Pytorch implementation of "Self-Attention Gan" based on Self-Attention Generative Adversarial Networks
This is an attempt to build a SaGan, basically a Gan using Spectral Normalization and Attention Layers in both the discriminator and generator.
What I've uploaded here differs a bit from the paper, this version has no residual blocks, and uses deconvolutions and stride 2 convolutions as opposed to up-sample and down-sample blocks.
I've tested this using the GeoPose3k Dataset.
Usage instructions found here: user manual page.
