So, While plotting both the losses in vanilla gan and dcgans, what can we conclude with the graph…? when should we stop our training, since the loss plot is not always decreasing, it shows random behaviour, how should one get to know when is the right time to stop iterations.
I would like to get some insights on this.
Plotting Losses in GANs
Hey Mohit,
Just like what we have done in the past, we can apply Early-stopping and Callbacks in GANs as well. Using these techniques make it a lot easier to analyse our results and stop on the appropriate iteration accordingly. Our loss plot is more or less random (though, decreasing overall) and these methods can better see result-by-result of when the loss stops decreasing any further.
Moreover, a very important thing to note here is that there is nothing like overfitting in a GAN because it’s an unsupervised learning technique. So, when we talk about stopping midway through the iterations it is because of no improvement in the model, i.e., a stationary curve.
I hope this clears your doubt.
I completely understand over fitting part.
But in the first part when u r saying this, I cannot understand which loss we should consider for stoping the training.
Since it is a minimax game, if loss for one model decreases, the loss for other will increase.
Also, Here generator loss is constantly increasing, so is this model performing bad? (This is DCGAN MNIST)
In GANs do we get very different types of graphs everytime??
I was talking about the Generator Loss, because we want it to win the min-max game in the end.
A general appropriate curve in DC GANs:
That’s what it seems. Though it may give descent enough results on the MNIST Dataset. Is that the case?
Yes, more fluctuating losses and curves than the other models we have seen.
Yes, though the loss is increasing on generator part, it gives good results on MNIST. (Bcoz it is very simple dataset)
The image you have just shown, I guess this is the ideal error plot, and I never got this type of plot. 
So, where to improve in that case, the architecture should be change or something else?
Also, I’ve heard that GANs are difficuilt to train is it because of minimax game? and again same ques. if the error plot is not appropriate what changes should I make.
In the same context, I would like you to guide me on simpson dataset, model is not performing well.
what changes can I make here?
https://www.kaggle.com/muniyal/simpson-faces/
I hope I’ve cleared your doubt. I ask you to please rate your experience here
Your feedback is very important. It helps us improve our platform and hence provide you
the learning experience you deserve.
On the off chance, you still have some questions or not find the answers satisfactory, you may reopen
the doubt.