Rounding scheme cannot be turned off sometimes

[npatki]

Environment Details

• SDV version: 0.17.0
• Python version: 3.7
• Operating System: Linux

Error Description

The learn_rounding_scheme parameter is available for all single table models. By default it is True.

When I specify the learn_rounding_scheme=False, the rounding scheme still gets enforced sometimes. In particular, it seems to happen when the original data has >14 digits.

Steps to reproduce

from sdv.tabular import GaussianCopula
import pandas as pd
import numpy as np

test_data = pd.DataFrame(data={
    'column': [0.123456789012345]*10
})

model = GaussianCopula(
    learn_rounding_scheme=False)

model.fit(test_data.round(4))
model.sample(num_rows=5)

All outputs have 4 digits even though I explicitly turned the rounding scheme off. Note that this is probably related to #1039

[sharisiri]

Hi there, I'm trying to run the FAST_ML preset in the tabular model tutorial and I'm getting the same error:

__init__() got an unexpected keyword argument 'learn_rounding_scheme'

Is it possible to pass in the enforce_rounding_scheme=False flag somewhere when using the FAST_ML preset?

Btw, huge fan of the project! Thanks for your all your efforts.

[npatki]

Hi @sharisiri, great to hear!

FAST_ML is designed to be lightweight model with no additional arguments (they are all preset). If you'd like to change the flags to customize your model, I'd recommend using the GaussianCopula model instead.

Please note that the parameter is called learn_rounding_scheme (not enforce_rounding_scheme).

from sdv.tabular import GaussianCopula

model = GaussianCopula(learn_rounding_scheme=False)
model.fit(data)
synthetic_data = model.sample(num_rows=100)

[npatki]

Closing this off. The initial example only had a single, constant value. In this case, no rounding scheme is used -- the synthetic data just recreates the same value.

If you try the same example with multiple values, then you can see that the rounding scheme can be turned off.

from sdv.tabular import GaussianCopula
import pandas as pd
import numpy as np

test_data = pd.DataFrame(data={
    'column': [0.123456789012345]*5 + [1.123456789012345]*5
})

model = GaussianCopula(
    learn_rounding_scheme=False)

model.fit(test_data.round(4))
model.sample(num_rows=5)