Negative losses about ctgan

[k15z]

Here are the two loss functions:

loss_d = -(torch.mean(y_real) - torch.mean(y_fake))
loss_g = -torch.mean(y_fake) + cross_entropy

The generator loss consists of two components, the cross_entropy which can never be negative and -torch.mean(y_fake) which can be negative if the discriminator assigns a high score to the fake/synthetic data.

Assuming that the discriminator has assigned a high score to the fake/synthetic data, then the only way that the discriminator loss could be negative is if it assigns an even higher score to the real data.

Therefore, this suggests that the discriminator is, on average, assigning higher scores to real data than fake data which means it's doing its job.

Without more information, I don't see anything to be concerned about here.

[k15z]

Re: the random seed, this library use PyTorch so should be able to just add:

import torch
torch.manual_seed(42)

before you call sample to get reproducible samples.

[adityalahiri]

Also, can you tell where can we set the random seed to get reproducible samples?

[adityalahiri]

Thank you so much for both the detailed responses. Closing issue now :)