In the training loop:
print(
f"Epoch {epoch} "
f"Step {cur_step} "
f"Gen_loss: {gen_mean:.3f} "
f"Critic_loss: {crit_mean:.3f}"
)
returns negative losses for both generator and critic!
In DCGAN, however, we never got into such negative values.
- Is it something wrong occurring in the training loop?
- What is the practical explanation for negative in general in the context of GANs?
You list this as GANs C1 W1 in the title, but WGAN is Week 3, right?
Well the whole point of Wasserstein Loss is that it is a very different approach to defining the Loss Function, which of course affects the training and back prop. And one of the things that is different about it is that here the loss function is not constrained to be positive. This was covered in some detail in the lectures for this week. You might want to watch them again and hear what Prof Zhou says about all this.
1 Like
Yes, sorry it was c1_w3 indeed.
I notice that W-loss can return any value from the lecture slide. However, my point is to clarify, whether or not this loss must ultimately converge to some small value, probably close to zero, or it should never be expected as such!
