Sampling with conditions={column: 0.0} for float columns doesn't work

[shlomihod]

Environment Details

• SDV version: 0.11.0
• Python version: 3.8.5
• Operating System: Ubuntu 20.04.1 LTS

Error Description

Trying to apply a condition on sampling with the value 0. for a float column leads to an exception.

Steps to reproduce

from sdv.demo import load_tabular_demo
from sdv.tabular import GaussianCopula

data = load_tabular_demo('student_placements')
data['experience_years'] = data['experience_years'].astype(float)  # for the demonstration
model = GaussianCopula()
model.fit(data)
model.sample(1, conditions={'experience_years': 0.})

ValueError                                Traceback (most recent call last)
<ipython-input-147-1a9ae5e719f8> in <module>
      6 model = GaussianCopula()
      7 model.fit(data)
----> 8 model.sample(1, conditions={'experience_years': 0.})

/opt/conda/lib/python3.8/site-packages/sdv/tabular/base.py in sample(self, num_rows, max_retries, max_rows_multiplier, conditions, float_rtol, graceful_reject_sampling)

/opt/conda/lib/python3.8/site-packages/sdv/tabular/base.py in _conditionally_sample_rows(self, dataframe, max_retries, max_rows_multiplier, condition, transformed_condition, float_rtol, graceful_reject_sampling)

ValueError: No valid rows could be generated with the given conditions.

Ability to contribute

I think that I can fix the issue and I'll create a pull request.
I suspect that it happens because of rtol with zero. The method _filter_conditions checked for < instead of <= (In numpy, <= is used [source]):

if column_values.dtype.kind == 'f':
distance = value * float_rtol
sampled = sampled[np.abs(column_values - value) < distance]
sampled[column] = value

[csala]

Thanks for reporting this @shlomihod
I think you are right about the problem being related to the tolerance, which makes the distance become 0 when the value is 0.

I think that we will revisit this rtol at some point and rather convert it to an atol (absolute tolerance), so the distance does not depend on the actual value, but right now your proposal should work around the problem perfectly fine.

[katxiao]

Hi @shlomihod, we appreciate your interest in contributing and your change looks good! Would you have a chance to make the changes requested? Please let us know if you have any other questions.