Attached is a basic GAN, referenced to "MNIST FASION".
Since the discriminator also improves during training, I needed an external tool to evaluate the improvement of the genreted images.
To do this I used FID.
https://en.wikipedia.org/wiki/Fr%C3%A9chet_inception_distance
Following are few pictures to illustrate the training progress.
Generated fake images.
Generated Real images.
This GAN suffers from mode-collapse.
By using an Unrolled GAN, we can prevent the generator from optimizing for a single fixed discriminator, since the loss function incorporates not only the current discriminator's classifications, but also future discriminator versions.
So the generator can't over-optimize for a single discriminator.
Clone the repository into a local machine using
git clone https://github.com/anubhavanand1516/MSRF_GAN
These notebooks were developed on pytorch with cuda implementation, as they provide sufficient GPU resources to develop these models. To open these notebooks locally, run
python Teamx.py
Following is the FID graph:
Thanks.