Trying to modify batch size in source code (synthesizer.py) from 500 to 16 and getting 'Assertion error' about ctgan

[k15z]

The discriminator uses the PacGAN framework to prevent mode collapse with pack size 10. @csala The docstring can be updated to clarify this restriction, the batch size must be divisible by the pack_size.

[Baukebrenninkmeijer]

In my tests, batch size needs to be a multiple of 10. I don't recall why this issue was again specifically.

[elisim]

In addition, batch_size option should be moved to fit function instead of in the CTGANSynthesizer ctor.
Now:

ctgan = CTGANSynthesizer(batch_size=128)

How it should be (IMHO):

ctgan.fit(data, discrete_columns, epochs=3, batch_size=128)

[csala]

The discriminator uses the PacGAN framework to prevent mode collapse with pack size 10. @csala The docstring can be updated to clarify this restriction, the batch size must be divisible by the pack_size.

Thanks @k15z, I opened an issue on SDV to add this to the CTGAN model documentation, so I'm closing this one for now.

In addition, batch_size option should be moved to fit function instead of in the CTGANSynthesizer ctor.
Now:

ctgan = CTGANSynthesizer(batch_size=128)

How it should be (IMHO):

ctgan.fit(data, discrete_columns, epochs=3, batch_size=128)

Thanks for the suggestion @elisim, but this was discussed a long time ago and we decided to move in another direction, so we will try to keep the hyperparameter organization as it is right now.

Also, bear in mind that we are pushing it even further when using CTGAN from SDV, where all the hyperparameters and setup arguments are passed when the model is created and the only input argument for the fit method is the data to which the model is fitted.