peal.generators.tabular_ddpm.TabularDDPMConfig¶

class peal.generators.tabular_ddpm.TabularDDPMConfig[source]¶

Parameters:

generator_type (str) – Registry name, "TabularDDPM".
input_dim (int) – Number of non-label feature columns the score network denoises.
embed_dim (int) – Width of the timestep embedding and of the hidden MLP layers.
num_timesteps (int) – Length of the forward noising chain (T).
model_name (str) – Filename of the saved state dict inside base_path.
var_schedule (str) – "linear" or "cosine" beta schedule.
base_path (str) – Directory used as model_dir when none is passed to the generator.
num_epochs (int, float) – Adam training budget. The batch size comes from GeneratorConfig.batch_size.
learning_rate (int, float) – Adam training budget. The batch size comes from GeneratorConfig.batch_size.
grad_scales (list of float) – Classifier-gradient weights tried by TabularDDPM.edit().
noise_steps_for_counterfactuals (list of int) – Forward-diffusion depths (t) tried by TabularDDPM.edit().
attack_iterations (int) – How many times the (steps x scales) grid is repeated in edit.
data (DataConfig) – Dataset config used when no dataset object is passed.

Methods

Attributes

`attack_iterations`
`base_path`
`data`	Pydantic description of a dataset, loadable from yaml.
`embed_dim`
`generator_type`	The type of generator that shall be used.
`grad_scales`
`input_dim`
`learning_rate`
`model_name`
`noise_steps_for_counterfactuals`
`num_epochs`
`num_timesteps`
`var_schedule`

generator_type: str = 'TabularDDPM'¶: The type of generator that shall be used.