That’s a distinct problem. When training our GAN, we completely ignore the training labels. If you’d like to classify the output images of the generator, you would have to firstly train a different algorithm, e.g. a neural network, to classify MNIST images. This trained network will then be able to classify the generator’s output images.