[BUG] Incorrect computation of the supervised loss in train_supervisor of TimeGAN. about ydata-synthetic

[donseca]

As far as I've understood, the first two steps pertain to the training of the autoencoding components (embedding, recovery) in terms of their ability to encode into latent space and learning temporal dependencies. Thus some sort of pre-training (Chapter 4.1 in the original paper)

What remains unclear to me is why the g_loss_s is included as is in the train_supervisor function, but the generator_loss_supervised is multiplied by 0.1 in the train_embedder function and by 100 in train_generator (while taking the square root). Any insights on this?

Agree on the change of the g_loss_s (and this should also be valid for generator_loss_supervised I guess).

[gmbss0]

I believe that all hyperparameters are purely empirically identified parameters that seem to perform well on the datasets used by the authors. The relative weight (0.1 or 100*sqrt) may be used to balance the individual loss components. g_loss_s is added to two cross entropy components (G_loss_U and G_loss_Ue) for the generator loss, as opposed to the embedding loss. Different magnitudes of the loss components may require different scaling of g_loss_s. At least this is my intuition...Hope this helps!

[donseca]

Thanks! Good point on the cross entropy...I guess I'll leave it at intuition for now.