Implementing an early stopping parameter about ctgan HOT 5 CLOSED

Since this is a GAN, does it make sense to stop early when the loss is stable? If both the generator and discriminator are improving at the same rate, shouldn't it stay stable?

from ctgan.

In my forked ctgan repo, i added a couple new features that were of interest to me including logging, gpu_stats, early_stopping and patience.

any feedback appreciated.
e.g.
ctgan.fit(data, discrete_columns, epochs=500, early_stopping=True, gpu_stats=False, patience=40, logging=True)

https://github.com/oregonpillow/CTGAN/blob/master/ctgan/synthesizer.py

from ctgan.

I implemented early stopping as well here https://github.com/Diyago/GAN-for-tabular-data/blob/master/ctgan/synthesizer.py#L262

from ctgan.

@oregonpillow @Diyago I came upon this while reviewing old issues. Since it seems that there are already a couple of implementations out there, please feel free to open a PR against this issue to contribute your versions! We can then discuss a bit more the details over the proposed code.

from ctgan.

As @Baukebrenninkmeijer mentioned, using the discriminator/generator losses to decide whether a GAN has converged is an open research question, so we won't be implementing it at this time.

from ctgan.