Allow me to access loss values for GAN-based synthesizers

[npatki]

Problem Description

The GAN-based synthesizers work by iterating through many different epochs and optimizing a loss function. Currently, all of our core ML models store a DataFrame with the loss values per epoch under a parameter called .loss_values.

See:

• CTGAN #298
• CTGAN #300
• DeepEcho #80

Now that the underlying ML models are storing loss values, they can be exposed through the SDV synthesizers using a simple function.

Expected behavior

Each of the GAN-based synthesizers should include a function get_loss_values() that returns a copy of the loss values. If the synthesizer has not yet been fit, then this function could raise an error.

# available for CTGAN, TVAE, CopulaGAN and PAR
synthesizer.fit(data)
loss_values_df = synthesizer.get_loss_values()

# if a synthesizer has not yet been fit
unfit_synthesizer.get_loss_values()
SynthesizerProcessingError: Loss values are not available yet. Please fit your synthesizer first.

For multi-table synthesizes, it's possible to have a GAN-based model for single tables. (In both HSA and Independent). These synthesizers should then be able to return the loss values given the table name.

# for HSA and Independent, it's possible to use GAN-based synthesizers for individual table modeling
hsa_synthesizer.set_table_parameters(
  table_name='users',
  table_synthesizer='CTGANSynthesizer')

# error if you haven't fit yet
users_loss = hsa_synthesizer.get_loss_values(table_name='users')
SynthesizerProcessingError: Loss values are not available yet. Please fit your synthesizer first.

# after fitting, it should be possible to access the loss values for those tables
hsa_synthesizer.fit(data)
users_loss = has_synthesizer.get_loss_values(table_name='users')

# if you try to get loss values from something that's not GAN-based, then there should be an error
txns_loss = hsa_synthesizer.get_loss_values(table_name='transactions')
SynthesizerInputError: Loss values are not available for table 'transactions' because the table does not use a GAN-based model.

Additional context

In order to do this issue, we'll likely need to make a CTGAN release and a DeepEcho release first. (The PRs are merged, but the releases haven't been made, so the SDV won't be able to access loss values.)