It is very hard to interpret loss in huge variety of situations. GANs is one of these cases. You hardly can just look at G and D losses and say, yeah, this model is great.

But you need to estimate model. So, I have very simple solution. Just generate a batch of images and plot them every N epochs. Also save model weights. If the quality of images is good, then stop training and use model weights from last checkpoint.

Another thing can use Early Stopping Callback idea. If not improve for N epochs - stop.

Also, during experience of many researchers some common bounds for dcgan were estimated: G_loss 2 and D_loss 0.1.

By the way, training process for GANs is very unstable. There are some technics to stabilize model training.

So, I highly recommend visual estimation approach :)