,
        log_frequency: bool = True,
        verbose: bool = False,
        epochs: int = 300,
        epoch_callback: Optional[Callable[[EpochInfo], None]] = None,
        pac: int = 10,
        cuda: bool = True,
        learn_rounding_scheme: bool = True,
        enforce_min_max_values: bool = True,
        conditional_vector_type: ConditionalVectorType = ConditionalVectorType.SINGLE_DISCRETE,
        conditional_select_mean_columns: Optional[float] = None,
        conditional_select_column_prob: Optional[float] = None,
        reconstruction_loss_coef: float = 1.0,
        force_conditioning: bool = False,
    ):

Import

from gretel_synthetics.actgan.actgan_wrapper import ACTGAN

I/O Contract

Inputs

Name	Type	Required	Description
field_names	`Optional[List[str]]`	No	List of column names to model. If None, all columns in the data are used.
field_types	`Optional[Dict[str, dict]]`	No	Dictionary specifying SDV-compatible data types and subtypes per field.
field_transformers	`Optional[Dict[str, Union[BaseTransformer, str]]]`	No	Dictionary specifying which RDT transformers to use per field.
auto_transform_datetimes	`bool`	No	If True, auto-detect datetime columns and convert them to Unix timestamps before fitting. Default: False.
embedding_dim	`int`	No	Size of the random noise vector passed to the generator. Default: 128.
generator_dim	`Sequence[int]`	No	Tuple of output sizes for each generator residual layer. Default: (256, 256).
discriminator_dim	`Sequence[int]`	No	Tuple of output sizes for each discriminator linear layer. Default: (256, 256).
generator_lr	`float`	No	Learning rate for the generator Adam optimizer. Default: 2e-4.
discriminator_lr	`float`	No	Learning rate for the discriminator Adam optimizer. Default: 2e-4.
batch_size	`int`	No	Number of data samples per training step. Must be divisible by 2 and pac. Default: 500.
epochs	`int`	No	Number of training epochs. Default: 300.
pac	`int`	No	Number of samples grouped for the discriminator (packing). Default: 10.
cuda	`bool`	No	Whether to use CUDA for GPU computation. Default: True.
conditional_vector_type	`ConditionalVectorType`	No	Type of conditional vector: SINGLE_DISCRETE or ANYWAY. Default: SINGLE_DISCRETE.
epoch_callback	`Optional[Callable[[EpochInfo], None]]`	No	Optional callable invoked after each epoch with an EpochInfo instance.
force_conditioning	`bool`	No	If True, directly set conditioned column values in generated data. Default: False.

Outputs

Name	Type	Description
(instance)	`ACTGAN`	A fully configured ACTGAN model ready for fitting with `.fit(data)`.

Usage Examples

Basic Example

from gretel_synthetics.actgan.actgan_wrapper import ACTGAN

# Create an ACTGAN model with default hyperparameters
model = ACTGAN(
    epochs=100,
    verbose=True,
    cuda=True,
)

Advanced Example with Conditional Generation

from gretel_synthetics.actgan.actgan_wrapper import ACTGAN
from gretel_synthetics.actgan.structures import ConditionalVectorType

model = ACTGAN(
    embedding_dim=256,
    generator_dim=(512, 512),
    discriminator_dim=(512, 512),
    batch_size=1000,
    epochs=200,
    pac=10,
    conditional_vector_type=ConditionalVectorType.ANYWAY,
    conditional_select_mean_columns=2.0,
    reconstruction_loss_coef=2.0,
    auto_transform_datetimes=True,
    verbose=True,
)

Related Pages

Implements Principle

Principle:Gretelai_Gretel_synthetics_ACTGAN_Model_Initialization

Requires Environment

Uses Heuristic

Heuristic:Gretelai_Gretel_synthetics_Binary_Encoder_Cutoff

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment