Unsupervised Representation Learning with Deep Convolutional Generative Adversarial networks (Python)
In this project, we implemented two different Generative Adversarial Nets (DCGAN and WGAN), then train them on both MNIST and SVHN datasets. We evaluate GANs by tuning different sets of hyper-parameters and comparing generated images with real images. In the end, we find that both DCGAN and WGAN can have a relatively good as the number of epoches increases if the loss of generator and discriminator are close, and we need deeper and more powerful networks with multiple Convolutional layers to generate more complex images in the real world.