From the course: Introduction to Generative Adversarial Networks (GANs)

Final model output

From the course: Introduction to Generative Adversarial Networks (GANs)

Final model output

“

- [Presenter] We've spoken about the architecture but what exactly is the final structure of the output? Once we've gone through the entire flow, all we really need is that generator model afterwards. So the discriminator has definitely done its job and we can save those weights and use it for the future. But what we actually just need are the weights of the generator model because if the model has been trained properly or if the system has been trained properly, it's weights should be quite useful and quite generative. The application we can then use is we can start generating other images from different seeds or even other different sources. If we wanted to, we could keep both models and start retraining on new data sets. So the models can be improved technically by taking in new data and giving it more breadth. So by that I mean the ability to generate new types of images. We could also change the architecture of the model itself to be able to improve and reduce the overall loss. So in essence, the weights of the generator are all we really need to be able to create media, create images. These weights can be improved and by improved I mean the breadth of the image types that can be generated can be broadened with new data and the weights can easily be transferred. So they're essentially just matrices of float numbers and we can use those in other types of models, in other types of systems quite easily as long as they all have the same shape.

Contents