Incorrectly enforced rounding on numerical/float data columns

[dionman]

Error Description

I’m fitting CT-GAN and TVAE on the attached halfmoon dataset (using the default parameters). When sampling from the model I get the float variables discretised to the closest integer. Have you observed behaviour like this in any other settings? It seems to me it’s probably due to some transformation. (I’m using the fit_sample() function as defined in the SDGym repo to fit the model and sample from it). On the other hand, I can get sane output if I instead use fit() and sample() as implemented in ctgan.synthesizers.ctgan.CTGANSynthesizer()
halfmoon.zip

[npatki]

Hi @dionman thanks so much for filing this and providing the data/metadata. Confirmed that I can replicate.

The issue appears to be in how we're learning & enforcing rounding. Only the wrappers in sdv have this feature, so the underlying ML model in ctgan library is unaffected.

From some of my own experiments, I've found the following:

1. There is a learn_rounding_scheme parameter when creating the model. But turning it to False does not help.
2. The rounding is being done by the RDT library right now. There seems to be an issue when the # of digits >14 (see issue)

Workaround

If it doesn't affect quality too much, I'd suggest just rounding your data to the first 14 digits before training the model.

rounded_data = data.round(14)
model.fit(rounded_data)

Fixes

The RDT library should be fixed, we actually plan to stop using RDT for rounding in future SDV releases. We should continue to keep this issue open until we verify a fix.

Separately, I'm not sure why learn_rounding_scheme cannot be turned off right now. I'll file another issue to track this.

[npatki]

Let's keep this one open until we fix & verify the underlying bug all the way throughout the SDV.

[npatki]

Hi everyone, great news! This issue has now been resolved in the new, SDV 1.0 (Beta!) release.

Fore more information and to get started, see the SDV 1.0 demos.