GreaterThan constraint between Date columns raises TypeError

[MLjungg]

Environment Details

Please indicate the following details about the environment in which you found the bug:

• SDV version: 0.9.1
• Python version: 3.8
• Operating System: OSX

Error Description

I'm trying to define a greaterThan constraint with CopulaGAN using the handling_strategy="transform" between dates. This causes the following error:

"TypeError: ufunc 'exp' not supported for the input types, and the inputs could not be safely coerced to any supported types according to the casting rule ''safe''"

Steps to reproduce

The error can be reproduced with the following code:

from sdv.tabular import CopulaGAN
from sdv.constraints import GreaterThan
from sdv.demo import load_tabular_demo

data = load_tabular_demo('student_placements')
data.head()

date_constraint = GreaterThan(
    low='start_date',
    high='end_date',
    handling_strategy='transform'
)

constraints = [
    date_constraint,
]

model = CopulaGAN(epochs=50, constraints=constraints)
model.fit(data)
model.sample(100)

Reason for error

The problem occurs because np.exp() is computed on a DateType64[ns] as the column is reversed_transformed:

    def reverse_transform(self, table_data):
        """Reverse transform the table data.

        The transformation is reversed by computing an exponential of the given
        value, converting it to the original dtype, subtracting 1 and finally
        clipping the value to 0 on the low end to ensure the value is positive.

        Finally, the obtained value is added to the ``low`` column to get the final
        ``high`` value.

        Args:
            table_data (pandas.DataFrame):
                Table data.

        Returns:
            pandas.DataFrame:
                Transformed data.
        """
        table_data = table_data.copy()
        diff = (np.exp(table_data[self._high]).round() - 1).clip(0)
        low_column = table_data[self._low]

        if pd.api.types.is_datetime64_ns_dtype(low_column):
            diff = pd.to_timedelta(diff)

        table_data[self._high] = (low_column + diff).astype(self._dtype)

        return table_data

[npatki]

Hi! Thanks for catching this. I'm able to reproduce and it is, indeed a bug. We'll update when we have a fix.

In the meantime, I've had success using the reject_sampling strategy.

date_constraint = GreaterThan(
    low='start_date',
    high='end_date',
    handling_strategy='reject_sampling'
)

[npatki]

Resolved by #476