synthcity.plugins.privacy.plugin_decaf module

Reference: Boris van Breugel, Trent Kyono, Jeroen Berrevoets, Mihaela van der Schaar “DECAF: Generating Fair Synthetic Data Using Causally-Aware Generative Networks”(2021).

class DECAFPlugin(n_iter: int = 100, n_iter_baseline: int = 1000, generator_n_layers_hidden: int = 2, generator_n_units_hidden: int = 500, generator_nonlin: str = 'relu', generator_dropout: float = 0.1, generator_opt_betas: tuple = (0.5, 0.999), discriminator_n_layers_hidden: int = 2, discriminator_n_units_hidden: int = 500, discriminator_nonlin: str = 'leaky_relu', discriminator_n_iter: int = 1, discriminator_dropout: float = 0.1, discriminator_opt_betas: tuple = (0.5, 0.999), lr: float = 0.001, batch_size: int = 200, random_state: int = 0, clipping_value: int = 1, lambda_gradient_penalty: float = 10, lambda_privacy: float = 1, eps: float = 1e-08, alpha: float = 1, rho: float = 1, weight_decay: float = 0.01, l1_g: float = 0, l1_W: float = 1, grad_dag_loss: bool = False, struct_learning_enabled: bool = True, struct_learning_n_iter: int = 1000, struct_learning_search_method: str = 'tree_search', struct_learning_score: str = 'k2', struct_max_indegree: int = 4, encoder_max_clusters: int = 10, device: Any = device(type='cpu'), workspace: pathlib.Path = PosixPath('workspace'), compress_dataset: bool = False, sampling_patience: int = 500, **kwargs: Any)

Bases: synthcity.plugins.core.plugin.Plugin

Inheritance diagram of synthcity.plugins.privacy.plugin_decaf.DECAFPlugin

DECAF (DEbiasing CAusal Fairness) plugin.

Parameters

  • n_iter – int Number of training iterations.

  • generator_n_layers_hidden – int Number of hidden layers in the generator.

  • generator_n_units_hidden – Number of neurons in the hidden layers of the generator.

  • generator_nonlin – str Nonlinearity used by the generator for the hidden layers: leaky_relu, relu, gelu etc.

  • generator_dropout – float Generator dropout.

  • generator_opt_betas – tuple Generator initial decay rates for the Adam optimizer

  • discriminator_n_layers_hidden – int Number of hidden layers in the discriminator.

  • discriminator_n_units_hidden – int Number of neurons in the hidden layers of the discriminator.

  • discriminator_nonlin – str Nonlinearity used by the discriminator for the hidden layers: leaky_relu, relu, gelu etc.

  • discriminator_n_iter – int Discriminator number of iterations(default = 1)

  • discriminator_dropout – float Discriminator dropout

  • discriminator_opt_betas – tuple Discriminator initial decay rates for the Adam optimizer

  • lr – float Learning rate

  • weight_decay – float Optimizer weight decay

  • batch_size – int Batch size

  • random_state – int Random seed

  • clipping_value – int Gradient clipping value

  • lambda_gradient_penalty – float Gradient penalty factor used for training the GAN.

  • lambda_privacy – float Privacy factor used the AdsGAN loss.

  • eps – float = 1e-8, Noise added to the privacy loss

  • alpha – float Gradient penalty weight for real samples.

  • rho – float DAG loss factor

  • l1_g – float = 0 l1 regularization loss for the generator

  • l1_W – float = 1 l1 regularization factor for l1_g

  • struct_learning_enabled – bool Enable DAG learning outside DECAF.

  • struct_learning_n_iter – int Number of iterations for the DAG search.

  • struct_learning_search_method – str DAG search strategy: hillclimb, pc, tree_search, mmhc, exhaustive, d-struct

  • struct_learning_score – str DAG search scoring strategy: k2, bdeu, bic, bds

  • struct_max_indegree – int Max parents in the DAG.

  • encoder_max_clusters – int Number of clusters used for tabular encoding

  • device – Any = DEVICE torch device used for training.

  • arguments (# Core Plugin) –

  • workspace – Path. Optional Path for caching intermediary results.

  • compress_dataset – bool. Default = False. Drop redundant features before training the generator.

  • sampling_patience – int. Max inference iterations to wait for the generated data to match the training schema.

Example

>>> from sklearn.datasets import load_iris
>>> from synthcity.plugins import Plugins
>>>
>>> X, y = load_iris(as_frame = True, return_X_y = True)
>>> X["target"] = y
>>>
>>> plugin = Plugins().get("decaf", n_iter = 100)
>>> plugin.fit(X)
>>>
>>> plugin.generate(50)
class Config

Bases: object

arbitrary_types_allowed = True
validate_assignment = True
fit(X: Union[synthcity.plugins.core.dataloader.DataLoader, pandas.core.frame.DataFrame], *args: Any, **kwargs: Any) → Any

Training method the synthetic data plugin.

Parameters

  • X – DataLoader. The reference dataset.

  • cond –

    Optional, Union[pd.DataFrame, pd.Series, np.ndarray] Optional Training Conditional. The training conditional can be used to control to output of some models, like GANs or VAEs. The content can be anything, as long as it maps to the training dataset X. Usage example:

    >>> from sklearn.datasets import load_iris
>>> from synthcity.plugins.core.dataloader import GenericDataLoader
>>> from synthcity.plugins.core.constraints import Constraints
>>>
>>> # Load in `test_plugin` the generative model of choice
>>> # ....
>>>
>>> X, y = load_iris(as_frame=True, return_X_y=True)
>>> X["target"] = y
>>>
>>> X = GenericDataLoader(X)
>>> test_plugin.fit(X, cond=y)
>>>
>>> count = 10
>>> X_gen = test_plugin.generate(count, cond=np.ones(count))
>>>
>>> # The Conditional only optimizes the output generation
>>> # for GANs and VAEs, but does NOT guarantee the samples
>>> # are only from that condition.
>>> # If you want to guarantee that output contains only
>>> # "target" == 1 samples, use Constraints.
>>>
>>> constraints = Constraints(
>>>     rules=[
>>>         ("target", "==", 1),
>>>     ]
>>> )
>>> X_gen = test_plugin.generate(count,
>>>         cond=np.ones(count),
>>>         constraints=constraints
>>>        )
>>> assert (X_gen["target"] == 1).all()

Returns

self

classmethod fqdn() → str

The Fully-Qualified name of the plugin.

generate(count: Optional[int] = None, constraints: Optional[synthcity.plugins.core.constraints.Constraints] = None, random_state: Optional[int] = None, **kwargs: Any) → synthcity.plugins.core.dataloader.DataLoader

Synthetic data generation method.

Parameters

  • count – optional int. The number of samples to generate. If None, it generated len(reference_dataset) samples.

  • cond – Optional, Union[pd.DataFrame, pd.Series, np.ndarray]. Optional Generation Conditional. The conditional can be used only if the model was trained using a conditional too. If provided, it must have count length. Not all models support conditionals. The conditionals can be used in VAEs or GANs to speed-up the generation under some constraints. For model agnostic solutions, check out the constraints parameter.

  • constraints –

    optional Constraints. Optional constraints to apply on the generated data. If none, the reference schema constraints are applied. The constraints are model agnostic, and will filter the output of the generative model. The constraints are a list of rules. Each rule is a tuple of the form (<feature>, <operation>, <value>).

    Valid Operations:

    • ”<”, “lt” : less than <value>

    • ”<=”, “le”: less or equal with <value>

    • ”>”, “gt” : greater than <value>

    • ”>=”, “ge”: greater or equal with <value>

    • ”==”, “eq”: equal with <value>

    • ”in”: valid for categorical features, and <value> must be array. for example, (“target”, “in”, [0, 1])

    • ”dtype”: <value> can be a data type. For example, (“target”, “dtype”, “int”)

    Usage example:
    >>> from synthcity.plugins.core.constraints import Constraints
>>> constraints = Constraints(
>>>   rules=[
>>>             ("InterestingFeature", "==", 0),
>>>         ]
>>>     )
>>>
>>> syn_data = syn_model.generate(
        count=count,
        constraints=constraints
    ).dataframe()
>>>
>>> assert (syn_data["InterestingFeature"] == 0).all()

  • random_state – optional int. Optional random seed to use.

Returns

<count> synthetic samples

get_dag(X: pandas.core.frame.DataFrame, struct_learning_search_method: Optional[str] = None, as_index: bool = False) → Any
static hyperparameter_space(**kwargs: Any) → List[synthcity.plugins.core.distribution.Distribution]

Returns the hyperparameter space for the derived plugin.

static load(buff: bytes) → Any
static load_dict(representation: dict) → Any
static name() → str

The name of the plugin.

plot(plt: Any, X: synthcity.plugins.core.dataloader.DataLoader, count: Optional[int] = None, plots: list = ['marginal', 'associations', 'tsne'], **kwargs: Any) → Any

Plot the real-synthetic distributions.

Parameters

  • plt – output

  • X – DataLoader. The reference dataset.

Returns

self

classmethod sample_hyperparameters(*args: Any, **kwargs: Any) → Dict[str, Any]

Sample value from the hyperparameter space for the current plugin.

classmethod sample_hyperparameters_optuna(trial: Any, *args: Any, **kwargs: Any) → Dict[str, Any]
save() → bytes
save_dict() → dict
save_to_file(path: pathlib.Path) → bytes
schema() → synthcity.plugins.core.schema.Schema

The reference schema

schema_includes(other: Union[synthcity.plugins.core.dataloader.DataLoader, pandas.core.frame.DataFrame]) → bool

Helper method to test if the reference schema includes a Dataset

Parameters

other – DataLoader. The dataset to test

Returns

bool, if the schema includes the dataset or not.

training_schema() → synthcity.plugins.core.schema.Schema

The internal schema

static type() → str

The type of the plugin.

static version() → str

API version

plugin

alias of synthcity.plugins.privacy.plugin_decaf.DECAFPlugin