Tabular Models

SDV has all the necessary tools to model and sample data from individual tables with a collection of tabular models, such as Copulas and CTGAN, while keeping all the complexity associated with their usage hidden away and offering a very simple and intuitive interface to the users.

All this functionalities are available in the sdv.tabular.BaseTabularModel, which is inherited and used by all the models implemented in the sdv.tabular sub-package.

Tabular Modeling Overview

The BaseTabularModel class offers a very simple API with only two methods, fit and sample, which handle the processes of preparing the data for modeling, calling an underlying model instance to learn the data and later on sample it, and finally post-processing the generated data to make it as similar to the original one as possible.

A regular modeling and sampling cycle performed within one of the BaseTabularModel subclasses has the following steps:

  1. An instance of the BaseTabularModel subclass is created. In this step, the user can either pass information about the data, such as field types or constraints, or a fully configured Table metadata object.

  2. The method is called with raw data from the table. During this step, the BaseTabularModel will:

    1. Fit the Table metadata object, unless it has already been fitted before.

    2. Store the number of rows that exist in the given data.

    3. Use the Table metadata to transform the data into numerical data, ready for the model. If the Table had any constraints defined using the transform strategy, this will also apply transformations to the data.

    4. Call the self._fit method, which has been implemented in the subclass, passing the numerical data.

  3. The subclass _fit method will then fit an instance of the underlying model class. For example, an instance of a copulas.multivariate.GaussianCopula will be created and fitted to the numerical data.

  4. In a later step, the BaseTabularModel.sample method will be called, optionally with an indication of the number of rows to sample. During this step, the BaseTabularModel will:

    1. Decide a number of rows to sample, which will either be the number of rows provided or the number of rows that the original table had.

    2. Call the self._sample method implemented by the subclass, which will use the underlying model to generate the indicated number of rows.

    3. Use the Table metadata to transform the sampled data back to the original format by passing the data to its revert_transform method. This will also revert any transformations performed by the Constraints that use the transform strategy.

    4. If there is any Constraint that is using the reject_sampling strategy, use the Table metadata to drop the invalid rows and repeat steps b and c until enough valid rows have been generated.

A part from the previous steps, the BaseTabularModel also offers a couple of minor functionalities:

  • get_metadata: Returns the Table metadata object that has been fitted to the data.

  • save: Saves the complete Tabular Model in a file using pickle.

  • load: Loads a previously saved model from a pickle file.

Implementing a Tabular Model

In order to implement a new Tabular Model, all you need is to create a class that inherits from dsv.tabular.base.BaseTabularModel and implement at least these two methods:

  • _fit: Gets clean numerical data as input and fits an underlying model.

  • _sample: Samples the indicated number of rows from the fitted model.

BaseTabularModel Arguments

When creating a subclass we will probably want to capture some arguments to the __init__ method.

The base __init__ method implemented in the BaseTabularModel expects the following arguments:

  • field_names: List of names of the fields that need to be modeled and included in the generated output data. Any additional fields found in the data will be ignored and will not be included in the generated output. If None, all the fields found in the data are used.

  • field_types: Dictionary specifying the data types and subtypes of the fields that will be modeled. Field types and subtypes combinations must be compatible with the SDV Metadata Schema.

  • field_transformers: Dictionary specifying which transformers to use for each field. Available transformers are:

    • integer: Uses a NumericalTransformer of dtype int.

    • float: Uses a NumericalTransformer of dtype float.

    • categorical: Uses a CategoricalTransformer without gaussian noise.

    • categorical_fuzzy: Uses a CategoricalTransformer adding gaussian noise.

    • one_hot_encoding: Uses a OneHotEncodingTransformer.

    • label_encoding: Uses a LabelEncodingTransformer.

    • boolean: Uses a BooleanTransformer.

    • datetime: Uses a DatetimeTransformer.

  • anonymize_fields: Dict specifying which fields to anonymize and what faker category they belong to.

  • primary_key: Name of the field which is the primary key of the table.

  • constraints: List of Constraint objects or dicts.

  • table_metadata: Table metadata instance or dict representation. If given alongside any other metadata-related arguments, an exception will be raised. If not given at all, it will be built using the other arguments or learned from the data.

Subclasses can extend this list by adding their own arguments, or even simply implement their own __init__ method. However, capturing these explicitly and passing them to the super().__init__ method is the recommended way to initialize a BaseTabularModel subclass.

We can see such implementation in the __init__ method of the sdv.tabular.ctgan.CTGAN model, which adds a few arguments to the class but still captures all the other arguments and calls the super().__init__ method with them:

def __init__(self, field_names=None, field_types=None, field_transformers=None,
             anonymize_fields=None, primary_key=None, constraints=None, table_metadata=None,
             epochs=300, log_frequency=True, embedding_dim=128, generator_dim=(256, 256),
             discriminator_dim=(256, 256), l2scale=1e-6, batch_size=500):
    self._embedding_dim = embedding_dim
    self._generator_dim = generator_dim
    self._discriminator_dim = discriminator_dim
    self._l2scale = l2scale
    self._batch_size = batch_size
    self._epochs = epochs
    self._log_frequency = log_frequency

By doing these, not only can the CTGAN take advantage of all the functionalities from the base class, but also the signature and API reference exposes all the accepted arguments appropriately.

_fit method

The _fit method only expects one argument called table_data, which is a pandas.DataFrame that contains numerical data only.

Within this method, you can perform any steps necessary to fit your model. For example, we can see how the sdv.tabular.ctgan.CTGAN._fit method creates an instance of the underlying model, CTGANSynthesizer, and prepares the list of categorical columns that it expects alongside the data.

def _fit(self, table_data):
    """Fit the model to the table.

        table_data (pandas.DataFrame):
            Data to be learned.
    self._model = self._CTGAN_CLASS(
    categoricals = [
        for field, meta in self._metadata.get_fields().items()
        if meta['type'] == 'categorical'


Here you can also see that some of the hyperparameters for the CTGANSynthesizer class are being taken from the instance itself, where the __init__ method stored them beforehand.

_sample method

The _sample method only expects one argument called num_rows, which is an integer that indicates the number of rows that need to be sampled. In most cases, such as the CTGAN example shown below, all this method does is call the sample method of the underlying model:

def _sample(self, num_rows):
    """Sample the indicated number of rows from the model.

        num_rows (int):
            Amount of rows to sample.

            Sampled data.
    return self._model.sample(num_rows)