What I know so far in DCGAN is that a discriminator is trained using the labeled data (so maybe that occurs before training the generative model). Also, I know that there is race between the generator and the discriminator, so maybe training occur online. So I have some concerns here:

  1. How many outputs the discriminator should have (Is it one output that describes the probability, ex: P(x))?
  2. How do we chose its output when feeding fake data vs. real data?
  3. Is the discriminator trained before using it with the DCGAN or the training is done online (It is mentioned in the Original Paper: Generative Adversarial Nets https://arxiv.org/pdf/1406.2661.pdf, that the whole network is trained using the back propagation), hence I think its online?

Any help is much appreciated!!


In normal GANs, there are no labels, the training is completely unsupervised.

The role of the discriminator is to tell apart samples generated by the generator from those taken from the training dataset. The training dataset is just a bunch of images. The discriminator is trained to output 0 for data generated by the generator (i.e. fake data) and 1 for real data (so the discriminator has a single output). This should answer points 1 and 2.

The training of discriminator and generator takes place alternatively in a loop: first we train the discriminator, then the generator, then the discriminator again, etc. It is possible (and common) to train the discriminator a few times per each time we train the generator. This should answer point 3.

It is also possible to use labels, but not in the way you were suggesting. When labels are used, we have Conditional GANs (https://arxiv.org/abs/1411.1784). In this case, the label is supplied as input to both the generator and the discriminator. The generator has to generate data that is associated to the supplied label. The discriminator has to tell apart fake data from real data, given the label.

| improve this answer | |
  • $\begingroup$ Thank you so much for the nice answer. I am going to post a new question regarding the loss function used with GANs. $\endgroup$ – I. A Aug 8 '17 at 15:17

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.