graph LR
CTGAN["CTGAN"]
Generator["Generator"]
Discriminator["Discriminator"]
Residual["Residual"]
_apply_activate["_apply_activate"]
_gumbel_softmax["_gumbel_softmax"]
_cond_loss["_cond_loss"]
CTGAN -- "instantiates" --> Generator
CTGAN -- "instantiates" --> Discriminator
CTGAN -- "invokes" --> Generator
CTGAN -- "interacts with" --> Discriminator
CTGAN -- "utilizes" --> _cond_loss
Generator -- "uses" --> Residual
Discriminator -- "uses" --> Residual
Generator -- "calls" --> _apply_activate
Discriminator -- "calls" --> _apply_activate
_apply_activate -- "delegates to" --> _gumbel_softmax
The CTGAN subsystem is orchestrated by the CTGAN component, which serves as the primary interface and training manager. It is responsible for instantiating and coordinating the Generator and Discriminator neural networks, which form the core of the adversarial training process. The Generator creates synthetic data, while the Discriminator evaluates its authenticity, providing crucial feedback. Both the Generator and Discriminator leverage Residual blocks to build deeper and more stable network architectures. During their operation, both the Generator and Discriminator directly utilize the _apply_activate utility function to apply various activation functions, including the specialized _gumbel_softmax for handling categorical data. The CTGAN orchestrator also utilizes the _cond_loss function to compute conditional loss, enabling the generation of synthetic data based on specified conditions. This architecture highlights a clear separation of concerns, with CTGAN managing the overall training flow and the Generator and Discriminator handling the core data generation and evaluation tasks, supported by specialized utility components.
The primary orchestrator and public API for the CTGAN model. It manages the entire training lifecycle, including initializing the Generator and Discriminator, executing their adversarial training loops, and handling data sampling. It embodies the "Model Implementations," "Training Logic," "Sampling/Generation Logic," and "Public API/Interface" patterns.
Related Classes/Methods:
A neural network responsible for generating synthetic data samples. It learns to mimic the distribution of real data based on feedback from the Discriminator. This is a core "Model Implementation" component.
Related Classes/Methods:
A neural network that evaluates the authenticity of data samples, distinguishing between real data and synthetic data produced by the Generator. Its feedback is crucial for training the Generator. This is a core "Model Implementation" component.
Related Classes/Methods: None
A building block (likely a PyTorch module) that implements residual connections within the Generator and Discriminator networks. These connections help improve training stability and enable deeper network architectures.
Related Classes/Methods:
A utility function that applies various activation functions to the outputs of neural network layers, including specialized ones like Gumbel Softmax, ensuring non-linearity and appropriate output distributions.
Related Classes/Methods:
Implements the Gumbel Softmax activation function, specifically used for generating differentiable discrete samples. This is critical for CTGAN's ability to handle categorical data effectively.
Related Classes/Methods:
Computes the conditional loss component during the CTGAN training process. This loss function is essential for enabling conditional synthetic data generation, allowing users to specify conditions for the generated data.
Related Classes/Methods: