shuyi-bomi / gan

Generative Adversarial Networks

GAN

Generative Adversarial Networks

In GANs, G could be regarded as a team of counterfeiters, trying to make counterfeit currency without being detected. D could be regarded as police, trying to identify counterfeit currency and give corresponding probability being real for each currency. We train them alternatively by Backpropagation and min-max game. Both G and D are multilayer perceptrons originally, whereas DCGAN arises afterwards which implements convolutional neural network (CNN) as G and D. As expected, generative samples are similar to real samples as well as the discriminative model converges to 0.5.

data: The dataset we used is planet data-mat format with dimension 3909 ∗ 4734.

We intend to apply GAN and DCGAN to our data and see how it goes. In order to apply CNN as G and D, which is required in DCGAN, we firstly pad our data and transform to square image(69691). Also, we normalize the data into range [0,1] using Minmax normalization.

But things don't go well. We find loss exploded and will never converge. Because when I apply GANs&DCGAN(design MLP/CNN) to represent G and D separately, I find that it would be easy for D to distinguish fake data/images which G produces from all inputs. So this will lead to training difficulties for G since it doesn’t know how to improve. At this point, D couldn’t converge to 1/2 as well. Like the following figure obtained from Tensorboard visualization tool of loss(DCGAN), we could see that both G and D exploded. Result obtained from GAN is similar.

WGAN

In exploring the unstable behavior of GAN training, researchers demonstrate that Jensen-Shannon divergence between Pr and Pg is unable to provide useful gradient information to perform Gradient descent when Pr and Pg are disjoint[2, 1]. Then, Wasserstein distance is proposed and combined with Kantorovich-Rubinstein duality to generate new loss function. It’s continuous and differentiable almost everywhere under mild assumptions. To satisfy the mild assumptions, we need to clip the weight within [-c, c].

WGAN-GP

However, WGAN still exist problem like training instability. Finally researcher finds out gradients may vanish quickly in original weight clipping. Finally, WGAN-GP apply an alternative way in weight clipping by implementing a k-Lipshitz constraint via gradient penalty.

We tried all these methods and it turned out WGAN-GP with MLP works best. We provide its loss figure obtained from Tensorboard visualization tool for G and D:

So we could see for D it converged to 0.5, which testifies trained G and D work great since 0.5 is the convergence result for D in theoretic proof.

After we obtained the desired network, we then generated/simulated planet images from 'generate.py'. For the following figure, one image contains 16 data(16 ∗ 4761) simulated and we reshape each to 69 ∗ 69 dimension to visualize it. Every 1000 trainings we generated one image. And this is the one that we train 1000000 times.

The picture showed general characteristic for planet data, central value is relatively big and spherical is small.