Implementation:Gretelai Gretel synthetics DGANConfig

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

Overview

Concrete tool for configuring DGAN model hyperparameters provided by the gretel-synthetics library.

Description

DGANConfig is a Python dataclass that stores all hyperparameters for the DoppelGANger model. It groups parameters into five categories: model structure (sequence length, noise dimensions, layer/unit counts), data transformation (normalization mode, feature scaling, example scaling, binary encoder cutoff), model initialization (forget bias for LSTM), loss function (gradient penalty coefficients and attribute loss coefficient), and training (optimizer learning rates, beta1 values, batch size, epochs, discriminator/generator rounds, CUDA, and mixed precision). The class provides a to_dict() method for serialization that returns a dictionary suitable for re-initializing a new instance.

Usage

Create a DGANConfig instance before constructing a DGAN model. At minimum, supply max_sequence_len and sample_len. The config is passed to the DGAN constructor and controls all subsequent network building, training, and generation behavior.

Code Reference

Source Location

Repository: gretel-synthetics
File: src/gretel_synthetics/timeseries_dgan/config.py
Lines: 27-142

Signature

@dataclass
class DGANConfig:
    # Model structure
    max_sequence_len: int
    sample_len: int

    attribute_noise_dim: int = 10
    feature_noise_dim: int = 10
    attribute_num_layers: int = 3
    attribute_num_units: int = 100
    feature_num_layers: int = 1
    feature_num_units: int = 100
    use_attribute_discriminator: bool = True

    # Data transformation
    normalization: Normalization = Normalization.ZERO_ONE
    apply_feature_scaling: bool = True
    apply_example_scaling: bool = True
    binary_encoder_cutoff: int = 150

    # Model initialization
    forget_bias: bool = False

    # Loss function
    gradient_penalty_coef: float = 10.0
    attribute_gradient_penalty_coef: float = 10.0
    attribute_loss_coef: float = 1.0

    # Training
    generator_learning_rate: float = 0.001
    generator_beta1: float = 0.5
    discriminator_learning_rate: float = 0.001
    discriminator_beta1: float = 0.5
    attribute_discriminator_learning_rate: float = 0.001
    attribute_discriminator_beta1: float = 0.5
    batch_size: int = 1024
    epochs: int = 400
    discriminator_rounds: int = 1
    generator_rounds: int = 1

    cuda: bool = True
    mixed_precision_training: bool = False

    def to_dict(self):
        return asdict(self)

Import

from gretel_synthetics.timeseries_dgan.config import DGANConfig

I/O Contract

Inputs

Name	Type	Required	Description
max_sequence_len	int	Yes	Length of time series sequences; all training and generated data will share this length
sample_len	int	Yes	Time steps generated per LSTM cell; must be a divisor of max_sequence_len
attribute_noise_dim	int	No (default 10)	Length of noise vector for attribute generation
feature_noise_dim	int	No (default 10)	Length of noise vector for feature generation
attribute_num_layers	int	No (default 3)	Number of layers in attribute generator MLP
attribute_num_units	int	No (default 100)	Units per layer in attribute generator MLP
feature_num_layers	int	No (default 1)	Number of LSTM layers in feature generator
feature_num_units	int	No (default 100)	Units per LSTM layer in feature generator
use_attribute_discriminator	bool	No (default True)	Whether to use a separate attribute discriminator
normalization	Normalization	No (default ZERO_ONE)	Scaling range for continuous variables: ZERO_ONE or MINUSONE_ONE
apply_feature_scaling	bool	No (default True)	Scale continuous variables to normalized range before training
apply_example_scaling	bool	No (default True)	Add per-example midpoint/half-range as additional attributes
binary_encoder_cutoff	int	No (default 150)	Use binary encoding instead of one-hot for columns exceeding this many unique values
forget_bias	bool	No (default False)	Initialize LSTM forget gate biases to 1 (matching TF1 behavior)
gradient_penalty_coef	float	No (default 10.0)	Coefficient for WGAN gradient penalty on feature discriminator
attribute_gradient_penalty_coef	float	No (default 10.0)	Coefficient for WGAN gradient penalty on attribute discriminator
attribute_loss_coef	float	No (default 1.0)	Weight of attribute discriminator loss in generator objective
generator_learning_rate	float	No (default 0.001)	Adam learning rate for generator
generator_beta1	float	No (default 0.5)	Adam beta1 for generator
discriminator_learning_rate	float	No (default 0.001)	Adam learning rate for feature discriminator
discriminator_beta1	float	No (default 0.5)	Adam beta1 for feature discriminator
attribute_discriminator_learning_rate	float	No (default 0.001)	Adam learning rate for attribute discriminator
attribute_discriminator_beta1	float	No (default 0.5)	Adam beta1 for attribute discriminator
batch_size	int	No (default 1024)	Number of examples per batch for training and generation
epochs	int	No (default 400)	Number of training epochs
discriminator_rounds	int	No (default 1)	Discriminator update steps per batch
generator_rounds	int	No (default 1)	Generator update steps per batch
cuda	bool	No (default True)	Use GPU if available
mixed_precision_training	bool	No (default False)	Enable automatic mixed precision to reduce memory usage

Outputs

Name	Type	Description
DGANConfig instance	DGANConfig	Dataclass instance holding all hyperparameters
to_dict()	dict	Dictionary representation of all fields, suitable for serialization or re-initialization

Usage Examples

Basic Example

from gretel_synthetics.timeseries_dgan.config import DGANConfig

config = DGANConfig(
    max_sequence_len=20,
    sample_len=5,
    batch_size=1000,
    epochs=10,
)

Advanced Example

from gretel_synthetics.timeseries_dgan.config import DGANConfig, Normalization

config = DGANConfig(
    max_sequence_len=100,
    sample_len=10,
    attribute_noise_dim=16,
    feature_noise_dim=16,
    feature_num_layers=2,
    feature_num_units=256,
    normalization=Normalization.MINUSONE_ONE,
    apply_example_scaling=True,
    gradient_penalty_coef=10.0,
    generator_learning_rate=0.0002,
    discriminator_learning_rate=0.0002,
    batch_size=512,
    epochs=200,
    cuda=True,
)

# Serialize and restore
config_dict = config.to_dict()
restored_config = DGANConfig(**config_dict)

Related Pages

Page Connections

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

Principle

Implementation

Heuristic

Environment