Add a reset_anonymization method

[pvk-developer]

Problem Description

It would be useful to have a reset_generator parameter for the transformers that when reverse transforming generate data (AnonymizedFaker and RegexGenerator). This way if we are using enforce_uniqueness we would be able to start over again with new value.

This change should propagate to create_anonymized_columns as those transformers are mainly interacted from there.

Expected behavior

from rdt.transformers import AnonymizedFaker

anonymizer = AnonymizedFaker(provider_name='job', function_name='job', enforce_uniqueness=True)
# There are 639 unique jobs in english
data = pd.DataFrame({'job': np.arange(639)})
tr_data = anonymizer.fit_transform(data, 'job')

# With this run we empty the unique values from the current faker instance
rev_data = anonymizer.reverse_transform(tr_data)

# This will not generate new data and will fail. There is no way to reset it currently.
anonymizer.reverse_transform(tr_data)

File ~/.virtualenvs/RDT/lib/python3.8/site-packages/faker/proxy.py:320, in UniqueProxy._wrap.<locals>.wrapper(*args, **kwargs)
    319 else:
--> 320     raise UniquenessException(f"Got duplicated values after {_UNIQUE_ATTEMPTS:,} iterations.")
    322 generated.add(retval)

UniquenessException: Got duplicated values after 1,000 iterations.

The above exception was the direct cause of the following exception:

Error                                     Traceback (most recent call last)
Cell In [1], line 12
      9 rev_data = anonymizer.reverse_transform(tr_data)
     11 # This will not generate new data and will fail. There is no way to reset it currently.
---> 12 anonymizer.reverse_transform(tr_data)

File ~/Projects/sdv-dev/RDT/rdt/transformers/base.py:362, in BaseTransformer.reverse_transform(self, data)
    359 data = data.copy()
    361 columns_data = self._get_columns_data(data, self.output_columns)
--> 362 reversed_data = self._reverse_transform(columns_data)
    363 data = data.drop(self.output_columns, axis=1)
    364 data = self._add_columns_to_data(data, reversed_data, self.columns)

File ~/Projects/sdv-dev/RDT/rdt/transformers/pii/anonymizer.py:149, in AnonymizedFaker._reverse_transform(self, data)
    144     reverse_transformed = np.array([
    145         self._function()
    146         for _ in range(sample_size)
    147     ], dtype=object)
    148 except faker.exceptions.UniquenessException as exception:
--> 149     raise Error(
    150         f'The Faker function you specified is not able to generate {sample_size} unique '
    151         'values. Please use a different Faker function for column '
    152         f"('{self.get_input_column()}')."
    153     ) from exception
    155 return reverse_transformed

Error: The Faker function you specified is not able to generate 639 unique values. Please use a different Faker function for column ('job').

Additional context

If we need to reuse the same transformer to generate data, we currently are not able to restart the current state.

[npatki]

Proposed API

1. Each relevant transformer (AnonymizedFaker, PseudoAnonymizedFaker, RegexGenerator) should have a reset_anonymization() method that resets the generation. Next time you call reverse_transform it will start from the beginning
2. HyperTransformer should have a reset_anonymization() method that calls reset_anonymization() on each relevant transformer in step 3

transformer = AnonymizedFaker(provider_name='job', function_name='job')
t = transformer.fit_transform(data)
transformer.reverse_transform(t)
transformer.reset_anonymization()
transformer.reverse_transform(t)

SDV Usage Example: SDV single table models have a randomize_samples parameter (docs). When set to False, the HyperTransformer should reset anonymization before calling sample.