Conditional generation lets you produce examples from the classes you want. For doing that, you need to have a label data-set and somehow pass that class information to both the generator and the discriminator during training. Here you'll learn how to tell the generator which class to produce and how declassifying an example is passing through the discriminator along with the image to be classified. You've seen already with unconditional generation, the generator needs a noise vector to produce random examples. For conditional generation, you also need a vector to tell the generator from which class the generated examples should come from. Usually this is a one-hot vector, which means that there are zeros in every position except for one position corresponding to the class you want. Remember that the noise vector is random values as well, but it doesn't have to zero or one. As an example, let's say all of these different values in the one-hot class vector are different dog breeds; and this second cell here corresponds to a husky. Specifying a position of one here means that you want the generator to produce a husky. If you had a one somewhere else and a zero on the husky cell, then you would want it to produce a different class; for example, a golden retriever. Here the noise vector is the one that adds randomness in the generation, similar to before; to let you produce a diverse set of examples. But now it's a diverse set within the certain class, conditioned on the certain class and restricted by the second class. There's one-hot vector is now what lets you control what class to generate. The input to the generator in a conditional GAN is actually a concatenated vector of both the noise and the one-hot class information. So it's one huge vector. For example, with a class vector here representing a husky breed, the generator will ideally produce a husky dogs from one noise vector. But if you change that noise vector, it should produce a different husky dog; while the class information stays the same. So both are [inaudible] huskies, both are the same class. But how do you ensure the generator actually produces a husky? This seems like magic. Why doesn't it just produce any random dog? Well, that's actually because the discriminator will also be given this class information. The generator, needs to fool the discriminator; so you'll see how this happens. The discriminator in a similar way will take the examples, but now they're paired with the class information as inputs to determine if the examples are either real or fake representations of that particular class; and that's what's key here. For example, with this class of golden retriever, but in image of a beagle dogs. So not a golden retriever being fed in. This discriminated arches actually say this is fake because that is not a real looking golden retriever, even though it's a very real looking image of a beagle. This is just 5 percent real is what it's going to say. As a result, for the discriminator to predict that an example is real, it needs to look like the examples from that class in the training data set. The training data-set is key here because it'll get these real labels next to it. The generator, in order to fool it with a golden retriever, needs to produce a realistic golden retriever. Following the last example with a golden retriever class, the discriminator will learn to predict that the image is more real only if the example looks like a golden retriever. Again, this is because the real images will have the correct class label, and this encourages the generated images to take on the right class as well. Pretty cool, right? Digging a little bit deeper, the input to the discriminator is an image. How do you go about adding that class information? Well, the image is fed in as three different channels, RGB, red, green, blue, or just one channel if it's a gray-scale image. So that's this image component. Then the one-hot class information could also be fed in as additional channels where all the channels take on values of all zeros; so all these white blocks of the same height and width as the image take on values of all zeros. Whereas this black one here will take on values of ones. In contrast to that one-hot vector, these are typically much larger matrices where each channel is full of zeros at every position where it's not that class. There are many other less space consuming ways of doing this, like compressing this information in another format. You can even create a separate neural network head to do that for you, which would be prudent if you had many different classes. But this approach definitely works for the seven or so classes you have here; and the model will learn that information. To sum up in conditional generation, you pass the class information to both models. To the generator it's typically a one-hot vector concatenated with your noise vector to the discriminator when the desired output of the GANs are images, it's one-hot matrices representing the channels. The size of the class vector and the number of extra channels for the class information is just the same as the number of classes you'll be turning on.
